Applying Single-Molecule
Localisation Microscopy
to Achieve Virtual Optical Sectioning
and Study T-Cell Activation
Matthieu Grégoire Simon Palayret
Department of Chemistry
University of Cambridge
This dissertation is submitted for the degree of
Doctor of Philosophy
Gonville & Caius College 2015

To my beloved family, heroes who built the hull.
To the joyful crew of friends, fellows of a life-long cruise.
“Le savant n’étudie pas la nature parce que cela est utile ; il l’étudie parce qu’il y prend
plaisir et il y prend plaisir parce qu’elle est belle. (. . . ) nous nous complairons tantôt à
suivre la course gigantesque des astres, tantôt à scruter avec le microscope cette
prodigieuse petitesse qui est aussi une grandeur. . . ”
Science et méthode, H. Poincaré, 1908.
(“The scientist does not study nature because doing so is useful; he studies it because he
enjoys doing so and he enjoys doing so because nature is beautiful. (. . . ) we may
sometimes revel in tracing the gigantic course of the stars, sometimes in scrutinizing,
through a microscope, this prodigious littleness, which is grand too. . . ”)

Declaration
This dissertation summarises research carried out in the laboratory of Professor David Klen-
erman at the Department of Chemistry, University of Cambridge, between October 2011
and February 2015. The work described in the dissertation is my own. When work was
performed in collaboration with others, it has been explicitly specified in the text: any ex-
periments and protocols performed by others are denoted in the methods sections and/or
relevant chapters and any data provided by others is noted in the relevant chapters and fig-
ure legends. This thesis has not, either in part or as a whole, been submitted for a degree,
diploma, or other qualification at any other university. The length of this dissertation does
not exceed the word limit of the degree committee (60,000 words), including appendices,
bibliography, footnotes, tables and equations.
Matthieu Grégoire Simon Palayret
2015

Acknowledgements
Although a statement of originality must precede any PhD thesis, this work would have been
an impossible achievement without the support I received over the past three years.
First of all, I would like to thank my supervisor, Professor David Klenerman: Dave,
your humble, positive and proactive attitude to tackle scientific challenges and ideas have
inspired this work. I am deeply grateful for your generosity in support, encouragement and
ever-flowing ideas that have continuously fuelled and channelled my scientific curiosity.
All the T cell work in this thesis was done in close collaboration with the lab of Pro-
fessor Simon Davis at the University of Oxford: Simon, I would like to thank you for your
communicative enthusiasm that never despaired. I am also grateful for the biological focus
you brought to my work and your financial support during the last months of my PhD.
Many ideas in this thesis originated and were discussed during regular meetings with
the labs of Professor Ernest Laue at the Department of Biochemistry and Professor Anthony
Carr at the University of Sussex. These meetings were precious opportunities to get external
feedback and extend my curiosity to different perspective and exciting research. I would like
to thank you both to have maintained such a collaborative dynamic.
The Wellcome Trust must be thanked for both its generous funding without which this work
could not have been undertaken, but also for giving me the opportunity to briefly explore
three research topics in immunology and get priceless experience in T cell biology and
mathematical modelling before starting my PhD.
This work has truly benefited from the knowledge and the abilities of many scientists I had
the pleasure to work with over the past three years in Cambridge:
Dr Steven Lee: Steve, from my very first day in the lab I tried to follow your successful
steps in the windings of the basement. Many thanks for the infinite time you patiently spent
building Arabidopsis with me. This may not be the first thesis of the Lee lab, may it be its
first thèse!
viii
Dr Kristina Ganzinger: Kristina, I could not decide where your name would fit best in
this list of acknowledgements. You truly deserve my deepest thanks for your unconditional
support, both scientific and emotional, from the beginning until the very end of this adven-
ture. You are both a precious colleague and a critical friend. Thank you for your advice, the
many coffees and discussions we had, the travels we shared and your final proofreading. I
wish you the very best of luck on the next step of your journey!
Meike Aßmann: Meike, working with you on CD28 super-agonism has been intellectu-
ally motivating and rewarding. I really enjoyed constructively confronting ideas and plan-
ning focussed experiments with you. Although I will leave Cambridge, super-agonism and
the kinetic-segregation model will keep populating my thoughts at night! Thanks for careful
proofreading.
Elizabeth Huang: Thank you for your continual help with the TCR project, through its
many iterations!
Dr James McColl: James, thanks for your cheerful welcome in the Klenerman group
and your foolproof patience with my countless questions.
Dr Julien Godet: Many thanks for your discreet but efficient work on the light-sheet. I
must also thank you for your career advice which are leading me to Bordeaux. I have been
very honoured to share my nickname with such another ‘Frenchman’.
There is not enough room here to thank all the members of the Klenerman group that I would
like to personally thank. Countless lunches and coffees, lab discussions, code debugging and
instrument troubleshooting, have put rhythm and motivation in my life in the dark labs in
the last three years. Particular thanks to Rohan, Kristina and Laura who cosseted me during
the writing-up, and to the members of the lab fine dining society.
The end of this work coincides with the end of my time in Cambridge. Looking back at the
last four years, I realise how lucky I have been to cross the path of so many (new) friends.
A heartfelt thank you to:
Ana for our first awkward meeting and the inseparable four years that followed; Maddie
& Ross for enduring my chaotic presence in St Andrews Road; Daniel, Roeland, Elsa &
Marius, Jude & Tom, Daniel, Ben, Mike, Olga & Matteusz, Emma, Lala, Adèle, . . . and the
mysterious M* for long nights, friendship with style and trips to Andorra and Corsica; Dr
John Casey for intense and rich debates, fine dinners and precious friendship;
Ro’, because your British irony and Rohan’s dynamics are terra incognita in Bordeaux;
because you are a rare example of unpretentious precision; because these three years chan-
ged an indecisive ‘vous’ into a definite ‘tu’; (and thanks for proofreading!)
ix
Benoît and Jérôme for bridging Cambridge to Paris; Delphine and Élise for the provi-
sions of motivating chocolate; Clémence, Pierre and Charles for their constancy and fidelity;
Thierry Hubert for his joyful example.
The final thanks are reserved to my dearest family: my parents, Anne and Laurent, for
their attentive, constant and strong support for all my projects; my siblings, Basile for his
enthusiasm when I come back, Madeleine for frequently taking news and taking me abroad,
Ambroise for sharing his joys and sorrows, Julie and her whole family which expanded three
times during this work to my greatest delight.

Abstract
Single-molecule localisation microscopy (SMLM) allows imaging of fluorescently-tagged proteins
in live cells with a precision well below that of the diffraction limit. As a single-molecule technique,
it has also introduced a new quantitative approach to fluorescence microscopy.
In the Part A of this thesis, the design and building of three SMLM instruments, the implemen-
tation of a custom-developed image analysis package and the characterisation of the photo-physical
properties of the photo-activable fluorescent protein used in this thesis (mEos), are discussed. Then, a
new post-processing method for SMLM analysis is characterised: axial optical sectioning of SMLM
images is demonstrated by thresholding fitted localisations using their fitted width and amplitude to
reject fluorophores that emit from above or below a virtual ‘light-sheet’, a thin volume centred on
the focal plane of the microscope. This method provides qualitative and quantitative improvements
to SMLM.
In the Part B of this thesis, SMLM is applied to study T cell activation. Although the T cell recep-
tor plays a key role in immunity, its stoichiometry in the membrane of resting T cells is still a matter
of debate. Here, single-molecule counting methods are implemented to compare the stoichiometry
of TCRs fused with mEos2 in resting T cells to monomeric and dimeric controls. However, because
of the stochasticity of mEos2 photo-physics, results are inconclusive and new counting techniques
based on structural imaging are discussed. In addition to TCR triggering, T cells require the co-
stimulatory triggering of the CD28 transmembrane receptor to become fully activated. However,
some immobilised anti-CD28 antibodies, referred to as super-agonists (SA), can directly activate
T cells without triggering the TCR. In this thesis, single-molecule tracking techniques are used to
investigate the molecular mechanism of CD28 super-agonism in live T cells. The results indicate
that the diffusion of CD28 is slowed by SA binding. This effect is further discussed in light of the
kinetic-segregation model proposed for TCR triggering.
Quantitative SMLM as implemented and further developed in this work offers new tools to in-
vestigate the molecular mechanisms initiating T cell activation, ultimately facilitating the discovery
of novel approaches to target these pathways for therapeutic purposes.

Contents
Contents xiii
1 Thesis overview 1
2 Introduction to single-molecule localisation microscopy 5
2.1 Fluorescence microscopy . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.1.1 Fluorescence and Stokes’ shift . . . . . . . . . . . . . . . . . . . . 5
2.1.2 Abbe’s diffraction limit . . . . . . . . . . . . . . . . . . . . . . . . 7
2.1.3 An overview of far-field optical super-resolution techniques . . . . 8
2.2 Single-molecule localisation microscopy . . . . . . . . . . . . . . . . . . . 10
2.2.1 Principle: spreading spatial density in time . . . . . . . . . . . . . 10
2.2.2 Single-molecule photo-switching . . . . . . . . . . . . . . . . . . 12
2.2.3 Fluorophores for SMLM imaging . . . . . . . . . . . . . . . . . . 14
2.3 A quantitative approach to microscopy . . . . . . . . . . . . . . . . . . . . 16
2.3.1 Local diffusion analysis and detection of rare events . . . . . . . . 16
2.3.2 Single-molecule counting and cluster analysis . . . . . . . . . . . . 17
2.3.3 Single-molecule multi-dimensional imaging . . . . . . . . . . . . . 17
3 Building and characterising a single-molecule super-resolution platform 21
3.1 Illumination path . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
3.2 Imaging path . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
3.2.1 Stability of the frame . . . . . . . . . . . . . . . . . . . . . . . . . 28
3.2.2 The double-helix point-spread function . . . . . . . . . . . . . . . 30
3.2.3 Two-colour imaging: registration of two channels . . . . . . . . . . 34
3.3 Which labelling strategy? . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
3.3.1 Labelling requirements and strategy . . . . . . . . . . . . . . . . . 41
3.3.2 Ensemble characterisation of the dark states of mEos . . . . . . . . 43
xiv Contents
3.3.3 Single-molecule characterisation of the dark states of mEos . . . . 46
3.3.4 Optimisation of mEos photo-activation . . . . . . . . . . . . . . . 49
3.4 Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
3.4.1 Fitting routine . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
3.4.2 Post-processing . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
3.4.3 What is precision . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
3.5 Automation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
4 The virtual-‘light-sheet’ 69
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
4.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
4.2.1 Axial variations of the point-spread function . . . . . . . . . . . . 71
4.2.2 Sub-diffraction beads under low illumination mimic idealised single
fluorophores . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
4.2.3 Building of the virtual-‘light-sheet’ . . . . . . . . . . . . . . . . . 76
4.2.4 Robustness of the calibration . . . . . . . . . . . . . . . . . . . . . 79
4.2.5 Improvement in contrast and quantification using vlsSMLM . . . . 83
4.3 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
4.3.1 A quantitative optical sectioning method for existing 2D SMLM
nanoscopes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
4.3.2 vlsSMLM only requires a simple calibration step . . . . . . . . . . 87
4.3.3 On thresholding . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
4.3.4 vlsSMLM, a user-friendly ImageJ plugin . . . . . . . . . . . . . . 89
5 Introduction to the molecular basis of T cell activation 93
5.1 A brief introduction to the vertebrate immune system . . . . . . . . . . . . 93
5.2 The T cell, a key player of the adaptive immune system . . . . . . . . . . . 95
5.3 The complex T cell receptor . . . . . . . . . . . . . . . . . . . . . . . . . 97
5.4 T cell activation at the immunological synapse . . . . . . . . . . . . . . . . 99
5.5 Three molecular models for TCR triggering . . . . . . . . . . . . . . . . . 102
6 Counting T cell receptors 109
6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
6.1.1 Organisation of the TCR on a resting T cell . . . . . . . . . . . . . 110
6.1.2 Counting single-molecule localisation microscopy . . . . . . . . . 112
6.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
Contents xv
6.2.1 Yeast counting controls . . . . . . . . . . . . . . . . . . . . . . . . 115
6.2.2 Studying the dependence of counting analyses on localisation den-
sities using simulations . . . . . . . . . . . . . . . . . . . . . . . . 129
6.2.3 State of the TCR . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
6.2.4 Two-colour strategy . . . . . . . . . . . . . . . . . . . . . . . . . 143
6.3 Conclusion and future experiments . . . . . . . . . . . . . . . . . . . . . . 145
7 CD28 super-agonism 147
7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148
7.1.1 CD28, a co-stimulatory membrane protein . . . . . . . . . . . . . . 148
7.1.2 Super-agonism . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
7.1.3 A structural difference . . . . . . . . . . . . . . . . . . . . . . . . 151
7.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154
7.2.1 Super-agonistic antibodies slow CD28 down . . . . . . . . . . . . 154
7.2.2 Super-agonist activation requires divalent engagement of CD28 . . 164
7.2.3 Effect of different coverslip antibody-coatings . . . . . . . . . . . . 167
7.2.4 CD28-bound antibodies . . . . . . . . . . . . . . . . . . . . . . . 169
7.2.5 Variation of the height separating the plasma membrane from the
coated surface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 172
7.2.6 Evidence for a size-dependent segregation? . . . . . . . . . . . . . 174
7.3 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177
7.3.1 Clusters of localisations: protein clusters or trajectories? . . . . . . 177
7.3.2 Choice of the diffusion analysis . . . . . . . . . . . . . . . . . . . 179
7.3.3 Physical origin of the multiple diffusing populations . . . . . . . . 180
7.3.4 Differences of affinity and kinetic rates . . . . . . . . . . . . . . . 182
7.3.5 Requirements for cross-linking of CD28 . . . . . . . . . . . . . . . 184
7.3.6 Importance of the formation of a close-contact zone . . . . . . . . . 185
7.3.7 From correlation to causality? . . . . . . . . . . . . . . . . . . . . 186
7.3.8 Possible evidence for segregation . . . . . . . . . . . . . . . . . . 188
7.3.9 Temporal segregation in light of the kinetic-segregation model . . . 189
7.3.10 Conclusions and future experiments . . . . . . . . . . . . . . . . . 191
8 Materials and methods 195
8.1 Materials and methods common to all projects . . . . . . . . . . . . . . . . 195
8.1.1 Microscope set-up . . . . . . . . . . . . . . . . . . . . . . . . . . 195
8.1.2 General imaging protocol . . . . . . . . . . . . . . . . . . . . . . 197
xvi Contents
8.1.3 SMLM data representation . . . . . . . . . . . . . . . . . . . . . . 198
8.1.4 Imaging fluorescent beads and isolated fluorophores . . . . . . . . 199
8.2 Materials and methods for Chapter 3 . . . . . . . . . . . . . . . . . . . . . 199
8.2.1 Double-helix microscope . . . . . . . . . . . . . . . . . . . . . . . 199
8.2.2 Automation of SMLM imaging . . . . . . . . . . . . . . . . . . . 201
8.3 Materials and methods for Chapter 4 . . . . . . . . . . . . . . . . . . . . . 202
8.3.1 Bead calibration experiments . . . . . . . . . . . . . . . . . . . . . 202
8.3.2 Purification of mEos3.1 . . . . . . . . . . . . . . . . . . . . . . . 203
8.3.3 Imaging separated fluorophores . . . . . . . . . . . . . . . . . . . 204
8.3.4 S. pombe experiments . . . . . . . . . . . . . . . . . . . . . . . . 204
8.3.5 Embryonic stem cells . . . . . . . . . . . . . . . . . . . . . . . . . 205
8.4 Materials and methods for Chapter 6 . . . . . . . . . . . . . . . . . . . . . 206
8.4.1 S. pombe cell lines . . . . . . . . . . . . . . . . . . . . . . . . . . 206
8.4.2 T cell lines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 206
8.4.3 SMLM imaging . . . . . . . . . . . . . . . . . . . . . . . . . . . . 207
8.5 Materials and methods for Chapter 7 . . . . . . . . . . . . . . . . . . . . . 208
8.5.1 Molecular Cloning . . . . . . . . . . . . . . . . . . . . . . . . . . 208
8.5.2 Stable transfection with lentivirus . . . . . . . . . . . . . . . . . . 209
8.5.3 Cell Culture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 209
8.5.4 Sample preparation . . . . . . . . . . . . . . . . . . . . . . . . . . 210
8.5.5 IL-2 in vitro stimulation assay . . . . . . . . . . . . . . . . . . . . 211
8.5.6 Generating Atto655-labelled Anti-murine-CD45 Fabs . . . . . . . 211
8.5.7 Single-particle tracking PALM imaging . . . . . . . . . . . . . . . 212
References 215
Appendix A Supplementary Figures and Tables 243
Appendix B Supplementary Movies and Software 247
Abbreviations
Acronyms / Abbreviations
KD Dissociation constant
na Numerical aperture (of an objective lens)
2D Two dimensional (x,y)
3D Three dimensional (x,y,z)
E. coli Escherichia coli
S. pombe Schizosaccharomyces pombe
Ab Antibody
APC Antigen-presenting cell
ATP Adenosine triphosphate
BALM Binding-activated localisation microscopy
CA Conventional antibody
CBC Coordinate-based co-localisation
CD Cluster of differenciation
cSMAC Central SMAC
CTLA-4 Cytotoxic T-lymphocyte-associated protein 4
DBSCAN Density-based spatial clustering of applications with noise algo-
rithm
xviii Abbreviations
DH-PSF Double-helix PSF
DNA Deoxyribonucleic acid
dSMAC Distal SMAC
dSTORM Direct STORM
DySCo Dynamic single-molecule co-localisation
EMCCD Electron-multiplying charge-coupled device
eYFP Enhanced yellow fluorescent protein
F-actin Filamentous actin
Fab Fragment antigen-binding region of an antibody
Fc region Fragment crystallisable region
FLIM Fluorescence lifetime imaging microscopy
fPALM Fluorescence PALM
FRET Förster resonance energy transfer
FWHM Full width at half maximum
GFP Green fluorescent protein
HILO Highly inclined and laminated optical sheet
hsPALM High-speed PALM
ICAM-1 Intercellular adhesion molecule 1
ICOS Inducible T-cell costimulator
IgG Immunoglobulin G
IL-2 Interleukin 2
iPALM Interferometric PALM
ITAM Immunoreceptor tyrosine-based activation motifs
Abbreviations xix
JD Jump-distance
KS Kinetic-segregation
LAT Linker of activated T cells
LFA Lymphocyte function-associated antigen
mAb Monoclonal antibody
MAP Mitogen-associated protein
MBIC Membrane-bound immune complex
MHC Major histocompatibility complex
MHC Major histocompatibility complex
MSD Mean square displacement
NeNA Nearest-neighbour-based approach
PA-mCherry1 Photo-activable mCherry 1
PAINT Point accumulation imaging in nanoscale topography
PALM Photo-activated localisation microscopy
PBS Phosphate-buffered saline solution
PC-PALM Pair-correlation PALM
PD-1 Programmed cell-death protein 1
PDMS Polydimethylsiloxane
PKCθ Protein kinase C θ
pMHC Complex formed of a peptide presented in the groove of an MHC
PSCFP2 Photo-switchable cyon fluorescent protein 2
PSF Point-spread function
pSMAC Peripheral SMAC
xx Abbreviations
RESOLF Reversible saturable optical fluorescence transitions
RNA Ribonucleic acid
SA Super-agonist antibody
sCMOS Scientific complementary metal-oxide semi-conductor
SDS-PAGE Sodium-dodecyl-sulfate polyacrylamide-gel electrophoresis
SIM Structured illumination microscopy
SMAC Supra-molecular activation complex
SMACM Single-molecule active control microscopy
SMLM Single molecule localisation microscopy
SNR Signal to noise ratio
SPIM Selective plane illumination microscopy
sptPALM Single particle tracking PALM
SSIM Saturating SIM
STED Stimulated emission depletion microscopy
STORM Stochastic optical reconstruction microscopy
TCCD Two-colour co-localisation detection
TCR T cell receptor complex
TEM Transmission electron microscopy
TIRF Total internal reflection fluorescence
TMR Tetramethylrhodamine (orange dye)
uPAINT Universal PAINT
UV Ultraviolet
vls Virtual ‘light-sheet’
ZAP-70 ζ -chain-associated protein 70
Chapter 1
Thesis overview
Single-molecule localisation microscopy (SMLM) allows imaging of fluorescently-tagged
proteins in live cells with a precision well below that of the diffraction limit. It has also
introduced a new quantitative approach to fluorescence microscopy by enabling the tracking
and counting of single molecules.
In the first part of this thesis (Part A), the design and building of three SMLM instru-
ments, the implementation of a coherent custom-developed image analysis package and the
characterisation of the photo-physical properties of the photo-activable fluorescent protein
used in this thesis, i.e. mEos, are discussed (Chapter 3).
Then, in Chapter 4, a new post-processing method for SMLM analysis is developed and
characterised. Axial optical sectioning of SMLM images is demonstrated by making use of
fitting information that is usually discarded to reject fluorophores that emit from above or be-
low a virtual ‘light-sheet’ (vls), a thin volume centred on the focal plane of the microscope.
A universal easy-to-use ImageJ plug-in is implemented to quickly analyse a calibration sam-
ple to define and use such a vls. This optical sectioning of super-resolution images is achie-
ved by applying well-characterised width and amplitude thresholds to diffraction-limited
spots that can be used to tune the thickness of the vls. This allows qualitative and quanti-
tative imaging improvements: by rejecting out-of-focus fluorophores, the super-resolution
image gains contrast and local features may be revealed; by retaining only fluorophores
close to the focal plane, vlsSMLM improves the probability that all emitting fluorophores
will be detected, fitted and quantitatively evaluated.
In the second part of this thesis (Part B), SMLM is applied to study T cell activation.
The human body is constantly challenged with ever-mutating pathogens. The complex and
well-balanced capacity of the adaptive immune system to distinguish self from non-self is
2 Thesis overview
crucial for survival. One key player in this recognition mechanism is the T cell receptor
(TCR). The exact mechanism of TCR triggering, from the binding of the TCR ligand, to the
transduction of the signal across the plasma membrane, is however still an object of contro-
versy. Three main non-exclusive models have been proposed: aggregation, segregation and
conformational change of the TCR. In light of the aggregation model, the natural organisa-
tion of TCRs at the membrane of the resting T cell has been widely discussed. Therefore,
in Chapter 6, single-molecule counting methods are developed to try to determine the sto-
ichiometry of TCRs fused to mEos2 and expressed at low levels on a resting T cell fixed
in solution. Monomeric and dimeric proteins are expressed in both yeast and T cell mod-
els as quantitative controls. Although a variety of analyses are investigated, results are not
conclusive: the photo-physical properties of mEos2 are too stochastic for single-molecule
counting by sequential and irreversible photo-activation. New counting techniques based
on structural imaging methods are discussed to further study the natural state of the TCR,
before its triggering.
In addition to TCR triggering, T cells require a co-stimulatory signal to become fully
activated. This second signal is initially delivered by a transmembrane receptor, CD28,
expressed on the T cell. CD28 can also directly activate T cells without requiring TCR trig-
gering: resting T cells incubated with defined anti-CD28 monoclonal antibodies, referred
to as super-agonists (SAs), and immobilised on a surface, proliferate and release IL-2. In
2006, a tragic clinical trial also dramatically revealed the potency of CD28 SA in vivo. This
super-agonism effect is specific to antibodies directed against epitopes of the receptor close
to the plasma membrane. In Chapter 7, single-molecule tracking techniques are used, in
parallel with functional IL-2 assays, to investigate the mechanism of CD28 super-agonism
at the single-molecule level in live T cells. Results indicate a specific slowing down of
CD28 upon SA binding. Although critical, antibody cross-linking does not fully account
for this SA-specific slowing down effect. The possibility that avidity of the SA is increased
by geometric constraints imposed by its epitope location is discussed. Finally, simultaneous
imaging of both CD28 and the large phosphatase CD45 could suggest some size-dependent
segregation at the membrane of the T cell. Additional experiments are proposed to chal-
lenge a CD28 triggering model inspired by the kinetic-segregation model proposed for TCR
triggering.
Finally, in the last part of this thesis (C), material and methods used for all experiments
presented in this work are summarised.
PART A

Chapter 2
Introduction to single-molecule
localisation microscopy
This chapter provides the technical context for the work presented in this thesis, and es-
pecially in the next two chapters (Chapters 3 and 4), by introducing fluorescence micros-
copy and its physical limit in terms of resolution, single-molecule localisation microscopy
(SMLM) as a successful method to overcome it, and the new quantitative approach that
SMLM enables. This chapter was partially published as a review [1] and an introductory
article written for the special imaging issue of the Physiology News of Summer 2014 [2].
2.1 Fluorescence microscopy
2.1.1 Fluorescence and Stokes’ shift
First described by Stokes in 1852, fluorescence describes the emission of a photon of light
when an electron of a molecule relaxes from an excited high-energy state S1 to a lower-
energy ground state S0 with a temporal probability well-modelled by an exponential decay
and a characteristic fluorescence lifetime τ on the order of nanoseconds [3]. The energy
loss E during this event is converted in a photon of wavelength λem = h · c/E (where h =
6.62 ·10−34 J·s is Planck’s constant, and c= 2.998 ·108 m·s−1, the speed of light in a vacuum)
(cf. general Jablonski diagram in Figure 2.1).
In most applications, electrons in fluorescent molecules (also called fluorophores) are
pumped in their first excited singlet state S1 by absorption of a photon. This excitation
photon has a shorter wavelength λex than the emitted photon: because of the second law
of thermodynamics, no energy is created but rather dissipated between the time when the
6 Introduction to single-molecule localisation microscopy
excitation photon is absorbed and when fluorescence is emitted (in terms of energies of both
photons: h · c/λex ≥ h · c/λem). Mechanistically, during relaxation, some non-radiative (i.e.
non-fluorescent) energy is transferred by heat to the solvent, and lost: (1) the excitation pho-
ton generally populates a high vibrational energy level of the excited state S1 which quickly
relaxes to the lowest vibrational energy level of S1 before emission of the fluorescence. (2)
The electron also often relaxes to a higher vibrational energy level of the ground state S0,
emitting a photon and further relaxing to the ground state by a non-radiative way.
Figure 2.1 General Jablonski energy diagram of a fluorophore. Under illumination at a wavelength λex, an
orbital electron of the fluorophore can be excited by a photon of energy hνex and leave its ground state S0 to
a singlet excited state S1 (Absorbance). The electron then relaxes to the lowest vibrational energy level of
S1 in a non-radiative way, dissipating heat in solution. The electron can either relax to its ground state S0 by
emitting a photon of light of energy hνem (Fluorescence) or transition to other states, such as a triplet state T
in a non-radiative way. Generally, the electron finally relaxes from the triplet state T to its ground state S0 by
radiative (Phosphorescence) or non-radiative ways.
This spectral shift between excitation and emission light is called the Stokes’ shift of a
fluorophore. Magnitudes of Stokes’ shifts differ from fluorophore to fluorophore for some-
times complex reasons. The Stokes’ shift also depends on the local environment of the
fluorophore (e.g. Förster resonance energy transfer (FRET) uses the close proximity of a
donor and an acceptor fluorophores to dramatically increase the spectral shift of the donor).
However, fluorophores with large Stokes’ shifts enable the optical separation of their ab-
sorbance and emission spectra by use of beamsplitters. This explains the successful advent
of fluorescence microscopy that uses fluorophores to tag a specific target in the sample: the
fluorescent signal can then be collected without being convolved with the reflected illumi-
nation light.
One drawback of fluorescence imaging, especially when single molecules are observed
2.1 Fluorescence microscopy 7
is the inherent stochasticity of fluorescence (as ensemble averaging masks individual sto-
chastic behaviours). This is, moreover, usually convolved with a multitude of compet-
ing energy paths by which the excited fluorophore can relax to its ground energy state.
Thus, a fluorophore under constant excitation light will stochastically take radiative and
non-radiatives pathways, through different intermediary energy states. The probabilities or
kinetics associated to each pathway determine the quantum yield and photo-blinking (tran-
sient entry in a dark state, i.e. cycling in non-radiative paths) propensity of a fluorophore.
2.1.2 Abbe’s diffraction limit
Microscopes, like telescopes, have been aiming at increasing their magnification in order to
see deeper into the infinitely small or the infinitely large. However, both optical tools are
limited in this goal by the physical diffraction limit as described by Abbe [4]: all imaging
systems are limited in their frequency space by a high-frequency cut-off corresponding to
λ/(2 na) (where λ is the wavelength of the collected light, and na the numerical aperture of
the lens of optical device being used). The function that convolves a point-emitter imaged
through an optical system is called the point-spread function (PSF) of the system.
Abbe’s limit results in the fact that the PSF of a typical objective lens corresponds to
an Airy disk that is well approximated with a 2D-Gaussian for emitters close to the focal
plane. The width σGaussian of the corresponding PSF approximation is proportional to the
wavelength as follows [5]:
σGaussian =
1.323λ
2π na
(2.1)
In this thesis (as it is common in the field of single-molecule microscopy), the PSF of a
fluorophore refers to the image of the specific fluorophore through the objective lens of the
microscope, i.e. the convolution of the PSF of the instrument with the position and intensity
of the fluorophore of interest (cf. Figure 2.2, C for an illustration).
Equation 2.1 reveals the restriction imposed by Abbe’s diffraction limit on optical reso-
lution (typically limiting the resolution of modern fluorescence microscopes to ∼ 250 nm):
two fluorophores closer than 250 nm from each other will hardly be distinguished as their
PSFs will overlap. They cannot be directly resolved optically. Electron microscopy solves
this issue by using much smaller wavelengths (10−10 instead of 10−7 m). It however re-
quires fixation of the sample, both altering the sample and preventing live cell imaging.
However, the imaged PSF of a fluorophore can be thought of as a probability distribution
function of the position of the point-emitter. Therefore, through determination of the centre
of an individual PSF, the Abbe diffraction limit can be overcome and the emitter can be
8 Introduction to single-molecule localisation microscopy
localised with greater precision. The calculation of the centre of the PSF and consequently
the precision of the localisation depends on two main variables: the number N of collected
photons, via 1/
√
N, and the pixel size of the detector (cf. Equation 3.13) [6].
2.1.3 An overview of far-field optical super-resolution techniques
New methods, collectively grouped under the term super-resolution microscopy, have been
developed to overcome the physical limit of diffraction, gaining over two orders of mag-
nitude in precision [7], and allowing direct observation of processes at spatial scales much
more compatible with the regime that biomolecular interactions take place on. Different ap-
proaches have so far been proposed, including: (1) limiting spontaneous fluorescence emis-
sion of the sample to spatial regions smaller than the diffraction limit or frequency domains
overpassing Abbe’s limit (targeted switching and readout), (2) or stochastically separating
single fluorophores in time to gain resolution in space (stochastic switching and readout).
Various other techniques, such as the recent radically-innovative expanding microscopy [8],
are still emerging and may present complementary advantages which could be precious to
orthogonally validate controversial observations.
Structured illumination microscopy
Targeted switching and readout includes structured illumination microscopy (SIM) [9] and
stimulated emission depletion (STED) microscopy [10, 11]. Both techniques use specific
illumination geometries to gain sub-diffraction information. By only illuminating a grating
pattern and imaging the Moiré pattern resulting from its interference with the sample, SIM
enables a two-fold gain in resolution due to more precise information obtained in specific
domains of the Fourier transform of the recorded image. To fully compute and rebuild
the frequency domain of the sample, though, multiple frames needs to be recorded while
varying independently both the orientation and the phase of the sine illumination pattern.
The super-resolved figure is obtained by finally computing the reverse Fourier transform of
the recombined frequency domain of the sample.
Although fast, conventional SIM can only improve the resolution two fold. However, in
a similar strategy to STED, saturating SIM (SSIM) [12] uses non-linear optical saturation
of the sample to overpass this limitation: schematically, the saturation of the sample enables
to limit the size of the non-illuminated domains of the sample to specific sub-diffraction
values. Saturation induces a series of high-order harmonics in the frequency domain of the
imaged sample which analysis translates into sub-diffraction resolutions (e.g. 50 nm). SIM
2.1 Fluorescence microscopy 9
and SSIM present the advantage of being fast and generally compatible with all conventional
fluorophores. They are, however, less precise than other methods such as STED.
Stimulated emission depletion microscopy
Similarly to SSIM, STED uses non-linear optical saturation to overpass Abbe’s limit, but
the optical saturation here concerns fluorescence depletion in a configuration reminiscent of
scanning confocal microscopy. In STED, a traditional confocal illumination beam is aligned
with a ‘doughnut’-shaped depletion beam designed such as to overlap a zero-intensity spot
with the excitation focal spot. The illumination beam excites the fluorophores in the con-
focal volume and the powerful depletion beam depletes almost instantaneously the excited
states of the fluorophores located in the zone illuminated by the doughnut, through stimu-
lated emission.
Using high-power depletion beams, saturation of the depletion is reached for all the fluo-
rophores around the focal diffraction-limited excitation spot, thus limiting the spontaneous
fluorescence emission to a sub-diffraction area (i.e. all fluorophores further than a distance
r (<250 nm) from the centre of the illumination beams are trapped in their ground dark state
S0). The radius r of the emission area can directly be modulated by the intensity of the
depletion bead. As in confocal microscopy, the aligned beams are then scanned through the
sample to rebuild the final super-resolved figure.
The resolution achieved by STED microscopy, equal to the width 2r of the scanned of
spontaneous fluorescence emission area, is theoretically only limited by the power of the
depletion beam, but practically by the quality and precision of the alignment of both excita-
tion and depletion beams. Another important drawback of the technique is the requirement
for high-power illumination which makes live cell imaging challenging. However, the gen-
eralisation of STED to the concept of reversible saturable optical fluorescence transitions
(RESOLFT) dramatically reduces the amount of required illumination, thus potentially in-
creasing its compatibility with live cell imaging [13, 14].
The other important drawback of STED microscopy is its time resolution which is in-
versely proportional to the area of the region of interest that is imaged and limited by the
scanning speed and the mean time of the duty cycle of the fluorophore. For very small re-
gions of interest (∼ 1 µm2) though, STED has so far been the fastest of all super-resolution
methods (up to 80 frames per second for diffusing beads [15], or 28 frames per second
in living cells [16]). Interestingly, parallelisation of the illumination pattern dramatically
improves its speed and combines STED with SSIM: two orthogonal SIM-like interference
patterns were designed to produce an array of more than 100,000 doughnut-shaped deple-
10 Introduction to single-molecule localisation microscopy
tion beams that only need to scan a tiny fraction of the sample [17]. Finally, an advantage
of the traditional implementation of STED with pulsed illumination is its compatibility with
lifetime measurements, adding another dimension to the data (as spectrally separating pro-
teins tagged with different fluorophore is technically challenging in STED microscopy).
Single-molecule localisation microscopy
The second approach (stochastic switching and readout) was independently demonstrated by
three different groups in 2006: photo-activated localisation microscopy (PALM) [18], sto-
chastic optical reconstruction microscopy (STORM) [19], and fluorescence photo-activation
localisation microscopy (fPALM) [20]. These techniques, grouped under the name of single-
molecule localisation microscopy (SMLM), or single-molecule active control microscopy
(SMACM), allow the imaging of single molecules which cannot only be precisely localised,
but also followed through time and quantified. The next sections focus on this ‘pointillism-
based’ super-resolution imaging1.
2.2 Single-molecule localisation microscopy
2.2.1 Principle: spreading spatial density in time
The concept behind SMLM is that the position of an object can be known with greater
precision than its size. The size of a table is on the order of meters, but its position can be
easily known with a precision of centimetres. This is similar with fluorophores. Although
they are observed as 250 nm wide PSFs, each PSF can be very accurately fitted with a 2D-
Gaussian and its centre localised with great precision. The localisation precision mainly
depends on the pixel size of the camera and is inversely proportional to the square root of
the number of collected photons above the background (cf. Section 3.4.3) [6].
However, this strategy can only be applied if PSFs are distinguishable and separated
from each other. If the density of labelled proteins brings the fluorophores closer than the
diffraction limit (e.g. protein complexes), their PSFs merge and determining the number of
PSFs and the positions of their centres becomes almost impossible. This is made addition-
ally complicated by the fact that the intensity of a fluorophore considerably varies both in
a deterministic and non-deterministic way, depending on its position, its orientation and its
excitation. This means that a PSF twice brighter than another one cannot be inferred to be
the image of twice as many labels.
1Good extensive reviews for STED and SIM methods can be found in refs. [21] and [22], resp..
2.2 Single-molecule localisation microscopy 11
Figure 2.2 Concept of single-molecule localisation microscopy. A musketeer coat of arms structure was simu-
lated (A) from which 10,000 fluorophores were stochastically activated. A few fluorophores only are activated
in each frame (C). Each PSF is then fitted (D-H) and re-plotted as a 2D-Gaussian which width corresponds
to the precision of the localisation (F-G). Many frames allow the rebuilding of a super-resolved picture (E)
which is much more precise than the diffraction limited one (B).
To circumvent this density issue, the essential step leading to pointillism-based super-
resolution microscopy was the ability to control the activation of point-emitters to stochas-
tically separate them in time, so that only a few PSFs appear in each single frame. Thus, a
SMLM experiment consists in imaging thousands of frames, each of which only showing
sparse separated PSFs that are fitted with 2D-Gaussians and whose positions and precisions
are plotted in a final super-resolved figure (cf. Figure 2.2).
The various SMLM techniques now distinguish themselves based on how the emitters
are spatially isolated: although Betzig originally proposed to separate emitters spectrally
[23], dense fluorophores are now commonly separated by using a population of dark fluo-
rophores (i.e. initially non-fluorescent or pumped into a non-fluorescent state) and stochas-
tically turning a few of them in their fluorescent state, thanks to photo-blinking, photo-
activation or binding interactions (cf. Section 2.2.2). Thus, very few fluorophores are im-
aged at the same time, statistically allowing their spatial separation through time.
12 Introduction to single-molecule localisation microscopy
2.2.2 Single-molecule photo-switching
To separate dense fluorophores through time, SMLM requires methods for switching fluo-
rophores ‘On’ and ‘Off’. Many solutions to this exist. Here, three popular mechanisms used
in biological imaging are highlighted: exploiting the innate photo-physics of fluorophores,
chemical/spectral structural modification, and finally diffusion.
Photo-physical manipulation
When pumped in its first singlet excited state, a fluorophore usually relaxes to its ground
state by emitting fluorescence within nanoseconds. However, inter-system crossing to non-
fluorescent states, such as triplet states may also occur, and when in such states, the fluo-
rophores are unable to rapidly relax and enter further excitation-emission cycles; here the
fluorescence lifetime is typically on the order of micro- to milliseconds. This leads to the
molecules undergoing fluorescence intermittently, so-called blinking, as they cycle into and
out of the ‘dark’ state. Unfortunately, population of these dark states can often be part of
the photo-bleaching pathway, and so the number of switching cycles is limited.
However, this blinking can still be used to temporally separate the emitters, and therefore
allow for SMLM imaging [24–26]. This process is also observed in genetically encodable
fluorescent proteins, for example enhanced yellow fluorescent protein (eYFP), which has
been shown to undergo power-dependent blinking under irradiation of 515 nm laser excita-
tion. This has been used to study super-resolved spatial distributions of proteins in bacterial
models [27, 28].
Photo-blinking has also successfully been used to resolve close quantum dots [29]. How-
ever, quantum dot photo-blinking is difficult to control and spatial separation of single quan-
tum dots is often impossible in dense labelled samples. An alternative strategy described
by Hoyer et al. consists in using quantum dot ‘blueing’ to further separate a smaller frac-
tion of blinking quantum dots as their emission spectrum switches towards the blue under
photo-induced oxidation [30].
Chemical modification
In standard fluorescence microscopy, various methods have been employed to circumvent
the effect of photo-bleaching. Because of its triplet state, oxygen has the most understood
role in inducing photo-bleaching [31]. Thus, means to remove molecular oxygen from the
imaging buffer, for example by simple degassing, or through addition of enzymatic oxygen
scavenger systems, are commonly used in fluorescence microscopy [32–34]. However, re-
2.2 Single-molecule localisation microscopy 13
moval of molecular oxygen simultaneously increases of the lifetime of the triplet state, and
consequently enhances the photo-blinking behaviour of the fluorophore.
This oxygen-depletion-induced photo-blinking could be directly used for SMLM imag-
ing, but its poor tunability makes it hard to adjust the blinking rate to the density of fluoro-
phores to be separated. Thus, addition of triplet state quenchers at various concentrations
(i.e. micro- to millimolar concentrations of reducing and oxidizing agents) to the oxygen-
depleted buffer enables a finer control over the proportion of fluorophores in their triplet
dark state as they are able to relax the fluorophore back to its ground energy state2 [35–37].
By varying the buffer composition (i.e. the concentrations of the oxidizing and reducing
agents), long-lasting dark states can be generated, therefore allowing for SMLM imaging
[38–40]. The dark state generally exists as a reduced form of the chromophore, which ab-
sorbs at shorter wavelengths. Hydrogenated fluorophores can also act as dark states (by
addition of sodium borohydrate for example) which can be depopulated upon 405 nm il-
lumination [41, 42]. Such dark states can thus be depopulated through irradiation with a
405 nm illumination yielding the on-state of the dye once again. Although in direct STORM
(dSTORM − a generalisation of STORM to standard dyes) excitation at this wavelength is
achieved directly by irradiation at 350-550 nm [26, 43], STORM originally used a reporter-
activator dye pair, or cyanine switch (e.g. Cy5-Cy3), where the close proximity of the two
dyes leads to the on-switching of the reporter by illumination at a lower wavelength [19, 44].
Globally, this chemically-induced photo-switching has now been successfully imple-
mented with a very wide variety of cyanine, rhodamine and oxazine dyes [45, 46]. Finally,
Uno et al. recently proposed a strategy to avoid the use of buffers hardly compatible with
live imaging and designed dyes which spontaneously blink at physiological pH [47].
Diffusion
In addition to photo-switchable fluorophores, techniques that rely on fluorophores being
transiently bound to the protein of interest have been developed. Fluorophores freely diffus-
ing in solution are not observed as their PSFs are blurred by their rapid sub-frame diffusion.
However, when a fluorophore temporally binds to a protein of interest, its diffusion drasti-
cally decreases and its PSF is detected over the diffusing and out-of-focus background. In
point accumulation for imaging in nanoscale topography (PAINT) [48] and universal PAINT
(uPAINT) [49], fluorophores temporarily bind to surfaces, such as membranes, and are lo-
2Triplet state quenchers were initially rather used to deplete the triplet state, thus the photo-blinking of the
fluorophore in order to increase its quantum yield. They were later used at lower concentrations to control the
population of fluorophores in the dark triplet state.
14 Introduction to single-molecule localisation microscopy
calised one-by-one allowing for SMLM imaging. Binding-activated localisation microscopy
(BALM) has been used to generate super-resolved figures of DNA, using the intercalating
dye YOYO-1, the quantum yield of which increases threefold upon binding to DNA [50].
2.2.3 Fluorophores for SMLM imaging
Fluorescent dyes and proteins
Although photo-switching was mostly presented and discussed for conventional organic
(cyanine, rhodamine and oxazine) dyes in the previous sections, other photo-switching
fluorophores have been used in SMLM experiments. Quantum dots have the advantage
of high brightness and little if any photo-bleaching, but are neither photo-activable nor
cell-permeable, present some cytotoxicity and are prone to aggregation. Thus their use
in SMLM has been limited to the tracking of low density of membrane proteins using their
photo-blinking behaviour [29, 51, 52].
Light-inducible control of the spectroscopic properties of fluorescent proteins enabled
their use in SMLM imaging. Fluorescent proteins are genetically encoded and thus present
the advantages of high specificity, stoichiometric and intracellular labelling, and compatibil-
ity with live imaging. They, however, are usually less bright and photo-stable than organic
dyes, although brighter fluorescent proteins have now been described, some providing a
photon budget comparable to popular organic dyes (e.g. the different versions of the bright
and popular mEos − cf. Section 3.3).
Many different fluorescent proteins have been described; however, they all consist of a
single polypeptide chain of about 230 amino acids in length, each containing an 11-stranded
beta-barrel, which has a single distorted helix spiralling down its centre and contains three
amino acids that form the chromophore [53]. The chromophore engages in a wide-range of
interactions with surrounding amino acids and solvent environments, and single mutations
in the non-chromophore amino acids can strongly influence the spectroscopic properties
of the protein, shifting its absorption and emission spectra, or even leading to it becoming
non-fluorescent [54]. Some mutations induce remarkable photo-switching properties of the
fluorescent proteins which can be distinguished in two general categories of fluorophores:
reversible photo-switchable and irreversible photo-activable fluorophores. Both categories
and their distinction are discussed in next section.
Fluorescent proteins directly tag the protein of interest as a genetic fusion protein. In
terms of label size (critical for accurate SMLM measurements − cf. Section 3.4.3), fluo-
rescent dyes (≤1 nm) are smaller than quantum dots (∼2-10 nm) and fluorescent proteins
2.2 Single-molecule localisation microscopy 15
(cylindrical shape of diameter ∼2.8 nm and length ∼4 nm). However, to specifically label
a protein, most dyes need to be covalently attached to an antibody, making the total tag
actually larger (∼14 nm in their largest dimension). Nevertheless, means to limit the total
size of the tag (e.g. nanobodies − small single-domain antibodies from camelids or carti-
laginous fish [7, 55, 56] −, small target molecules like phalloidin [37] or the intercalating
dye YOYO-1 [50], ‘click’ chemistry strategy [57, 58]) or to get the specificity of fluorescent
proteins (by retro-engineering enzyme tags, e.g. the HaloTag [59] of the Snap-tag [60])
have been proposed to optimise sample labelling and get the best of both organic dyes and
fluorescent proteins.
Reversible and irreversible photo-switching
Photo-switchable fluorophores can be separated in two categories3:
Reversibly photo-switchable fluorophores. The photo-switching methods described in the
previous section involve reversible dark states. A wide range of fluorescent proteins
also displays a light-induced, reversible transition between a fluorescent and non-
fluorescent state [61–64]. Some of them present a very high contrast, good thermal
reversibility and high thermal stability (e.g. Dronpa [65], rsTagRFP [66], or mGeos
[67]). This type of photo-switching usually involves a transition to the off-state after
irradiation with the same wavelength that induces fluorescent emission, while irradi-
ation with UV light leads to recovery of the on-state.
Irreversibly photo-activable fluorophores. Although organic dyes are mostly used as re-
versible photo-switching fluorophores in SMLM experiments, some methods were
developed to use dyes in an irreversible photo-activation scheme. The dye is made
non-fluorescent by covalently either adding a hybrid anion in situ [68] or replac-
ing an amine group by an azide [69], both of which disturb the π-bonded network
responsible for the fluorescence properties of the chromophore. The π network of
these so-called caged dyes or fluorogens is released during imaging by a light-induced
chemical reaction which irreversibly photo-activate the dark molecule.
Fluorescent proteins can also display irreversible light-induced changes, either con-
sisting of an off-state to on-state conversion (e.g. PA-GFP [70], PS-CFP [71], and
3Incorrectly, but as a practical convention, the acronym ‘PALM’ is often used to refer to an experiment
using irreversibly photo-activable fluorophores and ‘STORM’ to an experiment that involves reversibly photo-
switchable fluorophores. However, the distinction sometimes rather refer to the use of fluorescent proteins
(‘PALM’) as opposed to organic dyes (‘STORM’).
16 Introduction to single-molecule localisation microscopy
PAmCherry [72]), or a conversion from shorter to longer wavelength emission (green
to red) (e.g. Kaede [73], EosFP [74] and its monomeric/tandem variants, Dendra2
[75]), when irradiated with near-UV light. The green to red photo-conversion is a
result of the conjugated system in the chromophore being extended due to cleavage
of a histidine side-chain [76, 77].
The major advantage of the latter type of fluorophores is that they allow single-molecule
counting as they are activated only once and thus each single burst of fluorescence theore-
tically corresponds to one fluorophore (cf. next section and the introduction of Chapter 6).
Single-molecule counting is one of the various quantitative approaches that SMLM brings,
together with improved resolution, to the field of fluorescence microscopy.
2.3 A quantitative approach to microscopy
One of the least known but major advantage of SMLM is the quantification it brings to
the microscopy field. Instead of the N (e.g. 512x512) pixel intensity values obtained in
any other bulk method, an SMLM experiment gives as primary results a multi-dimensional
matrix composed of >10,000 rows (i.e. the list of fitted localisations) and >10 columns
(frame of first appearance, (x, y) position, amplitude, width of the PSF, ellipticity, number of
re-appearances, etc.). These data require careful statistical and sometimes complex analysis,
but provide precious quantitative information about the sample.
2.3.1 Local diffusion analysis and detection of rare events
As a proportion of fluorophores does not photo-bleach in a single frame, a natural extension
of SMLM in live cells is to follow a single activated fluorophore over a few consecutive
frames. This proportion is dependent on the illumination power which can be lowered
to increase the average time that the fluorophores are imaged for before photo-bleaching.
Bridging SMLM and single particle tracking [78] (in single particle tracking PALM − spt-
PALM [79]) gives access to local diffusion maps [49] and to the precise description of
single-molecule kinetic states [80].
In this respect, at the core of SMLM is the ability to track and study low probability
events. This is illustrated by the American school bus problem: when looking from the
space at the Earth, looking specifically at the American school busses in average over a day
(as in bulk imaging), an observer would see them aligned in car parks, static and aggregated.
This observation would not help the observer to understand the function of such a device
2.3 A quantitative approach to microscopy 17
on Earth. However, if the observer could follow only a few busses over one day, he would
realise that they leave their static state and very regularly visit the same schools. These tracks
would help understand the very functionality of the school busses. This metaphor highlights
the help SMLM provides in understanding the functionality of rare events compared to
conventional fluorescence bulk imaging.
2.3.2 Single-molecule counting and cluster analysis
Another natural extension of SMLM, especially when using irreversibly photo-activable
fluorophores, is to consider single-molecule counting. Indeed, although most fluorophores
used in SMLM enter multiple times in a dark state and photo-blink, some fluorophores such
as caged dyes or the fluorescent protein mEos can be photo-activated once and imaged until
they bleach. Such fluorophores thus promise the possibility of counting the precise number
of labelled molecules in a cell, a cluster or a protein complex. The technique is still at
its beginning and faces major over- and under-counting challenges, but innovative methods
begin to propose partial solutions to tackle them and unravel the stoichiometry of protein
complexes in vivo (cf. the introduction of Chapter 6 for a complete description of the field).
The stoichiometry of a protein complex can also be determined by localising each of its
proteins with higher precision than the size of the complex. A remarkable example is the
repeated observation of the 8-fold symmetry of the nucleopore complex by different groups
[7]. Because of the reproducibility of the sample preparation and the robust stoichiometry of
the nucleopore complex, it has even recently become a molecular ruler for both localisation
and counting precision [81].
Finally, in a larger scale, the positions of the fitted localisations can be statistically anal-
ysed in pair-correlation PALM (PC-PALM) [82, 83] to quantitatively study the size and
population of a distribution of clusters.
2.3.3 Single-molecule multi-dimensional imaging
Beyond the initial two spatial dimensions collected in SMLM, additional single-molecule
information can be collected on each imaged single emitter such as its axial position, its
lifetime, its spectral distribution or its orientation. Concerning continuous structures such
as membrane domains or vesicles, information about their evolution over time in live cells
can also add a further dimension to the SMLM experiment.
18 Introduction to single-molecule localisation microscopy
3D imaging
As most biological processes happen in three dimensions, observing their 2D projection
is not always relevant or convincing to study specific phenomena. Several methods were
developed to code the information about the axial position of the fluorophore in some de-
formation of the shape of its PSF. This additional feature is implemented by the addition
of different optical components in the imaging path, allowing it to be compatible with all
SMLM methods. Four main methods have been developed, enumerated here in the order
of their ease of implementation and use (or, in the reverse order, on their axial localisation
precision) [84, 85]: biplane microscopy [86], astigmatism [87], the double-helix PSF (DH-
PSF) [88] and interferometric PALM (iPALM) [89] (a detailed review of the four strategies
can be found in ref. [1]).
All four methods however share the same drawbacks: the axial position is gained to the
detriment of both precision and resolution. Indeed, adding extra optical elements, dilating
the PSF over more pixels or the increase in out-of-focus noise automatically decreases the
signal-to-noise ratio which is inversely correlated to the precision (cf. Section 3.4.3). Sim-
ilarly, since a volume is observed, more labelled proteins have to be separated, and an in-
creased sampling is necessary to prevent a decrease of resolution in the final super-resolved
picture.
Video-rate imaging
If the movement of the process observed is an order of magnitude slower than the time
required to acquire enough frames to sufficiently sample the structure of interest, SMLM
can be compatible with live cell imaging. Hence, a super-resolved video or time-lapse
of the phenomenon can be rebuilt: a moving time-window may be used to rebuild the final
video as a way to balance sampling and fluidity of the final super-resolved movie. Until very
recently this technique was limited by the acquisition rate of the highly-sensitive cameras
required for SMLM imaging (electron multiplying charged-coupled devices − EMCCDs −
have a typical minimal exposure of 33 ms).
Thanks to the subtle adaptation of the localisation fitting algorithms to the noise charac-
teristics of scientific complementary metal-oxide semiconductor (sCMOS) cameras, much
higher frame rates are reachable (on the order of the millisecond over the same area) [90].
This makes the new limiting factor the physical photo-physics of the fluorophore, such as its
lifetime decay, its photon budget or its absorption cross-section. Indeed, in order to keep a
high signal-to-noise ratio, a constant number of photons is required per fluorophore and per
2.3 A quantitative approach to microscopy 19
frame. Thus decreasing the exposure amounts to increase the illumination power if the fluo-
rophore is not saturated. An increase in illumination power might also prove to be cytotoxic
which can be problematic when imaging live samples.
Although these limitations are challenging, one could propose to bridge SMLM with
fluorescence lifetime imaging microscopy (FLIM), recording the fluorescence lifetime of
each single fluorophore. Single-molecule FLIM is commercially available and proof of
concept of wide-field FLIM imaging [91] and combination of TIRF illumination and FLIM
(TIRFLIM) [92] were demonstrated, but together with the saturation of the fluorophores and
the cytotoxicity of the illumination, new design of cameras are required [93].
A nano-world in five dimension?
As for conventional fluorescence microscopy, spectrally-separated fluorophores can be used
to simultaneously label various proteins of interests [72, 94, 95]. Interestingly, emission
spectra of the fluorophores can overlap or be similar if the photo-induced activation involves
separated wavelength ranges [44]. Another strategy for fixed samples, so-called exchange-
PAINT, consists in using the transient binding of the fluorophores in uPAINT to wash the
labels from the samples at the end of acquisition and replace them with labels targeting
another proteins tagged with the same colour [96].
Instead of considering discretely separated spectra, single-molecule spectroscopy can
also be associated with SMLM (spectral-PAINT, developed by Godet et al., article in prepa-
ration): for each bright fluorophore, both its spatial position and its spectrum are recovered
thanks to the simultaneous imaging of both zero and first interference orders through a
diffraction grid.
Finally, other more exotic but complementary dimensions, such as fluorescence anisotropy
[97] that gives information about the rotational mobility or the orientation of each single
emitter [98, 99], have also successfully been explored and coupled to SMLM imaging pro-
tocols.
Thus, by active control of the switching of fluorophores, the physical limit of diffraction
can be overcome: using SMLM, biological samples have been imaged with sub-diffraction
precision on the nanometer scale. Multiple mechanisms to separate dense fluorophores
through time have been described, and numerous dyes and fluorescent proteins have suc-
cessfully been used in SMLM. By localising single fluorophores one-at-a-time, SMLM pro-
vides rich quantitative information about the sample and has inspired the development of
new statistical methods for data analysis.

Chapter 3
Building and characterising a
single-molecule super-resolution
platform
Building of the three super-resolution platforms was the work of M. Palayret. They were
designed in collaboration with Dr S.F. Lee. Dr J. Godet helped with setting up the light-sheet
on the double-helix instrument. All algorithms were designed and written by M. Palayret.
All data presented here were also imaged and analysed by M. Palayret. The experiment
described in Section 3.5 are published in ref. [100].
Aim of this chapter
This chapter sums up the considerations about the design, building, and characterisation
of three single-molecule super-resolution microscopes: two similar 2D platforms differing
mostly in the wavelengths available for illumination (Figure 3.1 describes the first micro-
scope built, the second one offering additional 514 nm, 594 nm and 730 nm laser lines),
and a 3D two-colour double-helix platform (Figure 3.2). These instruments are now heav-
ily used by many users on a variety of projects ranging from the study of the aggregation
of pathogenic proteins in neuro-degenerative disorders to early T cell triggering, to DNA
replication and repair in live fission yeast [101].
This chapter first describes the two parts of a microscopes, and their specific require-
ments for SMLM: the illumination and the imaging paths. It then characterises the photo-
physical properties of mEos, the fluorescent protein used in this thesis. Following, the
22 Building and characterising a single-molecule super-resolution platform
bespoke algorithms developed to process SMLM data is described. Finally, an example
of a ‘smart’ automated imaging protocol is presented, and advantages of automation are
discussed.
3.1 Illumination path
In fluorescence microscopy, to measure signal from a sample, the fluorophores need to be
excited at a wavelength resonant to their absorption peak. In SMLM, since signals from
single fluorophores are recorded, both excitation of the fluorophores and collection of their
emitted signal are critical. First, the specific requirements of the illumination for single-
molecule imaging are described and discussed.
A single dye or fluorescent protein typically emits 1,000-10,000 photons per acquisition
frame (∼50 ms). Because of the limited numerical aperture of the lens collecting the fluo-
rescent signal, and the imperfect optical elements used in the imaging path (that absorbs and
reflects part of the signal), ∼100-1,000 photons are detected above the background at the
plane of the detector. However, this signal is spread over typically nine pixels on the camera
chip [6], following a 2D-Gaussian distribution. This spread decreases the signal-to-noise
ratio (SNR), since each pixel of the EMCCD reads a similar electronic noise [103, 104]. As
discussed in Section 3.4.3, the localisation precision of each fluorophore is inversely pro-
portional to the square-root of the SNR (Equation 3.13). Thus, optimising the SNR of the
technique is not only necessary to be able to detect single fluorophores but also to localise
them with precision below the diffraction-limit of the light.
A main source of noise in microscopy is the detection of photons coming from out-
of-focus fluorophores: since in epifluorescence microscopy (Figure 3.3, A-B,D) the entire
field of view is illuminated, all fluorophores emit photons, both those at the focal plane of
the imaging lens and those below or above it. The bright contrasted signal coming from
the focal plane is consequently noised by the diffused light coming from out-of-focus fluo-
rophores. Confocal microscopy increases the resolution of epifluorescence microscopy by
limiting both the illumination of the sample to a small single confocal volume (which can
be approximated to a 3D-Gaussian of standard deviation ∼230 nm laterally and ∼650 nm
axially for a 1.49-na oil-immersion objective at λ =680 nm), and the collection of the emit-
ted fluorescent light to a similar volume (with a pinhole of diameter around one Airy unit,
i.e. 33.5 µm).
Other methods, compatible with wide-field imaging, have been described to limit the
illumination of the sample to the focal plane and drastically reduce the out-of-focus noise
3.1 Illumination path 23
Figure 3.1 Scheme of one of the two 2D single-molecule localisation microscope built in this thesis. The
illumination path is composed of four laser-lines (640, 561, 488 and 405 nm from left to right) that are prepared
and aligned to illuminate a sample under TIRF. For clarity, only the 561 nm illumination is shown. The
fluorescence signal emitted by the sample is collected by the objective lens, filtered and spectrally separated
in a dualview before being projected on the EMCCD.
Legend: Curved arrows indicate direct hardware control by the MicroManager piece of software [102]. The
640 nm diode laser (left) was controlled through MicroManager and thus directly acts as a computer-controlled
shutter. Actual shutters were required for the three other laser lines.
24 Building and characterising a single-molecule super-resolution platform
Figure 3.2 Scheme of the 3D two-colour double-helix microscope built in this thesis. The illumination path is
composed of two laser-lines (640 and 405 nm from left to right) that are prepared and aligned to illuminate
a sample under TIRF, and one laser-line (640 nm) that illuminates the sample with a thin sheet of light. The
light-sheet is created by focussing the laser beam in a single direction at the back aperture of an objective lens
thanks to a cylindrical lens: an ‘extended 1D confocal volume’ is thus created, perpendicular to the optical
axis of the imaging objective. For clarity, only the 405 nm TIRF illumination and the 640 nm light-sheet
are shown. The fluorescence signal emitted by the sample is collected by the imaging objective, filtered and
conjugated in a 4 f system. The signal is spectrally separated by a dichroic mirror after the first Fourier lens.
A phase mask adapted to the wavelength of the collected signal is placed at each Fourier image plane. Finally,
the second Fourier lens of each channel focusses the resulting signal on the chip of an EMCCD camera.
Legend: Curved arrows indicate direct hardware control by the MicroManager piece of software [102].
3.1 Illumination path 25
(Figure 3.3, A):
• Total internal reflection fluorescence (TIRF) uses the difference of refractive index at
the surface between the glass coverslip and the water-based imaging buffer to totally
reflect the illumination beam1. A non-propagating evanescent wave is locally created.
The intensity of the TIR wave decreases exponentially along the optical path, typically
limiting the illumination of the sample up to∼100 nm above the coverslip (Figure 3.3,
A,F);
• Highly inclined and laminated optical sheet (HILO) [105], also referred to as “near-
TIRF”, is based on a similar optical geometry than TIRF illumination. However,
the angle of the illuminating beam with the coverslip is kept below the critical angle
to prevent its total reflection. As a consequence, a collimated beam illuminates the
sample. But because the incident angle is very close to the critical angle, the refracted
beam, following Snell’s law, is almost parallel to the coverslip, thus illuminating a thin
inclined sheet of the sample above the coverslip. Nevertheless, an inconvenience of
HILO is the non uniformity of the illumination as the HILO sheet is not perpendicular
to the optical axis (Figure 3.3, A,E);
• Selective plane illumination microscopy (SPIM) [106], or commonly called light-
sheet microscopy, conversely to both previous methods, usually uses two different
lenses for illumination and imaging (a very recent and elegant method, single objective-
SPIM, has been proposed by the group of Dr J.B. Sibarita2 to combine both lenses
and avoid the complex geometrical issue of bringing two high-numerical aperture ob-
jectives at 90◦ and very close from each other). The collimated illumination beam
is focussed horizontally at the back focal plane of the illumination lens thanks to a
cylindrical lens in order to focus the illumination beam in a line, similar to a confocal
volume extended in one dimension. This ‘focussed line’ acts locally as a 2D-sheet of
light which is aligned with the focal plane of the imaging lens (the axes of the two
objectives being perpendicular to each other) (Figure 3.3, A,C).
1The refractive index of the oil or the glass is ne = 1.52, higher than the average refractive index of living
cells ne = 1.35 or water, allowing total internal reflection if the angle of incidence is higher than the critical
angle.
2Dr J.B. Sibarita gave a convincing presentation about this technique at the Single-Molecule Localisa-
tion Microscopy Symposium organised at King’s College, London in August 2014. The title of the talk was:
“Single-objective Selective Plane Illumination for 3D high- and super-resolution imaging of biological struc-
tures”.
26 Building and characterising a single-molecule super-resolution platform
Figure 3.3 Effect of varying illumination geometries on the imaging contrast. Different illumination geome-
tries have been proposed to increase the signal-to-noise ratio in large-field microscopy. A: Scheme of four
commonly used illumination strategies: epifluorescence (blue), HILO (green), TIRF (red) and SPIM (orange).
B-C: Diffusing sub-diffraction fluorescent beads were imaged a few micrometers above the coverslip under
either epifluorescence (B) or SPIM (C) illumination in the double-helix PSF (DH-PSF) configuration of the
microscope (cf. Section 3.2.2). The arrow in (B) indicates the possible DH-PSF of one bead. D-F: Similar
carboxylated beads were fixed on the surface of a poly-lysine-coated coverslip and the same field of view was
imaged under either epifluorescence (D), HILO (E) or TIRF (F). The contrast of images D-F is fixed to the
same value. G: Histograms of the pixel intensities of each sample shown in D-F (A.U.: arbitrary unit− y axis:
log scale).
3.2 Imaging path 27
Figure 3.3 shows the experimental difference of contrast obtained when imaging fluores-
cent beads in solution with either epifluorescence (Figure 3.3, B) or light-sheet (Figure 3.3,
C) illumination. The latter dramatically increases the contrast as only a plane is illuminated:
although four diffusing beads are distinctly observed in panel C3, only one bead might be
observed when the whole depth of the sample is illuminated (white arrow, Figure 3.3, B).
Immobilised beads were imaged under epifluorescence, HILO or TIRF illumination
(Figure 3.3, D-F, resp.). Interestingly, the signal-to-noise ratio increases strikingly for HILO
and TIRF illuminations, as observed in Figure 3.3, D-F shown with the same contrast. The
background noise does not increase significantly, as opposed to the signal (Figure 3.3, G,
higher intensities). Furthermore, beads slightly out-of-focus and visible under epifluores-
cence or HILO illuminations disappear under TIRF illumination (Figure 3.3, F).
The sub-diffraction axial width of the TIRF field makes it the best illumination technique
to limit out-of-focus fluorescence emission. However, the main drawback is that TIRF can
only illuminate the part of the sample close to the coverslip. It is the perfect method to
image receptors at the plasma membrane of a cell forming a contact with the coverslip;
but to observe resting receptors on the top of a cell, or the organisation of the nuclei of
living yeast or stem cells, a HILO or a light-sheet illumination will be preferred. In this
thesis, TIRF illumination is used in Chapters 4-7, HILO illumination in the Section 3.5
of this chapter, and a 4 µm-wide light-sheet is combined with 3D double-helix imaging in
Section 3.2 of this chapter.
3.2 Imaging path
When illuminated, the sample emits a fluorescence signal that is collected by the imaging
lens and focussed by the tube lens onto the detector. This imaging path is very sensitive
to any external instability and noise as the collected single-molecule signal over the back-
ground is on the order of magnitude of the limit of detection of the camera (10-100 pho-
tons/PSF). The signal can also be optically modified to get more information about the
single-emitter: changing its wavefront allows to additionally obtain the axial position of the
fluorophore; spectrally separating the signal in two channels allows simultaneous imaging
of two spectrally-separated fluorophores.
3Cf. Section 3.2.2 for more details about the shape of the PSFs.
28 Building and characterising a single-molecule super-resolution platform
3.2.1 Stability of the frame
Since SMLM allows to image single-emitters with a sub-diffraction precision typically on
the order of 50 nm, the stability of the imaging path, i.e. the variation of the relative position
of the sample to the camera, has to be an order of magnitude lower, or corrected for. To
prevent any major variation, each microscope was built on a heavy optic table supported by
vibration-isolator legs in a temperature-controlled environment (air conditioned at 20◦C).
Solid, heavy sample holders made of mild steel were designed to fit in nm-precise piezo-
electric microscope stages. To prevent any drift due to the dragging induced by oil after
a stage displacement, samples were maintained on the sample holder by small rare-earth
magnets.
In order to evaluate the stability of the microscopes and their resistance to different
external noise, fluorescent beads were immobilised on a coverslip and imaged for multiple
frames. Their trajectories were rebuilt (cf. Section 3.4) and their standard deviation analysed
as a metric for stability or drift of the microscope (Figure 3.4). It is worthy to notice that
this metric is a lower bound for the stability, as even for high SNR, the standard deviation of
the trajectory of a bead is the convolution of the stability of the instrument, the movement
of the bead and the precision of localisation of the bead. In optimal conditions, under
high illumination, the distribution of standard deviations peaked at 6-7 nm (Figure 3.4, C),
about 6-8-fold lower than the typical precision obtained for SMLM experiments (40-60 nm).
Although opening of the door or switching off the air condition during the course of the
experiment did not affect the stability (E), any perturbation of the imaging path, either of the
optic table or of the camera, induced major displacement of the sample (F), even impeding
correct analysis of the trajectories of the beads in extreme cases. Two different microscopic
frames (Olympus IX71 and IX73) were tested and, under similar illumination and with the
same imaging optical path the new frame IX73, although offering more access to the optical
path in infinity space at its base, was repeatably slightly more stable than the previous IX71
frame (6 nm vs. 7 nm, Figure 3.4, A-C).
As a control, the effect of a ten-fold decrease in illumination was observed (D). As ex-
pected from Equation 3.13, the SNR decreased and the standard deviation of the trajectories
increased from 6 nm to 11.5 nm for the IX73 frame. If illumination and signal were linearly
correlated, the precision should have varied of a factor
√
10 = 3.2. The observed difference
results from the non linearity of the relation between illumination power and emitted signal
and from the convolution of the stability of the instrument with both long-term drift and
localisation precision which worsen the latter for high illumination power.
Long-term drift (in time scales of minutes) is harder to avoid as it involves many phe-
3.2 Imaging path 29
Figure 3.4 Stability of the microscopes. Sub-diffraction carboxylated fluorescent beads were fixed on poly-
lysine-coated coverslips and imaged under TIRF illumination. Trajectories of single separated beads were
tracked and an average trajectory was calculated to measure the global translational (xy) drift of the sample.
A: Average trajectory, i.e. drift, of beads from a sample mounted on a microscope frame IX71. B: Drift of
the same sample under similar activation, on a microscope frame IX73. C: Comparison of the distribution of
experimental precisions (i.e. the standard deviation of the trajectory of fixed fiducial markers, cf. Section 3.4.3)
for both frames IX71 (A, grey bars) and IX73 (B, hatched bars). D: Same experiment as panel B, under a ten-
times lower illumination. E-F: Same experiment as panel B, some perturbations are introduced: in E, the door
of the room was closed; in F, the optic table was perturbed.
Legend: In panels A-B and D-F, positions are given in pixels. The colour of the trajectories codes for the time
of the experiment (i.e. the frame of the SMLM movie) to which the position corresponds (from red, to yellow,
to green, to blue). In panel C, the experimental precision is in nm.
30 Building and characterising a single-molecule super-resolution platform
nomena such as thermal local variations or mechanical relaxations [107, 108]. It is however
easy to correct for lateral drift thanks to fixed bright fiducial beads or markers [18]. Then,
the coordinates of the fitted localisations can be measured relative to the average position of
the markers over time. The main difficulty here is to properly and stochastically immobilise
a few markers in each field of view. The brighter and the more immobile the markers are,
the better their positions are known and the less precision is lost in each localisation by such
correction. However, consistently correcting for axial, or focus, drift requires some deep
understanding about the axial variation of the PSF and robust real-time algorithms to pre-
vent losing the signal of the single fluorophore as it leaves the focal plane of the objective
[109]. Correcting the varying distance between the objective lens and the bottom of the
coverslip by either monitoring the position of the reflected TIR beam or using a closed-loop
piezo-electric actuator proves to globally correct for only half of the focus drift (personal
communication from Dr E. Drier).
3.2.2 The double-helix point-spread function
The double-helix PSF (DH-PSF) is an optical method allowing to gain axial information
about the position of an imaged single molecule thanks to the deformation of its 2D-PSF.
Its principle resides in a phase change of the Fourier transform of the imaged PSF. In order
to explain how the Fourier decomposition of a 2D image can be obtained and modified
optically, some basic theoretical concepts, such as the spherical-plane wave duality, will
be approached in this section, followed by experimental optical calculations of a Fourier
decomposition. Then, the specific implementation of the DH-PSF will be described, and
preliminary results presented.
In classical optics, a point emitter located at the origin r⃗0 uniformly radiates light in
all directions. The sine electromagnetic wave of wavelength λ that propagates through the
transparent medium of refractive index n at any point r⃗ and any time t can be described as:
f (⃗r, t) = Ar sin(k r−ω t)
= ℑ
(A
r e
ı k r e−ı ω t
) (3.1)
where r = ∥⃗r∥, k = 2 π/λ and ω = k c/n (with c the speed of light in vacuum). The
total intensity of the spheric wave for any spherical surface centred at the source,∫∫∫
| f (⃗r, t)|2 r2 sin(θ) dθ dϕ dt = 4π A2 (3.2)
in spherical coordinates (r,θ ,ϕ) originating from the origin r⃗0, is constant over time and
3.2 Imaging path 31
independent from the distance r from the source, which is consistent with the law of con-
servation of energy. For calculus simplification, the wave function is often defined as the
imaginary part of a complex function (Equation 3.1). Also, when working with a single
wavelength λ , the dependence in time e−ı ω t is similar for all propagating waves and can be
eliminated to study the geometry of a wavelength. The simplified complex wave function is
thus:
φ (⃗r) =
A
r
eı k r (3.3)
If the point emitter r⃗0 is far from the point of observation r⃗, i.e. r ≫ λ , the wave fronts
can be approximated with parallel planes. The simplified complex wave function of a plane
wave of constant amplitude
∣∣∣B(⃗k)∣∣∣ and vector of propagation k⃗ = k e⃗⊥ is:
φ (⃗r) = B(⃗k) eı k⃗·⃗r (3.4)
Interestingly, an optical lens of focal length f transforms a spherical wave originating
from a point emitter r⃗0 = (x0,y0,− f ) in its object plane to a plane wave of vector of propa-
gation:
k⃗ = k [cos(θx) cos(θy)e⃗z+ sin(θx)e⃗x+ cos(θx) sin(θy)e⃗y]
= 2πλ
[
e⃗z− x0f e⃗x− y0f e⃗y
]
in the approximation of small angles (x0 ≪ f and y0 ≪ f ).
(3.5)
Conversely, and following Helmholtz reciprocity principle, the lens transforms a plane
wave of vector of propagation k⃗ to a spherical wave converging to a single point in its image
plane which coordinates are:
 xy
f
=

f tan(θx)cos(θy)
f tan(θy)
f
= f
 θxθy
1
= f

kx
kz
ky
kz
1
 (3.6)
Both Figures 3.5 C and D are two interpretations of the same physical phenomenon. In
Figure 3.5 C, the complex 2D electromagnetic wave from the object plane is interpreted in
the Huygens sense as a superposition of spherical waves originating from a continuum of
point emitters. Each spherical wave is collimated by the lens and forms a plane wave of
specific vector of propagation (cf. Equation 3.5). So each point (x,y, f ) of the image plane
32 Building and characterising a single-molecule super-resolution platform
Figure 3.5 A lens is an optical Fourier transformer. A point-emitter emits light uniformly in all directions,
producing a spherical wave of wavelength λ (A). Looking from afar (∥⃗r∥ >> λ , e.g. in the boxed region),
the spherical wave is approximated with a plane wave (B). Examples of wavefronts and light rays are resp.
shown as dotted red lines and red arrows. An optical lens, by conjugating a spherical waves originating from
its object plane to plane waves (C) and vice versa (D), acts as an optical Fourier transformer. Indeed, an image
in the object plane can be seen as a superposition of point emitters or of plane waves corresponding to its
decomposition. Experimentally, the image of a grating was focused on the object plane of a lens (E). The
Fourier transform of (E), calculated in silico, (F) is observed at the image plane of the lens (G).
3.2 Imaging path 33
receives a superposition of plane waves corresponding to point sources in the object plane:
φ(x,y, f ) ∝
∫∫
B(⃗k) eı k⃗·⃗r d⃗k from Equation 3.4
∝
∫∫
B(x0,y0) e
−2πı
(
x0
λ f x+
y0
λ f y
)
dx0 dy0 from Equation 3.5− cf. note 4
(3.7)
In Figure 3.5 D, the complex 2D electromagnetic wave from the object plane is inter-
preted in the Fourier sense as a superposition of plane waves interfering in the object plane
(each plane wave representing a 2D spatial frequency). Each plane wave is focussed on a
single point of the image plane. So each point (x,y, f ) of the image plane receives all the
energy that departed from the object plane at the spatial frequency k⃗ = kxe⃗x+ kye⃗y+ kze⃗z:
φ(x,y, f ) ∝ TFourier((x0,y0) ∈ object plane)(⃗k)
∝
∫∫
φ(x0,y0,− f )eı(x0 kx+y0 ky) dx0 dy0 cf. note 5
∝
∫∫
φ(x0,y0,− f )e−
2πı
λ
(
x0 xf +y0
y
f
)
dx0 dy0 from Equation 3.6
(3.8)
Experimentally, a Ronchi ruling slide (Figure 3.5 E), a grating of parallel lines, was
placed at the focal plane of the microscope. A lens of focal length f = 20 cm was aligned in
the imaging path of the microscope, at a distance f from the image plane of the microscope.
The camera of a smartphone was placed at the image plane of the lens and the image in
Figure 3.5 G was recorded. The 2D Fourier transform of Figure 3.5 E was performed in
silico and its 2D intensity is shown in Figure 3.5 F. As the image of the grating can be
approximated with a sine wave, the Fourier transform shows three components: the average
intensity of the image at the centre (order 0) and the positive and negative symmetric spatial
frequencies of the sine (first order). A very similar geometry can be appreciated after the
optical transformation through the lens. Interestingly, the first order is more shifted for
red wavelengths compared to shorter ones. Indeed, the image of the grating in the image
plane of the microscope is diffraction-limited, smoothing the square signal of the grating
into an approximated sine wave. But diffraction is intrinsically achromatic, and the size of
4After simplification of the constant term e2πı
f
λ in the multiplicative factor.
5After simplification of the constant term eı f kz in the multiplicative factor, in the approximation of small
angles
(
kz ∼ 2πλ =
∥∥∥⃗k∥∥∥ for θ ≪ 1).
34 Building and characterising a single-molecule super-resolution platform
the diffraction pattern, or of the PSF, is inversely proportional to the wavelength of the light
(Equation 2.1). Thus, the red signal will be observed in the image plane of the microscope
as a wider sine, which is translated in the Fourier domain as a shift of the first order away
from the order 0.
The DH-PSF method consists in modifying the phase of the emitted light in the Fourier
domain in order to change the 3D-PSF of the instrument in the real domain. Indeed, phase
cannot be directly recorded on an EMCCD, although it carries additional axial information
to amplitude about the position of the observed fluorophore. A phase mask was engineered
by Pavani et al. to change the diffraction-limited PSF of the emitter into two diffraction-
limited spots which angle with the lateral axes is proportional to the axial position of the
emitter (Figure 3.6 C-E) [88]. The phase mask is applied at the Fourier plane of a 4 f system,
i.e. two collimated lenses of focal length f (Figure 3.6 A-B). Introduction of the phase mask
induces a modification of the PSF of the instrument in the plane of the camera chip.
Immobilised fluorescent beads were imaged at different axial positions in order to cal-
ibrate the microscope set-up (Figure 3.6 D-F). As a proof of principle to show that the tra-
jectories of sub-diffraction particles could be followed in three dimensions, freely diffusing
beads were also imaged in water (Figure 3.6 G).
Other methods have been implemented to follow sub-diffraction particles in 3D (cf.
Section 2.3.3). The DH-PSF method has the main advantage of presenting uniform axial
and lateral precisions over its depth of field (Figure 3.6 F, bottom panels), compared to
bi-plane microscopy or astigmatism. Compared to iPALM, it is also easier to implement
and allows imaging of thicker samples as DH-PSF does not require dual-objective imaging.
However, iPALM has better lateral and axial precisions and collects more signal from the
single emitters6.
3.2.3 Two-colour imaging: registration of two channels
Pairing fluorescent beads
In order to simultaneously image two different proteins, two spectrally-separated fluoro-
phores can be used in a dual-labelling strategy (cf. Section 3.3). Under the illumination
of two aligned lasers, the fluorescence signal of both fluorophores is spectrally separated
by a dichroic mirror which reflects signal of all wavelengths above a characteristic value,
6In DH-PSF, the photons collected by the objective is separated in two lobes that both needs to be detected
with high precision when in iPALM, the photons are collected by two objectives and divided between three
cameras.
3.2 Imaging path 35
Figure 3.6 The double-helix point-spread function allows three-dimensional localisation of diffraction-limited
particles. A 4 f optical system is composed of two conjugated lenses (A) and allows the modification of the
phase of the relayed image from the object plane of the first lens in the Fourier domain. A 4 f system was
implemented to relay the image plane of a microscope to the chip of an EMCCD (B). In the Fourier plane,
the DH-PSF phase mask (C, diameter: 7.25 mm) was aligned with the optical axis, optically modifying the
PSF of the microscope to two lobes forming a double-helix (D) which rotation angle is proportional with the
axial localisation of the emitter (E, and top graph in F). Fluorescent sub-diffraction fixed beads were imaged
at different axial position to calibrate the set-up. Uniform localisation lateral and axial precisions of resp. ∼30
and ∼80 nm were achieved over a depth of field of 4 µm (F). Freely diffusing fluorescent beads in solution
were also imaged to test the possibility of 3D trajectory analysis (G, montage of a small field of view, from
left to right, then top to bottom, every 60 ms).
36 Building and characterising a single-molecule super-resolution platform
and transmits all other wavelengths (Figure A.1). Each separated channel can be indepen-
dently handled and focussed on two different cameras (Figure 3.2) or focussed back on
the two halves of a single camera. The latter technique is referred to as dualview imaging
(Figure 3.7, A).
To compare the localisation of the proteins from both channels, the channels have to be
optically aligned, computationally registered and the precision of their alignment has to be
assessed. Because new optical elements are added (lenses, mirrors, dichroic mirror as in
Figure 3.7, A), the channels need to be finely aligned by using an alignment sample such
as a regular grid of gold dots deposited on a glass coverslip, or separated dual-coloured
fluorescent beads (Figure 3.7, B) [110–113]. However, due to chromatic aberrations and
fine sub-diffraction translation and rotation which are difficult to assess and correct by eye,
a further correction step is calculated and corrected for in silico.
To do so, immobilised dual-coloured separated fluorescent beads are simultaneously
imaged in both channels (Figure 3.7, B-C). Intensity maxima in each channel are fitted with
2D-Gaussians thanks to a SMLM routine (see Section 3.4). The localisations of the beads
detected in one channel are paired with their nearest neighbours in the other channel in
order to match the positions of the same beads in both channels. Remarkably, for this step
to work, a relatively good initial optical alignment is necessary. Also, to prevent the miss-
match of close miss-aligned beads, a localisation Ai is paired with its nearest-neighbour in
the other channel B only if their distance is smaller than a fraction (here, 0.3) of the distance
separating Ai from its nearest neighbour in the original channel A. This pairing process can
be repeated for more than one field of view to increase the density of paired localisations,
i.e. the sampling, and consequently the precision of the registration without increasing the
density of the imaged beads (Figure 3.7, D).
A global linear geometric correction
In geometry, any 2D affine transformation can be sequentially decomposed in a translation,
a uniform scaling and a rotation. The calculation and determination of each of these con-
secutive steps of the transformation which moves any point from channel A to channel B
is described next. First, the respective barycentres (Ω and O) of the paired localisations
(Ai, Bi) in each channel are calculated:
Ω=
1
N
N
∑
i=0
(Ai) and O =
1
N
N
∑
i=0
(Bi)
3.2 Imaging path 37
The translation
−→
T =
−→
ΩO (Figure 3.7, G) is then applied to all localisations of channel A:
A′i = Ai+
−→
ΩO.
The scale factor is calculated as the following mean (Figure 3.7, H):
h =
1
N
N
∑
i=0
OBi
OA′i
The uniform scaling is then applied to the localisations of channel A′: OA′′i = h OA′i.
The angle of the rotation is calculated as the mean angle between
−−→
OA′′i and
−−→
OBi (Figure 3.7,
I):
θ =
1
N
N
∑
i=0
sin−1
∥∥∥−−→OA′′i ∧−−→OBi∥∥∥
OBi 2
The rotation is finally applied to the localisations of channel A′′:
−−→
OA′′′i =
[
cosθ −sinθ
sinθ cosθ
]
−−→
OA′′i
Both channel B and the corrected channel A′′′ are plotted in Figure 3.7, E. The regis-
tration process allows a significant increase in co-localisation position (Figure 3.7, J): the
distance between localisations of each pairs after correction is greatly reduced (blue his-
togram, compared to the orange one). The precision of registration is defined as the value at
the peak of the histogram in Figure 3.7, J.
Such registration correction allows sub-pixel registration precisions of typically∼55 nm.
It successfully corrects geometrical miss-alignment of the fluorescence signals, but it is lim-
ited by non-linear achromatic distortions between the two channels.
A locally-weighted mean of linear transformation correction
To obtain a finer registration and account for the small non-linear distortions (due to non-
linear achromatism) that cannot be dealt with thanks to the previous global geometrical
method, a method originating from image processing to register two images of the same ob-
jects taken from different angles and detectors (tilted, rotated, scaled, with different intensity
values, etc.) was implemented [114, 115].
For each pair of calibration localisations (Ai, Bi), the n nearest neighbours of Ai are
selected and used to build two separate polynomials (P xAi , P
y
Ai) of degree d that interpolate
the distance between both channels at these (n+ 1) points, projected on each coordinate
38 Building and characterising a single-molecule super-resolution platform
Figure 3.7 Two channel imaging requires registration for co-localisation studies. A dualview system, made
of a dichroic mirror spectrally separates the fluorescent signal into two channels that are projected on both
halves of the camera (A). After manual alignment of the dualview channels, multiple fields of view of fixed
separated dual-coloured beads fixed on a coverslip are simultaneously imaged in both channels (B, one field of
view; C, the superposition of three fields of view; long wavelength channel in red, short wavelength channel
in green). For each field of view and each channels, each intensity maximum is fitted with a 2D-Gaussian
and paired with its nearest neighbour in the other corresponding channel. All successful pairs of localisations
for the three fields of view in (C) are plotted as coloured circles in (D). A geometrical combination of a
translation (G), a scaling transformation (H) and a rotation (I) are calculated to match the green localisations
with the red ones. A locally-weighted mean of linear transformation is also calculated for the same purpose
(F, cf. text for details). As a result of the corrections applied to the green channel, the distribution of the
distance separating the localisations in each channel is thinned and its centre is displaced to a lower value
(J, orange before correction; blue after global linear correction; pink after locally-weighted linear correction).
The average precision of registration achieved was 12 and 55 nm, resp.
3.2 Imaging path 39
axis:
∀ j ∈ [[0, n]], P xAi(x j,y j) = x′j− x j (3.9)
and n =
(d+1)(d+2)
2
−1
where (x j, y j) is the j-th nearest neighbour of Ai (Ai being its 0-th nearest neighbour), and
(x′j, y′j) its corresponding localisation in channel B. The range of both polynomes is defined
as the circle centred on Ai and of radius RAi corresponding to the distance to its n-th nearest
neighbour (e.g. Figure 3.7, F, red circle, n = 4).
Once the calibration step is done, for each new position (x, y) from channel A, a cor-
rected position (x′, y′) is calculated as the weighted sum of all the T polynomials that it
intersects in channel A (e.g. Figure 3.7, F, orange dot intersecting four orange circles):
x′ = x+
∑Tj=0W ( j, x, y)P xA j(x, y)
∑Tj=0W ( j, x, y)
(3.10)
with W ( j, x, y) = 1−3R2+2R3 and R = d(A j, (x, y))
RA j
The weight function W was chosen so that (1) the weight is null outside of the range RA j
of its polynomial PA j ; (2) it increases up to a value of 1 at its centre A j; and (3) its derivative
is null in its centre and at the limit of its range, allowing the transformation to be continuous
and smooth on all points of the surface.
If the position to correct does not intersect any range of any polynomial, a pair of inter-
polating polynomials (P x, P y) is calculated with its (n+ 1) nearest neighbours, and their
values at the position (x, y) determines the correction to apply.
The precision of the registration is estimated and calculated by imaging m separate cal-
ibration fields of view: (m− 1) fields are used to build polynomials, and the last i-th field
of view is used as a control measurement, which corrected positions from channel A are
compared to the real positions of the corresponding beads in channel B. This is repeated
for all fields of view i ∈ [[1, m]]. At the end of the calibration, all fields of view are used to
build the final set of polynomials that are used for correction of experimental data. Thus,
the precision calculated is a worse estimate of the real registration precision.
Interpolation polynomials are calculated with a quick orthogonal polynomial method
as described in ref. [115]. Quadratic polynomials (d = 2) allow a better precision than
linear ones (d = 1), but induce high distorting fluctuations out of the convex envelope of the
calibration localisations. Linear polynomials were thus preferred in this thesis.
40 Building and characterising a single-molecule super-resolution platform
Finally, to tackle loose interpolations due to colinear nearest neighbours, two extra
nearest neighbours are used than what is theoretically needed for an exact interpolation
(n′ = n+ 2, Figure 3.7, F, d = 1 and n′ = 4). Strictly speaking, such polynomials are not
exact interpolators but good local approximators. This is also a very good way to make the
method robust to outliers due to miss-paired calibration localisations. Lastly, it increases
the range of the polynomials, i.e. the cover of the surface that can be corrected with this
method.
Such registration method allows registration precisions almost five-fold better than the
global linear method (typically ∼12 nm, Figure 3.7, J), i.e. less than the average locali-
sation precision of typical SMLM experiments. Furthermore, compared to the global lin-
ear method, neither the calibration nor the correction steps noticeably slowed the analysis.
This method is adapted to SMLM imaging as it calculates smooth corrections for discrete
localisations [112, 116]. For all these reasons, it was chosen for registration of any single-
molecule dual-colour experiments in this thesis.
Conclusion
The registration corrections presented here allow sub-pixel registration precisions7 (∼55 nm
in the case of a global linear transformation, ∼12 nm in the case of a locally-weighted
mean of linear transformations). Although these methods correct for both misalignment
(global linear transformation) and lateral achromatism (locally-weigthed mean transforma-
tion), they do not correct for axial achromatism that affects dualview imaging: the micro-
scope imaging lenses have slightly different focal planes for different wavelengths. If each
channel is projected on different cameras, axial achromatism is easily accommodated for by
axially adapting the camera position. However, as both channels share a similar camera in
the dualview geometry, a small correcting lens can be added in one channel to correct the
mismatched focal planes.
3.3 Which labelling strategy?
The possibility to precisely tag proteins with fluorophores in many different ways enables
the use of SMLM in biology: genetically, immunologically and chemically. This allows to
image non-naturally-fluorescent proteins with single-molecule precision and high accuracy.
7To get the squared precision of the co-localisation, the squared registration precision needs to be added to
the squared precisions of localisation of both co-localised fluorophores (cf. Section 3.4) [107, 117].
3.3 Which labelling strategy? 41
3.3.1 Labelling requirements and strategy
The labelling strategy in SMLM is critical to the precision, the resolution and the accu-
racy of the localisations (cf. Section 3.4.3). Good reviews extensively reference existing
labelling methods in fixed [45, 52] and live [118] cell SMLM experiments, and compare
photo-transformable fluorescent proteins [119]. Here, seven characteristics of the labelling
strategy that require consideration are distinguished:
1. Size of the label: the protein of interest is not directly imaged but the fluorophore that
labels it. The accuracy of the measured position is thus decreased with the distance
separating the label to the protein. Antigen-binding fragments (Fab) of antibodies or
nanobodies [120] should thus be favoured over full antibodies. The use of noncanon-
ical amino acid and ‘click’ chemistry, although genetically heavy to use, is the best
approach limiting the size of the label described so far [57, 58].
2. Specificity: as for classical fluorescence microscopy, specificity of the fluorescent la-
bel to the protein of interest is critical for accuracy of the detection. Antibodies with
very low dissociation constants are thus required. Also, genetically-encoded fluores-
cent proteins (e.g. mEos [121], PA-mCherry [122]) or enzyme-based tags (e.g. Snap-
tags [60], HaloTags [59]) have the advantage of the genetic control and specificity of
the tagging.
3. Perturbation: the addition of a label to a protein can perturbate its function, its fold-
ing, its expression or its cellular localisation. This is particularly known for fluores-
cent proteins that tend to self-aggregate, especially at high densities [123]. Most of
them indeed are monomeric mutants of tetrameric proteins. This characteristic is crit-
ical, as it has led to extensive studies of aggregation and miss-localisation artefacts in
fluorescence microscopy [121, 123, 124].
4. Density: the labelling density directly affects the resolution (i.e. the sampling) of
the experiment (cf. Section 3.4.3). A high density is usually restrained by sterical
hindrance of dense labels (e.g. for labelled antibodies, or fluorescent proteins in virus
particles), or the expression levels reachable for fluorescently-tagged proteins [125].
In the latter case of fluorescent proteins, the folding efficiencies of the protein, of the
tag and of the fluorophore are three factors limiting the final density of proteins that
can be imaged. The aim of a labelling strategy, especially if counting is at stake, is
to label each protein of interest with exactly one and only one fluorophore. Over-
42 Building and characterising a single-molecule super-resolution platform
labelling is also frequent, when antibodies targeting the protein of interest are used
and labelled with dyes through their lysines.
5. Activation control: in SMLM, fluorophores are observed one by one. To reach super-
resolution, dense fluorophores are successively and stochastically switched to their
fluorescent states. Fluorophores are required to be activable or switchable in some
way (chemically, optically, or through thermodynamic binding events). In the case
of photo-activable fluorescent proteins, the activation efficiency is intrinsically lim-
ited [126]. Furthermore, in general, precise switching of the dye is not entirely con-
trollable. This is particularly the case in two-colour imaging, when two different
fluorophores present at different concentrations are to be switched independently at
different rates, in an orthogonal manner. Over- or under-switching limits the resolu-
tion of the experiment by decreasing the number of detected localisations (because of
overlapping PSFs or of under-sampling, resp.).
6. Brightness: because in an SMLM experiment single molecules are imaged one by
one (single molecule detection requires high signal-to-noise ratios), and since the
localisation precision is inversely correlated with the number of collected photons
(Equation 3.13), high photon-flux fluorophores are required. Quantum dots and or-
ganic dyes, compared to fluorescent proteins, are generally brighter and more photo-
stable. The HaloTags and Snap-tags combine the advantages of both the specificity of
a genetically-encoded tag and the brightness of organic dyes, but also the drawbacks
of limited folding efficiency and target accessibility of the tag to the organic dye.
Moreover, the brightness of fluorophores can be increased in specific buffer condi-
tions. For instance, the fluorescence quantum yield (i.e. the probability that a pho-
ton is emitted for each excitation photon) of far-red fluorophores is typically low;
however, non-radiative decay routes can be selectively turned off through the use of
deuterated (heavy) water based buffers, leading to an increase in the quantum yield
by over 200%. This has been shown in multiple common classes of dyes, including
oxazines and cyanines [127, 128].
7. Live-cell compatibility: the last characteristic specifically concerns live-cell imaging.
Not only some fluorophores are cytotoxic (e.g. quantum dots) but most dSTORM
buffers, required to control the blinking of organic fluorophores, are too. Addition-
ally, if the protein of interest lives inside the cell, cell-permeable fluorophores are
required. Intracellular live-cell labelling with organic dyes or quantum dots is still a
3.3 Which labelling strategy? 43
real challenge in SMLM. The compatibility of photo-activable or -switchable fluores-
cent proteins with intracellular labelling and live-cell imaging makes these convenient
to use.
Because fluorescent proteins are compatible with live-cell intracellular imaging, and
are highly specific labels of moderate size, mEos was chosen as the main fluorophore for
the experiments presented in this thesis. Indeed, mEos is an efficient photo-activable fluo-
rophore [74, 123], and the brightest described so far [18, 46, 94, 129]. Although early
constructs (mEos, mEos2) tended to dimerise at high concentration, true monomeric mEos
proteins (mEos3.1 and mEos3.2) were engineered [121], making mEos the most used photo-
activable fluorescent protein in SMLM applications [101, 123, 130, 131]. The choice of an
irreversibly photo-activable fluorophore is also important for the quantitative PALM ap-
proach investigated in Chapter 6 [132].
However, mEos occupies a large space in the wavelength domain [52]: it is activated
by 405 nm illumination; its green inactivated form absorbs around 488 nm and emits in the
yellow domain; its red activated form absorbs around 561 nm and emits in the red domain.
Thus, if the distribution of a second protein of interest needs to be simultaneously observed,
a photo-switchable fluorophore absorbing and emitting in the far red needs to be found. In
this domain, no bright fluorescent proteins allowing single-molecule detection are known
[119]. Thus, in this thesis, labelling with a second colour is obtained by Fabs targeting
the protein of interest and labelled with Atto655. This dye has the advantage of not being
affected by the 405 nm illumination required for mEos activation, but also the advantage
of naturally blinking under 640 nm illumination [25, 133, 134]. Atto655 photo-blinking
is also increased in 40 µM ascorbic acid (i.e. vitamin C) buffer, which is compatible with
live-cell imaging8 [37, 139, 140]. Thus, the control of both Atto655 and mEos activations
are deconvolved and orthogonal, enabling the optimisation of the densities of both activated
fluorophores that are imaged during acquisition.
3.3.2 Ensemble characterisation of the dark states of mEos
Since most experiments of this thesis involve mEos as the fluorophore of choice for single-
molecule experiments, some initial characterisation of its single-molecule photo-physical
8Ascorbic acid is naturally and abundantly present in cells at concentrations of 0.1 to 0.4 mM [135].
Although millimolar concentrations of ascorbic acid are known to affect cell viability, the sub-millimolar
concentrations used to induce Atto655 blinking failed to exhibit any cytotoxic effect after 24 h exposure for
various cell lines [136–138].
44 Building and characterising a single-molecule super-resolution platform
Figure 3.8 Ensemble photo-physics of mEos. Purified mEos2 (A-B) and mEos3.1 (C-D) proteins were acti-
vated by a pulse of 405 nm laser beam and imaged under 561 nm only (A and C) or both 561 and 405 nm
(B and D) illumination. The normalised intensity over 4,000 frames are plotted (grey) and linear combina-
tions of one (red), two (blue) and three (green) exponential decays were fitted. The residuals for each fit are
also plotted. Arrows: region poorly described by a biexponential decay. Inset: two- (ref), three- (blue) and
four-state (green) models for mEos fluorescence. The 405 nm-activated red-emitting fraction of mEos (‘On’)
is photo-bleached (‘Off’) over time, but could also be, under constant illumination, in equilibrium with one
(‘DS1’) or two (‘DS1’ and ‘DS2’) dark states.
3.3 Which labelling strategy? 45
properties were investigated. Indeed, although bulk experiments had been originally pub-
lished [123], very little was known about the molecular behaviours of this new fluorescent
protein.
mEos2 and mEos3.19 [121] proteins were expressed in E. coli and purified by Dr A.
Watson. After dilution in a gelatin gel, each protein was imaged in bulk concentration after
activation with a pulse of 405 nm laser (5 s; 66.3 mW), under either only 561 nm (227 mW)
illumination or both 561 nm and 405 nm (162 µW) illuminations. The variations of the mean
intensities over 4,000 frames (50 ms exposure) were analysed with models of one, two or
three exponential decays (Figure 3.8, with (B, D) or without (A, C) 405 nm illumination).
Consistently with the supplementary figures of ref. [123], the photo-bleaching curve of
the red form of mEos2 is not well fitted to a simple exponential decay. Instead, both mEos2
and mEos3.1 are well and parsimoniously described by three or two exponential decays,
depending on whether the proteins are continuously illuminated with a 405 nm laser beam
or not, respectively. Remarkably, both fluorescent proteins present quantitatively similar
photo-bleaching lifetimes (Table 3.1).
Table 3.1 Models of mEos photo-bleaching. (Values are given ± one standard deviation of the mean for two
fields of view and two biological repeats.)
mEos2 mEos3.1
No 405 nm illumination
f (t) = A e−
t
τ1 +B e−
t
τ2
τ1 = 116 ± 46 s
τ2 = 9,600 ± 8,200 s
τ1 = 135 ± 18 s
τ2 = 10,600 ± 400 s
Continuous 405 nm illumination
f (t) = A e−
t
τ1 +B e−
t
τ2 +C e−
t
τ3
τ1 = 42 ± 27 s
τ2 = 360 ± 60 s
τ3 = 10,800 ± 3,900 s
τ1 = 26 ± 11 s
τ2 = 310 ± 110 s
τ3 = 7,100 ± 1,300 s
The different exponential models can be interpreted as Jablonski diagrams of the fluores-
cent protein with different numbers of non-fluorescent states [141]: the simple exponential
decay characterises a two-state model, for which a initially populated visible ‘On’ state (the
activated red form of mEos) is depleted with a fixed rate (i.e. photo-bleached) over time
(Figure 3.8, inset in B, red rectangle); the addition of a dark state (DS1, blue rectangle) in
equilibrium with the ‘On’ state induces a biexponential decay of the occupation of the fluo-
rescent state; and a second dark state (DS2, green rectangle) is manifested by the ‘On’ state
population following a linear combination of three exponential decays [142].
9Since mEos2 partially dimerises at high concentrations, mEos3.1 or mEos3.2 were preferred when the
constructs were designed after the publication of mEos3.1 and mEos3.2 sequences in ref. [123], mostly in
Chapter 7.
46 Building and characterising a single-molecule super-resolution platform
Although a linear combination of two exponential decays describes well the occupancy
of the red fluorescent form of both mEos2 and mEos3.1 in the absence of 405 nm illumi-
nation (Figure 3.8, A,C), the addition of a third exponential decay is necessary to properly
fit this population when illuminated by both a 405 and a 561 nm beams (B,D − black ar-
rows indicate specific regions that the biexponential decay does not satisfactorily account
for). This 405 nm-dependence could be explained by the 405 nm-induced additional photo-
activation of mEos that re-populates the red fluorescent form of mEos (‘On’ population)
from the green-emitting non-activated pool of mEos acting as a dark state ‘DS2’ under
561 nm excitation. The fitted values of the amplitudes and exponential time constants of the
combination of the two or three exponential decays are given in Tables 3.1 and A.1.
In order to validate this working model of a single dark state, experiments to directly
study the photo-physical properties of separated single-fluorophore were designed and per-
formed.
3.3.3 Single-molecule characterisation of the dark states of mEos
Purified mEos2 and mEos3.1 were diluted in a gelatin gel so that they could be imaged
as spatially separated PSFs under constant 561 nm (227 mW) and 405 nm (power varying
from 7 to 910 µW) illuminations. PSFs were fitted (cf. Section 3.4) and spatially grouped in
trajectories (all localisations closer than 150 nm from each other). The distributions of the
‘on’ and ‘off’ times (tON : number of consecutive frames for which a PSF is detected, and
tOFF : number of frames separating the last detection of a PSF and its next detection in a non-
consecutive frame, respectively) were calculated for thousands of trajectories (Figure 3.9).
Each distribution was then approximated by a linear combination of one or two expo-
nential decays (Figure 3.9, red and blue, respectively − in A-B, for the white histogram).
Although a single exponential approximated well the ‘on’ time distributions (A,C − in C,
the biexponential fit did not converge), the combination of two exponential functions were
required to correctly account for the shapes of the tOFF histograms (B,D − values of the
extracted amplitudes and exponential time constants are summed up in Tables 3.2 and A.2).
In accordance with previous work on the photo-physics of Dronpa [143] and mEos2
[144, 145], the biexponential decay of the ‘off’ time distributions reveals the existence of
two dark states in addition to the irreversible photo-bleached state10. The single exponen-
10The ‘off’ time histogram can be seen as the evolution of a global dark state initially populated and in
equilibrium with the ‘on’ state over time. The biexponential decay indicates two different activation energies
or kinetic mechanisms linking the global dark state and the fluorescent state. Since a usual hypothesis for
fluorescent proteins is that there is a single fluorescent state, two activation energies would reveal the existence
3.3 Which labelling strategy? 47
Figure 3.9 Single-molecule photo-physics of mEos. Distribution of ‘On’ (A and C) and ‘Off’ (B and D)
times of trajectories of spatially-separated purified mEos2 (A-B) and mEos3.1 (C-D) proteins immobilised in
a gelatin gel under both constant 561 nm and varying 405 nm illumination. A,C: normalised distribution of
the number of consecutive frames (of 50 ms) for which single mEos proteins were detected. B,D: normalised
distribution of the time spent by single mEos proteins in a non-fluorescent state between two following fluo-
rescent detections. Experiments were done at different 405 nm illumination power: varying from 7 to 910 µW
for mEos2 (from white to dark grey, A-B) and at 162 µW for mEos3.1 (C-D). The number of trajectories
recorded for each distribution is indicated in parentheses in the legends. Distributions (white in A-B, grey in
C-D) were fitted with both a single exponential decay (red) or a combination of two exponential decays (blue).
48 Building and characterising a single-molecule super-resolution platform
Table 3.2 Models of mEos photo-physics. Exponential time constants τ of the exponential decays fitted to
the distributions of ‘on’ and ‘off’ times as described in Figure 3.9 under different 405 nm activation powers.
Values are given for the fit of a single exponential decay for tON , and a linear combination of two exponential
decays for tOFF .
mEos2 mEos3.1
405 nm tON tOFF 405 nm tON tOFF
7 µW τ1 = 0.77 s
τ1 = 3.20 s
τ2 = 26.1 s 162 µW τ1 = 1.00 s
τ1 = 1.88 s
τ2 = 14.6 s
31 µW τ1 = 0.83 s
τ1 = 2.79 s
τ2 = 20.0 s
86 µW τ1 = 0.87 s
τ1 = 2.62 s
τ2 = 15.6 s
910 µW τ1 = 1.01 s
τ1 = 1.91 s
τ2 = 10.9 s
tial decay fitting the ‘on’ time distributions can be explained by three scenarios: (1) not
enough trajectories were recorded to precisely detect a more complex shape of the distribu-
tion; (2) direct depopulation of the fluorescent state by both dark states, together with the
photo-bleached state, have indistinguishable activation energies; (3) the fluorescent state
is depopulated by the single short-lived dark state that is further depopulated in either the
second dark state, the fluorescent state or the photo-bleached state.
Figure 3.10 Lifetimes of the fluorescent and dark states of mEos2. The exponential time constants of the ‘on’
(A) and ‘off’ (B) time distributions of mEos2 plotted in Figure 3.9, A-B and summed up in Table 3.2 are
plotted in function of the 405 nm illumination power (log scale). For each lifetime, a linear regression (in the
log scale of the x axis) is shown.
of two different dark states composing the global dark state. It has to be underlined that the photo-bleached
state does not appear in the diagram as if photo-bleached, the protein never re-enters the ‘on’ state, which is
necessary to calculate an ‘off’ time.
3.3 Which labelling strategy? 49
Lifetimes of the fluorescent and dark states varied with the power of 405 nm illumi-
nation (Table 3.2). The lifetime τ1 of the fluorescent state increased exponentially with the
405 nm laser power (R2 = 0.981) whereas the lifetimes τ1 and τ2 of the dark states decreased
exponentially (R2 = 0.989 and 0.941, resp.) (Figure 3.10). These observations are consis-
tent with a 405 nm-induced depletion of one dark state [62, 143, 144], inducing longer ‘on’
times and shorter ‘off’ times of the fluorescent protein.
To refine the Jablonski diagram of mEos, more experiments would be required. How-
ever, these experiments justify the use of a low continuous 405 nm activation during PALM
imaging [146], as opposed to pulses of activation [100], especially when counting, and not
localising, is at stake. Indeed, depletion of one of the dark states induces less photo-blinking
(or re-activation) of the same single fluorescent protein, preventing initial over-counting.
Also, an important result is the very similar photo-physical behaviour of mEos3.1 compared
to mEos2. Both are thus used indifferently in this thesis.
3.3.4 Optimisation of mEos photo-activation
The concept of SMLM is based on the idea of stochastically separating in time fluorophores
that are spatially closer from each other than the diffraction limit. In PALM, this is achieved
by turning the fluorescent protein from a dark state (here, the native green form of mEos) to
a fluorescent state (the red form of mEos) with near-UV illumination. The photo-activation
efficiency (i.e. the combination of the folding efficiency of the protein and of the fluo-
rophore, and the photo-conversion efficiency from the dark to the fluorescent state) of the
most used fluorescent proteins were recently measured [126]: only 61 and 41% of expressed
mEos2 and mEos3.2 (resp.) are actually activated and can be observed during a PALM ex-
periment. This intrinsically limits the accuracy of both localisation (lowering the labelling
density) and counting (inducing under-counting that needs to be accounted for) in PALM.
This is a main drawback of PALM imaging.
In order to optimise the photo-activation of mEos during a PALM experiment, the power
of the 405 nm illumination needs to be adapted so that a similar density of fluorophores
gets activated in each frame of the experiment, theoretically equal to the number of photo-
bleaching events per frame. As densities of non-activated fluorophores are high at the be-
ginning of an experiment, a low 405 nm illumination is required so that the probability
of activating one fluorophore stays low; however, at the end of the experiment, very few
fluorophores are left and, in order to optimise the time of the experiment and activate all ac-
tivable fluorophores by the end of it, the power of the activation beam should be increased.
50 Building and characterising a single-molecule super-resolution platform
This optimisation is of critical importance when counting single molecules is at stake, to
prevent both initial under-counting by activating too many fluorophores at once, and late
over-counting by imaging no fluorophore in most of the final frames but single-molecule-
like background noise.
Figure 3.11 Activation rate of mEos2. Purified mEos2 proteins were immobilised in a gelatin gel and imaged
under 561 (200 mW) and 405 nm (powers as indicated) illumination. For each detected trace of the red form
of mEos, the time of its first appearance was calculated. Their distributions for three repeats and four different
405 nm illumination powers were fitted with an exponential decay. The corresponding exponential rates (or
activation constants) are plotted. A linear regression is shown (R2 = 0.983).
Lee et al. describe a Fermi activation curve that activates a constant number of fluoro-
phores during an entire SMLM experiment [145]. Such an imaging protocol also optimises
the imaging time, and consequently the size of the data to be stored and analysed. In order
to apply this shape of activation, the calibration curve of the effect of the activation power
on the activation rate was measured (Figure 3.11). Consistently with supplementary data
from ref. [145], a linear correlation between the activation rate ka of mEos2 and the 405 nm
illumination power (R2 = 0.983) was observed. This enabled the optimisation of mEos
photo-activation in counting experiments (cf. Chapter 6), varying the power of the 405 nm
3.4 Analysis 51
illumination over time using the following formula (Equation 5 from ref. [145]):
ka(t) =
1
T
e−(t−tF )/T
[1+ e−(t−tF )/T ] log[1+ e−(t−tF )/T ]
(3.11)
with T the fall-off time of the Fermi function, and tF its half-time.
3.4 Analysis
Because of the wave nature of the light, the distribution of photons emitted by a single
fluorophore interferes with itself. This mathematical description of the propagation of a
light ray as a wave explains the diffraction patterns of the intensity of a single fluorophore
recorded on a detector: the point-like fluorophore is observed through any focussed optical
system as a blurred spot, or point-spread function (PSF), of width ∼200 nm (Equation 2.1,
for visible wavelengths). In order to localise the position of the fluorophore in the imaging
plane with greater precision than the observed blur, each PSF needs to be detected and can
be fitted with a model PSF. This section describes and discusses the algorithm used for
SMLM processing in this analysis, and the post-processing tools that were developed to
further analyse the list of localisations obtained into trajectories or clusters, or to compare
the localisations of two differently-labelled proteins.
3.4.1 Fitting routine
The concept of SMLM relies on separation of fluorophores in time. Thus, multiple frames
are recorded to rebuild one single super-resolved picture. In each frame, a sub-set of fluo-
rophores are activated and visible as separated PSFs. A fitting routine is used to find and fit
each PSF with a 2D-Gaussian and store its precise lateral position. Good reviews describe
and compare existing algorithms [147, 148]. Below, two routines are described and com-
pared: a bespoke piece of software that was first written to analyse data and understand the
challenges and limitations of fitting; and a quick and efficient fitting routine called Peak Fit
and developed by Dr Alex Herbert at the University of Sussex [149] that was later used and
incorporated in all further post-processing scripts for convenience and rapidity reasons. All
results in this dissertation have been analysed with Peak Fit, using the parameters listed in
Table A.3. Both routines are written as ImageJ [150] plugins and follow a very similar se-
quential approach (Figure 3.13, A): each frame is first filtered through a Gaussian band-pass
filter in order to detect local maxima; the position of each maxima is then used as an initial
guess to fit a PSF function to the original non-filtered data.
52 Building and characterising a single-molecule super-resolution platform
Table 3.3 Differences between home-made and Peak Fit fitting routines.
Home-made Peak Fit
Algorithm Simplex optimiser Levenberg Marquardt
Method Maximising log-likelihood Minimising non-linear least square error
Parameters Θ1−5,7 (Θ6 = 0) Θ1−7
Bounds
{
All positive parameters,
(x,y) in fitting box Positive intensity
Fit check None Position shift, signal strength and width
Stop decision All selected maxima When 10 successive fits failed
• Gaussian band-pass filtering: two copies of each frame (Figure 3.13, C) are made,
and each copy is respectively convolved with a 2D-Gaussian kernel smaller than the
expected width of the PSF to get rid of high-frequency electrical noise, or with a 2D-
Gaussian kernel larger than the width of the largest expected PSF (Figure 3.13, D).
The latter is subtracted from the former in order to remove low-frequency biological
noise, giving a filtered frame as shown in Figure 3.13, E. This convolution process
amounts filtering the Fourier transform of the frame with a Gaussian band-pass filter,
but runs faster than calculating two Fast Fourier transforms for small kernels [151].
• Maxima selection: local maxima of the filtered frame are listed (Figure 3.13, F). Two
different strategies followed: in the home-made routine, local maxima were rejected
if they were below an empirically-defined threshold (1.2 µBackground + 5 σBackground,
where µ is the mean background of the frame and σ , the standard deviation of the
frame); in Peak Fit, local maxima are ranked by their intensity value.
• Fitting routine: for each local maximum, a region of the unfiltered initial data around
the maximum was selected and further fitted with a theoretical PSF. Although an Airy
disc PSF can be used, fitting a Gaussian PSF is quicker while it gives similar results
[152], and was thus preferred (Figure 3.13, G):
PSFΘ1−6(x,y) =
Θ5
2π Θ3 Θ4
e
−(x
′−Θ1)2
2 Θ23
− (y
′−Θ2)2
2 Θ24 (3.12)
with
(
x′
y′
)
=
(
cosΘ6 sinΘ6
−sinΘ6 cosΘ6
) (
x
y
)
where (Θ1, Θ2) is the centre of the PSF, (Θ3, Θ4), its two elliptical half-widths, Θ6,
3.4 Analysis 53
the rotation angle between its two perpendicular axes and the (xy) axes, and Θ5, its
integrated intensity. For the home-made routine, Θ6 was fixed to 0, limiting the num-
ber of parameters to fit. The background noise, or offset of the Gaussian function, is
added as a last additive fitted parameter in Equation 3.12 in Peak Fit. Then, the least
squared error between the estimated Gaussian function and the data is minimised by
a Levenberg Marquardt algorithm. Differently, the home-made routine makes the hy-
pothesis that the background noise is Poissonian and consequently incorporates it in
the log-likelihood estimator described by Abraham et al. [152], which is maximised
by a Simplex optimiser. The log-likelihood estimator theoretically allows a more pre-
cise estimation of the positions [152] but requires a good model of the noise of the
camera.
A last difference between the two fitting algorithms resides in the decision of which
maxima to fit. The home-made routine a priori selects the maxima thanks to an empirical
threshold before fitting them all. Peak Fit orders the maxima by intensity and fit them until
ten successive fits fail. A fail is defined by empirical thresholds on different parameters such
as the signal-to-noise ratio, the width, or the displacement of the position of the PSF after
fitting (see Table A.3 for the thresholds that were used in this thesis).
Finally, both fitting plug-ins output a list of fit positions and parameters to be post-
processed (e.g. Figure 3.13, H), before rebuilding the final super-resolved picture.
3.4.2 Post-processing
Once the SMLM data has been processed, the list of localisations usually requires some
post-processing: localisations can be linked in trajectories to determine their diffusion be-
haviour, or grouped in clusters to spatially compare effects of different conditions. If two
proteins labelled with two different fluorophores are simultaneously imaged, determining
how much co-localised they are is also commonly expected, although not well-defined.
Single particle tracking
Single particle tracking is of key importance for quantitative analysis of live-cell SMLM
data [49, 79, 153, 154]. Many algorithms, varying in (1) how they detect particles and
(2) how they link the particles from different frames in trajectories, have been proposed
(compared and reviewed in ref. [78]). Comparison of the different existing methods reveals
that the choice of the tracking algorithm is critical when the signal-to-noise ratio (SNR,
defined in [78] as SNR = (I0− Ib)/
√
I0, I0 denoting the amplitude of the PSF and Ib, the
54 Building and characterising a single-molecule super-resolution platform
Figure 3.12 Single-particle tracking algorithm. The three validating steps of the single-particle tracking analy-
sis used in this thesis are shown for data from one representative experiment (from Chapter 7). A: Distribution
of the signal-to-noise ratios (SNR = (I0 − Ib)/
√
I0, I0 denoting the amplitude of the PSF and Ib, the mean
background intensity) for all localisations fitted in the SMLM experiment. B: Distribution of the step sizes of
all built trajectories (◦). A characteristic lognormal distribution was fitted to the data (grey line, R2 = 0.988).
C: One representative frame of the SMLM time-lapse experiment with several activated but separated mEos
proteins. The trajectories of all localisations detected in the frame are plotted (one colour per trajectory). Tra-
jectories are shown up to the frame shown. D: Zoom-in of the white square in C. All the 1,555 frames are
available in Movie B.1.
3.4 Analysis 55
mean background intensity) is lower than 4. In the experiments of this thesis, SNR ≥ 7
(Figure 3.12, A), thus a simple tracking algorithm was chosen, in which localisations in
successive frames are linked in a trajectory if a localisation in one frame falls within some
maximum distance (here, 300 nm) of a localisation in the previous frame (as described in
ref. [154]).
The low density of the localisations in the SMLM experiments, together with the slow
diffusion of the transmembrane proteins of interest (cf. Chapter 7, for which measured
diffusion coefficients fall below 0.08 µm2/s), makes also the choice of a complex algorithm
based on prior knowledge of the data less relevant. The choice of the trajectory radius
(300 nm, less than the dimension of three pixels of the camera) was confirmed a posteriori
by examining the resultant step distribution, which could always be fitted with a log-normal
distribution (R2 ≥ 0.95, Figure 3.12, B). Furthermore, the trajectory radius was chosen to
allow the detection of diffusion coefficients up to 0.1 µm2/s (at a frame rate of 20 Hz), which
is in the upper range of the diffusion coefficients of membrane receptors [79].
More specifically, trajectories were defined as follows (Figure 3.13, B): for each loca-
lisation, all n localisations within 300 nm from its positions in the previous frame were
identified; then, three cases were found: (1) if n = 0 (no localisation detected in the previ-
ous frame), the process was repeated with the second previous frame (up to four frames of
memory, to allow for short blinking of the fluorophore and partial efficiency of the fitting
algorithm); (2) if n = 1, the localisation was linked to the single localisation detected in
the previous frame; (3) if n > 1 (i.e. the tracks merge), the localisation was linked to all
detected localisations from the previous frame and all the corresponding trajectories were
terminated to avoid artefacts. A characteristic example of CD28-mEos tracks is given in
Figure 3.12, C-D and Movie B.1.
Cluster grouping
Once a SMLM experiment is analysed, the list of localisations can be plotted in a super-
resolved picture or quantitatively analysed: a cluster analysis of the distribution of localisa-
tions is often essential, either to group multiple occurrences of a single emitter (PC-PALM)
[82, 155], or to compare the aggregation or multimeric states of proteins in different condi-
tions. Such algorithms are based on either:
1. the iterative search and grouping of all localisations within a defined distance d (Fig-
ure 3.13, K− around an iterative weighted mean position of the grouped localisations
[130, 131, 145, 155, 156], or by coalescing all localisations closer to at least one
56 Building and characterising a single-molecule super-resolution platform
Figure 3.13 Single molecule experiments require intensive data processing for any data analysis. In SMLM, a
fitting routine is used to extract the positions of single blinking fluorophores from a diffraction-limited movie
(A, flow chart of the fitting process used in this thesis): the background (D) of each diffraction-limited frame
(C) is filtered out (E) for its local maxima to be detected (F, local maxima represented as red crosses). Each
local maximum is then fitted with a 2D-Gaussian (G, fitted 2D-Gaussians). Finally, the output localisations
are post-processed (H, fitted 2D-Gaussians after amplitude and width thresholding). Together with rebuilding
the final super-resolution picture, post-processing can require tracking fluorophores that appears in successive
frames (B, flow chart of the simple but robust algorithm used in this thesis) or grouping fluorophores in clusters.
For such latter analysis, efficient representation of the sometimes very large data (> 100, 000 localisations) is
necessary: a QuadTree representation of the data (I, close-up of the squared region in J) was developed. This
allowed fast grouping of localisations within a defined distance (K) or fast searching of kth nearest neighbour
(L) (speed in N lnN instead of N2).
Legend: green arrows in A-B, path if true; red arrows, path if false; transparent red rectangles in K-L, sub-
regions explored by the algorithm; red line in L, path followed by the algorithm.
3.4 Analysis 57
localisation of the group than d [146, 157]);
2. or on a nearest-neighbour density approach (Figure 3.13, L). This was described in
a density-based spatial clustering of applications with noise algorithm (DBSCAN)
[158] which uses the distance of a point p to its kth nearest-neighbour to decide
whether p is part of a cluster, or part of the data noise. This is a robust, fast, size-,
density- and shape-independent clustering method.
For either approach, a brute force algorithm exploring and comparing the distance be-
tween each pair of N localisations has a complexity, and speed, varying in O(N2). For small
dataset, such speed is manageable but when the list of localisations increases (N > 100,000),
clustering can take a few hours up to days. A way to increase the speed of the analysis is to
re-organise the sparse data using its spatial information. In this thesis, a quad tree [159, 160]
or Q-tree approach is proposed to efficiently represent the data and to speed the spatial clus-
tering analysis up to a complexity in O(N lnN).
A quad tree consists of a tree whose leaves contain between 0 and m localisations. Each
branch of the tree divides itself in four branches or leaves that correspond to four spatial
quadrants (north-west, north-east, south-east and south-west). Each branch of a quad tree is
represented in Figure 3.13, I-J as a blue cross dividing the branch it is rooted in, in four. The
last branches, or leaves, contain the localisations (up to m = 10 black circles). The value
of the size m of the leaf is a compromise between a fast memory-consuming ordered tree
(m = 1) and a slow list-like tree-leaf (m≥ N); in this thesis, m was fixed to 10.
When a new localisation l is inserted, its leaf is quickly found (O(lnN), i.e. the average
depth of the tree) by following, from the root of the tree, the consecutive branch it spatially
belongs to until the algorithm finds a leaf. If the leaf contains less then m localisations, the
new localisation l is added to the leaf. Otherwise, the leaf becomes a branch and spatially
splits into four new leaves, re-arranging the (m+ 1) localisations it previously contained
(Figure 3.13, J).
Searching for all localisations closer than d from a localisation l consists in measuring
the distance between all localisations in the leaves which spatially intersect the circle of
centre l and radius d (Figure 3.13, K). All branches and their subsequent nodes and leaves
that do not intersect this circle can be rejected as the quad tree is visited from its root up to
its leaves. Such grouping has a complexity in O(d lnN) per localisation to be grouped.
Querying the kth nearest-neighbour of a localisation l can be seen as the search for all
localisations closer than a varying distance d: initially, d =+∞ and the first k visited loca-
lisations are stored in a circular buffer; for each new localisation l′ closer than d from l, l′ is
58 Building and characterising a single-molecule super-resolution platform
inserted in place of the localisation lfurthest in the buffer that is the furthest from l, and d is
modified to d(l, lfurthest). This algorithm allows a dramatic decrease in the number of locali-
sations (or the spatial surface) explored: only the nine red leaves are explored in Figure 3.13,
L (the red path underlines the consecutive localisations lfurthest that actualised the distance
d), instead of the whole plane composed of more than 48 leaves. Consequently, instead
of visiting the N localisations, such algorithm only visits an average of O(k lnN ln(lnN))
localisations per nearest-neighbour query.
Coordinate-based co-localisation
Figure 3.14 Coordinate-based co-localisation. After registration of two channels A and B, SMLM analysis
outputs a list of localisations (A, green and red, resp.). For each localisation Ai from channel A, the local
densities DAi, A(r) and DAi, B(r) in either channel within a radius r ≤ Rmax of Ai are computed. Both densities
are normalised to their values at r = Rmax (B). Spearson’s correlation coefficient between them is calculated
(0.129) and weighted by the distance of the point Ai to its nearest neighbour in the other channel B (A, distance
of 5.37 nm in the example shown). This final value is defined as the coordinate-based co-localisation (CBC)
value of the point Ai with respect to the channel B (here, 0.125).
SMLM is inherently different from classical microscopy or confocal techniques (e.g.
STED, SIM): it outputs spatially sparse and discrete data, a list of localisations (i.e. 0-
dimensional points in a multi-dimensional continuous space). To further analyse SMLM
data, methods that are specifically designed for such types of datasets have to be used [132].
A good example is found for two-colour co-localisation analysis: the super-resolved fig-
ure rebuilt from the list of fitted localisations cannot be used as the input for classical co-
localisation methods (e.g. conventional coefficient calculation by pixel-per-pixel overlay, or
intensity correlation analysis) because the way localisations are rebuilt would importantly
affect their co-localisation [52]. Conceptually, it makes more sense to keep a coordinate-
based approach for the entire analysis, as proposed by Malkush et al. [116].
3.4 Analysis 59
Although co-localisation techniques have been extensively used in multi-colour fluo-
rescence microscopy to gain functional insights from the comparison of the distribution of
various proteins, the concept of co-localisation can reach its limit in super-resolution mi-
croscopy: at the molecular scale, two bound proteins that can be localised with an infinite
precision are not superposed but are located at different positions. Thus, in SMLM, when
localisation precisions reach tens of nanometers (i.e., the size of a large molecular com-
plex), interacting molecules may not be observed as co-localised. Then, a nearest-neighbour
approach may be more valuable than a direct co-localisation analysis. Furthermore, co-
localisation studies are also affected at this molecular scale by the sampling, the size and
position of the label on the protein of interest, its localisation precision and its sub-frame
diffusion [107]. This is why a specific implementation to study the spatial correlation of
dual-colour SMLM data, as proposed by Malkush et al. [116] is required. By extension
and analogy to the bulk microscopy concept, this algorithm is called ‘coordinate-based co-
localisation’ (CBC) in this thesis.”
For each localisation Ai of the channel of reference A, the local densities DAi, A(r) and
DAi, B(r) of the localisations of each channel, within an increasing radius r ≤ Rmax, are cal-
culated and normalised to their values at r = Rmax (Figure 3.14). Spearman’s correlation
coefficient SAi between both normalised densities is computed and weighted with a coeffi-
cient e−EAi, B/Rmax (EAi, B being the distance between Ai and its nearest neighbour in channel
B) to prevent false positive co-colocalisations if the local density in the channel B is null.
Through the weighting factor, Rmax defines the resolution of the CBC analysis, espe-
cially for low-density data for which the nearest neighbour of the other species could be
potentially far away for many localisations. Therefore, its value should be chosen to re-
flect the average localisation precision of both channels and their registration precision (see
Section 3.4.3), giving an overall expected CBC resolution.
The final weighted correlation coefficient, taking values between −1 (anti-correlated
distributions), through 0 (non correlated distributions), to +1 (perfectly correlated distribu-
tions), is defined as the CBC value of the localisation Ai with respect to channel B. This
allows to affect a co-localisation value for each super-resolved localisation separately, for
each channel. Moreover, this co-localisation value is independent of the representation of
the single-molecule data.
60 Building and characterising a single-molecule super-resolution platform
3.4.3 What is precision
The purpose of SMLM is to image structures with a precision better than the diffraction
limit of visible light. A clear definition of precision is worth being carefully discussed,
together with its limits and its common misuses. The precision of a localisation is exper-
imentally accessed by imaging a fiducial marker multiple times. Because of the electrical
noise of the detector, the PSF is fitted slightly differently in each frame, inducing a vari-
ation in the estimation of its localisation, which is by definition the precision of the lo-
calisation. The position of the localisation is usually assumed to be normally distributed
(Figure 3.15, A). Although in part of the literature a full-width-at-half-maximum (FWHM)
is sometimes used, in this thesis the precision σ of a localisation refers to the standard de-
viation of the distribution of its fitted position. In case of a normal distribution, however,
FWHM = 2
√
2 ln2σ ∼ 2.35σ .
Importantly, the precision of a localisation has to be distinguished with the fitting error:
although both depends on the intensity of the fluorophore and the background noise, the
fitting error depends on the estimator that is maximised (e.g. least squared error, likelihood)
and on the number of degrees of freedom of the fitting function (e.g. symmetric, or asym-
metric Gaussian). A high number of fitting parameters can lead to some over-fitting of the
fluorescent signal. In this scenario, the fitting error will be very low, but will not represent
any physical precision.
However, since a high signal-to-noise ratio is required in SMLM and because most ex-
periments have a time-constraint, the position of each single fluorophore cannot be assessed
in enough consecutive frames to get a distribution of localisation positions: (1) in fixed
PALM or STORM experiments, the aim is to record a full duty cycle in a single frame to
limit the noise accumulated on each frame recorded by the EMCCD; (2) in sptPALM, live
STORM or uPAINT, each diffusive fluorophore is imaged for a few frames, but because
of its diffusion, the position of the fluorophore changes over time. Therefore, a theoretical
formula which, from one single localisation, describes the statistical distribution it origi-
nates from and gives its precision from accessible parameters is essential. Mortensen et al.
developed such a formula summarised in the following equation [149, 161]:
σ2th = 1.83
σ˜2
N
(
16
9
+
8π σ˜2 b2
N a2
)
(3.13)
where σ˜ is a length constant defined as:
σ˜2 = σxσy+
a2
12
3.4 Analysis 61
and
N =
Amplitude ·2π σxσy
Gain
Interestingly, the precision σth is inversely correlated with the total integrated intensity
N and correlated with the noise of the background b. Indeed, the more detected photons
above the background, the better the distribution of the PSF is evaluated and the better the
position of the localisation can be known. The dependence on the pixel size a, i.e. the
physical size of the conjugated image of one pixel of the camera in the focus plane, reveals
the existence of an optimal pixel size (σth diverges when a → 0 and a → +∞): physically,
if the pixel size a is too large, the whole PSF of the single fluorophore will fall into a
single pixel and no information above the background noise will be available to determine
the sub-pixel localisation of the centre of the PSF; inversely, if the pixel size is too small,
the whole PSF will be divided between a large number of pixels which will only receive a
low number of photons. If this number is too low, it will quickly fall below the constant
background noise observed per pixel and no spatial information will be available to localise
the sub-pixel position of the centre of the PSF.
Ober et al. [162] demonstrate that the optimal pixel size is ∼100 nm/pixel (correspond-
ing to a magnification of 50x for pixels of 5 µm) for a GFP-like emitter observed through a
1.4-na objective lens. Although the objective lens used in this thesis has a slightly higher nu-
merical aperture (na = 1.49), the optimal pixel size, which varies similarly than the width of
the PSF of the instrument (cf. Equation 2.1), will not differ significantly. Since the physical
size of the pixels of the EMCCD cameras in this thesis is 16 µm, with the magnification of
the 60x objective only, the pixel size would be 267 nm/pixel. Adding a 2.5x beam expander
leads to an optimised pixel size of 107 nm/pixel (110 nm/pixel as measured with a Ronchi
slide).
Nevertheless, Equation 3.13 is an estimate of the best theoretically accessible precision.
Although this equation is used as such in many publications to calculate super-resolved pre-
cisions (e.g. in PC-PALM), it seems important to compare it to actual data and possibly
adapt it to the characteristics of the imperfect optics of the microscope. Diffraction-limited
beads labelled with fluorophores were imaged for several consecutive frames under constant
excitation. After PSF fitting and grouping the localisations of the fixed fiducials, the stan-
dard deviations of the successive positions of each bead were plotted against their average
fitted intensity (Figure 3.15, B, black dots).
As expected, this experimental precision follows a similar curve than the theoretical
precision, calculated with Equation 3.13 for the average fitted intensity, noise and PSF width
of each trajectory (grey dots). Both precisions are well fitted and described by a function
62 Building and characterising a single-molecule super-resolution platform
Figure 3.15 Experimental localisation precision and bias of the fitting algorithm. Separated fluorescent sub-
diffraction carboxylated beads were fixed on a poly-lysine coated coverslip and imaged for 1,000 frames
(33 ms exposure) under varying 561 nm illumination (A-B). A: All localisations of one fiducial marker, orig-
inating from the 1.000 frames of one experiment, are plotted (•). B: Localisations for each experiments were
spatially grouped in trajectories (all localisations within 110 nm from each other). For each trajectory, a the-
oretical precision (grey dots) was calculated using Equation 3.13 and the average intensity, width and noise
of all the localisations of the trajectory. The experimental precision (black dots) was also calculated as the
standard deviation of the successive positions of all localisations of the trajectory. Both precisions were fitted
with the following function: y =
√
A/x+B/x2 (grey and black lines, resp.). Finally, Equation 3.13 is also
plotted for the average value of intensity, width and noise of all plotted trajectories (blue line).
Fluorophores were simulated 100 times in silico at the positions of a 11x11 square grid sampling a quarter of a
pixel, in order to test the fitting algorithm for its sub-pixel accuracy and possible structural bias. C: Simulated
positions are plotted as black dots (ground truth). Blue and orange circles are centred on the average position
of the corresponding fitted localisations and their radii are proportional to the standard deviations (on the x and
y axes, resp.) of the fitted positions. D: Distribution of the distance between the ground truth and the average
fitted position for all 121 simulated positions. E: Distributions of the standard deviations (or bias) of the fitted
positions.
3.4 Analysis 63
with same dependencies on the intensity than the previous equation: σ =
√
A/N+B/N2
(black and grey lines, resp.). Thus, the fitted function describing the experimental precision
in function of the number of collected photons could be used instead of the theoretical
estimate for all SMLM data taken on a similar microscope.
However, although used to determine the theoretical precision for each trajectory inde-
pendently, Equation 3.13 does not describe either theoretical or experimental precisions if
plotted as function of the intensity N only (σ = f (N), blue line) with fixed background noise
and PSF width (fixed as the average value over all trajectories). This reveals the importance
of both noise and PSF width variables, critical in this equation. Noticeably, the fitted ex-
perimental function does not take into account these dependencies. Thus, in this thesis, the
localisation precision was defined as the theoretical precision. Although under-estimating
the error in the localisation, it describes precisely the main dependencies of the localisation
precision on the parameters of the PSF fit. Moreover, with the calibration curve presented
here (Figure 3.15, B), an estimation of the error made with this approximation of the loca-
lisation precision with the theoretical precision can be determined: an under-estimation of
∼ 5-10 nm is made.
Experimentally, the localisation precision can also be affected by different parameters
such as diffusion blurring [163] if the sample is not fixed, or the dipole orientation of the
fluorophore [98] if it does not rotate freely. Additionally, if the positions of localisations in
two different channels are compared, the square of the precision of the registration of both
channels needs to be added to the squares of the localisation precisions of both fluorophores
(cf. Section 3.2.3) [107].
Finally and importantly, distinction between sometimes confused concepts needs to be
clarified [52, 164, 165]:
1. The localisation precision concerns how reproducible the repeated localisation of a
fluorophore is (it depends on the number of collected photons, on how they are spread
on the detector and on the background noise);
2. The resolution of an experiment concerns the size of the smaller structures or features
observed in the final rebuilt super-resolved figure (it depends on both the precision and
the density of the localisations [121]). Fourier ring correlation analysis, inspired from
a common method in electron microscopy, is a robust way to estimate the resolution
of an experiment [166].
3. The accuracy of an SMLM experiment concerns the mean distance between the es-
timated localisation and the ground truth, i.e. the position of the protein of interest,
64 Building and characterising a single-molecule super-resolution platform
and measures inherent bias of the method (it depends on both the size of the label
and its linker to the protein of interest, and systematic bias of the fitting localisation
estimator).
These three parameters are important in an SMLM experiment and should be kept in mind.
The accuracy of the fitting localisation estimator was briefly studied: fluorophores were
estimated at varying positions on a 11x11 square grid spanning a quarter of a pixel of the
camera (Figure 3.15, C). At each position, 100 frames were simulated, then fitted with Peak
Fit (cf. Section 3.4.1). No symmetric bias was observed in either the average localisation
position or its distribution (Figure 3.15, C). A symmetric error would be a marker for a
systematic intrinsic error induced by the symmetric fitting estimator. An estimation of the
average error made (Figure 3.15, D, less than 1 nm) and of its spread (Figure 3.15, E,
distribution of the standard deviations of the x and y errors for all simulated positions, peak
at ∼ 3.2 nm) reveals that the systemic bias of the algorithm is negligible compared to the
typical localisation precision of SMLM experiments presented in this thesis (on the order of
30-40 nm).
To conclude, in this thesis, the localisation precision is defined as the theoretical preci-
sion σth (Equation 3.13). The precision of an SMLM experiment is the value at the peak of
the histogram of σth for all localisations. However, when possible, the distribution of theo-
retical precision is compared to the distribution of the standard deviations of trajectories of
fixed fluorophores (i.e. an estimation of the experimental precision). An interesting new ap-
proach has recently been proposed by Endesfelder et al. [167]: the nearest-neighbour-based
approach (NeNA). Because of time constraint, this direct, robust and useful method, with no
need for calibration or prior knowledge of the sample or fluorophores, was not implemented
in this thesis. This would be, however, an important tool to use in further experiments.
3.5 Automation
One major improvement allowed by SMLM over classical microscopy is the possibility
to observe rare behaviours of molecules that would otherwise be missed in the ensemble
averaging of the entire population. However, a major drawback of SMLM compared to
conventional fluorescence microscopy is that very few cells are imaged as each acquisition
takes much longer. For each imaged cell, much more data is recorded and must be stored.
Also, since magnifications are higher, fields of view are usually much smaller. Therefore,
as distributions of proteins in a single cell can be observed more precisely, separating each
single behaviour, the variation of these behaviours in a population of cells, which is required
3.5 Automation 65
Figure 3.16 Designing an automated imaging procedure. A: Flow chart of the automation algorithm designed
for ‘smart’ large-scale data acquisition (Software B.3). Capital letters indicate the corresponding panels in
Figure 3.17. fw, fc: Focal planes of the well and of the cell, resp.. B: White-light transmission snapshot of
part of a micro-fluidic device with five rows of trapping sites. Direction of the flow is shown by white arrows.
Scale bar: 50 µm
66 Building and characterising a single-molecule super-resolution platform
to draw any conclusion, proves to be a real challenge, especially if the cellular population
of interest is itself rare in the general population of cells that is observed.
This is the case of yeast cells in SMLM live imaging. Yeast cells are usually mounted
between an agarose pad and the glass coverlips, to immobilise and image them live [130,
168, 169]. However, when interested in a specific phase of cell division, one needs to
manually find and select such cells one by one (the division phases of S. pombe cells are
strongly correlated with a specific cell size [170]) and acquisition cannot be automated.
An important evolution in the method was the design and introduction of a micro-fluidic
device through which separated yeast cells can be flown until they get individually trapped
at specific positions where they grow while they can be imaged [100]. By knowing the
positions of the discreet sites where cells can be trapped, an automated microscope stage
can visit and thus allows to image many trapped cells. The aim of the algorithm described
in this section (Software B.3) is, once initially focussed on the first well in the first row of the
device, to visit each trapping site, determine whether a cell is trapped and if so, whether it
has a pre-determined length of interest (i.e. whether it is in a specific phase of the cell cycle)
(Figure 3.16, A). If this is the case, then the cell can be imaged in a PALM experiment.
To account for the fact that both (1) the position of the trap site within the well varies
along a row (Figure 3.16, B, the depth of the sites varies in each row to keep a same trapping
pressure in each site) and (2) the axis of the trapping device may be at a slight angle relative
to the detector (e.g. Figure 3.17, A-C), moving from one well to the next is done in two
consecutive steps: firstly the (xy) stage moves to the theoretical place of the next well; then
a white-light transmission snapshot (to prevent any pre-activation of the fluorophores from
light of wavelength below 610 nm, a long-pass filter is used, Figure 3.17, A) is taken and
binarised to detect the main characteristics of the well (four horizontal lines and one vertical
one, Figure 3.17, B); the position of the (xy) stage is finally adjusted to precisely place the
centre of the trap in the centre of the camera chip (Figure 3.17, C). In this way individual
S. pombe cells always appear in the centre of the field of view independently of the trap the
cell happens to be trapped in.
For various reasons (stability of the optical components of the PALM platform, tem-
perature or oil drift, non-orthogonal sample mounting, etc.), the focus can repeatedly drift
on large spatial scales (typically over more than five to ten wells). An autofocus function
was thus added to the algorithm to correct for this instability: a transmission z-scan (Fig-
ure 3.17, D-E) is taken to define the focal plane of the well and, if a cell is present, of the cell
(which usually differs from the former, since the cell can be trapped at slightly different ax-
ial planes in the trapping site). The autofocus analysis is based on contrast maximisation on
3.5 Automation 67
Figure 3.17 Automated imaging of live yeast cells. Live yeast cells were trapped in several sites of a micro-
fluidic device. The different stages of the the automation algorithm are described with one representative
cycle observed during the course of an experiment. For each cycle, the (xy) computer-controlled microscope
stage moves to the theoretical position of the new well, takes a transmitted light snapshot (A), analyses it by
recognising characteristics of the borders of the well shown as fine lines (B) and adjusts the (xy) position of
the stage to move the centre of the trap to the centre of the field of view and therefore the camera sensor (C).
A transmitted light z-stack is then taken (D-E), allowing the determination of both focal planes of the well and
the possible cell by minimising the contrast (F) of both sub-regions of the well (blue rectangle, E) and of the
potential cell (red rectangle). This analysis determines whether a cell has been trapped in the centre of the well
and, if so, the length of the cell is calculated. Then, in the case of the detection of a cell, the algorithm focuses
the microscope on the focal plane of the cell and starts a PALM acquisition (G: standard deviation projection
of 50 frames). Otherwise, it focuses on the focal plane of the well before moving to the next well. Movie B.2
displays the visit of four consecutive trapping sites in real-time. Scale bars: 10 µm.
68 Building and characterising a single-molecule super-resolution platform
sub-regions of the field of view over 9 µm above and below the initial z0 plane (Figure 3.17,
E-F) [171]. A trapped cell is detected if the contrast variation of the sub-region of the area
of interest is greater than 1.5.
Then, if a cell has been detected, its width and total length are measured. Finally, if the
cell has the specific size of interest (corresponding, for example, to the important but short-
lived phase G2 of the yeast division cycle [130]), PALM acquisition is started (Figure 3.17,
G). This ‘smart’ automation algorithm allows one to ‘search out’ and image a specific sub-
population of cells from a large sample while optimising acquisition time, computer mem-
ory use and analysis computing time (Movie B.2).
This section describes a relatively simple and specific SMLM experiment for which an
automated design increases the size of the data that can be collected per sample by more than
an order of magnitude and decreases requirements in term of resources (number of samples
to prepare, data storage, time). This is a single example to underline the importance of
automation in SMLM. Bridging single-molecule techniques and high-throughput automa-
tion [102, 172, 173] will allow to gain a deep and quantified understanding of rare molecular
events on the scale of cell populations, while avoiding two of the major drawbacks of current
microscopy: the user bias in selecting cells to image and the low-throughput of the technique
[174]. Moreover, with the coming of faster and faster algorithms [149, 175, 176], an on-
the-go SMLM analysis is almost presently available, opening the door to decision-making
algorithms that will decide when to stop imaging (when local densities of localisations reach
a determined value) [177], how to adapt, in real time, the laser powers of the activation and
illumination beams in order to activate an optimal fraction of fluorophores and track them
for a given number of frames (cf. Section 3.3.4), etc.
However, such an automation procedure requires the efficient positioning of tens of
cells in known sites or wells. The design of such wells is itself a complex field of research,
especially in the case of mammalian cells. Indeed, contrarily to yeast cells, mammalian
cells present a wide variation of sizes and shapes, and develop strong, diverse and complex
interactions with the surfaces they meet. These two separate reasons explain why this au-
tomation step, although demonstrated here in the case of yeast cell imaging, was not further
used in this thesis which focusses on applications of SMLM to T-cell biology. Nevertheless,
were a device available for efficiently positioning T cells without activating them (cf. Sec-
tion 6.2.3), this automation procedure would greatly benefit the results of Chapters 6 and
7.
Chapter 4
The virtual-‘light-sheet’
Most data from this chapter has been published in PLOS ONE as an article entitled "Virtual-
‘Light-Sheet’ Single-Molecule Localisation Microscopy Enables Quantitative Optical Sec-
tioning for Super-Resolution Imaging" [2]. The original idea was from Dr S.F. Lee. All
experimental design, data analysis, and code writing was undertaken by the author of this
thesis. Dr S.F. Lee and M. Palayret designed and built the super-resolution instrument used
in this chapter. H. Armes imaged the S. pombe cells made by Dr A.T. Watson and presented
in Figure 4.8. Dr S. Basu imaged the stem cell presented in Figure 4.9. Dr U. Endesfelder
and T.J. Etheridge took some of the calibration z-scans analysed in Figure 4.6. Dr S.F. Lee
imaged the Tau-labelled monomers analysed in Figure 4.2, B. All other experiments were
done by M. Palayret. Finally, Dr A. Herbert gave precious advice in algorithm optimisation,
and M. Tracey is responsible for the final design of Figure 4.1, A and 4.9, B.
4.1 Introduction
Two limiting factors have driven technical developments in fluorescence microscopy: con-
trast optimisation and precision. Both innovative illumination and emission techniques have
been developed to increase the signal-to-noise ratio of the image: restricting the illumina-
tion to a strict portion of the in-focus sample imaged by the detector (e.g. confocal illu-
mination, TIRF, HILO, SPIM) decreases collection of light originating from out-of-focus
emitters; secondly, spatially restricting the fluorescence signal to reject all light originating
from above or below the focal plane thanks to a pinhole or a slit placed in the image plane of
the microscope (confocal detection) drastically increases the signal-to-noise ratio. Contrast
optimisation has become a main challenge with the observation of ever-smaller biological
structures, down to single-molecule tracking and imaging: structures with fewer proteins
70 The virtual-‘light-sheet’
of interests are tagged with fewer fluorophores which results in lower signals for similar
background noise.
Interest in smaller biological structures was challenged by the second limiting factor of
fluorescence microscopy: the Abbe diffraction-limit of the light (cf. Section 2.1.2). The
resolution, and the precision, of a microscope is intimately linked to the Abbe limit as two
objects cannot be separated and distinguished if their relative distance is smaller than the
width of their PSF, up to a prefactor (2.90 for the Rayleigh criterion, 1.44 for the Sparrow
criterion, cf. Section 3.4.3). One of the successful techniques developed to bypass the Abbe
limit is to sequentially image the single emitters one at a time in SMLM (cf. Section 2.2.1).
However, in SMLM, optimising the signal-to-noise ratio is essential for any substantial
increase in precision (the precision is inversely proportional to the square-root of the signal
above background − Equation 3.13). To improve it, restricting the sample illumination is
necessary (TIRF, HILO or SPIM). But as SMLM is a wide-field technique, a pinhole cannot
be used in the emission path1.
In this chapter, a method that enables 3D sectioning with SMLM is demonstrated that
uses fitting information that is usually discarded to reject fluorophores that emit from above
or below a virtual-‘light-sheet’ (vls), a thin volume centred on the focal plane of the micro-
scope. An easy-to-use routine (implemented as an open-source ImageJ plug-in) is described
that quickly analyses a calibration sample to define and use such a vls. In addition, the plug-
in is easily applicable on almost any existing 2D super-resolution instrumentation. This op-
tical sectioning of super-resolution images is achieved by applying well-characterised width
and amplitude thresholds to diffraction-limited spots that can be used to tune the thickness
of the vls. This allows qualitative and quantitative imaging improvements: by rejecting out-
of-focus fluorophores, the super-resolution image gains contrast and local features may be
revealed; by retaining only fluorophores close to the focal plane, vlsSMLM improves the
probability that all emitting fluorophores will be detected, fitted and quantitatively evalu-
ated.
1This is theoretically possible but impracticable since the separation of dense fluorophores in time that is
needed for SMLM would make the method extensively slow if additional confocal scanning was required,
multiplying the duration of an experiment by the number of confocal volumes to visit.
4.2 Results 71
4.2 Results
4.2.1 Axial variations of the point-spread function
In geometrical optics, a point-emitter in the object plane of the objective lens is conjugated
with a point on the plane of the detector located in the image plane of the tube lens (green
rays, Figure 4.1, A). Conversely, a point-emitter located above or below the object plane of
the objective lens, is conjugated with a point located respectively in front of or behind the
plane of the detector (red rays). The image recorded on the detector is consequently larger
and dimmer (for the same number of emitted photons, as they are distributed over a larger
surface on the detector). Because of the diffraction of the light, images of both the in-focus
and out-of-focus emitters are further ‘blurred’: e.g. the image of the in-focus single-emitter
is recorded as a blurred spot, or PSF, rather than a mathematical point (Figure 4.1, D). The
diffracted image of a point-emitter located at an axial position z through a microscope is
defined as the PSF(z) of the microscope.
In order to measure the 3D shape of the PSF of the instrument, sub-diffraction green-
fluorescent beads were imaged in axial steps of 10 nm (Figure 4.1, B). As expected, both
the width and the amplitude of the PSF dramatically vary in z: the PSF of an in-focus bead
is both thin (Figure 4.1, D, left panel, contrast-adjusted (xy) plane) and bright (right panel,
intensity plot), whereas the PSF of the same bead located 500 nm above or below the focal
plane is broader and dimmer (Figure 4.1, C and E). The 2D shape of the PSF of an object
therefore holds axial information. Bi-plane microscopy [86], a technique that compares the
widths of the PSFs of single emitters simultaneously imaged at two different axial positions,
is based on this same conclusion.
Most SMLM analyses fit each detected single PSF with a 2D-Gaussian function. This
means that characteristics of the PSF such as the width or the amplitude are de facto usually
extracted, although not quantitatively further used. This chapter offers to quantitatively
study the axial dependence of such parameters in order to increase contrast of the super-
resolved image by rejecting out-of-focus emitters.
An in-focus thin plane, or virtual-‘light-sheet’ (vls), is defined as an arbitrary volume
around the focal plane. The focal plane is itself defined here as the plane for which the
integrated intensity of a box drawn around the 2D-PSF of each imaged bead is maximal.
Due to the axial variations of most high-numerical aperture super-resolution microscopes,
the height of the vls will typically be of ∼600 nm (full width at half maximum − FWHM
72 The virtual-‘light-sheet’
Figure 4.1 The point-spread function of a SMLM instrument varies axially. In geometrical optics, an in-focus
point-emitter is conjugated with a point on the plane of the detector (A, green rays) whereas the light coming
from an out-of-focus emitter focusses in front of or behind the same plane, distributing the same intensity
on a larger surface (A, red rays); it is therefore imaged as a dimmer and larger spot. Imaging a z-scan of
separated 40 nm diameter 505 nm emitting beads confirms this conclusion: the 3D intensity image of 28
beads were superposed and rendered (B, the contrast of each plane being adjusted) to obtain a finer model of
the 3D-PSF. Close-ups of the section of the 3D-PSF are given for the focal plane (z = 0, D) and symmetric
out-of-focus planes (z =±500 nm, C and E). Left panels (C-E) present the contrast adjusted intensity image
(1pixel = 110 nm); right panels show the intensity surface plot of the same −though non contrast-adjusted−
data.
Legend: (Average) width σ = (Θ3 +Θ4)/2 of the PSF as fitted in Equation 3.12; Integrated intensity Θ5;
Amplitude Θ5/(2πΘ3Θ4); Signal to noise ratio defined as the amplitude over the noise of the background b;
Ellipticity defined as the ratio of max(Θ3,Θ4)/min(Θ3,Θ4); Precision σth defined as in Equation 3.13.
4.2 Results 73
−, corresponding to a Gaussian standard deviation σ = 255 nm2); in the experiments de-
scribed in this chapter, with a 1.49-na oil-immersion objective, the variation of the inte-
grated intensity of the 2D-PSF along the optical axis was fitted with a Gaussian function
(σ = 283 nm; FWHM = 666 nm). However, the width of the vls is highly dependent on
−and anti-correlated with− the numerical aperture (NA) of the objective lens used [178]:
in Figure 4.6, a calibration z-scan of beads was imaged with a 1.2-na water-immersion lens;
the thickness of the vls was consequently adjusted to 2,000 nm.
4.2.2 Sub-diffraction beads under low illumination mimic idealised sin-
gle fluorophores
To build the vls, a z-scan of single emitters is needed to calibrate the effect of width and
amplitude thresholding on typical localisations located at different z positions. A common
way to measure the PSF of a microscope in fluorescence microscopy is to use sub-diffraction
fluorescently-labelled beads as is done in deconvolution microscopy [179–182]. Conversely
to isolated fluorophores, they are photostable for many excitation/emission cycles (the same
multi-labelled beads can be sampled many times at different z positions without photo-
blinking and photo-bleaching), cheap and commercially available.
In order to validate this hypothesis, additional experiments were performed by purifying
the photo-activable fluorescent protein mEos3.1 from S. pombe, which was immobilised on
a glass coverslip and imaged in PBS. Similarly, single TMR* fluorophores were adsorbed
on a coverslip and imaged. Since it is impossible to repeatedly image a single dye (un-
der SMLM conditions) at many different axial positions, the protocol was adapted: a high
density of fluorescent protein was immobilised on the coverslip but very low irradiance
activation light was used in order to image only a few separated fluorophores per frame.
Therefore, different fluorophores were imaged at each axial position of the stage. After fit-
ting each detected 2D-PSF, the width vs. amplitude parameter plot was computed for the
localisations detected in the vls (Figure 4.2 A, purified mEos3.1 in green, TMR* in orange).
Separated sub-diffraction beads (40 nm in diameter) were also imaged under low illumina-
tion so that the beads were excited to yield a similar photon count as single fluorophores
(between 500 and 1,500 photons detected per frame) [45]. The beads were imaged at differ-
ent axial positions and, after PSF fitting, their localisations were plotted in the same width
vs. amplitude parameter plot (navy blue). The distributions of localisations of the single
fluorophores and of the bead in the parameter plot are indeed very similar, confirming the
2FWHM = 2
√
2ln2σ = 2.35σ
74 The virtual-‘light-sheet’
Figure 4.2 Images of single sub-diffraction beads under adapted low illumination have similar widths to those
of single fluorophores. A: Single fluorophores (purified mEos3.1 in green, TMR* in orange) were immobilised
on poly-lysine-coated coverslips and imaged under 561 nm illumination. Additional very low 405 nm irradi-
ance was used to activate a few separated fluorophores in each frame. 40 nm-diameter 505 nm emitting beads
(navy blue) were also immobilised on a similar coverslip and imaged under low 488 nm illumination to obtain
similar signal-to-noise levels as for the single fluorophores. For each sample, ten frames were imaged every
10 nm axially. After fitting, each localisation was plotted in a width vs. amplitude parameter plot. B: Sepa-
rated fixed Alexa-647-labelled Tau monomers were imaged in focus until they photo-bleached under different
illumination powers. After fitting, each localisation was plotted in a similar parameter plot.
use of beads as idealised single fluorophores in such low illumination conditions.
A fitting artefact is observed in Figure 4.2 at small width values: a line of localisations of
various amplitudes but of very similar width (∼30 nm) reveals the fitting of high frequency
noise or ‘hot’ pixels (one pixel corresponds to 110 nm). This observation points out the per-
missive width threshold value used in Peak Fit to allow further informative post-processing
during vlsSMLM (cf. ‘Width Factor’ in Table A.3).
The effect of illumination power was also investigated: another sample (separated and
immobilised Alexa647-labelled Tau monomers) was imaged until total photo-bleaching un-
der illumination powers varying from 5 to 80 mW (corresponding to 0.430 to 10.9 mW
measured at the objective). For illuminations higher than 5 mW, the distribution of locali-
sations in the width vs. amplitude plot was similar to the distribution previously obtained
with beads (Figure 4.2, B). The distribution contracted on the amplitude axis when the il-
lumination decreased, indicating that the illumination did not saturate the fluorophore. At
the lowest illumination power (5 mW), the single fluorophores could only be qualitatively
guessed after averaging multiple frames. Thus, this last imaging condition does not repre-
sent an illumination that can be used in an SMLM experiment.
Furthermore, a simple simulation confirms that the sub-diffraction bead can have a di-
ameter of up to 100 nm without increasing the width of its PSF by more than 2.10 nm: in an
4.2 Results 75
array of 512x512 pixels (1 pixel= 110 nm), disks of different diameters (1, 40 and 100 nm)
of constant intensity −as a crude model of a fluorescent bead− were convolved with a 2D-
Gaussian of width σGaussian = 71.36 nm corresponding to the theoretical approximation of
the in-focus PSF of a 505 nm emitting point-emitter (Equation 2.1). These simulations of
the PSF of the beads were fitted with a Gaussian function and their widths and R2 values
stored in Table 4.1. Interestingly, the simulated width of the PSF of the largest nano-object
(100 nm diameter) is only 2.10 nm wider than the theoretical width of the PSF of a point-
emitter. This difference is negligible when compared with the experimental variation of the
widths of the fitted PSFs (e.g. the in-focus ∼50 nm-thick band in Figure 4.2, A-B).
Table 4.1 Effect of the diameter of sub-diffracted emitters on the width of their image.
Disk diameter (nm) σfitted (nm) R2
0 (theoretical point-emitter) 71.36 N/A
1 71.40 1.0000
40 71.49 1.0000
100 72.46 1.0000
However, as described by Equation 2.1, the width of the PSF of an emitter depends on
the wavelength of its emitted photons. Table 4.2 gives the expected width of the PSF of a
point-emitter for three typical wavelengths. Although a 10 nm chromatic variation of the
width of the PSF between 505 nm emitting beads and 580 nm emitting mEos proteins is
predicted, Figure 4.2 reveals that this variation is experimentally not a limiting factor. The
pixilation of the camera, chromatic imperfections of the lenses, background noise and the
fitting algorithm are all factors that can introduce additional variation of the width of the
PSF above the theoretical value given by Equation 2.1. However, whenever possible, the
spectrum of the beads should be matched to the spectrum of the single emitters that they are
mimicking.
Table 4.2 Effect of the wavelength of sub-diffraction emitters on the width of their image.
Wavelength λ (nm) σtheoretical (nm)
510 72.07
580 81.96
670 94.68
76 The virtual-‘light-sheet’
4.2.3 Building of the virtual-‘light-sheet’
In order to calibrate and build the vls, separated idealised single fluorophores, i.e. sub-
diffraction fluorescent beads, were immobilised on a poly-lysine coated coverslip and suc-
cessively imaged at different axial positions. The axial and lateral variations of different
fitted parameters that characterise the shape of the PSF were plotted in Figure 4.3. The
three parameters which vary the most in the axial direction are the amplitude, the width and
the theoretical precision of the localisations. The variations of these parameters at a given
plane are much smaller than their axial variations. This validates the possibility of rejecting
out-of-focus localisations via thresholding some specific parameters.
Figure 4.3 Axial variations of the PSF. Separated idealised single fluorophores were immobilised on a coverslip
and imaged at different axial positions. After SMLM fitting, the axial variations of different fitted parameters
(width, amplitude) or commonly-used metrics (integrated intensity, signal to noise ratio, ellipticity, precision)
were compared. The variation of each parameter at a given axial plane z is represented as a shaded region
(mean ± standard deviation of the mean).
In principle, multi-parameter vlsSMLM thresholding is possible. However, the precision
is not a parameter which is directly fitted, and is calculated using both amplitude and width
parameters (Equation 3.13). The calculated precision is therefore mathematically correlated
4.2 Results 77
to both width and amplitude, and cannot be independently thresholded. Thus, the width and
amplitude are the best suited parameters for thresholding in vlsSMLM.
Figure 4.4 Building the vls. A calibration z-scan of separated and immobile idealised single fluorophores is
imaged. After SMLM fitting, all localisations are plotted in a width vs. amplitude parameter plots (A), in
red if they originate from a plane outside of the vls (left and right panels); in green otherwise (middle and
right panels). vlsSMLM consists of successively applying an upper width and lower amplitude thresholds
(B-C, respectively left and middle panels). The value of each threshold is chosen as a compromise between
the confidence (plain grey line) and the recall (dashed) ratios. The effect of thresholding can be seen in the
original parameter plot (B-C, right panels: localisations from greyed areas are rejected). B and C are two
different sets of vlsSMLM thresholds for two different applications: structural imaging and confidence modes
as defined and used in Section 4.2.5.
78 The virtual-‘light-sheet’
To visualise the axial variations of the width and amplitude, localisations were plotted
in a width vs. amplitude parameter plot, in green when imaged in the defined vls (Sec-
tion 4.2.1), otherwise in red (Figure 4.4, A). The vlsSMLM strategy can here be seen as
selecting the region of the parameter plot that contains most green localisations and the
fewest red ones. To quantitatively measure the effect of thresholding, confidence and recall
ratios were defined as follow:

confidence =
TP
TP + FP
recall =
TP
TP + FN
(4.1)
where TP represents the true positives (green localisations kept after thresholding), FP,
the false positives (red localisations kept after thresholding), and FN, the false negatives
(green localisations that are thresholded out). The confidence ratio is the probability that
a localisation that has been kept after thresholding actually comes from the vls. The recall
ratio measures the fraction of emitters from the vls that are kept after thresholding.
First, the effect of applying a width threshold to the localisations was studied. A thresh-
old eliminating all localisations whose widths were larger than a value varying from 500
to 100 nm was applied and the corresponding confidence and recall ratios were calculated
(Figure 4.4, B-C, left panels). As the threshold increases, more and more localisations are
rejected and the recall ratio drops (solid line). However, at thresholds ≥ 206 nm, most
of the rejected localisations come from outside the vls; thus the confidence ratio increases
until a maximum is reached (dashed line). In order to choose a width threshold, a compro-
mise needs to be made between high confidence (to increase the contrast of the final rebuilt
picture) and a high enough recall to correctly sample the structure of interest.
Once the first width threshold is defined, the confidence and recall ratios are calculated
for different amplitude thresholds (Figure 4.4, B-C, middle panels). Again, the amplitude
threshold is adjusted to obtain a high confidence ratio together with a high enough recall
rate (depending on the sampling required for the considered applications). Since the beads
were imaged under conditions at which their intensities were similar to the intensities of
single fluorophores in SMLM experiments (Section 4.2.2), similar thresholds can be used
for the analysis of SMLM experiments.
Finally, the result of the applied thresholds can be schematically observed in the param-
eter plot (Figure 4.4, B-C, right panels). The aim of vlsSMLM is to select a large fraction of
green localisations with widths of∼175 nm (for the 505 nm emitting beads), while rejecting
4.2 Results 79
most of the red localisations.
4.2.4 Robustness of the calibration
The robustness of the calibration was then tested: (1) its sampling, both in the number
of detected localisations per plane and in the number of axial planes to be imaged; (2)
its robustness over different calibration samples and instruments; (3) finally, the effect of
different fitting algorithms.
Figure 4.5 Optimisation of the vls calibration sampling. Ten frames of 32 fixed and separated idealised single
fluorophores were recorded every 10 nm axially. A: Varying in silico the number of frames taken into account
in the vlsSMLM analysis (top horizontal axis) was used as a metric to vary the sampling −the total number
of imaged physical fiducial markers− of the calibration data (bottom horizontal axis). B: The effect of the
variation of the axial sampling was studied by varying in silico the size of the z-steps used in the vlsSMLM
analysis. The values of both confidence and recall ratios for the threshold values corresponding to the structural
imaging mode (width threshold: 223 nm; amplitude threshold: 1,502 photons/µm2) were monitored (A-B).
Since both confidence and recall ratios are statistical metrics, the effect of the number
of detected localisations on their value was analysed. A calibration z-scan was imaged,
recording ten frames of the same region for each axial positions. As 32 beads were imaged
ten times in the field of view, an average of 31.5 beads were detected at the focal plane.
Selecting only one to ten frames per axial positions in silico (from the ten recorded), the
sampling of detected localisations was consequently varied and its effect on both ratios
observed (Figure 4.5, A). As few as 32 beads (or one frame per axial step) were sufficient
to characterise both confidence and recall ratios.
A similar analysis was used to assess the importance of axial sampling of the calibration.
A calibration z-scan was imaged, stepping every 10 nm. Calibration z-scans stepping every
20, 40, 80, 160, 320 and 640 nm were computationally extracted from the z-scan and the
80 The virtual-‘light-sheet’
confidence and recall ratios calculated for each of them (Figure 4.5, B). As expected, and
since the height of the vls is ∼600 nm (σ = 283 nm, Section 4.2.1), the size of the step
of the z-scan has a strong impact on the calibration efficiency: the confidence rate (dashed
line) dropped significantly as soon as the sampling involved z-steps of more than 40 nm.
The recall rate (plain line) was flatter, but also dropped for z-steps of 80 nm or higher. It is
therefore recommended to sample∼35 beads at least every 40 nm along the optical axis; this
can be achieved in multiple ways including piezo, mechanical or even manual positioning.
Figure 4.6 Robustness of the vlsSMLM calibration to different samples and instruments. Independent z-
scans of different isolated fluorescent particles were imaged following various protocols on different SMLM
nanoscopes to test the robustness of the vlsSMLM calibration. Analyses of five experiments were compared:
their in-vls (green) and out-of-vls (red) parameter plots are drawn in A. The vlsSMLM thresholds correspond-
ing to the structural imaging mode (223 nm, 1,502 photons/µm2) are respectively plotted as black horizontal
and vertical lines. The confidence (B) and recall (C) curves for width thresholding are also plotted and com-
pared.
To test the robustness of the vls calibration to different SMLM instruments and different
protocols, Dr U. Endesfelder and T.J. Etheridge immobilised diffraction-limited fluorescent
fiducial markers (i.e. fluorescent photo-stable particles of diameter smaller than the width
4.2 Results 81
of the PSF of the microscope at the same wavelength) on a coverslip and record a z-scan on
their own SMLM instruments. Calibration using beads emitting at different wavelengths,
or quantum dots, with different frame and z-step samplings, on different TIRF instruments,
were analysed (Figure 4.6, A-C). Since all microscopes tested had a 1.4-na oil-immersion
objective lens, a calibration z-scan was also performed with a radically different 1.2-na
water-immersion lens, leading to the definition of a ‘thicker’ vls (σ = 2,000 nm instead of
283 nm − see Section 4.2.1). The parameter plots (Figure 4.6, A) look relatively similar,
although they vary in density; the localisations located in the vls are remarkably denser
below the 223 nm width threshold (black horizontal line). The variations of the confidence
(Figure 4.6, B) and recall (Figure 4.6, C) ratios for width thresholding qualitatively follow
a similar shape. They however peak (Figure 4.6, B) and drop (Figure 4.6, C) at slightly
different threshold values depending on both the wavelength of the fluorescent particles and
on the specific instrument (Section 4.2.2). A calibration on the same microscope used for
the SMLM experiment of interest, with fiducial markers whose spectrum is matched with
the spectrum of the fluorophores used in the SMLM experiment is therefore recommended.
Interestingly, the confidence ratio for 560 nm emitting beads is globally lower. This is due to
some diffusing beads in solution which are out of the vls but are considered in it during the
calibration. Finally, the noticeable shift of the ratio curves for quantum dots may be linked
to the very small sampling in terms of the number of fiducial markers (only one frame was
imaged every 20 nm and the field of view only allowed the imaging of 8 quantum dots).
As vlsSMLM is a post-processing method, the effect of different fitting algorithms has
to be assessed. Five commonly-used algorithms (Peak Fit [149], QuickPALM [183], rapid-
STORM [156], ThunderSTORM [184] and M2LE [185]) were used to fit the same calibra-
tion data. The resulting list of localisations was then analysed with the vlsSMLM plugin
described here. Their difference is reflected in the parameter plots shown in panel A of
Figure 4.7. However, localisations of fluorophores within the vls (green) are mostly found
just below the 223 nm width threshold (black horizontal line). The confidence (Figure 4.7,
B) and recall (Figure 4.7, C) curves for the width threshold follow a similar shape. They
however peak (Figure 4.7, B) and drop (Figure 4.7, C) at different values, underlining the
importance of processing the calibration z-scan in the same way as the SMLM data, us-
ing the same fitting routine with the same parameters. Outlier behaviour is noticeable for
QuickPALM. This could be explained by the imprecise fitting of localisation widths by
QuickPALM.
Each fitting algorithms uses slightly various 2D-Gaussian PSF models defined by dif-
ferent numbers of parameters. To increase the speed of the analysis, rapidSTORM only fits
82 The virtual-‘light-sheet’
Figure 4.7 Effect of various fitting algorithms on vlsSMLM calibration. Five 2D-Gaussian fitting plugins were
used to analyse a single calibration z-scan. The list of localisations obtained were fed to the vlsSMLM plugin
and the outputs compared: in-vls (green) and out-of-vls (red) parameter plots are drawn (A). The vlsSMLM
thresholds of the structural imaging mode (223 nm, 1,502 photons/µm2) are shown as black horizontal and
vertical lines. The confidence (B) and recall (C) curves for width thresholding are also plotted and compared.
4.2 Results 83
as few parameters as possible, inferring a fortiori many characteristics of the PSFs from
these few fitted values thanks to a constrained model of the PSF [156]. This is visible in
Figure 4.7, where localisations fitted with rapidSTORM, but not with Peak Fit or Thun-
derSTORM, are observed on a thin 1D-curve in the width vs. amplitude parameter plots.
Although rapidSTORM performs remarkably well in the calibration of the vlsSMLM, it may
be less robust to variations of the PSF shape in vivo due to deformations of the wavefront in
deep samples or sub-frame diffusion of the single fluorophores.
Moreover, Figure 4.7 also underlines the pertinence of vlsSMLM: the limit of vlsSMLM
would be observed if no localisation appeared in the out-of-vls parameter plot (A, right
panels). Conversely, vlsSMLM could be used as a way to determine optimal thresholds used
as fitting checks in fitting algorithms. And even so, vlsSMLM would still give the possibility
to slightly adjust the thickness of the vls for different applications and quantitatively select
thresholds for specific needs (Section 4.2.5).
4.2.5 Improvement in contrast and quantification using vlsSMLM
Two different sets of thresholds −or vlsSMLM imaging modes− for two different poten-
tial applications in vlsSMLM are described in this section. The first −‘structural imaging
mode’− (Figure 4.4, B) allows controlled enhancement of contrast, while retaining as many
of the fitted localisations as possible, in order to maintain the sampling resolution in struc-
tures of interest. The width threshold for this mode was defined as the intersection of the
confidence and recall curves (223 nm). An amplitude threshold that gave a confidence ratio
of 90% for the structural imaging mode was then selected (1,502 photons/µm2). The second
imaging mode −‘confidence mode’− (Figure 4.4, C), favours the confidence ratio over the
recall, to ensure that the retained localisations are inside the vls volume. The width thresh-
old was thus defined as the width for which the confidence ratio is maximal (206 nm). The
amplitude threshold was then fixed to obtain a confidence ratio of 95% (2,164 photons/µm2).
To illustrate the two imaging modes, two different biological structures were imaged.
vlsPALM analysis was subsequently performed both to increase contrast and to look at
the distribution of 3D clusters. These examples were chosen to exemplify the types of
problem addressable by the different imaging modes. In both examples (the yeast vacuoles,
or the nucleus of a mammalian stem cell), the cellular structures of interest were located
more than 100 nm above the coverslip, i.e. outside of the field that can be measured using
TIRF illumination. To image such structures, in the absence of light-sheet capability, HILO
illumination [105] has to be used. Such broad illumination excites a large proportion of out-
84 The virtual-‘light-sheet’
of-focus fluorophores which contributes to an increased fluorescent background. vlsPALM
thresholds were applied to reject these out-of-focus fluorophores.
Structural imaging mode
Figure 4.8 Structural imaging mode. Fixed S. pombe cells expressing cytoplasmic Cdc22-mEos proteins were
imaged using PALM. 5,000 frames were analysed with Peak Fit and the resulting list of localisations was used
to produce a super-resolved picture directly after fitting (A), or after applying the vlsPALM thresholds defined
in Figure 4.4, B (B). The corresponding diffraction-limited image of the two cells is shown as an inset in A.
Close-ups of the white rectangles in A-B are shown in C-D. The contrast of the large intracellular vesicles of
the yeast is increased after vlsPALM filtering (white arrows in C-D).
Fixed fission yeast cells expressing the cytosolic protein Cdc22 fused to mEos2 at its
C-terminus were imaged to demonstrate the increase in contrast that the structural imaging
mode allows and reveal a cellular organelle. The fission yeast Schizosaccharomyces pombe
(S. pombe) is a powerful and highly tractable eukaryotic model organism, often used to study
the cellular responses to DNA damage and the process of DNA replication. In response to
nitrogen starvation or to osmotic stress, large vacuoles appear in S. pombe cells in order to
4.2 Results 85
restore the concentration of the cytosol [186]. Imaging Cdc22, a protein which is highly
expressed and evenly distributed in the cytoplasm of the yeast, allowed visualisation of, in
contrast, this organelle.
Applying the vlsPALM thresholds in structural imaging mode, an increase in contrast
is observed in the super-resolved image (Figure 4.8, B and D): some vacuoles only appear
after thresholding because out-of-focus localisations from below or above the vacuoles are
detected and plotted in the non-thresholded picture (Figure 4.8, A and C).
Confidence mode
Fixed mouse embryonic stem cells stably expressing mEos3.2-tagged centromere protein A
(Cenp-A) that form distinctive clusters were imaged to show how quantification of super-
resolved clusters benefits from the confidence mode of the vlsPALM analysis. Cenp-A is a
histone H3-like protein that is present in nucleosomes at the centromeres in eukaryotic cells.
Cenp-A forms foci at a number of defined points in a nucleus and determining the structure
of such foci or their stoichiometry is of interest in the yeast genetics field [130].
However, out-of-focus fluorophores have larger and dimmer PSFs (Figure 4.1), making
them more difficult to detect and fit over the background. Thus, only clusters which are
Figure 4.9 Confidence mode. Embryonic stem cells expressing Cenp-A-mEos proteins were fixed and imaged.
The corresponding movie (summed in A) was analysed with Peak Fit and the resulting list of localisations
was separated between in-vls (green) and out-of-vls (red) localisations using the vlsPALM thresholds defined
in Figure 4.4, C. vlsPALM allows the identification of the in-focus localisations (B). All localisations were
plotted either as fitted (C) or in a super-resolved picture (D), but coloured according to the vlsPALM filtering.
Three categories of Cenp-A clusters were observed: some were almost entirely within the vls (D-F, 1); others
were spanning one extremity of the vls, partly in the vls (D-F, 2); the last ones were entirely out of the vls
(D-F, 3). E: Diffraction-limited and super-resolved close-ups of the Cenp-A clusters defined in D. F: Number
of localisations inside (green) and outside (red) of the vls for each cluster. Such classification allows selecting
in-focus clusters for further quantification and preventing under-counting due to undetectable out-of-focus
emitters.
86 The virtual-‘light-sheet’
within the focal plane have all their fluorophores correctly detected, while those outside
the focal plane have some of their fluorophores undetected. Using vlsPALM (Figure 4.9,
A-D) selects only in-focus clusters (Figure 4.9, E) in order to analyse them for further
quantification (Figure 4.9, F), preventing initial under-counting.
4.3 Discussion
This chapter presents a simple, cost-effective, powerful method for selecting in-focus fluo-
rophores to increase the contrast in super-resolved images that can be used in conjunction
with any 2D super-resolution microscope and any super-resolution PSF-fitting algorithm.
This method however is incompatible with basic centroid-finder algorithms that do not give
precise information about width and amplitude of the localisations. By virtue of the fact that
each PSF is imaged, one by one, a thin, in-focus slice from a super-resolved image can be
tuned and selected, and in doing so a virtual-‘light-sheet’ can be created.
In the experiments described here, using fluorescent proteins (mEos2 and mEos3.2)
which emit at∼580 nm, and thresholds of 220 nm for width (standard deviation of the Gaus-
sian fit) and 1,500 photons/µm2 for amplitude (Figure 4.4), the contrast of super-resolved
images was significantly increased, allowing the observation of cellular organelles (Fig-
ure 4.8) and the detection of in-focus protein clusters (Figure 4.9).
4.3.1 A quantitative optical sectioning method for existing 2D SMLM
nanoscopes
In SMLM, most analysis softwares fit the detected PSFs with a 2D-Gaussian (rapidSTORM,
Peak Fit, QuickPALM, M2LE, ThunderSTORM, etc.). For most applications, from all the
fitted parameters (up to seven parameters) only two−the (x,y) position of the localisation−
are used. The width and amplitude of localisations are also often used to estimate the lo-
calisation precision. Although the other fitted parameters (i.e. the 3D shape of the PSF)
contain specific information about the axial position of the fluorophore, they are usually
disregarded or only heuristically used as fitting checks in algorithms. The post-analysis
method described in this chapter consists of quantitatively applying a threshold on the fitted
width and amplitude of all localisations. The width threshold can be thought of as a virtual
wide-field pinhole, rejecting any fluorophore located outside the focal plane. Globally, such
thresholding amounts to limiting imaging to a sheet of light −the vls− around the focal
plane of the microscope.
4.3 Discussion 87
Similarly to most 3D SMLM techniques (e.g. astigmatism [87], DH-PSF [88]) that opti-
cally modify the 3D-PSF of the microscope to extract additional information about the axial
position of the emitters, one can extract additional −though less precise− axial information
from the original 3D-PSF of the instrument. vlsSMLM does not compete with 3D-SMLM
techniques, as they have much better axial resolution and are non symmetric across the fo-
cal plane, but allows more information to be extracted from existing data. Such analysis
enhances the contrast of super-resolved pictures and provides additional quantified optical
sectioning around the focal plane.
Interestingly, Figure 4.7 shows why and how vlsSMLM is pertinent: existing fitting soft-
wares would be optimal and vlsSMLM would be useless if no localisation appeared in the
out-of-vls parameter plot (A, right panels). Conversely, vlsSMLM could be used as a way
to determine optimal width or amplitude thresholds used as fitting checks in fitting algo-
rithms. Even so, vlsSMLM enables fine adjustement of the thickness of the vls for different
applications and quantitative selection of thresholds for specific needs (Section 4.2.5).
As in light-sheet illumination, vlsSMLM imaging allows an increase in contrast in the
final super-resolved image, although any illumination geometry can be used (TIRF, HILO or
epifluorescence). Moreover, the width of the focal volume, or vls, can be tuned by changing
the thresholds applied. However, although a main advantage of light-sheet illumination is
the minimisation of photo-bleaching of the fluorophores and photo-damage of the sample as
only the imaged plane is illuminated, vlsSMLM imaging does not reduce photo-bleaching
or photo-damage as it is a post-process analysis. It is therefore not suitable for experiments
requiring multiple successive axial slicing to observe multiple planes in a sample.
4.3.2 vlsSMLM only requires a simple calibration step
This method only requires a simple z-scan calibration of fluorescent beads, normally achie-
ved via a controlled z piezo-driven stage. The calibration is used to build up a ‘parameter-
plot’ (Figure 4.4) that is specific for each individual imaging system or microscope, and
then to define thresholds with known confidence and recall rates that can then be applied to
the SMLM data. The calibration does not require extensive sampling: as few as 31 fiducial
markers and steps of up to 40 nm do not quantitatively affect the results (Figure 4.5).
Also virtually any fixed and separated sub-diffraction fluorescent point source (e.g.
beads, gold particles or quantum dots) can be used for the calibration. Interestingly, Mutch
et al. [187] found that “a log-normal distribution [of intensities] such as is observed for the
single fluorophores [Alexa Fluor 488] is also observed even when larger numbers of fluo-
88 The virtual-‘light-sheet’
rophores are present in a [region of interest]. Indeed, [they] found the measured intensity
distributions of 100 nm fluorescent beads (which contain hundreds of fluorophores per bead)
to also follow a log-normal distribution.” In other words, the authors find that the intensity
distribution of single dyes has the same shape as the intensity distribution of beads. There-
fore, if the illumination is such that the average intensities collected for both single dyes
and beads are equal, the intensity distributions will be identical. This is further confirmed
by comparing the parameter plots of diffraction-limited fluorescent beads imaged under low
illumination and immobilised single fluorophores (Figure 4.2).
The calibration proved to be robust to very different imaging protocols on the various
different instruments tested so far (Figure 4.6). However, the actual values of the vlsSMLM
thresholds depend on both the wavelength of the fluorophores and on the specific optics
used. Thus, a calibration should be measured on the instrument that is used for the SMLM
experiment with beads whose spectrum is matched to the wavelength of the fluorophore
used in the SMLM experiment.
4.3.3 On thresholding
To even better threshold in-focus from out-of-focus localisations in the parameter plot and
maximise the confidence ratio, complex polygons can be drawn in the parameter plot, as
is commonly used in flow cytometry analysis. Other parameters could also be added to an
n-dimensional parameter plot, increasing the precision, robustness and sophistication of the
vlsSMLM analysis. However, in the interest of simplicity and universality of the imple-
mented ImageJ plugin, a successive single-parameter thresholding approach is described in
this chapter.
Localisations were thresholded first according to their width, then according to their am-
plitude. Indeed, the amplitude of a PSF does not depend only on whether the fluorophore
is in the vls, but also on many different parameters such as its (x,y) position in the inho-
mogeneous illumination field, which fluorescent state it occupies, or the orientation of its
dipole [98]. The width of a localisation, on the contrary, depends only on its axial position,
and its diffusion during the exposure time of the frame [163, 188]. Therefore, the width
parameter is less convolved with phenomena other than the axial position of the emitter
and consequently allows better discrimination between emitters that are inside and outside
the vls. Furthermore, the distribution of intensities from single fluorophores follows a log-
normal distribution [187]. The long tail of such a distribution means that the intensity of a
fluorophore, whether or not it originates from the focal plane, varies a lot. Thus, in a given
4.3 Discussion 89
z plane, the amplitude of a population of fluorophores will follow the same distribution (as
the amplitude of a Gaussian is directly proportional to its intensity, at constant width). A
bright fluorophore out of the vls has a similar amplitude as a dimmer fluorophore in the vls.
Thus it is not optimal to use intensity (or amplitude) as the only or primary threshold.
4.3.4 vlsSMLM, a user-friendly ImageJ plugin
vlsSMLM was implemented as a freely available ImageJ [150] plug-in, which acts as a
post-processing layer that can use any 2D-Gaussian fitting results (Figure 4.7) and is thus
independent of fitting algorithms. The actual values of the vlsSMLM thresholds are never-
theless dependent on the fitting algorithm used for fitting both calibration and SMLM data.
The calibration step should thus be analysed with the specific fitting routine used to fit the
SMLM data.
The vlsSMLM plug-in is attached as a supplementary software to this thesis (Soft-
ware B.4). It offers two functions: the first uses a calibration z-scan to calculate (and display)
parameter plots with the corresponding confidence and recall rates specific to the instrument
used, as shown in Figure 4.4, B-C. The second function takes the list of fits from a SMLM
experiment and interactively displays a super-resolved image with user-defined thresholds.
It also outputs the list of thresholded fits for further analysis or reconstruction (see the doc-
umentation attached to the Software B.4 for more details). The vlsSMLM plug-in has been
designed to be an easy-to-use tool and is available both as a stand-alone ImageJ plug-in
or directly integrated into the Peak Fit plug-in [149], a quick efficient SMLM fitting algo-
rithm, making the vlsSMLM plug-in able to analyse any SMLM stack without any other
pre-processing PSF-fitting step.

PART B

Chapter 5
Introduction to the molecular basis of T
cell activation
This chapter provides the biological context for the work presented in the next two chap-
ters (Chapters 6 and 7) by outlining the organisation of the vertebrate immune system, the
key role of T cells in adaptive immunity, and the molecular processes underpinning T cell
activation.
5.1 A brief introduction to the vertebrate immune system
Since the development of life, a multitude of living species have evolved and share common
ecological niches which are characterised by competition for resources and inter-dependent
interactions between species. At the ends of the inter-species interaction spectrum, strategies
of symbiosis1 and immune defence are found.
The term immune system encompasses of all active and passive processes that a species
has developed to protect itself from invasion of parasites and infection with pathogens.
These mechanisms include the passive skin barrier in mammals or selected apoptosis in
plants. Since the immune system is inherently destructive, the main challenge it faces con-
sists in distinguishing self from non-self. Although immune systems take various creative
approaches in the living kingdom, a complex and powerful immune system, composed of
two highly intertwined and complementary branches, evolved in jawed vertebrates to protect
themselves against and quickly adapt their defence to ever-mutating pathogens.
1For example, the endo-symbiotic theory postulates that organelles such as chloroplast or mitochondria
originate from the uptake and integration of former free-living bacteria inside another cell.
94 Introduction to the molecular basis of T cell activation
The vertebrate immune system is formally divided in two components or legs, charac-
terised by distinct specificity, delay and strength of their response to an infection and the
memory of the infection:
Innate immunity. The first leg of the vertebrate immune system is the innate immune sys-
tem, responsible for the detection and routine elimination of pathogens. If a pathogen
overcomes the passive barriers such as the skin or mucosa, its intrusion is first de-
tected in a non-specific way by the innate immune system: some cells, such as macro-
phages or dendritic cells, express pattern recognition receptors that recognise regular
patterns of molecular structures common to a whole class of micro-organisms but ab-
sent from the body’s own cells, the so-called pathogen-associated molecular patterns.
Examples of such receptors are the mannose and glucan receptors and the scavenger
receptor that bind cell-wall carbohydrates of bacteria, yeast and fungi, respectively.
The family of Toll-like receptors provides recognition of cell-wall components of
both Gram-positive and Gram-negative bacteria. When these receptors are triggered,
macrophages release cytokines and chemokines that trigger a local inflammatory re-
sponse, signalling a zone of infection to the whole immune system.
The inflammatory response induces a great number of processes (e.g. vasodilata-
tion of the nearby blood vessels, migration, differentiation and activation of effector
immuno-competent cells, blood clotting and wound healing) aiming at both contain-
ing the invasion of the pathogen, preventing its spreading to the whole body through
the blood and lymph circulation systems, and clearing the infection. This initial innate
response is characterised by its un-specificity to the pathogen, its immediacy, and its
binary activation irrespectively of the history of previous infections.
Adaptive immunity. The inflammatory response also triggers another type of processes
forming the adaptive immune system. Contrarily to the innate immune response,
the adaptive response is characterised by its specificity to the pathogen, its delayed
effects (∼ 4-day post infection), its strength compared to the initial response of the
innate immune system and its memory upon re-infection of a pathogen.
Because pathogens reproduce and evolve at a higher rate than the human species, in-
vasion strategies overcoming the innate immune response have appeared over time.
The innate immune system can thus be overwhelmed and evaded. This is where adap-
tive immunity plays its essential role of strong and targeted clearing of the pathogen.
After detection and phagocytosis of the pathogen, professional antigen-presenting
cells (APC) of the innate immune system present short peptides from the lysis of
5.2 The T cell, a key player of the adaptive immune system 95
the pathogen in the groove of the major histocompatibility complex (MHC) class II,
a membrane protein which expression at their plasma membrane is increased during
an inflammatory response. This then triggers the clonal expansion and differentiation
of a specific subset of cells from a theoretically infinite repertoire. This specific acti-
vation is explained in detail in the case of the T cell in Section 5.2. These mature T
and B lymphocytes orchestrate the delayed but strong immune answer leading to the
successful clearing of the specific pathogen.
Once the pathogen has been cleared from the organism, a long-term memory of the
infection enables a fast and effective answer of the adaptive immune system upon re-
infection by the same pathogen. The 4 day delay is reduced as the specific lympho-
cytes only require limited maturation and their greater initial population exponentially
increases the speed of their new clonal expansion.
Adaptive and innate immunities, although traditionally distinguished and opposed, are
the two complementary and inter-dependent legs of the same system; they have co-evolved
and their differences are sometimes debated (e.g. recently, the possibility of memory associ-
ated to natural killer cells, cytotoxic lymphoid-like cells of the innate immune system, was
postulated [189]). Innate immunity plays an essential role in detecting the initial infection
and presenting the pathogen for specific selection and maturation of the lymphocytes initiat-
ing the adaptive immune response. It is also responsible for routine clearance of pathogens
that do not require an energy-expensive and complex adaptive response, and for containing
the infection until an effective adaptive response is orchestrated.
In addition to being energy-expensive, a major drawback of the adaptive immune sys-
tem resides in its accidental dramatic dysfunction: by mis-distinguishing self from non-self,
a false-negative detection may miss a life-threatening infection, while a false-positive de-
tection may lead to the development of an allergy or an auto-immune disease. Stringent
safe-guards and negative feedback loops have evolved to decrease the error rate and sta-
bilise the fine balance of sensitivity of adaptive immunity.
5.2 The T cell, a key player of the adaptive immune system
The adaptive immune system is organised around two types of lymphoid cells, the B and
the T cells, named after the organ they grow in (the bone marrow and the thymus, resp.)
[190]. Both types of lymphocytes express a large repertoire of immunoglobin-like antigen-
recognition receptors: although the repertoire, based on stochastic gene rearrangement,
96 Introduction to the molecular basis of T cell activation
junctional recombination and somatic hypermutations, is theoretically infinite, each cell
only expresses a single and unique receptor, which specifically recognises a unique epitope.
The antigen-recognition receptor of the B cell is the Y-shaped B cell receptor, and, upon B
cell activation and differentiation into plasma cell, large amounts are secreted in its soluble
form referred to as monoclonal antibody (mAb).
The antigen-recognition receptor of the T cell is the T cell receptor (TCR, cf. following
section). After careful selection to avoid any recognition of self peptides, each naive T cell
expresses a single and unique type of TCR that recognises a specific combination of non-self
peptide presented in the groove of a specific MHC protein. The type of the MHC protein
recognised by a T cell determines its sub-type, characterised by the specific expression of
a co-receptor: CD4+ T cells recognise non-self peptides bound to MHC class II molecules
(which are only expressed on professional APC); CD8+ T cells, those bound to MHC class
I molecules (which are expressed on the membrane of virtually all cells in the body, consti-
tutively presenting peptides which represent the diversity of the proteins degraded by their
proteasomes).
Like B cells, CD8+ T cells are the effector cells of the adaptive immune system: while
plasma cells (activated B cells) are the mAb factories, enabling neutralisation and opsonisa-
tion of, and complement activation against, the extracellular pathogen or toxin they target,
activated CD8+ T cells (also called cytotoxic T cells) specifically recognise and kill cells (by
inducing programmed cell death, i.e. apoptosis) that have been infected with the targeted
pathogen (e.g. intracellular bacteria or virus).
CD4+ T cells, however, have little direct effector function but orchestrate the whole
adaptive immune response. Several specialised subsets, defined by the cytokines they re-
lease and induced by different cytokine signals, have been described: TH1 activate macro-
phages; TH1 and TH2, also called helper T cells, activate the maturation of B cells targeted
against the same pathogen into plasma cells; TH17 recruit neutrophils to sites of infection
early in the adaptive immune response; lastly, Treg suppresses the activation and devel-
opment of naive T cells. Finally, CD4+ T cells stimulate the activation of CD8+ T cell
targeting the same pathogen.
Because of their essential role, both as cytotoxic effector cells and as coordinator in-
tegrating and organising the communication between the various cells of both innate and
adaptive immune systems, T cell development, selection and activation have been studied
since the 1960s. Understanding the fine mechanisms of T cell specific but sensitive acti-
vation and regulation is a major interest to design immuno-therapeutic strategies against
auto-immune diseases or cancers.
5.3 The complex T cell receptor 97
5.3 The complex T cell receptor
Stoichiometry
Although the stoichiometry of the TCR complex is still an object of controversy (cf. Chap-
ter 6), it is known to be composed of several subunits: ligand recognition is achieved by the
highly variable ectodomain of the hetero-dimer TCRαβ (or, in less understood and rarer
cases, TCRγδ ) whose intracellular domain is almost inexistent; signalling is conducted
by invariant transmembrane proteins collectively called CD3 and organised in two hetero-
dimers CD3εδ and CD3γε and one homo-dimer CD3ζζ . Without CD3, TCRαβ does not
form a complete cell-surface receptor and is degraded before reaching the cell membrane.
The cytoplasmic tail of each CD3 chain displays one (CD3δ , ε and γ) or three (CD3ζ )
immuno-receptor tyrosine-based activation motif (ITAM) domains that can be phosphory-
lated and which initiates downstream signalling of the receptor.
Positive and negative selections
Each T cell expresses many copies of a unique combination of recombined genes coding
for the TCRαβ subunit. This genetic recombination mechanism enables the recognition
of theoretically any epitope by a T cell, given an infinite number of T cell clones. This is
why the adaptive immune system has the potential to adapt and initiate a specific immune
answer to any newly mutated pathogen. Nevertheless, because the adaptive immune system
should not turn against cells of its own body, a mechanism must exist that enables T cells to
identify self from non-self.
Sequentially, genes coding for TCRα and TCRβ chains, are recombined until a pre-TCR
(validating an in-frame recombination of the gene coding for TCRα), then a full TCR (idem
for TCRβ ) is expressed on the plasma membrane of the CD4−CD8− T cell precursor. A
successful TCR expression at the membrane induces the expression of both CD4 and CD8
co-receptors. At this stage, double-positive CD4+CD8+ T cells undergo two selections
based on the affinity of their TCRs to self-peptides presented on MHC molecules by APCs
present in the thymus:
1. A positive selection induces apoptosis of T cells expressing TCRs that fail to bind to
any peptide-MHC (pMHC) molecule. This selection is also responsible for the dif-
ferentiation of the T cells into the CD4+ and the CD8+ subsets (cf. previous section):
basically, a T cell expressing TCRs recognising peptides presented by MHC class I
98 Introduction to the molecular basis of T cell activation
molecule will decrease its CD4 expression and become a cytotoxic CD8+ T cell; con-
versely, if it recognises peptides presented by MHC class II, its CD8 expression will
decrease and the T cell will become a CD4+ T cell.
2. Independently, a negative selection insures that T cells whose TCRs bind to self-
peptide:MHC ligands with high affinity are killed in the thymus.
Together, these positive and negative selection mechanisms enable the generation of a
mature T cell repertoire that is both restricted to the recognition of peptides presented by
self MHC molecules and tolerant to self peptides. Interestingly, this double selection allows
for TCR degeneracy that nuances the specificity of the TCR for a unique non-self pMHC:
each TCR can interact with, and is thus specific to, a set of different ligands [191]. TCR
degeneracy enlarges the total repertoire of pMHC antigens that can be targeted by a limited
number of T cells, although it may enable cross-reactivity of the T cell.
Sensitivity and specificity
The unmatched sensitivity and specificity of the TCR remains a puzzling challenge to the
research community [192]. The TCR discriminates a very weak signal in the presence of
considerable noise: (1) each TCR recognises and distinguishes both low-affinity (>500µM)
self-pMHC ligand for naive T cell survival and homeostasis, and high-affinity (though still
only in the 1-10 µM range) non-self pMHC ligands for T cell activation. (2) Even more
strikingly, the degree of activation of a T cell depends on the affinity of its TCR to the non-
self pMHC ligands that it is activated with. Different results suggest that (3) T cells can
recognise and respond to very few (if not a single) non-self pMHC complexes [193–196]
and that a single TCR micro-cluster can activate a T cell [196]. (5) Finally, these 1-10 non-
self pMHC molecules are recognised in presence of considerable noise: each APC presents
105-106 self pMHC complexes on its plasma membrane, most of them being irrelevant to
any T cell.
Altogether, these characteristics underline the extraordinary sensitivity and specificity of
the TCR for its ligand. It explains the major attention given to the TCR since the 1960s as
both a key signal receptor in adaptive immunity and a remarkable object of study in the field
of ligand-receptor interaction. Indeed, in most systems, a trade-off exists between sensitivity
and specificity, which the TCR somehow resolves, suggesting that TCR triggering is the
fruit of the evolution of a complex mechanism that integrates signals at various levels, from
the molecular level of the receptor, to the level of the T cell, and to the level of the whole
population of responding cells [197].
5.4 T cell activation at the immunological synapse 99
5.4 T cell activation at the immunological synapse
The activation of the TCR is thus a key determining step in immune responses, deciding on
a single-cell scale whether to initiate a response that has strong repercussion on the scale of
organs or even the whole individual. This decision begins at the contact between two cells:
either between a professional APC and a naive T cell in the lymph node deciding on clonal
selection and initiation of the adaptive immune response, or between an infected cell and an
activated cytotoxic CD8+ T cell deciding on inducing selective apoptosis of the former cell.
Figure 5.1 The immunological synapse. Conjugates of primary CD8+ T cells from a transgenic OT-1 mouse
were loaded with their pMHC ligands (ovalbumin peptides) to act as APCs for each others. Conjugates
were fixed in solution and immuno-stained with markers of the pSMAC (Lck in red, A) or of the dSMAC
(actin in red, B). C: Examples of T cells forming large oriented contacts characteristics of the formation of
immunological synapses. D: Scheme of the ‘bull-eyed’ organisation of the immunological synapse at the
plasma membrane of a cytotoxic T cell (cf. Section 5.4).
Legend: Nuclei in blue; CD8 in white to label the plasma membrane. Upper panels show (xy) sections and
lower panels show (xz) (en face) reconstruction. Scale bars: 5 µm in A-B, 10 µm in C.
100 Introduction to the molecular basis of T cell activation
The importance of cell-cell contact reveals the essential role played by membrane re-
ceptors, and especially the TCR, in this process. Several consecutive steps are involved that
finely tune and modulate T cell activation, starting with the formation of an adhesive con-
tact between the APC and the T cell and ending with changes in the transcription patterns
of the naive T cell inducing its differentiation in an effector T cell, or with the secretion of
cytotoxic granules to the infected cell by the activated cytotoxic T cell.
Initial integrin-mediated cell-cell adhesion. Cell-cell contact is initiated by binding of in-
tegrins and adhesion proteins such as the intracellular adhesion molecule 1 (ICAM-1,
or CD54) and CD58 (or lymphocyte function-associated antigen 3 − LFA-3) on the
APC, to LFA-1 and CD2 (or LFA-2) on the T cell. Interestingly, LFA-1 expression
is increased during T cell activation, leading to stronger and more frequent cell-cell
contacts with APCs. Similarly, maturation of the professional APC promotes T cell
adhesion by constraining ICAM-1 mobility [198]. It is suggested that initial adhesion
contact primarily involves long integrins (i.e. ICAM-1:LFA-1), allowing stochastic
engagement of the smaller CD58:CD2 interaction in a tighter contact [199]. These
closer contacts, of the size of the TCR:pMHC interaction, would enable the nucle-
ation of a zone where the TCRs could probe the membrane of the APC and which
could further develop into the formation of a stable cluster [200].
Extracellular ligand recognition. Once a tight contact is formed between the APC and the
T cell, various receptors can start interacting with their ligands. The integration of two
signals is required for a T cell to trigger: (1) signal 1, responsible for the specificity of
the activation to the antigen, is achieved when a pMHC is simultaneously recognised
by both a TCR and a CD4 or CD8 co-receptor. Signal 1 induces activation of the
T cell and guarantees the recognition of a non-self peptide presented by a self-MHC
molecule. (2) However, to avoid entry into an anergic state, the T cell needs a second
signal, signal 2, that promotes cell survival. Signal 2 is mostly achieved by interaction
of the CD28 co-receptor with its CD80 or CD86 ligands, expressed on the membrane
of professional APCs during an inflammatory response (cf. Chapter 7). This second
signal guarantees that T cells only activate during an infection, and prevents the killing
of professional APCs by activated cytotoxic T cells.
TCR triggering: transduction of the signal across the plasma membrane. Both signal 1
and 2 are detected by transmembrane receptors on the extracellular side of the plasma
membrane. How the signal is transduced inside the T cell is still unclear [190, 201,
202]. Several models for this very step, referred to as TCR triggering, have been
5.4 T cell activation at the immunological synapse 101
proposed and are discussed in the following section. TCR triggering leads the phos-
phorylation of the ten CD3 ITAMs of the TCR by Lck and Fyn, two protein tyrosine
kinases of the Src family.
Formation of the immunological synapse. Topologically, after TCR triggering, microclus-
ters of TCRs form [203] and concentrate towards the contact region between the
two cells, in a centripotent motion conducted by filamentous actin (F-actin) dragging
[204, 205]. A dramatic reorganisation and polarisation of the T cell then follows,
resulting in the formation of a dynamic but stable bull-eyed structure in the “small
space between the two interacting cells”: the so-called immunological synapse [206].
Monks et al. described three co-axial structures (or supramolecular activation com-
plexes − SMAC) forming the immunological synapse in 1998 [207]: in the cen-
tre resides the central SMAC (cSMAC), where the TCR micro-clusters merge. It is
also enriched in short adhesion molecules such as CD2, co-receptors CD4, CD8 and
CD28 and signalling kinases such as Lck, protein kinase C-θ (PKC-θ ) and ZAP-70
(ζ -chain-associated protein 70).
The cSMAC is surrounded by a ring of integrins, larger cell-adhesion molecules (e.g.
LFA-1 and ICAM-1), proteins with a large ectodomain such as the non-specific phos-
phatase CD45, and the cytoskeletal protein talin. This region is called the peripheral
SMAC (pSMAC). Finally, as actin clears from the cSMAC, a larger ring of actin
forms the distal SMAC (dSMAC), which is thought to bring the necessary force for
the final docking of the centrosome (or microtubule organising centre) to the plasmic
membrane. Indeed, a mature immunological synapse is also characterised by the mi-
gration of the centrosome of the cell towards the cSMAC in a process reminiscent of
cilia and flagella formation [206, 208]. In the case of cytotoxic T cells, the secretory
lysosomes are brought towards the minus end of microtubules located at the centro-
some, and secrete their contents at the immunological synapse next to the cSMAC.
At the cSMAC, cortical actin is actually not entirely cleared but rather remodelled so
that lytic granules might migrate through actin pores [209, 210]. This polarisation of
the T cell is essential for the focal secretion of toxic granzymes and perforin and thus
for the specific killing of the recognised target cell.
The immunological synapse may provide several functions such as stabilising the
contact, orientating granule secretion in the case of cytotoxic T cells, ending the T-
cell activation and recycling of receptors.
Activation of intracellular signalling cascades. The first signalling steps induced by TCR
102 Introduction to the molecular basis of T cell activation
triggering are the phosphorylation of the ten ITAMs of the CD3 chains of the TCR by
Lck and Fyn kinases. Lck exists under various activated forms and its pre-activation
and the mechanism by which it phosphorylates the TCR have recently been chal-
lenged [211, 212]. However, phosphorylation of the ITAMs trigger the recruitment
of ZAP-70 kinase to the plasma membrane, which is a common marker for TCR trig-
gering. In turn, ZAP-70 phosphorylates the scaffold proteins LAT (linker of activated
T cells), which indirectly leads to the clivage of PIP2 (phosphatidylinositol biphos-
phate).
Then, three signalling cascades are triggered [190, 192]: (1) Ca2+ ions from the en-
doplasmic reticulum are released, increasing the intracellular concentration of Ca2+
which finally activates the transcription factor NFAT (nuclear factor of activated T
cells). (2) The activation of PKC-θ induces the activation of the transcription factor
NFκB. (3) The activation of the small G-protein Ras triggers the MAP (mitogen-
associated protein) cascade, which is involved in the regulation of the transcription
factor AP-1. Integration of these three transcription factors stimulate the expression
of interleukin 2 (IL-2) which promotes survival and division of the lymphocyte and
regulate the global immune response.
Other signals are integrated and modulate the signalling cascades. Activation of the
co-receptor CD28 promotes cell survival, stimulates the metabolism of the cell but
also directly enhances TCR signalling by recruiting Lck to the TCR or by activating
the MAP cascade. On the top of signal 1 and 2, a third signal, signal 3, is provided by
cytokines secreted by other immuno-competent cells. Signal 3 modulates the activity
of the T cell and orientates the differentiation of naive CD4+ T cell in specific TH or
Treg subsets enabling an efficient and specific targeting of the immune response to the
pathogen (e.g. humoral or cell-mediated immunity).
5.5 Three molecular models for TCR triggering
In signalling pathways, membrane receptors play an important role as signal initiators, trans-
ducing the signal (i.e. the information about ligand binding) across the plasma membrane.
In the specific case of TCR triggering, three types of mechanism have been proposed to
explain signal transduction, namely, the aggregation, conformational change and kinetic-
segregation models. Evidence exists for all three mechanisms during TCR triggering (re-
viewed in ref. [202]) and each mechanism might play a role in TCR triggering. It is unclear,
5.5 Three molecular models for TCR triggering 103
however, whether any on its own is sufficient to induce triggering [190, 201, 202]. These
models attempt to fill the gap between TCR binding to its pMHC ligand, and T cell activa-
tion (as revealed by the formation of TCR micro-clusters or the triggering of the downstream
signalling cascades). All three models are introduced and discussed in the following para-
graphs.
Aggregation
Like the B cell receptor, the TCR is a tyrosine kinase-associated receptor. Most receptors
of this type are activated by passive cross-linking: upon binding of a multivalent ligand (or
clustered ligands), two or more receptors cluster. This aggregation locally increases the con-
centration of their cytoplasmic tails in the cytoplasm. Close-by tails can cross-phosphorylate
their tyrosines, either directly, if they have an intrinsic kinase activity (e.g. the stem cell fac-
tor receptor Kit [213]), or indirectly, by recruiting a tyrosine kinase (e.g. the B cell receptor
[190, 214]), and become activated.
Artificial aggregation of TCRs using either soluble antibodies or soluble multimeric
forms of pMHC is also sufficient to initiate TCR triggering [215, 216]. However, pMHC
are usually very rare on the target cell surface and TCR triggering can be observed in the
presence of very few pMHC ligands. It is therefore thought that extra CD4 or CD8 co-
receptors can bind either to (1) the same non-self pMHC (co-receptor hetero-dimerisation
model [217]) or to (2) another self-pMHC ligand (pseudo-dimer model [218]). Either way,
aggregation of TCRs would be induced via binding to CD4 or CD8. Nevertheless, TCR
triggering can occur in cells completely devoid of co-receptors indicating that this mecha-
nism is not sufficient to explain TCR triggering [219]. Furthermore, Jiang et al. recently
provided evidence that the CD8 co-receptor is only recruited to the TCR after initial TCR
triggering [220, 221].
It is clear, however, that TCR triggering leads to aggregation of TCRs into microclusters
of 10 to 100 TCRs [222, 223]. But whether it is a direct consequence of TCR binding to
pMHC or a further process downstream of TCR triggering is still unknown [203, 224].
Conformational change
The conformational change model is a powerful model to describe some ligand-receptor
interactions. It can be compared to a key-lock system: when a receptor (the lock) binds to
its ligand (the key), the interaction stabilises a different conformation of the receptor (the
key ‘opens’ the lock) that presents a catalytic locus, hidden in the close conformation of the
104 Introduction to the molecular basis of T cell activation
Figure 5.2 Four molecular models for TCR triggering have been proposed. A: The aggregation model postu-
lates that TCR complexes aggregate upon the engagement of a TCR with its ligand (through the co-receptor-
dependent formation of hetero- or pseudo-dimers). B: The conformation change model postulates a deep
conformational change of the whole TCR complex upon binding, releasing CD3 ITAMs that were buried in
the plasma membrane. It is proposed that binding under force may dramatically enhance and stabilise a ac-
tivated conformation of the complex. C-D: The kinetic-segregation model postulates that at equilibrium, the
TCR is constantly phosphorylated and dephosphorylated by freely diffusing Lck and CD45 proteins in the
plasma membrane. When a contact forms with an APC, phosphatases with large ecto-domains such as CD45,
but not the small Lck, are segregated from the contact-zone. If a TCR is held in the contact-zone long enough
(upon binding with its ligand), its net phosphorylation increases and the TCR is triggered.
Legend: T cell in green; APC in red; MCH complex in yellow, bound to either an antigen peptide (green
circle) or a self peptide (dark red triangle); TCR in blue (TCRαβ in light blue, CD3 in dark blue); ITAMs as
dark blue open circles, orange full circles when phosphorylated; CD4/CD8 co-receptor in orange; Lck in red
in the membrane of the T cell (phosphorylation of the ITAMs is shown by a yellow arrow and a ‘+’ symbol);
CD45 in dark green (dephosphorylation of the ITAMs is shown by a green arrow and a ‘−’ symbol).
5.5 Three molecular models for TCR triggering 105
unbound receptor. This is the principle behind the molecular mechanism of protein motors
(e.g. the kinesins or the ATP synthase) and the triggering mechanism of the family of the G
protein-coupled receptors [225].
Concerning TCR triggering, little evidence for a conformational change has been found
so far: although the crystallographic structure of the intact TCR-CD3 complex has not yet
been solved, the structure of the TCR in its bound and unbound state suggests a possible
subtle conformational change in the membrane-proximal loop of the TCRα constant domain
[226]. Adding to this argument, mutation of residues in this loop abrogates TCR triggering,
which is suggested to hinder TCR dimerisation [227]. However, the classical static con-
formational change model seems to poorly describe the mechanism behind TCR triggering:
the failure of several crystallographic studies in revealing global large-scale conformational
changes between a bound and unbound structure speaks against it [228].
With this unsuccessful model as an argument, Ma et al. introduced a dynamic conforma-
tional model, the receptor deformation model, where binding under forces, instead of a gain
in stability, provides the energy for a deep conformational change of the receptor [229].
Since CD3γε hetero-dimer is relatively rigid [230], a conformational change in the TCR
ectodomain could be transduced to the CD3γε ectodomain and induce mechanical piston-
like forces. Those forces could free the CD3ε cytoplasmic ITAMs which might otherwise
be buried in the lipid bilayer of the plasma membrane. Thus, new ITAMs would be exposed
to phosphorylation and the TCR would become activated [231].
Such a dynamic model is all the more attractive as it takes into account the mechanical
forces at play between two contacting cells in vivo where adhesion molecules and dynamic
cytoskeleton proteins induce fast filapodia and lamellipodia movements [200, 204, 205].
Furthermore, massive cytoskeleton re-arrangement is observed during the immunological
synapse formation that could provide the force required [206, 209]. Finally, this model
would give a convincing explanation for the observed fine specificity of the TCR for its non-
self ligand as binding under force (or catch bonds) can dramatically increase the stability of
an interaction [232, 233]. Variations of this dynamic model involving pulling, shearing or
pushing forces are reviewed in length in ref. [202]. Nevertheless, although attractive, this
model is entirely putative and is supported by no experimental data so far [229].
Kinetic-segregation
A recent model, first described seventeen years ago and based on the passive spatial re-
organisation of the receptor, was introduced to challenge the aggregation model of TCR
106 Introduction to the molecular basis of T cell activation
triggering: “Provocatively, we have argued that receptor cross-linking is not sufficient by
itself to activate the TCR, but rather that it is phosphatase exclusion mediated by cell–cell
contact that is critical.” [199]
Because of its two main requirements, this model was called the kinetic-segregation
(KS) model [199, 234]. It postulates that, when two cells contact each other, adhesion mole-
cules form a close-contact zone from which membrane proteins with large extracellular do-
mains are physically excluded. This leads to the formation of a local micro-environment de-
pleted of such proteins (e.g. unspecific phosphatases). Nevertheless, this micro-environment
should be small enough to prevent the activation of an unbound freely diffusing receptor. But
when bound to an external ligand on the contacting cell, the receptor will stay, in average,
longer in this micro-environment, leading to its activation (e.g. through phosphorylation).
Hence, it is the kinetics of these contact zones and the chemistry which occurs within them
that transfers the signal across the membrane through the receptors.
Different lines of evidence reviewed in [234] suggest that segregation of the TCR is
likely to have an important role in TCR triggering: (1) molecules with large ectodomains,
such as the inhibitory tyrosine phosphatases CD45 and CD148, are excluded from areas
of TCR triggering [235]. (2) Modifications of the intercellular distance at the putative
close-contact zone strikingly inhibits TCR triggering: the truncation of CD45 or CD148
ectodomains [235, 236]; the elongation of the pMHC complex [237, 238]; the use of soluble
vs. surface-associated pMHC ligand [239]; or the use of engineered TCRs with ectodomain
differing in nature and size [240, 241] obstruct T cell activation.
Moreover, recent data indicate that in resting T cells, Lck is mainly present in an ac-
tive form [211]. This is consistent (although debated in [212]) with the concept of a con-
stant equilibrium between unspecific inhibitory phosphatases and kinases such as Lck. The
close-contact zones would then locally shift the equilibrium towards phosphorylation and
eventually activation if a TCR stayed long enough in this micro-environment deprived of
the large CD45 and CD148 phosphatases.
Furthermore, Contento et al. reveal a new TCR-independent LFA-1-dependent pathway
that polarises T cells and shape them for later antigen recognition [242]. They propose
that the first adhesive contacts between T cells and APCs prepare and arrange the TCR
signalling micro-environment. This would complete the KS model in which the first step
involves adhesion contacts, that could also signal on their own, to prepare for further steps
such as TCR triggering or the later immunological synapse formation.
Finally, the KS model has recently regained interest and some of its key prediction have
been successfully tested: ligand-specific T cell-like activation was reconstituted in a non-
5.5 Three molecular models for TCR triggering 107
immune cell and requirement for segregation of CD45 and Lck was demonstrated in this
artificial system [243]; ligand-independent TCR triggering was observed in T cell forming
artificially large contacts with various adhesive surfaces from which CD45 was segregated
(Chang, Fernandes, Ganzinger, Lee et al., submitted).

Chapter 6
Counting T cell receptors
“In conclusion, three decades after the finding that αβ and CD3 subunits form a
single complex, we have to admit that the stoichiometry of the TCR still remains
to be resolved.” [224]
The original idea developed in this chapter was a collaboration between Prof. S.J. Davis,
Prof. D. Klenerman, Prof. A. Carr and Dr S.F. Lee. All T cell lines were transfected, and
antibodies were labelled and provided by E. Huang and Dr R.A. Fernandes. All yeast cell
lines were designed, transfected, grown and provided by H. Armes and Dr A.T. Watson. All
imaging experiments were done by M. Palayret, with the help of H. Armes when yeasts were
imaged. All data analysis, and code writing was undertaken by M. Palayret. Dr A. Herbert
brought important insights and critical discussion about carefully analysing the data. Dr S.F.
Lee and M. Palayret designed and built the super-resolution instrument used in this chapter.
Aim of the experiments described in this chapter
The organisation of the TCR on the membrane of the resting T cell still needs to be de-
termined. In this chapter, it was therefore tested whether a robust and quantitative single-
molecule counting tool could be developed to this aim if carefully designed controls were
used to calibrate the measurements. As SMLM techniques appeared in 2006, single-mole-
cule counting was an important topic of discussion and development in the SMLM field
when this project started in 2010. Several research groups published counting methods
during the course of the project and these are discussed at the relevant points in the chapter.
110 Counting T cell receptors
6.1 Introduction
6.1.1 Organisation of the TCR on a resting T cell
The TCR is composed of several subunits: a heterodimer composed of TCRα and TCRβ
chains (or TCRγ and TCRδ , depending on the specific gene recombination that led to the
mature T cell), responsible for the recognition of a specific peptide antigen presented in
the groove of the MHC complex; and the dimers CD3γε , CD3δε and CD3ζζ , contain-
ing the intracellular ITAM motifs which, upon phosphorylation, trigger the downstream
signalling cascade which results in T cell activation. However, the stoichiometry of the
peptide-recognition TCRαβ subunit in the complex is still unresolved and subject of con-
troversy [224, 244]. The stoichiometry of the TCRαβ subunit will be referred to as the
stoichiometry of the TCR in this chapter.
More than a theoretical academic dispute, the stoichiometry of the TCR on the resting
T cell has important implications on the models of TCR triggering: the understanding of
the structure of the TCR is likely to impose important constraints on theories of antigen
recognition and TCR triggering [224, 244]. Although so far very little is known about T cell
triggering after ligand binding, micro-clusters of TCRs are observed on the plasma mem-
brane upon early T cell activation [203]. TCR micro-clusters are even thought to be the
actual in vivo cSMAC where T cell activation occurs1. Understanding the precise struc-
ture of the TCR right before and after ligand binding will provide more insights into how
these micro-clusters form and how the antigen-induced signal crosses the plasma membrane
to induce downstream signalling. The theoretical implications of the stoichiometry of the
TCR before T cell activation and the lack of robust evidence explain why it is an object of
controversy in the field.
Adding to the confusion, recent and conflicting evidence were reported:
1. TCRs organised in ‘protein islands’ (i.e. nano-domains of the plasma membrane en-
riched in specific membrane proteins) were observed on the bottom membrane of live
T cells in contact with a non-activating supported-lipid bilayer by high-speed PALM
(hsPALM) [248]. This clustered behaviour was confirmed by further transmission
electron microscopy (TEM). These data were also consistent with previous Blue-
Native-polyacrylamide gel electrophoresis and electron microscopy observations of
1Stable and late immunological synapses have been proposed to be the place for secondary signals and
effector functions of the T cell after activation [245, 246]. The initial TCR micro-clusters would then be the
place of initial TCR triggering. The function of the immunological synapse is, however, subject of discussion
[247].
6.1 Introduction 111
multivalent linear TCR complexes, referred to as TCR nano-clusters [249].
2. An opposite conclusion was drawn by the group of D. Klenerman: using two single-
molecule fluorescence microscopy techniques (two-colour coincidence detection −
TCCD − and dynamic single-molecule co-localisation − DySCo) and monomeric
and dimeric controls, TCRs were observed as monomeric species [250–252].
Technical bias or biological artefacts were proposed to reconcile these results [244].
First, technically, TCCD is biased towards diffusing species and only observes a small
confocal volume at the membrane of the cell. Thus, less mobile clusters, which are less
uniformly distributed on the membrane, might have been missed. DySCo, by comparing
the co-localisation of trajectories of single molecules over time, is similarly biased towards
diffusing species. Interestingly, Dunne et al. use cells expressing ten-fold fewer TCRs than
in the parental cell line. Although the cells still activate normally, lower densities at the cell
surface could favour monomeric species over multivalent ones.
Conversely, hsPALM in live cells, because of sub-frame diffusion blurring [163], is
biased towards immobile species, as Lillemeier et al. acknowledge. Moreover, the photo-
blinking behaviour of the fluorescent protein PSCFP2 is not discussed [71], although most
fluorescent proteins were shown to reversibly enter dark states [62, 144]. Photo-blinking
of the fluorophore induces clustering artefacts in SMLM analyses and should consequently
be considered and discussed [82, 144]. Moreover, Lillemeier et al. are the first and only
ones to use PSCFP2 in a single-molecule experiment, so its single-molecule photo-physical
behaviour has not been described and should not be simply assumed. Also, the regions of
the cell membrane enriched with TCRs may reflect the observation of the regions of tighter
contact of the live cells with the supported lipid bilayer, which are the only regions that
are imaged under TIRF illumination [253–255]. Finally, TEM and gel electrophoresis are
prone to aggregation artefacts because both require fixing and labelling of the sample. Thus,
it is technically not entirely surprising that TCCD and DySCo experiments tend to observe
mobile monomeric TCRs when hsPALM and TEM experiments are in favour of protein
islands or nano-clusters of TCRs.
A biological artefact was also proposed to reconcile the contradictory observations: the
clustered structures observed by Lillemeier et al. could be an intermediate in micro-cluster
formation due to partial activation of the T cells when forming a contact, even on non-
activating surfaces [244, 256]. Indeed, a major difference in the TCCD and the hsPALM
experiments resides in the region of the membrane imaged: James et al. imaged a small
region of the apical membrane, on the top of the cell, while Lillemeier et al. imaged the
112 Counting T cell receptors
region forming a contact on different surfaces, at the bottom of the cell. In light of the recent
observation of ligand-independent TCR triggering (Chang, Fernandes, Ganzinger, Lee et
al., submitted, and supplementary information of ref. [257]), T cells forming a contact on
poly-lysine coated glass for 45’ at 37◦C are expected to activate, even in the absence of TCR
ligand. This suggests that the protein islands may reflect some early and partial activation
of the T cells but not the actual distribution of the TCR on a resting T cell.
Last, the controversy about the stoichiometry of the TCR can be solved by the simple
idea that both monomeric and protein island structures are not exclusive: TCRs could ex-
ist as monomers inside protein islands [224, 248, 249]. The important conclusion of the
dual-colour fluorescence cross-correlation spectroscopy experiments performed by Lille-
meier et al. is often forgotten in reviews summarising the controversy: a high fraction of
mobile TCR complexes was detected, which “indicate[d] that protein islands [we]re indeed
membrane domains and not protein complexes” [248]. As underlined and developed by
Schamel et al., protein islands do not explain the mechanism behind TCR triggering, but
rather how TCR avidity to its ligand is increased by proximity effects (through concentra-
tion of the TCRs in protein islands, or segregation zones [234]) and enables high sensitivity
and wide range of response of the receptor [249]. Reciprocally, Dunne et al. observed some
cytoskeleton-dependent confinement of TCR diffusion, that would be consistent with TCR
concentration in membrane domains such as protein islands. In that respect, simulations us-
ing a statistical-mechanical model of diffusing monomeric TCRs successfully reproduced,
for various hypotheses, such TCR-enriched membrane domains [258].
6.1.2 Counting single-molecule localisation microscopy
When counting single molecules, errors originating from different phenomena can be di-
vided into two general categories: over-counting and under-counting [52, 132]. Over-
counting results from imperfect photo-physical behaviour of the fluorophores: many dyes
and photo-activable fluorescent proteins used in SMLM switch reversibly between a fluo-
rescent and a dark state. Such a reversible behaviour implies that a single fluorophore might
be re-activated during the time of an experiment and imaged, and thus counted, more than
once. Although this is a sampling advantage in localisation microscopy, it is a drawback
when counting is at stake as controlling and correcting for the average number of switching
events in an experiment is a real challenge. This is why (theoretically)2 irreversibly photo-
2Although the bulk photo-physical properties of fluorophores are well-described in the literature, the com-
plex and stochastic photo-physical states of single fluorophores have rarely been studied. With the advent
of single-molecule techniques, the study of fluorophore at the single-molecule level is critical to under-
6.1 Introduction 113
activable fluorophores, and especially the bright and monomeric mEos protein, are used in
most, if not all, single-molecule counting methods described so far.
Photo-blinking is the other phenomenon inducing over-counting. On the single-mole-
cule level, most photo-activable fluorophores do not stay in a fluorescent state once activated
until photo-bleaching: mEos2, for example, has been shown to stochastically enter two dark
states, one of which could be depleted by 405 nm illumination [144]. A counter-example,
however, is PA-mCherry1, which is irreversibly photo-activable and does not blink [130].
However, compared to mEos, the lower brightness of PA-mCherry1 makes it less suitable
for single-molecule detection and thus precise counting [46, 72, 130]. Over-counting due to
photo-blinking is corrected by methods either grouping localisations together on the basis
of a characteristic dark time (Figure 6.1, A-B; Table 6.2) or statistical analysis using the
pair-correlation function g(r) (Figure 6.1, C-D; refs. [82, 83]).
Similarly, various phenomena result in under-counting errors. Since fluorophores are
stochastically activated, two fluorophores separated by less than the diffraction limit may be
activated simultaneously, and hence they are not resolved as two localisations (but one, or
none if the shape of the resulting PSF is too deformed to be fitted correctly). To decrease the
probability of these events, low 405 nm activation is preferred, limiting the number of PSFs
observed in each frame. Conversely, the finite time of the experiment limits the probability
of activating all available probes at the end of imaging. Increasing the 405 nm activation
power at the end of the experiment to increase the chance of activating most of the remaining
non-activated fluorophores partly solves this issue (i.e. Fermi activation as proposed by Lee
et al. [145]; Section 3.3.4). If the proteins of interest are imaged under constant activation
at high density (e.g. a protein of the envelope of a virus), the cumulative distribution of the
time of detection of the localisations can be used to estimate the number of fluorophores
that were missed at the end of the experiment [146].
A last and essential factor inducing under-counting is the photo-activation efficiency
[126, 260, 261]: even in the case of the perfect labelling efficiency of a fusion protein,
not all fluorescent proteins fold and mature correctly. In addition, in the specific case of
SMLM, the probability that a correctly folded fluorophore can be photo-converted to its
activated form is lower than 1 [259]. The photo-activation efficiency accounts for all these
effects and measures the probability of a fluorescent protein to be photo-activated into a
fluorescently detectable form [126]. If the stoichiometry n of the protein of interest is uni-
form, such under-counting can be corrected by fitting a binomial distribution B(n, pPE) to
stand, tackle and optimise behaviours such as photo-bleaching [31], photo-blinking [144] or photo-activation
[122, 259].
114 Counting T cell receptors
Figure 6.1 Two types of single-molecule counting methods can be distinguished: (1) A-B: Localisations can
be grouped in spatial clusters (A) defined by a threshold radius dT hresh. Clusters are then temporally divided in
groups called ‘molecules’ thanks to a dark time threshold tT hresh: successive localisations of a cluster belong to
the same ‘molecule’ if they are separated by a dark time shorter than tT hresh (B). (2) C-D: The pair-correlation
function g(r) (also referred to as radial distribution function) of the localisations can be calculated. g(r)
measures the density of localisations between distances of r and r+dr from each localisation (C), divided by
the density of a uniform distribution of the same total number of localisations, averaged over all localisations.
The pair-correlation reveals radial random distribution is close to 1 (D, red line), and radial clustering at
distance r if g(r)≫ 1 (orange line between 0 and 60 nm).
6.2 Results 115
the distribution of the number of fluorophores detected per cluster, pPE being estimated by
a monomeric control [126, 131].
Most of the methods mentioned were developed during the time of the work presented
in this chapter. They were implemented and assessed on monomeric, dimeric and trimeric
mEos3.1 cytoplasmic constructs in Schizosaccharomyces pombe, and monomeric and di-
meric mEos2 membrane constructs in T cells. The advantages and drawbacks of each
method are discussed in the following results section.
6.2 Results
6.2.1 Yeast counting controls
To develop and test a single-molecule counting method, control constructs of known stoi-
chiometry were designed: a single, two and three repeats of the photo-activable fluorescent
protein mEos3 were expressed at low concentration in S. pombe cells by H. Armes. Puch-
ner et al. independently developed a very similar strategy, published in 2013, expressing
the same constructs in fusion with a membrane protein to target the fluorophores at the
membrane of the S. pombe cells [131].
The original counting idea consists in grouping, or clustering, localisations in both space
and time in so-called ‘molecules’3: this is based on the observation that activated fluo-
rophores stochastically enters dark (i.e. non-fluorescent) states before getting irreversibly
photo-bleached (cf. Section 3.3 and ref. [144]), which is commonly referred to as blinking4.
Thus, a single fluorophore may appear in non-consecutive frames during SMLM imag-
ing and induce over-counting. However, although different blinking behaviours and dark
times up to minutes were observed, the distribution of the time spent in a dark state is
well described by an exponential decay (of constant τOFF ): a single blinking fluorophore is
manifested by multiple localisations closely correlated in space (within the localisation pre-
cision) and in time (within a few τOFF ). The challenge in single-molecule counting consists
in choosing optimal grouping thresholds (dT hresh, tT hresh) and validating the specificity, pre-
3In this thesis, different terms are distinguished for clarity: (1) ‘fluorophore’ refers to the physical molecule
fluorescing in the sample; (2) ‘PSF’ refers to its image detected on the plane of the camera and recorded for
further analysis; (3) ‘localisation’ refer to the fit of a PSF with a 2D-Gaussian in a specific frame. It is defined
by the frame number, and all the fitted parameters ((xy) position, intensity, width, etc.). (4) A ‘cluster’ refers
to the spatial grouping of all localisations of the experiment within a specific distance dT hresh. (5) A ‘molecule’
finally refers to a sub-division of a cluster in time: a molecule groups all localisations of a cluster separated by
dark times smaller than a specific duration tT hresh.
4A blink is defined as the period a fluorophore spends in its fluorescent state following a first fluorescent
emission followed by its entry in a dark state.
116 Counting T cell receptors
cision of the method and its robustness to single-molecule noise and to the stochastic nature
of fluorophore photo-physics.
Mono-, di- and trimeric controls
S. pombe cells were stably transfected with constructs of one to three repeats of mEos3. A
weak promoter and a sequence destabilising its RNA messenger were used to limit the ex-
pression of the constructs to very low levels in the cytoplasm of the cells [262]. Therefore,
the observation and characterisation of spatially separated monomers, dimers and trimers of
mEos was possible (Table 6.1). Cells were fixed with formaldehyde and immobilised under
an agarose gel pad on the coverslip before being imaged under 561 nm TIRF illumination.
Each cell was imaged for 4,000 frames (50 ms exposure) under 405 nm Fermi TIRF illu-
mination (cf. Section 3.3.4) to optimise the number of spatially-separated fluorescent PSFs
observed in each frame.
Table 6.1 Number of cells and localisations.
Monomer Dimer Trimer Negative control
Number of cells 340 114 339 16
Number of thresholded localisations 1,389 666 7,280 11
Average localisation precision
(value at peak) 11 nm 11 nm 11 nm 8 nm
After PSF fitting, localisations were first spatially grouped in clusters using a centroid-
linkage algorithm: (1) all localisations were initialised in clusters of size one. (2) The closest
pair of clusters within the defined distance threshold dT hresh were grouped and the position
of the cluster was updated as the intensity-weighted average of all the localisations the new
cluster was composed of. (3) This step was then repeated until the two closest clusters were
separated by a distance larger than dT hresh.
Then, after spatial grouping, each cluster was divided in molecules, or groups of locali-
sations separated by dark times larger than tT hresh. In the following sections, different ways
of determining the optimal couple of parameters (dT hresh, tT hresh) for correct counting of
the three mEos repeats are described, as published by several groups during the time of this
project (Table 6.2).
6.2 Results 117
Table 6.2 Grouping parameters (dT hresh, tT hresh) used in counting experiments. †Long-lived dark time in paren-
theses when two dark times are described.
Reference Yearof publication Dark time τ (s) t T hresh (s) dT hresh (nm)
Annibale et al. [263] 2011 0.1 N/A 133
Coltharp et al. [264] 2012 0.1 0.4 60
Gunzenha¨user et al. [146] 2012 0.3 50
Lando et al. [130] 2012 N/A 40
Lee et al. [145] 2012 0.06 (2.3†) N/A σth?
Puchner et al. [131] 2013 0.58 2.66 150
Durisic et al. [126] 2014 0.66 4.1 σth?
Figure 6.2 Rejection of localisations upon different localisation thresholdings. Monomeric mEos3.1 proteins
were expressed at low concentration in S. pombe cells. After fixation in formaldehyde, cells were imaged under
561 nm TIRF illumination and low continuous Fermi 405 nm activation. SMLM movies were analysed using
Peak Fit (cf. Section 3.4). A-B: width vs. amplitude parameter plots of all analysed localisations. The effect
of two post-processing thresholdings of the localisations to be rejected are shown (A, loose thresholding; B,
stringent). C: After either thresholding, localisations were grouped in spatial clusters of varying radius dT hresh
and the average number of discrete appearance of molecules per cluster αblinks was calculated. Black vertical
line: dT hresh = 50 nm. D: Distribution of the theoretical precision (cf. Equation 3.13) of the localisations after
either thresholding. Green: Loose thresholding (peak at 11 nm). Orange: Stringent thresholding (peak at
10 nm).
118 Counting T cell receptors
Determining the optimal thresholds by estimating the true number of fluorophores
In 2012, Coltharp et al. proposed a way to select optimal dT hresh and tT hresh with no other
knowledge of the sample than the average number of blinks αblinks5 of the fluorophore [264].
In the work presented in this chapter, αblinks was estimated in situ as the average number
of blinks (number of molecules for tT hresh = 1 frame) per cluster in the monomeric sam-
ple. Figure 6.2, C represents this value αblinks calculated for different values of dT hresh and
two different localisation thresholding (to select real localisations over noise, cf. Table 6.6
and the following section). A plateau is observed at αblinks = 1.2, for dT hresh ∼ 40 nm. The
plateau indicates a regime for which the average number of blinks does not depend on spatial
grouping in clusters. The regime (dT hresh ∼ 40 nm) is furthermore consistent with the loca-
lisation precision of the experiment (σexp = 2 ·σth = 22 nm, cf. Table 6.1, Figure 6.2, D and
Section 3.4.3). As the value of αblinks was unchanged for either localisation thresholding,
the more stringent thresholds were kept, in order to limit the number of false localisations,
for further analysis.
As described in ref. [264], the total number of fluorophores is thus estimated as Nre f =
Nunprocessed/αblinks, where Nunprocessed is the number of localisations. For each sample, the
matrix |(N −Nre f )/Nre f | was calculated for different values of dT hresh and tT hresh, where
N is the number of molecules obtained for specific values of (dT hresh, tT hresh) (Figure 6.3).
From such plots, the intersections of the two optimal valleys (dark regions, indicating of
good estimates of N to Nre f ) were identified as the optimal values for dT hresh = 40 nm
and tT hresh = 150 ms. In order to validate these thresholds, both values were then used
to group the localisations of each sample. The pair-correlation function g(r), representing
the degree of clustering of a distribution (a random distribution would give g(r) = 1), was
compared before (tT hresh = 50 ms = 1 frame) and after grouping in time (tT hresh = 150 ms):
in the case of the monomer, after grouping, the molecules were expected to be randomly
distributed. The very little variation in the pair-correlation function after grouping in time
revealed that the threshold values chosen did not group all localisations originating from the
single blinking fluorophore together.
This difficulty underlines the drawback of this method: its robustness highly depends
on the prior knowledge or the good estimation of the true number of fluorophores, which
is precisely what needs to be assessed in single-molecule counting (the total number of
fluorophores gives the average number of blinking events, thus the average stoichiometry of
each cluster). Here, it is approximated to the total number of localisations divided by the
5Actually, αblinks refers to the average number of ‘On’ times of the fluorophore, so αblinks = nblinks+1.
6.2 Results 119
Figure 6.3 Determination of optimal thresholds. Following the method of Coltharp et al. for single-molecule
counting, a single, two, or three mEos3.1 repeats were expressed at low concentration in yeast cells and imaged
after fixation under 561 nm TIRF illumination and low continuous Fermi 405 nm activation. After SMLM
analysis, localisation were spatially (if closer from the mean centre position than dT hresh) and timely (if sepa-
rated from another localisation by less than tT hresh) grouped in ‘molecules’. A-C: matrix of the total number
of molecules N calculated for pairs of thresholds (dT hresh, tT hresh) relative to an estimation of the expected true
number of fluorophore Nred : |(N−Nre f )/Nre f |. D-F: Pair-correlation function g(r) of the localisations before
grouping (white bars) and after grouping in molecules (pair of thresholds (40 nm, 150 ms), striped grey bars).
A,D: Monomeric mEos3.1 construct. B,E: Dimeric mEos3.1 construct. C,F: Trimeric mEos3.1 construct.
120 Counting T cell receptors
average number of blinking events. Since fluorophores sometimes appear in consecutive
frames before entering a dark state or photo-bleaching, the number of localisations is an
over-estimation of the total number of ‘On’ times of all fluorophores.
To get a better estimation of the total number of fluorophores Nre f , localisations of the
same fluorophores appearing in consecutive frames need to be grouped. This requires the
definition of a spatial threshold dT hresh: localisations closer from each other than dT hresh and
appearing in consecutive frames originate from the same fluorophore. Unfortunately, dT hresh
is exactly what the method tries to estimate. To describe it differently, to find the optimal
dT hresh, a good estimate of Nre f is needed, which itself requires a prior estimation of dT hresh.
One could study the dependence of this prior dT hresh threshold on the estimation of the
optimal threshold, or start with an estimate of dT hresh and apply the method iteratively until
it converges to an optimal dT hresh. It both implies making an initial empirical assumption.
Therefore, the next section describes a semi-empirical method.
Semi-empirical estimation of the blinking rate
In 2011, Annibale et al. proposed to fit a semi-empirical function to the first values of the
curve N(tT hresh), the total number of molecules obtained after grouping spatial clusters in
time with tT hresh [263]. This fit allows to estimate the total number of fluorophores in the
sample Ntrue, the average number of blinking events nBlinks, and the dark time constant τOFF
of the fluorophore (the exposure of the camera was set to exposure = 50 ms):
N(tT hresh) = Ntrue ·
(
1+nBlinks · e
(
exposure−tT hresh
τOFF
))
(6.1)
This method requires to first group localisations spatially. Although the authors em-
pirically choose a value of dT hresh = 133 nm, a rational estimate is defined as a multiple
of the localisation precision of the experiment [145]. Indeed, if the localisation error is
Gaussian, 99.7% of successive localisations of an immobile fluorophore will be localised
within 3 ·σexp of the average position of the localisations. Thus, considering that for the
microscopes used in this thesis, σexp ∼ 2 ·σth (cf. Section 3.4.3), dT hresh was defined as
3 ·σexp ∼ 60 nm (cf. Figure 6.2, D).
After grouping localisations in spatial clusters, clusters were divided in molecules for
different values of tT hresh. The total number of molecules N(tT hresh) was calculated and
Equation 6.1 was fitted for each sample. The fitted parameters for the three samples are
summarised in Table 6.3. Noticeably, the average number of blinking events per fluorophore
nBlinks is much lower than previously measured [263]. This imaging protocol would thus
6.2 Results 121
Figure 6.4 Semi-empirical determination of the optimal dark time threshold. Localisations from the samples
described in Figure 6.3 were grouped spatially in clusters of radius dT hresh = 60 nm. Such clusters were divided
in ‘molecules’ separated by dark times higher than different tT hresh threshold values. The number of total
molecules obtained were plotted against the threshold value tT hresh (A-C, •). The semi-empirical Equation 6.1
was fitted to the first ten points (tT hresh ≤ 0.5 s) (grey curve). Fitted parameters are summarised in Table 6.3.
A: Monomeric mEos3.1 construct. B: Dimeric mEos3.1 construct. C: Trimeric mEos3.1 construct.
Table 6.3 Semi-empirical fit of Equation 6.1 to the number of molecules obtained after grouping localisations
both spatially (dT hresh = 60 nm) and for different dark time thresholds tT hresh.
Monomer Dimer Trimer
nblinks 0.118 0.156 0.222
τOFF (s) 0.109 0.128 0.209
Ntrue 1,046 479 5,181
122 Counting T cell receptors
be theoretically optimal to count single fluorophores as little correction would need to be
applied. However, this would imply that a very low tT hresh dark time threshold, if any, would
correctly group all localisations from a single fluorophore together. Such a pair of thresholds
(dT hresh, tT hresh) is very similar to the previous one (40 nm, 150 ms) which, when applied
to the monomeric sample, proved to describe poorly spatially separated single fluorophores.
Similarly, after grouping localisations with the pair of thresholds (60 nm, 50 ms) for the
monomeric sample, the pair-correlation function g(r) of the grouped molecules was not flat
but showed some short-range clustering very similarly to Figure 6.3, D.
The poor estimation of nBlinks could be explained by some random single-molecule-like
noise that would be detected as a correct PSF and localisation (i.e. of correct width and
intensity) but that would not blink. Thus, nBlinks represents the average number of blinking
events for the whole populations of localisations. Non-blinking single-molecule-like noise
would proportionally lower nBlinks, therefore inducing false clustering of localisations in
molecules.
Although this method poorly accounts for noise, it allows to gain access to in situ esti-
mation of the exponential decay τOFF of the time spent by the fluorophore in its dark state.
Indeed, the presence of non-blinking single-molecule noise does not affect this fitted value
that only describes behaviour of blinking species, thus the fluorophore. Values of τOFF ob-
tained here are consistent with values described in the literature for the short-lived dark time
of mEos (τOFF = 0.06−0.66 s, cf. Table 6.2). There are however quite different from the in
vitro experiments described in Section 3.3 with purified mEos2 and mEos3.2 proteins under
similar imaging protocol (τOFF ∼ 1 s).
Empirical estimate of the thresholds
Knowing the value of the time constant τOFF of the exponential decay of the dark times
of in situ mEos allows to estimate a sensible tT hresh [126, 131, 146]. As proposed in 2013
by Puchner et al., the dark time threshold was defined as the time tT hresh, which takes into
account 99.8% of the measured dark times, i.e. 870 ms6.
As the pair of thresholds (60 nm, 870 ms) did not affect the clustering observed in the
pair-correlation function of the grouped molecules for the monomeric sample, tT hresh was
empirically increased until the pair-correlation function flattened, correctly describing the
random distribution of the monomeric fluorophores in the sample. This was the case for
tT hresh = 3 s (Figure 6.5, A), which is remarkably consistent with values used by Durisic et
al. and Puchner et al. (Table 6.2).
6tT hresh =−τOFF ln(1−0.998)
6.2 Results 123
Figure 6.5 Empirical determination of the optimal dark time threshold. Localisations from the samples de-
scribed in Figure 6.3 were grouped spatially in clusters of radius dT hresh = 60 nm. Such clusters were divided
in ‘molecules’ separated by dark times higher than increasing tT hresh threshold values. The threshold value
tT hresh was increased until clustering effect disappeared from the pair-correlation function of the correspond-
ing grouped molecules in the case of the monomeric construct (A). A-C: Pair-correlation function g(r) before
cluster division in molecules (white bars) or after dividing clusters in molecules (for tT hresh = 3 s, striped grey
bars). A: Monomeric mEos3.1 construct. B: Dimeric mEos3.1 construct. C: Trimeric mEos3.1 construct. D:
Distribution of n-mers (i.e. numbers of molecules in each cluster) for the three mEos3.1 constructs.
Table 6.4 Fit of binomial distributionsB(n, pPE) to the number of molecules per cluster.
pPE
n Monomer Dimer Trimer
2 2.6% 5.4% 14.2%
3 1.3% 2.8% 7.6%
4 0.9% 1.9% 5.2%
5 0.7% 1.4% 4.0%
124 Counting T cell receptors
For each sample, the distribution of molecules per cluster after grouping with (60 nm,
3 s) thresholds was calculated (Figure 6.5). In order to account for under-counting due
to imperfect photo-activation efficiency pPE , binomial distributions B(n, pPE) for n-mers
were fitted to these histograms [126, 131] (Table 6.4). The small fraction of counts higher
than 1 for the monomeric sample, 2 for the dimeric control, and 3 for the trimeric control,
are caused by randomly overlapping fluorophores. Noticeably, the number of miss-folded
undetectable fluorophores (column 0 of the histogram) cannot be observed. Thus, only
binomial distributions representing populations of n-mers for n ≥ 2 can be fitted (with the
help of an additional normalising parameter).
Interestingly, the binomial fits for 2-mers for the dimer sample, and for 3-mers for the
trimer sample, output comparable values for the photo-activation efficiency parameter pPE .
This percentage is however more than ten-fold smaller than what has been previously char-
acterised for a similar mEos2 protein [126, 131]. This difference could be explained by the
slightly different fluorophore mEos3.1 used in this work, difference in the imaging protocol
(and a possibly more stringent localisation thresholding) or additional single-molecule noise
proportionally increasing the proportion of 1-mers in the histograms.
A possible idea to determine the stoichiometry of a single population of proteins is
to calculate the histogram of the numbers of molecules per cluster and try to fit several
binomial distributionsB(n, pPE) for different values of n. Then, the comparison of the value
of the fitted parameter pPE with the experimental value obtained in the case of the dimer and
trimer controls unravels the best description of the histogram. To say it differently, fitting
the histogram withB(n, pPE), where pPE is fixed to its value experimentally determined by
the dimer and trimer controls, allows to determine the best value n describing the data.
Although theoretically attractive, this counting method assumes a single stoichiometry
in the population observed, and is thus not robust to single-molecule-like noise which would
be considered as a second monomeric population. This explains why, in the light of this
method, the trimer control would be best described as a 4-mer population (pPE = 5.2%)
rather than a 3-mer (7.6%). Nevertheless, this counting method, although quantitatively
limited in presence of single-molecule noise, allows to qualitatively separate a monomeric
population from a multimeric one.
Pair-correlation study
Because single-molecule-like noise was limiting all previous counting approaches, a method
robust to such a noise was proposed. In 2011, a statistical clustering method, so-called pair-
correlation-PALM (PC-PALM), first proposed and developed by Veatch et al., and applied
6.2 Results 125
by Sengupta et al. to PALM used the pair-correlation function g(r) to detect and correct
over-counting due to photo-blinking of the fluorophores (cf. Section 6.1.2). Although g(r)
was used in the previous counting methods as a tool to assess the quality of the pair of
thresholds (dT hresh, tT hresh), PC-PALM directly corrects the pair-correlation function for
blinking of the fluorophore in order to access to the real clustering, or stoichiometry, of the
protein of interest.
Figure 6.6 Pair-correlation PALM approach. Localisations from the samples described in Figure 6.3 were
analysed with PC-PALM. A: comparison of the pair-correlation function g(r) for the three samples (monomer
in orange, dimer in red, trimer in dark red). B: gstoch(r) (Equation 6.2, green curve) was fitted to the pair-
correlation g(r) (black curve) of the monomeric mEos3.1 sample. The residual is also plotted (red curve). C-
D: gstoch(r) (green curve) was subtracted from the pair-correlation function g(r) (black curve) of the dimeric
(C) or trimeric (D) mEos3.1 samples. The resulting curve was fitted by gprotein(r) ∗ gPSF(r) (Equation 6.2).
The sum of the fit and gstoch(r) is plotted in grey; gprotein(r) is plotted in red, but takes values higher than 3.
Fitted parameters are summarised in Table 6.5.
This method proposes to separate, in the pair-correlation function g(r), the contribution
of the stochastic blinking of each fluorophore (g(r)stoch) that induces artificial clustering of
localisations, from the real clustering of the fluorophores (g(r)protein), which is itself blurred
126 Counting T cell receptors
(i.e. convolved) with the localisation precision σexp:
g(r) = g(r)stoch+g(r)protein ∗g(r)PSF
with, g(r)stoch = 14πσ2exp ρaverage · exp
(
−r2
4σ2exp
)
g(r)protein = A · exp
(
−r
ξ
)
+1
and, g(r)PSF = 14πσ2exp · exp
(
−r2
4σ2exp
) (6.2)
The method is composed of two consecutive steps: first, monomeric fluorophores are
imaged to characterise the stochastic blinking of the fluorophore. The pair-correlation func-
tion of the localisations is fitted with the theoretical function g(r)stoch, and both the average
blinking rate αblinks of the fluorophore and the average experimental localisation precision
σexp are extracted. The pair-correlation function of the monomeric mEos3 construct in yeast
cells was well fitted with the equation of g(r)stoch (R2 = 0.90, Figure 6.6). Furthermore, the
experimental localisation precision of ∼ 22 nm (Table 6.5) is consistent with the theoretical
precision of the experiment (11 nm, Table 6.1). The average number of discrete appearances
of a fluorophore αblinks is, however, much below one (0.064). This counter-intuitive value
could reflect the incertitude in the estimation of the average localisation density over the
whole field of view as αblinks is calculated in reference to it (cf. Section 6.2.2).
Table 6.5 Fitted parameters of pair-correlation PALM analysis as described in Equation 6.2.
g(r)stoch fit g(r)protein fit
Sample σ exp ρ average α blinks A ξ N clusters
Monomer 21.6 nm 249 µm−2 0.064 − − −
Dimer 21.6 nm 267 µm−2 0.074 83.3 2.9 µm 1.3 ·106
Trimer 29.9 nm 44.4 µm−2 0.399 160.5 245 nm 16,700
In the case of a population in a uniform multimeric state (i.e. only n-meric species), this
first step can be used and the resulting average number of blinks αblinks value compared to
the monomer control: a dimer should be observed with a similar localisation precision σexp
but in average the double number of blinking events. Nevertheless, as previously discussed,
the incertitude in the determination of αblinks makes this method not very quantitatively pre-
cise nor qualitatively robust: very little difference is observed between the mEos3 monomer
and dimer controls (0.064 vs. 0.074), and αblinks is more than six-fold higher for the trimer
than for the monomer construct (0.399 vs. 0.064, Table 6.5).
In PC-PALM, a second step consists in subtracting the contribution from the pair-corre-
lation function due to the stochastic blinking of single fluorophores as revealed by the first
6.2 Results 127
step, then in fitting the resulting function g(r)− g(r)stock as developed in Equation 6.2.
The fitted parameters A and ξ respectively are a measure of the protein density in the
average protein cluster observed, and an estimate of the radius of such a cluster [82].
From these parameters, the average number of proteins per clusters is approximated by
Nclusters ∼ 2 Aπ ξ 2 ρaverage (Table 6.5).
Although PC-PALM is powerful to statistically and quantitatively describe complex pro-
tein clusters on the plasma membrane of cells [82] and does not depend on the presence of
non-blinking random noise, it requires a high number of localisations to be robust: in the
experiments presented in this chapter, very low densities of fluorophores were expressed in
cells to easily spatially separate them. Therefore, even when combining the pair-correlation
from a large number of imaged cells, the total number of localisations stays much lower
than what is advised for significant PC-PALM results (nloc ∼ 1,000 ≪ 100,000, cf. Ta-
ble 6.1 and ref. [155]), explaining the poorly relevant results obtained for the experiments
presented here. Another drawback of PC-PALM analysis is that it sums up the complex
multi-dimensional localisation data ((xy) position, frame of first appearance, intensity, etc.)
as an average 1D measurement: it thus better characterises clusters of circular-symmetry
shapes and populations of uniform stoichiometry.
Single-molecule-like noise
Even more than for classic SMLM experiments, counting single-molecule requires a very
low background noise. Two types of noise have to be distinguished: (1) out-of-focus diffuse
background noise due to an average of lowly fluorescent sources in the cell. This noise can
be photo-bleached and impact mostly both the ability to detect single fluorophores and the
localisation precision achievable (cf. Equation 3.13). (2) Less described in the literature
is single-molecule-like noise (or ‘single-molecule noise’): some endogenously-expressed
fluorescent proteins (e.g. riboflavin7, and proteins of the flavin group of molecules) emits
around 550 nm, in the same domain as the red activated form of mEos [265, 266].
In order to avoid the first category of noise, general SMLM precautions were taken (e.g.
TIRF illumination, use of filtered buffers, plasma-cleaning of coverslips). Additionally,
the effect of two different localisations thresholdings were observed in the width-amplitude
parameter plot (cf. Chapter 4) and further analyses were performed with the more stringent
set of parameters (Figure 6.2, A-B and Table 6.6). Localisations that were kept are bright
and have an expected symmetric Gaussian shape of well-characterised width. Thus, false
7Commonly known as vitamin B2.
128 Counting T cell receptors
Figure 6.7 Single-molecule-like noise. Untransfected yeast cells were fixed and imaged under 561 nm TIRF il-
lumination and low continuous Fermi 405 nm activation. A: Intensity trace of a single-molecule PSF observed
in the untransfected sample (A.U.: arbitrary units; inset: close-up of the region indicated by the grey dotted
rectangle). B: Time-lapse of the region of the camera where the PSF is observed (from top to bottom, left to
right, frames 265 to 284; exposure: 50 ms). The PSF appeared and disappeared in a single frame, which is the
characteristics of a single-molecule. Scale bar: 500 nm.
PSF-fits of background events are rejected. The stringent thresholding allows for better
rejection of false localisations but also decreases the number of localisations observed.
Table 6.6 Two localisation sets of thresholds were compared to reject out-of-focus noise.
Stringent parameters Loose parameters
Maximal width (σmax) 125 nm 150 nm
Minimal width (σmin) 60 nm 50 nm
Maximal width ratio (σmax/σmin) 1.6 2.3
Minimal amplitude 2,000 photons/µm2 1,000 photons/µm2
The effect of single-molecule noise is even more critical in single-molecule counting.
Indeed, single-molecule noise is almost indistinguishable from the signal of a fluorescent
probe: such noise also originates from a single fluorescent molecule emitting at a simi-
lar wavelength [251]. Figure 6.7 gives the example of such noise: a bright, PSF-shaped,
non blinking PSF is observed. Thus, the presence of single-molecule noise implies that
the counting analyse needs to deal with two different populations: the mEos-tagged pro-
tein of interest and some randomly-distributed (in first approximation) non-blinking single-
molecule.
This heavily limits most of the counting techniques studied in this chapter. Puchner
6.2 Results 129
et al. propose to correct their molecule numbers by subtracting the mean number of n-
mers that they counted per cell in an un-transfected negative control sample [131]. For
the work presented here, a similar approach gave an average number of 0.69 localisations
attributed to noise per cell, corresponding to 0.31 1-mers, and no larger n-mers, detected
per cell (cf. Table 6.1). Such low values do not significantly change any previous analysis.
Nevertheless, the difficulty to focus on and image non-blinking single-molecule noise that
tends to bleach during the exposure of a single frame has not been discussed by Puchner et
al., although it proved to be a real limitation to image single-molecule noise (particularly in
negative control cells expressing no fluorophore, as fluorophores tend to appear for a few
consecutive frames). This implies that the corrections reported here are a low estimate of
the noise to subtract. It is however difficult to meliorate the estimation or get an idea of how
far it may stand from its true value.
Finally, fixing the cells may also increase single-molecule noise while decreasing the
diffuse background noise. It indeed fixes all cellular proteins: although the fluorescent
endogenous proteins responsible for single-molecule noise may diffuse quickly during the
exposure time of the camera and be consequently imaged as blurred in the diffuse back-
ground noise, when fixed, they would appear as non-blurred single emitters. However, non
fixing the yeast cells would also induce diffusion of the mEos3 repeats, decrease the loca-
lisation precision (due to blurring), and raise non-trivial questions about tracking blinking
fluorophores over time, especially during long dark times between two discrete activations
of two mEos3 of the same fusion construct.
6.2.2 Studying the dependence of counting analyses on localisation den-
sities using simulations
Both types of analyses described here are dependent on the density of imaged fluorophores:
(1) in order to avoid false grouping of separate fluorophores which happen to be within
dT hresh of each other, low densities of labelled proteins are required for correct grouping;
(2) conversely, PC-PALM requires high sampling to obtain statistically relevant informa-
tion about the clustering state of the protein of interest, which consequently needs to be
expressed at higher densities. To study the density thresholds below and above which each
type of analysis is preferred, single fluorophores were simulated, fitted and analysed.
Fluorescent point-emitters were simulated as 2D-Gaussian PSFs whose parameters were
randomly drawn from statistical distributions obtained from an experimental list of locali-
sations (horizontal and vertical widths σx and σy were drawn from a Gaussian distribution
130 Counting T cell receptors
of mean 129.7 nm and standard deviation 21.5 nm, with the conditions of their maximal
ratio being lower than 1.7; total integrated intensities I were drawn from a Gaussian distri-
bution of mean 698 photons and standard deviation 590 photons; Figure 6.8, A-B). For each
new frames, a list of m (drawn from a Poisson distribution of mean 5) PSFs was simulated,
whom (x, y) positions were drawn from a uniform distribution in a 121 µm2 square. Finally,
a Poissonian noise of mean 19 photons was added on each pixel of the frame. The number
of simulated frames was defined such as to get a specific density of localisation C:
Frame number =
C · (Total area)
Mean number of PSFs per frame
(6.3)
Localisation densities of 0.01, 0.1, 0.5, 1, and 5 localisations/pixel2 (i.e. 0.83, 8.3, 41,
83, and 413 localisations/µm2, resp.) were simulated, and further fitted. Because of the op-
timised but imperfect fitting algorithm (detecting some noise and rejecting some simulated
PSFs), the variation of the simulated PSFs (some of which simulated with an intensity below
the average noise), the nearest-neighbour-based measure of density of the fitted localisation,
and the estimation of the number of frames to simulate the given density C, localisation den-
sities of 4.6, 8.4, 29, 52, and 241 localisations/µm2 were observed after fitting (Figure 6.8,
C-G, resp.).
For each simulation, both pair-correlation function (Figure 6.8, H) and the distribution of
number of molecules found per cluster, for the previous thresholds (dT hresh = 60 nm, tT hresh =
3 s = 60 frames) (Figure 6.8, I) were computed. As expected, the grouping analysis is
very sensitive to the density of localisations and 2-mers are observed for densities as low
as 8.4 localisations/µm2. However, the simulation magnifies this dependence since, if an
average of five PSFs are observed in an SMLM experiment, most fluorophores appear in
consecutive frames and will be considered as single localisations. Thus, in SMLM ex-
periments, PSFs are usually more separated in time, decreasing the number of incorrectly
grouped localisations. This is why only experiments with localisation densities lower than a
25-30 localisations/µm2 threshold are here considered for counting analysis by (60 nm, 3 s)
grouping. This localisation density is only valid for monomeric samples, and the threshold
should be doubled for the dimeric control.
Conversely, the pair-correlation function is poorly defined for low densities (Figure 6.8,
H, red and orange) and closely approaches its true value (g(r) = 1 for a uniform distribution)
for higher densities (H, blue and purple). Thus, PC-PALM (or more qualitatively comparing
pair-correlation functions) will be here only considered for localisation densities higher than
a 30 localisations/µm2 threshold.
6.2 Results 131
Figure 6.8 Simulations provide insights on the dependence of counting analyses to localisation densities.
Simulated PSFs were drawn from experimental distributions of fitted localisations (distributions of intensity,
width, ellipticity, and noise) for various number of frames in order to simulate different localisation densities.
A: Region of a presentative frame of the experimental data which provided the distributions of localisations.
B: Region of a representative simulated frame. C-G: Super-resolved figures of the localisations obtained after
fitting the simulated data for increasing localisation densities of 0.83, 8.3, 41, 83, and 413 localisations/µm2,
resp. Because of various pitfalls in simulation, PSF fitting and estimation of the localisation densities, densities
of 4.6, 8.4, 29, 52, and 241 localisations/µm2 were recovered after PSF fitting, resp. H: Pair-correlation
functions of the recovered list of fitted localisations for the different localisation densities. I: Normalised
distributions of number of molecules per cluster after spatio-temporally grouping the recovered localisations
using the pair of thresholds (dT hresh = 60 mm, tT hresh = 3 s) for each simulated density of localisations. Scale
bar: 2 µm (same scale in panels A-G).
132 Counting T cell receptors
6.2.3 State of the TCR
Although none of the counting methods tested here proved to be very robust, their drawbacks
have two origins: the stochasticity of photo-blinking of the fluorophore and single-molecule
noise detected in the sample. For practical reasons, yeast cells were used as a control model
to express multiple repeats of mEos3.1 and test a few counting methods. But the single-
molecule noise originating from the fixed yeast cells might not be representative of the
noise observed in fixed T cells. This is why a monomeric (CD86), and a dimeric (CD28)
membrane protein, or the β chain of the TCR were fused to mEos2 and stably expressed in
a TCR-deficient hybridoma T cell line at low densities.
Designing a monomeric and a dimeric controls enables the direct comparison of the state
of the TCR to them. Since robust methods for counting absolute stoichiometry of complexes
are not available yet, the relative stoichiometry of the TCR to known multimers is studied
here.
Pre-activation and fixation
Observing the state of the TCR on a resting T cell presents challenges that must be carefully
addressed: (1) TCRs form micro-clusters at the membrane of the T cell shortly after being
triggered [203]; (2) recent work from Chang, Fernandes, Ganzinger, Lee et al. (submitted)
reveals that T cells trigger on many surfaces that used to be considered as non-activating
(e.g. IgG-coated glass, and as discussed in the introduction). This implies that imaging the
membrane of a live resting T cell might be more challenging than previously thought. This
could also account for the difference in the state of the TCR observed on the apical [250]
vs. the contact-forming [248] side of the cell.
Live T cells stably expressing a TCRβ -mEos2 fusion construct were dropped on a poly-
lysine coated coverslip and imaged under 561 nm and low 405 nm TIRF illumination. In
such conditions, bright clusters of TCRs were observed (Figure 6.9, C, arrows). The inten-
sity traces over time of these structures did not exhibit the photo-blinking or photo-bleaching
single-step behaviour characteristic of single fluorophores (Figure 6.9, A). In comparison,
single-molecule behaviour was detected outside of these clusters (Figure 6.9, B and D, arrow
heads). These observations indicate that micro-clusters are formed on IgG-coated surfaces
which thus activate the T cells. These data are consistent with observations of Williamson
et al. (ref. [257], supplementary information) that “adherence, even onto non-activating
antibodies results in a partial T cell activation”. They consequently conclude that to study
resting T cells, cells have to be fixed in suspension.
6.2 Results 133
Figure 6.9 Live T cells form TCR micro-clusters on a non-activating antibody-coated surface. Live T cells
stably expressing a TCRβ -mEos2 fusion protein were dropped on a glass coverslip coated with non-specific
antibodies and imaged under both 561 nm TIRF illumination and low and continuous 405 nm activation.
A-B: The pixel intensity traces of three bright and immobile PSFs (arrows in C), and three single-molecule
PSFs stochastically appearing during the experiment (the black arrow head designs the curve corresponding
to the PSF shown by the white arrow head in D) were plotted (exposure: 50 ms). C: Average intensity of
the 250 frames of the experiment for one representative T cell. D: One frame of the movie averaged in C,
during which a single-fluorophore (arrow head) appeared. Scale bar: 5 µm (same scale in panels C-D). A.U.:
arbitrary unit.
134 Counting T cell receptors
Figure 6.10 Assessing the fixation of T cells. Live T cells stably expressing a TCRβ -mEos2 fusion construct
were fixed in suspension in 4% formaldehyde and no (A) or 0.2% (B) glutaraldehyde in T cell medium for
30’ at 37◦C. After fixation, T cells were washed three times in PBS and immobilised under a 2% agarose gel
pad on a glass coverslip. Finally, cells were imaged under 561 nm TIRF illumination and 405 nm continuous
Fermi TIRF activation. For each fixation condition, a white light transmission snapshot of a representative cell
is shown, followed by seven consecutive frames of the SMLM experiment. Along the seven frame, arrows
track two activated fluorescent emitters. Scale bars: 5 µm (same scale for all panels).
6.2 Results 135
In order to image resting T cells, T cells were fixed in solution in 4% formaldehyde and
immobilised on the coverslip under a 2% agarose gel pad. This also had the effect of slightly
squeezing the cell on the coverslip, enabling to image a larger surface of the membrane of
the cells compared to imaging fixed cells sitting (by gravity) on a poly-lysine coated cover-
slip [251]. However, activated TCRβ -mEos2 proteins that could be followed during more
than a single frame were still mobile in the membrane (Figure 6.10, A, arrows). Consistent
with this observation, the shapes of the PSFs were also deformed, which is characteristic of
sub-exposure diffusion [188]. All movement and asymmetry of the PSFs disappeared when
adding 0.2% glutaraldehyde to the fixation buffer (Figure 6.10, B, arrows). Other fixation
protocols (methanol, different incubation time and temperature) as proposed by Tanaka et al.
for fixing membrane proteins were also tried [267]. Optimal activation efficiency, brightness
of the activated form of mEos2 and limited movement of the imaged proteins were obtained
when incubating the cells in 4% formaldehyde and 0.2% glutaraldehyde in T cell medium
for 30’ at 37◦C. This fixation protocol was therefore used in the experiments presented here.
Although fixed in solution, some cells showed some large-scale clustering, reminiscent
of a ‘protein island’ or ‘lipid raft’ structure (e.g. comparing Figures 6.11 and 6.13). This
effect might reflect on pre-activation state of the T cell, in which TCR micro-clusters start
to form. As this phenomenon was rarely observed, it is speculated that such cells might
have been activated while in contact with surfaces in the flask they were growing in, or in
the plastic tube they were fixed in. This underlines how sensitive T cells may be to most
surfaces for activation. Further research in the interaction of cells with surfaces of different
stiffness and its implication in cell growth and differentiation is required to provide better
understanding in the mechanism at stake [268]. Stiffness of surfaces is known to specifically
impact T cell growth and activation (cf. ref. [269] and personal communication from M.
Aßman, who observed different T cell activation behaviour on plastic and glass).
Population variability of TCRβ -mEos2 expression levels
As discussed in Section 6.2.2, the level of expression of the mEos2-tagged protein of inter-
est is important for all counting analyses studied here. Thus, some control over the level of
expressed fusion proteins was required. Moreover, because of the stochastic photo-physics
of single fluorophores and the stringent thresholds used to reject false localisation (cf. Ta-
ble 6.6), data from many cells have to be combined to draw any conclusion. Therefore, a
uniform population of cells with similar expression levels was seeked.
Examples of the spread of expression levels observed in cells stably expressing the
TCRβ -mEos2 construct are given in Figure 6.12. Similar spread was observed for the con-
136 Counting T cell receptors
Figure 6.11 Large-scale clustering suggest pre-activation of T cells. As in Figure 6.10, B, T cells expressing
a TCRβ -mEos2 fusion construct were fixed in suspension in 4% formaldehyde and 0.2% glutaraldehyde in T
cell medium for 30’ at 37◦C. After fixation, T cells were washed three times in PBS and immobilised under
a 2% agarose gel pad on a glass coverslip. Finally, cells were imaged under 561 nm TIRF illumination and
405 nm continuous Fermi TIRF activation. From the population of cells imaged, about 1-2% cells showed
some large-scale clustering in their super-resolved picture. Two representative examples are shown. A,C:
White light transmission snapshots of the cells. B,D: Corresponding super-resolved figures. Scale bar: 5 µm
(same scale for all panels).
6.2 Results 137
Figure 6.12 Variation of the expression levels of mEos2-fused proteins. T cells expressing a TCRβ -mEos2
fusion construct were fixed and imaged as described in Figure 6.11. Various expression levels of the protein on
the membrane of the T cells were observed. About 5% of the cells highly expressed the fluorescent construct.
Three examples of two to three nearby cells underline the range of expression levels observed in a single
population. A,C,E: White light transmission snapshots of the cells. B,D,F: Corresponding super-resolved
figures. Scale bar: 5 µm (same scale for all panels). Super-resolved images are convolved with an additional
Gaussian blur (σ = 110 nm) to increase the contrast of the figure.
138 Counting T cell receptors
Figure 6.13 Cells sorted for a specific expression level reverted to the diversity of expression levels of their
parental population. T cells expressing a TCRβ -mEos2 fusion construct were sorted in four populations by
flow cytometry depending on their fluorescence levels (as a measure for the expression level of the fusion
construct). After recovery, cells from the lowest- (A-C) and the highest-expressing (D-F) populations were
fixed and imaged as described in Figure 6.11. Very similar expression levels of the protein on the membrane
of the T cells were observed. For each population, the super-resolved figures of two representative examples
of nearby cells that were imaged by PALM are shown. Scale bar: 10 µm (same scale for all panels). Super-
resolved images are convolved with an additional Gaussian blur (σ = 110 nm) to increase the contrast of the
figure.
6.2 Results 139
trol cell lines expressing either CD86 or CD28 constructs. The fraction of high-expression
cells represented about 5% of the cells. In order to homogenise the population of cells, each
cell line was sorted in four populations, based on their green fluorescence (as a measure of
mEos2 expression) by flow cytometry. However, within a week (which was the time re-
quired for recovering a population large enough to perform the experiment), the expression
levels in each population had spread and usually did not differ from the parental population
(cf. the highest and lowest expressing populations in Figure 6.13, A-C and D-F, resp.). A
similar homogenisation and relaxation phenomenon has been described as a transcriptome
noise in stem cell populations [270]. Here, since the population variability could not be eas-
ily controlled, sorted populations were screened and, if their expression levels differed, the
populations with most comparable densities were further imaged. Finally, cells expressing
noticeably high densities of fusion proteins were avoided.
A rather monomeric TCR?
Despite the difficulties and limitations of the experiment, the three constructs (monomeric
CD86-, dimeric CD28-, and TCRβ -mEos2) were imaged under 561 nm illumination and
continuous Fermi 405 nm activation. Because the TIRF angle was optimised on each day
of experiment, only results from samples imaged on the same day were compared. For each
SMLM movie, both the pair-correlation function g(r) and, after grouping the localisations
with the parameters (60 nm, 3 s), the distribution of molecules per cluster were computed.
Results from different fields of view from the same biological sample were combined (as a
weighted mean for the pair-correlation function −Figure 6.14, left−, and as a sum for the
distribution of molecules per cluster −Figure 6.14, right). Finally, the average density of
localisations was estimated for the whole sample (Figure 6.14, numbers in parenthesis in
the legends in the left panels). Results of five independent days of experiments are plotted
in Figures 6.14 and 6.15.
In the three experiments presented in Figure 6.14, the localisation densities vary around
the threshold value of 30 localisations/µm2 (cf. Section 6.2.2). Thus, both pair-correlation
and grouping analyses have to be considered and sometimes compared. Interestingly, the
TCRβ -mEos2 construct is, in almost all experiments, the construct the least expressed. Al-
though sharing the same promoter than both other constructs, the requirement of TCRβ to
pair with a TCRα to be expressed at the membrane of the cell might explain its globally
lower expression. However, in the first two experiments, localisation densities for the three
constructs are similar (especially when correcting the number of localisations for the di-
meric control CD28-mEos2), validating the qualitative comparison of the results. Since the
140 Counting T cell receptors
Figure 6.14 Comparing TCR stoichiometry to a monomer and a dimer controls. Three repeats of the same
experiments were done in three different days and with independent batches of cells. Results from each
experiment are presented in each of the three rows. Each day, T cells expressing a monomeric CD86- (orange),
a dimeric CD28- (red), a TCRβ -mEos2 (green), or no (‘WT’, grey, in the second experiment) constructs were
fixed and imaged as described in Figure 6.11. Results from more than five cells were combined for each
sample (as a weighted mean for the pair-correlation function; as a sum for the distribution of molecules per
cluster). A,C,E: Plot of the pair-correlation functions of the constructs. Numbers in parentheses in the legend:
Density of localisations for each construct. B,D,F: Normalised distributions of the number of molecules per
cluster for all constructs after spatio-temporal grouping of the localisations with the pair of thresholds (60 nm,
3 s). Numbers in parentheses in the legend: Total number of molecules detected for each construct and number
of fields of view combined.
6.2 Results 141
densities are lower than the threshold value of 30 localisations/µm2, although close to it, the
distribution of molecules per cluster is discussed first.
In the three first experiments, the distribution of molecules per cluster for TCRβ -mEos2
is closer to CD86-mEos2 than to CD28-mEos2 (Figure 6.14, B, D and F). Interestingly,
the same conclusion can be drawn from the comparison of the pair-correlation functions
(Figure 6.14, A, C and E). In the second experiment, untransfected cells (‘WT’), with no
mEos2 expression, were fixed and imaged. The localisation density of this negative control
is very low compared to any other sample (1.8 ≪ 16.7 localisations/µm2); its distribution
of molecules per cluster is even more shifted to lower numbers of molecules than it is in
the case of the monomeric CD86-mEos2. Finally, the pair-correlation of the negative con-
trol stands below all other pair-correlation functions. This indicates that both CD86- and
TCRβ - (and therefore CD28-)mEos2 constructs are well expressed and imaged: for these
constructs, results successfully detect some signal above the single-molecule noise.
During the third experiment, a large variation of localisation densities was observed
between the three samples (more than a three-fold difference between TCRβ - and CD86-
mEos2). As the distribution of molecules per cluster is sensitive to differences in localisation
densities (cf. Section 6.2.2), and since the densities of both CD86- and CD28-mEos2 are
high above the threshold value for which the grouping analysis is valid, the distribution of
molecules per cluster should be disregarded (Figure 6.14, F). However, the pair-correlation
graph (Figure 6.14, E), consistent with both previous experiments, confirms that the pair-
correlation analysis is less sensitive to differences of localisation densities.
Altogether, these three first experiments are in favour of a monomeric TCRβ , consistent
with previous work [250, 251]. This conclusion however needs to be balanced by contrast-
ing results obtained in two other experiments (Figure 6.15). The fourth experiment (first
row in Figure 6.15) shows very little difference between the pair-correlation functions of
both monomeric and dimeric controls (Figure 6.15, A). However, the low localisation den-
sities, much lower than the threshold of 30 localisations/µm2, mean that the pair-correlation
analysis should be disregarded. The grouping analysis, though, does not reveal much dif-
ference between any of the samples, not even the negative control (‘WT’, Figure 6.15, B).
Noticeably, the densities of localisations for TCRβ -mEos2 and the negative controls are
almost identical, which indicate that the expression levels are too low for any signal to be
statistically detected over the single-molecule noise.
The fifth experiment (Figure 6.15, C-D) is contradictory. The grouping analysis is in
favour of a monomeric state of TCRβ -mEos2, for any of the three sorted populations im-
aged. But the localisation densities are almost all higher than the threshold of 30 locali-
142 Counting T cell receptors
Figure 6.15 Robustness and repeatability of relative counting of the TCR. Two additional repeats of the ex-
periments described in Figure 6.14 were performed in two different days and with independent batches of
cells. Results from each experiment are presented in each of the three rows. Each day, T cells expressing a
monomeric CD86- (orange), a dimeric CD28- (red), a TCRβ -mEos2 (green, purple and cyan, for different
expression levels), or no (‘WT’, grey, in the first experiment) constructs were fixed and imaged as described
in Figure 6.11. Results from more than five cells were combined for each sample (as a weighted mean for
the pair-correlation function; as a sum for the distribution of molecules per cluster). A,C: Plot of the pair-
correlation functions of the constructs. Numbers in parentheses in the legend: Density of localisations for
each construct. B,D: Normalised distributions of the number of molecules per cluster for all constructs after
spatio-temporal grouping of the localisations with the pair of thresholds (60 nm, 3 s). Numbers in parentheses
in the legend: Total number of molecules detected for each construct and number of fields of view combined
together.
6.2 Results 143
sations/µm2 above which the grouping analysis looses its validity. Conversely, the pair-
correlation analysis shows very little difference between the monomeric and dimeric con-
trols and indicates higher blinking rates (interpreted as clustering state) than the dimeric
control for the three sorted populations of cells expressing TCRβ -mEos2.
This last experiment could reveal some pre-activated cells, forming micro-clusters of
TCRs. However, the probability to observe such an artefact in three different samples is
very low. It nevertheless underlines the poor robustness of the experiment and its difficult
reproducibility due to the convolution of many stochastic and uncontrolled parameters: (1)
the stochastic photo-blinking of mEos2, (2) stochastic single-molecule noise, (3) the low
transcription of a gene, and the resulting expression of low densities of proteins [271, 272],
and (4) the possible pre-activation of the T cells before or during the fixation protocol.
6.2.4 Two-colour strategy
In an attempt to only select TCRβ -mEos2 PSFs and reject all single-molecule noise, a two-
colour strategy was designed [126]: an antigen-binding fragment (Fab) of a monoclonal an-
tibody directed against the ectodomain of TCRβ was labelled with Atto655 (cf. Section 3.3).
After fixation, cells were labelled with such Atto655-Fabs, washed, and immobilised under
an agarose pad. Samples were then sequentially imaged under 640 nm TIRF illumination,
and under both 561 nm TIRF illumination and continuous Fermi 405 nm activation.
For each field of view, localisations of each channel were extracted and their positions
channel-corrected (cf. Section 3.2.3). The super-resolved image of a representative cell ex-
pressing CD3δ -mEos28 on its plasma membrane and of a representative TCR-deficient neg-
ative control cell are shown in Figure 6.16, B and E, resp. Strikingly, a lot of single-molecule
noise is observe in both channels, and the CD3δ -mEos2 expressing cell is indistinguishable
from the negative control. Also, CD3δ -mEos2 and α-TCR Fabs do not co-localise signif-
icantly. This is confirmed by coordinate-based co-localisation analysis (cf. Section 3.4.2,
Figure 6.16, C and F-H).
These results underline that the two-colour experiment requires some optimisation of
the labelling strategy in relation with the fixation of the cells, of the CD3δ -mEos expression
level, and of the imaging protocol. However, before investing time on complex optimisation
8After a few unsuccessful attempts at labelling anti-human TCR antibodies with Atto655, the experiment
was transferred from the human model − J.RT3 cell line − to the mouse model − BW5147 cell line. Both
cell line do not express a full TCR complex at their surface: although J.RT3 is deficient in TCRβ subunit,
BW5147 cells lack both CD3δ and CD3ζ subunits. Consequently, the transfected subunit of the TCR, labelled
with mEos2, changed from human TCRβ to mouse CD3δ . And BW5147 cells were co-transfected with an
unlabelled mouse CD3ζ chain.
144 Counting T cell receptors
Figure 6.16 Two-colour labelling to reject single-molecule noise. CD3δ -deficient T cells expressing either a
CD3δ -mEos2 (top row) or no (‘WT’, middle row) fusion construct were fixed as described in Figure 6.11.
After three washes in PBS, cells were incubated for 30’ on ice with anti-TCRβ Fabs labelled with Atto655.
Cells were then washed three times in PBS and immobilised between an agarose gel pad and a glass cov-
erslip. Samples were sequentially imaged under 640 nm TIRF illumination, and under both 561 nm TIRF
illumination and continuous Fermi TIRF activation. For each sample, a representative cell is shown. A,D:
White light transmission snapshot of the cell. B,E: Super-resolved figure of the localisations detected in both
green (mEos2) and red (Atto655) channels. Localisations were corrected for channel miss-alignment (cf.
Section 3.2.3). C,F: Coordinate-based co-localisation (CBC) analysis of the SMLM localisations of the red
channel (cf. Section 3.4.2), with respect to the green channel. For each localisation of the red channel, a co-
localisation value between -1 (anti-correlation; blue) and 1 (positive correlation; orange) is computed. G-F:
Normalised distribution of the CBC values for all localisations of each sample (CD3δ -mEos2: plain bars;
TCR−: empty bars) with respect to either channel (G: green; H: red). Scale bar: 5 µm (same scale for all
panels).
6.3 Conclusion and future experiments 145
experiments, one might consider and evaluate the benefits and drawbacks of this approach:
by removing one stochastic unknown (single-molecule noise), the two-colour SMLM actu-
ally adds a few new unknowns that will each decrease the accuracy of the co-localisation
such as the labelling efficiency of the Fab, the sampling ratios of both labels, the switching
efficiency of Atto655, the compatibility of the imaging conditions for mEos and Atto655,
the registration error between both channels, or the accuracy of the labelling of two differ-
ent subunits on each side of the plasma membrane [52, 132]. In this respect, since TCRs
are expressed at the surface of the fixed cells at low density, Fabs could be labelled with
conventional non-switching dyes such as AlexaFluor647: most TCRs could be resolved and
part of the complexity of the two-colour SMLM experiment would be avoided.
6.3 Conclusion and future experiments
Counting absolute number of proteins by single-molecule fluorescence microscopy is still
at its beginning and important challenges are still to be overcome. Although it may confirm
other results or point towards possible stoichiometries, the robustness and reproducibility
of single-molecule counting remains unconvincing. This is due to the convolution of many
stochastic parameters that are still not well controlled and corrected for. Spatio-temporal
grouping-based counting methods require low densities of complexes which reduce the
signal-to-noise ratio and decrease the amount of data recorded per field of view (and because
of the stochasticity of photo-blinking, many fields of view are needed). Pair-correlation
based methods require higher sampling and densities of complexes. Although less sensitive
to single-molecule noise because they involve extensive fitting and estimation of parameters,
these methods are more optimised for clustering description than determination of absolute
stoichiometry.
Concerning the state of the TCR on a resting T cell, the stochasticity of the photo-
blinking of the fluorophore [141] is additionally obscured by issues raised by the specific
role of the TCR in T cell activation, since the state of the TCR changes during its triggering
which induces clustering. Moreover, recent work showed that T cells actually activate in the
absence of ligand when contacting a stiff surface, such as a poly-lysine coated glass surface
(considered until recently as a non-activating surface) (Chang, Fernandes, Ganzinger, Lee
et al., submitted).
Strategies to tackle the challenges of single-molecule counting need to carefully con-
sider the advantages and drawbacks of increasing the complexity of an already leading-
edge experiment, with all the technological complexity and subtle balances that such ex-
146 Counting T cell receptors
periments imply. In this chapter, exploring a two-colour co-localisation strategy to limit
single-molecule noise was a convincing idea to investigate but it added suboptimal and un-
satisfactory conditions to the experiment instead of increasing its accuracy. In a nutshell,
while rejecting noise, many new limitations were faced, introduced by the second label (e.g.
labelling efficiency, sampling, co-localisation).
Keeping the idea of using two colours to limit single-molecule noise, and because the
TCR is expressed at low densities on the plasma membrane, a repeat of the DySCo ex-
periment could compare the state of the TCR in a live cell at both the bottom of the cell
contacting the coverslip [251] and the apical membrane with a 3D single-molecule micro-
scope. This would investigate the hypothesis suggesting that the TCR would be monomeric
on a resting T cell and form micro-clusters as soon as it would contact a stiff activating
surface.
In a longer term, to avoid the intrinsic stochasticity of single-fluorophore photo-blinking
[273], recent methods based on bright fluorophores and high localisation precision are
emerging that may allow a subunit precision enabling the determination of the stoichiome-
try of complexes [7, 44]. Szymborska et al. confirmed the precise stoichiometry of the very
homogeneous nucleopore complex and partially resolved its structure by class averaging.
Another interesting method is quantitative DNA-PAINT, a version of uPAINT micros-
copy where proteins are tagged with DNA probes that are stochastically binding comple-
mentary DNA probes labelled with a dye in solution [96, 274]. Since ‘On’ times are con-
trolled by the affinity of the complementary probes and ‘Off’ dark times are controlled by
the concentration of DNA probes in solution and the number of tagged proteins at the same
position, the acquisition of many binding and dissociation events and the observation of
the resulting distribution of dark times at each observed position, for fixed samples, would
theoretically enable absolute counting. However, once again, careful study of the impact of
the wide distributions (e.g. intensity, dark times) characterising most properties of single-
fluorophore behaviour on single-molecule counting has to be undertaken [273].
Chapter 7
CD28 super-agonism
The original idea developed in this chapter was from Prof. S.J. Davis. All cell lines were
transfected, and most antibodies were labelled and provided by M. Aßmann. IL-2 assays
were also performed by M. Aßmann. All other experiments were done by M. Palayret. Ex-
perimental design was the fruit of the collaboration of Prof. S.J. Davis, Prof. D. Klenerman,
Dr S.F. Lee, Dr K.A. Ganzinger, M. Aßmann and M. Palayret. All data analysis, and code
writing was undertaken by M. Palayret. Dr S.F. Lee and M. Palayret designed and built the
super-resolution instrument used in this chapter.
Aim of the experiments described in this chapter
This chapter aims at studying the mechanism of T cell super-agonism activation (cf. Sec-
tion 7.1.2) at the single-molecule level for the first time, in parallel with functional IL-2
assays, in order to rule out, validate or refine the different models that have been proposed
for super-agonism (cf. Section 7.1.3). In the light of previous works by Chang, Fernandes,
Ganzinger, Lee et al. (submitted) unravelling ligand-independent TCR triggering in accor-
dance with the KS model, predictions from this model are especially targeted by the design
of the experiments.
Studying the mechanism of super-agonism is of key importance for theoretical and prac-
tical reasons. Any credible model of T cell activation has to explain related phenomena such
as antibody-induced receptor triggering. As such, understanding the mechanism of CD28
super-agonism helps the more general comprehension of T cell triggering. The mechanism
of CD28 triggering might also apply to other receptors of the immune system [275] and
beyond. More practically, in light of the dramatic effect of the TGN1412 SA in vivo, a
more complete understanding of CD28 triggering and super-agonism will allow to target
148 CD28 super-agonism
new signalling proteins to develop innovative drugs for immunotherapy.
7.1 Introduction
7.1.1 CD28, a co-stimulatory membrane protein
When recognising an antigen-presenting cell (APC), CD4+ T cells become activated and
start proliferating, allowing for a strong and specific immune response against the pathogen.
Activation of a T cell requires two different signals delivered by an APC: signal 1, or trigger-
ing of the TCR (see Chapter 6), and signal 2, induced by the triggering of a co-stimulatory
receptor.
Signal 1 is delivered by the TCR recognising its specific antigen (a non-self peptide
presented on a MHC receptor on the APC membrane). This signal alone is not sufficient to
activate a T cell. Indeed, without a co-stimulatory signal, the cell becomes anergic, i.e. it
becomes harder to activate in the future.
Signal 2, i.e. the co-stimulatory signal, is not antigen-specific and induced by membrane
proteins of the B7 family expressed on the APC upon stimuli specific to an infection, such
as products of pathogens or of dead or necrotic cells, through the engagement of Toll-like
receptors on the APC. This signal does not activate the T cell on its own, but modifies the
interpretation of signal 1. This mechanism prevents inappropriate activation of a T cell that
recognises self-antigens, which would lead to the development of an auto-immune reaction.
The CD28 (cluster of differentiation 28) receptor is the only co-stimulatory receptor consti-
tutively expressed on T cells. Engagement of both CD28 on naive T cells by either CD80
(B7-1, KD = 4 µM) or CD86 (B7-2, KD = 20 µM) ligands on APCs and the TCR induces IL-
2 transcription, expression of high-affinity IL-2 receptor and T cell proliferation [276, 277].
While CD80 is a homo-dimer, CD86 is monovalent and, in vivo, binds favourably to CD28
[277]. Each CD28 homo-dimer engages only one CD86 monomer. Upon activation, T
cells express other receptors such as CTLA-4 (cytotoxic T-lymphocyte-associated protein
4), ICOS (inducible T-cell co-stimulator) or PD-1 (programmed cell-death protein 1) on the
T cell membrane to further modulate this signal, either through interaction with the same B7
ligands (CTLA-4, inhibitory) or B7-like ligands (ICOS, co-stimulatory in a subset of cells)
[276].
CD28 is expressed on the plasma membrane as a homo-dimer thanks to a disulfide bond
on the ectoplasmic side of the protein, close to the plasma membrane [277]. Its cytoplasmic
tail contains different signalling motifs: a YMNM PI-3 kinase binding site whose activation
7.1 Introduction 149
leads to the BCL-XL pro-survival pathway, and a separate PYAP SH3-kinase binding site
critical for proliferation and IL-2 production [276]. These motifs are phosphorylated by ex-
ternal Src tyrosine kinases, but the mechanism by which triggering happens through ligand
recognition is unknown. Also, during TCR triggering, the TCR recruits kinases such as Lck
and ZAP-70 that become activated and, if recruited to the vicinity of CD28, can phospho-
rylate the tyrosines on CD28, introducing a cross-talk between the signalling pathways of
both signal 1 and signal 2. This idea gets very interesting in the light of ligand-independent
triggering of the TCR seeming to play a role in the anti-CD28 SA mechanism. However,
the precise mechanism by which TCR and CD28 cooperate, whether at the cell surface or
further downstream, is still unknown [278].
7.1.2 Super-agonism
Similar to the TCR, the natural ligand of CD28 can be replaced by a monoclonal antibody
(mAb) which binds to and cross-links CD28 in experimental systems [279]. Most antibodies
against CD28, referred to as ‘conventional antibodies’ (CA), similarly to the natural ligand
CD86, are co-stimulatory as they amplify an anti-TCR signal, but do not induce T cell
activation on their own. Surprisingly, some specific antibodies, subsequently called ‘super-
agonist’ (SA), can fully activate T cells without TCR engagement both in vitro and in vivo
(Figure 7.1) [278, 280]. A similar class of mAbs was discovered targeting the TCR that
activate T cells without requirement for signal 2 [281]. Interestingly, as for SA targeting the
TCR, soluble SAs targeting CD28 do not activate T cells. Immobilisation of the SA antibody
is necessary for activation of cells independently of TCR triggering. Immobilisation of the
SA via Fc receptors is also required in vivo [282].
In the rat animal model, two antibodies (one SA and one CA) have been thoroughly
studied [278]. A couple of differences between them were observed: (1) the CA blocks
SA-binding to CD28, suggesting that both antibodies bind related epitopes on CD28 or that
they sterically exclude each other’s binding, and that the CA binds CD28 in a more efficient
way than the SA; (2) although interaction of either antibody with CD28 is very stable, the
CA binds much faster to CD28 than the SA, confirming the previous conclusion. This is,
however, specifically observed in primary T cells, but not in CD28-transfected fibroblast
cells. This indicates that the binding rate of the SA depends on its local environment. Tacke
et al. propose a model where CD28 is expressed on the plasma membrane with two hypo-
thetical conformations in equilibrium. The SA would only bind to and stabilise a signalling-
competent conformation; the CA would indifferently bind to either conformation. Binding
150 CD28 super-agonism
Figure 7.1 Super-agonistic antibodies do not require TCR engagement (signal 1) for activating T cells. The
IL-2 secretion from cells incubated both with (right panel) or without (left panel) an anti-TCR SA (to trigger
the TCR without need of signal 2), and with either no anti-CD28 antibody or one of two anti-CD28 antibodies
(a so-called super-agonist−SA− and a conventional−CA− antibody) is recorded. IL-2 production is used as
a marker for T cell activation.
7.1 Introduction 151
of the SA would indeed be slower than binding of the CA. It would also induce a shift in
the conformational equilibrium of CD28, explaining the ability of the SA to activate T cells
without TCR co-engagement. This model, however, does not explain the necessity of SA
immobilisation.
Super-agonists are not known natural ligands of CD28 in vivo. Super-agonism could
therefore be regarded as an in vitro artefact. In preclinical murine model, the stimulation
of CD28 with a SA activated and expanded CD4+ T cells, and preferentially regulatory T
cells, without detectable toxic or proinflammatory effect. Anti-CD28 SAs were hoped to be
useful as anti-inflammatory immune modulators. In 2006, however, the unfortunate Phase I
clinical trial of the anti-CD28 SA TGN1412 drew attention to the unexpected potency of this
class of antibodies in vivo, which resulted in cytokine release syndrome, severe lymphopenia
and life-threatening multi-organ failure in six trial volunteers [283]. Therefore, the medical
interest for understanding the mechanism of super-agonism is evident.
7.1.3 A structural difference
A systematic study unravelled a convincing correlation between the functional class of SA
and topology of epitope binding [284]: in rat and human cell lines, all studied mitogenic
antibodies bind to the C"D loop of CD28, whereas all CAs bind close to the CD86 binding
motif. The resolution of the crystal structure of the ectoplasmic domain of CD28 bound to
an SA Fab, together with cryo-EM data, confirmed the predicted topology of both epitopes
[285]: SAs bind on the ‘side’ of the extra-cellular domain of CD28, close to the plasma
membrane, orthogonally to the epitope of CAs, localised on the ‘top’ of the homo-dimer.
Topologically, an SA Fab is consequently∼ 75 Å closer to the plasma membrane than a CA
Fab [234]. Importantly, the correlation between mitogenic potential and epitope position is
not only verified for anti-CD28 antibodies, but also for anti-TCR and anti-PD1 antibodies
(personal communication from Prof SJ Davis).
In summary, different models were proposed to investigate the functional difference of
the two classes of mAbs [284, 285]:
Preferential engagement of signalling-competent CD28: this model is based on the hy-
pothesis of an oligomerisation heterogeneity of the CD28 population. The SA would
favourably bind and cross-link a signalling-competent mobile fraction, the ‘side’ epi-
tope of CD28 being more difficult to access than the ‘top’ one in signalling-incompetent
immobile micro-clusters [284]. It is, however, hypothetical whether these two sub-
populations exist and this model does not explain the requirement for antibody immo-
152 CD28 super-agonism
bilisation in super-agonism activation.
Differential cross-linking: whereas the physiological ligand CD86 is monomeric, and each
CD28 dimer engages only one CD86 monomer (probably due to steric constraints),
both SA and CA bind CD28 divalently and are thus able to cross-link CD28 in pos-
sibly large-scale linear lattices through periodic CD28-mAb interactions [285]. Such
lattice structures were described for the interaction of CTLA-4 with CD80. Through
avidity, they are thought to dramatically increase the stability of the CD80-CTLA-4
interaction and the downstream inhibitory signalling responsible for ending the sig-
nalling cascade. Antibody-induced cross-linking of CD28 could similarly amplify
CD28 signalling. The position of the SA epitope, on the ‘side’ of the CD28 homo-
dimer, could hypothetically favour the binding of two different CD28 homo-dimers
over the binding of both epitopes of the same homo-dimer, favouring cross-linking
and explaining the increased triggering effect of the SA over the CA. However, no
difference in cross-linking ability has been described so far. Nevertheless, a system-
atic difference of binding kinetics or dissociation constants between both classes of
antibodies could also induce differential cross-linking. However, although this has
not been investigated yet, the binary ‘all-or-nothing’ activation and the number of an-
tibodies tested for each class makes this possibility unlikely. Furthermore, this model
does not account for the requirement for antibody immobilisation in super-agonism
activation, since cross-linking is favoured in solution.
Antibody-induced conformational change: binding of SAs could favour an ‘active’ con-
formation of CD28. Indeed, since all SAs bind to the same epitope, as in the key-lock
model of ligand interaction, binding of the SA, but not the CA, could stabilise an
otherwise less stable but signalling-competent conformation of CD28. This model
is nevertheless unlikely as it requires long range conformational changes spanning
across the plasma membrane, which is rare in interactions between antibodies and
protein antigens [286]. Crystallographic evidence suggests that a rigid-body interac-
tion with very little conformational change characterises the binding of CTLA-4 with
its ligand [287]. Also, antibodies binding to shared epitopes rarely bind in exactly
the same way [288]. However, a novel physical model involving a piston-like move-
ment induced by the binding of the SA close to the plasma membrane, inspired by the
mechanism of the aspartate receptor [289], could explain the activation of CD28 but
still needs to be investigated [290]. Against this model, Evans et al. argue that the
recent crystal structure is in favour of a single stable state of the CD28 homo-dimer
7.1 Introduction 153
[285]. Lastly, as proposed, the piston model also fails to account for the requirement
for antibody immobilisation.
Figure 7.2 Kinetic-segregation model for CD28 super-agonism. The 75 Åof height difference between the SA
and the CA could be responsible for differential segregation of phosphatases with large ecto-domains such as
CD45. When a T cell (top, light green), decorated with CA forms a contact on a glass surface coated with
secondary antibodies, the size of the contact does not prevent CD45 to diffuse in the contact. Thus, a null
net phosphorylation of the CD28 and TCR receptors would be observed, leading to no activation of the cell.
However, when the cell is decorated with the SA, the 75 Åshorter contact may hinder the entry and diffusion
of membrane proteins with large ecto-domains, such as CD45. Thus, CD45 is excluded from the contact zone
where the small Lck kinase is free of diffusing. This local segregation shifts the balance of kinase/phosphatase
and the net phosphorylation of CD28 and the TCR get positive. If the receptors do not diffuse out of this local
segregation zone, downstream signalling induces activation of the cell.
Proximity effect: the SA could also alter the arrangement of signalling proteins, either by
physically interfering with signalling components and influencing the proximity of
intracellular effector proteins, or by segregating surface receptors in a size-dependent
manner analogous to the kinetic-segregation (KS) model proposed for TCR triggering
[234] (Figure 7.2). Indeed, by binding to an epitope closer to the plasma membrane,
SAs could either (1) sterically and very locally seclude all transmembrane protein with
an ectodomain from getting near CD28 in a more efficient way than CAs, or (2) induce
a closer contact zone inducing better size-dependent segregation of large-ectodomain
transmembrane proteins. The requirement of antibody immobilisation also suggests
the importance of the formation of a contact between the T cell and a surface (either
the coverslip coated with anti-Fc secondary antibodies or a cell membrane presenting
Fc receptors). This requirement is a typical prediction of the KS model.
154 CD28 super-agonism
7.2 Results
A dramatic increase in T cell activation, as measured by IL-2 production, was reproduced
in mouse T cells (Figure 7.1) when stimulating with an indirectly immobilised SA antibody
against CD28, compared to a CA. The super-agonistic signal, i.e. the binding of SA antibod-
ies to CD28, consequently needs to pass from the extracellular surface through the plasma
membrane to activate the downstream signalling pathways. Similarly to TCR triggering,
this information can travel by different means: a change of conformation, a binding-induced
aggregation, or a change of localisation.
In order to understand the effect of anti-rat-CD28 SA antibodies on CD28 at the mech-
anistic scale, a chimera CD28 (ectodomain of rat-CD28 fused to the transmembrane and
signalling domain of the mouse-CD28) was fused to the photo-activable fluorescent protein
mEos and stably expressed in the mouse BW5147 cell line. Therefore, the precise sub-
diffraction distribution and diffusion of CD28 could be studied and compared under SA or
CA activation.
7.2.1 Super-agonistic antibodies slow CD28 down
Stably transfected T cells were incubating for 15’ at 37◦C with 50 µg/mL of either SA
or CA, dropped onto a coverslip coated with polyclonal secondary antibodies, and imaged
under 561 nm TIRF illumination (Figure 7.3). A very small amount of additional 405 nm
illumination was necessary for mEos activation.
Figure 7.3 Scheme of the diffusion experiments. TCR-deficient cells were stably transfected with a construct
of a mouse CD28 protein with the ectoplasmic domain of rat CD28, fused to a 30 amino acid linker and the
mEos3.2 fluorescent protein at its intracellular end. Cells were incubated at 37◦C for 15’ with either SA,
CA or no antibody (left), dropped on a coverslip coated with secondary antibodies and imaged under TIRF
illumination (right).
7.2 Results 155
At 37◦C, activation by SAs induced the formation of clusters of localisations in the
super-resolution image (Figure 7.4, top left panel) compared to activation by CAs (middle
left panel), or cells left incubating in medium without antibody (bottom left). Since all lo-
calisations from all frames were plotted, such clusters can originate from two non exclusive
effects: (1) real aggregation of separate CD28 fluorescently-tagged proteins, or (2) slowly
diffusing CD28 proteins observed during multiple successive frames.
To study both effects, and because single fluorophores could be visually followed during
several frames, localisations within 300 nm from each other were grouped into trajectories,
allowing blinking events of up to four frames (as described in ref. [154] and Section 3.4.2).
The trajectory radius of 300 nm (less than the dimension of three pixels of the camera),
together with a frame exposure of 50 ms (frame rate of 20 Hz) allowed detection of diffusion
coefficients up to 0.1 µm2/s [79]. A characteristic example of CD28-mEos tracks is given in
Movie B.1.
Mean square displacement (MSD) analysis [291, 292] of the long trajectories (longer
than ten frames, the five first time-points being used for the linear regression) showed a
qualitative slowing down of CD28 induced by the SA (Figure 7.4, top middle panel) com-
pared to activation by the CA or absence of activation (resp. middle and bottom middle
panels). However, in order to quantify this behaviour, and because many trajectories were
short [293], an approach using the probability distribution of square displacements was per-
formed [294]: the probability P(r2,τ) that a particle is found within a circle of radius r from
its initial position after a time-lag τ was calculated from the trajectories of at least five cells
for ten multiples of the exposure time (τ ∈{50k ms; k∈ [[1, 10]]}). P(r2,50 ms) corresponds
to the cumulative jump-distance distribution described by Gru¨nwald et al. [293, 295, 296].
For each time-lag τ , models of p diffusing populations with diffusion constants Di can
be fitted to the distributions [294, 297]:
P(r2,τ) = 1−
p
∑
i=1
[
fi exp−r
2(τ)
SDi
]
(7.1)
This is exemplified for p ∈ [[2, 4]] populations for the distributions at τ = 100 ms in
Figure 7.4 (right column). All conditions (top to bottom: SA, CA or no antibody activation)
are well-approximated by the models of three or four diffusing populations (resp. orange and
yellow) as revealed by the residuals that show very little systematic deviation. Conversely, a
model of two populations (red) does not efficiently describe the data. The best parsimonious
description is thus a model of three diffusing populations.
For each population, the mean square displacement SDi(τ) is proportional to the time-
156 CD28 super-agonism
Figure 7.4 Analysis of CD28 sptPALM trajectories. Live TCR− T cells transfected with a [full-length CD28]-
mEos construct were decorated with either a SA (A-C), a CA (D-F) or no antibody (G-I) and then dropped
onto and imaged on a coverslip coated with secondary antibodies. All PSFs were fitted with 2D-Gaussians.
All localisations were plotted as 2D-Gaussians of similar intensity and of width equal to the average precision
of the localisations (A,D,G, intensity increasing from yellow to red to black). Localisations were also grouped
in trajectories (localisations within 300 nm from each other, allowing blinking events of up to four frames)
and trajectories grouping more than ten localisations were plotted (B,E,H). Colours of the trajectories code
for the MSD diffusion coefficient of the trajectory (calculated with the five first time-lags) as indicated in the
log-scale. C,F,I: the cumulative jump-distance distribution for the time-lag τ = 100 ms (black) and the fit of
models of two (red), three (orange) and four (yellow) diffusing populations. Inset: zoom-in on the inflexion
of the distribution with different scales for the x-axis. Residuals from the fits are also plotted.
7.2 Results 157
Figure 7.5 Choice of the number of diffusing populations. A model of two (left column), three (middle) and
four (right) diffusing populations were fitted to the data shown in Figure 7.4, top row. For each time-lag τ , the
full distribution P(r2,τ) (◦) and five separate parts randomly drawn from it (, horizontal bar: mean, whiskers:
one standard deviation of the mean) were fitted with Equation 7.1. For each model, both SDi(τ) (top rows)
and fi(τ) (bottom row) are drawn for each population i (colours for each population i ∈ [[1, 4]]: dark blue,
orange, purple, pink).
158 CD28 super-agonism
lag τ if the motion of the population is purely Brownian. However, if the localisation error
σ is assumed to be normally distributed, the relation between SDi and τ is linear because of
the localisation precision [298, 299]:
SDi(τ) = 4 Di τ+4σ2 (7.2)
The curves SDi(τ) are shown for the case of the SA activation in Figure 7.5. Curves
for each of the populations of three models (two, three and four diffusing populations −
left, middle, right columns, resp.) are given, together with the variation of the relative
proportions fi(τ) of the populations (bottom row). In order to test the robustness of the
jump-distance distribution fit − from which the parameters SDi(τ) and fi(τ) are extracted
−, all trajectories (ncells ≥ 5) are used to calculate the cumulative jump-distance distribution
(distribution of distances separating two positions in trajectories separated by τ in time).
This new distribution, for each time-lag τ is randomly divided into five distributions. The
same analysis is then performed for the initial jump-distance distribution (◦) and for its five
separate parts randomly drawn ().
Whereas the values of the parameters (SDi and fi) are reproducibly well fitted in the case
of two or three populations, the poor robustness of the fit of four populations reveals that
the latter has excessive degrees of freedom for the available quantity of data. This confirms
the choice of a model of three diffusing populations (Figure 7.5, middle column) as the best
parsimonious model for the data described here.
Since the mean precision of the experiments is σ ∼ 40 nm, squared displacements be-
low σ2 = 1,600 nm2 are to be considered null, and the corresponding tracked particles,
immobile. Interestingly, a linear regression of SD1(τ) following Equation 7.2 confirms the
range of the precision: σ = 66 nm. Consequently, the movement of the third and slowest
population falls under the precision of the experiment (Figure 7.5, middle column, SD3).
Furthermore, this fixed population is always observed, even for mEos-tagged proteins other
than CD28, such as CD86-mEos (Section 7.2.2), and is therefore not specific to the super-
agonism experiment. This population might reflect a non-CD28-specific immobile fraction
of transmembrane proteins, the presence of dust on the coverslip, protein aggregates from
dead cells, or antibody aggregates sticking on the glass. It is henceforth ignored and the
relative fraction of the first population to the second is defined as:
f ′1 =
f1
f1+ f2
(7.3)
The SDi(τ) plots are characterised by a linear slope for τ ≤ 200 ms, plateauing for
7.2 Results 159
larger values of τ , typical of a spatially-restricted diffusing population of molecules [292,
297, 300]. A linear regression of the first four points determines the diffusion coefficients
Di (Equation 7.2). The time-lag-mean value F ′1 of the relative fraction f
′
1(τ) is also calcu-
lated. Final values D1, D2 and F ′1 are given in Figure 7.6, both for the total jump-distance
distribution (◦) and for its five random sub-divisions (plotted as lower-upper-quartile bars
±1 standard deviation).
At 37◦C, as expected, the diffusion of the fast population of CD28 is slowed down
when applying either SA or CA (Figure 7.6, ‘Full-length at 37◦C’). The effect of SA
activation on the diffusion coefficient, however, is significantly larger than CA activation
(DSA1 = 0.026 µm
2/s < DCA1 = 0.053 µm
2/s; p = 5 ·10−7). The diffusion of the slower pop-
ulation is, however, not significantly affected by the type of activation (p = 0.07). In the
absence of activation, diffusion of the slow population is indistinguishable from being null,
indicating that a model of two diffusing populations would best describe this experimental
condition. Finally, the relative fraction of the fast population F ′1 specifically increases in the
case of CA activation: the CA slows down the fast population of CD28 qualitatively less,
but quantitatively more, than the SA.
It is expected that passive diffusion is correlated with temperature. Indeed, at 20◦C,
globally lower diffusion behaviour is observed (‘Full-length at 20◦C’); this is in particular
visible for the fast diffusing population. Interestingly, at this temperature the slow popula-
tion mimics the behaviour of the fast population at 37◦C: since all diffusion is slowed down,
part of the slowing effect observed in the fast population at 37◦C has consequently moved to
the slow population at 20◦C in the analysis. The increase of the relative fraction of the fast
population in the case of no activation indicates that more of the fast diffusing population
is detected by the analysis. Indeed, this is a thresholding effect due to the radius parameter
of 300 nm chosen for single molecule tracking: molecules diffusing faster than 0.1 µm2/s
cannot be tracked [79]. Therefore, a decrease in temperature slows these non tracked mole-
cules such that they can now be detected, consequently increasing the proportion of the fast
population.
Deconvolving diffusion from the downstream effects of early T cell activation
T cell activation happens within five minutes of contact with an activating surface, even
in the absence of a TCR ligand (Chang, Fernandes, Ganzinger, Lee et al., submitted). To
separate the dramatic downstream effects of T cell activation (which induces immunological
synapse formation, involving deep reorganisation of the cell architecture [301]) from the
direct early mechanical effect of the SA, BW5147 cells which do not express TCR and
160 CD28 super-agonism
Figure 7.6 Robust decrease of diffusion of CD28. The effect of SA (red), CA (green) or no antibody (white) ac-
tivation on a full-length and a signalling-incompetent (‘truncated’) CD28-mEos constructs were compared. A:
Scheme of the diffusion experiments shown for CA incubation. Purple arrows symbolise 405 nm illumination
stochastically switching mEos to its red form; yellow arrows, 561 nm illumination. B: IL-2 production of the
cells expressing either construct at 37◦C in TCR+ and TCR− cells, measured by M. Aßmann (three repeats,
± one standard deviation of the mean). C-D: Diffusion coefficients of the fast (D1) and slower (D2) diffusing
populations, calculated for both CD28 constructs at both 20◦C and 37◦C. E: Fraction F ′1 of the first population
relative to the first and second populations (Equation 7.3). Values for the total jump-distance distribution (◦)
and for its five random sub-divisions (lower-upper-quartile bars±1 standard deviation of the mean) are shown.
7.2 Results 161
therefore cannot be activated were used. As the diffusion experiments are on this time-
scale, the effects of ligand-independent TCR triggering and super-agonistic activation would
be convolved and difficult to separate were the TCR expressed.
To focus on the initial passive reorganisation of receptors in isolation from downstream
signalling-induced reorganisation processes, a truncated version of CD28 without its intra-
cellular signalling domain was engineered by M. Aßmann. The absence of TCR and CD28
signalling were checked by IL-2 assay (Figure 7.6, B). In both cases, IL-2 production was
depleted.
Slightly faster, but very similar diffusion behaviours to the full CD28 protein were ob-
served both at 37◦ and 20◦C (Figure 7.6, ‘Truncated’ at ‘37◦C’ and at ‘20◦C’). Indeed,
because part of the intra-cellular domain is truncated, the apparent size of CD28 is reduced
which consequently increases its diffusion motion [302].
An important increase in the diffusion coefficient of the fast moving population for the
CA activation is noticeable, compared to the full-length CD28, especially at 37◦C. This
indicates that the truncated version of CD28, while preventing convolution of the SA ef-
fect with any downstream signalling effect, increases the dynamic range of the diffusion
analysis. Truncated signalling-incompetent CD28 is therefore used as the standard CD28
construct for all following diffusion experiments in this thesis.
Finally, in order to directly and formally validate the comparison between the IL-2 assay
and the diffusion experiments, a single cell line was grown to be studied with both proto-
cols: TCR-sufficient BW cells were transfected with a [full-length CD28]-mEos construct.
Both IL-2 assay (M. Aßmann, personal communication) and the diffusion experiments (Fig-
ure 7.7) were in accordance with previous IL-2 results for TCR+ cells transfected with a
full-length CD28 construct (Figure 7.6, B), and diffusion results for TCR− cells transfected
with a full-length CD28-mEos construct (Figure 7.6, C-E), respectively.
Specificity of the decrease of CD28 diffusion
Since CD28 is dimeric [276], both antibodies (SA or CA) and CD28 present each two
corresponding binding sites [285, 303]. Large scale linear structures can therefore form
if the ratio of antibody bound to CD28 and CD28 itself is close to one. Such a structure,
similar to polymers, could form a large meshwork increasing the viscosity of CD28 in the
fluid plasma membrane and sterically decreasing the diffusivity of its components. To test
whether the SA-specific decrease in diffusivity of CD28 affects all transmembrane proteins,
both truncated but unlabelled CD28, and another transmembrane protein, CD86, fused to
mEos, were stably transfected in T cells and imaged as previously described.
162 CD28 super-agonism
Figure 7.7 Compatibility of IL-2 and diffusion experiments. A single cell line was used in both IL-2 and
diffusion experiments in order to formal validate their compatibility. The diffusion of the full-length CD28-
mEos construct expressed in TCR-sufficient T cells is observed, after either SA (red), CA (green) or no
antibody (white) incubation at 37◦C. A-C: Diffusion coefficients of the fast and slower diffusing populations
and their relative fractions for the three incubation conditions. Same data representation as for Figure 7.6.
7.2 Results 163
Figure 7.8 Specificity of the decrease of CD28 diffusion. The diffusion of the truncated signalling-incompetent
CD28-mEos construct is compared to the diffusion of a CD86-mEos construct co-transfected with the unla-
belled truncated CD28 protein, after either SA (red), CA (green) or no antibody (white) incubation at 37◦C.
A-C: Diffusion coefficients of the fast and slower diffusing populations and their relative fractions for the three
incubation conditions. Same data representation as for Figure 7.6. D: Scheme of the diffusion experiments
shown for CA activation.
164 CD28 super-agonism
The diffusion of CD86 was affected by neither of the activation antibodies (Figure 7.8).
The SA, therefore, − at least to some extent − specifically slows down CD28 diffusion.
7.2.2 Super-agonist activation requires divalent engagement of CD28
CD28 monomers do not trigger
To test the hypothesis whether a meshwork of SA and CD28 is responsible for the effect of
the SA on CD28 diffusion, a monomeric version of CD28 was engineered: the ectodomain
of a monomeric transmembrane protein, human PD-1, was fused to replace the ectodomain
of CD28 in the truncated CD28-mEos construct. Two monoclonal antibodies against distal
epitopes of PD-1, were used either separately or in combination during the 15’ incubation
in place of the previous mAbs.
Each of the antibody separately induced a decrease in diffusion of PD-1-CD28 proteins,
very similar to the CA antibody (Figure 7.9, ‘CA 1’ and ‘CA 2’). Strikingly, applying
both antibodies simultaneously caused a drop in diffusivity reminiscent of the effect of SA
on CD28 diffusion (‘CA 1 + CA 2’). Simplistically, this result can be interpreted as two
different roles played by the two sets of antibodies: one forms CD28 dimers from two PD-
1-CD28 monomers; the other one induces cross-linking.
However, since a different system is used (different set of antibodies and antigen), the
values obtained are not directly comparable in a quantitative way. For better comparison,
three populations were used although parameters are noticeably correlated, indicating that
a model of two populations would have been a better parsimonious model to describe the
data. Nevertheless, this results, together with its IL-2 equivalent (Figure 7.9, B), emphasises
the importance of the divalence of CD28 for super-agonism activation.
Effect of the antibody concentration
In order to further compare the behaviour of both SA and CA, the effect of their concen-
trations on CD28 diffusivity was investigated. The diffusivity of both populations of CD28
importantly decreased with higher concentration of SAs through the whole range from 0.1
to 50 µg/mL (Figure 7.10, B-D). This is consistent with previous dilution experiments mea-
suring the effect of SA on proliferation of T cells [278] or IL-2 assays (Figure 7.10, A).
Strikingly, the concentration of the CA has quite a different effect on the diffusion of
CD28: the diffusion of the fast population of CD28 reaches a minimum at an intermediary
concentration of 1 µg/mL (bottom row).
7.2 Results 165
Figure 7.9 Effect of cross-linking. The effect of the incubation of two mAbs (‘CA 1’ and ‘CA 2’), separately
(red, green) or in combination (green and red stripes), or no antibody (white) on the diffusion of a monomeric
CD28-mEos-like construct, PD-1-CD28-mEos, is used to study the effect of cross-linking. A: Scheme of
the diffusion experiments shown for activation with either antibody. B: IL-2 production measured in all four
conditions by M. Aßmann. C-E: Diffusion coefficients of the fast and slower diffusing populations and their
relative fractions for the three incubation conditions. Same data representation as for Figure 7.6. F: Crystal
structure of PD-1 ectoplasmic domain (the plasma membrane would be at the bottom) showing the epitopes
of both mAbs (figure courtesy of S.J. Davis).
166 CD28 super-agonism
Figure 7.10 Effect of the antibody concentration. The effect of the incubation of different concentrations of SA
(red, A-D) and CA (green, E-G) mAbs on the diffusion of the transfected signalling-incompetent CD28-mEos
construct was studied. A: IL-2 production, measured by M. Aßmann for a series of SA concentrations. B-D
and E-G: Diffusion coefficients of the fast and slower diffusing populations and their relative fractions for the
three incubation conditions. Same data representation as for Figure 7.6.
7.2 Results 167
By comparing levels of surface expression of CD28 between the stably transfected cell
lines and primary cells by flow cytometry, R. Knox roughly estimated that between 0.5
and 1 million of CD28 homo-dimers were expressed per cell (personal communication).
From this value, a quick estimation of the ratio of antibody over CD28 homo-dimer can
be calculated (Table 7.1). This indicates that the range of concentrations chosen probably
spans conditions of excess of CD28 and of excess of antibody.
Table 7.1 Approximative ratio of mAb to CD28 expressed on the surface of the T cells in the experiments for
different concentrations of antibodies.
[mAb] (µg/mL) 0.1 1 5 50
[mAb] (mol/L) 7 ·10−10 7 ·10−9 3.5 ·10−8 3.5 ·10−7
# mAb/CD28 ∼ 0.8 ∼ 8 ∼ 40 ∼ 400
7.2.3 Effect of different coverslip antibody-coatings
Together with its concentration during incubation, the interaction of the activation antibody
with CD28, which differs when the antibodies are decorating the cells in solution or coated
on the coverslip, is critical to observe super-agonism. As for TCR triggering [235], M.
Aßmann has shown that immobilisation on a surface is required for SA activation by mea-
suring IL-2 after incubating cells with F(ab′)2 fragments of SA that do not bind to anti-Fc
secondary antibodies (personal communication, also ref. [278]). However, IL-2 produc-
tion was recovered when cells were dropped on coverslips coated with secondary antibodies
directed against the κ chain of F(ab′)2.
Here, the effect of different activation protocols is tested: first, instead of decorating
the cells in solution (Figure 7.11, ‘mAb in solution’), activation antibodies (either SA or
CA) were used to coat the coverslip (‘mAb coated’), in place of secondary antibodies. For
non-decorated cells placed on mAb coated coverslips, the SA-specific decrease of diffusion
of CD28 is not observed any more; conversely, the diffusion of the slower CD28 population
speeds up and the fraction of the faster population increases. Indeed, when the cells are
not decorated, the formation of a meshwork of CD28 is not favoured because of combined
geometric and mobility effects due to the immobility of the antibodies.
However, when cells are decorated in solution with the activation antibodies and dropped
on coverslips also coated with the same activation antibodies (‘mAb coated & in solution’),
the SA-specific effect is recovered. This suggests that the secondary antibodies play the
role of immobilising the activation antibodies on the coverslip. Both immobilisation and
cell-decoration are therefore required to observe SA-specific CD28 slowing down.
168 CD28 super-agonism
Figure 7.11 Effect of different coverslip antibody-coatings. The effects of different antibody-coatings on the
signalling-incompetent CD28-mEos protein were compared. A: Scheme of the different coating protocols
represented for the CA. Different combinations of steps were experimented: cells were either decorated with
the mAb in solution (SA, red; CA, green; or no antibody, white) or left in warm medium; in one case (‘mAb in
solution, then washed’), unbound mAbs were then washed off; finally, they were dropped on coverslips coated
with either the mAb or a secondary polyclonal antibody against the mAb. B: IL-2 production, measured by
M. Aßmann for the four immobilisation conditions described in A. Since the CA do not induce IL-2 secretion,
experiments only involved the SA. C-E: Diffusion coefficients of the fast and slower diffusing populations and
their relative fractions for the four immobilisation conditions. Same data representation as for Figure 7.6.
7.2 Results 169
A more complex case consists in repeating the initial experiment, but washing out
the excess of activation antibodies after their incubation with the cells (‘mAb in solution,
washed’). Strikingly, the slowing down effect is even more increased, and affects both
CA and SA activation. A possible interpretation for this observation is that the activation
antibodies are either better immobilised on the coverslip or better decorating the T cells. In-
deed, by washing out the unbound activated antibody, there are no antibodies left in solution
that compete with CD28-bound antibodies and that effectively partially block the secondary
antibodies. The immobilisation of activation antibodies is thus increased.
7.2.4 CD28-bound antibodies
Because the SA and the CA used in this study have different binding kinetics, the ther-
modynamic equilibrium is probably not reached when imaging begins in the case of SA
activation. To investigate whether this would affect the distribution of the antibodies bound
to CD28, both antibodies were labelled with Atto655, a far-red photo-switching fluores-
cent dye. After incubation with either mAb (20% labelled), cells were thoroughly washed
from any excess of antibody, and imaged in a simultaneous two-colour PALM-dSTORM
experiment.
The labelling and imaging protocols were optimised to obtain sufficient, but not too
high, density of both labels (mEos and Atto655) for thousands of frames. Atto655 was
chosen as a probe absorbing and emitting in separate wavelength domains than mEos, and
whose photo-blinking is not controlled by near-UV light. Thus, the densities of both labels
can be controlled independently: mEos through 405 nm illumination [121], Atto655 by the
percentage of labelled over unlabelled antibodies used. A reducing agent was not used for
live-cell imaging because of cytotoxic effects [134, 304]. Both labelling strategies are thus
orthogonal.
For both antibody activations, the average projection of each channel (Figure 7.12) re-
veals a cluster distribution similar to the one previously observed in the PALM-only ex-
periments (Figure 7.4, ‘SA’ and ‘CA’). This distribution is induced by the antibodies as a
uniform distribution of CD28 proteins is observed in their absence (‘No mAb’). Almost no
cross-talk of mEos in the far-red channel was recorded (except for high-intensity intracel-
lular vesicles). The untransfected original cell line was also imaged (‘WT’): very little if
any background was recorded in both channels. These observations validate the two-colour
labelling and imaging protocol.
At this diffraction-limited scale, a good overall co-localisation is observed between both
170 CD28 super-agonism
Figure 7.12 Bulk co-localisation of CD28 and anti-CD28 mAbs. The signalling-incompetent CD28-mEos
construct was stably expressed in TCR-deficient T cells, incubated for 15’ at 37◦C with either a SA, a CA or no
antibody. 20% of the activation mAbs were labelled with Atto655. Cells were washed to remove any labelled
but unbound mAb and resuspended in the same incubation buffer containing only unlabelled mAbs. Both
fluorophores were simultaneously imaged in two spectrally separated channels. A white-light transmission of
a representative cell at the end of the imaging (first column), the average projection of each channel (second
and third columns) and their superposition (last column) are given for all three incubation conditions and for
an untransfected cell incubated with no mAb as a negative control (‘WT’). Arrows: examples of co-localised
clusters.
7.2 Results 171
channels (arrows for specific characteristics, and ‘Combined’ column). The non-trivial reg-
istration of the channels, the movement of the live cells and the slight difference of illumi-
nations makes the projection of both channels not identical. But many features and clusters
are definitely co-localised.
Figure 7.13 Super-resolved co-localisation of CD28 and anti-CD28 mAbs. The data from the experiments
described in Figure 7.12 were analysed for super-resolution: for each cell, the PSFs of each channel were
fitted. Localisation positions were corrected for channel registration and superposed in a combined super-
resolved figure where each localisation is represented as a 2D-Gaussian of similar intensity and of width
equal to the average localisation precision (green for CD28-mEos, red for mAb-Atto655). Top row corre-
sponds to the representative T cell activated by the SA in Figure 7.12; bottom row, to the one activated by
the CA. A,F: Super-resolved figures. B,D,G,I: Zoom of the indicated region of the super-resolved figure.
C,E,H,J: Coordinate-based co-localisation representation of the green channel of the corresponding super-
resolved zoomed region with respect to the red channel. Each localisation is plotted as a pixel which colour
codes its CBC value (calculated with Rmax = 200 nm, ranging from -1−blue−, to 0−black−, to 1−yellow−).
Each channel was processed for SMLM, the list of localisations was corrected with
172 CD28 super-agonism
channel registration, and a corresponding super-resolved picture was rebuilt (Figure 7.13).
Consistent with faster diffusion, clusters in the case of CA activation (bottom) are on average
larger and less circular than for SA activation (top). For both activations, however, clusters
of CD28 are co-localised with clusters of mAbs. To quantify the co-localisation and avoid
biases introduced by image processing steps and intensity thresholding, a coordinate-based
co-localisation analysis was performed [116] (right column, CD28 with respect to CD45):
for each localisation detected, a score is calculated, ranging from -1 (blue) to 0 (black), to
1 (yellow), meaning that the localisation is respectively anti-correlated, not correlated, or
correlated with the localisations from the other channel.
Most CD28 clusters show some co-localisation with either mAb. However, partial la-
belling (20% of the antibodies only are labelled, and not all CD28 molecules are imaged
as only a proportion of all mEos proteins are correctly folded and can be activated [126]),
and the limited number of frames recorded restrict the sampling required for detection of
total co-localisation. Furthermore, cells are imaged live. Imaging is only started when
cells have settled on the coverslip to limit their movement. However, some membrane- and
intracellular reorganisation happen during the course of the experiment, and prevent per-
fect co-localisation. It can also explain small cluster displacements, that are observed even
after non-linear achromatism correction between channels, within and between channels,
depending on when the localisations of the cluster were imaged during the experiment.
7.2.5 Variation of the height separating the plasma membrane from
the coated surface
To test the KS model, and specifically its size-dependency hypothesis, a taller CD28 protein
was engineered: a∼ 75 Å-high Fc fragment of a human IgG1 was genetically fused between
the mouse transmembrane domain of CD28 and its rat ectodomain. In functional assays,
this addition blocks the SA-induced IL-2 production and T cell activation (Figure 7.14, B).
Remarkably, incubating the cells with a putative SA antibody which targets the ‘side’ of the
inserted Fc domain (S.J. Davis, personal communication), close to the plasma membrane,
allows the recovery of IL-2 production, confirming the importance of the relative proximity
of the targeted epitope of CD28 and the plasma membrane.
A similar experiment was performed in TCR− cells with a signalling-incompetent trun-
cated version of the Fc-CD28 construct in order to observe the effect of the different acti-
vations on CD28 diffusivity (Figure 7.14). Incubation with anti-CD28 SA or CA induced
very similar diffusion effects to the previous truncated CD28 construct. Compared to the
7.2 Results 173
Figure 7.14 Variation of the height separating the plasma membrane from the coated surface. A ∼ 75 Å-tall
human Fc domain was added between the transmembrane and the ectodomain of the signalling-incompetent
CD28-mEos construct to increase the distance of the epitopes of the SA and the CA to the plasma membrane.
The effect of a putative SA targeting the ‘side’ of the added Fc domain on the diffusion of the new construct
was compared to effect of the original anti-CD28 mAbs. A: Scheme of the experiment with the new construct
Fc-CD28, drawn for the original anti-CD28 SA and the anti-Fc putative SA. B: IL-2 production, measured by
M. Aßmann for the original anti-CD28 SA for both original and new constructs, and for the anti-Fc putative
SA for the new construct. C-E: Diffusion coefficients of the fast and slower diffusing population and their
relative fraction for the four incubation conditions. Same data representation as for Figure 7.6.
174 CD28 super-agonism
original signalling-incompetent CD28-mEos construct, the fast population of the taller Fc-
CD28-mEos is diffusing slightly faster when CA-activated, and slightly slower when SA-
activated, but those variations of ∼0.01µm/s2 are not significant. Even more strikingly, the
application of the putative anti-Fc SA reduced the diffusion of Fc-CD28-mEos to a similar
level as the anti-CD28 CA. These results suggest that slowing down of CD28 diffusion, on
its own, is not sufficient to induce IL-2 release.
7.2.6 Evidence for a size-dependent segregation?
In view of the KS model, the specific effect of SA incubation on CD28 diffusion could
play a role in T cell triggering (through the kinetic arm of the model, by locally increasing
the time CD28 stays in a contact zone). To test the other requirement of the KS model,
the distribution of CD28 was observed simultaneously with the distribution of CD45, a
large transmembrane phosphatase which segregation from the contact zones in T cells is
necessary for T cell triggering (Chang, Fernandes, Ganzinger, Lee et al., submitted). Small
anti-CD45 Fabs were labelled with Atto655 dyes and added to the incubation antibodies.
Cells were washed from any unbound Fab before being resuspended in their activation buffer
and imaged using a sptPALM-dSTORM approach.
The combined average of both channels (Figure 7.15, left column) reveal some co-
localisation of CD45 and CD28 when neither SA nor CA is added in solution, and an in-
creased segregation of both proteins when SA or CA are used. The super-resolved figures
underline a more complex situation (middle and right columns): in the case of incubation
in the absence of any mAb, a uniform distribution of both CD28 and CD45 is observed,
consistent with the previous results of random diffusion of CD28. Also consistent with re-
sults described above, the membrane of the cell in contact with the coated coverslip presents
CD28 clusters when the cells are incubated with SA or CA. Two observations are noticeable
about the distribution of CD45: (1) in comparison with the uniform distribution of CD45 in
the absence of antibody (bottom row), some clusters of CD45 appear on the membrane of
the T cells in the presence of SA or CA. (2) the distribution of CD45 is less uniform in the
presence of SA (top row) than in presence of CA (middle row).
In order to compare the different conditions in a quantitative manner, the coordinate-
based co-localisation (CBC− see Section 3.4.2) analysis of more than five cells with respect
to either channel was performed (Figure 7.15, right column, bottom close-ups, CBC image
with respect to the more homogeneous CD45 channel). For each cell, the median CBC
value was analysed (Figure 7.16). Although no significant difference can be observed in
7.2 Results 175
Figure 7.15 Simultaneous imaging of CD28 and CD45. Live TCR-deficient T cells expressing the signalling-
incompetent CD28-mEos construct were incubated with anti-CD45 Fab fragments labelled with Atto655 and
either the SA (A-D), the CA (E-H) or no antibody (I-L). Unbound Fab fragments were washed off and cells
were resuspended in their incubation buffer and dropped on a coverslip coated with secondary polyclonal
antibodies. For each incubation condition, a representative cell is shown. A,E,I: Combined average projection
of each channel (CD28-mEos in green, anti-CD45 Fab-Atto655 in red). B,F,J: Combined channel-corrected
super-resolved figure (each localisation is shown as a 2D-Gaussian of unit intensity and of width equal to the
average localisation precision). C,G,K: Zoom of the region marked as a white rectangle. D,H,L: Coordinate-
based co-localisation (CBC) representation of the red channel of the corresponding super-resolved zoomed
region with respect to the green channel. Each localisation is plotted as a pixel whose colour codes its CBC
value (calculated with Rmax = 200 nm, ranging from -1 −blue−, to 0 −black−, to 1 −yellow−).
176 CD28 super-agonism
Figure 7.16 Quantitative coordinate-based co-localisation analysis. For each incubation condition (SA, CA
or no antibody), five live cells were imaged. For each live cell, the CBC values of the localisations of each
channel were calculated with respect to the other channel. The median CBC values of the five cells are plotted
(mean of the distribution ± one standard deviation of the mean) with respect to CD28 (left panel) and with
respect to CD45 (right panel).
7.3 Discussion 177
co-localisation with respect to either channel between the conditions, a slight trend can be
hypothesised: CD28 and CD45 are less co-localised when cells are incubated with SA than
in the presence of CA.
These results suggest that there might be some SA-induced segregation of CD45. The
difference is, however, not significant and more evidence is required to confirm this conclu-
sion.
7.3 Discussion
The results presented in this chapter reveal that the SA specifically slows down the diffusion
of CD28 on the membrane of a live T cell immobilised on a surface. A critical description
of the hypotheses of the diffusion model is essential to understand the limits of the conclu-
sions that can be drawn from these data, and the additional questions they raise. Can this
phenomenon be explained on the mechanistic scale by a difference of avidity, affinity or
kinetics of the mAbs? Is this correlation between a decrease in diffusion and activation of T
cells causal? What does it reveal about CD28 activation?
7.3.1 Clusters of localisations: protein clusters or trajectories?
Clusters of localisations were observed when incubating the T cells with different anti-
CD28 antibodies: the clustering was more pronounced in the case of an SA compared to a
CA activation (Figure 7.4). This observation can originate from two, non-exclusive, types of
protein behaviour: (1) clusters of localisations can reflect a clustered distribution of proteins;
(2) it can also reflect the transient trace of a single fluorophore over a few frames, which,
if fixed or diffusing slowly, will be observed as multiple localisations clustered together in
time and space.
To distinguish between both scenarii, and because single CD28 molecules were observed
in consecutive frames, the localisations were grouped into trajectories (if closer than 300 nm
from each other within up to four frames). If the clustering of localisations was only due
to diffusion (hypothesis 2), the distribution of trajectories (represented by their mean posi-
tion in time) should not differ between SA and CA activation. Interestingly, although the
distribution of localisations is more uniform when the mean positions of the trajectories are
plotted, clusters do not disappear (Figure 7.17, A-F), indicating that clusters of localisations
are induced by both protein clustering and diffusion effects.
To further check that the grouping algorithm does not artificially group trajectories of
178 CD28 super-agonism
Figure 7.17 Differentiating protein clusters from single particle trajectories. Trajectories of representative
cells (corresponding to Figure 7.4) transfected with a full length CD28-mEos construct and incubated with
either SA (A-B), CA (C-D) or no mAb (E-F), were further analysed. A,C,E: Each trajectory grouping more
than two localisations is plotted as a single pixel at its time-average position. The colour of the pixel code for
the time of first appearance of the trajectory during the experiment (from 0 −blue−, to 3’20" −red−). B,D,F:
Zoom of the region of the cell indicated by a white rectangle. G: Distribution of the number of localisations
per trajectory (also referred as their ‘length’). The total number of trajectories detected for each cell is given
in parentheses in the legend.
7.3 Discussion 179
physically different molecules, the distributions of the trajectory lengths were also com-
pared. For both antibody activations, the distributions were almost identical (Figure 7.17,
G). In the absence of mAb activation, lengths of the trajectories were much smaller, due
to the quicker diffusion of the proteins which cannot be followed if they diffuse more than
300 nm between consecutive frames. Finally, for each cell, the distribution of displacements
were plotted and a good fit with a log-normal function was obtained (R2 = 0.98, 0.99 and
0.92 for SA, CA and no mAb activation, resp.), further confirming the correct grouping of
localisations in trajectories [154].
7.3.2 Choice of the diffusion analysis
Different methods have been proposed to study the diffusivity of single proteins, all based
on relating the squared displacement of the trajectories with the diffusion constant of a mo-
lecule undergoing Brownian motion (Equation 7.2). The mean-square displacement (MSD)
method calculates a diffusion constant for each single trajectory [291, 292]. This linear re-
gression method proves to be very accurate and efficient when long trajectories are available,
especially for multi-tagged or long-lived fluorescent particles (e.g. intracellular trafficking
vesicles or quantum dots). However, for small trajectories with few localisations, such as
single-molecule trajectories, the regression is done on much fewer data points and is conse-
quently less accurate and consistent [293].
For such trajectories, a jump-distance (JD) analysis was proposed: all square displace-
ments from consecutive frames (or ‘jump distances’) from all trajectories are binned to-
gether and the overall probability of square displacements is fitted with a model of multiple
populations or Brownian-diffusing proteins (Equation 7.1) [295, 296]. This analysis does
not directly allow to calculate a diffusion coefficient for each trajectory, but it can unravel
transitory changes of diffusion behaviours within all trajectories, since all square displace-
ments are binned regardless of their belonging to a specific trajectory. Also, no data is lost,
since even for the smallest trajectory a square displacement can be calculated [293]. Nev-
ertheless, this method only takes into account displacements between consecutive frames,
thus accumulating errors due to the time resolution of the acquisition: the localisation in
each frame basically represents the average position of the diffusing molecule during the
exposure time of the camera.
A more complete method, based on the analysis of the distribution of square displace-
ments for different times, generalises the JD analysis [294]. In this method, a JD analysis is
performed for square displacements from multiple time windows between frames. Instead
180 CD28 super-agonism
of calculating the mean square displacement for each trajectory, this method proposes to
calculate the mean square displacement SDi for each population i detected in the JD analy-
ses. This analysis thus cumulates the efficiency of JD analysis for small trajectories and the
robustness and precision of the MSD analysis, and was therefore chosen for the diffusion
analysis in this chapter.
For each analysis, five fields of view containing at least one cell were recorded for
4,000 frames (200 s). In order to measure the quality of the analysis (involving two fit-
ting steps), which depends on the number and length of the trajectories, the analysis could
have been repeated on each field of view or each cell. This would rather reveal the cell-
to-cell variability, due to differences in CD28 expression, cell landing, etc. Conversely, all
cells were analysed together and a bootstrap approach was preferred. Because in the JD
analysis all displacements are considered independently from each other and from the tra-
jectories they belong to, the distribution of all squared displacements was randomly divided
in five distributions that were separately analysed. The distribution of the extracted results
is plotted as a box plot (25 and 75 percentiles, and mean ± one standard deviation of the
mean) as in Figure 7.6 and similar.
This ‘bootstrap’ analysis can be seen as a means to computationally homogenise the five
cells. It allows to test the robustness of the fit (which depends on the size of the dataset) and
possibly extract statistical tests. Finally, the whole distribution was also fitted (◦) as it con-
tains the whole dataset and is thus more precise than the mean of the bootstrap distribution.
7.3.3 Physical origin of the multiple diffusing populations
The diffusion analysis separates populations with different diffusion behaviour. The mean-
ing of these populations needs to be carefully discussed. First, the assumptions of the two
fitting models used in the analysis should be recalled:
1. All proteins are freely diffusing, following a 2D Brownian motion (Equation 7.2). Bi-
ologically, this might not be the case, as CD28 could be secluded into micro-clusters,
lipid rafts or membrane regions related to the close cell cortex structure. It is, how-
ever, a good model in general for short displacements compared to the diameter of the
seclusion area. Indeed, for short time scales (τ ≤ 200 ms), the square displacements
and the time-lapse τ are well described by a linear correlation (Figure 7.5).
2. Only a discrete number of homogeneous populations of diffusing CD28 were fit-
ted (Equation 7.1). Biologically a large continuum of diffusing CD28 populations
7.3 Discussion 181
is rather expected since the antibody-CD28 complexes may form long linear struc-
tures, referred as to membrane-bound immune complexes (MBICs). These structures
are expected to present a large variation of sizes.
The most parsimonious model fitting the data was a model of three populations. For
some experiments, though, the populations were highly correlated (e.g. Figure 7.9 or Fig-
ure 7.14) and a model of two populations would be more parsimonious to describe the data.
For the sake of comparison, the model of three populations was consistently used for all
experiments in this chapter.
The third ‘fixed’ population (SD3(τ)≤ σ ) is always observed in each cell line and for all
tested constructs (CD28 as well as CD86 fused to mEos). This is interpreted as an artefact
of the imaging protocol, probably due to cell debris or fluorescent dust. This population is
therefore fitted but disregarded in all results. The two other populations can be interpreted
in two complementary ways:
1. The diffusion analysis, indifferently on the validity of its hypotheses, is a theoretical
mathematical means to compare different activation conditions in a quantitative, sta-
tistical and parsimonious manner. The populations are in this sense a way to describe
the data, statistically compare results, and draw conclusions about trends.
2. A second interpretation gives more biological insights on the putative mechanism at
play, but is highly speculative and subject to criticism, It would confer some biological
meaning to the division into two diffusing populations. For example, the two popu-
lations could refer to antibody-bound (D2) and antibody-unbound (D1) CD28 sub-
populations. A more plausible interpretation would consider CD28 proteins bound to
an MBIC that is itself bound to a secondary antibody coating the coverslip (D2) or not
(D1). Since the SA and the CA bind similarly to the secondary antibodies, the number
of MBICs bound to the coverslip (D2) is similar for both, but the sizes or interactions
of the MBICs could be different between both mAbs and this would only significantly
affect the diffusion of the unbound MBICs (D1).
A CD28-protein that belongs to an MBIC bound to an immobilised secondary anti-
body can still move around the antibody fixed on the coverslip: the linear MBIC chain
can be seen as an ideal polymer chain with one fix point. On scales smaller than
√
n · l
(n being the number of antibodies linking the CD28 to the fix point, and l the size of
a CD28-antibody element) [305], the CD28 protein, although part of a bound MBIC
complex, still diffuses. This shows that the CD28 population bound to bound MBICs
are not fixed and thus do not belong to the third population (D3).
182 CD28 super-agonism
Although appealing, there is not enough biological information available to support this
latter interpretation. Also, not all of the above results are consistent with it: e.g. in the
cross-linking experiments (Figure 7.9), MBICs only form when both antibodies are added
to the cells. No second population should then be observed (only diffusive CD28 proteins
or CD28-antibody complexes −D1− and fixed CD28-antibodies complexes −D3). The
absence of the third population (D2) is probably why results are highly correlated and a
model of two population would have described the data better. However, in the case of
the dual-antibody activation, large MBICs are restored and three independent populations
should then be observed, instead of the observed correlated diffusion coefficients.
This is why, although interesting, no final biological interpretation of the CD28 popu-
lations is given in this chapter. The diffusion results are rather used as a quantitative way
to compare and analyse trends. More biological knowledge is required to draw any solid
conclusion about the nature of the diffusive populations.
However, although little can be said about CD28 populations, and since CD28 expres-
sion is not trivial to vary, the concentration or immobilisation of the mAbs was varied to
gain more knowledge and evidence about the molecular mechanism of T cell SA-mediated
activation.
7.3.4 Differences of affinity and kinetic rates
Although unlikely between two classes of antibodies defined by a binary behaviour, it is not
impossible that together with the difference of epitope positions, a difference of thermody-
namic or kinetic constants characterises SAs compared to CAs. Therefore, M. Aßmann is
purifying the Fab fragments of both SA and CA antibodies used in this study and is deter-
mining their precise KD by Biacore surface plasma resonance.
Since the KD is equal to the ratio of the dissociating ko ff to the association kon kinetic
rates, an approximative comparison of the thermodynamic constants can be estimated from
flow cytometry experiments. And because dissociating rates are relatively similar for both
mAbs studied here (flow cytometry experiments by M. Aßmann and ref. [278]), the compar-
ison can be restricted to the association rates kon. Furthermore, although the thermodynamic
binding equilibrium is reached in the IL-2 experiments − after 48 h of mAb incubation −
an estimation of the association rate of the SA indicates that only ∼ 2/3 of the final bound
fraction of SAs are bound after the 15’ incubation before imaging occurs in the sptPALM
experiments described in this chapter [278]. The CA, however, has reached its thermody-
namics binding equilibrium after 15’.
7.3 Discussion 183
The dilution experiments (Figure 7.10) can be studied as precipitin reaction curves
[306]: as the concentration of mAb increases, larger MBICs are formed, slowing down their
diffusion, until mAbs are introduced in excess and favourably bind monovalently CD28
proteins, decreasing the size of the MBICs and increasing their diffusivity. At equivalence
(when the fraction of antibodies that is cross-linked is maximal) and at equilibrium, Dembo
et Goldstein demonstrate that the mAb concentration is equal to half of the monovalent KD
[307].
Since the thermodynamic equilibrium is reached for CA binding in the diffusion exper-
iment, from Figure 7.10.E the equivalence was extracted (∼ 1 µg/mL) and the estimation
of the monovalent equilibrium constant was calculated: KCAD ∼ 3.5 ·10−9 mol/L. The ther-
modynamic equilibrium is not reached in the case of SA incubation for the diffusion ex-
periments, but it was for the IL-2 experiments, which had a 48 h incubation period. Thus,
similarly, the equivalence was extracted from Figure 7.10, A (∼10 µg/mL) and the dissoci-
ation constant was calculated: KSAD ∼ 3.5 ·10−8 mol/L = 0.1 KCAD .
These values are noticeably consistent with the observation of Tacke et al. that the SA
"had to be used at a 10-fold higher concentration [than the CA] to achieve a comparable
binding over time" [278]. This also confirms that after 15’ of incubation, the SA has not
reached its binding equilibrium: indeed, the equivalence is not observed at 10 µg/mL in
Figure 7.10, B.
In summary, the thermodynamic monovalent dissociation constants of the CA to CD28
is approximately an order of magnitude smaller than the dissociation constant of the SA
to CD28. This is driven by a 10-fold difference of their respective binding rates. The
precise values measured by surface plasma resonance should validate these results. Since
TCR− cell lines and signalling-incompetent CD28 proteins are used, cells are locked in
a pre-activation state. This, repeating the dilution experiments at later time-points when
SA-binding equilibrium is reached (e.g. after 2 h incubation) is possible and would also
cross-validate these conclusions.
Similar experiments with different pairs of SAs and CAs would be necessary to deter-
mine whether this difference in affinity is specific to the mAbs studied in this chapter or a
new characteristic defining both classes of antibodies. In this sense, Tacke et al. postulated
that the shared epitopes of SAs, closer to the plasma membrane, are less accessible, ex-
plaining the lower association rate of the SA, and thus its higher KD. But if the monovalent
affinity is different, does it affect cross-linking of CD28 by the mAbs? Does it affect the
avidity of the antibodies?
184 CD28 super-agonism
7.3.5 Requirements for cross-linking of CD28
Since both CD28 and mAbs are divalent, the mAbs can cross-link CD28 homo-dimeric
receptors and form long linear structures, also called membrane-bound immune complexes
(MBICs). Consistent with previous studies showing that only monovalent anti-CD28 mAbs
can blockade CD28 signalling [308], soluble SA Fabs were unable to induce IL-2 production
[309]. Using the opposite approach, bivalent SA were also unable to induce IL-2 production
by monomeric CD28-like receptors (Figure 7.9, B). Strikingly, T cell activation was strongly
recovered by cross-linking the monomeric CD28-like receptors with two different bivalent
antibodies and allowing the formation of MBICs.
As expected, similar results were observed in diffusion experiments (Figure 7.9): little
slowing down in the monomeric CD28-like receptors is induced by incubation with either
mAb alone, but a strong decrease in diffusivity appears when both antibodies are used in
combination. Interestingly though, even in the absence of any second-order cross-linking,
one antibody (‘CA 1’) significantly slows down CD28-like monomers more than the other
one (‘CA 2’).
Incidentally, the epitopes of both anti-PD-1 antibodies have been mapped (Figure 7.9,
F, courtesy of S.J. Davis). The epitope of ‘CA 1’ antibody is located near the plasma mem-
brane, on the ‘side’ of PD-1, whereas the epitope of ‘CA 2’ is on ‘top’ of PD-1. Therefore,
‘CA 1’ falls into the antibody class of putative SA. This classification is furthermore fully
consistent with the previous observation of a significant decrease of PD-1-CD28 diffusivity
compared to ‘CA 2’ putative CA. It also underlines that although cross-linking has the major
effect on slowing down CD28 diffusion, another independent phenomenon plays also some
part in it.
This observation confirms that although it may decrease the affinity of SAs, the specific
epitope location of SAs may also increase their avidity, playing a complex role in slow-
ing the diffusion of their ligands. Such a phenomenon was theoretically described by Å.
Larsson when studying divalent binding of mAb to a monomeric cell surface antigen [310].
Binding of a divalent mAb is described as a two-step reaction between (1) soluble diffusing
antibodies and monomeric receptors, and (2) monovalently bound antibodies anchored on
the membrane and monomeric receptors, thus separating the effects of monovalent affinity
and avidity.
Moreover, although not entirely convincing as the two Fab arms of an antibody are
considered to be independent because of the flexible hinge linking them to the Fc domain
[311, 312], Larsson suggests that if the monomeric receptor has a rigid structure, "the best
fit for the divalent binding is obviously obtained when the angle between the epitope and the
7.3 Discussion 185
membrane surface, together with the predominant Fab-Fab angle, amounts to 180◦." Prac-
tically, IgG1 antibodies such as those used in this chapter have a Fab-Fab angle estimated
to 30−180◦ [313]. For such antibodies, cross-linking would then be critically favoured for
antigens oriented with an angle to the plasma membrane of 0−75◦, i.e. not on ‘top’ of the
receptor (see Figure 3 in ref. [310]).
A stronger argument preserving the large distribution of Fab-Fab angles is that once
bound to one receptor, the plasma membrane sterically restricts the overall accessible Fab-
Fab angles. The restriction is more stringent when the epitope is located close to the plasma
membrane. Interestingly, the plasma membrane blocks Fab-Fab angles that do not allow
binding of another receptor. Thus, thermodynamically, the closer the epitope is to the plasma
membrane, the more cross-linking is favoured because the proportion of accessible states
that allow binding is increased. This would directly and mechanistically link epitope loca-
tion (and thus the SA class of antibody) to its effect on CD28 diffusion.
Finally, as not all transmembrane proteins (e.g. CD86-mEos in Figure 7.8, co-expressed
with CD28 proteins in cells activated with anti-CD28 mAbs) having an extracellular domain
see their diffusivity reduced in presence of the SA, cross-linking of CD28 receptors is not
forming a large mesh-work sterically limiting the diffusion of all transmembrane proteins
as in a fence and picket model of the plasma membrane [314], but rather separated ball-like
clusters as indeed visible in Figures 7.4 and 7.13.
7.3.6 Importance of the formation of a close-contact zone
As for antibody-induced TCR triggering, antibody-induced CD28 co-stimulation requires
immobilisation of the cross-linking antibody on a surface both in vitro and in vivo (M. Aß-
mann, personal communication and ref. [278, 282]). The Fc part of the activating antibody
also seems to play a role in T cell activation [308, 315]. Together, these conclusions suggest
a critical role for a zone of contact between the membrane of the T cell and the surface on
which anti-CD28 antibodies are immobilised.
Because SAs bind to epitopes on CD28 ∼ 75 Å closer to the plasma membrane than
CAs, it has been proposed that the distance separating the T cell from the activating surface
was critical for SA-mediated T cell activation [234]. To test this hypothesis, a new construct,
inserting a ∼ 75 Å-tall human Fc domain between the transmembrane domain and the ec-
toplasmic domain of the previous CD28 construct, was engineered. Thus, the previously
used anti-CD28 SA and CA target the same epitopes which are located a little further from
the T cell membrane. An anti-Fc mAb that binds an epitope on the ‘side’ of the added Fc
186 CD28 super-agonism
domain (S.J. Davis, personal communication) was also used as a putative SA that can bind
at a similar distance from the membrane than the anti-CD28 SA in the case of the original
CD28 construct (Figure 7.14, A).
Strikingly, as predicted in the proposed hypothesis, the addition of the 75 Å domain
depleted IL-2 production. Furthermore, targeting the base of the new construct with the
anti-Fc putative SA allowed the recovery of T cell activation (Figure 7.14, B). Interestingly,
although the additional domain increases CD28 diffusion from 0.025 to 0.04 µm2/s for anti-
CD28 SA incubation, the effect of incubation with the anti-Fc putative SA was comparable
to the effect of the CA and did not recover the slow diffusivity of the original CD28 construct
activated by the anti-CD28 SA (0.07≫ 0.025 µm2/s) (Figure 7.14, C-E). The original anti-
CD28 SA is no longer an SA for Fc-CD28 in terms of IL-2, but still looks like a SA for
Fc-CD28 in terms of diffusion.
This is the only experiment for which IL-2 and diffusion experiments do not strongly
correlate in this chapter. A first explanation is that the cell line used for the diffusion ex-
periments did not correctly express the new construct1. Repeating the experiment with
newly transfected cells should allow for testing this hypothesis. Otherwise, a critical dif-
ference in kinetic rates between the anti-Fc putative SA and the anti-CD28 SA could also
explain the difference observed: the thermodynamic equilibrium may be reached much more
slowly for the putative SA, and the diffusion experiment might only describe an early out-
of-equilibrium situation at the difference of the IL-2 experiments. Repeating the experiment
at a later time-point (e.g. after 2 h of incubation) will test this possibility.
7.3.7 From correlation to causality?
A very strong correlation between SA-induced IL-2 production and the slowing down of
CD28 receptors is revealed by almost all experiments in this chapter. An outlier is the exper-
iment for which the extracellular domain of CD28 is separated from the plasma membrane
by an extra ∼ 75 Å, as discussed in the previous section. Can any information be inferred
from such a correlation? What does it mean in terms of the mechanism of SA-mediated
leucocyte receptor triggering?
The previous discussions suggest a mechanism of a two-step binding reaction model
1Different cell lines are used: for the IL-2 experiment, BW5147 cells are simultaneously transfected with
CD3δ and CD3ζ chains to recover TCR expression and with the following construct: [rat ectodomain of
CD28]-[human Fc domain]-[mouse transmembrane and signalling domains of CD28]; for the diffusion ex-
periment, BW5147 cells are only transfected with the mEos-tagged and signalling-incompetent version of the
previous construct: [rat ectodomain of CD28]-[human Fc domain]-[mouse transmembrane domain of CD28]-
[30 amino acid linker]-[mEos3.2].
7.3 Discussion 187
linking epitope binding site to avidity and thus to cross-linking and CD28 diffusivity (Sec-
tion 7.3.5).
Nevertheless, such conclusions are mostly based on experiments done with a unique pair
of SA and CA. An important exception, however, concerns the cross-linking experiment
(Figure 7.9) where two putative SA and CA (respectively ‘CA 1’ and ‘CA 2’) are used.
Noticeably, this independent pair of SA and CA shows all similar characteristics compared
to the original pair: (1) the putative SA is so-called because of the location of its epitope
on its ligand close to the plasma membrane, in opposition to the putative CA (Figure 7.9,
F); (2) the binding rate kon of the putative SA is 10-fold lower than that of the putative
CA, but have relatively similar ko ff (S.J. Davis, personal communication); (3) the putative
SA decreases the diffusivity of its receptor more than the putative CA (Figure 7.9, C-D); (4)
finally, it might also induce a slightly higher IL-2 production, although large cross-linking is
prevented, than the putative CA (Figure 7.9, B). Importantly, this suggests that the previous
binding model and conclusions could apply to the whole classes of SA and CA rather than
being specific for a pair of mAbs. Repeating similar experiments with other pairs of SA and
CA will challenge the possibility of such a generalisation.
This model mechanistically links epitope binding location with diffusion behaviour but
how CD28-mediated signalling occurs remains elusive. Although a correlation does not au-
tomatically implies a link of causality, it suggests a possibility of a shared common cause if
not a direct causal effect. Cross-linking of receptors is known to locally accumulate intracel-
lular domains of receptors that may directly or indirectly induce their cross-phosphorylation
and cross-activation (e.g. receptor tyrosine kinases, reviewed in ref. [316]). CD28, how-
ever, cannot directly cross-activate (i.e. trans-autophosphorylate) as it lacks intrinsic kinase
activity. CD28 depends on extrinsic kinases such as Lck or ZAP-70. If a CD28 receptor
recruits kinases, a local accumulation of CD28 would, however, favour the likelihood of
neighbouring receptors to become phosphorylated by these kinases. Lastly, cross-linking
alone does not account for the requirement of antibody immobilisation.
The KS model presents the advantage of offering an explanation for both a causal link
between a decrease in diffusivity and signalling (the receptor needs to stay long enough in
the close-contact zone to get activated), and the requirement for antibody immobilisation
(to form a close-contact zone that segregates activating and deactivating proteins). Indeed,
an accumulation of cross-linked CD28 could potentially create a large zone of close-contact
between the T cell and the APC (or the secondary antibody-coated surface). Transmembrane
molecules with large extodomains would be excluded from such close-contact zones. Such
a segregation would locally change the equilibrium between cytoplasmic (e.g. ZAP-70) or
188 CD28 super-agonism
small transmembrane (e.g. Lck) kinases and abundant transmembrane phosphatases with
large ectodomains (e.g. CD45 or CD148) [317], locally increasing the net phosphorylation
of the CD28 receptors immobilised in the close-contact zone, which would induce their
triggering.
The importance of the proximity of the epitope to the plasma membrane for cell acti-
vation (Figure 7.14, A-B) also suggests a mechanism sensitive to the height of the contact
between the T cell and the coated coverslip. Therefore, to test the mechanism proposed
by the KS model and specifically the existence of height-dependent segregation zones (cf.
Section 7.2.6), the localisations of both CD28 and CD45, a transmembrane non-specific
phosphatase with a large ectodomain and which plays a critical role in TCR triggering,
were studied and compared.
7.3.8 Possible evidence for segregation
A dual-colour PALM-dSTORM protocol was optimised to simultaneously observe both
CD28 and CD45, a transmembrane protein with a large rigid extracellular domain that is
segregated from close-contact zones in TCR triggering (Chang, Fernandes, Ganzinger, Lee
et al., submitted). The dual tagging of CD28 intra-cellularly by the photo-activable fluores-
cent protein mEos and extra-cellularly by anti-CD28 mAbs labelled with a far red dye served
as a positive control to test the possibilities and limits of the method. Partial co-localisation
of mEos and anti-CD28 mAbs was observed in both diffraction-limited and super-resolved
modes (Figure 7.12 and 7.13). Total co-localisation was prevented by multiple factors typ-
ical of single-molecule experiments: partial labelling, limited sampling, and movements
due to live-cell imaging (as revealed by the displacement of clusters of trajectories over
the acquisition time in Figure 7.17, B and D). However, although finding a protocol to im-
age two spectrally-separated independently-photo-switchable fluorophores simultaneously
at correct densities with buffer conditions compatible with live cell imaging was non-trivial
[107], the signal-to-noise ratio of single emitters and the quality of the recorded data posi-
tively validated the method.
Such drawbacks could account for the non-significantly different average co-localisation
of CD28 and CD45 observed after SA and CA incubation (Figure 7.15 and 7.16) [52, 107].
Nonetheless, the trend might suggest a decrease in co-localisation in the case of SA acti-
vation, which would be an evidence for SA-induced increase of segregation of CD45 from
CD28, as postulated by the KS model. Further optimisation of the protocol (e.g. modifi-
cation of the imaging buffer, longer imaging time), as well as investigating the possibility
7.3 Discussion 189
to fix T cells without disrupting the contact formed with the coverslip will allow to tackle
some of the drawbacks.
In parallel, functionally testing predictions of the KS model in a top-down approach (e.g.
testing the effect of CD45 proteins with a truncated ectodomain on both IL-2 production
and CD28 diffusion) and imaging of other components of the signalling cascade (e.g. Src
tyrosine kinases such as Lck, or the TCR) will allow a better understanding of mechanism
behind SA-mediated CD28 signalling. Indeed, the segregation of phosphatases with large
ectodomains, such as CD45, and small kinases, such as Lck, is a direct prediction of the
KS model. Furthermore, since the SA induces T cell activation in absence of signal 1,
in a ligand-independent but TCR-dependent manner (Figure 7.6, IL-2 production is only
observed in TCR-sufficient cells), the SA might induce ligand-independent TCR triggering.
Observation of the distribution, the diffusion and the possible segregation of the TCR on the
membrane of SA-decorated cells is therefore of great interest.
7.3.9 Temporal segregation in light of the kinetic-segregation model
In case of CD28 super-agonism T cell activation, the segregation of CD45 may turn out to
be on much smaller length scales than previously observed on supported-lipid bilayers and
on glass coated with a single layer of activating mAb (ref. [243] and Chang, Fernandes,
Ganzinger, Lee et al., submitted). In the experiments presented in this chapter, effectively
two randomly-oriented layers of antibodies separate the cell from the glass surface (the
activating mAb decorating the cell, and the secondary antibodies coating the glass). This
configuration may not favour the observation of clear segregation at length scales larger than
the precision of the SMLM imaging. Segregation of CD45 phosphatases from contact-zones
may only be very transient and may never be observable in the co-localisation experiments.
However, even transient CD45 segregation would result in an overall decrease of CD45
and CD28 diffusion rates, since the segregation would impose an additional barrier to dif-
fusion. Another way to look at it, is to consider the chance of encounter between CD28
and CD45: it can be decreased by reducing the number of CD45 proteins (segregation side
of the KS model) or (independently and non-exclusively) by reducing their relative speed
(kinetic side) (Figure 7.18). Measuring the variation of CD45 diffusion rates between CA
and SA activations would enable the estimation of such an effect.
Thus, if CD45-CD28 encounter rate was indeed decreased in the case of the SA, the
dephosphorylation rate of CD28 would decrease whereas the phosphorylation rate would
be expected to remain constant (since intracellular Lck should be unaffected, since Lck has
190 CD28 super-agonism
Figure 7.18 Temporal segregation of membrane receptors. Illustration of the effect of diffusion decrease of
two receptors on their encounter rates. A: Slow diffusion of CD45 (red) and CD28 (green) decreases the
rates of encounter of both species which need much more time (varying with the square of the distance) to
explore the same region. Thus this temporal segregation, in light of the KS model, would induce long-lived
phosphorylation of CD28. B: Conversely, fast diffusion observed on non antibody-activated cells induces
frequent encounters between both CD45 and CD28 proteins. Thus, only short-lived phosphorylation could be
observed. C: Temporal segregation induces the same effect as conventional spatial segregation. If CD45 are
prevented to diffuse in a close-contact zone (orange circle), encounters between both CD45 and CD28 will
be prevented in this zone, for any diffusion rate. Thus, long-lived phosphorylation, as postulated by the KS
model, will be observed.
7.3 Discussion 191
no ectoplasmic domain). Thus, in line with the KS model, the kinase-phosphatase balance
would shift; therefore the number of CD28 molecules that remain phosphorylated for a time
sufficient to activate the cell would increase.
Seen in this light, the significantly and reproducibly increased confinement, or decreased
diffusion rates, of CD28 for the SA compared to the CA are very interesting, since a slowing-
down of both ‘partners’ (CD28, CD45) may effectively amplify their segregation. Repeating
the final co-localisation experiment (Figure 7.15) at labelling densities and laser powers that
allow simultaneous single-particle tracking of both CD28 and CD45 would enable to test
this hypothesis. Also, as this last experiment may be technically challenging, observing
CD45 diffusion alone, after CA or SA activation, on a single-molecule level or by flu-
orescence recovery after photo-bleaching will be sufficient to initially validate a specific
decrease of CD45 diffusion by the SA.
7.3.10 Conclusions and future experiments
Results of this chapter indicate a strong correlation between SA-mediated T-cell activation
and lower diffusivity of CD28 proteins. They also suggest a two-step binding reaction of
anti-CD28 mAbs to CD28 homo-dimers: (1) soluble mAbs first bind monovalently to a
CD28 receptor, the rate of the reaction depending on the proximity of the epitope on CD28
to the plasma membrane (lower rates for membrane-proximal epitopes); (2) once bound to
one receptor, mAbs bind to a second receptor, the binding constant of the reaction decreasing
with the proximity of the epitope to the plasma membrane for geometric and thermodynamic
reasons. Experiments need to be repeated at later time-points and with other pairs of SA and
CA mAbs to validate the generalisation of this working model to both classes of antibodies.
Although this model suggests the importance of differential cross-linking for super-
agonism, it neither accounts for the SA-specific slowing down of CD28 in the absence of
cross-linking (Section 7.2.2) nor for the requirement for antibody immobilisation and the
possible importance of the size of the gap separating the plasma membrane from the con-
tacted surface. From the four originally proposed models, only the KS model accounts for
these observations. Some initial but non-significant evidence for size-dependent segrega-
tion of transmembrane proteins is in line with the predictions of the KS model. Repeating
the Fc-CD28 diffusion experiment is critical to understand the possible link between both
differential cross-linking and KS models.
Additional imaging of other components of the signalling cascade of CD28 and func-
tional experiments further challenging the KS model are necessary to confirm these re-
192 CD28 super-agonism
sults and elaborate the missing steps bridging CD28 cross-linking and CD28 signalling.
Finally, in a long-term perspective, similar experiments would need to be repeated in a more
biologically-relevant system of cell-cell interaction such as a supported lipid bilayer func-
tionalised with Fc receptors [318].
PART C

Chapter 8
Materials and methods
8.1 Materials and methods common to all projects
8.1.1 Microscope set-up
Two microscope set-ups of similar design were used in the experiments presented here (cf.
Figure 3.1). The set-up used in Chapters 3, 4 and 6 is described here. Where both set-ups
differ, specificities of the microscope used in Chapter 7 are specified in brackets.
Collimated 640 nm, 561 nm , 488 nm, and 405 nm laser beams were spectrally filtered,
aligned and focussed at the back aperture of an Olympus UPlanApo 1.49 na 60x oil TIRF
objective mounted on an IX71 [IX73] Olympus inverted microscope frame (cf. Table 8.1
for the specific lasers and filters). The power of the collimated beams at the back aperture
of the microscope were typically of 20, 50, 5, <0.01 mW, resp., corresponding to power
densities of 40-400 W/cm2 (<60 mW/cm2 for the 405 nm activation line) at the sample
(cf. Table 8.1). The lens objective was regularly cleaned off oil left-overs thanks to three
organic solvents: methanol (Sigma Aldrich, 154903), m-xylene (Sigma Aldrich, 95672),
and methanol (Sigma Aldrich, 154598).
The fluorescent signal was separated from the illumination beams by a four-band di-
chroic (Semrock, Di01-R405/488/561/635) and filtered with either a 488 nm long-pass fil-
ter (Semrock, BLP01-488R, ‘green channel’), a combination of a 561 nm long-pass (Sem-
rock, BLP01-561R [BLP02-561R]) and a band-pass filters (Semrock, FF01-593/40 for yeast
imaging; Semrock, FF01-587/35 for mammalian cell imaging) (‘red channel’), or a 640 nm
long-pass filter (Semrock, FF01-642/LP, ‘far red channel’). The filtered signal was then
expanded through a 2.5x achromatic beam expander (Olympus, PE 2.5x 125) and finally
projected on an EMCCD camera (Photometrics, Evolve 512 [Evolve Delta]).
196 Materials and methods
Table 8.1 Optical characteristics of the microscope set-ups. Characteristics of the set-up used in Chapters 3,
4 and 6 are given. Where characteristics differ, specificities of the microscope used in Chapter 7 are specified
in brackets. All filters are from Semrock.
640 nm line 561 nm line 488 nm line 405 nm line
Laser
Coherent,
Cube 640-100C
Cobolt, Jive 200 Toptica, iBeam
Smart 488 100 mW
Oxxius,
LaserBoxx 405
[Toptica, iBeam
Smart CD_11005]
[Cobolt, Jive 500] [Cobolt, MLD 405]
Clean-up filter
FF01-640/14 FF01-561/14 LL01-488 FF01-417/60
[LD01-640/8] [LL02-561] [LL01-488] [FF01-405/10]
Beam combiner
N/A FF605-Di02 FF552-Di02 FF458-Di02
[FF705-Di01] [Di02-R561] [Di02-R488] [Di02-R405]
Typical power at
20 mW 50 mW 5 mW <10 µWthe back aperture
Beam diameter ∼10 mm ∼7 mm ∼6.5 mm ∼8 mmafter collimation
Power density ∼100 W/cm2 ∼400 W/cm2 ∼40 W/cm2 <60 mW/cm2at the sample
8.1 Materials and methods common to all projects 197
For simultaneous two-colour imaging, the signal was only filtered with the 561 nm long-
pass filter. A dualview was inserted in the optical path between the beam expander and the
camera to split the collected signal in two channels projected on two halves of the cam-
era (Photometrics, DV2 mounted with a Semrock, FF635-Di01 dichroic). The short-pass
channel was further filtered with a 580 nm band-pass filter (Semrock, FF01-587/35), and
the long-pass channel with both the 640 nm long-pass filter and an additional 700 nm band-
pass filter (Semrock, FF01-692/40). Before imaging, the dualview was aligned thanks and
multi-colour beads were simultaneously imaged in both channels for their registration (cf.
Section 3.2.3).
8.1.2 General imaging protocol
Samples were directly imaged on square glass coverslips (VWR International, borosilicate
glass thickness #1, 631-0122). Before any use, auto-fluorescence from the coverslips were
eliminated by their washing in an argon plasma for >30’ (Harrick Plasma, PDC-002). Sam-
ples were resuspended and washed in buffers filtered twice with 200-nm filters (Millex GP,
PR04064) to limit any fluorescence background in the buffer.
Samples were typically imaged for 2,000-4,000 frames of 30-50 ms exposure under
TIRF illumination. Automated sequences written as plug-ins or macros for µManager [173]
were programmed to ensure unbiased and consistent imaging protocols. The quality and
reproducibility of the TIRF illumination was qualitatively monitored by the angle of the
reflected total-internal reflection beam at the back aperture of the microscope.
Two modes of 405 nm activation were used for switching mEos fluorescent proteins on:
Continuous constant activation , used by default. The intensity of the 405 nm illumina-
tion was chosen on a first adjustment sample such as PSFs were separated in each
frame, before preparing another sample to take the measurements.
Continuous Fermi activation , in order to optimise the number of PSFs appearing in each
frame along the whole imaging time (cf. Section 3.3.4). This type of activation was
used in all the experiments of Chapter 6 and when indicated as such in the other
chapters. The maximum intensity of the 405 nm illumination was chosen as for the
constant activation, although a denser PSF coverage per frame were aimed at.
Recorded SMLM movies were analysed with Peak Fit [149] using the parameters ac-
cording to Table A.3. Further post-processing algorithms are described in Section 3.4 and
7.2.1.
198 Materials and methods
8.1.3 SMLM data representation
From the final list of localisations was rebuilt a super-resolved image. Four different repre-
sentations are used in this work:
Same amplitude and width. Each localisation is rebuilt as a 2D-Gaussian of same inten-
sity (arbitrary) and same width (the theoretical precision of the experiment, i.e. the
value at the peak of the distribution of theoretical precisions of all localisations as
defined in Section 3.4.3). This representation is the default one and was used in Fig-
ure 4.9, all figures from Chapter 6 except Figure 6.16, and Figure 7.4.
Individual amplitude and width. Again, each localisation is rebuilt as a 2D-Gaussian, but
both intensity and width are specific to each localisation: the intensity of the 2D-
Gaussian is equal to the fitted intensity of the localisation, and the width to the the-
oretical precision of the estimation of its position (cf. Section 3.4.3). Figures 3.13
(G-H) and 4.8 were rebuilt following this protocol.
Time of first appearance. Localisations appear at a given frame during the SMLM ac-
quisition. To distinguish between clustering due to transient photo-blinking of the
fluorophore and real clustering of multiple aggregated fluorophores, each localisation
can be plotted as a small circle which colour codes for its frame of appearance: thus,
uni-colour clusters indicate photo-blinking, and multi-colour ones suggest aggrega-
tion. If localisations are grouped, the time of first appearance, i.e. the earlier frame
of the localisations belonging to the group, is used instead of the frame. This was the
representation chosen for Figure 7.17.
Trajectories. When fluorophores are tracked and localisations grouped in trajectories, a
mean-square displacement (MSD) analysis allows to estimate diffusion coefficients
for each trajectory [291, 292]. The MSD of the long trajectories (>10 frames) are
computed and the five first time-points are used for linear regression and to estimate
the diffusion coefficient. Trajectories are then plotted as lines joining the consecutive
positions of the localisations belonging to the specific trajectory. The colour of the
trajectory codes for the estimated diffusion coefficient (on a log-scale value). This
representation is used in Figure 7.4. However, in Figure 3.4, the colour of each sec-
tion of the rebuilt trajectories codes for the frame of appearance of the localisations
during the acquisition (orientating the trajectory time-wise). Also, in Figure 3.12 and
Movie S1, each trajectory is defined by a different colour independently on its mobil-
ity.
8.2 Materials and methods for Chapter 3 199
For two-colour co-localisation observations, two complementary representations were
implemented and used in Figures 6.16, 7.13 and 7.15:
Superposition of the super-resolved images. As in two-colour conventional fluorescence
microscopy, the data from each channel of the dualview can be independently anal-
ysed. After registration of the lists of localisations to correct for channel mis-align-
ment (cf. Section 3.2.3), both super-resolved images can be rebuilt by one of the
methods previously described and superposed using two different colours (e.g. green
for the short-pass channel, and red for the long-pass one).
Coordinate-based co-localisation images. Similarly, after registration of the lists of lo-
calisations, a coordinate-based co-localisation (CBC) can be computed (cf. Sec-
tion 3.4.2): for each localisation of each channel, a CBC value between -1 (anti-
correlation) and +1 (perfect co-localisation) is calculated and defines the degree of
co-localisation of the localisation with the other channel. These values are then used
to rebuild two images (one for each channel) where each localisation of one channel
is plotted as a small circle which colour codes for its CBC value.
8.1.4 Imaging fluorescent beads and isolated fluorophores
The stock of beads or fluorophore of interest (cf. Table 8.2) were sonicated for 30 s and
diluted 1/100 in MilliQ water. The diluted particles were then sonicated for 2-10’ and
diluted again 1/1,000 in MilliQ water. An imaging hydrophobic gasket (Bio-rad, SLF0201)
was applied on each coverslip. The coverslip was then coated with 100 µL of fresh sterile
poly-L-lysine (Sigma, P4832) for 15’ at room temperature, before being washed three times
with 100 µL of MilliQ water. 50 µL of the diluted particles were dropped on the coverslip
for 10’ at room temperature, washed three times with 100 µL of MilliQ water and imaged
under TIRF illumination.
8.2 Materials and methods for Chapter 3
8.2.1 Double-helix microscope
Collimated 640 nm (Toptica, iBeam Smart 405-S_10924) and 405 nm (Toptica, iBeam
Smart 640-S_10925) laser beams were spectrally filtered (Semrock, FF01-640/14 and FF01-
405/10, resp.), combined (Semrock, dichroic mirror Di02-R405), aligned and focussed at
the back aperture of a Nikon Plan Apo 1.20 na 60x water objective mounted on a Nikon
200 Materials and methods
Table 8.2 Separated beads and fluorophores used in this thesis. †: high concentration (1/10 dilution of the
stock of purified proteins) were used to record bulk measurements.
Figure Separated fluorophores
Figure 3.3, 3.4, 3.6 and 3.15 Invitrogen, 0.1 µm TetraSpeck Microspheres (T-7279)
Figure 3.7 id.
Figure 3.8†, 3.9, 3.10 and 3.11 Purified mEos2 and mEos3.1
Figure 4.1, 4.2 (A), 4.3, 4.4,
4.5, 4.6 (A, 1st row), and 4.7
Molecular Probes, 0.04 µm FluoSpheres
NeutrAvidin-Labeled Microspheres (F-8771)
Figure 4.2 (B) Alexa647-labelled Tau monomer
Figure 4.6 (A, 2nd , 3rd , 5th rows) Invitrogen, 0.1 µm TetraSpeck Microspheres (T-7279)
Figure 4.6 (A, 4th row) Invitrogen, Qdot 655 ITK (Q21321MP)
Eclipse Ti inverted microscope frame. The power of the collimated beams at the back aper-
ture of the microscope were typically of 50 mW and <50 µW, resp., corresponding to power
densities of 600 W/cm2 and <600 mW/cm2, resp., at the sample. The lens objective was
regularly cleaned off oil left-overs as described in Section 8.1.
The beam of a second 640 nm laser (qioptiq, iFlex-2000) was focussed to create a thin
sheet of light to illuminate a plan of the sample perpendicular to the optical axis of the imag-
ing objective lens. The collimated laser beam was expanded up to a diameter of ∼10 mm.
A cylindrical lens (Thorlabs, LJ1567RM-A, f = 100 mm) was collimated with the back
aperture of the illumination objective (Mitutoyo, M Plan Apo 0.28 na 10x air objective) to
horizontally focus the laser beam in the focal object plane of the objective. A horizontal
light-sheet was consequently formed at the focal image plane of the illumination objective
which was aligned with the optical axis of the imaging objective.
Figure 8.1 3D-printed L-shaped plastic support to hold the sample and allow side light-sheet illumination.
8.2 Materials and methods for Chapter 3 201
To hold the sample, two cleaned glass coverslips (cf. Section 8.1) were glued on a
3D-printed L-shaped plastic support (Figure 8.1), perpendicularly to each other. To avoid
optical artefacts at the entry of the light-sheet in the solution, slight protrusion of the vertical
coverslip over the horizontal one at their joint was preferred to the opposite. Once set, the
sealing of the sample holder was checked by filling it with buffer. Finally, the buffer was
removed, the sample was added on the bottom coverslip and topped up with some buffer
to fill the height of the light-sheet at when entering in the solution. The geometry of this
home-made holder enables at the same time entry of the light-sheet in the buffer (through the
vertical coverslip) a few micrometers above the bottom horizontal coverslip and absorption
of the light-sheet when leaving the solution on the other side.
The fluorescent signal emitted by the sample was separated from the illumination beams
by a quad-band dichroic (Semrock, Di01-R405/488/561/635) and a 4 f Fourier system (New-
port, two collimated lenses PACO87AR.14, f = 200 mm) was collimated with the im-
age plane of the microscope. At the intermediate Fourier image plane, a DH-PSF phase
mask was placed and aligned (DoubleHelix LLC, matched for 580 nm wavelength, and of
7.25 mm diameter; cf. Section 3.2.2). Finally, the signal was filtered by the combination of a
640 nm long-pass (Semrock, BLP01-635R) and a band-pass filters (Semrock, FF02-675/67)
and projected on the chip of an EMCCD camera (Photometrics, Evolve Delta) placed at the
focal image plane of the second Fourier lens (cf. Figure 3.2).
Calibration z-stacks (thanks to a piezo-stage) and SMLM acquisitions were recorded on
µManager [173] and analysed with the easy-dhpsf algorithm (Matlab routine available at
https://code.google.com/p/easy-dhpsf/source/checkout [319]).
8.2.2 Automation of SMLM imaging
Polydimethylsiloxane (PDMS) micro-fluidic devices were obtained as described in ref. [100]
and bonded to a glass coverslip after incubation of the coverslip in an oxygen plasma for 1’.
The design and growing conditions of the strain of S. pombe imaged in Figure 3.17 are also
described in ref. [100], as is the preparation of the sample. In short, live cells were loaded
into the device at a rate of 10 µL/h and maintained at 28◦C (± 2◦C) by pumping heated wa-
ter through the control layer of the PDMS device. Once the cells had reached the trapping
region the pump was switched off and the residual pressure in the tubing maintained a slow,
smooth flow.
For the data presented in Figure 3.17 and Movie B.2, images were acquired under
561 nm HILO excitation (at power densities of ∼4 kW/cm2) and 405 nm pulsed HILO
202 Materials and methods
activation (∼5 W/cm2). A pulse sequence was chosen to ensure the irreversible bleaching
of single mEos2 molecules in the cluster prior to the next imaging cycle. Acquisitions of 10
frames (with 33 ms exposure) were performed between two activation pulses. Although in
SMLM experiments, acquisition cycles would be repeated until no further fluorescence ac-
tivation is observed (∼5,000 frames), in this proof-of-concept experiment, acquisition was
stopped after the first PALM imaging cycle. Illumination beams and position of sample
were controlled using mechanical shutters and a motorised (xyz) stage (Prior, Optiscan III,
V31XYZEF) and a bespoke script written as a µmanager [173] plug-in (Software B.3).
8.3 Materials and methods for Chapter 4
8.3.1 Bead calibration experiments
The main bead calibration experiments were performed as described in Section 8.1 and
Table 8.2 by M. Palayret. The acquisition protocol consisted in recording a z-stack of
201 steps, each axially separated by 10 nm. For each step, 10 frames were imaged with
a 33 ms exposure.
Additionally, in order to test the robustness of the vlsSMLM method, ‘quick & dirty’
calibration experiments were repeated by T.J. Etheridge and Dr U. Endesfelder on two other
TIRF instruments (Figure 4.6, A, 2nd and 3rd-4th rows, resp.). Consequently, additional
precisions about these instruments and the imaging protocols follow:
Figure 4.6, A, 2nd row. 560 nm emitting TetraSpeck beads were fixed on a coverslip and
imaged every 20 nm using a z piezo-stage on a replica of the instrument described in
Section 8.1 (cf. Figure 3.1). The fluorescence signal was spectrally separated from the
illumination light by a 561 nm long-pass (Semrock, BLP01-561R) and a band-pass
filters (Semrock, FF01-593/40). The exposure of each frame was set to 100 ms.
Figure 4.6, A, 3rd and 4th rows. 660 nm emitting TetraSpeck beads (Invitrogen) or 655 nm
emitting quantum dots were fixed on a coverslip. Chamber slides were mounted on a
Nikon Ti Eclipse inverted microscope equipped with a 100x oil immersion objective
lens (Apo TIRF 100X 1.49 Oil, Nikon). The 647 nm beam of a multi-line argon
krypton mixed gas ion laser (Coherent, Innova70C) was projected into the microscope
on a quad-band dichroic mirror (F73-888, AHF). Fluorescence light was collected by
the objective, spectrally separated from the excitation light by a band-pass filter (F47-
700, AHF) and projected on an EMCCD (Andor, iXon DU897). Videos of the traces
8.3 Materials and methods for Chapter 4 203
of the different fluorescent particles were recorded at different axial position in 20 nm
steps using a z piezo-stage.
8.3.2 Purification of mEos3.1
The purification of mEos3.1 from Escherichia coli was adapted from Fang et al. by Dr A.T.
Watson [320]. The E. coli codon-optimised mEos3.1 gene was cloned into the N-terminal
6xHis-tagging expression vector pTWO-E (a kind gift from the Pearl lab) and the resulting
plasmid was transformed into E. coli BL21(DE3) cells (Novagen). Three litres of culture
were grown to an A600 nm absorbance of 0.6 and cooled to 20◦C. IPTG1 (Fisher Scientific)
was then added to a final concentration of 0.5 mM to induce protein expression, and the
cultures were further incubated in a shaker for 14–16 h. Collected cell pellets were placed
at -20◦C overnight (not snap-frozen).
Pellets were resuspended in buffer A (10 mM Tris2, 100 mM NaCl, 50 mM Na2HPO4, at
pH 8.0) with EDTA3-free protease inhibitor mix (Roche) (one tablet per 3 L of culture) and
[4-(2-Aminoethyl) benzenesulfonyl fluoride hydrochloride] at 10 µg/mL, to a total volume
of 35 mL in a 50 mL centrifuge tube. Cells were lysed by sonication using a Sonics Vibra-
Cell for 5 s pulses spaced by 10 s resting time, to a total sonication time of 3’. Lysed
cells were centrifuged at 18,500 rpm (32,000g) in an Allegra 64R ultracentrifuge (Beckman
Coulter) with a F0650 rotor for 1 h. The supernatant was applied to a 5 mL Ni-NTA agarose
resin (Qiagen) (pre-equilibrated using buffer A) and incubated for 1 h at 4◦C with gentle
agitation. The column was washed with eight column volumes of buffer B (50 mM KH2PO4,
300 mM NaCl, 20 mM imidazole, 5% glycerol, at pH 7.9), and the protein was eluted with
two column volumes of buffer C (50 mM KH2PO4, 300 mM NaCl, 250 mM imidazole, 5%
glycerol, at pH 7.9).
The protein was further purified on an SD200 size exclusion column pre-equilibrated
with buffer D (10 mM Tris, 100 mM NaCl, 5% glycerol, at pH 8.0 and degassed) and protein
aliquots were snap-frozen in liquid nitrogen and stored at -80◦C. The protein purification
was performed at 4◦C. The exposure to light was minimised by covering tubes and columns
in foil, using red light-emitting diodes and growing cultures in the dark.
1IPTG: Isopropyl b-D-thiogalactopyranoside.
2Tris: tris(hydroxymethyl)aminomethane.
3EDTA: ethylenediaminetetraacetic acid
204 Materials and methods
8.3.3 Imaging separated fluorophores
For Figure 4.2, purified mEos3.1 fluorescent protein and SNAP-Cell TMR-Star dye (New
England BioLabs, S9105S) were diluted 1/105 in twice-200-nm-filtered PBS (Invitrogen,
003002) and fixed for 5’ on poly-L-lysine-coated plasma-cleaned coverslips as described in
Section 8.1. For each acquisition, 4,000 frames were imaged with a 50 ms exposure. The
FF01-587/35 (Semrock) emission band-pass filter was used.
The δK280 mutant of the neuronal microtubule-associated Tau protein was labelled
by Dr Magnus Kjaergaard. The two natural cysteine residues at positions 291 and C322
were mutated to alanine, and a cysteine was introduced at position 260. This enabled
fluorescently-labelling of the protein without introducing bulky residues in the repeat do-
main that is responsible for aggregation. The construct was expressed in E. coli, and la-
belled with either AlexaFluor647 using maleimide chemistry. Mass spectroscopy, SDS-
PAGE (sodium dodecyl sulfate polyacrylamide gel electrophoresis) and analytical size ex-
clusion all showed the protein labelling to have been successful, and the material to be
apparently free of pre-formed aggregates (data not shown). Labelled protein was kept in a
50 mM ammonium acetate buffer (pH 7) for 90’ at 37◦C without shaking. The sample was
then prepared as described in Section 8.1, altough the coverslip was coated with negatively
charged poly-L-glutamate, to immobilise the positively charged Tau proteins, for 30’. Dr
S.F. Lee imaged the samples under constant 640 nm TIRF illumination at various power
densities.
8.3.4 S. pombe experiments
For the experiment described in Figure 4.8, Dr A.T. Watson established the S. pombe cell
line, H. Armes prepared the sample and imaged it. Below is summed a brief description of
both steps.
The gene sequence of mEos2 was sub-cloned into the pFA6a-kanMX6 C-terminal gene
tagging plasmid [321] to create pFA6a-mEos2-kanMX6. The mEos2-kanMX6 sequence
was then integrated directly downstream of the S. pombe cdc22 gene open-reading frame, es-
sentially as described in ref. [321], to create the AW709 strain (h-, cdc22-mEos2-kanMX6,
leu1-32, ura4D18). A PCR against the inserted fragment showed that the diploid cells were
heterozygote for cdc22-mEos2.
Cells were grown from frozen stocks on YEA (yeast extract 0.5% w/v, glucose 3.0%
w/v, adenine, leucine and uracil all at 225 mg/L, and 2% Agar) plates then cultured in
Edinburgh Minimal Media (EMM2) (filter sterilised) in a shaking incubator at 30◦C. Fresh
8.3 Materials and methods for Chapter 4 205
plates were prepared for new experiments and the cultures set up 12-24 hours before the
experiment. Samples for fixing were taken at an optical density (OD) at 595 nm of 0.1-0.2.
Samples were spun down at 8,000 g. Cell pellets were fixed in 1% formaldehyde (made
from 16% methanol-free stock, Fisher Scientific) in MilliQ water at room temperature for
15’. The fixed sample was spun down and washed three times in MilliQ water before being
re-suspended in 10-20 µL of MilliQ water.
Finally, 5 µL of the resuspended sample were added between an agarose pad (100 µL
of 1% agarose −Sigma, A0169− in MilliQ water pipetted between two cleaned coverslips)
and a cleaned coverslip (cf. Section 8.1). For each acquisition, 5,000 frames were recorded
with a 50 ms exposure.
8.3.5 Embryonic stem cells
For the experiment described in Figure 4.9, Dr S. Basu established the embryonic stem cell
line, prepared the sample and imaged it. Below, a brief summary describes both steps.
mEos3.2 (mEso3.2 plasmid kindly given by Tao Xu [121]) was first cloned at the N-
terminus of Cenp-A (cDNA plasmid from Thermo Scientific, clone MMM1013-64851) us-
ing a standard cloning vector. PCR of mEos3.2 and Cenp-A were carried out with primers
described in Table 8.3. Then, mEos3.2-Cenp-A was cloned into the NcoI/XbaI site of a
mammalian expression vector (pEF.myc.ER-E2-Crimson; Addgene plasmid 38770) [322].
Sequencing of the vector was carried out using the pEF forward primer and the BGH reverse
primer (Table 8.3).
Table 8.3 DNA primers used for Cenp-A-mEos3.2 cloning and sequencing. Restriction sites are written in
lower case.
Primer name Sequence
NcoI_mEos3_F TGAACACGTGGCCAccatggGTAGTGCGATTAAGCCAGACATGAAGATCA
XhoI_mEos3_R AGGctcgagTCGTCTGGCATTGTCAGGCAATC
XhoI_NheI_Cenp-A_F AGGctcgaggctagcATGGGCCCGCGTCGCAAA
Xba1_Cenp-A_R AGGtctagaTTAGGGGAGTCCGCCCTCG
pEF_F TCTCAAGCCTCAGACAGTGGTTCAAAGT
BGH_R TAGAAGGCACAGTCGAGG
Mouse embryonic Mbd3Flox/- [323] stem cells were cultured in standard serum and LIF
(leukaemia inhibitory factor) conditions as previously described [324]. These cells were
transfected with the mEos3.2-Cenp-A plasmid using lipofectamine 2000 (Life Technolo-
gies), and after 24 h, a stable cell line was selected and maintained using 500 µg/mL ge-
206 Materials and methods
neticin selection. These cells were then trypsinised and resuspended in PBS, fixed in 1%
methanol-free formaldehyde (ThermoScientific, 28908) for 5’ and imaged on clean cover-
slips (cf. Section 8.1). The acquisition consisted of 5,000 frames recorded with a 50 ms
exposure.
8.4 Materials and methods for Chapter 6
The S. pombe strains used in Chapter 6 were designed and cloned by Dr A.T. Watson, and the
corresponding samples prepared by H. Armes. The T-cell lines were designed and cloned
by E. Huang. Sample preparation for the T cells and all imaging were performed by M.
Palayret.
8.4.1 S. pombe cell lines
Dr A.T. Watson initially designed these constructs using the sequence for mEos3.1 with the
codons optimised for expression in S. pombe. Three constructs were designed consisting of
a single, double or triple repeat of mEos3.1 separated by a TGS linker and under the control
of the uracil inducible promoter Urg1. The Urg1 promoter, when induced, gives a very
high expression level, thus the sequence of the determinant of selective removal (DSR),
which reduces expression level through transcript removal [325], an effect that increases
with repeats, was added at the C-terminal end of the gene construct at varying levels (1-6
repeats).
The expression and size of the three constructs were confirmed on a Western blot. Ex-
pression levels of both induced and un-induced samples from the varying DSR levels were
also examined. Even un-induced, the Urg1 promoter is ‘leaky’: expression can still be ob-
served indicating that DSR induction is necessary to fully prevent expression, or, in the case
of the experiments described in this work, to obtain very low expression. Consequently,
constructs containing six DSRs were transformed into leu− AW459 yeast from the pAW8
plasmid using cassette exchange and further imaged [326].
8.4.2 T cell lines
The genes encoding human CD28, human CD86, human TCRβ , murine CD3δ , murine
CD3ζ and mEos2 fluorescent protein were amplified by PCR. PCR fragments were lig-
ated to form CD28-mEos2, CD86-mEos2, TCRβ -mEos2 and CD3δ -mEos2 constructs,
which together with CD3ζ were each cloned into the pHR-SIN lentiviral vector. Human
8.4 Materials and methods for Chapter 6 207
hybridoma J.RT3-T3.5 cells [327], deficient for TCRβ , were infected with VSV-G pseu-
dotyped viruses produced from transiently transfected HEK-293T cells with either of the
three CD28-mEos2, CD86-mEos2 or TCRβ -mEos2 constructs as described in ref. [250].
Similarly, mouse hybridoma BW5147 cells, lacking expression of CD3δ and CD3ζ , were
co-infected with similar viruses carrying the CD3δ -mEos2 and CD3ζ constructs. Correct
surface expression of the appropriate membrane receptor was confirmed by flow cytometry
(using the blue emission of the inactivated form of mEos2 and appropriate staining antibod-
ies). Finally cells were sorted by flow cytometry.
The mouse anti-mouse-TCRβ antibodies (clone F23.1) were purified and fragmented
as described in ref. [250]. Fabs were labelled with Atto655 according to the kit labelling
protocol (Sigma Aldrich).
8.4.3 SMLM imaging
Circa 2 · 106 T cells were washed (2x) in warm JMEM and resuspended in 1 mL of fix-
ing buffer (4% formaldehyde (Lancaster, 14190), 0.2% glutaraldehyde (Sigma Aldrich,
G6257)). Cells were fixed for 30’ at 37◦C, then washed (3x) in 200 µL cold filtered PBS
and finally resuspended in 50 µL. Agarose pads were formed by dropping 100 µL of 2%
agarose (Sigma, A0169) in filtered PBS4 between two plasma-cleaned coverslips. Agarose
pads were left to set at 4◦C for >10’. Then, one of the two coverslips sandwiching each
agarose pad was carefully removed thanks to a scalpel and 20 µL of the fixed sample were
dropped on the pad. A new plasma-cleaned was added on the top of the pad, over the cells,
and the whole sample was mounted upside down on the objective (so that the cells were
immobilised on the bottom coverslip in contact with the objective thanks to the weight of
the agarose pad).
Yeasts were grown in Edinburgh Minimal Medium supplemented with 0.1 g/L adenine
and leucine up to an OD of 0.3-0.4 at 595 nm absorption. 8 mL of cells were washed in
MilliQ water and resuspended in 1 mL of 1% formaldehyde (Lancaster, 14190) in MilliQ
water and fixed for 30’ at room temperature. Fixed cells were washed (5x) in MilliQ water
and finally resuspended in ∼20 µL MilliQ water. Before imaging, 5 µL of the sample were
pipetted onto the agarose pad and plasma-cleaned coverslip was placed on the top, as above.
During the whole preparation, for either yeast and T-cell samples, samples were strictly
protected from UV or blue light. Only red light-emitting diodes were used whenever the
4To prevent any contamination, metallic tools were forbidden from the agarose stock. Similarly, a critical
step to avoid single-molecule noise in the sample is to thoroughly melt and mix the agarose. Un-melted
agarose crystals might explain the high background noise observed in some original experiments.
208 Materials and methods
sample was not completely covered with aluminium foil. A transmission red light was
used to search for immobilised cells and adjust the focus before starting the acquisition.
To insure repeatability, the acquisition was entirely led by a macro written in µManager
controlling the shutters, the camera and the 405 nm laser power. For each cell, 1,000 frames
with a 50 ms exposure were imaged under 561 nm illumination only to photo-bleached pre-
activated mEos proteins and the background noise. Then, an acquisition of 4,000 frames
(50 ms exposure) was taken under both constant TIRF 561 nm illumination and Fermi TIRF
405 nm activation (cf. Section 3.3.4). Finally, a transmission red light snapshot of the cell
was recorded.
For the two-colour imaging experiment presented in Figure 6.16, the sample preparation
varied slightly: after fixation, cells were washed (3x) in cold filtered PBS and incubated in
50 µL of 6.25 µg/mL Atto655-labelled αmurine-TCRβ Fabs for 30’ on ice. Cells were then
washed (3x) in cold filtered PBS and immobilised as previously between an agarose gel pad
and a glass coverslip. Samples were sequentially imaged under 640 nm TIRF illumination
(2,000 frames; 50 ms exposure), and under both 561 nm TIRF illumination and continuous
Fermi TIRF activation (2,000 frames; 50 ms exposure).
8.5 Materials and methods for Chapter 7
Molecular cloning of the various constructs used in Chapter 7, transfection of the cell lines
with the various constructs, IL-2 experiments, anti-murine-CD45 (αmCD45) mAb purifi-
cation and Fab generation and labelling were the work of M. Aßmann. M. Palayret cultured
the cell lines and prepared the samples for SMLM imaging and performed the imaging
experiments presented in this work and their analyses.
8.5.1 Molecular Cloning
Different genes of interest (cf. Table 8.4) were inserted into the pHR-CSGWdNotI/dEcoRI
(pHR) vector via the MluI/BamHI/NotI cloning sites [328]: the structure of the constructs
were designed as [pHR-MluI-(gene of interest)-BamHI-(10 amino-acid linker)-mEos3.2].
After cloning each gene in the pHR vector, competent Top10 bacteria were transformed with
each construct for building working stocks of plasmids (Invitrogen, HiPure plasmid mini-
prep kit). Each construct was sequenced to confirm the correct insertion and the absence of
undesirable mutations.
8.5 Materials and methods for Chapter 7 209
Table 8.4 Gene constructs used in Chapter 7. Legend: rCD28: ecto-domain of rat CD28; mCD28: transmem-
brane and intracellular domain of murine CD28; trunc-mCD28: mCD28 which signalling motifs have been
truncated; linker: a 30-amino-acid linker; hPD-1: the ecto-domain of human-PD-1.
Name Gene construct Figures
CD28 [rCD28]-[mCD28] Figure 7.8
CD28-mEos [rCD28]-[mCD28]-[linker]-mEos3.2
Figures 7.4, 7.6,
7.7 & 7.17
Truncated CD28-mEos [rCD28]-[trunc-mCD28]-[linker]-mEos3.2
Figures 7.6, 7.8, 7.10, 7.11,
7.12, 7.13 & 7.15
PD-1-mEos [hPD-1]-[trunc-mCD28]-[linker]-mEos3.2 Figure 7.9
CD86-mEos [murine-CD86]-[linker]-mEos3.2 Figure 7.8
Fc-CD28-mEos [human-Fc domain]-[trunc-mCD28]-[linker]-mEos3.2 Figure 7.14
8.5.2 Stable transfection with lentivirus
Stable transfection of cell lines was achieved by transducing them with lentivirus, and was
confirmed by flow cytometry. HEK-293Tcells [329] were used for virus production. They
were transfected (using the GeneJuice transfection reagant) with the p8.91 vector, contain-
ing the structural proteins and enzymes needed to generate infectious viral particles, the
pMDG vector, containing viral coat proteins for packaging, and the pHR vector, containing
the gene of interest (cf. previous section), flanked by long terminal repeats, under control of
the CMV promoter as proposed in ref. [328].
Viruses were harvested 48 h post-transfection, filtered with a 0.45 µL syringe filter and
added to 106 BW5147 cells in 0.5 mL JMEM (cf. next section). The following day the
medium was replaced with fresh JMEM. After 5-7 days, the expression of the gene of in-
terest was tested by flow cytometry. Expression was matched for different constructs within
one experiment, e.g. full-length and truncated CD28-mEos3.2, or truncated CD28 and trun-
cated Fc-CD28.
8.5.3 Cell Culture
All cell lines imaged in Chapter 7 originate from the BW5147 cell line that is deficient in
CD3δ and CD3ζ chains. Thus, if not transfected with adequate constructs, they do not
express any TCR on their plasma membrane.
Mouse leukaemia BW5147 T-cell lines were cultured in JMEM medium (RPMI-1640
medium without phenol red (Gibco, 11835-063) supplemented with 10% foetal calf serum
210 Materials and methods
(Thermo Scientific, HyClone SV30160.03), 10 mM HEPES5 (Sigma, H0887), 1 mM sodium
pyruvate (Gibco, 11360-039), 2 mM L-glutamine (Gibco, 2503-024) and antibiotics [50 units
penicillin and 50 µg streptomycin per mL] (Gibco, 15140-122)). They were split∼1/10 into
fresh medium every 2-3 days to keep them at a density between 105 and 106 cells/mL. Cells
were grown at 37◦C in a 5% CO2 atmosphere and regularly counted in a haemocytometer.
Trypan blue (Sigma, T8154), a dye that penetrates dead but not viable cells, was mixed 1:1
(v/v) with cells before counting to assess the general health of the cultures.
8.5.4 Sample preparation
In both IL-2 experiments and sptPALM imaging, cells were decorated with mAbs when
incubating in an activation solution for ∼15’ and directly dropped (with no washing step)
on a glass surface coated with 500 µg/mL donkey anti-mouse secondary antibodies (Jackson
ImmunoResearch Laboratories Inc, 715-001-003) (cf. Figure 7.3). The activation buffer
was made fresh for each experiment and consisted in 50 µg/mL (except where indicated,
especially Figure 7.10) mAb in JMEM buffer. The list of mAbs used in this thesis is given in
Table 8.5. In the experiment presented in Figure 7.11, the coating and decorating protocols
were varied as indicated in the legend and the accompanying text.
Table 8.5 Monoclonal antibodies used in Chapter 7. †: putative SA, based on the location of the epitope of the
mAb. All mAbs (except from the rat mAb KT3) are recognised by donkey anti-mouse secondary antibodies.
Clone Protein target SA / CA behaviour Source
JJ316 rat-CD28 SA BD Pharmingen, 554992
JJ319 rat-CD28 CA eBioscience, 16-0280
Clone 2 human-PD1 CA Prof. S.J. Davis
Clone 10 human-PD1 CA Prof. S.J. Davis
Clone 19 human-PD1 SA† Prof. S.J. Davis
IC10 human-Fc domain SA†
kind gift from Dr John Young,
Institute of Animal Health, Compton, UK
KT3 mouse-CD3 SA Bio-Rad AbD Serotec Limited, MCA500XZ
Once the cells were decorated and dropped on the coated glass surface, their IL-2 secre-
tion was either monitored by M. Aßmann (with TCR-sufficient BW5147 cells, except when
indicated in Figure 7.6) or the cells forming a contact with the glass coverslip were imaged
with sptPALM by M. Palayret (with TCR-deficient BW5147 cells, except when indicated in
Figure 7.7).
5HEPES: 4-(2-hydroxyethyl)-1-piperazineethanesulfonic acid
8.5 Materials and methods for Chapter 7 211
8.5.5 IL-2 in vitro stimulation assay
Thermo Scientific Nunc MicroWell 96-well optical-bottom plates with coverglass base were
wet-coated in 1x coating buffer (from the Mouse IL-2 ELISA Ready-SET-Go! kit, eBio-
science) overnight at 4◦C with donkey anti-mouse polyclonal secondary antibody at 500 µg/mL.
Wells were washed three times with 250 µL sterile PBS (a blocking step did not affect re-
sults, and was thus omitted).
107 cells/mL resuspended in the activation solution (50 µg/mL mAb in JMEM). Cells
were incubated at room temperature for ∼15’. Each condition was preformed in dupli-
cates or triplicates, i.e. 2-3 wells per condition. Each well received 5 · 105 cells in 100 µL
activation solution. Wells were then incubated at 37◦C in a 5% CO2 atmosphere for 48 h.
The cell supernatant was analysed for released IL-2 with the Mouse IL-2 ELISA Ready-
SET-Go! kit (eBioscience), following the kit protocol. An experimental calibration curve
was used to convert the measured absorption at 450 nm into concentration of released IL-2
in the cell supernatant. IL-2 concentrations were corrected for day-to-day variations (up to
2-4 fold differences observed between replicates performed on different days) by normali-
sation to a positive control (activation of the cells by 10 µg/mL KT3 mAb directly coated
with the secondary antibody on the glass; no decorating mAb). To be able to display IL-2
concentration in pg/mL units, the normalised values were multiplied with a typical (average)
experimental IL-2 concentration of the positive control. All error bars display the standard
deviation of the mean.
8.5.6 Generating Atto655-labelled Anti-murine-CD45 Fabs
The YW62.3.20 [330, 331] monoclonal antibody is a pan-reactive anti-mouse CD45 IgG2b.
Hybridoma expressing the antibody in the supernatant was grown. Specific binding of the
mAb present in the supernatant to CD45-expressing but not CD45-deficient BW5147 cells
was confirmed by flow cytometry. After three weeks, the mAb was purified on protein-G
coupled sepharose beads from the supernatant of 2 L of hybridoma culture. Ten fractions
were collected, pooled, concentrated on Vivaspin 20 columns (Sartorius Stedium Biotech)
and purified by fast protein liquid chromatography (FPLC) with a SuperDex 200 HR column
in an ÄKTA FPLC system. Purity of the mAb was confirmed by SDS-PAGE. As no con-
taminant bands were detected, all antibody-containing fractions were pooled, concentrated
to 2 mg/mL as above, and frozen in 1 mL aliquots.
To produce Fabs, 6 mg of the mAb was digested with immobilised papain slurry (Ther-
moScientific). The supernatant and two washes of the slurry were purified on protein-G
212 Materials and methods
coupled sepharose beads, pooled and further purified by FPLC with a SuperDex 75 HR col-
umn. No Fc or whole antibody contamination was observed in the sepharose bead elution.
Purity of the Fab fragments was confirmed by SDS-PAGE.
To label the anti-mouse CD45 Fab fragments, the protocol of the Atto655 protein la-
belling kit (Sigma Aldrich) was optimised. Freshly made 0.2 M sodium bicarbonate at
pH 9.0 was added to 100 µL of Fab at 1 mg/mL in HEPES buffered saline (150 mM NaCl,
20 mM HEPES in MilliQ water) to achieve a final pH of 8.3. Then, 6.7 nmol of Atto655-
NHS ester in PBS were added to the Fabs and the reaction was incubated for 2 h at room
temperature, protected from light. The labelled Fabs were purified from unbound dye with
gel filtration columns from the antibody conjugate purification kit (LifeTechnologies). Spe-
cific binding of the labelled anti-CD45 Fab was confirmed by flow cytometry.
8.5.7 Single-particle tracking PALM imaging
For sptPALM imaging, 1.25 · 105 cells were washed twice in 200 µL JMEM, resuspended
in 100 µL of the considered warm activation solution, and incubated for 10’ at 37◦C (except
where indicated in Figure 7.6; the incubation was then done at room temperature). Square
hydrophobic imaging gaskets (Bio-rad, SLF0201) were applied on plasma-cleaned cover-
slips and 100 µL of secondary antibodies were dropped on each coverslip. After 15’, one
coverslip was washed (3x) with warmed filtered PBS (37◦C) and 50 µL of the sample were
dropped on the coated coverslip. The sample was then directly mounted on the objective
and cells were imaged as soon as they formed a contact with the coverslip (followed until
them with white transmission light).
A plexiglass chamber was designed to enclose the sample, the stage of the microscope
and the objective, but not the cooled EMCCD camera, and stably heated to 37◦C thanks to
two vibration-free heaters (DigitalPixel, DP2000 Intelligent Heater). To avoid thermal drift
and obtain a stable temperature, heaters were switched on 2 h before imaging. In Figure 7.6,
where indicated, the heaters were kept off and the samples were imaged at 20◦C.
For each cell, the acquisition consisted in 3,000 frames with a 50 ms exposure under
constant TIRF 561 nm illumination (∼130 W/cm2) and 405 nm activation (<0.61 W/cm2).
Analysis is thoroughly described and discussed in Section 7.2.1.
For two-colour imaging (Figures 7.12, 7.13 and 7.15), the sample preparation slightly
differed:
Labelling with activation mAb. Cells were incubated for 15’ at 37◦C in 200 µL of warm
activation buffer made of 40 µg/mL unlabelled mAb and 20 µg/mL Atto655-labelled
8.5 Materials and methods for Chapter 7 213
mAb in JMEM. After incubation, cells were washed (3x) in 100 µL warm JMEM and
finally resuspended in 100 µL warm JMEM (with no mAb), before being dropped on
the coated coverslip.
Labelling CD45. Cells were incubated for 15’ at 37◦C in 200 µL of warm activation buffer
made of 10 µg/mL activating mAb and 2.5 µg/mL α-mCD45 Fab-Atto655. After
incubation, cells were washed (3x) in 200 µL warm JMEM and finally resuspended in
100 µL warm incubation buffer as originally described (50 µg/mL of the considered
mAb in JMEM), before being dropped on the coated coverslip.
However, the acquisition protocol was similar for both experiments: after alignment
and calibration of the dualview (cf. Section 8.1), cells were imaged after stabilisation of
the formation of a contact with the coated coverslip (∼1-5’ after apparition in the TIRF
field), under constant TIRF 640 nm (∼150 W/cm2), 561 nm (∼130 W/cm2) illuminations
and 405 nm (<0.5 W/cm2) activation. Both colours were simultaneously imaged in the two
channels of the dualview. Because of some achromatism, the focal planes for both channels
were not identical. The focus was thus carefully adjusted in an intermediary position for
which in-focus PSFs were observed in both channels. Acquisitions consisted in movies of
3,000 frames with a 50 ms exposure.

References
[1] M. H. Horrocks, M. Palayret, D. Klenerman, and S. F. Lee, “The changing point-
spread function: single-molecule-based super-resolution imaging.,” Histochemistry
and cell biology, vol. 141, pp. 577–85, June 2014.
[2] M. Palayret, “Single molecule super-resolution microscopy: All for one and one for
all,” Physiology News, no. 95, pp. 24–27, 2014.
[3] G. G. Stokes, “On the Change of Refrangibility of Light,” Philosophical Transactions
of the Royal Society of London, vol. 142, pp. 463–562, Jan. 1852.
[4] E. Abbe, “Beiträge zur Theorie des Mikroskops und der mikroskopischen
Wahrnehmung,” 1873.
[5] A. Santos and I. T. Young, “Model-based resolution: applying the theory in quantita-
tive microscopy.,” Applied optics, vol. 39, pp. 2948–58, June 2000.
[6] R. E. Thompson, D. R. Larson, and W. W. Webb, “Precise nanometer localization
analysis for individual fluorescent probes.,” Biophysical journal, vol. 82, pp. 2775–
2783, 2002.
[7] A. Szymborska, A. de Marco, N. Daigle, V. C. Cordes, J. A. G. Briggs, and J. Ellen-
berg, “Nuclear pore scaffold structure analyzed by super-resolution microscopy and
particle averaging.,” Science (New York, N.Y.), vol. 341, pp. 655–8, Aug. 2013.
[8] F. Chen, P. W. Tillberg, and E. S. Boyden, “Expansion microscopy,” Science,
pp. science.1260088–, Jan. 2015.
[9] M. G. L. Gustafsson, “Surpassing the lateral resolution limit by a factor of two using
structured illumination microscopy,” Journal of Microscopy, vol. 198, pp. 82–87,
May 2000.
[10] S. W. Hell and J. Wichmann, “Breaking the diffraction resolution limit by stimulated
emission: stimulated-emission-depletion fluorescence microscopy.,” Optics letters,
vol. 19, pp. 780–2, June 1994.
[11] S. W. Hell, “Far-field optical nanoscopy.,” Science (New York, N.Y.), vol. 316,
pp. 1153–8, May 2007.
[12] M. G. L. Gustafsson, “Nonlinear structured-illumination microscopy: wide-field flu-
orescence imaging with theoretically unlimited resolution.,” Proceedings of the Na-
tional Academy of Sciences of the United States of America, vol. 102, pp. 13081–6,
Sept. 2005.
216 References
[13] S. W. Hell, “Toward fluorescence nanoscopy.,” Nature biotechnology, vol. 21,
pp. 1347–55, Nov. 2003.
[14] M. Hofmann, C. Eggeling, S. Jakobs, and S. W. Hell, “Breaking the diffraction barrier
in fluorescence microscopy at low light intensities by using reversibly photoswitch-
able proteins.,” Proceedings of the National Academy of Sciences of the United States
of America, vol. 102, pp. 17565–9, Dec. 2005.
[15] V. Westphal, M. A. Lauterbach, A. Di Nicola, and S. W. Hell, “Dynamic far-field
fluorescence nanoscopy,” New Journal of Physics, vol. 9, pp. 435–435, Dec. 2007.
[16] V. Westphal, S. O. Rizzoli, M. A. Lauterbach, D. Kamin, R. Jahn, and S. W. Hell,
“Video-rate far-field optical nanoscopy dissects synaptic vesicle movement.,” Science
(New York, N.Y.), vol. 320, pp. 246–9, Apr. 2008.
[17] A. Chmyrov, J. Keller, T. Grotjohann, M. Ratz, E. D’Este, S. Jakobs, C. Eggeling,
and S. W. Hell, “Nanoscopy with more than 100,000 ’doughnuts’.,” Nature methods,
vol. 10, pp. 737–40, Aug. 2013.
[18] E. Betzig, G. H. Patterson, R. Sougrat, O. W. Lindwasser, S. Olenych, J. S. Bonifa-
cino, M. W. Davidson, J. Lippincott-Schwartz, and H. F. Hess, “Imaging intracellular
fluorescent proteins at nanometer resolution.,” Science (New York, N.Y.), vol. 313,
pp. 1642–1645, Sept. 2006.
[19] M. J. Rust, M. Bates, and X. Zhuang, “Sub-diffraction-limit imaging by stochastic
optical reconstruction microscopy (STORM).,” Nature methods, vol. 3, pp. 793–795,
2006.
[20] S. T. Hess, T. P. K. Girirajan, and M. D. Mason, “Ultra-high resolution imaging by
fluorescence photoactivation localization microscopy.,” Biophysical journal, vol. 91,
pp. 4258–72, Dec. 2006.
[21] T. Müller, C. Schumann, and A. Kraegeloh, “STED microscopy and its applications:
new insights into cellular processes on the nanoscale.,” Chemphyschem : a European
journal of chemical physics and physical chemistry, vol. 13, pp. 1986–2000, June
2012.
[22] J. R. Allen, S. T. Ross, and M. W. Davidson, “Structured illumination microscopy
for superresolution.,” Chemphyschem : a European journal of chemical physics and
physical chemistry, vol. 15, pp. 566–76, Mar. 2014.
[23] E. Betzig, “Proposed method for molecular optical imaging.,” Optics letters, vol. 20,
pp. 237–9, Feb. 1995.
[24] S. W. Hell and M. Kroug, “Ground-state-depletion fluorscence microscopy: A con-
cept for breaking the diffraction resolution limit,” Applied Physics B Lasers and Op-
tics, vol. 60, pp. 495–497, May 1995.
[25] J. Fölling, M. Bossi, H. Bock, R. Medda, C. A. Wurm, B. Hein, S. Jakobs,
C. Eggeling, and S. W. Hell, “Fluorescence nanoscopy by ground-state depletion
and single-molecule return.,” Nature methods, vol. 5, pp. 943–5, Nov. 2008.
References 217
[26] M. Heilemann, S. van de Linde, M. Schüttpelz, R. Kasper, B. Seefeldt, A. Mukher-
jee, P. Tinnefeld, and M. Sauer, “Subdiffraction-resolution fluorescence imaging with
conventional fluorescent probes.,” Angewandte Chemie (International ed. in English),
vol. 47, pp. 6172–6, Jan. 2008.
[27] J. L. Ptacin, S. F. Lee, E. C. Garner, E. Toro, M. Eckart, L. R. Comolli, W. E. Moerner,
and L. Shapiro, “A spindle-like apparatus guides bacterial chromosome segregation.,”
Nature cell biology, vol. 12, pp. 791–8, Aug. 2010.
[28] S. F. Lee, M. A. Thompson, M. A. Schwartz, L. Shapiro, and W. E. Moerner, “Super-
resolution imaging of the nucleoid-associated protein HU in Caulobacter crescen-
tus.,” Biophysical journal, vol. 100, pp. L31–3, Apr. 2011.
[29] K. Lidke, B. Rieger, T. Jovin, and R. Heintzmann, “Superresolution by localization of
quantum dots using blinking statistics.,” Optics express, vol. 13, pp. 7052–62, Sept.
2005.
[30] P. Hoyer, T. Staudt, J. Engelhardt, and S. W. Hell, “Quantum dot blueing and blinking
enables fluorescence nanoscopy.,” Nano letters, vol. 11, pp. 245–50, Jan. 2011.
[31] C. Duan, V. Adam, M. Byrdin, J. Ridard, S. Kieffer-Jaquinod, C. Morlot, D. Arcizet,
I. Demachy, and D. Bourgeois, “Structural evidence for a two-regime photobleaching
mechanism in a reversibly switchable fluorescent protein.,” Journal of the American
Chemical Society, vol. 135, pp. 15841–50, Oct. 2013.
[32] S. W. Englander, D. B. Calhoun, and J. J. Englander, “Biochemistry without oxy-
gen.,” Analytical biochemistry, vol. 161, pp. 300–6, Mar. 1987.
[33] C. E. Aitken, R. A. Marshall, and J. D. Puglisi, “An oxygen scavenging system
for improvement of dye stability in single-molecule fluorescence experiments.,” Bio-
physical journal, vol. 94, pp. 1826–35, Mar. 2008.
[34] M. Swoboda, J. Henig, H.-M. Cheng, D. Brugger, D. Haltrich, N. Plumeré, and
M. Schlierf, “Enzymatic oxygen scavenging for photostability without pH drop in
single-molecule experiments.,” ACS nano, vol. 6, pp. 6364–9, July 2012.
[35] I. Rasnik, S. A. McKinney, and T. Ha, “Nonblinking and long-lasting single-molecule
fluorescence imaging.,” Nature methods, vol. 3, pp. 891–3, Nov. 2006.
[36] Y. Harada, K. Sakurada, T. Aoki, D. D. Thomas, and T. Yanagida, “Mechanochemical
coupling in actomyosin energy transduction studied by in vitro movement assay.,”
Journal of molecular biology, vol. 216, pp. 49–68, Nov. 1990.
[37] J. Vogelsang, T. Cordes, C. Forthmann, C. Steinhauer, and P. Tinnefeld, “Controlling
the fluorescence of ordinary oxazine dyes for single-molecule switching and super-
resolution microscopy.,” Proceedings of the National Academy of Sciences of the
United States of America, vol. 106, pp. 8107–12, May 2009.
[38] M. Heilemann, E. Margeat, R. Kasper, M. Sauer, and P. Tinnefeld, “Carbocyanine
dyes as efficient reversible single-molecule optical switch.,” Journal of the American
Chemical Society, vol. 127, pp. 3801–6, Mar. 2005.
218 References
[39] G. T. Dempsey, M. Bates, W. E. Kowtoniuk, D. R. Liu, R. Y. Tsien, and X. Zhuang,
“Photoswitching mechanism of cyanine dyes.,” Journal of the American Chemical
Society, vol. 131, pp. 18192–3, Dec. 2009.
[40] N. Olivier, D. Keller, P. Gönczy, and S. Manley, “Resolution doubling in 3D-STORM
imaging through improved buffers.,” PloS one, vol. 8, p. e69004, Jan. 2013.
[41] K. Kundu, S. F. Knight, N. Willett, S. Lee, W. R. Taylor, and N. Murthy, “Hydro-
cyanines: a class of fluorescent sensors that can image reactive oxygen species in
cell culture, tissue, and in vivo.,” Angewandte Chemie (International ed. in English),
vol. 48, pp. 299–303, Jan. 2009.
[42] L. Carlini, A. Benke, L. Reymond, G. Lukinavicˇius, and S. Manley, “Reduced
dyes enhance single-molecule localization density for live superresolution imaging.,”
Chemphyschem : a European journal of chemical physics and physical chemistry,
vol. 15, pp. 750–5, Mar. 2014.
[43] S. van de Linde, A. Löschberger, T. Klein, M. Heidbreder, S. Wolter, M. Heilemann,
and M. Sauer, “Direct stochastic optical reconstruction microscopy with standard
fluorescent probes.,” Nature protocols, vol. 6, pp. 991–1009, July 2011.
[44] E. Lubeck and L. Cai, “Single-cell systems biology by super-resolution imaging and
combinatorial labeling.,” Nature methods, vol. 9, pp. 743–8, July 2012.
[45] G. T. Dempsey, J. C. Vaughan, K. H. Chen, M. Bates, and X. Zhuang, “Evaluation
of fluorophores for optimal performance in localization-based super-resolution imag-
ing.,” Nature methods, vol. 8, pp. 1027–36, Dec. 2011.
[46] J. Lippincott-Schwartz and G. H. Patterson, “Photoactivatable fluorescent proteins
for diffraction-limited and super-resolution imaging.,” Trends in cell biology, vol. 19,
pp. 555–65, Nov. 2009.
[47] S.-n. Uno, M. Kamiya, T. Yoshihara, K. Sugawara, K. Okabe, M. C. Tarhan, H. Fujita,
T. Funatsu, Y. Okada, S. Tobita, and Y. Urano, “A spontaneously blinking fluorophore
based on intramolecular spirocyclization for live-cell super-resolution imaging,” Na-
ture Chemistry, vol. 6, pp. 681–689, July 2014.
[48] A. Sharonov and R. M. Hochstrasser, “Wide-field subdiffraction imaging by accumu-
lated binding of diffusing probes.,” Proceedings of the National Academy of Sciences
of the United States of America, vol. 103, pp. 18911–18916, 2006.
[49] G. Giannone, E. Hosy, F. Levet, A. Constals, K. Schulze, A. I. Sobolevsky, M. P.
Rosconi, E. Gouaux, R. Tampé, D. Choquet, and L. Cognet, “Dynamic superresolu-
tion imaging of endogenous proteins on living cells at ultra-high density.,” Biophysi-
cal journal, vol. 99, pp. 1303–10, Aug. 2010.
[50] I. Schoen, J. Ries, E. Klotzsch, H. Ewers, and V. Vogel, “Binding-activated local-
ization microscopy of DNA structures.,” Nano letters, vol. 11, pp. 4008–11, Sept.
2011.
References 219
[51] J. B. Delehanty, H. Mattoussi, and I. L. Medintz, “Delivering quantum dots into cells:
strategies, progress and remaining issues.,” Analytical and bioanalytical chemistry,
vol. 393, pp. 1091–105, Feb. 2009.
[52] H. Deschout, F. Cella Zanacchi, M. Mlodzianoski, A. Diaspro, J. Bewersdorf, S. T.
Hess, and K. Braeckmans, “Precisely and accurately localizing single emitters in
fluorescence microscopy.,” Nature methods, vol. 11, pp. 253–66, Mar. 2014.
[53] P. Dedecker, F. C. De Schryver, and J. Hofkens, “Fluorescent proteins: shine on, you
crazy diamond.,” Journal of the American Chemical Society, vol. 135, pp. 2387–402,
Feb. 2013.
[54] R. Y. Tsien, “The green fluorescent protein.,” Annual review of biochemistry, vol. 67,
pp. 509–44, Jan. 1998.
[55] U. Rothbauer, K. Zolghadr, S. Tillib, D. Nowak, L. Schermelleh, A. Gahl, N. Back-
mann, K. Conrath, S. Muyldermans, M. C. Cardoso, and H. Leonhardt, “Target-
ing and tracing antigens in live cells with fluorescent nanobodies.,” Nature methods,
vol. 3, pp. 887–9, Nov. 2006.
[56] M. H. Kubala, O. Kovtun, K. Alexandrov, and B. M. Collins, “Structural and ther-
modynamic analysis of the GFP:GFP-nanobody complex.,” Protein science : a pub-
lication of the Protein Society, vol. 19, pp. 2389–401, Dec. 2010.
[57] P. J. M. Zessin, K. Finan, and M. Heilemann, “Super-resolution fluorescence imaging
of chromosomal DNA.,” Journal of structural biology, vol. 177, pp. 344–8, Feb.
2012.
[58] I. Nikic´, T. Plass, O. Schraidt, J. Szyman´ski, J. A. G. Briggs, C. Schultz, and E. A.
Lemke, “Minimal tags for rapid dual-color live-cell labeling and super-resolution
microscopy.,” Angewandte Chemie (International ed. in English), vol. 53, pp. 2245–
9, Feb. 2014.
[59] G. Los and L. Encell, “HaloTag: a novel protein labeling technology for cell imaging
and protein analysis,” ACS chemical . . . , vol. 3, no. 6, pp. 373–382, 2008.
[60] S. A. Jones, S.-H. Shim, J. He, and X. Zhuang, “Fast, three-dimensional super-
resolution imaging of live cells.,” Nature methods, vol. 8, pp. 499–508, June 2011.
[61] N. C. Shaner, M. Z. Lin, M. R. McKeown, P. A. Steinbach, K. L. Hazelwood, M. W.
Davidson, and R. Y. Tsien, “Improving the photostability of bright monomeric orange
and red fluorescent proteins.,” Nature methods, vol. 5, pp. 545–51, June 2008.
[62] R. M. Dickson, A. B. Cubitt, R. Y. Tsien, and W. E. Moerner, “On/off blinking
and switching behaviour of single molecules of green fluorescent protein.,” Nature,
vol. 388, pp. 355–8, July 1997.
[63] G. Jung, C. Bräuchle, and A. Zumbusch, “Two-color fluorescence correlation spec-
troscopy of one chromophore: Application to the E222Q mutant of the green fluores-
cent protein,” The Journal of Chemical Physics, vol. 114, p. 3149, Feb. 2001.
220 References
[64] G. Jung, J. Wiehler, B. Steipe, C. Bräuchle, and A. Zumbusch, “Single-molecule mi-
croscopy of the green fluorescent protein using simultaneous two-color excitation.,”
Chemphyschem : a European journal of chemical physics and physical chemistry,
vol. 2, pp. 392–6, June 2001.
[65] R. Ando, H. Mizuno, and A. Miyawaki, “Regulated fast nucleocytoplasmic shuttling
observed by reversible protein highlighting.,” Science (New York, N.Y.), vol. 306,
pp. 1370–3, Nov. 2004.
[66] F. V. Subach, L. Zhang, T. W. J. Gadella, N. G. Gurskaya, K. A. Lukyanov, and V. V.
Verkhusha, “Red fluorescent protein with reversibly photoswitchable absorbance for
photochromic FRET.,” Chemistry & biology, vol. 17, pp. 745–55, July 2010.
[67] H. Chang, M. Zhang, W. Ji, J. Chen, Y. Zhang, B. Liu, J. Lu, J. Zhang, P. Xu, and
T. Xu, “A unique series of reversibly switchable fluorescent proteins with beneficial
properties for various applications.,” Proceedings of the National Academy of Sci-
ences of the United States of America, vol. 109, pp. 4455–60, Mar. 2012.
[68] J. C. Vaughan, S. Jia, and X. Zhuang, “Ultrabright photoactivatable fluorophores
created by reductive caging.,” Nature methods, vol. 9, pp. 1181–4, Dec. 2012.
[69] S. J. Lord, N. R. Conley, H.-l. D. Lee, R. Samuel, N. Liu, R. J. Twieg, and W. E.
Moerner, “A photoactivatable push-pull fluorophore for single-molecule imaging in
live cells.,” Journal of the American Chemical Society, vol. 130, pp. 9204–5, July
2008.
[70] G. H. Patterson and J. Lippincott-Schwartz, “A photoactivatable GFP for selective
photolabeling of proteins and cells.,” Science (New York, N.Y.), vol. 297, pp. 1873–7,
Sept. 2002.
[71] D. M. Chudakov, V. V. Verkhusha, D. B. Staroverov, E. A. Souslova, S. Lukyanov,
and K. A. Lukyanov, “Photoswitchable cyan fluorescent protein for protein track-
ing.,” Nature biotechnology, vol. 22, pp. 1435–9, Nov. 2004.
[72] F. V. Subach, G. H. Patterson, S. Manley, J. M. Gillette, J. Lippincott-Schwartz, and
V. V. Verkhusha, “Photoactivatable mCherry for high-resolution two-color fluores-
cence microscopy.,” Nature methods, vol. 6, pp. 153–9, Feb. 2009.
[73] R. Ando, H. Hama, M. Yamamoto-Hino, H. Mizuno, and A. Miyawaki, “An opti-
cal marker based on the UV-induced green-to-red photoconversion of a fluorescent
protein.,” Proceedings of the National Academy of Sciences of the United States of
America, vol. 99, pp. 12651–6, Oct. 2002.
[74] J. Wiedenmann, S. Ivanchenko, F. Oswald, F. Schmitt, C. Röcker, A. Salih,
K.-D. Spindler, and G. U. Nienhaus, “EosFP, a fluorescent marker protein with
UV-inducible green-to-red fluorescence conversion.,” Proceedings of the National
Academy of Sciences of the United States of America, vol. 101, pp. 15905–10, Nov.
2004.
References 221
[75] N. G. Gurskaya, V. V. Verkhusha, A. S. Shcheglov, D. B. Staroverov, T. V. Chep-
urnykh, A. F. Fradkov, S. Lukyanov, and K. A. Lukyanov, “Engineering of a
monomeric green-to-red photoactivatable fluorescent protein induced by blue light.,”
Nature biotechnology, vol. 24, pp. 461–5, Apr. 2006.
[76] H. Mizuno, T. K. Mal, K. I. Tong, R. Ando, T. Furuta, M. Ikura, and A. Miyawaki,
“Photo-induced peptide cleavage in the green-to-red conversion of a fluorescent pro-
tein.,” Molecular cell, vol. 12, pp. 1051–8, Oct. 2003.
[77] K. Nienhaus, G. U. Nienhaus, J. Wiedenmann, and H. Nar, “Structural basis for
photo-induced protein cleavage and green-to-red conversion of fluorescent protein
EosFP.,” Proceedings of the National Academy of Sciences of the United States of
America, vol. 102, pp. 9156–9, June 2005.
[78] N. Chenouard, I. Smal, F. de Chaumont, M. Maška, I. F. Sbalzarini, Y. Gong, J. Car-
dinale, C. Carthel, S. Coraluppi, M. Winter, A. R. Cohen, W. J. Godinez, K. Rohr,
Y. Kalaidzidis, L. Liang, J. Duncan, H. Shen, Y. Xu, K. E. G. Magnusson, J. Jaldén,
H. M. Blau, P. Paul-Gilloteaux, P. Roudot, C. Kervrann, F. Waharte, J.-Y. Tinevez,
S. L. Shorte, J. Willemse, K. Celler, G. P. van Wezel, H.-W. Dan, Y.-S. Tsai, C. Ortiz
de Solórzano, J.-C. Olivo-Marin, and E. Meijering, “Objective comparison of particle
tracking methods.,” Nature methods, vol. 11, pp. 281–9, Mar. 2014.
[79] S. Manley, J. M. Gillette, G. H. Patterson, H. Shroff, H. F. Hess, E. Betzig, and
J. Lippincott-Schwartz, “High-density mapping of single-molecule trajectories with
photoactivated localization microscopy.,” Nature methods, vol. 5, pp. 155–7, Feb.
2008.
[80] F. Persson, M. Lindén, C. Unoson, and J. Elf, “Extracting intracellular diffusive states
and transition rates from single-molecule tracking data.,” Nature methods, vol. 10,
pp. 265–9, Mar. 2013.
[81] A. Löschberger, C. Franke, G. Krohne, S. van de Linde, and M. Sauer, “Correlative
super-resolution fluorescence and electron microscopy of the nuclear pore complex
with molecular resolution.,” Journal of cell science, vol. 127, pp. 4351–5, Oct. 2014.
[82] P. Sengupta, T. Jovanovic-Talisman, D. Skoko, M. Renz, S. L. Veatch, and
J. Lippincott-Schwartz, “Probing protein heterogeneity in the plasma membrane us-
ing PALM and pair correlation analysis,” 2011.
[83] S. L. Veatch, B. B. Machta, S. A. Shelby, E. N. Chiang, D. A. Holowka, and B. A.
Baird, “Correlation functions quantify super-resolution images and estimate apparent
clustering due to over-counting.,” PloS one, vol. 7, p. e31457, Jan. 2012.
[84] B. Rieger and S. Stallinga, “The lateral and axial localization uncertainty in super-
resolution light microscopy.,” Chemphyschem : a European journal of chemical
physics and physical chemistry, vol. 15, pp. 664–70, Mar. 2014.
[85] M. Badieirostami, M. D. Lew, M. A. Thompson, and W. E. Moerner, “Three-
dimensional localization precision of the double-helix point spread function versus
astigmatism and biplane.,” Applied physics letters, vol. 97, p. 161103, Oct. 2010.
222 References
[86] M. F. Juette, T. J. Gould, M. D. Lessard, M. J. Mlodzianoski, B. S. Nagpure, B. T.
Bennett, S. T. Hess, and J. Bewersdorf, “Three-dimensional sub-100 nm resolution
fluorescence microscopy of thick samples.,” Nature methods, vol. 5, pp. 527–9, June
2008.
[87] B. Huang, W. Wang, M. Bates, and X. Zhuang, “Three-dimensional super-resolution
imaging by stochastic optical reconstruction microscopy.,” Science (New York, N.Y.),
vol. 319, pp. 810–3, Feb. 2008.
[88] S. R. P. Pavani, M. A. Thompson, J. S. Biteen, S. J. Lord, N. Liu, R. J. Twieg, R. Pies-
tun, and W. E. Moerner, “Three-dimensional, single-molecule fluorescence imaging
beyond the diffraction limit by using a double-helix point spread function.,” Proceed-
ings of the National Academy of Sciences of the United States of America, vol. 106,
pp. 2995–9, Mar. 2009.
[89] G. Shtengel, J. A. Galbraith, C. G. Galbraith, J. Lippincott-Schwartz, J. M. Gillette,
S. Manley, R. Sougrat, C. M. Waterman, P. Kanchanawong, M. W. Davidson, R. D.
Fetter, and H. F. Hess, “Interferometric fluorescent super-resolution microscopy re-
solves 3D cellular ultrastructure.,” Proceedings of the National Academy of Sciences
of the United States of America, vol. 106, pp. 3125–30, Mar. 2009.
[90] F. Huang, T. M. P. Hartwich, F. E. Rivera-Molina, Y. Lin, W. C. Duim, J. J. Long, P. D.
Uchil, J. R. Myers, M. A. Baird, W. Mothes, M. W. Davidson, D. Toomre, and J. Be-
wersdorf, “Video-rate nanoscopy using sCMOS camera-specific single-molecule lo-
calization algorithms.,” Nature methods, vol. 10, pp. 653–8, 2013.
[91] J. Siegel, D. S. Elson, S. E. D. Webb, D. Parsons-Karavassilis, S. Lévêque-Fort, M. J.
Cole, M. J. Lever, P. M. W. French, M. A. A. Neil, R. Juskaitis, L. O. Sucharov, and
T. Wilson, “Whole-field five-dimensional fluorescence microscopy combining life-
time and spectral resolution with optical sectioning,” Optics Letters, vol. 26, p. 1338,
Sept. 2001.
[92] P. Blandin, S. Lévêque-Fort, S. Lécart, J. C. Cossec, M.-C. Potier, Z. Lenkei,
F. Druon, and P. Georges, “Time-gated total internal reflection fluorescence micros-
copy with a supercontinuum excitation source,” Applied Optics, vol. 48, p. 553, Jan.
2009.
[93] X. Michalet, R. A. Colyer, G. Scalia, S. Weiss, O. H. W. Siegmund, A. S. Trem-
sin, J. V. Vallerga, F. Villa, F. Guerrieri, I. Rech, A. Gulinatti, S. Tisa, F. Zappa,
M. Ghioni, and S. Cova, “New photon-counting detectors for single-molecule flu-
orescence spectroscopy and imaging.,” Proceedings of SPIE, vol. 8033, p. 803316,
May 2011.
[94] H. Shroff, C. G. Galbraith, J. A. Galbraith, H. White, J. Gillette, S. Olenych, M. W.
Davidson, and E. Betzig, “Dual-color superresolution imaging of genetically ex-
pressed probes within individual adhesion complexes.,” Proceedings of the National
Academy of Sciences of the United States of America, vol. 104, pp. 20308–13, Dec.
2007.
References 223
[95] F. V. Subach, G. H. Patterson, M. Renz, J. Lippincott-Schwartz, and V. V. Verkhusha,
“Bright monomeric photoactivatable red fluorescent protein for two-color super-
resolution sptPALM of live cells.,” Journal of the American Chemical Society,
vol. 132, pp. 6481–91, May 2010.
[96] R. Jungmann, M. S. Avendaño, J. B. Woehrstein, M. Dai, W. M. Shih, and P. Yin,
“Multiplexed 3D cellular super-resolution imaging with DNA-PAINT and Exchange-
PAINT.,” Nature methods, vol. 11, pp. 313–8, Mar. 2014.
[97] I. Testa, A. Schönle, C. v. Middendorff, C. Geisler, R. Medda, C. A. Wurm, A. C.
Stiel, S. Jakobs, M. Bossi, C. Eggeling, S. W. Hell, and A. Egner, “Nanoscale separa-
tion of molecular species based on their rotational mobility,” Optics Express, vol. 16,
p. 21093, Dec. 2008.
[98] M. P. Backlund, M. D. Lew, A. S. Backer, S. J. Sahl, and W. E. Moerner, “The
role of molecular dipole orientation in single-molecule fluorescence microscopy and
implications for super-resolution imaging.,” Chemphyschem : a European journal of
chemical physics and physical chemistry, vol. 15, pp. 587–99, Mar. 2014.
[99] A. S. Backer, M. P. Backlund, A. R. von Diezmann, S. J. Sahl, and W. E. Moerner, “A
bisected pupil for studying single-molecule orientational dynamics and its application
to three-dimensional super-resolution microscopy.,” Applied physics letters, vol. 104,
p. 193701, May 2014.
[100] L. Bell, A. Seshia, D. Lando, E. Laue, M. Palayret, S. F. Lee, and D. Klenerman,
“A microfluidic device for the hydrodynamic immobilisation of living fission yeast
cells for super-resolution imaging,” Sensors and Actuators B: Chemical, vol. 192,
pp. 36–41, Mar. 2014.
[101] T. J. Etheridge, R. L. Boulineau, A. Herbert, A. T. Watson, Y. Daigaku, J. Tucker,
S. George, P. Jönsson, M. Palayret, D. Lando, E. Laue, M. A. Osborne, D. Klener-
man, S. F. Lee, and A. M. Carr, “Quantification of DNA-associated proteins inside
eukaryotic cells using single-molecule localization microscopy.,” Nucleic acids re-
search, pp. gku726–, Aug. 2014.
[102] A. Edelstein, N. Amodaj, K. Hoover, R. Vale, and N. Stuurman, “Computer control
of microscopes using µManager.,” Current protocols in molecular biology / edited by
Frederick M. Ausubel ... [et al.], vol. Chapter 14, p. Unit14.20, Oct. 2010.
[103] Photometrics, “Evolve 512 datasheet,” http://www.photometrics.com/products/
datasheets/evolve_512.pdf, 2015.
[104] M. Hirsch, R. J. Wareham, M. L. Martin-Fernandez, M. P. Hobson, and D. J. Rolfe,
“A stochastic model for electron multiplication charge-coupled devices–from theory
to practice.,” PloS one, vol. 8, p. e53671, Jan. 2013.
[105] M. Tokunaga, N. Imamoto, and K. Sakata-Sogawa, “Highly inclined thin illumination
enables clear single-molecule imaging in cells.,” Nature methods, vol. 5, pp. 159–61,
Feb. 2008.
224 References
[106] J. Huisken, J. Swoger, F. Del Bene, J. Wittbrodt, and E. H. K. Stelzer, “Optical sec-
tioning deep inside live embryos by selective plane illumination microscopy.,” Sci-
ence (New York, N.Y.), vol. 305, pp. 1007–9, Aug. 2004.
[107] P. Annibale, M. Scarselli, M. Greco, and A. Radenovic, “Identification of the fac-
tors affecting co-localization precision for quantitative multicolor localization mi-
croscopy,” Optical Nanoscopy, vol. 1, p. 9, Nov. 2012.
[108] C. Geisler, T. Hotz, A. Schönle, S. W. Hell, A. Munk, and A. Egner, “Drift estima-
tion for single marker switching based imaging schemes.,” Optics express, vol. 20,
pp. 7274–89, Mar. 2012.
[109] Mad City Labs Incorporation, “Nano-Cyte, 3D stability for imaging,”
http://www.madcitylabs.com/pdfs/NanoCyteBrochure.pdf, 2015.
[110] D. Baddeley, D. Crossman, S. Rossberger, J. E. Cheyne, J. M. Montgomery, I. D.
Jayasinghe, C. Cremer, M. B. Cannell, and C. Soeller, “4D super-resolution micros-
copy with conventional fluorophores and single wavelength excitation in optically
thick cells and tissues.,” PloS one, vol. 6, p. e20645, Jan. 2011.
[111] M. Bates, G. T. Dempsey, K. H. Chen, and X. Zhuang, “Multicolor super-resolution
fluorescence imaging via multi-parameter fluorophore detection.,” Chemphyschem :
a European journal of chemical physics and physical chemistry, vol. 13, pp. 99–107,
Jan. 2012.
[112] L. S. Churchman, Z. Okten, R. S. Rock, J. F. Dawson, and J. A. Spudich, “Sin-
gle molecule high-resolution colocalization of Cy3 and Cy5 attached to macro-
molecules measures intramolecular distances through time.,” Proceedings of the Na-
tional Academy of Sciences of the United States of America, vol. 102, pp. 1419–23,
Feb. 2005.
[113] M. Lehmann, S. Rocha, B. Mangeat, F. Blanchet, H. Uji-I, J. Hofkens, and V. Piguet,
“Quantitative multicolor super-resolution microscopy reveals tetherin HIV-1 interac-
tion.,” PLoS pathogens, vol. 7, p. e1002456, Dec. 2011.
[114] A. D. Maude, “Interpolation–mainly for graph plotters,” The Computer Journal,
vol. 16, pp. 64–65, Jan. 1973.
[115] A. Goshtasby, “Image registration by local approximation methods,” Image and Vi-
sion Computing, vol. 6, pp. 255–261, Nov. 1988.
[116] S. Malkusch, U. Endesfelder, J. Mondry, M. Gelléri, P. J. Verveer, and M. Heile-
mann, “Coordinate-based colocalization analysis of single-molecule localization mi-
croscopy data.,” Histochemistry and cell biology, vol. 137, pp. 1–10, Jan. 2012.
[117] I. Koyama-Honda, K. Ritchie, T. Fujiwara, R. Iino, H. Murakoshi, R. S. Kasai, and
A. Kusumi, “Fluorescence imaging for monitoring the colocalization of two single
molecules in living cells.,” Biophysical journal, vol. 88, pp. 2126–36, Mar. 2005.
[118] J. Xu, J. Chang, Q. Yan, T. Dertinger, M. Bruchez, and S. Weiss, “Labeling Cytoso-
lic Targets in Live Cells with Blinking Probes.,” The journal of physical chemistry
letters, vol. 4, pp. 2138–2146, July 2013.
References 225
[119] V. Adam, “Phototransformable fluorescent proteins: which one for which applica-
tion?,” Histochemistry and cell biology, Feb. 2014.
[120] J. Ries, C. Kaplan, E. Platonova, H. Eghlidi, and H. Ewers, “A simple, versatile
method for GFP-based super-resolution microscopy via nanobodies.,” Nature meth-
ods, vol. 9, pp. 582–4, June 2012.
[121] M. Zhang, H. Chang, Y. Zhang, J. Yu, L. Wu, W. Ji, J. Chen, B. Liu, J. Lu, Y. Liu,
J. Zhang, P. Xu, and T. Xu, “Rational design of true monomeric and bright photoac-
tivatable fluorescent proteins.,” Nature methods, vol. 9, pp. 727–9, July 2012.
[122] F. V. Subach, V. N. Malashkevich, W. D. Zencheck, H. Xiao, G. S. Filonov, S. C.
Almo, and V. V. Verkhusha, “Photoactivation mechanism of PAmCherry based on
crystal structures of the protein in the dark and fluorescent states.,” Proceedings of the
National Academy of Sciences of the United States of America, vol. 106, pp. 21097–
102, Dec. 2009.
[123] S. A. McKinney, C. S. Murphy, K. L. Hazelwood, M. W. Davidson, and L. L. Looger,
“A bright and photostable photoconvertible fluorescent protein.,” Nature methods,
vol. 6, pp. 131–3, Feb. 2009.
[124] D. R. Whelan and T. D. M. Bell, “Image artifacts in Single Molecule Localization
Microscopy: why optimization of sample preparation protocols matters.,” Scientific
reports, vol. 5, p. 7924, Jan. 2015.
[125] B. Huang, M. Bates, and X. Zhuang, “Super-resolution fluorescence microscopy.,”
Annual review of biochemistry, vol. 78, pp. 993–1016, Jan. 2009.
[126] N. Durisic, L. Laparra-Cuervo, A. Sandoval-Álvarez, J. S. Borbely, and
M. Lakadamyali, “Single-molecule evaluation of fluorescent protein photoactivation
efficiency using an in vivo nanotemplate.,” Nature methods, vol. 11, pp. 156–62, Mar.
2014.
[127] S. F. Lee, Q. Vérolet, and A. Fürstenberg, “Improved super-resolution microscopy
with oxazine fluorophores in heavy water.,” Angewandte Chemie (International ed.
in English), vol. 52, pp. 8948–51, Aug. 2013.
[128] K. Klehs, C. Spahn, U. Endesfelder, S. F. Lee, A. Fürstenberg, and M. Heilemann,
“Increasing the brightness of cyanine fluorophores for single-molecule and super-
resolution imaging.,” Chemphyschem : a European journal of chemical physics and
physical chemistry, vol. 15, pp. 637–41, Mar. 2014.
[129] H. Shroff, C. Galbraith, J. Galbraith, and E. Betzig, “Live-cell photoactivated local-
ization microscopy of nanoscale adhesion dynamics,” Nature methods, vol. 5, no. 5,
pp. 417–423, 2008.
[130] D. Lando, U. Endesfelder, H. Berger, L. Subramanian, P. D. Dunne, J. McColl,
D. Klenerman, A. M. Carr, M. Sauer, R. C. Allshire, M. Heilemann, and E. D.
Laue, “Quantitative single-molecule microscopy reveals that CENP-A(Cnp1) depo-
sition occurs during G2 in fission yeast.,” Open biology, vol. 2, p. 120078, July 2012.
226 References
[131] E. M. Puchner, J. M. Walter, R. Kasper, B. Huang, and W. A. Lim, “Counting mo-
lecules in single organelles with superresolution microscopy allows tracking of the
endosome maturation trajectory.,” Proceedings of the National Academy of Sciences
of the United States of America, vol. 110, pp. 16015–20, Oct. 2013.
[132] A. Shivanandan, H. Deschout, M. Scarselli, and A. Radenovic, “Challenges in quan-
titative single molecule localization microscopy.,” FEBS letters, vol. 588, pp. 3595–
602, Oct. 2014.
[133] S. van de Linde, R. Kasper, M. Heilemann, and M. Sauer, “Photoswitching micros-
copy with standard fluorophores,” Applied Physics B, vol. 93, pp. 725–731, Oct.
2008.
[134] S. van de Linde, U. Endesfelder, A. Mukherjee, M. Schüttpelz, G. Wiebusch,
S. Wolter, M. Heilemann, and M. Sauer, “Multicolor photoswitching microscopy
for subdiffraction-resolution fluorescence imaging.,” Photochemical & photobiolog-
ical sciences : Official journal of the European Photochemistry Association and the
European Society for Photobiology, vol. 8, pp. 465–9, Apr. 2009.
[135] C. M. Hardaway, R. B. Badisa, and K. F. A. Soliman, “Effect of ascorbic acid and hy-
drogen peroxide on mouse neuroblastoma cells.,” Molecular medicine reports, vol. 5,
pp. 1449–52, June 2012.
[136] J. D. Campbell, M. Cole, B. Bunditrutavorn, and A. T. Vella, “Ascorbic acid is a
potent inhibitor of various forms of T cell apoptosis.,” Cellular immunology, vol. 194,
pp. 1–5, May 1999.
[137] H. Sakagami and K. Satoh, “Modulating factors of radical intensity and cytotoxic
activity of ascorbate (Review),” Jan. 1997.
[138] R. J. Schmidt, L. Y. Chung, A. M. Andrews, and T. D. Turner, “Toxicity of L-ascorbic
acid to L929 fibroblast cultures: relevance to biocompatibility testing of materials
for use in wound management.,” Journal of biomedical materials research, vol. 27,
pp. 521–30, Apr. 1993.
[139] W. C. Duim, B. Chen, J. Frydman, and W. E. Moerner, “Sub-diffraction imaging
of huntingtin protein aggregates by fluorescence blink-microscopy and atomic force
microscopy.,” Chemphyschem : a European journal of chemical physics and physical
chemistry, vol. 12, pp. 2387–90, Sept. 2011.
[140] R. Henriques, C. Griffiths, E. Hesper Rego, and M. M. Mhlanga, “PALM and
STORM: unlocking live-cell super-resolution.,” Biopolymers, vol. 95, pp. 322–31,
May 2011.
[141] S. van de Linde and M. Sauer, “How to switch a fluorophore: from undesired blinking
to controlled photoswitching.,” Chemical Society reviews, vol. 43, pp. 1076–87, Feb.
2014.
[142] L. Goldman, “Quantitative analysis of a fully generalized four-state kinetic scheme.,”
Biophysical journal, vol. 91, pp. 173–8, July 2006.
References 227
[143] C. Flors, J.-i. Hotta, H. Uji-i, P. Dedecker, R. Ando, H. Mizuno, A. Miyawaki, and
J. Hofkens, “A stroboscopic approach for fast photoactivation-localization micros-
copy with Dronpa mutants.,” Journal of the American Chemical Society, vol. 129,
pp. 13970–7, Nov. 2007.
[144] P. Annibale, M. Scarselli, A. Kodiyan, and A. Radenovic, “Photoactivatable Fluo-
rescent Protein mEos2 Displays Repeated Photoactivation after a Long-Lived Dark
State in the Red Photoconverted Form,” The Journal of Physical Chemistry Letters,
vol. 1, pp. 1506–1510, May 2010.
[145] S.-H. S.-H. Lee, J. Y. Shin, A. Lee, and C. Bustamante, “Counting single photoacti-
vatable fluorescent molecules by photoactivated localization microscopy (PALM).,”
Proceedings of the National Academy of Sciences of the United States of America,
vol. 109, pp. 17436–17441, Oct. 2012.
[146] J. Gunzenhäuser, N. Olivier, T. Pengo, and S. Manley, “Quantitative super-resolution
imaging reveals protein stoichiometry and nanoscale morphology of assembling HIV-
Gag virions.,” Nano letters, vol. 12, pp. 4705–10, Oct. 2012.
[147] A. Small and S. Stahlheber, “Fluorophore localization algorithms for super-resolution
microscopy,” Nature Methods, vol. 11, pp. 267–279, 2014.
[148] EPFL, “Single-Molecule Localization Microscopy Challenge ISBI 2013,”
http://bigwww.epfl.ch/smlm/software/, 2015.
[149] A. Herbert, “Peak Fit, a Single-molecule Plugins,” http://www.sussex.ac.uk/gdsc/ in-
tranet/microscopy/imagej/smlm_plugins, 2014.
[150] C. A. Schneider, W. S. Rasband, and K. W. Eliceiri, “NIH Image to ImageJ: 25 years
of image analysis,” 2012.
[151] S. W. Smith, “Linear Image Processing,” in The Scientist and Engineer’s Guide to
Digital Signal Processing, ch. 24, pp. 397–422, 1997.
[152] A. V. Abraham, S. Ram, J. Chao, E. S. Ward, and R. J. Ober, “Quantitative study of
single molecule location estimation techniques.,” Optics express, vol. 17, pp. 23352–
73, Dec. 2009.
[153] P. Winckler, L. Lartigue, G. Giannone, F. De Giorgi, F. Ichas, J.-B. Sibarita, B. Lou-
nis, and L. Cognet, “Identification and super-resolution imaging of ligand-activated
receptor dimers in live cells.,” Scientific reports, vol. 3, p. 2387, Jan. 2013.
[154] S. A. Shelby, D. Holowka, B. Baird, and S. L. Veatch, “Distinct stages of stimulated
fcεri receptor clustering and immobilization are identified through superresolution
imaging,” Biophysical Journal, vol. 105, pp. 2343–2354, 2013.
[155] P. Sengupta, T. Jovanovic-Talisman, and J. Lippincott-Schwartz, “Quantifying spa-
tial organization in point-localization superresolution images using pair correlation
analysis.,” Nature protocols, vol. 8, pp. 345–54, Feb. 2013.
228 References
[156] S. Wolter, A. Löschberger, T. Holm, S. Aufmkolk, M.-C. Dabauvalle, S. van de Linde,
and M. Sauer, “rapidSTORM: accurate, fast open-source software for localization
microscopy.,” Nature methods, vol. 9, pp. 1040–1, Nov. 2012.
[157] J. Hoshen and R. Kopelman, “Percolation and cluster distribution. I. Cluster multiple
labeling technique and critical concentration algorithm,” Physical Review B, vol. 14,
pp. 3438–3445, Oct. 1976.
[158] M. Ester, H. Kriegel, J. Sander, and X. Xu, “A density-based algorithm for discover-
ing clusters in large spatial databases with noise.,” KDD, 1996.
[159] R. A. Finkel and J. L. Bentley, “Quad trees a data structure for retrieval on composite
keys,” Acta Informatica, vol. 4, no. 1, pp. 1–9, 1974.
[160] D. Baddeley, M. B. Cannell, and C. Soeller, “Visualization of localization microscopy
data.,” Microscopy and microanalysis : the official journal of Microscopy Society of
America, Microbeam Analysis Society, Microscopical Society of Canada, vol. 16,
pp. 64–72, Feb. 2010.
[161] K. I. Mortensen, L. S. Churchman, J. A. Spudich, and H. Flyvbjerg, “Optimized
localization analysis for single-molecule tracking and super-resolution microscopy.,”
Nature methods, vol. 7, pp. 377–381, 2010.
[162] R. J. Ober, S. Ram, and E. S. Ward, “Localization accuracy in single-molecule mi-
croscopy.,” Biophysical journal, vol. 86, pp. 1185–200, Feb. 2004.
[163] D. J. Rowland and J. S. Biteen, “Top-hat and asymmetric Gaussian-based fitting func-
tions for quantifying directional single-molecule motion.,” Chemphyschem : a Euro-
pean journal of chemical physics and physical chemistry, vol. 15, pp. 712–20, Mar.
2014.
[164] R. B. Murphy, “On the meaning of precision and accuracy,” Precision measurement
and calibration: Selected NBS papers on statistical concepts and procedures, NBS
Special Publication, vol. 300, pp. 357–360, 1969.
[165] J. C. Waters, “Accuracy and precision in quantitative fluorescence microscopy.,” The
Journal of cell biology, vol. 185, pp. 1135–48, June 2009.
[166] N. Banterle, K. H. Bui, E. A. Lemke, and M. Beck, “Fourier ring correlation as a
resolution criterion for super-resolution microscopy.,” Journal of structural biology,
vol. 183, pp. 363–7, Sept. 2013.
[167] U. Endesfelder, S. Malkusch, F. Fricke, and M. Heilemann, “A simple method to esti-
mate the average localization precision of a single-molecule localization microscopy
experiment.,” Histochemistry and cell biology, vol. 141, pp. 629–38, June 2014.
[168] D. R. Rines, D. Thomann, J. F. Dorn, P. Goodwin, and P. K. Sorger, “Live cell imag-
ing of yeast.,” Cold Spring Harbor protocols, vol. 2011, pp. pdb.top065482–, Sept.
2011.
References 229
[169] J. A. Waddle, T. S. Karpova, R. H. Waterston, and J. A. Cooper, “Movement of
cortical actin patches in yeast.,” The Journal of cell biology, vol. 132, pp. 861–70,
Mar. 1996.
[170] P. Nurse, “Genetic control of cell size at cell division in yeast.,” Nature, vol. 256,
pp. 547–51, Aug. 1975.
[171] T. Yeo, S. Ong, and R. Sinniah, “Autofocusing for tissue microscopy,” Image and
Vision Computing, vol. 11, pp. 629–639, Dec. 1993.
[172] M. Held, M. H. A. Schmitz, B. Fischer, T. Walter, B. Neumann, M. H. Olma, M. Pe-
ter, J. Ellenberg, and D. W. Gerlich, “CellCognition: time-resolved phenotype an-
notation in high-throughput live cell imaging.,” Nature methods, vol. 7, pp. 747–54,
Sept. 2010.
[173] A. D. Edelstein, M. A. Tsuchida, N. Amodaj, H. Pinkard, R. D. Vale, and N. Stu-
urman, “Advanced methods of microscope control using µManager software,” Nov.
2014.
[174] S. J. Holden, T. Pengo, K. L. Meibom, C. Fernandez Fernandez, J. Collier, and
S. Manley, “High throughput 3D super-resolution microscopy reveals Caulobacter
crescentus in vivo Z-ring organization.,” Proceedings of the National Academy of
Sciences of the United States of America, vol. 111, pp. 4566–71, Mar. 2014.
[175] L. Zhu, W. Zhang, D. Elnatan, and B. Huang, “Faster STORM using compressed
sensing.,” Nature methods, vol. 9, pp. 721–3, July 2012.
[176] N. Brede and M. Lakadamyali, “GraspJ: an open source, real-time analysis package
for super-resolution imaging,” Optical Nanoscopy, vol. 1, p. 11, Dec. 2012.
[177] A. Kechkar, D. Nair, M. Heilemann, D. Choquet, and J. B. Sibarita, “Real-Time Anal-
ysis and Visualization for Single-Molecule Based Super-Resolution Microscopy,”
PLoS ONE, vol. 8, no. 4, 2013.
[178] X. Chen, L. Ren, Y. Qiu, and H. Liu, “New method for determining the depth of field
of microscope systems.,” Applied optics, vol. 50, pp. 5524–33, Oct. 2011.
[179] Y. Hiraoka, J. Sedat, and D. Agard, “The use of a charge-coupled device for quanti-
tative optical microscopy of biological structures,” Science, vol. 238, pp. 36–41, Oct.
1987.
[180] J. G. McNally, T. Karpova, J. Cooper, and J. A. Conchello, “Three-dimensional imag-
ing by deconvolution microscopy.,” Methods (San Diego, Calif.), vol. 19, pp. 373–85,
Nov. 1999.
[181] J.-B. Sibarita, “Deconvolution microscopy.,” Advances in biochemical engineer-
ing/biotechnology, vol. 95, pp. 201–43, Jan. 2005.
[182] D. S. C. Biggs, “3D deconvolution microscopy.,” Current protocols in cytometry / ed-
itorial board, J. Paul Robinson, managing editor ... [et al.], vol. Chapter 12, pp. Unit
12.19.1–20, Apr. 2010.
230 References
[183] R. Henriques, M. Lelek, E. F. Fornasiero, F. Valtorta, C. Zimmer, and M. M. Mh-
langa, “QuickPALM: 3D real-time photoactivation nanoscopy image processing in
ImageJ.,” Nature methods, vol. 7, pp. 339–40, May 2010.
[184] M. Ovesný, P. Krˇížek, J. Borkovec, Z. Svindrych, and G. M. Hagen, “Thunder-
STORM: a comprehensive ImageJ plug-in for PALM and STORM data analysis and
super-resolution imaging.,” Bioinformatics (Oxford, England), pp. 1–2, Apr. 2014.
[185] R. Starr, S. Stahlheber, and A. Small, “Fast maximum likelihood algorithm for local-
ization of fluorescent molecules.,” Optics letters, vol. 37, pp. 413–5, Feb. 2012.
[186] N. Bone, J. B. Millar, T. Toda, and J. Armstrong, “Regulated vacuole fusion and
fission in Schizosaccharomyces pombe: an osmotic response dependent on MAP
kinases.,” Current biology : CB, vol. 8, pp. 135–44, Jan. 1998.
[187] S. A. Mutch, B. S. Fujimoto, C. L. Kuyper, J. S. Kuo, S. M. Bajjalieh, and D. T.
Chiu, “Deconvolving single-molecule intensity distributions for quantitative micros-
copy measurements.,” Biophysical journal, vol. 92, pp. 2926–43, Apr. 2007.
[188] H. Deschout, K. Neyts, and K. Braeckmans, “The influence of movement on the lo-
calization precision of sub-resolution particles in fluorescence microscopy,” Journal
of Biophotonics, vol. 5, pp. 97–109, 2012.
[189] J. C. Sun, S. Lopez-Verges, C. C. Kim, J. L. DeRisi, and L. L. Lanier, “NK cells
and immune "memory".,” Journal of immunology (Baltimore, Md. : 1950), vol. 186,
pp. 1891–7, Feb. 2011.
[190] K. Murphy, P. Travers, and M. Walport, Janeway’s Immunobiology, vol. 7. 2008.
[191] L. B. Nicholson and D. C. Wraith, “T-cell receptor degeneracy: the dog that did not
bark. Adaptation of the self-reactive T-cell response to limit autoimmune disease.,”
Molecular immunology, vol. 40, pp. 997–1002, Feb. 2004.
[192] G. P. Morris and P. M. Allen, “How the TCR balances sensitivity and specificity for
the recognition of self and pathogens.,” Nature immunology, vol. 13, pp. 121–8, Feb.
2012.
[193] Y. Sykulev, M. Joo, I. Vturina, T. J. Tsomides, and H. N. Eisen, “Evidence that a
single peptide-MHC complex on a target cell can elicit a cytolytic T cell response.,”
Immunity, vol. 4, pp. 565–71, June 1996.
[194] D. J. Irvine, M. A. Purbhoo, M. Krogsgaard, and M. M. Davis, “Direct observation
of ligand recognition by T cells.,” Nature, vol. 419, pp. 845–9, Oct. 2002.
[195] C. V. Harding and E. R. Unanue, “Quantitation of antigen-presenting cell MHC class
II/peptide complexes necessary for T-cell stimulation.,” Nature, vol. 346, pp. 574–6,
Aug. 1990.
[196] B. N. Manz, B. L. Jackson, R. S. Petit, M. L. Dustin, and J. Groves, “T-cell triggering
thresholds are modulated by the number of antigen within individual T-cell receptor
clusters.,” Proceedings of the National Academy of Sciences of the United States of
America, vol. 108, pp. 9089–94, May 2011.
References 231
[197] A. J. T. George, J. Stark, and C. Chan, “Understanding specificity and sensitivity of
T-cell recognition.,” Trends in immunology, vol. 26, pp. 653–9, Dec. 2005.
[198] W. A. Comrie, S. Li, S. Boyle, and J. K. Burkhardt, “The dendritic cell cytoskele-
ton promotes T cell adhesion and activation by constraining ICAM-1 mobility,” The
Journal of Cell Biology, vol. 208, pp. 457–473, Feb. 2015.
[199] A. S. Shaw and M. L. Dustin, “Making the T Cell Receptor Go the Distance: A
Topological View of T Cell Activation,” Immunity, vol. 6, pp. 361–369, Apr. 1997.
[200] B.-C. Chen, W. R. Legant, K. Wang, L. Shao, D. E. Milkie, M. W. Davidson, C. Jane-
topoulos, X. S. Wu, J. A. Hammer, Z. Liu, B. P. English, Y. Mimori-Kiyosue, D. P.
Romero, A. T. Ritter, J. Lippincott-Schwartz, L. Fritz-Laylin, R. D. Mullins, D. M.
Mitchell, J. N. Bembenek, A.-C. Reymann, R. Bohme, S. W. Grill, J. T. Wang,
G. Seydoux, U. S. Tulu, D. P. Kiehart, and E. Betzig, “Lattice light-sheet micros-
copy: Imaging molecules to embryos at high spatiotemporal resolution,” Science,
vol. 346, pp. 1257998–1257998, Oct. 2014.
[201] M. M. Davis, “A new trigger for T cells.,” Cell, vol. 110, pp. 285–7, Aug. 2002.
[202] P. A. van der Merwe and O. Dushek, “Mechanisms for T cell receptor triggering.,”
Nature reviews. Immunology, vol. 11, pp. 47–55, 2011.
[203] M. L. Dustin and J. T. Groves, “Receptor signaling clusters in the immune synapse.,”
Annual review of biophysics, vol. 41, pp. 543–56, Jan. 2012.
[204] M. L. Dustin, “Cell adhesion molecules and actin cytoskeleton at immune synapses
and kinapses.,” Current opinion in cell biology, vol. 19, pp. 529–33, Oct. 2007.
[205] A. T. Ritter, K. L. Angus, and G. M. Griffiths, “The role of the cytoskeleton at the
immunological synapse.,” Immunological reviews, vol. 256, pp. 107–17, Nov. 2013.
[206] G. M. Griffiths, A. Tsun, and J. C. Stinchcombe, “The immunological synapse: a
focal point for endocytosis and exocytosis.,” The Journal of cell biology, vol. 189,
pp. 399–406, May 2010.
[207] C. R. Monks, B. A. Freiberg, H. Kupfer, N. Sciaky, and A. Kupfer, “Three-
dimensional segregation of supramolecular activation clusters in T cells.,” Nature,
vol. 395, pp. 82–6, Sept. 1998.
[208] J. C. Stinchcombe and G. M. Griffiths, “Communication, the centrosome and the
immunological synapse.,” Philosophical transactions of the Royal Society of London.
Series B, Biological sciences, vol. 369, Sept. 2014.
[209] J. C. Brown and W. W. Newcomb, “Herpesvirus capsid assembly: insights from struc-
tural analysis.,” Current opinion in virology, vol. 1, pp. 142–9, Aug. 2011.
[210] G. D. Rak, E. M. Mace, P. P. Banerjee, T. Svitkina, and J. S. Orange, “Natural killer
cell lytic granule secretion occurs through a pervasive actin network at the immune
synapse.,” PLoS biology, vol. 9, p. e1001151, Sept. 2011.
232 References
[211] S. J. Davis and P. A. van der Merwe, “Lck and the nature of the T cell receptor
trigger.,” Trends in immunology, vol. 32, pp. 1–5, Jan. 2011.
[212] O. Ballek, J. Valecˇka, J. Manning, and D. Filipp, “The pool of preactivated Lck in
the initiation of T-cell signaling: a critical re-evaluation of the Lck standby model.,”
Immunology and cell biology, Nov. 2014.
[213] M. A. Lemmon and J. Schlessinger, “Cell signaling by receptor tyrosine kinases.,”
Cell, vol. 141, pp. 1117–34, June 2010.
[214] J. Yang and M. Reth, “Oligomeric organization of the B-cell antigen receptor on
resting cells.,” Nature, vol. 467, pp. 465–9, Sept. 2010.
[215] R. Schwinzer, R. A. Franklin, J. Domenico, H. Renz, and E. W. Gelfand, “Mon-
oclonal antibodies directed to different epitopes in the CD3-TCR complex induce
different states of competence in resting human T cells.,” Journal of immunology
(Baltimore, Md. : 1950), vol. 148, pp. 1322–8, Mar. 1992.
[216] L. Wooldridge, A. Lissina, D. K. Cole, H. A. van den Berg, D. A. Price, and A. K.
Sewell, “Tricks with tetramers: how to get the most from multimeric peptide-MHC.,”
Immunology, vol. 126, pp. 147–64, Feb. 2009.
[217] A. Trautmann and C. Randriamampita, “Initiation of TCR signalling revisited,”
Trends in Immunology, vol. 24, pp. 425–428, Aug. 2003.
[218] M. Krogsgaard, Q.-J. Li, C. Sumen, J. B. Huppa, M. Huse, and M. M. Davis, “Ago-
nist/endogenous peptide-MHC heterodimers drive T cell activation and sensitivity.,”
Nature, vol. 434, pp. 238–43, Mar. 2005.
[219] R. M. Locksley, S. L. Reiner, F. Hatam, D. R. Littman, and N. Killeen, “Helper T
cells without CD4: control of leishmaniasis in CD4-deficient mice.,” Science (New
York, N.Y.), vol. 261, pp. 1448–51, Sept. 1993.
[220] P. A. van der Merwe and S.-P. Cordoba, “Late arrival: recruiting coreceptors to the T
cell receptor complex.,” Immunity, vol. 34, pp. 1–3, Jan. 2011.
[221] N. Jiang, J. Huang, L. J. Edwards, B. Liu, Y. Zhang, C. D. Beal, B. D. Evavold,
and C. Zhu, “Two-stage cooperative T cell receptor-peptide major histocompatibility
complex-CD8 trimolecular interactions amplify antigen discrimination.,” Immunity,
vol. 34, pp. 13–23, Jan. 2011.
[222] R. Varma, G. Campi, T. Yokosuka, T. Saito, and M. L. Dustin, “T cell receptor-
proximal signals are sustained in peripheral microclusters and terminated in the cen-
tral supramolecular activation cluster.,” Immunity, vol. 25, pp. 117–27, July 2006.
[223] T. Yokosuka, K. Sakata-Sogawa, W. Kobayashi, M. Hiroshima, A. Hashimoto-Tane,
M. Tokunaga, M. L. Dustin, and T. Saito, “Newly generated T cell receptor micro-
clusters initiate and sustain T cell activation by recruitment of Zap70 and SLP-76.,”
Nature immunology, vol. 6, pp. 1253–62, Dec. 2005.
[224] W. W. A. Schamel and B. Alarcón, “Organization of the resting TCR in nanoscale
oligomers,” Immunological Reviews, vol. 251, pp. 13–20, Jan. 2013.
References 233
[225] G. Lebon, T. Warne, and C. G. Tate, “Agonist-bound structures of G protein-coupled
receptors.,” Current opinion in structural biology, vol. 22, pp. 482–90, Aug. 2012.
[226] L. Kjer-Nielsen, M. A. Dunstone, L. Kostenko, L. K. Ely, T. Beddoe, N. A. Mifsud,
A. W. Purcell, A. G. Brooks, J. McCluskey, and J. Rossjohn, “Crystal structure of the
human T cell receptor CD3 epsilon gamma heterodimer complexed to the therapeutic
mAb OKT3.,” Proceedings of the National Academy of Sciences of the United States
of America, vol. 101, pp. 7675–80, May 2004.
[227] T. Beddoe, Z. Chen, C. S. Clements, L. K. Ely, S. R. Bushell, J. P. Vivian, L. Kjer-
Nielsen, S. S. Pang, M. A. Dunstone, Y. C. Liu, W. A. Macdonald, M. A. Perugini,
M. C. J. Wilce, S. R. Burrows, A. W. Purcell, T. Tiganis, S. P. Bottomley, J. Mc-
Cluskey, and J. Rossjohn, “Antigen ligation triggers a conformational change within
the constant domain of the alphabeta T cell receptor.,” Immunity, vol. 30, pp. 777–88,
June 2009.
[228] M. G. Rudolph, R. L. Stanfield, and I. A. Wilson, “How TCRs bind MHCs, peptides,
and coreceptors.,” Annual review of immunology, vol. 24, pp. 419–66, Jan. 2006.
[229] Z. Ma, P. A. Janmey, and T. H. Finkel, “The receptor deformation model of TCR trig-
gering.,” FASEB journal : official publication of the Federation of American Societies
for Experimental Biology, vol. 22, pp. 1002–8, Apr. 2008.
[230] Z. J. Sun, K. S. Kim, G. Wagner, and E. L. Reinherz, “Mechanisms contribut-
ing to T cell receptor signaling and assembly revealed by the solution structure of
an ectodomain fragment of the CD3 epsilon gamma heterodimer.,” Cell, vol. 105,
pp. 913–23, June 2001.
[231] C. Xu, E. Gagnon, M. E. Call, J. R. Schnell, C. D. Schwieters, C. V. Carman,
J. J. Chou, and K. W. Wucherpfennig, “Regulation of T cell receptor activation by
dynamic membrane binding of the CD3epsilon cytoplasmic tyrosine-based motif.,”
Cell, vol. 135, pp. 702–13, Nov. 2008.
[232] B. T. Marshall, M. Long, J. W. Piper, T. Yago, R. P. McEver, and C. Zhu, “Direct
observation of catch bonds involving cell-adhesion molecules.,” Nature, vol. 423,
pp. 190–3, May 2003.
[233] C. D. Buckley, J. Tan, K. L. Anderson, D. Hanein, N. Volkmann, W. I. Weis, W. J.
Nelson, and A. R. Dunn, “The minimal cadherin-catenin complex binds to actin fila-
ments under force,” Science, vol. 346, pp. 1254211–1254211, Oct. 2014.
[234] S. J. Davis and P. A. van der Merwe, “The kinetic-segregation model: TCR triggering
and beyond.,” Nature immunology, vol. 7, pp. 803–9, Aug. 2006.
[235] J. Lin and A. Weiss, “The tyrosine phosphatase CD148 is excluded from the im-
munologic synapse and down-regulates prolonged T cell signaling,” Journal of Cell
Biology, vol. 162, pp. 673–682, Aug. 2003.
[236] C. Irles, A. Symons, F. Michel, T. R. Bakker, P. A. van der Merwe, and O. Acuto,
“CD45 ectodomain controls interaction with GEMs and Lck activity for optimal TCR
signaling.,” Nature immunology, vol. 4, pp. 189–97, Feb. 2003.
234 References
[237] K. Choudhuri, D. Wiseman, M. H. Brown, K. Gould, and P. A. van der Merwe, “T-
cell receptor triggering is critically dependent on the dimensions of its peptide-MHC
ligand.,” Nature, vol. 436, pp. 578–82, July 2005.
[238] K. Choudhuri, M. Parker, A. Milicic, D. K. Cole, M. K. Shaw, A. K. Sewell,
G. Stewart-Jones, T. Dong, K. G. Gould, and P. A. van der Merwe, “Peptide-major
histocompatibility complex dimensions control proximal kinase-phosphatase balance
during T cell activation.,” The Journal of biological chemistry, vol. 284, pp. 26096–
105, Sept. 2009.
[239] Z. Ma, K. A. Sharp, P. A. Janmey, and T. H. Finkel, “Surface-anchored monomeric
agonist pMHCs alone trigger TCR with high sensitivity.,” PLoS biology, vol. 6,
p. e43, Feb. 2008.
[240] C. Bluemel, S. Hausmann, P. Fluhr, M. Sriskandarajah, W. B. Stallcup, P. A.
Baeuerle, and P. Kufer, “Epitope distance to the target cell membrane and antigen size
determine the potency of T cell-mediated lysis by BiTE antibodies specific for a large
melanoma surface antigen.,” Cancer immunology, immunotherapy : CII, vol. 59,
pp. 1197–209, Aug. 2010.
[241] S. E. James, P. D. Greenberg, M. C. Jensen, Y. Lin, J. Wang, B. G. Till, A. A.
Raubitschek, S. J. Forman, and O. W. Press, “Antigen sensitivity of CD22-specific
chimeric TCR is modulated by target epitope distance from the cell membrane.,”
Journal of immunology (Baltimore, Md. : 1950), vol. 180, pp. 7028–38, May 2008.
[242] R. L. Contento, S. Campello, A. E. Trovato, E. Magrini, F. Anselmi, and A. Viola,
“Adhesion shapes T cells for prompt and sustained T-cell receptor signalling.,” The
EMBO journal, vol. 29, pp. 4035–47, Dec. 2010.
[243] J. R. James and R. D. Vale, “Biophysical mechanism of T-cell receptor triggering in
a reconstituted system.,” Nature, vol. 487, pp. 64–9, July 2012.
[244] J. J. Illingworth and P. Anton van der Merwe, “Dissecting T-cell activation with high-
resolution live-cell microscopy.,” Immunology, vol. 135, pp. 198–206, Mar. 2012.
[245] S. J. Davis and P. van der Merwe, “The immunological synapse: required for T cell
receptor signalling or directing T cell effector function?,” Current Biology, vol. 11,
pp. R289–R290, Apr. 2001.
[246] S. Cemerski, J. Das, E. Giurisato, M. A. Markiewicz, P. M. Allen, A. K. Chakraborty,
and A. S. Shaw, “The balance between T cell receptor signaling and degradation at
the center of the immunological synapse is determined by antigen quality.,” Immunity,
vol. 29, pp. 414–22, Sept. 2008.
[247] J. L. Rodríguez-Fernández, L. Riol-Blanco, and C. Delgado-Martín, “What is an
immunological synapse?,” Microbes and infection / Institut Pasteur, vol. 12, pp. 438–
45, June 2010.
[248] B. F. Lillemeier, M. a. Mörtelmaier, M. B. Forstner, J. B. Huppa, J. T. Groves, and
M. M. Davis, “TCR and Lat are expressed on separate protein islands on T cell mem-
branes and concatenate during activation.,” Nature immunology, vol. 11, pp. 90–6,
Jan. 2010.
References 235
[249] W. W. A. Schamel, I. Arechaga, R. M. Risueño, H. M. van Santen, P. Cabezas,
C. Risco, J. M. Valpuesta, and B. Alarcón, “Coexistence of multivalent and mono-
valent TCRs explains high sensitivity and wide range of response.,” The Journal of
experimental medicine, vol. 202, pp. 493–503, Aug. 2005.
[250] J. R. James, S. S. White, R. W. Clarke, A. M. Johansen, P. D. Dunne, D. L. Sleep,
W. J. Fitzgerald, S. J. Davis, and D. Klenerman, “Single-molecule level analysis of
the subunit composition of the T cell receptor on live T cells.,” Proceedings of the
National Academy of Sciences of the United States of America, vol. 104, pp. 17662–
7, Nov. 2007.
[251] P. D. Dunne, R. A. Fernandes, J. McColl, J. W. Yoon, J. R. James, S. J. Davis, and
D. Klenerman, “DySCo: quantitating associations of membrane proteins using two-
color single-molecule tracking.,” Biophysical journal, vol. 97, pp. L5–7, Aug. 2009.
[252] J. R. James, J. McColl, M. I. Oliveira, P. D. Dunne, E. Huang, A. Jansson, P. Nilsson,
D. L. Sleep, C. M. Gonçalves, S. H. Morgan, J. H. Felce, R. Mahen, R. A. Fernandes,
A. M. Carmo, D. Klenerman, and S. J. Davis, “The T cell receptor triggering appa-
ratus is composed of monovalent or monomeric proteins.,” The Journal of biological
chemistry, vol. 286, pp. 31993–2001, Sept. 2011.
[253] M. A. Purbhoo, H. Liu, S. Oddos, D. M. Owen, M. A. A. Neil, S. V. Pageon, P. M. W.
French, C. E. Rudd, and D. M. Davis, “Dynamics of subsynaptic vesicles and surface
microclusters at the immunological synapse.,” Science signaling, vol. 3, p. ra36, Jan.
2010.
[254] I. Parmryd and B. Onfelt, “Consequences of membrane topography.,” The FEBS jour-
nal, vol. 280, pp. 2775–84, June 2013.
[255] D. M. Owen, C. Rentero, J. Rossy, A. Magenau, D. Williamson, M. Rodriguez, and
K. Gaus, “PALM imaging and cluster analysis of protein heterogeneity at the cell
surface.,” Journal of biophotonics, vol. 3, pp. 446–54, July 2010.
[256] P. A. van der Merwe, P. D. Dunne, D. Klenerman, and S. J. Davis, “Taking T cells
beyond the diffraction limit.,” Nature immunology, vol. 11, pp. 51–2, Jan. 2010.
[257] D. J. Williamson, D. M. Owen, J. Rossy, A. Magenau, M. Wehrmann, J. J. Gooding,
and K. Gaus, “Pre-existing clusters of the adaptor Lat do not participate in early T
cell signaling events.,” Nature immunology, vol. 12, pp. 655–62, July 2011.
[258] Y. Neve-Oz, Y. Razvag, J. Sajman, and E. Sherman, “Mechanisms of localized acti-
vation of the T cell antigen receptor inside clusters.,” Biochimica et biophysica acta,
Oct. 2014.
[259] V. Adam, K. Nienhaus, D. Bourgeois, and G. U. Nienhaus, “Structural basis of en-
hanced photoconversion yield in green fluorescent protein-like protein Dendra2.,”
Biochemistry, vol. 48, pp. 4905–15, June 2009.
[260] M. F. Garcia-Parajo, M. Koopman, E. M. van Dijk, V. Subramaniam, and N. F. van
Hulst, “The nature of fluorescence emission in the red fluorescent protein DsRed,
revealed by single-molecule detection.,” Proceedings of the National Academy of
Sciences of the United States of America, vol. 98, pp. 14392–7, Dec. 2001.
236 References
[261] M. H. Ulbrich and E. Y. Isacoff, “Subunit counting in membrane-bound proteins.,”
Nature methods, vol. 4, pp. 319–21, Apr. 2007.
[262] A. T. Watson, P. Werler, and A. M. Carr, “Regulation of gene expression at the fission
yeast Schizosaccharomyces pombe urg1 locus.,” Gene, vol. 484, pp. 75–85, Sept.
2011.
[263] P. Annibale, S. Vanni, M. Scarselli, U. Rothlisberger, and A. Radenovic, “Quanti-
tative photo activated localization microscopy: unraveling the effects of photoblink-
ing.,” PloS one, vol. 6, p. e22678, Jan. 2011.
[264] C. Coltharp, R. P. Kessler, and J. Xiao, “Accurate construction of photoactivated
localization microscopy (PALM) images for quantitative measurements.,” PloS one,
vol. 7, p. e51725, Jan. 2012.
[265] B. Chance, B. Schoener, R. Oshino, F. Itshak, and Y. Nakase, “Oxidation-reduction
ratio studies of mitochondria in freeze-trapped samples. NADH and flavoprotein flu-
orescence signals.,” The Journal of biological chemistry, vol. 254, pp. 4764–71, June
1979.
[266] S. Huang, A. A. Heikal, and W. W. Webb, “Two-photon fluorescence spectroscopy
and microscopy of NAD(P)H and flavoprotein.,” Biophysical journal, vol. 82,
pp. 2811–25, May 2002.
[267] K. A. K. Tanaka, K. G. N. Suzuki, Y. M. Shirai, S. T. Shibutani, M. S. H. Miyahara,
H. Tsuboi, M. Yahara, A. Yoshimura, S. Mayor, T. K. Fujiwara, and A. Kusumi,
“Membrane molecules mobile even after chemical fixation.,” Nature methods, vol. 7,
pp. 865–6, Nov. 2010.
[268] V. Vogel and M. Sheetz, “Local force and geometry sensing regulate cell functions.,”
Nature reviews. Molecular cell biology, vol. 7, pp. 265–75, Apr. 2006.
[269] R. S. O’Connor, X. Hao, K. Shen, K. Bashour, T. Akimova, W. W. Hancock, L. C.
Kam, and M. C. Milone, “Substrate rigidity regulates human T cell activation and
proliferation.,” Journal of immunology (Baltimore, Md. : 1950), vol. 189, pp. 1330–
9, Aug. 2012.
[270] H. H. Chang, M. Hemberg, M. Barahona, D. E. Ingber, and S. Huang,
“Transcriptome-wide noise controls lineage choice in mammalian progenitor cells.,”
Nature, vol. 453, pp. 544–7, May 2008.
[271] L. Cai, N. Friedman, and X. S. Xie, “Stochastic protein expression in individual cells
at the single molecule level.,” Nature, vol. 440, pp. 358–62, Mar. 2006.
[272] A. Raj and A. van Oudenaarden, “Nature, nurture, or chance: stochastic gene expres-
sion and its consequences.,” Cell, vol. 135, pp. 216–26, Oct. 2008.
[273] G. C. Rollins, J. Y. Shin, C. Bustamante, and S. Pressé, “Stochastic approach to
the molecular counting problem in superresolution microscopy.,” Proceedings of the
National Academy of Sciences of the United States of America, vol. 112, pp. E110–8,
Dec. 2014.
References 237
[274] R. Iinuma, Y. Ke, R. Jungmann, T. Schlichthaerle, J. B. Woehrstein, and P. Yin, “Poly-
hedra self-assembled from DNA tripods and characterized with 3D DNA-PAINT.,”
Science (New York, N.Y.), vol. 344, pp. 65–9, Apr. 2014.
[275] K. Zimmermann, T. Liechti, A. Haas, M. Rehr, A. Trkola, H. F. Gunthard, and A. Ox-
enius, “The Orientation of HIV-1 gp120 Binding to the CD4 Receptor Differentially
Modulates CD4+ T Cell Activation,” The Journal of Immunology, vol. 194, pp. 637–
649, Dec. 2014.
[276] B. M. Carreno and M. Collins, “The B7 family of ligands and its receptors: new
pathways for costimulation and inhibition of immune responses.,” Annual review of
immunology, vol. 20, pp. 29–53, Jan. 2002.
[277] A. V. Collins, D. W. Brodie, R. J. Gilbert, A. Iaboni, R. Manso-Sancho, B. Walse,
D. I. Stuart, P. van der Merwe, and S. J. Davis, “The Interaction Properties of Cos-
timulatory Molecules Revisited,” Immunity, vol. 17, pp. 201–210, Aug. 2002.
[278] M. Tacke, G. Hanke, T. Hanke, and T. Hünig, “CD28-mediated induction of prolifer-
ation in resting T cells in vitro and in vivo without engagement of the T cell receptor:
evidence for functionally distinct forms of CD28.,” European journal of immunology,
vol. 27, pp. 239–47, Jan. 1997.
[279] M. Tacke, G. J. Clark, M. J. Dallman, and T. Hünig, “Cellular distribution and cos-
timulatory function of rat CD28. Regulated expression during thymocyte maturation
and induction of cyclosporin A sensitivity of costimulated T cell responses by phor-
bol ester.,” Journal of immunology (Baltimore, Md. : 1950), vol. 154, pp. 5121–7,
May 1995.
[280] R. Siefken, R. Kurrle, and R. Schwinzer, “CD28-mediated activation of resting hu-
man T cells without costimulation of the CD3/TCR complex.,” Cellular immunology,
vol. 176, pp. 59–65, Feb. 1997.
[281] J. P. Van Wauwe, J. R. De Mey, and J. G. Goossens, “OKT3: a monoclonal anti-
human T lymphocyte antibody with potent mitogenic properties.,” Journal of im-
munology (Baltimore, Md. : 1950), vol. 124, pp. 2708–2713, 1980.
[282] C. Ball, B. Fox, S. Hufton, G. Sharp, S. Poole, R. Stebbings, D. Eastwood, L. Find-
lay, P. W. H. I. Parren, R. Thorpe, A. Bristow, and S. J. Thorpe, “Antibody C region
influences TGN1412-like functional activity in vitro.,” Journal of immunology (Bal-
timore, Md. : 1950), vol. 189, pp. 5831–40, Dec. 2012.
[283] G. Suntharalingam, M. R. Perry, S. Ward, S. J. Brett, A. Castello-Cortes, M. D. Brun-
ner, and N. Panoskaltsis, “Cytokine storm in a phase 1 trial of the anti-CD28 mono-
clonal antibody TGN1412.,” Tech. Rep. 10, Sept. 2006.
[284] F. Lühder, Y. Huang, K. M. Dennehy, C. Guntermann, I. Müller, E. Winkler,
T. Kerkau, S. Ikemizu, S. J. Davis, T. Hanke, and T. Hünig, “Topological require-
ments and signaling properties of T cell-activating, anti-CD28 antibody superago-
nists.,” The Journal of experimental medicine, vol. 197, pp. 955–66, Apr. 2003.
238 References
[285] E. J. Evans, R. M. Esnouf, R. Manso-Sancho, R. J. C. Gilbert, J. R. James, C. Yu,
J. A. Fennelly, C. Vowles, T. Hanke, B. Walse, T. Hünig, P. Sø rensen, D. I. Stuart, and
S. J. Davis, “Crystal structure of a soluble CD28-Fab complex.,” Nature immunology,
vol. 6, pp. 271–279, 2005.
[286] D. R. Davies and G. H. Cohen, “Interactions of protein antigens with antibodies.,”
Proceedings of the National Academy of Sciences of the United States of America,
vol. 93, pp. 7–12, Jan. 1996.
[287] C. Yu, A. F.-P. Sonnen, R. George, B. H. Dessailly, L. J. Stagg, E. J. Evans, C. A.
Orengo, D. I. Stuart, J. E. Ladbury, S. Ikemizu, R. J. C. Gilbert, and S. J. Davis,
“Rigid-body ligand recognition drives cytotoxic T-lymphocyte antigen 4 (CTLA-4)
receptor triggering.,” The Journal of biological chemistry, vol. 286, pp. 6685–96,
Feb. 2011.
[288] S. J. Davis, G. A. Schockmel, C. Somoza, D. W. Buck, D. G. Healey, E. P. Rieber,
C. Reiter, and A. F. Williams, “Antibody and HIV-1 gp120 recognition of CD4 under-
mines the concept of mimicry between antibodies and receptors.,” Nature, vol. 358,
pp. 76–9, July 1992.
[289] K. M. Ottemann, W. Xiao, Y. K. Shin, and D. E. Koshland, “A piston model for trans-
membrane signaling of the aspartate receptor.,” Science (New York, N.Y.), vol. 285,
pp. 1751–4, Sept. 1999.
[290] D. H. Margulies, “CD28, costimulator or agonist receptor?,” The Journal of experi-
mental medicine, vol. 197, pp. 949–53, Apr. 2003.
[291] L. S. Barak and W. W. Webb, “Diffusion of low density lipoprotein-receptor complex
on human fibroblasts.,” The Journal of cell biology, vol. 95, pp. 846–52, Dec. 1982.
[292] A. Kusumi, Y. Sako, and M. Yamamoto, “Confined lateral diffusion of membrane
receptors as studied by single particle tracking (nanovid microscopy). Effects of
calcium-induced differentiation in cultured epithelial cells.,” Biophysical journal,
vol. 65, pp. 2021–40, Nov. 1993.
[293] L. Weimann, K. A. Ganzinger, J. McColl, K. L. Irvine, S. J. Davis, N. J. Gay, C. E.
Bryant, and D. Klenerman, “A quantitative comparison of single-dye tracking analy-
sis tools using Monte Carlo simulations.,” PloS one, vol. 8, p. e64287, Jan. 2013.
[294] G. J. Schütz, H. Schindler, and T. Schmidt, “Single-molecule microscopy on model
membranes reveals anomalous diffusion.,” Biophysical journal, vol. 73, pp. 1073–80,
Aug. 1997.
[295] G. Guigas and M. Weiss, “Size-dependent diffusion of membrane inclusions.,” Bio-
physical journal, vol. 91, pp. 2393–8, Oct. 2006.
[296] D. Grünwald, R. M. Martin, V. Buschmann, D. P. Bazett-Jones, H. Leonhardt, U. Ku-
bitscheck, and M. C. Cardoso, “Probing intranuclear environments at the single-
molecule level.,” Biophysical journal, vol. 94, pp. 2847–58, Apr. 2008.
References 239
[297] C. Tardin, L. Cognet, C. Bats, B. Lounis, and D. Choquet, “Direct imaging of lat-
eral movements of AMPA receptors inside synapses.,” The EMBO journal, vol. 22,
pp. 4656–65, Sept. 2003.
[298] S. Wieser, M. Moertelmaier, E. Fuertbauer, H. Stockinger, and G. J. Schütz,
“(Un)confined diffusion of CD59 in the plasma membrane determined by high-
resolution single molecule microscopy.,” Biophysical journal, vol. 92, pp. 3719–28,
May 2007.
[299] X. Michalet, “Mean square displacement analysis of single-particle trajectories with
localization error: Brownian motion in an isotropic medium.,” Physical review. E,
Statistical, nonlinear, and soft matter physics, vol. 82, p. 041914, Oct. 2010.
[300] B. Wang, J. Kuo, S. C. Bae, and S. Granick, “When Brownian diffusion is not Gaus-
sian.,” Nature materials, vol. 11, pp. 481–5, June 2012.
[301] K. L. Angus and G. M. Griffiths, “Cell polarisation and the immunological synapse.,”
Current opinion in cell biology, vol. 25, pp. 85–91, Feb. 2013.
[302] A. Einstein, “Über die von der molekularkinetischen Theorie der Wärme geforderte
Bewegung von in ruhenden Flüssigkeiten suspendierten Teilchen,” Annalen der
Physik, vol. 322, no. 8, pp. 549–560, 1905.
[303] J. Charles A Janeway, P. Travers, M. Walport, and M. J. Shlomchik, “The structure
of a typical antibody molecule,” Garland Science, 2001.
[304] R. Wombacher, M. Heidbreder, S. van de Linde, M. P. Sheetz, M. Heilemann, V. W.
Cornish, and M. Sauer, “Live-cell super-resolution imaging with trimethoprim con-
jugates.,” Nature methods, vol. 7, pp. 717–9, Sept. 2010.
[305] P.-G. de Gennes, Scaling Concepts in Polymer Physics. Cornell University Press,
1979.
[306] J. Marrack, The chemistry of antigens and antibodies,. London: H.M. Stationery
Off., 1938.
[307] M. Dembo and B. Goldstein, “Theory of equilibrium binding of symmetric bivalent
haptens to cell surface antibody: application to histamine release from basophils.,”
Journal of immunology (Baltimore, Md. : 1950), vol. 121, pp. 345–53, July 1978.
[308] C. Mary, F. Coulon, N. Poirier, N. Dilek, B. Martinet, G. Blancho, and B. Vanhove,
“Antagonist properties of monoclonal antibodies targeting human CD28,” mAbs,
vol. 5, no. 1, pp. 47–55, 2013.
[309] B. Vanhove, G. Laflamme, F. Coulon, M. Mougin, P. Vusio, F. Haspot, J. Tiollier, and
J.-P. Soulillou, “Selective blockade of CD28 and not CTLA-4 with a single-chain Fv-
alpha1-antitrypsin fusion antibody.,” Blood, vol. 102, pp. 564–70, July 2003.
[310] A. Larsson, “Divalent binding of monoclonal antibody to a cell surface antigen. Mod-
elling of equilibrium data.,” Molecular immunology, vol. 26, pp. 735–9, Aug. 1989.
240 References
[311] R. C. Valentine and N. Green, “Electron microscopy of an antibody-hapten complex,”
Journal of Molecular Biology, vol. 27, pp. 615–617, Aug. 1967.
[312] N. Wrigley, E. Brown, and J. Skehel, “Electron microscopic evidence for the axial ro-
tation and inter-domain flexibility of the fab regions of immunoglobulin G,” Journal
of Molecular Biology, vol. 169, pp. 771–774, Sept. 1983.
[313] L. Gregory, K. G. Davis, B. Sheth, J. Boyd, R. Jefferis, C. Nave, and D. R. Burton,
“The solution conformations of the subclasses of human IgG deduced from sedimen-
tation and small angle X-ray scattering studies.,” Molecular immunology, vol. 24,
pp. 821–9, Aug. 1987.
[314] K. Ritchie, R. Iino, T. Fujiwara, K. Murase, and A. Kusumi, “The fence and picket
structure of the plasma membrane of live cells as revealed by single molecule tech-
niques (Review).,” Molecular membrane biology, vol. 20, no. 1, pp. 13–18, 2003.
[315] S. L. Shiao, J. M. McNiff, T. Masunaga, K. Tamura, K. Kubo, and J. S. Pober, “Im-
munomodulatory properties of FK734, a humanized anti-CD28 monoclonal antibody
with agonistic and antagonistic activities.,” Transplantation, vol. 83, pp. 304–13, Feb.
2007.
[316] S. R. Hubbard, “Structural analysis of receptor tyrosine kinases,” Progress in Bio-
physics and Molecular Biology, vol. 71, pp. 343–358, Apr. 1999.
[317] S.-P. Cordoba, K. Choudhuri, H. Zhang, M. Bridge, A. B. Basat, M. L. Dustin, and
P. A. van der Merwe, “The large ectodomains of CD45 and CD148 regulate their
segregation from and inhibition of ligated T-cell receptor.,” Blood, vol. 121, pp. 4295–
302, May 2013.
[318] D. J. Irvine and J. Doh, “Synthetic surfaces as artificial antigen presenting cells in the
study of T cell receptor triggering and immunological synapse formation.,” Seminars
in immunology, vol. 19, pp. 245–54, Aug. 2007.
[319] M. D. Lew, A. R. S. von Diezmann, and W. E. Moerner, “Easy-DHPSF open-source
software for three-dimensional localization of single molecules with precision be-
yond the optical diffraction limit.,” Protocol exchange, vol. 2013, Feb. 2013.
[320] C. Fang, R. R. Frontiera, R. Tran, and R. A. Mathies, “Mapping GFP structure evolu-
tion during proton transfer with femtosecond Raman spectroscopy.,” Nature, vol. 462,
pp. 200–4, Nov. 2009.
[321] J. Bähler, J. Q. Wu, M. S. Longtine, N. G. Shah, A. McKenzie, A. B. Steever,
A. Wach, P. Philippsen, and J. R. Pringle, “Heterologous modules for efficient and
versatile PCR-based gene targeting in Schizosaccharomyces pombe.,” Yeast (Chich-
ester, England), vol. 14, pp. 943–51, July 1998.
[322] R. L. Strack, B. Hein, D. Bhattacharyya, S. W. Hell, R. J. Keenan, and B. S. Glick,
“A rapidly maturing far-red derivative of DsRed-Express2 for whole-cell labeling,”
Biochemistry, vol. 48, pp. 8279–8281, 2009.
References 241
[323] K. Kaji, I. M. Caballero, R. MacLeod, J. Nichols, V. A. Wilson, and B. Hendrich,
“The NuRD component Mbd3 is required for pluripotency of embryonic stem cells.,”
Nature cell biology, vol. 8, pp. 285–92, Mar. 2006.
[324] J. Nichols, J. Silva, M. Roode, and A. Smith, “Suppression of Erk signalling pro-
motes ground state pluripotency in the mouse embryo.,” Development (Cambridge,
England), vol. 136, pp. 3215–3222, 2009.
[325] Y. Harigaya, H. Tanaka, S. Yamanaka, K. Tanaka, Y. Watanabe, C. Tsutsumi,
Y. Chikashige, Y. Hiraoka, A. Yamashita, and M. Yamamoto, “Selective elimination
of messenger RNA prevents an incidence of untimely meiosis.,” Nature, vol. 442,
pp. 45–50, July 2006.
[326] A. T. Watson, V. Garcia, N. Bone, A. M. Carr, and J. Armstrong, “Gene tagging and
gene replacement using recombinase-mediated cassette exchange in Schizosaccha-
romyces pombe.,” Gene, vol. 407, pp. 63–74, Jan. 2008.
[327] A. Weiss and J. D. Stobo, “Requirement for the coexpression of T3 and the T cell
antigen receptor on a malignant human T cell line.,” The Journal of experimental
medicine, vol. 160, pp. 1284–99, Nov. 1984.
[328] C. Demaison, K. Parsley, G. Brouns, M. Scherr, K. Battmer, C. Kinnon, M. Grez,
and A. J. Thrasher, “High-level transduction and gene expression in hematopoietic
repopulating cells using a human immunodeficiency [correction of imunodeficiency]
virus type 1-based lentiviral vector containing an internal spleen focus forming virus
promoter.,” Human gene therapy, vol. 13, pp. 803–13, May 2002.
[329] F. L. Graham, J. Smiley, W. C. Russell, and R. Nairn, “Characteristics of a human cell
line transformed by DNA from human adenovirus type 5.,” The Journal of general
virology, vol. 36, pp. 59–74, July 1977.
[330] S. M. Watt, D. J. Gilmore, J. M. Davis, M. R. Clark, and H. Waldmann, “Cell-surface
markers on haemopoietic precursors. Reagents for the isolation and analysis of pro-
genitor cell subpopulations.,” Molecular and cellular probes, vol. 1, pp. 297–326,
Dec. 1987.
[331] Y. M. Kong, H. Waldmann, S. Cobbold, A. A. Giraldo, B. E. Fuller, and L. L. Simon,
“Pathogenic mechanisms in murine autoimmune thyroiditis: short- and long-term
effects of in vivo depletion of CD4+ and CD8+ cells.,” Clinical and experimental
immunology, vol. 77, pp. 428–33, Sept. 1989.

Appendix A
Supplementary Figures and Tables
Table A.1 Models of mEos photo-bleaching (Fitted amplitude). (Values of the fitted amplitudes are given ±
one standard deviation of the mean for two fields of view and two biological repeats. A.U.: arbitrary units.)
mEos2 mEos3.1
No 405 nm illumination
f (t) = A e−
t
τ1 +B e−
t
τ2
A = 390 ± 240 A.U.
B = 360 ± 230 A.U.
A = 500 ± 140 A.U.
B = 440 ± 110 A.U.
Continuous 405 nm illumination
f (t) = A e−
t
τ1 +B e−
t
τ2 +C e−
t
τ3
A = 340 ± 110 A.U.
B = 280 ± 100 A.U.
C = 220 ± 75 A.U.
A = 470 ± 85 A.U.
B = 250 ± 80 A.U.
C = 330 ± 120 A.U.
Table A.2 Models of mEos photo-physics (Fitted amplitude). Amplitudes A and B of the exponential decays
fitted to the distributions of ‘on’ and ‘off’ times as described in Figure 3.9 under different 405 nm activation
powers. Values are given for the fit of a single exponential decay for tON , and a linear combination of two
exponential decays for tOFF .
mEos2 mEos3.1
405 nm tON tOFF 405 nm tON tOFF
7 µW A = 6.98
A = 1.86
B = 0.16 162 µW A = 4.50
A = 3.26
B = 0.16
31 µW A = 6.12
A = 2.05
B = 0.19
86 µW A = 5.62
A = 2.13
B = 0.21
910 µW A = 4.38
A = 3.18
B = 0.18
244 Supplementary Figures and Tables
Figure A.1 Spectra of the dichroic mirrors and filters used in both imaging channels of the dualview. Black:
Transmission of the quad-band dichroic mirror separating the illumination beams from the the fluorescent
signal. Blue: Transmission of the long-pass emission filter inserted before the dualview. Orange: Reflection
(A) and transmission (B) of the beam splitter of the dualview. Green: Transmission of the band-pass filter
inserted in the short-pass channel of the dualview. Red and Pink: Transmission of the band-pass and long-
pass filters (resp.) inserted in the long-pass channel of the dualview..
245
Table A.3 Parameters values used in Peak Fit. Table summarising the parameters used in the ImageJ plug-in
Peak Fit [149] integrated in the bespoke pieces of Software B.4 (‘General’) and B.5 (‘vls’) to detect and fit
PSFs.
Name Value (General) Value (vls)
Calibration 110 nm /pix
Gain EM Gain / Camera Gain
Exposure time 50 ms
Initial StdDev0 Theoretical σPSF (Eqn 2.1)
Initial StdDev1 Theoretical σPSF (Eqn 2.1)
Initial Angle 0
Smoothing 0.5
Smoothing2 3 8
Search Width 3 8
Fitting Width 3
Fit Solver LVM
Fit Function Free
Fit Criteria Least squared error
Significant Digits 5
Lambda 10
Max Iterations 20
Fail Limit 10
Include Neighbours true
Neighbour Height 0.3
Residual Threshold 1
Duplicate Distance 0.5
Shift Factor 1.5 3
Signal Strength 20
Width Factor 5 30
Precision 0

Appendix B
Supplementary Movies and Software
All movies and software listed here can be found on the accompanying CD.
Movie B.1 Example of sptPALM analysis. An example of the single-particle tracking analysis used in this the-
sis is shown for data from one representative experiment (from Chapter 7). T cell expressing CD28-mEos3.2 at
their membrane and decorated with anti-CD28 mAbs are dropped on a glass coverslip coated with secondary
antibodies. The live cell forms a contact with the coated surface and is imaged under TIRF 561 nm illumi-
nation. A low-power constant TIRF 405 nm illumination is additional applied to activate a sub-population
of mEos3.2 fluorophores. For the representative cell shown in this movie, 1,555 frames were acquired and
analysed as described in Section 7.2.1 (the single-particle tracking algorithm is developed in Section 3.4.2).
The trajectories of all localisations detected in the frame are plotted (one colour per trajectory). Trajectories
are shown up to the frame shown and with a maximum of 10 frames of history. The video is played back at
real-time (20 frames per second). The scale bar is 1.5 µm and data was acquired at 37◦C.
Movie B.2 Real-time demonstration of the automation algorithm. Some yeast cells were trapped in the
microfluidic device introduced in Section 3.5 and imaged using the automation algorithm without its auto-
focusing step (Software B.3). First, the different steps of a cycle of the algorithm (Figure 3.16) are presented
and described. Then, a real-time movie of the computer screen during four cycles of the algorithm follows.
248 Supplementary Movies and Software
Software B.3 Automation µManager plug-in. The automation of the microscope was operated through the
control of various elements (lasers, shutters, (xy) stage, focus and EMCCD) with µManager [173]. A bespoke
library (‘MicroManager_MP.jar’) was designed to operate the microscope with easy-to-use lines of command
adapted to the experiments. However, this means that the source code is adapted to the configuration of the
instrument in µManager and has to be modified for each new instrument. Nevertheless, it is designed to work
on four instruments, so its implementation can be easily generalised to other microscope set-ups. Both the
source code and its compiled jar file are provided.
This library was used in both µManager plugins (e.g. to calibrate the double-helix in Section 3.2.2) or macros
written for specific imaging protocols in the Beanshell script feature of µManager. A few examples of macros
are given, such as the one used for ‘smart’ automation imaging of live fission yeast in micro-fluidic devices
(‘MicroFluidicDeviceAutomation.bsh’; Section 3.5) or the Fermi 405 nm activation protocol described in
Section 3.3.4 and used in Chapter 6 (‘Fermi405Activation.bsh’).
Software B.4 vlsSMLM ImageJ plug-in. An ImageJ plug-in to implement the vlsSMLM optical sectioning
method described in Chapter 4 was designed. The plug-in is divided in two consecutive steps: (1) the cali-
bration of the vls by analysing a z-scan of immobilised and separated sub-diffraction fluorescent beads. (2)
This enables the determination of width and amplitude thresholds that can be applied to SMLM data imaged
under similar conditions. A ‘ReadMe’ text file provides the information to install the plug-in. An example
of calibration file and the SMLM data analysed in Figure 4.8 are also provided. Both the source code and its
compiled jar file are provided.
Software B.5 ImageJ plug-ins for SMLM data analysis. All analysis methods presented in this thesis were
implemented as a coherent ImageJ library (‘SMLM_.jar’) that evolved during the course of this thesis. It is
organised in sub-packages and and various plug-ins implementing different analyses (e.g. batch analyses, reg-
istration of two channels (Section 3.2.3), CBC analysis (Section 3.4.2), single-particle tracking (Section 3.4.2),
pair-correlation analysis (Section 6.2.1), clustering (Section 3.4.2)). Although a simple interface asks for the
most frequently used parameters in most plug-ins, some plug-ins were slightly modified from one analysis
to the other ((un)commenting a few lines) directly in the source code. A ‘ReadMe’ text file provides the
information to install the plug-in. Both the source code and its compiled jar file are provided.