RNA sequencing
for the study of splicing
Mar Gonzàlez-Porta
European Molecular Biology Laboratory
European Bioinformatics Institute
Darwin College
A dissertation submitted to the University of Cambridge
for the degree of Doctor of Philosophy
April 2014

To my mum.

This dissertation is the result of my own work and includes nothing which is the
outcome of work done in collaboration except where specifically indicated.
The text in this thesis does not exceed the specified word limit of 60,000 words as
defined by the Biology Degree Committee.
April, 2014 Mar Gonzàlez-Porta

Abstract
Amongst the many processes that shape the final set of RNA molecules present
in eukaryote cells, splicing emerges as the most prominent mechanism for mes-
sage diversification. In recent years, applications of high throughput sequencing
to RNA, known as RNA sequencing, have opened new avenues for the study
of transcriptome composition, and have enabled further characterisation of such
mechanism. In this thesis, I focus on the application of this technology to the study
of human transcript diversity and its potential impact on the protein repertoire.
In the first results chapter, I explore the extent of transcriptome diversity by asking
whether there is a preference for the production of specific alternative transcripts
within each given gene. I show that while many alternative transcripts can be
detected, the expression of most protein coding genes tends to be dominated by
one single transcript (major transcript). Such findings are validated in the second
chapter, and are further used to explore changes in splicing patterns in a disease
context. By analysing healthy and tumor samples from kidney cancer patients, I
show that most of the detected splicing alterations do not lead to big changes in
the relative abundance of major transcripts, at least in a recurrent manner. In ad-
dition, I introduce a framework to visualise the most extreme changes in splicing
and to evaluate their potential functional impact. In the third chapter, I investi-
gate the role of spliceosome assembly dynamics on the regulation of splice site
choice. I show that depletion of PRPF8, a core spliceosomal component, leads to
the preferential retention of a subset of introns with weaker splice sites, and also
introduces alterations in the rate of co-transcriptional splicing. Finally, in the last
chapter, I explore the validation of changes in alternative transcript abundance at
the protein level, through the integration of results derived from RNA sequencing
datasets with those obtained from proteomics experiments.
Altogether, the findings described in this thesis provide a global picture on the
extent of alternative splicing in the diversification of the transcriptome, expand
current knowledge on the splicing reaction, and open new possibilities for the
integration of transcriptomics and proteomics data.

Preface
I would like to start by thanking my supervisor Alvis Brazma, for giving me the
opportunity to get involved in this experience. My PhD has been a long journey
of almost four years, and I am extremely grateful for his support and guidance
throughout all this time. I am also deeply indebted to Johan Rung, who behaved
like a second supervisor to me, and to John Marioni, without whom half of this
dissertation would have not been possible.
I cannot thank enough all the members of the Functional Genomics research team,
who have made it easy to go to the office every day. Special thanks to Mitra,
Liliana, Natalja, Gabry, Angela, Shyama, Wan, Jing, Aurora, Sandra and Yuedan
for their continuous encouragement and interest in my research. Likewise, I shall
not forget all of those with whom I have not necessarily shared offices, but who
have always kept a smile on their faces when crossing me in the corridor.
I wish to acknowledge the members of my Thesis Advisory Committee, John Mar-
ioni, Jan Korbel and Simon Tavaré, for their insightful comments, as well as all my
collaborators, including Adam Frankish, Jennifer Harrow, David Perera, Arthur
Bartolozzi, Ashok Venkitaraman and Yansheng Liu, for the many thoughtful dis-
cussions. I would like to pay special thanks to Vi Wickramasinghe, for making the
end of my PhD such a great experience. I am also thankful to all the people who
have provided feedback along the path, specially Wolfgang Huber, Nicholas Lus-
combe, Roderic Guigó and Ernest Turró, and to all of those who have helped in
improving this dissertation with their comments. Looking back, I ammost grateful
to all the people who contributed to my scientific growth and education, partic-
ularly Antoni Romeu, Julien Roux, Marc Robinson-Rechavi and Roderic Guigó.
I also need to thank Gabriella Rustici for dynamising my PhD experience by in-
vii
volving me in teaching. Similarly, I wish to thank all the people with whom I
have been involved in this endeavor, including Angela Gonçalves, Mitra Barzine,
Liliana Greguer and Laura Emery, as well as Francesco Paolo Casale for his great
patience. Finally, I would like to express special thanks to Lynn French and Tracey
Andrew, for making my bureaucratic life that much easier.
I have been extremely fortunate to share my time at the EBI with a great group
of people, from which I would like to give special thanks to Robert, Alexa, Nidhi,
Matthias, Senay, Sergio, Rita, Cathy and Afonso. Likewise, I cannot thank enough
all those friends who have remained a part of my life, despite the long distances
and the course of time.
Last, I am deeply grateful to all the people who have supported me in my personal
life, specially my family, who have contributed the most in making this possible.
In particular, I would like to thank my brother, for keeping the family conversa-
tions alive; my dad, for his routine visits and continuous care; and my mum, for
pushing us all to continue ahead. Finally, I wish to thank Pol, for the countless
trips and hours in Skype, which have now finally been replaced for conversations
over dinner. I would not have made it this far without him.
viii
Contents
Contents ix
List of Figures xiii
List of Tables xvii
1 Introduction 1
1.1 A day in the life of an mRNA . . . . . . . . . . . . . . . . . . . . . . . 2
1.1.1 Transcription . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.1.2 mRNA processing . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.1.3 The cytosolic life of mRNAs . . . . . . . . . . . . . . . . . . . 5
1.2 The splicing reaction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.2.1 Key players in mRNA splicing . . . . . . . . . . . . . . . . . . 6
1.2.2 Splicing by the major spliceosome . . . . . . . . . . . . . . . . 7
1.2.3 Diversifying the message: alternative splicing and beyond . . 10
1.2.4 The regulation of splicing . . . . . . . . . . . . . . . . . . . . . 12
1.3 Studying the transcriptome with RNA sequencing . . . . . . . . . . . 14
1.3.1 A typical sequencing workflow . . . . . . . . . . . . . . . . . . 15
1.3.2 Read mapping strategies . . . . . . . . . . . . . . . . . . . . . 18
1.3.2.1 Alignment to the genome or transcriptome . . . . . 19
1.3.2.2 De novo assembly . . . . . . . . . . . . . . . . . . . . 21
1.3.3 The estimation of expression levels . . . . . . . . . . . . . . . 21
1.3.3.1 Gene expression levels . . . . . . . . . . . . . . . . . 21
1.3.3.2 Transcript expression levels . . . . . . . . . . . . . . 23
1.3.3.3 De novo transcript identification . . . . . . . . . . . . 26
1.3.4 Read count normalisation . . . . . . . . . . . . . . . . . . . . . 27
1.3.5 Differential expression . . . . . . . . . . . . . . . . . . . . . . . 30
ix
1.3.6 Differential splicing . . . . . . . . . . . . . . . . . . . . . . . . 33
1.4 Aims of the thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
2 The extent of transcriptome diversity 37
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
2.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
2.2.1 Most protein coding genes express one dominant transcript . 41
2.2.2 The evaluation of different methods leads to a consistent out-
come . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
2.2.3 Major transcripts from coding genes do not always code for
proteins . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
2.3 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
2.4 Computational methods . . . . . . . . . . . . . . . . . . . . . . . . . . 56
3 The prevalence of splicing changes in cancer 63
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
3.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
3.2.1 Splicing is largely altered in ccRCC . . . . . . . . . . . . . . . 67
3.2.2 Large and recurrent changes in splicing are rare . . . . . . . . 73
3.2.3 Identification, annotation and visualisation of switch events
with SwitchSeq . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
3.2.4 Pathways appear as broadly disrupted when integrating dif-
ferent layers of information . . . . . . . . . . . . . . . . . . . . 78
3.2.5 Splicing patterns in cancer cell lines are different from those
in primary tissues . . . . . . . . . . . . . . . . . . . . . . . . . 82
3.3 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
3.4 Computational methods . . . . . . . . . . . . . . . . . . . . . . . . . . 88
4 The regulation of splicing by core spliceosomal factors 93
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
4.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
4.2.1 PRPF8 knock-down causes accumulation of cells in mitosis . 96
4.2.2 The mitotic arrest phenotype is driven by disruptions in splic-
ing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
4.2.3 Modulating splice site strength can compensate for the down-
regulation of PRPF8 . . . . . . . . . . . . . . . . . . . . . . . . 103
x
4.2.4 PRPF8 knock-down has an impact on the rate of co-transcriptional
splicing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
4.3 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
4.4 Computational methods . . . . . . . . . . . . . . . . . . . . . . . . . . 112
5 The impact of splicing at the protein level 117
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
5.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
5.2.1 Extreme changes in splicing can be detected at the protein
level . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
5.2.2 RNA-seq fold-change estimates correlate with those obtained
from SWATH-MS . . . . . . . . . . . . . . . . . . . . . . . . . . 126
5.2.3 The correlation estimates can be improved by incorporating
information on major transcripts . . . . . . . . . . . . . . . . . 127
5.2.4 The detected correlation is not driven by differences in gene
expression levels . . . . . . . . . . . . . . . . . . . . . . . . . . 128
5.3 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
5.4 Computational methods . . . . . . . . . . . . . . . . . . . . . . . . . . 131
6 Conclusions 133
Full list of publications 139
A RNA-seq datasets used in the thesis 141
B Supplementary Material for Chapter 2 149
C Supplementary Material for Chapter 3 155
D Supplementary Material for Chapter 4 167
E Supplementary Material for Chapter 5 173
References 177
xi

List of Figures
1.1 Key steps in the regulation of eukaryotic gene expression . . . . . . 3
1.2 Splicing by the major spliceosome . . . . . . . . . . . . . . . . . . . . 10
1.3 Mechanisms for the formation of alternative transcripts from the
same genomic locus . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.4 The regulation of splicing . . . . . . . . . . . . . . . . . . . . . . . . . 13
1.5 Overview of library preparation and sequencing steps in an Illu-
mina platform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
1.6 Overview of the mapping algorithm implemented in TopHat . . . . 20
1.7 Overview of htseq-count . . . . . . . . . . . . . . . . . . . . . . . . . . 22
1.8 Overview of the analysis workflow implemented in MISO for the
estimation of transcript abundances . . . . . . . . . . . . . . . . . . . 25
1.9 Overview of the de novo transcript identification algorithm imple-
mented in Cufflinks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
1.10 Limitations on the use of RPKMs for differential expression analysis 29
1.11 Overview of the steps required for differential expression analysis
using DESeq2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
1.12 Strategies for the study of changes in the abundance of alternative
transcripts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
2.1 Most protein coding genes express one dominant transcript . . . . . 43
2.2 Evaluation of two hypothetical scenarios on major transcript abun-
dance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
2.3 Length distribution for major transcripts . . . . . . . . . . . . . . . . 47
2.4 Example of non-canonical major transcript common to all the 16
tissues analysed . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
2.5 Summary of Cufflinks de novo quantification results . . . . . . . . . . 49
xiii
2.6 Major non-coding transcripts in protein coding genes . . . . . . . . . 51
2.7 Focus on retained introns . . . . . . . . . . . . . . . . . . . . . . . . . 52
3.1 Splicing patterns in ccRCC tumours vs. healthy matched samples . . 68
3.2 Number of expressed transcripts per gene in normal vs. tumour
samples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
3.3 Most protein coding genes express one dominant transcript in both
normal and tumour samples . . . . . . . . . . . . . . . . . . . . . . . 71
3.4 Major transcript expression patterns in ccRCC tumours vs. healthy
matched samples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
3.5 Switch events between tumours vs. healthy matched samples in pro-
tein coding genes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
3.6 Examples of recurrent switch events . . . . . . . . . . . . . . . . . . . 76
3.6 Examples of recurrent switch events (continued) . . . . . . . . . . . . 77
3.7 Patient-specific landscape of alterations for a subset of the genes
with recurrent somatic mutations . . . . . . . . . . . . . . . . . . . . . 80
3.8 Summary of detected alterations in the VHL/HIF pathway . . . . . . 81
3.9 Most protein coding genes express one predominant transcript in
cell lines. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
3.10 Splicing patterns in cancer cell lines vs. primary tissues . . . . . . . . 83
3.11 Switch events in cancer cell lines vs. primary tissues . . . . . . . . . . 85
4.1 Characterisation of the mitotic arrest phenotype obtained after PRPF8
KD . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
4.2 Phenotype recovery experiments in PRPF8 KD cells . . . . . . . . . . 98
4.3 Results from Sm iCLIP experiments for the 5’ splice site . . . . . . . 99
4.4 Characterisation of splicing alterations in PRPF8 KD cells . . . . . . 101
4.5 Validation of the predicted splicing changes in a subset of genes
with a role in cell division . . . . . . . . . . . . . . . . . . . . . . . . . 103
4.6 Splice site strength for the top retained introns following PRPF8 KD 104
4.7 Mini-gene experiments for the CDC20 gene . . . . . . . . . . . . . . . 105
4.8 GC content of top retained vs. non-retained introns . . . . . . . . . . 106
4.10 Experimental validation of the predicted differences in the rate of
co-transcriptional splicing after PRPF8 KD . . . . . . . . . . . . . . . 107
4.9 Co-transcriptional splicing evidence from intronic reads . . . . . . . 108
5.1 Integration of the RNA-seq and SWATH-MS data . . . . . . . . . . . 122
xiv
5.2 Examples of switch events with supporting peptide evidence . . . . 124
B.1 Transcript relative abundances across all the studied datasets . . . . 150
B.2 Major transcript dominance across all the studied datasets . . . . . . 151
C.1 SwitchSeq analysis workflow . . . . . . . . . . . . . . . . . . . . . . . 156
C.2 Summary of detected alterations in the PI(3)K-AKT-MTOR signalling
pathway . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161
C.3 Summary of detected alterations in the focal adhesion pathway . . . 165
D.1 Validation of PRPF8 down-regulation . . . . . . . . . . . . . . . . . . 168
D.2 Cell cycle analysis following knock-down of splicing factors from
several spliceosomal complexes . . . . . . . . . . . . . . . . . . . . . . 169
D.3 Results from Sm iCLIP experiments for the 3’ splice site . . . . . . . 170
D.4 Differences in intron expression between first and last introns . . . . 171
E.1 RNA-seq and SWATH-MS fold-change estimates for the differen-
tially used transcripts with peptide evidence . . . . . . . . . . . . . . 174
E.2 RNA-seq and SWATH-MS fold-change estimates for the differen-
tially used transcripts with peptide evidence after excluding differ-
entially expressed genes. . . . . . . . . . . . . . . . . . . . . . . . . . . 175
xv

List of Tables
2.1 Consistency in the transcript abundance estimates across different
software . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
5.1 Summary of the results obtained from the RNA-seq and SWATH-
MS datasets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
5.2 Validation of switch events . . . . . . . . . . . . . . . . . . . . . . . . 123
5.3 Validation of switch events when filtering by peptide fold-change
significance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
5.4 Validation of differential transcript usage events . . . . . . . . . . . . 127
5.5 Validation of differential transcript usage events after excluding dif-
ferentially expressed genes . . . . . . . . . . . . . . . . . . . . . . . . 128
A.1 RNA-seq data used in Chapter 2 . . . . . . . . . . . . . . . . . . . . . 142
A.2 RNA-seq data used in Chapter 3 . . . . . . . . . . . . . . . . . . . . . 144
A.3 RNA-seq data used in Chapters 4 and 5 . . . . . . . . . . . . . . . . . 148
B.1 GO enrichment analysis for recurrent 5-fold dominant genes . . . . 152
B.2 GO enrichment analysis for genes that tolerate splicing . . . . . . . . 152
B.3 mRNA pool estimates for the cell line dataset . . . . . . . . . . . . . 153
B.4 Number of dominant transcripts in the cell line dataset . . . . . . . . 153
B.5 GO enrichment analysis for genes with a major retained intron both
in the nucleus and the cytosol . . . . . . . . . . . . . . . . . . . . . . . 154
B.6 Re-annotation of major processed transcripts . . . . . . . . . . . . . . 154
C.1 Pathway enrichment analysis for differentially spliced genes in cell
lines vs. tumour samples . . . . . . . . . . . . . . . . . . . . . . . . . . 165
xvii

Chapter 1
Introduction
When the first human genome sequence was declared completed in 2003, it es-
tablished the order of the billions of letters that contain the instructions for the
survival of each of the cells in our body [International Human Genome Sequenc-
ing Consortium, 2001, 2004; Venter et al., 2001], an exercise that was compared to
reading the book of life1. In fact, the comparison of DNA to a book of instructions
has been widely used as a metaphor to explain the flow of genetic information2,
and far from analogies, it has been recently shown that this molecule is able to
encode for text and even more [Goldman et al., 2013]. Since then, vast efforts have
been devoted to interpreting the content of the human genome sequence, best ex-
emplified by the ENCODE project [ENCODE Project Consortium et al., 2012]. It
has become clear from those that the flow of information is not as simple as it was
initially envisioned [Crick, 1970], and that the book of instructions might actually
look closer to those from the series "Choose Your Own Adventure"3, in which one
text can lead to different outcomes.
Overall, from the interpretation of the genetic information stored in the DNA to
protein synthesis, many regulatory processes contribute to shape the final reper-
toire of molecules present in the cell. In multicellular organisms such as humans,
those processes constitute the intricate basis upon which a large diversity of cell
1http://www.sanger.ac.uk/about/press/2003/030414.html
2http://ed.ted.com/lessons/dna-the-book-of-you-joe-hanson
3http://en.wikipedia.org/wiki/Choose_Your_Own_Adventure
1
Chapter 1. Introduction
types and states are derived from the same genetic message. In this context, the
study of transcriptome composition has emerged as a valuable strategy to un-
derstand such diversity, and while the mere existence of an RNA molecule does
not ensure the production of a functional protein product, this type of knowledge
has helped in the identification of links between phenotype and genotype [Vogel
and Marcotte, 2012]. Furthermore, recent technological advances have enabled
the characterisation of RNA samples in a more reproducible and high throughput
fashion than is yet possible for their final products [Wang et al., 2009].
In the present chapter, I provide an overview of the different regulatory processes
that control the identity and abundance of the set of RNAs found in a given cell,
with special emphasis on splicing and its potential for message diversification. In
addition, I introduce key concepts on the use of RNA sequencing for the study of
gene and transcript expression in a high throughput manner.
1.1 A day in the life of an mRNA
Amongst the diversity of identified RNA species, messenger RNAs (mRNAs) are
regarded as those that contain the necessary information for the synthesis of pro-
teins. The half-lives of mRNAs are short compared to those of other molecules
in the cell (e.g. most mammalian mRNAs remain in the cell for approximately 9h,
compared to 46h in the case of proteins [Schwanhausser et al., 2011]). Nonethe-
less, from birth to death, these molecules are controlled by complex regulatory
systems that determine which messages are eventually expressed (Figure 1.1). In
this section, I provide an overview of the key steps behind such control.
1.1.1 Transcription
Transcription is the first phase in determining the set of RNAs expressed in a
given cell [Alberts et al., 2002]. During transcription, stretches of DNA (i.e. genes)
are used as templates for the synthesis of complementary single stranded RNA
molecules (i.e. transcripts). In eukaryote cells, such reaction can be catalysed by
three different enzymes (i.e. RNA polymerases I, II and III), depending on the
type of gene being targeted. RNA polymerase II is commonly regarded as the
most prominent player in the transcription reaction, since it is responsible for
the synthesis of RNAs derived from the majority of genes, including those that
encode for proteins. On the other hand, RNA polymerase I and III are specifically
2
Chapter 1. Introduction
AAAAA
DNA
pre-mRNA
nucleus
cytosol
AAAAA
mRNA
TRANSCRIPTION
RNA PROCESSING
    5’ capping
    splicing
    3’polyadenylation
EXPORT
LOCALISATION
DEGRADATIONTRANSLATION
protein
Figure 1.1| Key steps in the regulation of eukaryotic gene expression. mRNA
expression starts with the nuclear transcription of specific DNA loci which con-
tain the information required for the synthesis of proteins (i.e. genes). Following
several processing steps, some of which occur co-transcriptionally, the transcrip-
tion products are further transformed into mature mRNAs that can then be ex-
ported to the cytosol and localised to sub-cellular compartments. mRNA export
is linked to strict quality control mechanisms, and both unprocessed RNAs and
debris from the previous processes will be degraded previously to this stage.
Once in the cytosol, mRNAs can be recognised by ribosomes and translated into
proteins, and will be eventually degraded.
involved in the transcription of ribosomal RNAs (rRNAs), transfer RNAs (tRNAs)
and several small RNAs [Paule and White, 2000].
Transcription by RNA pol II is a multi-step process that starts with the binding
of several proteins to a regulatory region located upstream of the gene, known as
the promoter [Fuda et al., 2009]. These proteins enable the subsequent assembly
of the polymerase and the formation of the transcription initiation complex, and
given that they participate in the recognition of the majority of promoters, they are
commonly regarded as general transcription factors (TFs). Conversely, more spe-
3
Chapter 1. Introduction
cific TFs also exist, which are able to modulate the fate of the reaction by binding
to DNA regions that promote or inhibit polymerase assembly (i.e. enhancers and
silencers, respectively), hence contributing to the regulation of expression levels
[Vaquerizas et al., 2009]. Following the assembly steps and further conformational
rearrangements, RNA pol II releases from the large complex of proteins and aban-
dons the promoter region (i.e. promoter clearance), thus entering the elongation
phase [Kwak and Lis, 2013]. However, the transition to this stage is not immediate,
and in most cases the polymerase remains at the promoter generating short trun-
cated transcripts (i.e. abortive initiation). During elongation, RNA is synthesised
from the transcription start site (TSS), and nucleotides are incorporated in a com-
plementary basis in the 5’ to 3’ direction. Eventually, the polymerase transcribes
through the cleavage and polyadenylation signals that mark the end of the gene,
and it is released from the DNA template [Kuehner et al., 2011].
1.1.2 mRNA processing
All mRNA molecules undergo several modifications before they are exported to
the cytosol, which include the addition of a 5’ cap, the polyadenylation of the 3’
end and the removal of introns via splicing [Darnell, 2013].
Shortly after the RNA pol II has entered the elongation phase, a methylated gua-
nine nucleotide is added to its 5’ end through an enzymatic reaction (i.e. cap-
ping). Such cap not only delimits the 5’ end of transcripts, but also enables the
distinction of mRNAs from other RNA species (e.g. RNA pol I and III produce
uncapped RNAs). Furthermore, it protects the RNA molecule from degradation,
and aids in the initiation of translation by enabling the binding of ribosomes to
the mRNA. Similarly, the 3’ end of mRNAs is also modified through the addition
of a polyAdenosine (polyA) tail, which is typically 200-250 nucleotides long in
mammalian cells. Polyadenylation serves the function of extending the mRNA
half-life, and given that it is a modification shared by all mRNA molecules, it is
commonly exploited for their study (i.e. one of the most common RNA extraction
protocols in use relies on the specific selection of polyA-tailed RNA species).
Splicing, however, is a much more complex reaction. During this process, some
regions of the pre-mRNA are lost (i.e. introns), and the stretches of sequence that
contain the necessary information for protein synthesis (i.e. exons) are brought
together, as detailed later on in this chapter. As a result, a mature mRNA product
is obtained.
4
Chapter 1. Introduction
1.1.3 The cytosolic life of mRNAs
Once a mature mRNA molecule is obtained, it undergoes selective export through
the nuclear pore. mRNA export is linked to strict quality control mechanisms
that ensure that immature RNAs remain in the nucleus [Porrua and Libri, 2013].
Those mechanisms rely on the recognition of protein complexes that accompany
the RNA molecules (i.e. RNA binding proteins; RBPs), which act as markers on
the completion status of the processing steps mentioned in the previous section.
For example, the cap-binding and polyA binding-complexes serve as indicators of
successful capping and polyadenylation reactions, respectively, and other protein
complexes mark the end of the splicing in a similar fashion (i.e. the exon-junction
complex - EJC; see next section). Conversely, the presence of RBPs involved in the
execution of each of these steps marks the mRNA molecule as immature, hence
preventing its export.
Unprocessed mRNAs, together with the remainders from the transcription and
splicing reactions, will be eventually degraded by the exosome, a large complex
of RNA exonucleases [Pérez-Ortín et al., 2013]. In the cases when they are erro-
neously exported, or when intact mRNAs become damaged in the cytosol, further
quality control mechanisms prevent their translation. Most of these are intrinsic
to the steps required for the initiation of protein synthesis, as is the case for the
recognition of the 5’ cap and polyA tail by the translation initiation machinery.
However, a separate surveillance system also exists, which actively seeks aberrant
mRNAs for degradation, before efficient translation occurs. This system is referred
to as nonsense-mediated decay (NMD) and specifically targets the presence of pre-
mature stop codons in the transcript, which might arise from errors in the splicing
reaction [Pérez-Ortín et al., 2013]. During NMD, a first round of translation starts
as soon as the 5’ end of the mRNA emerges from the nuclear pore, during which
the exon-junction complexes that surround each splice-site are detached from the
mRNA. Under the presence of nonsense codons, the mRNA remains bounded to
such complexes and is rapidly degraded.
Following export, mRNAs are then localised within the cytosol according to the
signals encoded in their 3’ UTR regions, and are eventually recognised by ribo-
somes and translated [Alberts et al., 2002]. Such binding of ribosomes to the mR-
NAs is in direct competition with mRNA decay, a process that starts as soon as
transcripts are exported into the cytosol and which consists of the gradual short-
5
Chapter 1. Introduction
ening of the polyA tail. Once the polyA tail reaches a critical length, mRNAs are
ultimately degraded, either through the continuation of the digestion from the 3’
end or through the removal of the 5’ cap (i.e. decapping) and subsequent 5’ to
3’ decay [Schoenberg and Maquat, 2012]. Alternatively, cytosolic polyadenylation
can also occur, thus having a positive impact on the mRNA half life [Villalba et al.,
2011]. Altogether, this evidences a critical role for both the 5’ and 3’ UTRs in
regulating mRNA stability and translation efficiency.
1.2 The splicing reaction
Splicing was first discovered in 1977 by Phillip Sharp and Richard J. Roberts, who
independently observed that genes could be encoded across split segments in the
DNA. In their experiments, they hybridised adenoviral mRNAs with complemen-
tary single stranded DNA fragments, and following observation with electron
microscopy (EM), they detected alternate double stranded and single stranded
stretches in the resulting hybrid. They concluded that certain regions of the mRNA
are removed during its maturation (i.e. the introns), hence bringing together sep-
arate parts of the RNA [Berget et al., 1977; Chow et al., 1977]. This observation
was soon extended to all domains of life, even though splicing is most prevalent
in eukaryotes (although some instances of splicing have been detected in prokary-
otes, they lack the major pathway through which this process is achieved) [Alberts
et al., 2002]. Both Sharp and Roberts were awarded a Nobel Prize in 1993 for their
contribution.
In the next section, I describe further details regarding this RNA processing step,
with emphasis on the key players and steps required for its completion. I also
discuss the potential of this process for the diversification of the message encoded
in the DNA, and outline the elements that contribute to its regulation.
1.2.1 Key players in mRNA splicing
In eukaryotes, splicing is most commonly carried out through the spliceosomal
pathway, whereby a large complex of proteins and RNAs orchestrates the process
of intron removal. Such complex is known as the spliceosome, and has been cate-
gorised as one of the most complicated machineries in the cell [Nilsen, 2003]. Two
different types of spliceosomes have been identified (i.e. the major and minor),
which differ in their components and the properties of the introns that they tar-
6
Chapter 1. Introduction
get. Specifically, the major spliceosome is involved in >99% of the splicing events,
and is responsible for the removal of introns that harbour consensus splice site
sequences (i.e. canonical splicing; Figure 1.2a) [Matera and Wang, 2014]. By con-
trast, a small set of introns display sequences that differ from the consensus ones,
and are targeted instead by the minor spliceosome (i.e. non-canonical splicing)
[Turunen et al., 2013].
In addition to spliceosomal introns, a separate class of introns that undergo splic-
ing in a protein-independent fashion have also been detected. Such introns, re-
ferred to as self-splicing introns, are able to mediate the splicing reaction through
rearrangements in the RNA structure by acting as ribozymes [Alberts et al., 2002].
While these are rarely detected in eukaryotes, they contribute to the splicing of
some organelle genes and rRNAs, thus providing evidence that splicing is not lim-
ited to mRNAs (i.e. in fact, other non-coding RNAs such as tRNAs, micro-RNAs
and long non-coding RNAs can also undergo such reaction). Their existence is
considered to support the RNAworld hypothesis, which states that self-replicating
RNA constituted the initial building blocks of life [Robertson and Joyce, 2012].
Given the relevant contribution of the major spliceosome in catalysing the removal
of the vast majority of introns, this section will further focus on the process of
canonical splicing.
1.2.2 Splicing by the major spliceosome
The major spliceosome is a highly dynamic complex that undergoes significant
conformational and compositional rearrangements during the several steps of the
splicing reaction [Will and Luhrmann, 2011]. It is composed of five different small
nuclear RNA molecules (snRNAs: U1, U2, U4, U5 and U6), as well as several
hundreds of proteins. Each of these RNA molecules associate with several pro-
teins and form complexes called small nuclear ribonucleoproteins (snRNP). Such
snRNPs form the core of the spliceosome and are directly involved in the recog-
nition of splice sites and branch-point sequences, as well as the catalysis of the
splicing reaction. Indeed, one of the most striking features about the spliceo-
some is that the actual intron removal reaction is catalysed by RNA molecules
rather than proteins [Fica et al., 2013]. More specifically, splicing is the result of
two transesterification reactions, which are based on nucleophilic attacks between
RNA nucleotides (Figure 1.2b) [Will and Luhrmann, 2011]. In the first reaction,
7
Chapter 1. Introduction
the 5’ exon is cleaved from the intron through a nucleophilic attack by the 2’ hy-
droxyl group of the A branch-point residue on the phosphate group of the GU
dinucleotide at the 5’ splice site, thus forming a lariat intermediate. Next, the two
exons become ligated via a similar reaction that involves the 3’ hydroxyl group of
the previously released exon and the phosphate group of the last nucleotide of the
3’ end of the intron, thus causing the release of the intron in the form of a lariat.
However, before the actual intron removal can occur, the active catalytic site of
the spliceosome needs to be created, an event that requires many changes in its
composition and conformation (Figure 1.2c) [Matera and Wang, 2014]. The first
step towards the accomplishment of the splicing reaction is splice site recognition.
This task is achieved trough base pairing between the U1 and U2 snRNPS and the
5’ and 3’ splice sites, respectively, and results in the formation of complex E. Sim-
ilarly, the branch-point sequence is also recognised by the U2 snRNP, which even-
tually interacts with the U1 snRNP and forms the pre-spliceosome (complex A),
thus bringing together both splice sites. Further recruitment of a pre-assembled
tri-snRNP formed by the U4, U5 and U6 snRNPs (the U4-U6·U5 tri-snRNP) leads
to the creation of complex B, which then becomes activated as a result of several
conformational and compositional changes. The activated complex B (complex B*)
mediates the first catalytic step in the splicing reaction and leads to the creation of
complex C, which contains the detached exon (exon 1) and the lariat intermediate
(intron - exon 2). Following extra rearrangements, complex C then performs the
second catalytic step and forms a post-spliceosomal complex, which contains the
spliced exons and the lariat. The latter is eventually released, together with the
remaining snRNPs, which are then recycled for the next round of splicing. Impor-
tantly, these steps are followed by the binding of a new complex of proteins to the
newly created exon junction (i.e. the exon junction complex, ECJ), which mark the
successful completion of splicing at that specific location and further contribute in
shaping the fate of the mRNA molecule.
8
Chapter 1. Introduction
5’ splice site branch
point
3’ splice site
EXON5’ 3’INTRON EXON
polyY
tract
(a)
G
U
A AG
GU A AG
5’ 3’
OH
first step
OH
G
U
A AG
second step
lariat
spliced mRNA
(b)
TSS
U1
complex E
U2
pol II
complex B*
complex C
(catalytic
spliceosome)
post-spliceosomal
complex
lariat
first step
second step
complex A
(pre-spliceosome)
U1
U4
U6 U4
U5complex B
(pre-catalytic
spliceosome)
5’ 3’(c)
(n)
9
Chapter 1. Introduction
Figure 1.2| Splicing by the major spliceosome. Adapted from Will and
Luhrmann [2011].
(a) Core splicing signals recognised by the major spliceosome. The spliceosome recog-
nises several core signals in the pre-mRNA: the 5’ and 3’ splice sites, the branch-
point sequence (typically located between 15 and 50 nucleotides upstream of the
3’ intron) and the polypyrimidine tract. Here, R indicates a purine (A or G), Y
represents a pyrimidine (U or C) and N refers to any nucleotide. Introns that
harbour these consensus sequences are referred to as U2-type introns, because
they are recognised by the U2 snRNP.
(b) Steps in the splicing reaction. Splicing is the result of two transesterification re-
actions that involve the nucleotides from the pre-mRNA and snRNA molecules.
(c) Spliceosomal rearrangements during splicing. As transcription proceeds, several
components of the spliceosome are transferred from the polymerase tail to the
nascent pre-mRNA, thus facilitating the process of splice site recognition. Fol-
lowing this step, the spliceosome undergoes several compositional and confor-
mational changes that lead to the formation of the catalytic site, the cleavage of
the intron and the eventual release of the splicing products.
1.2.3 Diversifying the message: alternative splicing and beyond
During the splicing process, particular exons might become excluded from the
final mRNA product, and similarly, some introns might fail to be removed, thus
leading to the formation of alternative mRNA products from a given genomic
locus (Figure 1.3). This process is known as alternative splicing and plays a key
role in the diversification of the message encoded within a gene [Kornblihtt et al.,
2013].
In humans, over 95% of multi-exon genes have been detected to result in more
than one transcript, and such observation has been suggested as an explanation
for the low number of genes compared to other lower eukaryotes (e.g. the human
genome contains only ⇠30% more genes than that of Caenorhabditis elegans) [Pan
et al., 2008; Wang et al., 2008]. Given the potential differences in biological func-
tion between the resulting alternative transcripts, alternative splicing can result in
the generation of proteins with different biological function, structure, localisation
and interaction capabilities [Keren et al., 2010; Nilsen and Graveley, 2010]. In
this context, the detection of splicing products at the protein level confirms the
potential of such process in increasing the protein repertoire [Tress et al., 2008b].
On the other hand, alternative splicing of pre-mRNAs can also contribute to the
regulation of expression levels, through the formation of transcripts that will be
targeted by the nonsense-mediated decay pathway, as well as non-coding mRNA
10
Chapter 1. Introduction
Exon skipping
Alternative 3’ss selection
Alternative 5’ss selection
Intron retention
Mutually exclusive exons
Alternative promoters
Alternative polyA
(a)
(b)
(c)
(d)
(e)
(f )
(g)
Figure 1.3| Mechanisms for the formation of alternative transcripts from the
same genomic locus. Differences in the execution of the splicing reaction can
lead to message diversification (a-e). Similarly, the usage of alternative first and
last exons also emerges as a common mechanism for the generation of alternative
transcripts from the same gene (f-g). Adapted from Keren et al. [2010].
products that arise from intron retention events [McGlincy and Smith, 2008; Yap
et al., 2012]. Finally, it has also been suggested that a considerable amount of the
detected alternative splicing products result simply from noisy splicing and have
no function at all [Melamud and Moult, 2009].
The biological significance of alternative splicing becomes evident in light of the
detection of tissue-specific events [Buljan et al., 2012; Ellis et al., 2012; Merkin et al.,
2012; Wang et al., 2008], as well as its relevant role in dynamic processes such as
development [Kalsotra and Cooper, 2011] and cellular differentiation [Trapnell
et al., 2010]. Hence, the annotation and functional characterisation of alternative
mRNA products is an important task. A prominent effort towards that goal is the
GENCODE project, which aims to annotate all evidence-based features in the hu-
man genome [Harrow et al., 2012]. Similarly, several methods have been devised
in order to predict the impact of alternative splicing at the protein level (e.g. AP-
PRIS [Rodriguez et al., 2013], AS-EAST [Shionyu et al., 2012], MAISTAS [Floris
et al., 2011], AltAnalize [Emig et al., 2010]).
11
Chapter 1. Introduction
Nonetheless, message diversification does not stop with alternative splicing. For
example, exons from different genes can be combined during the splicing reaction
in a process called trans-splicing [Lasda and Blumenthal, 2011]. Similarly, RNA
editing constitutes a separate mechanism through which the information con-
tained within an mRNA can be altered, via the substitution, insertion or deletion
of specific nucleotides [Maas, 2012]. Finally, post-translational modifications also
contribute in shaping protein diversity. Amongst those, the removal of internal
protein segments (i.e. inteins) emerges as an interesting example, given its analogy
to alternative splicing [Volkmann and Mootz, 2013].
1.2.4 The regulation of splicing
Apart from the core splicing signals, further elements contribute to the definition
of exon-intron boundaries and the regulation of splicing. This is the case for
other cis-regulatory sequences that are typically present in the pre-mRNA
(i.e. splicing regulatory elements, SREs), which can vary in terms of location and
effect (Figure 1.4a) [Matera and Wang, 2014]. In general, SREs contribute to the
recruitment of trans-acting splicing factors (SFs), a set of proteins that can act as
repressors or activators of splicing, typically by influencing spliceosome assembly.
A prominent example of the role of SREs is their contribution to the recognition
of exon-intron boundaries through a process called exon definition (Figure 1.4b)
[De Conti et al., 2013]. Such mode of splice site recognition is especially common
in higher eukaryotes, where intron size exceeds that of exons. This could lead
to splicing errors because of the existence of cryptic splice sites. Hence, a class
of SFs called SR proteins (Serine-Rich proteins) promote the binding of snRNPs
to the splice sites located at both ends of the same exon, by binding to exonic
splicing enhancers (ESE). As a result, a cross-exon recognition complex is formed,
which will eventually lead to intron-spanning interactions through spliceosomal
rearrangements. Conversely, in the case of short introns and in lower eukary-
otes, intron definition emerges as the prevalent mode of splice site recognition
(Figure 1.4b) [De Conti et al., 2013]. In this case, splice sites situated on both ends
of the same intron are directly identified without the help of specific SFs.
12
Chapter 1. Introduction
exon
3’ splice site
exonic splicing
enchancer (ESE)
exonic splicing
silencer (ESS)
intronic splicing
enhancer (ISE)
intronic splicing
silencer (ISS)
5’ splice site
intron
(a)
(b)
EXON DEFINITION INTRON DEFINITION
SR
U2 U1 U1 U2
Figure 1.4| The regulation of splicing.
(a) Cis-acting sequences involved in the regulation of intron removal. In addition to
the core splicing signals (i.e. 5’ splice site, branch-point and 3’ splice site), sev-
eral regulatory sequences shape the splicing decision by recruiting trans-acting
splicing factors (SFs). Common SFs include SR proteins and hnRNPs, which
typically promote and inhibit splicing, respectively. However, SRE activity is
highly context dependent: albeit recruiting the same SF, the same sequence can
have opposite roles, depending on whether it is located within exonic or intronic
boundaries. Adapted from Matera and Wang [2014].
(b) Exon vs. intron definition. Intron definition is the prevalent mode of splice site
recognition in lower eukaryotes, which consists of the pairing of the 5’ and 3’
splice site located at each end of the intron. In higher eukaryotes, the longer
intron size could lead to the inclusion of cryptic splice sites, and exon definition
is used instead. In this case, the recognition of the splice sites that surround a
given exon is mediated by SR proteins, which promote the formation of a cross-
exon recognition complex. Next, further spliceosomal rearrangements ensure
intron-spanning interactions. Adapted from Ast [2004].
Together with the competition amongst splice sites and cis-acting SREs based
on their sequence composition, the accessibility to those elements also plays a
key role in alternative splicing regulation. Such accessibility can be influenced
by the pre-mRNA secondary structure, chromatin arrangements and nucleosome
positioning, and can be also dynamically controlled as a result of the coordination
of the transcription and splicing processes [Brown et al., 2012; Plass and Eyras,
2014]. In humans, most of the splicing events occur before transcription termi-
nation, a phenomenon known as co-transcriptional splicing [Tilgner et al., 2012].
This implies that transcription elongation rates can have an impact on splice site
choice: for example, a slow elongation will provide a window of opportunity
for the recognition of weak splice sites, whilst a fast elongation will generally
13
Chapter 1. Introduction
promote the recognition of strong splice sites instead [Bentley, 2014]. Finally, the
regulation of splicing decisions is not limited to the role of specific SFs, since
fluctuations in the concentration of core components of the spliceosome are also
known to influence the splicing outcome [Saltzman et al., 2011].
Overall, the above mentioned processes guarantee that splicing occurs in an ac-
curate albeit flexible fashion. The accuracy of splicing is further increased by
the many rearrangements that are required before the actual intron removal reac-
tion can occur, and splicing errors are eliminated by the nonsense-mediated decay
pathway. On the other hand, the accumulation of splice site mutations or the
alteration in the function of spliceosomal components can lead to serious pheno-
typic consequences and, in fact, dysregulation of splicing has been linked to many
diseases, including cancer [Ladomery, 2013; Padgett, 2012; Tazi et al., 2009].
1.3 Studying the transcriptome with RNA sequencing
In the past few years, RNA sequencing (RNA-seq) has become the method of
choice for the study of transcriptome composition [Mortazavi et al., 2008; Wang
et al., 2009]. Compared to microarrays, which constituted the first technology for
the high throughput comparison of expression levels across conditions, RNA-seq
offers a much bigger dynamic range to study gene expression patterns, and en-
ables a much broader set of analyses without the need for intricate experimental
designs [Malone and Oliver, 2011]. For example, besides standard differential
gene expression analysis, popular applications of RNA-seq comprise the identi-
fication of novel transcribed regions, including fusion genes, the deconvolution
of allele specific expression, and, as further explored in this thesis, the possibility
to estimate transcript expression levels and to study differential splicing across
conditions.
Since the introduction of the first sequencing machines in 2005, this technology
has seen the rise and fall of many companies; however, following the acquisition
of Solexa, Illumina’s platforms have consolidated as the most commonly used.
The reason behind such wide adoption of Illumina’s systems is the large volume
of information obtained from a typical sequencing run (i.e. sequencing depth),
which, at a good ratio with the cost, compensates for the lower accuracy compared
to other competitors [Mardis, 2013]. Thus, although microarrays can still be a
14
Chapter 1. Introduction
cheaper option to perform routine differential expression analysis at the gene
level, the larger scope of applications and the decrease in the costs of sequencing
(just announced to have reached the target of 1,000$ per human genome by
Illumina while writing this thesis) explain the increasing popularity of RNA-seq.
In this section, I introduce the typical steps required to sequence a transcriptome
with an Illumina platform, since this is the one that has been used to produce all
the datasets analysed here. Moreover, I provide a detailed description of the most
commonly used methods to study the transcriptome composition from RNA-seq
data, with special emphasis on the analysis approaches used within the different
chapters.
1.3.1 A typical sequencing workflow
The first step in transcriptome sequencing is library preparation, and consists of
obtaining the starting material and converting it into a cDNA library that can be
loaded into the sequencing machine (Figure 1.5) [van Dijk et al., 2014]. Following
RNA extraction, the RNA species of interest are typically enriched through
either polyA selection or ribodepletion. In both cases, the aim is to diminish
the concentration of rRNAs, i.e. the most abundant species of RNA in the cell.
With the first method, this is achieved through the use of oligo-dT beads, which
enable the specific extraction of polyAdenylated RNAs, hence ensuring a good
representation of mRNAs (Figure 1.5 - step 1). Conversely, ribodepletion relies
on the use of ribonucleases to specifically digest rRNAs, and has the advantage
of not restricting the analyses to a specific type of RNA. Indeed, the term total
RNA is typically used to refer to datasets produced with such protocol, while
those obtained with the former method are commonly known as polyA-selected.
Due to the simpler protocol and its lower price, polyA selection emerges as
the most popular choice amongst the currently available RNA-seq datasets,
with the notable exception of those studies aimed at characterising non-coding
RNA species, which typically lack a polyA tail. The extracted RNA is then
fragmented via hydrolysis with divalent cations and retro-transcribed into double
stranded cDNA by using random hexamer primers, since the sequence of the
obtained fragments is not known at this point (Figure 1.5 - step 2). These steps
are followed by the ligation of adapter sequences at both ends of each cDNA
fragment (Figure 1.5 - step 3). Such adapters satisfy two different purposes: on
the one hand, they enable the hybridisation of those fragments into the flow cell,
15
Chapter 1. Introduction
where the sequencing takes place; on the other hand, they serve as primers for
the sequencing reaction. Then, the resulting cDNA fragments are size-selected
through gel electrophoresis to fit within the range required by the sequencing
machine (typically 300-500 bp). Fragments outside this range will be missed;
hence the existence of alternative protocols for the study of small RNAs [Zhuang
et al., 2012]. Finally, the cDNA library is amplified by PCR.
AAAAA
polyA selection1
Fragmentation
Random hexamer priming
First and second strand synthesis
2
Adapter ligation
Size selection
PCR amplification
3
Hybridisation
Bridge amplification4
Sequencing5
A1 SP1
A2SP2
A T G
...
A
Figure 1.5| Overview of library preparation and sequencing steps in an Illu-
mina platform. A typical paired-end workflow is illustrated here, which consists
of ligating different adaptors at each end of the initial cDNA molecule. This en-
ables sequencing each cDNA fragment from both ends, in two separate reactions,
and has further advantages for the downstream bioinformatic analyses compared
to single-end approaches. Adapted from Mardis [2013].
Once the library preparation procedure has finished, samples can be loaded
into a flow cell for sequencing [Mardis, 2013]. Such flow cell is saturated with
adapters that are complementary to the ones ligated at both ends of the cDNA
fragments, consequently promoting the hybridisation of the denatured double
16
Chapter 1. Introduction
strand molecules. After this step, the starting material needs to be amplified once
again in order to increase the signal for the sequencing reaction, this time through
bridge amplification (Figure 1.5 - step 4). Such process consists of the synthesis
of fragments that are complementary to the hybridised cDNA molecules, which
will in turn bend and hybridise with adjacent adapters, thus enabling subsequent
rounds of synthesis. As a result, a large number of clusters with identical
sequences will be formed, now ready to undergo sequencing. Illumina platforms
rely on sequencing by synthesis to read the base pair composition of each cDNA
cluster (Figure 1.5 - step 5) [Bentley et al., 2008]. This reaction is based on the use
of modified versions of the four bases, which differ from the standard nucleotides
in the fact that they incorporate a reversible terminator, as well as a fluorescent
dye. Hence, during each sequencing cycle, and following the addition of the
necessary reagents, elongation will be blocked after a successful base incorpora-
tion, the identity of which can be recorded by measuring its fluorescent signal.
Repetition of this process will lead to a set of images, which after interpretation
with a base calling software, will be converted into a set of sequences or reads
[Das and Vikalo, 2013]. Such reads represent the set of molecules expressed in the
initial sample, and their length corresponds to the number of cycles performed
during the sequencing reaction. Eventually, the obtained sequence information,
together with the probability of a wrong base call at each given position of the
read (i.e. Phred score), are stored in a plain text file in FASTQ format [Cock et al.,
2010].
RNA-seq is not such an established technology as microarrays, and in spite
of its many advantages, it still has some challenges. For example, regarding
library preparation, it is known that the random hexamer priming step is not as
random as initially proposed, since certain fragments have been observed to be
preferentially converted to cDNA due to sequence composition [Hansen et al.,
2010]. In the same category of sequence-dependent biases, the PCR amplification
step has also been described to lead to differential amplification of fragments with
higher or lower GC content [Benjamini and Speed, 2012], and it is known that
failure to block the elongation reaction or to remove the fluorescent dye during
the sequencing step can lead to wrong base calls [Metzker, 2010]. In most cases,
the identification of such biases has been accompanied by the introduction of
alternative protocols or analysis methods to overcome them. For example, several
algorithms now try to take into account potential biases derived from the random
17
Chapter 1. Introduction
hexamer amplification step (e.g. Cufflinks [Trapnell et al., 2010], MMSEQ [Turro
et al., 2011]). Alternative library preparation methods have also been proposed
to account for PCR bias, whereby random barcodes (i.e. molecular identifiers)
are used to quantify the absolute number of molecules [Shiroguchi et al., 2012].
Finally, some downstream analysis algorithms also incorporate information on
the probability of a wrong base call at a given position of the read, as reported by
the Phred score (e.g. Kim et al. [2013]).
On the other hand, alternative library preparation strategies can also add further
information to the experiment. This is the case of strand-specific protocols, which
are able to provide information on the strand from which each read originates
[Levin et al., 2010]. Similarly, multiplexing emerges as a widely used approach
to optimise the amount of data that can be obtained from each sequencing run,
by enabling pooling of several samples into a single lane of the flow cell through
the use of sequence identifiers (i.e. sample-specific barcodes) [Wong et al., 2013b].
Lastly, a very common strategy to overcome limitations on the read length and try
to span larger regions consists of sequencing each cDNA fragment from both ends
(i.e. paired-end sequencing, as opposed to the single-end strategy), which can be
achieved through the use of modified adapters (Figure 1.5 - step 3) [Mardis, 2013].
1.3.2 Read mapping strategies
The next step in a typical RNA-seq analysis pipeline consists of identifying, for
each read, the genomic region from which it has originated. In RNA-seq, this
task is equivalent to discovering the loci that are expressed in a given sample. In
general, two different strategies exist to perform this task: on the one hand, reads
can be aligned to the reference genome or transcriptome, provided that such
information is available for the species of interest; on the other hand, they can be
directly assembled into contigs (i.e. contiguously expressed regions) with the aim
of reconstructing the set of expressed transcripts. The first strategy constitutes a
much simpler approach, and it is typically the method of choice when working
with model organisms.
Independently of the strategy used, read mapping is typically the most time con-
suming step of the analysis workflow, and the available tools make use of heuristic
parameters such as the maximum number of allowed mismatches per read in or-
der to speed up this task. Such processing can lead to information loss given a
18
Chapter 1. Introduction
decrease of quality at the 3’ end of the read, which emerges as a common profile
when working with Illumina platforms due to the increased difficulty in inter-
preting the fluorescent signal as sequencing cycles accumulate [Minoche et al.,
2011]. Thus, in order to avoid such reads being discarded, it is often useful to
first perform a quality control and pre-filtering step, whereby read sequences can
be shortened (i.e. trimmed) in terms of their quality (e.g. Andrews [2010]). Simi-
larly, reads with overall low quality can be also removed, in order to speed up the
subsequent mapping process.
1.3.2.1 Alignment to the genome or transcriptome
A commonly used approach in the cases where a reference genome is available,
is to align the reads directly to that sequence. Similarly, reads can be aligned
to the transcriptome instead, provided that a good annotation exists. The main
advantage of using this second strategy is that the alignment task is simplified
due to the lack of intronic sequences; but this comes at the price of limiting the
number of downstream analysis that can be performed (e.g. alignment to the
transcriptome is not compatible with the identification of novel expressed regions
nor the study of intronic expression levels). Thus, a good compromise is the use
of hybrid approaches, as implemented in TopHat [Kim et al., 2013].
TopHat is a read mapping tool specially intended for RNA-seq data, since it en-
ables alignment of the reads to the genome while taking into consideration the
existence of splice junctions (Figure 1.6). It is based on Bowtie [Langmead and
Salzberg, 2012], an independent algorithm for the alignment of short reads, and
its main strength is the ability to detect exon-exon junctions without the need for
any a priori knowledge on the annotation. However, the search can be simplified by
providing such information, and in that case TopHat will first attempt to map the
reads to the derived transcriptome. Those that fail to align will be subsequently
queried against the genome (Figure 1.6 - step 1). Alternatively, reads can also be
mapped to the genome directly (Figure 1.6 - step 2). In both cases, the goal is to
assemble the initially aligned reads into exons, which might eventually become
connected through spliced alignment (Figure 1.6 - step 2). Reads that fail to align
in this initial phase, as well as those that map with low alignment scores, are sub-
sequently used to build a database of possible splice junctions, by splitting them
into smaller segments and re-aligning those independently (Figure 1.6 - step 3). In
this context, a splice junction is reported whenever a read appears to span multi-
19
Chapter 1. Introduction
ple exons, i.e. in the cases in which an internal segment fails to align, or when two
consecutive segments from the same read do not align contiguously on a given
genomic locus. Next, the identified splice sites and their flanking sequences are
concatenated into a novel transcriptome, which is then used to re-align the set
of unmapped reads (Figure 1.6 - step 4). In the case of paired-end data, each
read is processed separately, and the alignments obtained are evaluated in a final
phase by taking into account additional sources of information such as fragment
length and orientation of the reads. Finally, all the information gathered during
the mapping process is reported in SAM/BAM format [Li et al., 2009].
Mapping to the transcriptome
(optional)1
unmapped
reads
+
+
identified exons
x
identified splice junction
Mapping to the genome
Exon definition2
Splice site identification3
Split read mapping4
Figure 1.6| Overview of the mapping algorithm implemented in TopHat. In
the presence of an annotation file, TopHat uses a hybrid approach to uncover the
genomic loci from which the detected reads could have originated. Alternatively,
TopHat can directly align the reads to a reference genome. In both cases, the
first step consists of identifying a set of expressed exons, and this is followed by
the detection of splice junctions by using information from those reads that span
multiple exons. Adapted from Kim et al. [2013].
20
Chapter 1. Introduction
1.3.2.2 De novo assembly
De novo assembly emerges as an advantageous strategy in the cases where the
species of interest lacks a reference genome. Additionally, it can be used in sit-
uations where the genome composition of a given sample is expected to differ
largely from that of the reference assembly (e.g. cancer samples). The goal here
is to assemble the reads into sets of expressed regions (i.e. contigs), by relying
on their overlap. Nonetheless, the short read length adds to the non-triviality of
the task, and even though the use of paired-end data can simplify this process,
lowly expressed regions are often difficult to solve. In terms of available software,
Trinity [Grabherr et al., 2011] emerges as the most popular tool to perform this
task; however, such methods are not used in this thesis and are covered elsewhere
[Martin and Wang, 2011].
1.3.3 The estimation of expression levels
Once the reads have been assigned to a specific location in the genome or tran-
scriptome, the next step in the RNA-seq analysis pipeline consists of estimating
expression levels for the features of interest, typically genes and transcripts. Sim-
ilarly to the scenarios encountered during the mapping step, the quantification
of expression levels can be achieved by relying on existing information, but it can
also be performed independently from any annotation, thus enabling de novo iden-
tification of transcribed regions (i.e. novel genes or unnanotated transcripts within
known gene loci).
1.3.3.1 Gene expression levels
When working at the gene level, and provided a complete annotation exists, abun-
dance estimation can be easily achieved by counting how many reads overlap each
given locus (Figure 1.7a). Such count-based approach constitutes the starting point
for many downstream analysis algorithms (e.g. DESeq2 [Love et al., 2014], DEXSeq
[Anders et al., 2012]), and can be easily performed with the popular tool htseq-
count [Anders et al., 2014]. However, despite this apparent simplicity, there are
some challenges that need to be considered. First, in order not to over-estimate
expression levels, reads that map to multiple locations in the genome, and which
arise from repetitive or duplicated loci, need to be handled with care. In this
situation, htseq-count adopts the most conservative approach and discards them,
but other alternative strategies have been proposed in order to attempt to keep
21
Chapter 1. Introduction
the information from such multi-mapping reads. Generally, these consist of uni-
formly distributing them to all the mapped positions (e.g. Trapnell et al. [2010]), or
probabilistically assigning them depending on the coverage at each mapping lo-
cus (e.g. Trapnell et al. [2010]; Turro et al. [2011]; first proposed by Mortazavi et al.
[2008]). Second, special attention is required in the case of overlapping features.
htseq-count offers several execution modes to deal with this scenario, even though
in some cases reads remain ambiguously assigned (Figure 1.7b). Finally, despite
not being intended for de novo quantification, htseq-count also gives the user some
flexibility on how strictly the provided feature coordinates should be taken into
account (Figure 1.7b).
c1 = 7 c2 = 6 c3 = 4
gene A gene B
union
A
A
A
ambiguous
ambiguous
intersection
nonempty
A
A
A
ambiguous
A
intersection
strict
no feature
no feature
A
ambiguous
A
(a)
(b)
Figure 1.7| Overview of htseq-count.
(a) Illustration of the read counting concept. Expression estimation with htseq-count
consists of counting the reads that overlap with the features of interest. In this
example, any reads that fall outside the grey areas will not be considered.
(b) The three different execution modes available in htseq-count. htseq-count provides
different counting modes to rescue reads that do not strictly overlap with the
provided coordinates. These modes differ in how strictly the annotation is taken
into account and in the behaviour adopted in the case of overlapping features.
Adapted from Anders et al. [2014].
22
Chapter 1. Introduction
Alternatively, gene expression can be calculated after estimation of transcript
expression levels, by aggregating the corresponding individual transcript abun-
dances, as implemented for example in Cufflinks [Trapnell et al., 2010] and MM-
SEQ [Turro et al., 2011].
1.3.3.2 Transcript expression levels
On the other hand, the task of estimating expression levels becomes far more
complicated when focusing on individual transcripts, since many reads will over-
lap with exons that are shared across multiple isoforms of the same gene. In this
scenario, the question translates into attributing reads to specific transcripts, and
further inference approaches are needed. The available algorithms typically rely
on different sources of information in order to probabilistically estimate transcript
expression levels, the most valuable one being those reads that map uniquely
to one of the annotated transcripts within the loci. Moreover, reads that span
two different exons (i.e. split reads) become especially informative. For example,
splice junctions that involve cassette exons tend to provide unambiguous support
for their inclusion or skipping. This is where paired-end information becomes
most relevant: sequencing both ends of the initial cDNA fragment facilitates
covering larger genomic regions, thus increasing the probability that a given read
pair is mapped across different exons (i.e. spliced reads). Similarly, information
on the fragment length distribution can also be used to deconvolute ambiguous
assignments, by attributing a lower likelihood to those that would require extreme
distances between the paired reads.
One of the most popular tools to estimate transcript expression levels is MISO
[Katz et al., 2010], which formulates this task as a Bayesian inference problem
[Beaumont and Rannala, 2004], whereby the goal is to find a probability distribu-
tion (the posterior) over transcript abundances (Y) given the observed RNA-seq
data (Figure 1.8). Such distribution can be computed in terms of two quantities:
the expectation about the value of Y before observation of the reads (the prior,
set by MISO as a uniform distribution), and the probability of observing the data
given a fixed value of Y (likelihood of the reads). Thus, following sampling across
the space of Y values, all possible assignments of every read to each isoform are
probabilistically evaluated, and this information is subsequently used to refine the
search of the optimal set of Y values that best explain the observed data. Finally,
an estimate of transcript abundances is obtained by calculating the mean over the
23
Chapter 1. Introduction
computed posterior distributions, and confidence intervals are also calculated as
a measure of the certainty of the estimate.
Similar approaches are taken by other alternative methods, including Cufflinks
[Trapnell et al., 2010] and MMSEQ [Turro et al., 2011], both of which are also
used in this thesis. The main difference among the mentioned methods relies
on the implementation of the inference approach, as well as the type of input
required. Similarly to MISO, Cufflinks requires the reads to be mapped to a ref-
erence genome, but relies on a frequentist approach to find the expression levels
that best explain the observed data, which does not allow quantification of the un-
certainty around the obtained expression estimates. On the other hand, MMSEQ
adopts a Bayesian model similar to the one of MISO, but requires mapping to the
transcriptome, which limits the scope of the downstream analysis that can be per-
formed. Furthermore, both MMSEQ and Cufflinks accommodate known sequence
biases in their models, and are also able to retain information from reads that map
to multiple genes.
24
Chapter 1. Introduction
read mapping results
TA
TB
R1
R2
R3
R4
RN
transcript compatibility
matrix
TA TB
... ...
R1 1 1
R2 1 0
R3 0 1
R4 1 1
RN 1 0
inference of transcript abundances
P (     | R1:N ) P ( R1:N |     )  x  P (     )
0 1
posterior
mean
pr
ob
ab
ili
ty
posterior distribution over
       given the reads
0 1
pr
ob
ab
ili
ty
prior over      
(uniform)
likelihood of reads
given     
pr
ob
ab
ili
ty
TA TB
R1
R2
RN
...
Figure 1.8| Overview of the analysis workflow implemented in MISO for the
estimation of transcript abundances. After alignment to the genome, MISO
evaluates the compatibility of each read with all the transcripts annotated within
a given gene. For example, in the scenario depicted here both read 2 (R2) and
read N (RN) can only be detected if transcript A (TA) is expressed, whilst read
3 (R3) uniquely supports transcript B (TB). Inference of expression levels is then
done by calculating a probability distribution (the posterior) over such expression
(Y) given the reads. Following Bayes’ rule, this distribution can be obtained
from the product of two terms: the likelihood of obtaining the observed set of
reads given a fixed value of Y and the expectation on the value of Y before
observation of the data (the prior). Hence, the inference problem translates into
sampling from the space of Y values and evaluating all possible assignments
of each read to each transcript. For example, given the larger number of reads
that support the expression of transcript A in comparison to transcript B, higher
expression of the latter will be probabilistically penalised, and this will contribute
to the preferential assignment of ambiguous reads to the former. Fragment length
information can also be used to deconvolute ambiguous assignments, as it is the
case for read 4 (R4): assigning such read to transcript B would imply an unusual
distance between the paired reads, hence increasing the likelihood that it can be
explained by transcript A. Finally, the overall probability of observing the reads
given the evaluated Y value is obtained by combining the information from all
reads, and this information is further used to calculate the posterior distribution.
Adapted from Wang et al. [2010].
25
Chapter 1. Introduction
1.3.3.3 De novo transcript identification
One of the main advantages of RNA-seq over microarrays is the possibility to
gather information on novel expressed loci in a more high throughput manner. In
this context, Cufflinks [Trapnell et al., 2010] emerges as one of the most popular
tools to achieve this task, thus complementing its aforementioned quantification
capabilities (Figure 1.9). By relying on the output provided by TopHat [Kim et al.,
2013], the Cufflinks assembler first identifies the expressed loci (i.e. genes) present
in a given sample. Then, for each of them, it evaluates the observed data in
search of a set of incompatible reads, i.e. reads which have necessarily originated
from different transcripts. This step is followed by the construction of an overlap
graph, whereby each read represents a node and each edge is used to connect
compatible reads. Finally, Cufflinks tries to identify the minimum set of paths
(i.e. transcripts) that explain such graph.
assembled transcripts
overlap graph
Figure 1.9| Overview of the de novo transcript identification algorithm imple-
mented in Cufflinks. Three different incompatible sets of fragments exist in the
example depicted here (i.e. yellow, blue, red). Black reads represent those that
are compatible with any of the sets. Following the construction of an overlap
graph that indicates the possible connections amongst the observed fragments,
Cufflinks assembles the data into the minimum set of paths required to explain
such graph. The identified transcripts can then be used for subsequent down-
stream analyses, including estimation of transcript expression levels. Adapted
from Trapnell et al. [2010].
26
Chapter 1. Introduction
Alternatively, Cufflinks can also be used in conjunction with the existing anno-
tation (i.e. annotation based transcript assembly). In this scenario, the annotated
transcripts are used to generate artificial data points that are combined with the
observed data during the assembly process, hence serving as a guide. Following
transcript assembly, the novelty of the obtained transcripts is evaluated by
comparing them to the annotation, and those that differ are reported as novel.
Overall, and similarly to de novo read assembly, the task of transcript assembly is
not a trivial one, although it becomes simplified by the existence of mappings to
the genome. Similarly to the situation encountered in the former scenario, lowly
expressed regions are difficult to analyse, given that in those cases the algorithm is
less likely to find a unique solution for the constructed graph. Finally, alternative
start and end sites also become difficult to characterise, since all the paths are
extended to the maximum.
1.3.4 Read count normalisation
Independently of the quantification approach followed, the result from such step
is going to be an estimate on the number of reads that can be attributed to a
certain feature, further referred to as counts. These counts will be proportional
to the expression levels of the feature of interest; however, they will also depend
on the length of the feature and the sequencing depth of the experiment (i.e. the
total number of sequenced reads). In addition, further experimental biases have
also been detected to have an impact on the counts detected in certain loci, as it
is the case for the previously mentioned sequence-dependent biases. Altogether,
these observations illustrate the need for normalisation in order to enable the
comparison of read counts across different samples and features.
One of the measures commonly used to report expression levels derived from
RNA-seq data is the Reads/Fragments per Kilobase per Million mapped reads
(RPKMs or FPKMs, in the case of single-end or paired-end data, respectively)
[Mortazavi et al., 2008]:
27
Chapter 1. Introduction
µˆij =
kij
Njli
· 109
where:
µˆij = normalised expression for gene i in sample j
kij = observed counts for gene i in sample j
Nj = total number of reads in sample j
(sequencing depth)
li = length for gene i
Given that this measure takes into account both the length of the feature of
interest and the total number of mapped reads in the dataset (i.e. sequencing
depth), it has become established as an intuitive measure of expression levels.
However, this method is based on the assumption that the overall RNA levels are
similar across samples, and hence it might fail to properly estimate the normal-
isation factors in cases where the compared libraries differ in their composition
[Robinson and Oshlack, 2010]. For example, let us imagine two samples that
express a common set of genes at similar levels, and let us consider an extra
small set of highly expressed genes in one of them. Since the sequencing step
can be understood as a sampling process, where it is more likely to detect reads
from genes with high expression levels, the signal from commonly expressed
genes will be lower in the latter sample, provided that both are sequenced
at a similar depth. Hence, using the above mentioned normalisation method
would lead to the interpretation that most genes undergo changes in expression
across conditions; whilst the observed differences could be better explained by
the isolated differential expression of the few non-overlapping genes (Figure 1.10).
The above described scenario evidences the need for more robust normalisation
methods, especially when the goal is to compare across libraries (e.g. in down-
stream analysis such as differential expression/splicing). An example of those
methods is the one provided within the DESeq2 Bioconductor package [Love et al.,
2014]. Such algorithm starts by calculating a geometric mean for each gene in or-
der to capture the variability of the observed measurements across all the libraries
(similar to obtaining a reference sample). Then, these values are used to nor-
malise the initial counts, and finally, the library-specific normalisation factors are
obtained from the median of the calculated ratios:
28
Chapter 1. Introduction
sj = median
i: kRi 6=0
kij
kRi
where:
sj = size factor for sample j
kij = observed counts for gene i in sample j
kRi = geometric mean for gene i across
the m samples: (’mv=1 kiv)
1/m
Other tools (e.g. Cufflinks [Trapnell et al., 2010] and MMSEQ [Turro et al., 2011])
enable also the correction of sequence-dependent biases, by attributing a weight to
each position in the expressed loci based on its sequence context. The calculated
weights are then used during the abundance inference step in order to model the
non-uniform location of reads along the transcripts [Li et al., 2010; Roberts et al.,
2011].
sample A
gene 1
gene 2
gene 100
gene 101
total number
of reads
...
100
100
100
0
10,000
...
80
80
80
2,000
10,000
...
sample B
A+B B
Figure 1.10| Limitations on the use of RPKMs for differential expression anal-
ysis. In this example, sample A and B express a common set of genes at sim-
ilar levels; however, sample B also contains a highly expressed gene that is not
present in the former. If both samples are sequenced at the same depth, the
observed counts for the common set of genes will be lower in sample B, given
the limited number of reads. In the context of differential expression analysis,
the RPKM normalisation method would lead to the interpretation that all genes
are differentially expressed, since it assumes homogeneity in library composi-
tion. In the scenario represented here such assumption does not hold, and the
observed differences are better explained by the isolated differential expression
of the gene unique to sample B. This highlights the need for more robust nor-
malisation methods when comparison across libraries is attempted.
29
Chapter 1. Introduction
1.3.5 Differential expression
One of the most common uses of RNA-seq data is the assessment of differences
in expression levels across conditions. Provided the corresponding counts have
been obtained, such analysis can be performed both at the gene and transcript
levels (differential gene/transcript expression), and one of the most popular tools
to achieve that is the Bioconductor package DESeq2 [Love et al., 2014].
In general terms, DESeq2 relies on the use of Generalised Linear Models (GLMs)
of the Negative Binomial (NB) family in order to address the significance of the
detected changes in expression levels. The implemented analysis workflow first
consists of normalising the observed counts in order to enable their comparison
across libraries (Figure 1.11 - step 1), as covered in the previous section. Next, for
each gene, an estimate on the amount of variability that can be expected on the
measurements from biological replicates is calculated (Figure 1.11 - step 2), and
finally, the differential expression test is performed (Figure 1.11 - step 3).
As with any counting process, one would not expect the detected counts for a
given gene to be exactly the same across all observations from a single condition.
Hence, the underlying question in differential expression analysis is whether the
counts observed across the two evaluated conditions are similar enough to be de-
rived from the same distribution (null hypothesis), or whether they are better ex-
plained by two separate ones (alternative hypothesis). Given the nature of the data
obtained from RNA-seq experiments, the Poisson distribution was first proposed
to model noise intrinsic to the counting process [Marioni et al., 2008]. However,
it was soon shown that while this approach works well for technical replicates, it
underestimates the variability in measurements across biological replicates [An-
ders and Huber, 2010; Robinson et al., 2010]. As a result, the negative binomial
distribution has been widely adopted to account for such over-dispersion:
30
Chapter 1. Introduction
kij = NB(µij, s2ij)
µij = sjqij
where:
kij = observed counts for gene i in sample j
µij = distribution mean for gene i in sample j
s2ij = dispersion for gene i in sample j
sj = size factor for sample j
qij = quantity proportional to the concentration
of cDNA fragments for gene i in sample j
The identification of the amount of variation across biological replicates is an
essential step in the aforementioned workflow, since it enables for the evaluation
of the significance of any changes detected. However, because of the low
number of replicates typically available in RNA-seq experiments, such variation
cannot be directly calculated, and needs to be estimated from the data instead.
Following the assumption that genes with similar expression levels have similar
sample-to-sample variance, DESeq2 obtains gene-specific variance estimates by
taking into account not only the observed dispersion for each given gene, but also
that of all other genes. This is achieved by fitting a regression curve to the data
(i.e. average normalised counts vs. observed dispersion), which is subsequently
used to modify the observed dispersion values.
Finally, by further decomposing the mean into a function of independent vari-
ables (i.e. the covariates), it is possible to take all known sources of variation into
account:
log2(µij) =Â
r
xjrbir
where:
µij = mean for gene i in sample j
xjr = independent variable r for sample j
bir = coefficient for gene i and variable r
Altogether, the algorithmic approach behind DESeq2 consists of fitting the model
defined in the aforementioned equations for both the null and alternative hy-
potheses (reduced vs. full model, respectively), followed by the evaluation of the
significance of the coefficient of interest (Figure 1.11).
31
Chapter 1. Introduction
ex
pr
es
sio
n 
le
ve
l
samples
Normalisation1
Differential expression test3
Variance estimation2
mean of normalised counts
di
sp
er
sio
n
Null hypothesis
log μij = β0 log μij = β0 + xj βT
fre
qu
en
cy
observed counts
Alternative hypothesis
fre
qu
en
cy
observed counts
Kij ~ NB(μij, αi)
Figure 1.11| Overview of the steps required for differential expression analysis
using DESeq2. First, reads are normalised in order to enable comparison across
libraries. Next, for each gene, an estimate of the amount of variability that can be
expected across biological replicates is obtained. Given the typical low number
of biological replicates in RNA-seq experiments, it is not possible to obtain such
information directly from the data. Hence, DESeq2 relies on the observed dis-
persion from all genes instead (black dots), and by fitting a regression curve that
explains the dependence of the dispersion on the mean (red line), further modi-
fies the initial values through a process called shrinkage estimation of variability
(blue dots). Finally, the obtained information is used to test the hypothesis that
the observed counts originate from different distributions (alternative hypothe-
sis). DESeq2 uses the negative binomial distribution to model both stochastic
and biological noise, and it further relies on the use of GLMs to take all known
sources of variation into account. Here, the independent variable xj represents
the experimental condition, and can be arbitrarily set to 0 in the case of controls
and to 1 in the case of treated samples. Under the alternative hypothesis, this
enables the existence of two different distributions that can explain potential dif-
ferences in expression levels of the studied gene i. Conversely, under the null
hypothesis, there is no need for such term, since in this case all counts arise from
the same distribution. Hence, the general idea behind the differential expression
test consists of deciding whether the inclusion of the variable xj adds meaningful
information to the model, and it translates into assessing the significance of the
bT coefficient (highlighted in grey).
32
Chapter 1. Introduction
Similarly to DESeq2, further available tools rely on the use of read counts in or-
der to make assessments of differential gene expression (e.g. edgeR [Robinson
et al., 2010] and baySeq [Hardcastle and Kelly, 2010]). These tools can also be
used to study differential expression at the transcript level, but a common alterna-
tive approach consists of relying on the algorithms implemented in the framework
of transcript abundance estimation, since those are able to take into account the
uncertainty in the read assignment process. For example, this is the case with
Cuffdiff2 [Trapnell et al., 2013] and MMDIFF [Turro et al., 2014], which can be
executed following estimation of transcript expression levels with Cufflinks [Trap-
nell et al., 2010] and MMSEQ [Turro et al., 2011], respectively. Moreover, and
consistent with the Bayesian model adopted, MMDIFF is also able to make use of
the uncertainty in the expression estimates, thus adding further sensitivity to the
differential transcript expression analysis.
1.3.6 Differential splicing
In the previous section, I have discussed briefly several approaches for the
assessment of differential transcript expression. However, differences in abso-
lute transcript abundance are not necessarily indicative of differential splicing
(Figure 1.12a), and alternative analysis strategies are preferred when the focus
lies on the latter. In general terms, changes in splicing patterns can be assessed
through the identification of either differential exon usage (DEU) or differential
transcript usage (DTU) events (Figure 1.12b and c, respectively), with advantages
inherent to both approaches. On the one hand, exon-centric analysis strategies are
completely independent from isoform reconstruction efforts, thus avoiding the
uncertainty intrinsic to that task. Furthermore, while those rely on the existing an-
notation, such dependence is limited to the exonic coordinates, and this approach
still enables the indirect identification of novel transcripts (i.e. novel alternatively
spliced isoforms). On the other hand, the results from such exon-centric analysis
are often difficult to interpret, and in this context transcript-centric analysis
strategies emerge as an attractive alternative.
Interestingly, in terms of algorithm development, much of the effort has been fo-
cused on the identification of differential transcript expression events, with limited
availability of tools for the study of changes in splicing. Amongst those, the Bio-
conductor package DEXSeq [Anders et al., 2012], was the first tool to account for
33
Chapter 1. Introduction
diffferential transcript 
expression (DTE)
(a) diffferential transcript
usage (DTU)
(c)
tr
an
sc
rip
t
ex
pr
es
sio
n 
le
ve
ls
re
la
tiv
e 
tr
an
sc
rip
t
ex
pr
es
sio
n
condition A condition B
transcript A
transcript B
tr
an
sc
rip
t
ex
pr
es
sio
n 
le
ve
ls
re
la
tiv
e 
tr
an
sc
rip
t
ex
pr
es
sio
n
condition A condition B
transcript A
transcript B
diffferential exon
usage (DEU)
(b)
ex
on
 c
ov
er
ag
e
condition B
condition A
Figure 1.12| Strategies for the study of changes in the abundance of alternative
transcripts.
(a) Differential transcript expression concept. Differential transcript expression is
analogous to differential gene expression, and does not necessarily imply differ-
ences in splicing.
(b) Differential exon usage concept. Differences in the read coverage for a given
exon, relative to changes in the number of reads that overlap the other exons
within the same gene, can be used as an indicator of differential splicing.
(c) Differential transcript usage concept. Differential transcript usage refers to those
cases where there is a change in the transcript relative abundances, which is not
necessarily linked to an overall change in expression levels. It constitutes the
most direct strategy for the study of differential splicing.
biological variation in the analysis, a vital requirement for robust testing. Briefly,
by relying on the same algorithmic principles as the aforementioned DESeq2, this
method enables the identification of significant differences in the proportion of
reads that overlap each exon, relative to the total number of reads that overlap the
corresponding gene (DEU events; Figure 1.12b). On the other hand, the recently
introduced tool MMDIFF [Turro et al., 2014] provides a method for the analysis
34
Chapter 1. Introduction
of DTU events (Figure 1.12c). MMDIFF is based on the use of Bayesian mixed
models, whereby the uncertainty in transcript expression estimates can be incor-
porated into the regression models used for testing, thus improving the power
to detect the events of interest. Altogether, the scarcity of tools to deconvolute
differences in splicing from differences in expression, together with the lack of
methods to infer the functional impact of the identified events, evidence that the
computational pipelines for the analysis of RNA-seq data are still not completely
established.
1.4 Aims of the thesis
The work presented in this thesis focuses on the use of RNA sequencing for the
high throughput study of alternative transcript products in human samples. Over-
all, the goal is to improve the current understanding of splicing by addressing the
following questions:
• What is the extent of transcriptome diversity? Are specific alternative tran-
scripts preferentially produced within a given gene?
• How prevalent are changes in splicing patterns in cancer? How can we
assess the potential functional impact of such changes?
• How do core spliceosomal factors participate in the regulation of splicing?
What are the effects of disrupting such regulation on dynamic cellular pro-
cesses such as cell division?
• Can the differential splicing events identified from transcriptomics data be
recapitulated at the protein level?
The present chapter has provided an general introduction to the two central con-
cepts behind this thesis, i.e. the splicing reaction and RNA sequencing. Further
introductory remarks relevant to each of the aforementioned questions will be
covered in the following chapters.
35

Chapter 2
The extent of transcriptome
diversity
The number of transcripts annotated in the human genome exceeds by far the
number of genes, and evidence collected from RNA-seq experiments during the
last few years indicates that the vast majority of those transcripts can be detected
as expressed across a range of conditions [Pan et al., 2008; Wang et al., 2008].
This chapter aims to investigate the extent to which transcriptome diversity is
maintained in a given sample when considering transcript relative abundances
within each gene. In the first part, I evaluate whether, in a given sample, gene
expression tends to be dominated by one transcript, as opposite to observing
similar expression levels for several of them. Next, I explore different analysis
strategies in order to increase the robustness of the obtained results. Finally, I
dissect further the biological impact of the reported findings by incorporating
information on transcript biotypes.
All the computational analyses described here have been performed by myself
under the supervision of Dr. Alvis Brazma. Dr. Adam Frankish and Dr. Jennifer
Harrow from the Wellcome Trust Sanger Institute contributed to the results
described in section 2.2.3, as detailed in the Methods, and also provided general
feedback on the project.
37
Chapter 2. The extent of transcriptome diversity
Publications derived from this chapter
• Gonzàlez-Porta, Frankish, Rung, Harrow & Brazma. Transcriptome analysis
of human tissues and cell lines reveals one dominant transcript per gene.
Genome Biology 14, R70 (2013).
The present chapter includes a relevant portion of text from this manuscript.
• (invited talk) Eukaryotic mRNA processing 2013, CSHL.
• (poster prize) Integrative RNA Biology 2013 (ISMB/ECCB satellite meeting),
Berlin.
• (poster) Genomics Medicine in the Mediterranean 2013, Crete.
• (poster) European Conference on Computational Biology 2012, Basel.
38
Chapter 2. The extent of transcriptome diversity
2.1 Introduction
Although there are less than 22,000 protein coding genes known in the human
genome, they are transcribed into over 140,000 different transcripts (Ensembl
release 66 [Flicek et al., 2012]), more than 65% of which have protein coding
potential and thus may contribute to protein diversity. Following the introduction
of RNA-seq, significant evidence has accumulated showing that over 95% of
multi-exon genes have several alternative splice-forms expressed [Pan et al., 2008;
Wang et al., 2008], and that transcript expression is regulated [House and Lynch,
2008; Smith et al., 2008]. On the other hand, focusing on expressed sequence
tag (EST) data, Taneri et al. [2011] predicted that there is a single dominant
transcript per gene in primary tissues. Recently, the ENCODE project showed that
indeed, most genes have a major transcript in cell lines, although at the same time
noted that "genes tend to express many transcripts simultaneously, and as the
number of annotated transcripts per gene grows, so does the number of expressed
transcripts" [Djebali et al., 2012]. Despite these observations, it is still unclear
if and to what extent major transcripts are dominating the transcriptome, and
what proportion of transcript diversity is likely to contribute to protein diversity.
In addition, given the notable differences in gene expression between primary
tissues and cell lines [Lukk et al., 2010; Waks et al., 2011], transcriptome analysis
in cell lines can be extended to primary tissues only to some extent.
The present chapter aims to characterise the potentially coding transcriptome
from a functional perspective. By focusing on protein coding genes, I show
that, in primary tissues, almost 85% of the total mRNA from protein coding
loci originates from major transcripts (76% in cell lines). Notably, these major
transcripts are not always the longest possible for the gene (40% of the major
transcripts in primary tissues and 30% in cell lines are not the longest annotated),
nor always include the longest coding DNA sequence (CDS; approximately 50%
of the cases in both tissues and cell lines). I also observe that the ratio of the
number of expressed transcripts to genes in primary tissues is on average 1.12,
which corresponds to just over one transcript per gene. I further distinguish
between: (1) major transcript - the transcript with the highest expression level for
a given gene; and (2) dominant transcript - a major transcript that is expressed
at a considerably higher level than any minor transcripts of the gene. I show
that most protein coding genes in most conditions have one dominant transcript;
39
Chapter 2. The extent of transcriptome diversity
e.g. for almost 80% of the expressed genes in primary tissues the major transcript
is at least twice as abundant as the next one. Furthermore, I detect a clear
separation in expression levels between major and minor transcripts, as well as
a considerably higher overlap for major vs. minor transcripts when comparing
them to a set of transcripts predicted to be translated into functional proteins by
entirely independent means [Rodriguez et al., 2013]. Finally, I observe that for
almost 20% of the studied protein coding genes (n = 18,450) the major transcript
does not code for a protein, and that this percentage is considerably higher in the
nucleus than in the cytosol.
These analyses have been performed using three different computational meth-
ods [Katz et al., 2010; Trapnell et al., 2010; Turro et al., 2011], and additionally,
where sufficient coverage exists, alternative transcript abundances were assessed
directly from the reads spanning unique exon junctions. Furthermore, simulated
data [Griebel et al., 2012] was used to confirm that the methods can reliably dis-
tinguish between two hypothetical alternative scenarios - one dominant transcript
per gene vs. several transcripts per gene expressed at similar levels. All those
methods produced a consistent outcome, supporting the robustness of the pre-
sented conclusions.
2.2 Results
In this chapter, I describe the analysis of two different RNA-seq datasets in order to
quantify and analyse the overall contribution of major transcripts to the potentially
coding mRNA transcriptome, in comparison to minor transcripts. These include
data on 16 primary human tissues from the Illumina Body Map dataset (BM),
further referred to as tissue dataset, as well as from 5 ENCODE cell lines, further
referred to as cell lines dataset (see Methods and Table A.1). Moreover, the latter
also include data from different cellular compartments (i.e. whole cell, cytosol
and nucleus), thus making it possible to compare across them. All the results
mentioned in this section refer to the average across all the samples of the tissue
dataset unless otherwise indicated.
40
Chapter 2. The extent of transcriptome diversity
2.2.1 Most protein coding genes express one dominant transcript
Similarly to the results reported in previous RNA-seq transcriptome studies,
it is possible to detect more than one transcript for approximately 85% of the
expressed genes (83.70% to 89.95%, SD = 1.84). In line with this observation, a
total of 105,456 different transcripts can be detected as expressed in at least one
tissue, which corresponds to approximately 90% of the studied transcripts (n =
117,759; see Methods). However, when quantifying all the annotated transcripts
within a gene based on their relative abundance, the expression of most genes
appears to be dominated by a single transcript in most conditions, rather than
by a subset of similarly expressed transcripts (Figure 2.1a and Figure B.1). This
observation is also supported by the fact that, in primary tissues, the ratio of the
number of expressed transcripts to genes is 1.12 (0.98-1.40, SD = 0.11). Finally,
analysis of the mRNA pool derived from protein coding loci revealed that major
transcripts comprise approximately 85% of the coding mRNAs present in the cell
(79.98% to 86.49%, SD = 2.17; Figure 2.1b).
In order to address the impact of these observations at the protein level, I
plotted the distribution of expression levels for both major and minor transcripts
(Figure 2.1c). This analysis led to the observation that minor transcripts have a
tendency to be expressed below 1 FPKM, a threshold that has been suggested
as the minimum expression required for protein detection [Hebenstreit et al.,
2011; Vogel and Marcotte, 2012]. With the same goal in mind, I then evaluated
the overlap between the major/minor transcript predictions obtained from the
interpretation of the RNA-seq data and those obtained by an entirely independent
method (APPRIS [Rodriguez et al., 2013]). Briefly, APPRIS aims at identifying
principal isoforms amongst all the transcript annotated within a gene, by combin-
ing information on protein domains, structures and conservation across species.
Such predictions were overlapped with two different sets of coding transcripts:
on the one hand, recurrent major transcripts expressed above 1 FPKM; on the
other hand, recurrent minor ones expressed below the same threshold. Focusing
on common genes in these two scenarios (n = 6,082), the detected overlap
was significantly higher for the first set compared to the second one (45.61%
vs. 29.35%, respectively; Fisher’s exact test p-value < 2.2·1016). This was also the
case when taking into consideration only those genes that are expressed in all
the tissues (n = 1,682, 59.93% vs. 22.06% overlap, respectively; Fisher’s exact test
41
Chapter 2. The extent of transcriptome diversity
p-value < 2.2·1016). Altogether, these results suggest that major transcripts could
be preferentially translated, even though minor ones may still play a functional
role.
Next, I quantified major transcript dominance by calculating for every gene
the ratio of the expression levels between the major transcript and the second
most abundant one. Overall, such analysis revealed that 79% of the genes in the
studied tissues (74.21% to 81.94%, SD = 2.16) have a two-fold dominant major
transcript (i.e. expressed twice as much as the second most abundant one), and
that for 56% of the genes (43.39% to 61.60%, SD = 3.50) the major transcript is
five-fold dominant (Figure 2.1d and Figure B.2). This indicates that most genes
tend to express one dominant transcript in a given sample. Similarly, dominant
transcripts were estimated to account for most of the studied mRNA pool: 76.69%
for a two-fold dominance (70.04% to 80.74%, SD = 3.48) and 67.47% for a five-fold
dominance (59.97% to 73.83%, SD = 4.81; Figure 2.1b). GO enrichment analysis
of genes that consistently express a five-fold dominant transcript across the 16
studied tissues indicated that they are functionally involved in metabolism and
cellular respiration, protein transport and transcription regulation (n = 1,450;
Table B.1). In addition, the comparison of the fraction of dominant vs. non-
dominant major transcripts across different FPKM thresholds evidenced that the
observed dominance is accentuated in highly expressed genes (Figure 2.1e). On
the other hand, focusing on genes that tend to express several transcripts at a
similar level, it was possible to identify 463 genes for which the major transcript
was less than two-fold dominant in all the tissues analysed (only 17 for a five-fold
dominance threshold). GO enrichment analysis of those revealed that they are
involved in RNA splicing/processing, post-transcriptional regulation of gene
expression and regulation of translation (Table B.2).
Similar analyses were also performed in the cell lines dataset, which includes
different cellular compartments (Figure B.1 and Figure B.2). In this context, I ob-
served that major transcripts account for approximately 80% of the studied mRNA
pool in the cytosol (77.20% to 83.66%, SD = 1.98; Table B.3), even though overall
transcript dominance is less accentuated than in primary tissues: 69% (63.11%
to 71.17%, SD = 2.40) of genes expressed a two-fold dominant major transcript
and 42% (35.16% to 44.76%, SD = 2.90) a five-fold dominant one in the cytosol
(Table B.4). Such differences could reflect higher transcription and splicing rates
42
Chapter 2. The extent of transcriptome diversity
0.
0
0.
2
0.
4
0.
6
0.
8
1.
0
transcript position
in the ranking
tr
an
sc
rip
t r
el
at
iv
e 
ab
un
da
nc
e
(a)
1 2 3 4 5 +5
transcript position
in the ranking
0
20
40
60
80
10
0
%
 m
RN
A 
po
ol
(b)
1 2 3 4 5 +5
2-fold dominant 
major transcripts
5-fold dominant 
major transcripts
10-2 1 100 10,000
transcript expression level 
(log10 scale)
0.
0
0.
2
0.
4
0.
6
0.
8
de
ns
ity
(c) major transcripts
minor transcripts
(e)
FPKM threshold
1 5 10
0
2
4
6
8
pr
op
or
tio
n 
of
 d
om
in
an
t  
vs
. 
no
n-
do
m
in
an
t m
aj
or
 tr
an
sc
rip
ts 2-fold dominance
5-fold dominance
(d)
1 FPKM 10,410 (56.42%) 8,179 (78.51%) 5,864 (56.22%)
expressed genes
genes with a dominant major transcript
2-fold dominance 5-fold dominance
3,077 (66.27%)3,898 (83.64%)4,671 (25.32%)5 FPKM
1,794 (72.60%)2,146 (86.54%)2,486 (13.47%)10 FPKM
Figure 2.1| Most protein coding genes express one dominant transcript. All
the results presented here correspond to the tissue data.
(a) Relative abundance of the subset of transcripts in each position of the ranking. For
each gene, transcripts were ranked based on their relative abundances. There is
generally one predominant transcript over the rest.
(b) Percentage of the studied mRNA pool explained by each category of transcripts. The
mean percentage for all samples is represented here. Major transcripts represent
approx. 85% of the studied mRNA population and were further classified into
2-fold and 5-fold dominant, depending on whether they are expressed twice or
five times as much as the second most abundant transcript for the corresponding
gene.
(c) Expression distribution for major and minor transcripts. A total of 31,902 tran-
scripts are expressed above 1 FPKM in at least one tissue, including 26,641 dif-
ferent major transcripts.
(d) Average number of genes with dominant major transcripts. Different dominance
ratios and gene expression thresholds were considered in the quantification.
(e) Proportion of dominant vs. non-dominant major transcripts. Values of the ratio
above 1 indicate a higher proportion of dominant major transcripts with respect
to non-dominant ones. Different dominance ratios and gene expression thresh-
olds were considered for this analysis.
43
Chapter 2. The extent of transcriptome diversity
in cell lines, although they could also be due to technical variability between the
two datasets. Given that these datasets have been generated by different laborato-
ries, it is very difficult to distinguish between these two scenarios, but this question
is revisited again in the next chapter, where the appropriate data is available.
2.2.2 The evaluation of different methods leads to a consistent outcome
Given that estimating transcript expression from short reads constitutes a
challenging task, I performed additional analyses to test the reliability of the
aforementioned observations. First, these included the simulation of different
RNA-seq datasets to test whether the pipeline used can distinguish between two
hypothesised scenarios: one dominant transcript per gene vs. similar expression
levels of the different transcripts in each gene (see Methods). Such analysis
led to the conclusion that the methods used here reliably discriminate between
the two scenarios, even when taking into account different sequencing depths
(Figure 2.2a). In addition, it evidenced that they are not biased towards the
identification of a single transcript per gene and corroborated the previously
described findings about transcript dominance (Figure 2.2b).
Figure 2.2| Evaluation of two hypothetical scenarios on major transcript abun-
dance.
(a) Relative abundance of the subset of transcripts in each position of the ranking for the
simulated datasets, including different sequencing depths. Several datasets were sim-
ulated under the assumption of one predominant transcript (default) or similar
expression levels for the transcripts annotated within each expressed gene (con-
stant). In comparison to the real expression values used for the simulation (real),
major transcript abundance is underestimated after the execution of the analysis
pipeline (20M, 40M and 80M, corresponding to the different sequencing depths
considered). This suggests that the analysis pipeline used is not biased towards
the identification of a single transcript per gene.
(b) Number of dominant major transcripts for the simulated datasets. The results ob-
tained when running the simulation under the hypothesis of a predominant tran-
script per gene (default) resemble those of the real datasets.
44
Chapter 2. The extent of transcriptome diversity
(a)
(b)
0.
0
0.
2
0.
4
0.
6
0.
8
1.
0
0.
0
0.
2
0.
4
0.
6
0.
8
1.
0
tr
an
sc
rip
t r
el
at
iv
e 
ab
un
da
nc
e
transcript position in the ranking
1 2 3 4 5 +5 1 2 3 4 5 +5 1 2 3 4 5 +5 1 2 3 4 5 +5
20M 6,102 4,603 (75.43%) 2,958 (48.48%)
expressed genes
genes with a dominant major transcript
2-fold dominance 5-fold dominance
2,943 (50.59%)4,471 (76.86%)5,81740M
3,113 (51.19%)4,722 (77.65%)6,08180MD
EF
AU
LT
CO
N
ST
AN
T 20M 13,533 4,309 (31.84%) 1,145 (8.46%)
1,249 (9.19%)4,288 (31.56%)13,58540M
1,282 (9.41%)4,313 (31.66%)13,62280M
real 20M 40M 80M
D
EFAU
LT
one predom
inant  transcript
CO
N
STAN
T
sim
ilar relative abundance
Second, the evaluation of alternative methods to estimate transcript expression
levels revealed a strong agreement in the major transcript predictions. For
example, for highly expressed genes, up to 90% overlap was observed when using
Cufflinks (87.78% to 92.59%, SD=1.62; Table 2.1). Furthermore, a considerable
overlap with the predictions from MISO was also detected when using direct
evidence from junction reads (i.e. 66.35% - 61.44% to 69.60%, SD=2.66; see
Methods).
45
Chapter 2. The extent of transcriptome diversity
MISO vs Cufflinks MISO vs MMSEQ
mean correlation SD mean correlation SD
1 FPKM 0.89 4.89e-03 0.70 2.36e-02
5 FPKM 0.93 5.33e-03 0.78 1.96e-02
10 FPKM 0.94 4.34e-03 0.81 1.91e-02
MISO vs. Cufflinks MISO vs MMSEQ
mean overlap (%) SD mean overlap (%) SD
1 FPKM 84.29 1.14 54.66 1.54
5 FPKM 88.71 1.67 57.34 2.98
10 FPKM 90.43 1.62 56.50 4.12
Table 2.1| Consistency in the transcript abundance estimates across different
software. There is in general a good agreement between the different methods
evaluated, with the biggest differences detected when using MMSEQ, which re-
quires mapping to the transcriptome (see Discussion).
(top) Average Pearson correlation coefficient in the expression estimates.
(bottom) Overlap in the major transcript predictions.
Third, the length of major transcripts was observed to be widely distributed
(Figure 2.3). More specifically, in over 50% of the cases (50.98% to 55.46%,
SD = 1.53) the identified major transcript did not correspond to the longest
one annotated. The same trend was observed when taking into account CDS
length: in approximately 50% of the genes (44.42% to 48.23%, SD = 1.12), the
major transcript did not contain the longest CDS, thus not corresponding to
the canonical transcript as annotated in UniProt [UniProt Consortium, 2012].
An example of such cases is the AES gene, which presents a ubiquitous major
transcript that is shorter than the current reference (Figure 2.4).
46
Chapter 2. The extent of transcriptome diversity
0
1
2
3
4 BM
ENCODE cell
ENCODE cytosol
ENCODE nucleus
0.0 0.2 0.4 0.6 0.8 1.0
major transcript length
relative to the longest annotated
(a)
(b) expressed genes
BM 8654
ENCODE cell 8527
ENCODE cytosol 7877
ENCODE nucleus 9177
major transcripts which are not the 
longest one annotated %
5106 58.94
5806 68.12
5073 64.35
6096 66.48
expressed genes for which the major 
transcript has a CDS annotated
major transcripts for which the 
annotated CDS is not the longest %
BM 7999 3714 46.35
ENCODE cell 7347 3650 49.74
ENCODE cytosol 7299 3532 48.38
ENCODE nucleus 7049 3564 50.61
de
ns
ity
Figure 2.3| Length distribution for major transcripts. Genes with several tran-
scripts annotated and expressed above 1 FPKM were taken into account for the
analysis.
(a) Length distribution for major transcripts.
(b) Number of major transcripts that correspond to the longest one annotated or that con-
tain the longest CDS. Major transcript detection is not biased towards the longest
one annotated.
47
Chapter 2. The extent of transcriptome diversity
AES-201
AES-202
AES-201 264
length (aa)
197
1743
length (bp)
1676
ENST00000221561
transcript id
ENST00000327141AES-202
(a)
(b)
Figure 2.4| Example of non-canonical major transcript common to all the 16
tissues analysed.
(a) Read coverage for the gene AES (amino-terminal enhancer of split).
(b) Annotated transcripts structure and length. The transcript AES-202 is recurrently
identified as major in all the tissues of the BM dataset, and it does not correspond
to the longest CDS.
Finally, the impact of unnanotated transcripts in the above observations was ad-
dressed by performing de novo transcript quantification using Cufflinks (see Meth-
ods). As expected, such analysis led to the identification of a higher number of
transcripts per gene (6.38 in GENCODE v11 vs. 12.84 using Cufflinks), but it is still
possible to observe one dominant transcript for most of them (Figure 2.5).
48
Chapter 2. The extent of transcriptome diversity
GENCODE
18,450genes 18,635
117,759transcripts 239,350
6.38ratio 12.84
CUFFMERGE GENCODE
n=1 19.23 18.32
12.54 7.84n=2
9.78 6.74n=3
8.30 4.73n=4
7.06 4.76n=5
43.09 57.61n>5
% genes with
n transcripts CUFFMERGE
(a)
0.
0
0.
2
0.
4
0.
6
0.
8
1.
0
tr
an
sc
rip
t r
el
at
iv
e 
ab
un
da
nc
e
transcript position in the ranking
1 2 3 4 5 +5
(b)
1 FPKM 11,169 (59.94%) 7,399 (66.14%) 4,642 (41.47%)
expressed genes
genes with a dominant major transcript
2-fold dominance 5-fold dominance
2,972 (43.68%)4,669 (68.25%)6,830 (36.65%)5 FPKM
1,960 (44.61%)3,052 (69.33%)4,401 (23.62%)10 FPKM
(e)
0
20
40
60
80
%
 m
RN
A 
po
ol
(c)
transcript position in the ranking
1 2 3 4 5 +5
2-fold dominant 
major transcripts
5-fold dominant 
major transcripts
major transcripts
minor transcripts
0.
0
0.
2
0.
4
0.
6
de
ns
ity
(d)
10-2 1 100 10,000
transcript expression level 
(log10 scale)
10
0
Figure 2.5| Summary of Cufflinks de novo quantification results.
(a) Summary of the number of genes and transcripts considered in the study set (GEN-
CODE) and those obtained after de novo transcript identification in the BM dataset using
Cufflinks (Cuffmerge). As expected, the average number of transcripts per gene
increases after de novo transcript quantification.
(b) Relative abundance of the subset of transcripts in each position of the ranking. It
is possible to detect one predominant transcript over the rest, albeit with lower
relative abundance than reported in that Figure 2.1.
(c) Percentage of the studied mRNA pool explained by each category of transcripts.
Lowly abundant transcripts explain a larger fraction of the mRNA pool than
in Figure 2.1, due to the increase in the average number of transcripts per gene.
(d) Expression distribution for major and minor transcripts. It is still possible to ob-
serve a clear separation between the expression levels of major vs. minor tran-
scripts.
(e) Number of dominant major transcripts, taking into account all genes in the extended
annotation (n = 18,635). Most major transcripts are expressed in a dominant fash-
ion, and this becomes more evident for higher FPKM thresholds.
49
Chapter 2. The extent of transcriptome diversity
2.2.3 Major transcripts from coding genes do not always code for pro-
teins
Functional classification of major transcripts revealed that, for 17% of protein
coding genes expressed in primary tissues (15.26% to 20.64%, SD = 1.60), the
major transcript lacks an annotated CDS as indicated by GENCODE. Taking
into account expression levels, and focusing on cell lines data, major non-coding
transcripts were observed to be more abundant in the nucleus, where they
represent approximately 15% of the studied mRNA pool (12.99% to 16.66%, SD
= 1.10, Figure 2.6a). Genes with major non-coding transcripts are expressed at
higher levels in the nucleus, compared to those with major coding transcripts,
while this trend is inverted in the cytosol (Figure 2.6b). In addition, non-coding
major transcripts are less dominant than coding ones in both compartments
(Figure 2.6b). Finally, analysis of the annotation revealed that these major
non-coding transcripts correspond to retained introns and processed transcripts,
which lack an open reading frame (see Methods). The latter are more prevalent
in the cytosol, while the proportion of retained introns is higher in the nucleus
(Figure 2.6c).
In order to evaluate the hypothesis that incomplete splicing could explain the
higher proportion of major retained introns in the nucleus, I compared intron
expression levels across cellular compartments (see Methods for details on the
calculation of intron expression). As expected, intron expression was detected to
be higher in the nucleus compared to the cytosol (Figure 2.7a). In addition, such
analysis revealed a general trend in the location of major retained introns towards
the transcriptional 3’-end (Figure 2.7b), which has been previously linked to the
nonsense-mediated decay pathway (see Discussion). Interestingly, this trend is
more accentuated in the cytosol than in the nucleus, where it could be masked
by the higher intronic expression levels. Alternatively, the prevalence of retained
introns as a major transcript could point to a functional mechanism, since genes
with retained introns as the major transcript both in nucleus and cytosol were
detected to be expressed at lower levels in the latter (Figure 2.7c; see Discussion).
Those genes are associated with ribosomal components, consistent with previous
findings indicating that introns regulate the expression of ribosomal proteins in
yeast (Table B.5, see Discussion).
50
Chapter 2. The extent of transcriptome diversity
BM PE
ENCODE
cell
ENCODE
cytosol
ENCODE
nucleus
0 20 40 60 80 100
major non-coding
minor non-coding minor coding
major coding
% mRNA pool
17.11
23.10
17.06
31.34
%
 genes w
ith a m
ajor 
non-coding transcript
retained intron
processed transcript
(a)
(b)
(c)
BM PE
ENCODE
cell
ENCODE
cytosol
ENCODE
nucleus
0 0 0.2 0.4 0.6 0.85 10 15 20 25 30 35
gene expression (FPKM) major transcript dominance
1
Figure 2.6| Major non-coding transcripts in protein coding genes.
(a) Proportion of the mRNA studied represented by different categories of transcripts. Av-
erage proportions were calculated including all the samples from each dataset.
Major non-coding transcripts are more abundant in the nucleus compared to the
cytosol.
(b) Expression patterns across cellular compartments for major non-coding transcripts.
Protein coding genes for which the most abundant transcript is non-coding are
expressed at higher levels in the nucleus, whilst this trend becomes inverted in
the cytosol (left). Major transcript dominance becomes reduced both when the
major transcript is non-coding and in the nucleus (right).
(c) Transcript biotype categories for the major non-coding transcripts. Average pro-
portions were calculated including all the samples from each dataset. Processed
transcripts are more abundant in the cytosol, while retained introns represent the
major fraction in the nucleus. Other minor categories that represented less than
1% of the transcripts were also identified, but are not visible in the plots.
51
Chapter 2. The extent of transcriptome diversity
cellcytosolnucleus
0
0.2
0.4
0.6
0.8
1
no
rm
al
ise
d 
in
tr
on
 e
xp
re
ss
io
n
(a) (b)
M
A
JO
R RI
M
IN
O
R RI
protein coding transcript
retained intron
0 20 40 60 80 100
% transcript
nucleus cellcytosol
(c)
gene expression (FPKMs)
cytosol
nucleus
0 10 403020
425
nucleus
cytosol
81
42
Figure 2.7| Focus on retained introns.
(a) Normalised intron expression in different cellular compartments. FPKMs were cal-
culated for all the introns and normalised by gene expression levels (see Meth-
ods). Intron expression is higher in the nucleus than in the cytosol (Wilcoxon test
p-value < 2.2·1016).
(b) Location of the dominant retained introns within the context of protein coding tran-
scripts. Genes for which the major transcript is a retained intron (RI) were initially
considered in the analysis, and cases where the second most abundant transcript
is protein coding and overlaps with the RIs were further selected. Similar criteria
were applied to analyse minor RIs. The location of the RIs is obtained by measur-
ing the distance from their centre to the transcriptional start of the overlapping
coding transcript, as illustrated in the panel below the figure (red dots). Major
RIs are preferentially located towards the transcriptional end of protein coding
transcripts.
(c) Expression levels for genes with major retained introns. The number of genes for
which the most abundant transcript is a RI is represented in the left. Amongst
the genes with major RIs in both cellular compartments (n = 81), gene expression
is higher in the nucleus (Wilcoxon test p-value < 2.2·1016).
52
Chapter 2. The extent of transcriptome diversity
On the other hand, the term processed transcript constitutes an ambiguous cate-
gory. Manual inspection of a subset of processed transcripts that were consistently
identified across all samples as the major transcript indicated that they could po-
tentially be re-annotated to protein coding, nonsense-mediated decay or retained
intron (Table B.6). Together, these observations suggest that the true proportion
of non-coding major transcripts for protein coding genes may be lower than the
current annotation suggests, in line with recent evidence pointing to the existence
of peptides from non-coding RNAs [Hemberg et al., 2014].
2.3 Discussion
In this chapter, I combine RNA-seq data from different primary tissues, cell
lines and cellular compartments to characterise the human protein coding tran-
scriptome from a functional perspective. I show that, in a given condition, most
protein coding genes not only express one major transcript, as recently observed
by Djebali et al. [2012], but in most cases the major transcripts are dominating
the transcriptome. The described findings are consistent across the two datasets
studied here, and are supported by several quantification methods, including
direct evidence from junction reads. Interestingly, the biggest differences in
major transcript predictions were detected when using MMSEQ, and those even
exceeded the ones obtained from direct interpretation of the raw data. Similarly to
MISO, this software uses a Bayesian approach to infer transcript abundances, but
requires mapping the reads to the transcriptome instead (see Introduction - sec-
tion 1.3.3.2). Such differences expose the potential variability introduced during
the mapping step and emphasise the need for further workflow evaluation studies
similar to the ones undertaken through the RGASP project [Engstrom et al., 2013;
Steijger et al., 2013]. On the other hand, transcript quantification from short read
sequences is not a trivial task, and the current annotation is continuously updated
to include novel transcripts. De novo transcript quantification revealed consistent
results to those obtained when relying exclusively on the annotation, although
major transcript dominance was less accentuated due to the higher proportion
of genes with a large number of transcripts. Nonetheless, it is reassuring to
observe that the detected single transcript dominance becomes stronger for
highly expressed genes, for which transcript prediction and quantification have
been reported to be more reliable [Steijger et al., 2013]. Furthermore, similar
observations have been made at the single-cell level following publication of our
53
Chapter 2. The extent of transcriptome diversity
manuscript [Marinov et al., 2014; Shalek et al., 2013; Yan et al., 2013], and in the
long term, longer reads will shed more light on the topic.
The observation that a non-negligible fraction of protein coding genes express a
major non-coding transcript came as a surprising finding. However, non-coding
major transcripts are more prevalent in the nucleus, specifically in the case of
retained introns, and it was also possible to detect higher intronic expression lev-
els in this compartment. Altogether, these observations could reflect incomplete
splicing, consistent with existing evidence indicating that unspliced or incom-
pletely spliced mRNAs are not exported to the cytosol [Porrua and Libri, 2013].
In addition, the described results show that retained introns are preferentially
located towards the transcriptional end of transcripts. This observation has been
previously linked to the nonsense-mediated decay pathway [Kurmangaliyev
and Gelfand, 2008], a surveillance mechanism that ensures the degradation of
unspliced transcripts when they are transported to the cytosol (see Introduction
- The cytosolic life of mRNAs) [Pérez-Ortín et al., 2013]. Nevertheless, several
cases of functionally relevant retained introns have been described, either as
a mechanism to target mRNA molecules (e.g. [Buckley et al., 2011]), produce
alternative protein products (e.g. [Li et al., 2006]) or to regulate expression levels
(e.g. [Wong et al., 2013a; Yap et al., 2012]). In this context, I observe that genes with
retained introns as the major transcript in both nucleus and cytosol are expressed
at considerably lower levels in the latter, which could point to a regulatory role.
Finally, I also detect that those genes are associated to ribosomal components,
which is consistent with previous findings indicating that introns regulate the
expression of ribosomal proteins in yeast [Parenteau et al., 2011].
Overall, it is difficult to predict the impact of the described observations at
the protein level. There have been several studies addressing the relationship
between protein and transcript levels, which in general point at a modest, but
not insignificant correlation (i.e. the best estimates point at a range of 56%-64%
correlation [Li et al., 2014; Lundberg et al., 2010; Nagaraj et al., 2011]). Transla-
tional efficiency, mRNA and protein turnover rates are likely to have an impact
on protein levels [Vogel and Marcotte, 2012]. In addition, alternative splicing not
only has an impact on the protein repertoire, but also contributes to the control
of expression levels [McGlincy and Smith, 2008; Yap et al., 2012] and transcript
localisation [Keren et al., 2010; Nilsen and Graveley, 2010], which brings in other
54
Chapter 2. The extent of transcriptome diversity
potential roles for minor transcripts. On the other hand, proteomics studies
also show that detecting a protein is unlikely unless there are at least a certain
number of RNA molecules per cell [Ramakrishnan et al., 2009]. This may be
partly due to insufficient sensitivity of the methods used; nevertheless, the clear
separation in expression levels between major and minor transcripts, together
with the higher overlap of the former with an independent set of transcripts
predicted to be translated into functional products, suggest that the abundance
of proteins derived from minor transcripts is likely to be lower than the one from
dominant ones, and that minor transcripts could simply result from noisy splicing
[Melamud and Moult, 2009].
Further information on the functional impact of the observations described in
this chapter can be obtained from comparative analyses. Alternative splicing is
a widespread process among eukaryotes, where it has been regarded as a mech-
anism for the diversification of protein sequence, structure and function, hence
conferring flexibility in the regulation of functional products across different
tissues or developmental stages [Nilsen and Graveley, 2010]. Comparative studies
based on the analysis of exon usage patterns have provided evidence for the
existence of lineage-specific splicing patterns [Barbosa-Morais et al., 2012; Merkin
et al., 2012], and whether similar results can be observed from the evaluation of
trends in major transcript expression across species remains an open question.
This analysis would not only complement past efforts for the identification of
evolutionary conserved splicing programs, but would also help in assessing the
functional relevance of major transcripts, providing extra evidence for function for
those with a high degree of conservation. Similarly, it would also be interesting
to evaluate major transcript dominance in the context of gene properties that
have been shown to correlate with alternative splicing. For example, the number
of transcripts annotated per gene has been shown to correlate positively with
gene age, length and the number of cassette exons [Roux and Robinson-Rechavi,
2011], but the relationship of these features with major transcript expression
patterns has so far been unexplored. Altogether, the research questions proposed
here would contribute in improving the current mechanistical understanding of
splicing.
Knowledge on major transcripts can be used to build a catalogue for the reference
transcriptome, by focusing on dominant transcripts that are recurrent in many
55
Chapter 2. The extent of transcriptome diversity
samples. In this context, a closer inspection of such set of transcripts revealed
cases where they do not contain the longest CDS, a criteria often used in resources
like UniProt to define a reference transcript [UniProt Consortium, 2012]. These
results reinforce previous observations on the limitations of the current definitions
[Tress et al., 2008a], and point to potential advantages of taking into account
this type of data. Similarly, such information can also help in improving the
existing annotation, as exemplified here with the re-annotation of a subset of
major processed transcripts that were consistently identified across all samples.
Moreover, identification of changes in the major transcript across conditions can
lead to relevant findings, as shown in the next chapter. Finally, such observations
may also help proteome analysis by prioritising the candidate proteins that are
more likely to be present in a given sample, as further explored in Chapter 5.
In conclusion, the discovery of alternative splicing and many different classes of
non-coding RNAs, together with the establishment of RNA-seq, revealed that the
number of transcripts exceeds many times the number of genes in the human
genome. This has been used to argue that alternative splicing possibly explains
the low number of genes compared to what was believed before it was sequenced
[Nilsen and Graveley, 2010]. Despite such transcriptome diversity, the results de-
scribed here show that most protein coding genes express one dominant transcript
in a given condition, which, when combined, comprise most of the potentially
coding mRNA transcriptome. Hence, it is tempting to hypothesize that although
some minor transcripts may play a functional role, the major ones are likely to be
the main contributors to the proteome.
2.4 Computational methods
All the computational analyses described in this chapter have been performed by myself,
except for the re-annotation of major processed transcripts, which has been carried out by
Dr. Adam Frankish, as detailed below.
Datasets and mapping
Analyses were based on the Illumina Body Map (BM) dataset and a subset of
ENCODE cell lines [ENCODE Project Consortium et al., 2012] (ArrayExpress
accession ids: E-MTAB-513 and E-GEOD-26284, respectively), jointly covering
56
Chapter 2. The extent of transcriptome diversity
a total of 21 different tissues and cell lines, as well as different cellular com-
partments (see Table A.1). Raw FASTQ files were retrieved from the European
Nucleotide Archive1. In addition to the publicly available datasets, two RNA-seq
experiments were generated using the Flux Simulator [Griebel et al., 2012]. On
the one hand, the scenario of one dominant transcript per gene was simulated by
running this tool with the default parameters. On the other hand, a custom .pro
file with AFREQ_EXP=1000 was used to resemble similar expression of the several
transcripts annotated within each given gene.
FASTQ files in the BM dataset were filtered before mapping by trimming the last
five nucleotides of all reads. Raw data were mapped to the human genome and
transcriptome (Ensembl 66 [Flicek et al., 2012]) using Bowtie v0.12.7 [Langmead
et al., 2009] and TopHat v1.3.3 [Trapnell et al., 2009], respectively.
Gene and transcript study sets
Gene and transcript annotations used in the analyses correspond to those in GEN-
CODE v11 [Harrow et al., 2012]. This annotation is the result of the combination
of both computational and experimental evidence, as well as manual curation
efforts, and has been shown to offer a good compromise in terms of complexity
when compared to existing alternatives [Wu et al., 2013].
All the analyses discussed in this chapter are based on protein coding genes. Those
genes for which at least one of the annotated transcripts was shorter than 300 bp
were removed from the analyses (n = 1,638), since they would be lost during
the size selection step in the RNA-seq experiment (see Introduction - A typical
sequencing workflow). In total, the study set comprises 18,450 protein coding
genes, of which 14,902 have more than one transcript annotated.
Biotype definition for transcripts derived from protein coding genes
The transcript biotypes used in this chapter were obtained from the GENCODE
annotation, and are defined as follows:
• Protein coding: if the transcript contains a CDS.
• Nonsense-mediated decay (NMD): if the transcript contains a CDS but has one
or more splice junctions >50bp downstream of the stop codon.
1http://www.ebi.ac.uk/ena/
57
Chapter 2. The extent of transcriptome diversity
• Retained intron: if the transcript has an intronic sequence compared to a
reference variant and has no strong evidence for function.
• Processed transcript: if the transcript does not contain a CDS and does not
fulfil the previous criteria.
Counting reads overlapping exonic and intronic regions
Exonic coordinates were retrieved from the annotation and used to define intronic
regions. Formally, the definition of intron encompasses those regions that are
located inside genic boundaries and are not overlapped by any exon in any anno-
tated transcript. The number of reads overlapping known exons and introns was
computed using dexseq-count (DEXSeq v1.5.5 [Anders et al., 2012]) and converted
to FPKM values with custom scripts.
Estimating gene and transcript expression levels
For each gene, expression levels were calculated as the average FPKMs of all ex-
pressed exons. Transcript abundances were obtained using three different tools:
MISO v0.4.1 [Katz et al., 2010], Cufflinks v1.3.0 [Trapnell et al., 2010] and MMSEQ
v0.10.0 [Turro et al., 2011]. MISO and Cufflinks take as input alignments to the
genome, while MMSEQ requires mapping to the transcriptome, thus the need to
use two different mapping strategies (see Methods - Datasets and mapping). In all
three cases, the expression estimates were based on the existing transcript annota-
tion, cancelling any option for de novo inference, and were converted to transcript
relative abundances when necessary. In this chapter, I refer to the results obtained
by MISO and I use a default FPKM threshold of 1 to consider a gene/transcript
as expressed, a threshold that has been suggested as the minimum expression re-
quired for protein detection [Hebenstreit et al., 2011; Vogel and Marcotte, 2012].
In addition, higher expression thresholds are also included in the analyses (5 and
10 FPKM), since transcript quantification has been reported to be more reliable for
those [Steijger et al., 2013]. Finally, a transcript is considered as detected indepen-
dently of its expression level, provided that the corresponding gene is expressed.
mRNA pool estimates
mRNA pool estimates were calculated as introduced by Ramskold et al. [2009].
Briefly, the fraction of the studied mRNA pool that can be explained by the ex-
pression of major transcripts can be represented as the ratio of the sum of FPKMs
58
Chapter 2. The extent of transcriptome diversity
for major transcripts vs. the sum of FPKMs for all the transcripts in the study set.
All transcripts encoded within protein coding genes were taken into account in the
calculation, independently of their transcript biotype. Mitochondrial genes in the
study set were discarded (n = 11 in the study set), since they are present multiple
times in the cell and could bias the quantification.
Using direct evidence from junction reads
Genes that are part of the study set (see Methods - Gene and transcript study sets)
and for which all of the annotated transcripts can be uniquely identified by at least
one splice junction (n = 2,306) were considered in this analysis. Major transcripts
were then identified based on coverage evidence, by quantifying the number of
reads supporting each junction and taking the average in case of several splice
junctions. For each sample, the overlap with MISO was calculated.
De novo transcript discovery using Cufflinks
Cufflinks v1.3.0 [Trapnell et al., 2010] was used to discover novel transcripts in
each tissue from the BM dataset and all the obtained annotations were merged us-
ing cuffmerge [Trapnell et al., 2010]. Next, the subset of transcripts that overlaps
with known protein coding genes was used for analysis, and those genes with
transcripts shorter than 300 bp were filtered out (see Methods - Gene and tran-
script study sets). A summary of the number of genes and transcripts identified
can be found in Figure 2.5a.
Gene Ontology enrichment analyses
Gene Ontology (GO) enrichment analyses were performed with the DAVID and
WebGestalt [Huang et al., 2009a,b; Wang et al., 2013]. The reference population
was defined by our gene study set (see Methods - Gene and transcript study sets),
and an adjusted p-value of 0.05 was used as a threshold for the identification of sig-
nificant GO terms (Benjamini and Hochberg correction [Benjamini and Hochberg,
1995]).
Re-annotating major processed transcripts
Analysis performed by Dr. Adam Frankish.
Processed transcripts that were recurrently identified as the major isoform in a
consistent manner across all the studied samples were manually curated by using
59
Chapter 2. The extent of transcriptome diversity
information from the Zmap annotation interface1 and re-annotated as indicated in
Table B.4.
1http://www.sanger.ac.uk/resources/software/zmap
60
Chapter 3
The prevalence of splicing
changes in cancer
Splicing plays a central role in the expression of most genes, and its dysregulation
is known to contribute to the pathogenesis of several diseases, including cancer
[Padgett, 2012; Pedrotti and Cooper, 2014]. In this context, alterations in the
splicing reaction have been associated with oncogenesis, tumour suppression and
metastasis [Hagen and Ladomery, 2012; Kaida et al., 2012]. The goal of the present
chapter is to investigate the prevalence of changes in splicing patterns in the most
common type of kidney cancer, i.e. clear cell Renal Cell Carcinoma (ccRCC). The
analyses described here have been performed as part of the CAGEKID project
(CAncer GEnomics of the KIDney), a broad effort aimed at improving the current
understanding of this type of cancer through the combination of data from
genomics, transcriptomics and epigenomics experiments (DNA-seq, RNA-seq
and methylation arrays, respectively). In the first part of the chapter, I perform
a global analysis of the splicing changes that underlie ccRCC. Those are further
dissected in the second part, with special emphasis on the most extreme and
recurrent events. I next pursue integration of the described results with those
obtained from other types of data generated within the same Consortium. Finally,
I expand these analyses to a separate RNA-seq dataset in order to investigate
the resemblance of splicing patterns in primary tissues and cell lines, linking the
derived findings to the observations from Chapter 2.
The results from all the computational analyses reported here have been pro-
duced by myself under the supervision of Dr. Alvis Brazma. Further analyses
61
Chapter 3. The prevalence of splicing changes in cancer
performed by Dr. Johan Rung from EMBL-EBI and Dr. Louis Letourneau from
McGill University have also been included in sections 3.2.1 and 3.2.4, and their
contribution has been emphasised in the Methods.
Publications derived from this chapter
• Gonzàlez-Porta & Brazma. Identification, annotation and visualisation of
extreme changes in splicing from RNA-seq experiments with SwitchSeq.
bioRxiv. 10.1101/005967.
• (submitted) Scelo?, Riazalhosseini?, Greger?, Letourneau?, Gonzàlez-Porta?
et al. Whole-genome sequencing reveals variation in the genomic landscape
of clear cell Renal Cell Carcinoma in Europe.
• (poster) Systems biology: Global regulation of gene expression 2014, CSHL.
?shared first authors
62
Chapter 3. The prevalence of splicing changes in cancer
3.1 Introduction
Numerous studies have evidenced the role of alternative splicing in disease
development, including cancer (e.g. see Padgett [2012]; Pedrotti and Cooper
[2014] for a review), where its dysregulation has been linked to oncogenesis,
tumour suppression and metastasis [Hagen and Ladomery, 2012; Kaida et al.,
2012]. In general, changes in splicing patterns are known to contribute to
disease either by directly affecting the function of proteins (e.g. protein-protein
interactions, sub-cellular localization or catalytic ability [Keren et al., 2010; Nilsen
and Graveley, 2010]), or by subtly regulating gene expression levels [McGlincy
and Smith, 2008; Yap et al., 2012]. Moreover, several mechanisms that could lead
to the disruption of splicing have been identified, including mutations in core
elements of the splicing machinery, alterations on the concentration and function
of Splicing Factors and mutations in regulatory regions associated with splice site
selection [Tazi et al., 2009].
In this chapter, I aim at investigating the changes in splicing patterns associated
with clear cell Renal Cell Carcinoma (ccRCC), the most common and aggressive
type of kidney cancer [Jonasch et al., 2012]. ccRCC tumours develop in the renal
parenchyma and represent 70-80% of renal cancers, which are in general one of
the most commonly diagnosed cancers in adults (i.e. they represent the 7th most
common cancer in Europe and account for 3% of adult malignancies worldwide)
[Chow et al., 2010]. Key clinical needs for ccRCC include the identification of
new biomarkers and new therapeutic targets, as suggested by the fact that this
is one of the few tumour types for which there are currently no biomarkers in
routine clinical use [Jonasch et al., 2012]. Recently, several large scale sequencing
projects, including initiatives from The Cancer Genome Atlas (TCGA) and the
International Cancer Genome Consortium (ICGC), have shed light on the genomic
alterations that underlie this type of cancer [Sato et al., 2013; Scelo et al.; TCGA
Network, 2013]. The main findings derived from this large scale projects include
the consolidation of previous knowledge on the role of the tumour suppressor
gene VHL in ccRCC development (with mutations detected in approximately 90%
of spontaneous tumours), as well as on the relationship between the allelic loss
of chromosome 3p and disease pathogenesis [Brugarolas, 2013]. Interestingly, the
VHL gene is located at this locus, as it is the case for other driver genes that have
been identified amongst the most commonly mutated ones (i.e. PBRM1, SETD2
63
Chapter 3. The prevalence of splicing changes in cancer
and BAP1; all of them tumour suppressors) [Brugarolas, 2013]. In addition, these
studies have reinforced the idea that ccRCC is a metabolic disease, with many
genomic aberrations affecting several metabolic pathways [Linehan et al., 2010].
However, with much of the focus being placed on the genomics scale, a thorough
characterisation of the splicing alterations that affect this type of cancer is still
missing, in spite of data availability (i.e. RNA-seq data for hundreds of patients
was generated in all the three aforementioned studies). Valletti et al. [2013] and
Zhao et al. [2013] relied on the identification of differentially expressed exons to
assess exon skipping events, by using microarrays and the previously mentioned
TCGA RNA-seq data, respectively, but this is a limited approach that focuses only
on the study of a subset of splicing events.
Through the analysis of RNA-seq data for 45 matched ccRCC samples (i.e. tumour
and healthy peripheral tissue from the same patient), I show here that splicing
patterns are largely altered in ccRCC, with almost 40% of the expressed genes
being affected (n = 7,842 out of 19,944 genes). Consistent with the results reported
in Chapter 2, I identify gene expression to be dominated by one transcript in
most cases, both in tumour and healthy samples, and I detect a bigger overlap
in major transcripts for the latter condition. Following these observations, I
estimate that only ⇠8% of the genes with altered splicing manifest significant
changes in major transcript dominance across conditions (n = 602 genes). Thus, I
conclude that despite the large number of genes that undergo differential splicing
events, most of those do not lead to big changes in the abundance of major
transcripts, at least in a recurrent fashion. Further study of switch events in each
sample pair (i.e. changes in the identity of the major transcript for a given gene
across conditions) confirmed the idea that most of the big changes in splicing
are patient-specific. This behaviour becomes even more extreme for those events
that involve dominant transcripts (i.e. major transcripts that are expressed at a
considerably higher level in comparison to the second most abundant one from
the same gene). Nonetheless, following integration with genomic data for a
set of 36 matched samples, I observe that pathways are widely disrupted when
combining information across multiple regulatory layers. For example, from such
integrative analysis, I do not detect major transcriptomic alterations for VHL, but
I do observe switch events for PBRM1 and BAP1 in several patients. Interestingly,
for PBRM1, these do not overlap with those cases in which the gene is mutated.
Finally, I explore overall differences in the splicing patterns of primary tissues
64
Chapter 3. The prevalence of splicing changes in cancer
vs. cell lines by analysing RNA-seq data from a set of 6 ccRCC cell lines.
The analyses described in this chapter have been performed as part of the
CAGEKID project (Cancer Genomics of the Kidney), one of the aforementioned
large scale sequencing efforts that aimed at characterising this disease from a ge-
nomic, transcriptomic and epigenetic perspective. This project is part of ICGC and
comprises 14 partners from 6 EU countries and Russia, including our team at the
EBI, which has led the interpretation of the transcriptomic data.
3.2 Results
In spite of the diversity of data generated within the context of CAGEKID, my
analyses rely primarily on the available matched RNA-seq samples. This includes
a set of 45 patients diagnosed with ccRCC, from whom both tumour and healthy
kidney tissue were extracted and sequenced with good quality. In all the cases,
revision by multiple pathologists ensured a 70% content of tumour cells in the
ccRCC samples. RNA-seq was performed on polyA-selected extracts using the
Illumina HiSeq 2000 platform (100 bp paired-end reads; Table A.2). In addition,
DNA-seq data for 36 of the 45 matched samples was also produced with the same
platform (100 bp paired-end reads), thus enabling the integration of genomic and
transcriptomic evidence. Finally, I also had access to RNA-seq data from a separate
set of 6 ccRCC cell lines (Table A.2). Even though those samples were not formally
generated as part of CAGEKID, they were sequenced by one of the participating
laboratories following the same protocols, thus providing a good scenario for the
comparison of splicing patterns in primary tissues vs. cell lines.
3.2.1 Splicing is largely altered in ccRCC
First, genes with significant changes in splicing in ccRCC vs. healthy tissues were
identified by running DEXSeq on the 45 matched RNA-seq samples (see Methods).
Out of the 19,944 genes detected as expressed, a total of 7,842 genes were predicted
to manifest significant differences in exon usage (padj < 0.01; Figure 3.1a), the vast
majority of which are protein coding genes (n = 7,182). GO enrichment analysis
of this set of genes revealed that they are involved in processes such as cellular
respiration, the formation of cell junctions and the regulation of cell cycle and
apoptosis, amongst other functions (Figure 3.1b).
65
Chapter 3. The prevalence of splicing changes in cancer
(a) % expressed genes
0 20 40 60 80 100
differentially
spliced
not changeddifferentially
expressed
both
(c)
splicing variability
normal
tumour
0.0 0.1 0.2 0.3 0.4
(b) GO term adjusted p-value
protein biosynthesis 8.83E-012
DNA repair 8.81E-005
cytoskeleton organization 8.35E-006
cell death 5.46E-005
cell cycle 4.35E-005
intracellular protein transport 2.75E-014
cellular respiration 2.52E-008
generation of precursor
metabolites and energy
1.61E-008
actin binding 1.57E-006
kinase activity 1.33E-012
tight junction 6.3E-003
oxidative phosphorylation 5.7E-003
mitochondrial respiratory chain 5.4E-003
regulation of apoptosis 2.1E-003
9.6E-003RNA splicing,
via transesterification reactions
3.5E-003posttranscriptional regulation
of gene expression
1.6E-003adherens junction
pathway adjusted p-value
metabolic pathways 2.72E-191
pathways in cancer 1.82E-036
oxidative phosphorylation 1.10E-028
focal adhesion 4.37E-027
tight junction 1.90E-025
cell cycle 2.79E-018
MAPK signalling pathway 8.05E-018
6.00E-017
PPAR signalling pathway 1.87E-016
2.27E-016
RNA transport 4.20E-016
8.82E-015
renal cell carcinoma 1.34E-011
p53 signalling pathway 1.31E-010
cell adhesion molecules (CAMs) 6.21E-009
mRNA surveillance pathway 6.91E-009
gap junction 7.12E-009
RNA degradation 1.16E-008
mTOR signalling pathway 6.41E-008
drug metabolism - cytochrome P450 1.13E-007
apoptosis 4.35E-006
drug metabolism - other enzymes 2.51E-005
VEGF signalling pathway 4.67E-005
TGF-beta signalling pathway 4.95E-005
spliceosome
adherens junction
ErbB signalling pathway
Figure 3.1| Splicing patterns in ccRCC tumours vs. healthy matched samples.
(a) Percentage of genes that are differentially spliced, differentially expressed or both
amongst the set of expressed genes. A large proportion of the expressed genes
present differences in either splicing, expression or both.
(b) GO and pathway enrichment analysis for the set of differentially spliced genes. Genes
with significant differences in splicing have a role in cellular respiration, energy
metabolism, the formation of cell junctions and cell cycle progression. Represen-
tative terms for 77 enriched pathways are shown here.
(c) Splicing variability distribution for healthy and tumour samples. Transcript relative
abundances within each given gene are more variable across tumours than across
normal samples (p-value < 2.2 · 1016).
66
Chapter 3. The prevalence of splicing changes in cancer
Similar results were obtained following pathway enrichment analysis, which
expectedly pointed at an alteration of pathways involved in cancer and renal cell
carcinoma (Figure 3.1b). Overall, the above results suggest that, in almost 40%
of the expressed genes, alterations in splicing are broad enough to be detected
in a general comparison of normal vs. tumour samples. Such a large fraction of
altered genes is consistent with the high number of differentially expressed genes
detected in the same sample set (n = 12,231; padj < 0.01; see Methods). Notably,
there is a significant overlap between these two sets (p-value < 2.2·1016; common
genes = 5,090; Figure 3.1a).
On a separate note, transcript expression levels were detected to be more variable
within tumours compared to normal samples (p-value < 2.2 · 1016, Figure 3.1c;
see Methods). Given the positive correlation between the number of annotated
transcripts and splicing variability (rS = 0.75; p-value < 2.2 · 1016), part of this
increase in variability could be explained by a higher number of expressed tran-
scripts per gene in tumours (Figure 3.2).
0 20 40 60 80 100
% genes
number of expressed transcripts
normal
tumour
1 2 3 4 5 +5
(a)
0 1 2 3 4
expressed transcripts in normal vs. tumor
paired 
samples
random
(b)
Figure 3.2| Number of expressed transcripts per gene in normal vs. tumour
samples.
(a) Percentage of genes with each given number of expressed transcripts. 1 FPKM was
used as the expression threshold, following the analysis workflow introduced in
Chapter 2.
(b) Ratio of the number of expressed transcripts in paired samples compared to a ran-
domised set. Tumours tend to express a higher number of transcripts per gene
(p-value < 2.2 · 1016).
67
Chapter 3. The prevalence of splicing changes in cancer
Finally, and consistent with the results described in Chapter 2, gene expression
was observed to be dominated by one transcript in most cases, both in tumours
and in healthy tissue samples (Figure 3.3). The integration of information across
biological replicates evidenced that genes commonly express one major transcript
(Figure 3.4a); nonetheless, it was also possible to identify several major transcripts
for a considerable fraction of genes. Notably, the number of different major tran-
scripts per gene was detected to be higher in tumours, in line with the previous
observations on the overall larger number of expressed transcripts in this condi-
tion. Furthermore, major transcripts were detected to be recurrently expressed in
both conditions (e.g. 90% of the genes express the same major transcript in >50%
of the samples where they are expressed; Figure 3.4b). Lastly, the overlap in major
transcripts was identified to be higher within healthy tissue samples than within
tumours (Figure 3.4c), consistent with the higher splicing variability previously
detected for the latter.
68
Chapter 3. The prevalence of splicing changes in cancer
0.
0
0.
2
0.
4
0.
6
0.
8
1.
0
tr
an
sc
rip
t r
el
at
iv
e 
ab
un
da
nc
e
transcript position in the ranking
1 2 3 4 5 +5
transcript position in the ranking
1 2 3 4 5 +5
0
20
40
60
80
10
0
%
 m
RN
A 
po
ol
0.
0
0.
2
0.
4
0.
6
0.
8
1.
0
tr
an
sc
rip
t r
el
at
iv
e 
ab
un
da
nc
e
transcript position in the ranking
1 2 3 4 5 +5
0.
0
0.
2
0.
4
0.
6
0.
8
de
ns
ity
10-2 1 100 10,000
transcript expression level 
(log10 scale)
0.
0
0.
2
0.
4
0.
6
0.
8
de
ns
ity
major transcripts
minor transcripts
1 FPKM 10,802 (58.55%) 7,115 (65.88%) 4,187 (38.79%)
expressed genes
genes with a dominant major transcript
2-fold dominance 5-fold dominance
2,878 (44.42%)4,502 (69.41%)6,491 (35.18%)5 FPKM
1,922 (48.89%)2,836 (72.07%)3,937 (21.34%)10 FPKM
1 FPKM 10,368 (56.19%) 6,977 (67.31%) 4,222 (40.74%)
expressed genes
genes with a dominant major transcript
2-fold dominance 5-fold dominance
2,627 (47.34%)3,934 (70.91%)5,549 (30.07%)5 FPKM
1,715 (52.34%)2,445 (73.28%)3,336 (18.08%)10 FPKM
transcript position in the ranking
1 2 3 4 5 +5
0
20
40
60
80
10
0
%
 m
RN
A 
po
ol
2-fold dominant 
major transcripts
5-fold dominant 
major transcripts
major transcripts
minor transcripts
10-2 1 100 10,000
transcript expression level 
(log10 scale)
2-fold dominant 
major transcripts
5-fold dominant 
major transcripts
TUMOUR
(e) (f )
(h)
(g)
NORMAL
(a) (b) (c)
(d)
Figure 3.3| Most protein coding genes express one dominant transcript in both
normal and tumour samples.
(a,e) Relative abundance of the subset of transcripts in each position of the ranking.
(b,f) Percentage of the studied mRNA pool explained by each category of transcripts.
(c,g) Expression distribution for major and minor transcripts.
(d,h) Average number of genes with dominant major transcripts.
69
Chapter 3. The prevalence of splicing changes in cancer
10
0
0 20 40 60 80 100
0
20
40
60
80
expression breadth for the most 
common major transcript (%)
m
aj
or
 tr
an
sc
rip
t o
ve
rla
p 
(%
)
%
 g
en
es
normal
tumour
random
(b) (c)
N T R
20
40
60
80
normal
tumour
random
(a)
0
10
20
30
40
50
%
 g
en
es
number of different
major transcripts
+102 4 6 80
Figure 3.4| Major transcript expression patterns in ccRCC tumours vs. healthy
matched samples.
(a) Number of major transcripts detected for each gene in each of the evaluated condi-
tions. Genes have been stratified with regard to the number of different major
transcripts that can be detected across biological replicates, and a random dis-
tribution was obtained by randomising the set of major transcripts expressed in
each sample. One single major transcript can be detected for most genes; how-
ever, it is also possible to observe more than one major transcript for a large
percentage of cases.
(b) Expression breadth for major transcripts. Expression breadth is measured by the
number of samples where a given transcript is detected as major, relative to the
number of samples where the corresponding gene is identified as expressed. For
each gene, the most common major transcript was considered for the analysis.
Major transcripts tend to be recurrently detected in both normal and tumour
samples, and there is a negative correlation between the number of annotated
transcripts and the observed expression breadth (rS: -0.44, p-value < 2.2·1016).
(c) Overlap in the set of major transcripts across healthy and tumour samples. There
is higher consistency in major transcript identity in normal samples (p-value <
2.2 · 1016). N: normal; T: tumour.
70
Chapter 3. The prevalence of splicing changes in cancer
3.2.2 Large and recurrent changes in splicing are rare
As a next step, and in order to address the magnitude of the above reported splic-
ing alterations, I evaluated the extent of changes in the dominance and identity of
major transcripts across conditions. Analysis of major transcript dominance from
a patient-centric perspective revealed that, amongst the set of expressed genes,
the number of genes with a dominant transcript differs significantly between the
healthy and tumour tissues in all the 45 sample pairs (padj < 2.2 · 1016 in all the
comparisons; see Methods). However, from a gene-centric perspective, only 602
genes show significant changes in major transcript dominance across conditions
(see Methods), which represents a small fraction of the set of genes with reported
alterations in splicing (i.e. 7.68% of the 7,842 genes previously identified). Even
though the first observation is consistent with the broad alterations described
in the previous section, the results from the gene-centric analysis suggest that
most of those alterations do not lead to big changes in the abundance of major
transcripts, at least in a recurrent fashion.
On the other hand, ⇠50% of the genes with altered splicing were predicted to
undergo changes in the identity of their major transcript in at least one sample
pair; cases that are further referred to as switch events (Figure 3.5a; see Methods).
Switch frequency decreases rapidly as the stringency on their expression breadth
is increased, suggesting that genes do not tend to be recurrently switched
(e.g. ⇠65% of genes that undergo switch events do so in less than 50% of the
patients where they are expressed; Figure 3.5b-d). Nonetheless, a small fraction of
broadly switched genes can be detected, which are dominated by those expressed
in a low number of samples (Figure 3.5b-c). Additional filtering of switch events
that involve 2-fold and 5-fold dominant transcripts in each condition revealed
a decrease in the proportion of genes being affected (i.e. 25% and 2% of genes
in the study set, respectively; Figure 3.5a), as well as in the number of samples
involved (Figure 3.5d). Altogether, these findings add to the idea that most of the
big changes are not recurrently detected.
In the context of switch events, an interesting set of genes are those that are
recurrently switched. An arbitrary threshold of >50% expression breadth was
chosen to investigate such cases, leading to a total of 1,112 protein coding genes
that fulfil this criteria, which were further analysed based on the potential
71
Chapter 3. The prevalence of splicing changes in cancer
number of patients
(a)
(b)
(d)
(c)
0
20
0
40
0
60
0
80
0
10
00
0 10 20 30 40
0
50
10
0
15
0
0 10 20 30 40
0
10
0
20
0
30
0
40
0
50
0
0 10 20 30 40
nu
m
be
r o
f g
en
es
nu
m
be
r o
f g
en
es
nu
m
be
r o
f g
en
es
all SE
2-fold SE
5-fold SE
same major transcript
5-fold SE
2-fold SE
switch events (SE):
function gain
other
function loss
protein coding (same sequence)
protein coding (different sequence)
0
% differentially spliced genes
1020304050
DOMINANCE
BIOTYPE
0 20 40 60 80 100
0
20
40
60
80
10
0
expression breadth (%)
0 20 40 60 80 100
expression breadth (%)
recurrently
switched
genes
%
 sw
itc
he
d 
ge
ne
s
nu
m
be
r o
f p
at
ie
nt
s
0
10
20
30
40
45
Figure 3.5| Switch events between tumours vs. healthy matched samples in
protein coding genes.
(a) Percentage of differentially spliced genes involved in switch events. Half of the genes
with altered splicing present changes in the identify of major transcripts across
conditions in at least one sample pair. This percentage decreases when consider-
ing events that involve dominant transcripts. Most commonly, switch events in
protein coding genes involve two different coding transcripts that differ in their
protein sequence.
(b) Proportion of switched genes relative to their expression breadth. Genes are rarely
switched in all the samples where they are expressed.
(c) Number of samples involved in switch events relative to switch expression breadth.
Recurrently switched genes tend to be expressed in a lower number of samples.
(d) Frequency of switch events relative to the number of samples involved in the switch.
In most cases, genes are not recurrently involved in switch events. This is spe-
cially true for events that involve dominant transcripts.
72
Chapter 3. The prevalence of splicing changes in cancer
functional implications of the associated switch events. This information can
be inferred from the transcript biotype information available in Ensembl [Flicek
et al., 2012], which allows the classification of transcripts derived from protein
coding genes into four main categories: protein coding, nonsense-mediated
decay, retained intron and processed transcript (see Chapter 2 - Methods for a
description). Thus, switch events from the set of recurrently switched genes can
be classified depending on whether they involve (i) a protein coding transcript in
each condition, (ii) a protein coding transcript in the healthy tissue but not in the
tumour (function loss), (iii) the reverse situation (function gain), or (iv) any other
combination of the above mentioned biotypes (other).
This classification system revealed that, in about half of the cases (i.e. 49%),
those events involve two different protein coding transcripts which differ in their
coding sequence (Figure 3.5a). PPP2R4, a Ser/Thr phosphatase that participates
in the negative control of cell division and growth, stands out as a clear example
in this category of events (Figure 3.6a). Notably, the recurrent major transcript
in healthy samples is classified as the principal isoform by APPRIS [Rodriguez
et al., 2013], while this is not the case for the one recurrently detected in tu-
mours. A similar assessment can be obtained based on the crystal structure of
the corresponding protein, where the smaller peptide that would derive from
the major transcript detected in tumours is predicted to be non-functional by
MAISTAS [Floris et al., 2011]. Following with the functional classification of
switch events, it is also possible to identify cases where the two transcripts
involved encode the same protein and consequently differ only in their UTR
sequences. For example, a major transcript with longer UTRs was recurrently
detected in tumour samples for the ASPA gene (Figure 3.6b), an N-acetylaspartate
(NAA) scavenger. Alternatively, the splicing factor SRSF6 and the transcriptional
repressor PATZ1 are clear examples within the function loss and function gain
categories, respectively (Figure 3.6c and Figure 3.6d). Finally, switch events that
involve dominant transcripts also constitute an interesting category, even though
they are not as broadly detected (Figure 3.5d). For example, FGFR2, for which
splicing alterations had already been reported in the context of ccRCC [Zhao
et al., 2013], was detected to undergo a switch event in 22 out of 41 patients,
which in 7 cases involved 2-fold dominant and coding transcripts in both normal
and tumour samples.
73
Chapter 3. The prevalence of splicing changes in cancer

protein_coding
 ENST00000347048

protein_coding
 ENST00000393370
0.0 0.2 0.4 0.6 0.8 1.0
transcript relative abundance
(a)
PPP2R4
 
  
20 40 60 80 100 120
gene expression (FPKMs)
ENSG00000119383
T
N
T
N
T
N
protein_coding
 ENST00000347048
protein_coding
ENST00000347048
protein_coding
 ENST00000393370
protein_coding
ENST00000393370
total annotated transcripts: 27
A
B
C
D
(b)
ASPA
T
N
T
N

0 10 20 30 40
gene expression (FPKMs)
ENSG00000108381
 
protein_coding
 ENST00000263080

protein_coding
 ENST00000456349
0.0 0.2 0.4 0.6 0.8 1.0
transcript relative abundance
T
N
protein_coding
 ENST00000456349
protein_coding
 ENST00000263080
protein_coding
ENST00000456349
nonsense_mediated_decay 
ENST00000263080
total annotated transcripts: 2
Figure 3.6| Examples of recurrent switch events. Each panel contains infor-
mation on the summarised gene and transcript expression profiles in tumours
vs. healthy samples (A and B, respectively), as well as sample-specific expression
profiles for a subset of three patients (C). The latter plot constitutes a novel visu-
alisation strategy, where each sample is represented as a pie chart, and each of
the slices corresponds to a transcript. The size of each slice is proportional to the
transcript expression level, and the overall size of the plot is proportional to the
gene expression level, thus allowing for comparisons across samples. Exon and
intron structures for the highlighted transcripts are also included in D. None of
the genes depicted here is differentially expressed.
(a) A switch between two protein coding transcripts for the PPP2R4 gene.
(b) A switch event between two protein coding transcripts that differ only in the UTRs
for the ASPA gene.
74
Chapter 3. The prevalence of splicing changes in cancer
(d)
PATZ1
T
N
T
N

2 4 6 8 10 12
gene expression (FPKMs)
ENSG00000100105

protein_coding
 ENST00000215919
retained_intron
 ENST00000465287
0.0 0.2 0.4 0.6 0.8 1.0
transcript relative abundance
T
Nretained_intron
 ENST00000465287
protein_coding
 ENST00000215919
retained_intron
ENST00000465287
protein_coding
ENST00000215919
total annotated transcripts: 6total annotated transcripts: 2
T
N
T
N

20 30 40 50 60
gene expression (FPKMs)
ENSG00000124193

protein_coding
 ENST00000244020
protein_coding
 ENST00000244020
0.0 0.2 0.4 0.6 0.8 1.0
transcript relative abundance

T
N
nonsense
mediated decay
 ENST00000483871
(c)
SRSF6
protein_coding
ENST00000244020
nonsense_mediated_decay 
ENST00000483871
nonsense
mediated decay
 ENST00000483871
Figure 3.6| Examples of recurrent switch events (continued).
(c) A switch event from a protein coding transcript to a nonsense-mediated decay one for
the SRSF6 gene.
(d) A switch event from a retained intron to a protein coding one for the PATZ1 gene.
75
Chapter 3. The prevalence of splicing changes in cancer
3.2.3 Identification, annotation and visualisation of switch events with
SwitchSeq
The analysis workflow followed in the previous section, based on the detection
and interpretation of switch events that affect protein coding genes, has been
automated within the tool SwitchSeq [Gonzàlez-Porta and Brazma, 2014]. Briefly,
this tool facilitates the interpretation of the changes in splicing reported by tools
like DEXSeq [Anders et al., 2012] and MMDIFF [Turro et al., 2014], by letting the
user focus on the most extreme cases (i.e. switch events).
SwitchSeq input consists of a matrix of normalised transcript-level counts, as well
as several annotation files that can be easily obtained with the provided wrap-
per tool (Figure C.1). Based on the expression data provided by the user, the
tool identifies the most abundant transcript within each gene and detects cases
where its identity changes across conditions, here referred to as switch events.
The detected events are then annotated by incorporating information from sev-
eral public resources (Ensembl, APPRIS, UniPDB), and the results are reported
in a self-contained HTML report, as well as in txt and json format. Furthermore,
SwitchSeq produces plots for the visualisation of the identified events, similar to
those depicted in Figure 3.6, which can be easily accessed from the report. Such
plotting capabilities have been implemented in an independent R package (tviz),
and hence can be used independently of any SwitchSeq execution.
3.2.4 Pathways appear as broadly disrupted when integrating different
layers of information
Switch events that are not as commonly shared across patients are also interesting,
specially if further alterations have been reported for the same gene in the same
patient. With this question in mind, and focusing on the subset of 36 matched
samples for which there is both RNA-seq and DNA-seq data available, I inte-
grated sample-specific information on switch events with the results from other
analyses performed in the context of CAGEKID, namely somatic mutations and
gene expression changes. Out of the 585 genes with non-silent somatic mutations
identified (see Methods), 226 overlap with those reported in previous studies
[Sato et al., 2013; TCGA Network, 2013], and from those, a subset of 42 also
undergo switch events in at least one patient. Such patient-specific alterations,
including changes in gene expression, can be visualised in Figure 3.7. Notably,
76
Chapter 3. The prevalence of splicing changes in cancer
VHL, which has been reported as the most frequently mutated gene in ccRCC,
is not part of this subset of genes. On the other hand, PBRM1, the second most
mutated gene in this type of cancer, is affected by switch events in 5 patients,
which, interestingly, do not overlap with those in which this gene is mutated.
Other genes that have been reported as significantly mutated in ccRCC and which
are also affected by switch events include PTEN, MTOR and the recurrently
switched BAP1 [Sato et al., 2013; TCGA Network, 2013].
The same strategy can be used to investigate patterns in the alterations that af-
fect components of relevant pathways in ccRCC (see section 3.2.1). For example,
in the well characterised VHL/HIF pathway (Figure 3.8), such a visualisation ap-
proach makes it easy to discriminate genes that are predominantly altered at the
genomic level (e.g. VHL) vs. cases in which transcriptomic alterations prevail, either
as changes in gene expression (e.g. EGLN3, SLC2A1, TGFB1), splicing (e.g. PIK3R,
HIF1A, AKT2) or both (e.g. VEGFA, ETS1). Similar patterns can be observed when
focusing on the PI(3)K-AKT-MTOR pathway (Figure C.2), which also plays a rele-
vant role in ccRCC, and the focal adhesion pathway (Figure C.3), located upstream
of the former. Overall, such analyses evidence that, even though most genes are
not altered in a recurrent fashion, pathways are broadly disrupted when combin-
ing information across multiple layers of information.
77
78
KIAA1217
SPTBN1
LAMA5
ZEB2
NLRP6
YLPM1
MSH3
COL12A1
SLC5A12
LRP2
TET2
FAT1
MTOR
ZGPAT
DST
MDN1
SYNE1
BAP1
PTEN
MLL3
PBRM1
COL14A1
ANK3
EIF4G1
FNDC3A
DDX24
ASTN2
DHX9
RAPGEF5
MACF1
MYH9
DENND4C
MPDZ
WWP2
LRRK2
PLEKHA7
AKAP9
DOCK2
ACIN1
DDX46
NUP205
TRIP11
patients
switch events
gene expression
somatic mutationsge
ne
s
function gain
function loss
other
protein coding
downregulated
upregulated
SWITCH EVENTS
GENE EXPRESSION
codon alteration
combination
frameshift
missense
splice site
stop gain
SOMATIC MUTATIONS
Figure 3.7| Patient-specific landscape of alterations for a subset of the genes
with recurrent somatic mutations. A set of 42 genes with recurrent somatic
mutations across three different studies and which undergo a switch event in at
least one sample are represented here. For each gene, patient specific informa-
tion was obtained by comparing the healthy and tumour samples and has been
summarised in three tracks: switch event information (top), gene expression dif-
ferences (middle) and confirmed somatic mutations (bottom). A gene is reported
as either up or down-regulated in a given patient if there is at least a 2-fold
change in its expression levels across conditions (i.e. 2-fold log10 FPKMs). Genes
have been sorted regarding the total number of somatic mutations identified.
79
TCEB1
HGF
TGFB3
MAPK3
PIK3CD
SOS1
PAK1
BRAF
EPAS1
RAPGEF1
CUL2
AKT2
HIF1A
ETS1
PIK3R3
TGFA
MET
TGFB1
VHL
SLC2A1
VEGFA
EGLN3
PAK2
PTPN11
NRAS
MAP2K1
FLCN
KRAS
GRB2
CDC42
FH
AKT1
GAB1
TGFB2
AKT3
CREBBP
MAP2K2
ARNT2
JUN
PDGFB
EP300
EGLN1
MAPK1
PIK3R1
PIK3R5
(a)
(b)
Figure 3.8| Summary of detected alterations in the VHL/HIF pathway.
(a) The VHL/HIF pathway. Genes which have been identified as altered in any
of the three considered analyses (i.e. switch events, gene expression changes and
confirmed somatic mutations) have been highlighted in red.
(b) Patient-specific landscape of alterations for genes in the VHL/HIF pathway. Each
gene is represented by three tracks as described in Figure 3.7, and have been
sorted based on the total number of alterations of any kind. Only those previ-
ously highlighted in a are included here.
Chapter 3. The prevalence of splicing changes in cancer
3.2.5 Splicing patterns in cancer cell lines are different from those in
primary tissues
Cell lines are widely used as cancer models, and thus, it is important to un-
derstand how they differ from primary tissues. Several studies have evidenced
differences in their transcriptional programs, reporting alterations in metabolic
pathways, cell division and communication (e.g. Auman and Howard [2010];
Ertel et al. [2006]; van Staveren et al. [2009]), but changes in splicing have
been so far unexplored. In this section, I will focus on the comparison of splic-
ing patterns in primary tissues vs. cell lines by analysing a set of 6 ccRCC cell lines.
In general terms, and consistent with the results reported for primary tissues
both here and in Chapter 2, ccRCC cell lines also express one dominant transcript
per gene (Figure 3.9). Nonetheless, there are differences in the percentage
of the mRNA pool that can be attributed to dominant transcripts (Figure 3.9
vs. Figure 3.3), which constitute a first hint of differences in the splicing programs.
For example, 2-fold dominant transcripts account for 83% of the mRNA pool in
healthy tissues, while this percentage decreases to 74% and 67% for tumours and
cell lines, respectively. In the same line, clustering analysis based on transcript
relative abundances revealed that cell lines behave differently from primary
tissues in terms of splicing patterns (Figure 3.10). Notably, matched samples clus-
tered by disease status rather than by patient, thus indicating that the alterations
described in the previous sections dominate above inter-individual variability.
Closer inspection of splicing differences with DEXSeq revealed that there are
8,831 genes with at least one exon differentially used between cell lines and
tumour samples. This represents 26% of the expressed genes (n = 33,982), a lower
percentage than the one obtained when comparing healthy vs. tumour samples.
Pathway enrichment analysis of the set of differentially spliced genes suggested a
role in metabolism, cell cycle control and several signalling pathways (Table C.1),
consistent with previous findings regarding differences in gene expression [Ertel
et al., 2006].
80
0.
8
major transcripts
0.
0
0.
2
0.
4
0.
6
de
ns
ity
1 2 3 4 5 +5
0.
0
0.
2
0.
4
0.
6
0.
8
1.
0
tr
an
sc
rip
t r
el
at
iv
e 
ab
un
da
nc
e
transcript position in the ranking
10-2 1 100 10,000
transcript expression level 
(log10 scale)
1 2 3 4 5 +5
0
20
40
60
80
10
0
%
 m
RN
A 
po
ol
transcript position in the ranking
2-fold dominant 
major transcripts
5-fold dominant 
major transcripts
minor transcripts
1 FPKM 9,226 (50.01%) 5,962 (64.62%) 3,429 (37.16%)
expressed genes
genes with a dominant major transcript
2-fold dominance 5-fold dominance
2,480 (41.79%)4,017 (67.66%)5,936 (32.18%)5 FPKM
1,759 (44.54%)2,733 (69.21%)3,949 (21.40%)10 FPKM
CELL LINES
(a) (b) (c)
(d)
Figure 3.9| Most protein coding genes express one dominant transcript in cell
lines. Panels to be be interpreted as in Figure 3.3.
normal
tumour
cell lines
Figure 3.10| Splicing patterns in cancer cell lines vs. primary tissues. Samples
were clustered based on transcript relative abundances. There is a clear separa-
tion between cell lines (green) and primary tissues, as well as between matched
healthy (blue) and tumour samples (red). This suggests that cancer cell lines
differ from primary tumours in their splicing patterns and that disease status
prevails over inter-individual variability.
81
Chapter 3. The prevalence of splicing changes in cancer
Moreover, the extent to which ccRCC cell lines resemble primary tumours in terms
of switch events was also investigated. In this context, a potentially interesting set
of genes are those which had been previously identified as recurrently switched
(n = 1,112). Focusing on the most common switch event for each gene, the overlap
in the identity of the involved transcripts and the most recurrent major transcript
in cell lines was evaluated, thus leading to four possible different scenarios: (i)
cases where the most recurrent major transcript in cell lines coincides with the
one detected in tumours; (ii) cases where it coincides with the one identified in
normal samples; (iii) cases where there is no overlap or where there are several
recurrent transcripts that can be ambiguously classified; and (iv) cases where the
corresponding gene is not expressed in cell lines. Such classification revealed that
recurrently expressed major transcripts in ccRCC cell lines overlap significantly
more often with those detected in tumours than with those detected in healthy
tissues (Figure 3.11a; Chi-square test, p-value = 5.531 · 1008). For example, the
previously highlighted gene PPP2R4 has consistent splicing patterns between cell
lines and tumours (Figure 3.11b). However, this situation represents only 32%
of the cases, thus suggesting that, most of the time, major transcript expression
patterns are not conserved. Surprisingly, SRSF6, for which a clear switch event
has been reported earlier in this chapter, emerges as an example of the divergence
in splicing patterns between cell lines and tumours (Figure 3.11c). Finally, 19% of
the genes that had been identified as recurrently involved in switch events are not
detected as expressed in cell lines.
3.3 Discussion
The results described in this chapter constitute the first in-depth characterisation
of the splicing alterations that underlie ccRCC, the most common type of renal
cancer [Jonasch et al., 2012]. Formerly, most of the efforts towards the character-
isation of this disease have focused on the genomic level, with recent significant
advances derived from three different large scale sequencing projects [Sato et al.,
2013; Scelo et al.; TCGA Network, 2013]. The few existing attempts to understand
the potential disruption of splicing programs in ccRCC have mostly been based
on the use of differential exon expression as a measure of exon skipping [Valletti
et al., 2013; Zhao et al., 2013]. However, such approach does not cover the totality
of annotated splicing events, as evidenced by the differences in the number
of genes reported to undergo differences in splicing (i.e. hundreds in previous
82
Chapter 3. The prevalence of splicing changes in cancer
same major transcript as T
same major transcript as N
other
not expressed
% genes
0 5 10 15 20 25 30 35
(a)
(b) (c)
PPP2R4 SRSF6
protein_coding
 ENST00000347048
(same as T)
protein_coding
 ENST00000244020
(same as N)
Figure 3.11| Switch events in cancer cell lines vs. primary tissues.
(a) Overlap in the identity of the most recurrent major transcript in cell lines and those
involved in recurrently switch events. Major transcripts in ccRCC cell lines overlap
significantly more often with those detected in tumours than with those detected
in healthy tissues (Chi-square test, p-value = 5.531 · 1008). N: normal; T: tumour.
(b-c) Sample-specific transcript usage in the six cell lines analysed for a subset of genes
previously identified to undergo switch events. The major transcripts detected in cell
lines tend to coincide with those detected in primary tumours (e.g. PPP2R4, see
Figure 3.6a); however there are also genes for which the most abundant transcript
in cell lines corresponds to the one detected in healthy samples (e.g. SRSF6, see
Figure 3.6c).
studies vs. thousands in the present chapter). Thus, I show that splicing patterns
are largely altered in this type of cancer, consistent with the wide alterations
already described at the gene level [Sato et al., 2013; Scelo et al.; TCGA Network,
2013]. Furthermore, these observations link to the identification of the mRNA
processing pathway as commonly disrupted in ccRCC [Sato et al., 2013].
GO and pathway enrichment analysis of the set of genes that undergo differential
splicing in ccRCC points to a significant alteration of metabolic pathways,
including oxidative phosphorilation, the citrate cycle, glycolisis and the pentose
phosphate pathway. This is suggestive of a general metabolic shift and consistent
83
Chapter 3. The prevalence of splicing changes in cancer
with the Warburg effect, which constitutes a hallmark of ccRCC [Linehan et al.,
2010]. Moreover, such analyses identified a dysregulation of the focal adhesion
pathway, which contains genes that act upstream of PI(3)K. Abnormalities of the
extra-cellular matrix play key roles in tumour formation and progression [Cox
and Erler, 2011] either by affecting downstream growth-promoting pathways
[Levental et al., 2009] or by contributing to angiogenesis and metastasis by
influencing the tumour microenvironment [Lu et al., 2012]. Together with the
recurrently altered PI(3)K-AKT-MTOR signalling pathway [Sato et al., 2013; TCGA
Network, 2013], this novel finding highlights the relevance of this signalling
cascade as a therapeutic target for ccRCC.
In spite of the large number of differentially spliced genes identified, I observe
that big changes in splicing do not tend to be recurrently detected. These
observations are consistent with the abundant molecular heterogeneity identified
in ccRCC [Gerlinger et al., 2012; Martinez et al., 2013] and, more generally,
with the high levels of inter-tumour variability described for cancer [Cusnir and
Cavalcante, 2012; Shibata and Shen, 2013]. By focusing on major transcripts, I
estimate that, amongst the genes with differential splicing, only a small subset
manifest also differences in major transcript dominance between the healthy and
tumour tissues. Following the introduction of the concept of switch event as a
means to explore the most extreme changes in splicing, I detect that most of these
events can be detected only in a small number of patients (e.g. ⇠65% of genes
that undergo switch events do so in less than 50% of the samples where they
are expressed). Nonetheless, it is also possible to identify interesting cases of
recurrent switch events. For example, I predict a non-functional major transcript
in tumours for PPP2R4, a Ser/Thr phosphatase involved in the negative control
of cell division and growth. Similarly, I detect a switch between a protein coding
transcript and a nonsense-mediated decay one for the splicing factor SRSF6,
which has been reported as an oncogene [Michal et al., 2013]. Interestingly,
this is consistent with existing knowledge that levels of core components of the
splicesome (SmB/B’) can be auto-regulated through feedback loops that rely on
NMD [Saltzman et al., 2011]. Overall, these analyses evidence that many of the
detected splicing differences arise from lowly abundant transcripts, which could
potentially result from noisy splicing. Nonetheless, similarly to what has been
discussed in Chapter 2, one cannot rule out the possibility that minor transcripts
may have a functional impact on the cell’s fate. mRNA and protein levels are
84
Chapter 3. The prevalence of splicing changes in cancer
only correlated up to some extent, and hence it is possible that minor transcripts
are still translated. Even if this is not the case, they might still play a functional
role in the regulation of expression levels [McGlincy and Smith, 2008; Yap et al.,
2012]. Thus, evaluating the functional relevance of the annotated transcripts is
one of the challenging tasks ahead. In this context, incorporating information on
the conservation of protein sequences and features across species could help in
identifying potentially functional minor transcripts, and the disruption of key
protein domains and structures compared to major transcripts could indicate
functional defects [Rodriguez et al., 2013].
Beyond the general comparison of normal vs. tumour samples, patient-specific
analysis of transcriptomic and genomic alterations allowed for a more clear
picture of the complexity underlying this type of cancer. For example, I show
that VHL is predominantly altered at the genomic level. On the other hand, I
observe that PBRM1, the second most commonly mutated gene in ccRCC, is also
affected by switch events in a small number of patients. This is also the case
for BAP1, identified amongst the most mutated genes in this type of tumour,
and MTOR, a component of the PI(3)K-AKT-MTOR pathway that constitutes a
strong therapeutic target for ccRCC [Audenet et al., 2012]. In addition, I observe
that despite the low recurrence of the detected splicing changes, pathways are
broadly altered when aggregating information across different regulatory layers.
This evidences the need for integrative analysis in complex diseases such cancer,
and highlights the importance of sample-specific comparisons in uncovering such
complexity.
Finally, following analysis of a set of 6 ccRCC cell lines, I predict overall differ-
ences in the splicing patterns of cell lines vs. primary tissues when clustering
samples based on transcript relative abundances. Several studies had already
evidenced differences in transcriptional programs for cell lines (e.g. Auman
and Howard [2010]; Ertel et al. [2006]; van Staveren et al. [2009]), but changes
in splicing patterns have remained uncharacterised. Focusing on genes that
had been previously identified as recurrently switched, I show that recurrently
expressed major transcripts in ccRCC cell lines overlap significantly more often
with those detected in tumours than those from healthy tissues. However, in most
cases (i.e. 68%), major transcript expression patterns are not conserved between
cell lines and tumours, as exemplified by SRSF6. Overall, such observations
85
Chapter 3. The prevalence of splicing changes in cancer
reinforce the idea that transcriptome analysis in cell lines can be extended to
primary tissues only to some extent [Lukk et al., 2010].
Altogether, the findings described in this chapter constitute the tip of the iceberg
for a landscape of large transcriptomic alterations in ccRCC. By analysing alterna-
tive splicing patterns, it has been possible to detect changes that would not have
been picked up by a gene-level analysis, which might as well constitute actionable
therapeutic targets [Tang et al., 2013]. Nonetheless, there are still many more av-
enues to cover in order to gain a deep understanding of the ccRCC transcriptome,
including novel transcription and novel transcripts, as well as potential alterations
affecting the expression of long non-coding RNAs (lncRNAs) and micro-RNAs
(miRNAs). RNA-seq has undoubtedly facilitated the study of the transcriptome,
and as the costs of sequencing continue to fall down, having access to a more inte-
grated picture through the combination of different layers of information (e.g. ge-
nomics, transcriptomics, epigenetics, etc.) will not only enable the discovery of
novel biomarkers in research, but will soon constitute a reality in the clinics.
3.4 Computational methods
All the computational analyses described in this chapter have been performed by myself,
with the exception of the differential expression and somatic mutation analyses, which have
been carried out by Dr. Johan Rung and Dr. Louis Letourneau, as detailed below.
Datasets and mapping
RNA from 45 healthy kidneys and matched ccRCC tumours was extracted and
polyA-selected following Illumina’s standard protocols. Sequencing was per-
formed on an Illumina HiSeq 2000 platform (100 bp paired-end reads; Table A.2).
In addition, DNA-seq data for 36 of the 45 matched samples was also produced
with the same platform (100 bp paired-end reads). In all the cases, revision
by multiple pathologists ensured a 70% content of tumour cells in the ccRCC
samples. Finally, RNA from a separate set of 6 ccRCC cell lines was extracted and
prepared for sequencing following the same Illumina’s protocols and platform
(100 bp paired-end reads; Table A.2).
Raw RNA-seq reads were initially trimmed down to 95 nucleotides from the 3’
86
Chapter 3. The prevalence of splicing changes in cancer
end using PRINSEQ v0.19.5 [Schmieder and Edwards, 2011] and mapped to the
reference genome (Ensembl 66 [Flicek et al., 2012]) using TopHat v2.0.6 [Kim et al.,
2013]. The DNA-seq data were analysed by Dr. Mark Lathrop’s group at McGill
University and Genome Quebec Innovation Centre (see Integrative analyses).
Gene level analyses
Analyses performed by Dr. Johan Rung:
Gene counts were estimated using HTSeq v0.5.3p7 [Anders et al., 2014], using
the --intersection-nonempty and --stranded=no parameters. Genes with zero
counts for more than nine tumour or normal samples were filtered out, and
the rest were considered as expressed. A paired test for differential expression
was performed for the 45 matched sample pairs using edgeR [Robinson et al.,
2010], and significance was assessed with an FDR threshold of 0.01 (Benjamini
& Hochberg correction [Benjamini and Hochberg, 1995]). Gene FPKMs were
obtained from the calculated counts using custom scripts.
Analyses performed by myself:
Gene expression levels of ccRCC cell lines vs. primary tumours were compared
with edgeR as detailed above. Following the previous strategy, only genes with
non-zero counts for five or less samples were considered as expressed.
GO and pathway enrichment analyses were performed with DAVID [Huang et al.,
2009a,b] and WebGestalt [Wang et al., 2013], and significance was assessed with an
adjusted p-value of 0.01 (Benjamini & Yekutieli correction [Benjamini and Yeku-
tieli, 2001]). All the information on gene function was retrieved from Genecards
[Safran et al., 2010] unless otherwise indicated in the text.
Transcript level analyses
Differential splicing and splicing variability
DEXSeq v1.7.0 [Anders et al., 2012] was used to identify genes that undergo
differential exon usage across the two studied conditions (i.e. healthy kidney
vs. ccRCC samples). An FDR threshold of 0.01 was used to assess significance
(Benjamini & Hochberg correction [Benjamini and Hochberg, 1995]).
Splicing variability was calculated following the method introduced by Gonzàlez-
87
Chapter 3. The prevalence of splicing changes in cancer
Porta et al. [2012]. Briefly, for each gene, the relative abundances of the annotated
transcript isoforms for all the individuals in the study set (normal or tumour)
are represented in a multi-dimensional space, with the number of dimensions
corresponding to the number of transcripts. Then, the mean Hellinger distance to
the centroid of all the data points (individuals) is taken as a measure of splicing
variability (i.e. the higher the observed dispersion, the bigger the variability in
transcript relative abundances within individuals of the study set).
Major transcripts and switch events
Relative abundances for the transcripts annotated in Ensembl 66 were obtained
using MISO v0.4.1 [Katz et al., 2010]. Transcript FPKMs were then obtained
by multiplying those relative abundances with the corresponding gene FPKMs.
Similarly to the nomenclature introduced in Chapter 2, the most abundant
transcript within each gene is referred to as major transcript.
Major transcript dominance was assessed by calculating, for every gene, the ratio
of the expression levels between the second most abundant transcript and the
major one. For each sample, the number of expressed genes with a dominant tran-
script (i.e. 2-fold dominant major transcript) was calculated. Then, the difference
in the numbers obtained for the normal and tumour samples from each given
patient was assessed with a McNemar’s test (patient-centric analysis). Similarly,
for each gene in the set of differentially spliced genes, the difference in major
transcript dominance between all normal vs. tumour samples was interrogated
with a Wilcoxon test (gene-centric analysis). In both analyses, p-values were
corrected for multiple testing using the Benjamini-Hochberg method [Benjamini
and Hochberg, 1995].
Changes in major transcript identity across conditions were referred to as switch
events (see Figure 1.12c for an illustration). A given switch event was classified
as dominant if both transcripts involved are expressed in a dominant fashion. For-
mally, given a gene G, a pair of transcripts Ik and Il and two paired samples Si
and Sj from patient P, we say that gene G undergoes an x-fold switch between
transcripts Ik and Il in samples Si and Sj, if G is expressed in both Si and Sj and
the ratio of the expression of Ik to Il is at least x in Si and no more than 1/x in Sj.
In addition, switch events were also classified as recurrent if they were present in
>50% of the samples in which the corresponding gene was expressed.
88
Chapter 3. The prevalence of splicing changes in cancer
Cluster analysis
Spearman correlation coefficients were calculated by comparing transcript relative
abundances across samples in a pairwise fashion and used to cluster samples
using the R function heatmap.2.
Integrative analyses
Analyses performed by Dr. Louis Letourneau:
Non-silent somatic mutations for a total set of 585 genes were identified with
Genome MuSiC v0.4 [Dees et al., 2012] (FDR < 0.2) and validated by pooled
amplicon sequencing on an Illumina MiSeq or by Sanger sequencing.
Analyses performed by myself:
A gene is reported as either up or down-regulated for a given patient if the
observed difference in expression levels across matched samples is more than
2-fold (i.e. >2-fold log10 FPKMs).
Pathway information was retrieved from KEGG [Kanehisa and Goto, 2000; Kane-
hisa et al., 2014] and reproduced with the permission of Kanehisha Laboratories.
89

Chapter 4
The regulation of splicing by core
spliceosomal factors
Gaining a comprehensive knowledge of all the factors involved in intron removal
and their specific contribution in shaping splicing decisions is the key to under-
standing the basis of this reaction and how its disruption might contribute to
pathogenesis [Barash et al., 2010]. This chapter describes a collaborative effort to
further characterise the function of the splicing factor PRPF8, and to dissect the
role of core spliceosomal components in the regulation of the splicing reaction.
In the first part, I describe the phenotypic effects obtained following knock-down
of this splicing factor. Next, I link such effects to alterations in splicing, and I
identify common characteristics of the set of genes that are most affected by such
down-regulation. Finally, I evaluate the impact of PRPF8 knock-down on the rate
of splicing as a potential explanation for the observed effects.
All the computational analyses have been performed by myself under the su-
pervision of Dr. John Marioni from EMBL-EBI, unless otherwise indicated in the
Methods. The experimental work has been carried out by Dr. Vi Wickramasinghe,
Dr. David Perera and Arthur Bartolozzi in Dr. Ashok Venkitaraman’s laboratory
at the Hutchison Research Institute. Finally, Dr. Christopher W. Sibley from Dr.
Jernej Ule’s laboratory at University College London produced and analysed the
iCLIP dataset.
91
Chapter 4. The regulation of splicing by core spliceosomal factors
Publications derived from this chapter
• (submitted) Wickramasinghe?, Gonzàlez-Porta? et al. PRPF8 abundance
dictates the patterns of alternative and constitutive messenger RNA splicing.
?shared first authors
92
Chapter 4. The regulation of splicing by core spliceosomal factors
4.1 Introduction
The splicing reaction is a multi-step process in which over one hundred proteins
take part, typically referred to as splicing factors (SFs). Together with several
small nuclear RNAs (snRNAs), these proteins form a large protein complex -
the spliceosome - which has been widely investigated in terms of composition,
function and structure (e.g. reviewed in Hoskins and Moore [2012]; Matera and
Wang [2014]). Amongst the collection of SFs that participate in splicing, the
core spliceosomal component PRPF8 stands out because of its location at the
centre of the spliceosome [Galej et al., 2013; Grainger and Beggs, 2005]. This
SF is part of the U5 small nuclear ribonucleoprotein particle (snRNP) and the
B-complex, and has been suggested as a master regulator of splicing [Grainger
and Beggs, 2005]. More specifically, PRPF8 is known to act as a scaffold during
spliceosomal assembly, and to be involved in the activation of the B-complex
through the recruitment of the the U4-U6·U5 tri-snRNP ([Li et al., 2013]; see
Introduction - The splicing reaction). Notably, this SF has been established as a
candidate gene for autosomal dominant retinitis pigmentosa, a genetic disease
that affects specifically retinal cells and that is characterized by a reduction in
splicing activity [Tanackovic et al., 2011]. Furthermore, similarly to other SFs, it
has been detected to play an essential role in cell division [Neumann et al., 2010].
However, despite its key involvement in the splicing reaction, its clinical relevance
and its requirement for proper cell cycle progression, much of the structure and
mode of action of PRPF8 remain uncharacterised.
In this chapter, I describe the work carried out in collaboration with Dr. Ashok
Venkitaraman’s and Dr. Jernej Ule’s teams in order to further elucidate the func-
tion of PRPF8 in the context of its associated mitotic arrest phenotype. Following
down-regulation of this SF with small interference RNAs (siRNAs), we confirmed
and characterised the resulting disruption in cell cycle progression. Moreover, by
introducing properly spliced mRNAs into PRPF8 knock-down (KD) cells, we show
that such phenotype can be attributed to splicing defects, rather than to a direct
role of PRPF8 in cell division, as reported for other splicing factors [Hofmann
et al., 2010]. Further results from Sm iCLIP experiments indicate that spliceoso-
mal binding is affected at a global scale, hence pointing to an overall disruption
of splicing. We further generated and analysed an RNA-seq dataset in order to
compare splicing patterns in control vs. PRPF8 KD cells, and consistent with the
93
Chapter 4. The regulation of splicing by core spliceosomal factors
above results, we observed higher intronic expression levels in the latter. We also
detect differences in splicing for a subset of the expressed genes (n = 3,388 out
of 13,216 protein coding genes) that are enriched for those encoding mitotic cell
cycle factors, thus linking the identified alterations to the observed phenotype.
Focusing on intronic reads, we identify a subset of introns that remain preferen-
tially unspliced after PRPF8 KD, and find that they harbour weaker splice sites
and GC composition similar to that of adjacent exons, characteristics that have
been previously linked to intron retention [Amit et al., 2012; Sakabe and de Souza,
2007]. Mini-gene experiments further show that increasing splice site strength
can compensate for the down-regulation of PRPF8, hence suggesting that the ki-
netic competition across splicing signals could constitute an explanation for the
observed alterations. Finally, and consistent with this kinetic competition model,
we predict and validate changes in the rate of splicing following PRPF8 KD by re-
lying on novel computational analyses based on intronic reads. Altogether, these
findings highlight PRPF8 as a key player for the completion of splicing in tempo-
rally constrained processes such as cell division and reveal an intimate role of this
SF in regulating the dynamics of splicing.
4.2 Results
The findings described here are the result of a collaborative effort, which com-
prised the use of a variety of experimental and computational approaches for
the purposes stated above. Even though the present chapter includes an overall
description of the project, the emphasis has been placed on the analysis of the
RNA-seq data, in which I have been most involved.
4.2.1 PRPF8 knock-down causes accumulation of cells in mitosis
Following down-regulation of PRPF8 (Figure D.1), it was possible to observe that
cells with reduced expression of this splicing factor did not progress through the
cell cycle and tended to accumulate in mitosis (Figure 4.1a). Moreover, PRPF8
knock-down (KD) cells manifested mitotic abnormalities that could provide an
explanation for the failure to exit this phase, as exemplified by the severe chromo-
some misalignment observed from chromosome alignment assays (Figure 4.1b).
Finally, knock-down of other B-complex components also led to a mitotic arrest
phenotype, while this was not generally the case for A and C-complex components
(Figure D.2; see Discussion).
94
Chapter 4. The regulation of splicing by core spliceosomal factors
(a)
control siRNA PRPF8 KD control siRNA
time in mitosis
PRPF8 KD
ce
lls
 in
 m
ito
sis
 (%
)
ce
lls
 (%
)
tubulin centromeres DNA merge
control
PRPF8
KD
(b)
Figure 4.1| Characterisation of the mitotic arrest phenotype obtained after
PRPF8 KD.
(a) Results from the cell cycle analyses. Cells with reduced expression of this splic-
ing factor tend to accumulate in mitosis (p-value < 0.001).
(b) Results from the chromosome alignment assays. Chromosome alignment assays
measure whether chromosomes are aligned correctly at the metaphase plate.
Compared to controls, PRPF8 KD cells display lagging chromosomes.
Figure provided by Dr. Vi Wickramasinghe.
4.2.2 The mitotic arrest phenotype is driven by disruptions in splicing
Given that PRPF8 is a component of the catalytic core of the spliceosome, it is
likely that the observed phenotype is caused by alterations in splicing. In order
to evaluate this hypothesis, PRPF8 KD cells were treated with a cytosolic extract
from control cells that contained polyA-selected mRNAs. Such treatment led to
partial phenotype recovery in PRPF8 KDs (Figure 4.2), suggesting that the lack of
properly spliced mRNA is affecting cell cycle progression. Similar results were ob-
tained from further Sm iCLIP experiments, a technique that allows for the study of
spliceosomal protein-RNA interactions along the genome at nucleotide resolution
[Briese et al.; König et al., 2010]. These experiments evidenced lower spliceosomal
95
Chapter 4. The regulation of splicing by core spliceosomal factors
binding in PRPF8 KD cells, thus pointing to global defects in the splicing reaction
(Figure 4.3 and Figure D.3). Notably, most of the differences in spliceosome occu-
pancy emerged at the 5’ splice site, consistent with the preferential interaction of
PRPF8 with this splicing signal [Li et al., 2013].
control
siRNA
PRPF8 KD PRPF8 KD +
cytoplasmic
mRNA
ce
lls
 in
 m
ito
sis
 (%
)
Figure 4.2| Phenotype recovery experiments in PRPF8 KD cells. Cytoplasmic
mRNAs were isolated from control cells and introduced in PRPF8 KDs. Such
treatment revealed a decrease in the proportion of cells arrested in mitosis, sug-
gesting that the observed alterations in cell cycle progression are caused by splic-
ing defects (control vs. PRPF8 KD: p-value < 0.001; PRPF8 KD vs. PRPF8 KD +
cytoplasmic mRNA: p-value < 0.01). Figure provided by Dr. Vi Wickramasinghe.
To gain further insights into the splicing alterations that led to the observed
mitotic arrest, the transcriptome of PRPF8 KD and control cells was sequenced on
an Illumina HiSeq2000 using 100 bp paired-end reads (4 and 3 biological repli-
cates, respectively; see Methods and Table A.3). Analysis of this RNA-seq dataset
revealed that intronic expression levels are higher in PRPF8 KDs compared to
control cells (p-value < 2.2 · 1016; Figure 4.4a and Figure 4.4b; see Methods), thus
pointing to an overall misregulation of splicing patterns, as initially suggested by
the iCLIP experiments. Following such observation, DEXSeq was used to detect
intron retention events (see Methods), and this analysis led to the identification
of a set of 2,086 protein coding genes that contain at least one retained intron.
Similarly, from the execution of DEXSeq on annotated exons it was possible to
detect a set of 1,921 protein coding genes with significant differences in exon
usage across conditions (FDR < 0.01; see Methods). These results suggest that
down-regulation of PRPF8 not only impairs intron removal, but also affects
alternative splicing patterns. Notably, there is a significant overlap between this
96
Chapter 4. The regulation of splicing by core spliceosomal factors
-100 -50 0 50 100
0
50
0
10
00
15
00
20
00
25
00
-100 -50 0 50 100
0
50
0
10
00
15
00
20
00
25
00
su
m
 o
f n
or
m
al
ise
d 
cD
N
A 
co
un
ts
su
m
 o
f n
or
m
al
ise
d 
cD
N
A 
co
un
ts
position relative to junction position relative to junction
CONTROL PRPF8 KD(a)
normalised cDNA counts
control
PRPF8 KD
(b)
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7
Figure 4.3| Results from Sm iCLIP experiments for the 5’ splice site.
(a) RNA maps for control and PRPF8 KD samples. RNAmaps provide a summarised
view of the detected spliceosomal binding sites across all exon-intron junctions,
where each peak corresponds to different RNA-binding proteins [Briese et al.].
Such maps suggest that the down-regulation of PRPF8 leads to broad differences
in the studied RNA-protein interactions. Mean + SD across replicates have been
plotted for each condition (see Methods).
(b) Changes in the distribution of normalised cDNA counts between control and PRPF8
KD samples. Normalised cDNA counts in the shaded regions from a have been
plotted here for all exon-intron junctions. The overall decrease in spliceosomal
binding in PRPF8 KDs evidences that the previously detected differences are not
caused by a small subset of genes (p-value < 2.2 · 1016, Wilcoxon test).
set of genes and those previously detected to display retained introns (n = 637;
p-value < 2.2·1016). Most importantly, the set of genes with reported alterations
in splicing (either in the form of intron retention or differential exon usage events)
were detected to constitute only a subset of the expressed protein coding genes (n
= 3,388 out of 13,216; expression threshold = 1 FPKM; Figure 4.4c; see Methods),
suggesting a certain degree of specificity in the mode of action of PRPF8.
97
Chapter 4. The regulation of splicing by core spliceosomal factors
(a)
0.0 0.5 1.0 1.5 2.0 2.5
normalised intron expression
control
PRPF8
KD
(b)
control
PRPF8
KD
exons (%)
86.53
85.70
85.79
78.88
79.12
80.21
80.12
mapped
reads (M)
17.24
22.00
25.76
20.62
12.84
24.97
24.63
exon-intron
boundaries (%)
4.59
4.70
4.72
5.49
5.51
5.48
5.48
intron
bodies (%)
4.85
5.41
5.36
11.18
10.84
9.95
10.12
total
introns (%)
9.44
10.12
10.07
16.68
16.35
15.43
15.60
(c)
0 20 40 60 80 100
% expressed genes
intron
retention
both
differential
exon usage
GO term adjusted p-value
RNA processing 1.23E-20
RNA splicing 1.14E-10
ubiquitin conjugation pathway 3.66E-13
ubiquitin-dependent protein catabolic process 3.83E-10
cell cycle 1.05E-11
mitosis 5.73E-08
cell division 4.70E-08
98
Chapter 4. The regulation of splicing by core spliceosomal factors
Figure 4.4| Characterisation of splicing alterations in PRPF8 KD cells.
(a) Intron expression levels in PRPF8 KD vs. control cells. Down-regulation of PRPF8
leads to an increase in intronic expression (p-value < 2.2 · 1016). Intron expres-
sion levels were normalised to take into account gene expression and library size
as specified in Methods.
(b) Percentage of reads mapping to exons and introns in PRPF8 KD vs. control cells. The
proportion of reads mapping to both exon-intron boundaries and intron bodies
increases after PRPF8 KD. Paired reads are counted once.
(c) Gene Ontology enrichment analysis for protein coding genes with altered splicing
after PRPF8 KD. Genes that manifest significant differences in splicing patterns
after PRPF8 KD (n = 3,388) are involved in cell division and mitosis, amongst
other functions.
Altogether, the above described sets of genes were determined to be in-
volved in cell division and mitosis through Gene Ontology enrichment analysis
(Figure 4.4c), thus linking these findings to the observed phenotype. Focusing
on the common genes from these two sets, the predicted splicing changes were
further validated by RT-PCR in a subset of genes with a role in cell division
(Figure 4.5a and Figure 4.5b). These included ASPM, observed to undergo an
exon skipping event; CDC23, for which an alternative terminal exon was detected;
and CDC20 and NUDC, both predicted to contain retained introns. Furthermore,
for both ASPM and CDC23, the observed differences were big enough to lead to a
change in the identity of major transcripts across conditions (Figure 4.5c). Finally,
such changes could also be validated at the protein level by Western blot in the
cases where antibodies were available (Figure 4.5d). Overall, these findings con-
firm the hypothesis that PRPF8 is required for proper splicing of genes involved
in mitosis.
99
Chapter 4. The regulation of splicing by core spliceosomal factors
Alternative 
terminal exon, 
truncated protein
Wild type
CDC23
ASPM
Wild type
Exon skipping,
removal of 
largest exon
C KD
NUDC
Retained intron
Wild type
CDC20
Retained intron
Wild type
(a)
C
KD
C
KD
C
KD
KD
C
C KD
C KD
C KD
protein_coding − ENST00000294732
protein_coding − ENST00000367409
KD
C
ASPM(c) C KD
PRPF8
CDC23
CDC20
Actin
250
75
50
50
(d)
control CDC23
m
RN
A 
ex
pr
es
sio
n
(n
or
m
al
ise
d 
to
 c
on
tr
ol
)
m
RN
A 
ex
pr
es
sio
n
(n
or
m
al
ise
d 
to
 c
on
tr
ol
)(b)
protein_coding − ENST00000394884
protein_coding − ENST00000394886
KD
C
CDC23
100
Chapter 4. The regulation of splicing by core spliceosomal factors
Figure 4.5| Validation of the predicted splicing changes in a subset of genes
with a role in cell division.
(a) RT-PCR validation results. Each panel includes information on the read cov-
erage in the corresponding genomic region in a control vs. PRPF8 KD sample
(right) and the RT-PCR validation results (left). Genes have been selected to rep-
resent a variety of splicing events.
(b) qRT-PCR validation results. The relative abundance for the amplification prod-
ucts labelled with arrows in a is represented here. ASPM and CDC23: p-value <
0.001; CDC20 and NUDC: p-value < 0.01.
(c) Estimated transcript expression levels. Both ASPM and CDC23 are predicted to
undergo a switch event in PRPF8 KD vs. control samples (i.e. different major tran-
scripts are detected in each condition; see Chapter 3 - Methods).
(d) Western blot results. Predicted alternative splicing events in genes for which
antibodies were available could be successfully validated at the protein level.
Figures with experimental results have been provided by Dr. Vi Wickramasinghe.
4.2.3 Modulating splice site strength can compensate for the down-
regulation of PRPF8
Further characterisation of the most extreme intron retention events suggested
that those introns that are not as efficiently removed after PRPF8 KD tend to
have weaker 5’ splice sites (Figure 4.6a; p-value = 0.0220, Wilcoxon test). Motif
enrichment analysis on the same set of genes revealed consistent results: while
the top identified motifs correspond to the consensus 5’ splice site sequence both
for retained (RI) and non-retained introns (NRI), the percentage of targets with
such motif varies significantly between the two categories (Figure 4.6b; 68.50% for
RI vs. 91.03% for NRI; p-value < 2.2·1016, Fisher’s exact test). Such differences
were not observed at the 3’ splice site (Figure 4.6c and d). Finally, mini-gene
experiments further demonstrated that increasing splice site strength can revert
the splicing defects observed in PRPF8 KD cells (Figure 4.7).
On a separate note, analysis of GC composition for the most extreme RI revealed
that, contrary to NRI, the former display GC content similar to that of adjacent
exons (p-value < 2.2 · 1016, Wilcoxon test; Figure 4.8a), as well as higher GC
content overall (p-value = 0.0055, Wilcoxon test; Figure 4.8b).
101
Chapter 4. The regulation of splicing by core spliceosomal factors
5’ splice site strength
RI
NRI
0 5 10 15
RI
NRI
top motif p-value % targets
1E-145 68.50
1E-2577 91.03
(a)
(b)
RI
NRI
(c) 3’ splice site strength
0 5 10 15
top motif p-value % targets
1E-177 83.50
1E-2939 85.50
(d)
RI
NRI
Figure 4.6| Splice site strength for the top retained introns following PRPF8
KD. A set of 200 retained introns (RI) were selected based on fold-change differ-
ences (see Methods), and non-retained introns (NRI) within the same set of genes
were used as a contrast.
(a) Differences in 5’ splice site strength. Top retained introns have weaker 5’ splice
sites compared to non-retained introns from the same set of genes (p-value =
0.0220, Wilcoxon test).
(b) Results from the motif enrichment analysis on the 5’ splice site. The top motif iden-
tified corresponds to the consensus 5’ splice site sequence both for RI and NRI.
However, the percentage of targets displaying such sequence differs significantly
between the two sets (p-value < 2.2·1016, Fisher’s exact test).
(c) Differences in 3’ splice site strength. No significant differences were detected
between RI vs. NRI (p-value = 0.2068, Wilcoxon test).
(d) Results from the motif enrichment analysis on the 3’ splice site. The findings in c
are supported by these results.
102
Chapter 4. The regulation of splicing by core spliceosomal factors
Exon                                         Intron                                             Exon
AGAGGGTAAGACWT
5’ mut AGCAGGTAAGAC
     TCACTCCGTTGCCACAGGTTAT
     TTTTTTTTTTTTTACAGGTTAT
     WT
3’ mut
5’ splice site 3’ splice site
   CDC20 gene (a)
PRPF8 KD
Control
WT       5’ mut    3’ mut  Double mut
***
n.s.
n.s.
n.s.
(c)
 Con   PRPF8 
siRNA siRNA 
Un-spliced
RNA
Spliced
RNA
 Con   PRPF8 
siRNA siRNA 
Con   PRPF8 
siRNA siRNA 
 Con   PRPF8 
siRNA siRNA 
WT        5’ mut     3’ mut  Double mut(b)
%
 s
pl
ic
ed
 C
D
C
20
 R
N
A
Figure 4.7| Mini-gene experiments for the CDC20 gene.
(a) Illustration of the experimental design. The CDC20 gene had been previously
identified to contain a retained intron following PRPF8 KD (see Figure 4.5). In
this experiment, the splicing efficiency of such retained intron was evaluated
with several mini-gene constructs, which contained either the wild-type splice
sites, stronger 5’ and 3’ splice sites (5’ mutants and 3’ mutants, respectively), and
the combination of both (double mutants).
(b) RT-PCR results. Splicing of the intron of interest can be rescued by the intro-
duction of stronger 5’ and 3’ splice sites, as well as by the combination of both.
Most importantly, splicing recovery is obtained independently of the condition
studied (i.e. control and PRPF8 KD). The same results can be observed in c.
(c) qRT-PCR results. Mean + standard error are shown for triplicate reactions of
each experiment.
Figure provided by Dr. Vi Wickramasinghe.
103
Chapter 4. The regulation of splicing by core spliceosomal factors
0.2 0.3 0.4 0.5 0.6
RI
NRI
% GC content
differential GC content
RI
NRI
0.6 0.8 1.0 1.2
(a)
(b)
Figure 4.8| GC content of top retained vs. non-retained introns.
(a) Differential GC content. Differential GC content was calculated by dividing the
GC content of each intron to the average of its adjacent exons (see Methods).
Compared to NRI, the GC content in retained ones is closer to that of neighbour-
ing exons (p-value < 2.2 · 1016, Wilcoxon test).
(b) Overall GC content. Top retained introns have slightly higher GC content than
non-retained ones (p-value = 0.0055, Wilcoxon test).
4.2.4 PRPF8 knock-down has an impact on the rate of co-transcriptional
splicing
The above results suggest that the down-regulation of PRPF8 leads to increased
kinetic competition across splice sites; however, they do not provide an explana-
tion for what could be mediating such increase. In this context, an alteration in
the rate of splicing following PRPF8 KD emerges as a plausible cause for such
differences: splicing is known to occur co-transcriptionally for most human genes
[Tilgner et al., 2012], and thus any changes in the co-existence of this reaction
with transcriptional elongation could result in an increased competition across
splice sites.
Given that the analysed RNA-seq libraries are polyA-selected, it is theoretically
possible to obtain information on the kinetics of the splicing reaction from intronic
reads. More specifically, under a scenario of co-transcriptional splicing, introns
located towards the 5’ end of transcripts are more prone to be spliced out, since
they are transcribed first. Conversely, introns located towards the 3’ end are
more likely to be detected in the RNA-seq experiment, given that splicing has
104
Chapter 4. The regulation of splicing by core spliceosomal factors
not yet finalised after the addition of the polyA tail. Altogether, this suggests
that a difference in the number of reads that map to the first vs. last introns of
each given transcript could be used as an indicator of co-transcriptional splicing
(Figure 4.9a). Similarly, differences in such proportion across the two studied
conditions (i.e. control vs. PRPF8 KD) could serve as an indicator of changes in
the dynamics of splicing.
The results obtained following the above analysis strategy revealed that introns
located towards the 3’ end of transcripts are generally expressed at higher lev-
els compared to those situated towards the 5’ end (Figure 4.9b and Figure D.4),
thus suggesting that splicing is co-transcriptional for most genes. Interestingly,
there is a negative correlation between the differences in expression of first vs. last
introns and transcript length, which indicates that co-transcriptional splicing is
more prevalent in longer transcripts (average rS = -0.24, p-value < 5 · 104 in all
samples). On the other hand, such differences are less accentuated after down-
regulation of PRPF8, hence pointing at changes in splicing efficiency (Figure 4.9b;
p-value = 5.34 · 109). Further experimental validation on a subset of the pre-
viously highlighted genes with a prominent role in cell cycle progression showed
that, indeed, splicing is co-transcriptional overall and less efficient after PRPF8 KD
(Figure 4.10). Specifically, splicing recovery was considerably delayed in PRPF8
KDs compared to normal cells for CDC20 and CDC23, and even more extremely,
it could not be detected during the time frame of the experiment for ASPM.
105
Chapter 4. The regulation of splicing by core spliceosomal factors
AAAAA
AAAAA
r =
Fl
Fk
where:
ci = counts for feature i
li = length of feature i
N = library size
i = {k, k+1, l, l-1} ;
ci
li
fi = 
fl
fl-1 N
Fl = 
fk
fk+1 N
Fk =; CO-TRANSCRIPTIONAL SPLICING (r < 1)
POST-TRANSCRIPTIONAL SPLICING (r ≥ 1)
AAAAA
fk+1 fl-1fk fl5’ 3’
0.
0
0.
5
1.
0
1.
5
r
co
-t
ra
ns
cr
ip
tio
na
l s
pl
ic
in
g
control PRPF8 KD
(b)
(a)
Figure 4.9| Co-transcriptional splicing evidence from intronic reads.
(a) Intronic reads contain potential information on the dynamics of splicing. Introns lo-
cated towards the 5’ end of transcripts have a higher probability of being spliced
out under a scenario of co-transcriptional splicing, but this is not the case if
splicing occurs post-transcriptionally. A co-transcriptional splicing ratio was cal-
culated by comparing the intronic coverage of the first vs. last introns in each
transcript. Such coverage was normalised to take into account intron length and
the expression levels of adjacent exons.
(b) Distribution of the co-transcriptional splicing ratio in control vs. PRPF8 KD sam-
ples. Co-transcriptional splicing dominates in control samples as indicated by the
low ratio values. PRPF8 KDs present higher ratios, suggesting disruptions in the
rate of co-transcriptional splicing (p-value = 5.34 · 109). Only genes with one
transcript annotated in Ensembl 66 were considered for this analysis (n = 2,366;
see Methods).
106
Chapter 4. The regulation of splicing by core spliceosomal factors
unspliced
nascent transcript
fully spliced
transcript
partially spliced
transcript
CDC20 CDC23 ASPM
KD
C
Exon10 :: Intron10 (B :: C)
Exon09 :: Intron10 (A :: C)
Exon04 :: Intron04 (B :: C)
Exon03 :: Intron04 (A :: C)
Exon19 :: Intron19 (B :: C)
Exon18 :: Intron19 (A :: C)
(a)
(b)
107
Chapter 4. The regulation of splicing by core spliceosomal factors
Figure 4.10| Experimental validation of the predicted differences in the rate
of co-transcriptional splicing after PRPF8 KD.
(a) Illustration of the experimental design. Rates of transcription and splicing can be
measured in vivo by treating cells with DRB, a reversible kinase inhibitor [Arde-
hali and Lis, 2009; Singh and Padgett, 2009]. This drug targets specifically a
kinase that is involved in the phosphorilation of the carboxy-terminal domain
(CTD) of RNA pol II, thus blocking its entry into productive elongation. Follow-
ing DRB release, newly synthesised transcripts can be detected with the indicated
primers. Specifically, the primer pair BC provides information on the rate of tran-
scription, as it measures the production of pre-mRNAs. On the other hand, the
primer pair AC will anneal to both unspliced and partially spliced transcripts,
but it will fail to generate a PCR product in the former due to the large size of
the enclosed intron. Thus, it can be used as an indicator of the rate of splicing.
Eventually, once the intron that is targeted by primer C is removed, no amplifi-
cation product will be detected.
(b) Results of the DRB experiments for a set of genes with a role in cell division. Follow-
ing treatment of cells with DRB, the recovery of transcription (black) and splicing
(gray) were measured as described in a. In general, recovery of splicing can be
detected previously to the completion of transcriptional recovery, hence suggest-
ing that intron removal occurs in a co-transcriptional fashion. PRPF8 KD cells
manifest less accentuated recovery rates overall, as well as a slower recovery of
splicing. More extremely, for the CDC23 and ASPM genes, such recovery was
largely impaired. For all panels, mean + standard error are shown for triplicate
reactions of each experiment.
Figure provided by Dr. Vi Wickramasinghe.
4.3 Discussion
The present chapter describes a collaborative effort to further characterise the
function of the splicing factor PRPF8 and, more generally, to dissect the role
of core spliceosomal components in regulating the splicing reaction. Similarly
to previous studies that combined RNA interference and imaging techniques
for the large-scale identification of genes required for cell division, the findings
presented here indicate that the expression of this SF is necessary to exit mitosis
[Neumann et al., 2010]. Moreover, they reveal that the down-regulation of other
B-complex components can also lead to a mitotic arrest phenotype, whilst this is
not the case for components of the A and C complexes, which emphasises the
role of the B-complex in initiating the catalysis of intron removal. In spite of this
B-complex specificity, the reported observations show that knocking-down PRPF8
results in the strongest effects, consistent with its location at the centre of the
spliceosome and its proposed role as a master regulator of splicing [Galej et al.,
2013; Grainger and Beggs, 2005]. Moreover, the treatment of PRPF8 KD cells with
108
Chapter 4. The regulation of splicing by core spliceosomal factors
cytosolic extracts from control samples leads to a decrease in the proportion of
cells arrested in mitosis, thus indicating the the observed phenotype is driven
by defects in splicing. These findings are further corroborated by the results
from Sm iCLIP experiments, which point to an overall decrease in spliceosomal
RNA-protein interactions after PRPF8 down-regulation.
Consistent with the above observations, RNA-seq experiments show an overall
increase in intronic expression levels. In addition, analysis of exon usage pat-
terns reveals further differences in alternative splicing, adding up to existing
evidence in indicating that changes in the concentration of core components of
the spliceosome can regulate this process [Saltzman et al., 2011]. Overall, analysis
of the RNA-seq data allowed for the detection of alterations in splicing for a
subset of the expressed protein coding genes (n = 3,388 out of 13,216). Notably,
these genes are enriched for cell division regulators and mitotic factors, hence
linking the observed splicing defects to the mitotic arrest phenotype. Further
exploration of a set of introns especially sensitive to PRPF8 KD revealed that
they present typical characteristics of retained introns, including weaker splice
sites [Sakabe and de Souza, 2007] and lower differential GC content [Amit et al.,
2012]. Particularly, the results from the computational analyses indicate that
these differences only exist at the 5’ splice site, in line with existing evidence
that points to the preferential interaction of PRPF8 with this splicing signal
[Li et al., 2013]. However, experimental validation with mini-gene experiments
showed that, in PRPF8 KD cells, intron removal can be promoted by increasing
both 5’ and 3’ splice site strength. Altogether, these findings suggest that the
splicing alterations detected following a decrease in the concentration of this SF
are likely to arise from the kinetic competition between splice sites, a well known
mechanism of splicing regulation [House and Lynch, 2008; Wang and Burge, 2008].
Finally, analysis of intronic reads in control samples evidenced that slicing occurs
mostly in a co-transcriptional fashion in the studied set of genes. This is consistent
with existing knowledge suggesting that splicing occurs co-transcriptionally for
most human genes [Bentley, 2014], and in particular, it recapitulates previous
observations from Tilgner et al. [2012] that were based on the analysis of junc-
tion reads. Moreover, such finding shows that it is possible to gain biological
knowledge from the intronic data generated in an RNA-seq experiment, often
considered to be noise. Interestingly, co-transcriptional splicing was detected
109
Chapter 4. The regulation of splicing by core spliceosomal factors
to be more prevalent in longer transcripts, where the increased timespan of the
transcription reaction could provide an opportunity for splicing to start prior to
the addition of the polyA tail. Indeed, in yeast, in which the average gene length
is shorter than in humans, the RNA pol II has been detected to pause at the 3’
terminal end of transcripts and downstream of introns, hence providing extra time
for co-transcriptional splicing to occur [Alexander et al., 2010; Carrillo Oesterreich
et al., 2010]. Most importantly, co-transcriptional splicing was detected to be less
prevalent after PRPF8 KD, thus providing an explanation for the detected kinetic
competition amongst splice sites.
Altogether, the findings described in this chapter show that fluctuations in the
concentration of core spliceosomal factors can result in changes in the dynamics
of splicing, which can in turn manifest as changes in splice site choice. Even
though the kinetics of spliceosomal rearrangements had been already identified
to play a role in shaping splicing decisions [Query and Konarska, 2004; Saltzman
et al., 2011; Yu et al., 2008], the interplay between this process and the dynamics
of splicing had not been characterised to date, hence evidencing the novelty of the
described findings.
4.4 Computational methods
All computational analyses have been performed by myself unless otherwise stated. Exper-
imental analyses have been performed by Dr. Vi Wickramasinghe, Dr. David Perera and
Arthur Bartolozzi. Dr. Christopher W. Sibley carried out the Sm iCLIP experiments and
part of the derived data analysis. Given the computational focus of this thesis, details on
the experimental methods followed have been covered elsewhere [Wickramasinghe et al.].
Analysis of the Sm iCLIP data
Analyses performed by Dr. Christopher W. Sibley:
Sm iCLIP datasets for 3 control and 3 PRPF8 KD samples were generated
following the protocol described in Briese et al. Each biological replicate was
sequenced twice, hence resulting in a total of 12 samples. The obtained reads were
mapped to the human genome and analysed with the iCount software1, following
the steps detailed in König et al. [2010]. bed files that contain information on
1http://icount.biolab.si/
110
Chapter 4. The regulation of splicing by core spliceosomal factors
spliceosome occupancy along the genome were generated as a result, and further
used for downstream analysis.
Analyses performed by myself:
iCLIP counts were normalised by library size to account for variations in sequenc-
ing depth across samples. Similarly, for each junction, counts were divided by the
maximum value to allow for the comparison across different features. Normalised
counts were eventually used to produce Figure 4.3.
Datasets and mapping
RNA from control and PRPF8 KD Cal51 cell lines was extracted, polyA-selected
and prepared for sequencing following standard Illumina protocols. Sequencing
was performed on a HiSeq 2000 machine (100 bp paired-end reads), including
3 and 4 biological replicates for control and PRPF8 KD samples, respectively
(Table A.3). On average, 25M reads per sample were obtained.
Due to the high quality of the reads, raw data were directly mapped to the human
genome (Ensembl 66 [Flicek et al., 2012]) using TopHat v1.3.3 [Trapnell et al.,
2009] with the following options: --max-multihits 1 --no-novel-juncs --min-
isoform-fraction 0.0 --GTF $gtf_file.
Intron counts and normalised intron expression
Intron coordinates were obtained with custom scripts based on bedtools v2.17.0
[Quinlan and Hall, 2010], by considering exons from transcripts annotated as pro-
tein coding in Ensembl 66. Overlapping exons were merged and the longest exon
or combination of exons was kept. Intronic expression levels were then obtained
with dexseq-count (DEXSeq v1.7.0 [Anders et al., 2012]) and normalised as fol-
lows:
111
Chapter 4. The regulation of splicing by core spliceosomal factors
µˆi =
ci/li
cg/lg
· 10
9
N
where:
µˆi = normalised expression for intron i
ci = counts for intron i
li = length for intron i
cg = counts for gene g
lg = length for gene g
N = library size
Identification of differentially used exons
DEXSeq v1.7.0 was used to identify genes with differential exon usage across the
two studied conditions. Specific options include --aggregate=no for the prepa-
ration of the annotation and --paired=yes --stranded=no to count reads that
overlap exons. An FDR threshold of 0.01 was used to assess the significance of the
detected fold-changes (Benjamini & Hochberg p-value correction [Benjamini and
Hochberg, 1995]).
Identification of retained introns
Differential intron usage was assessed with DEXSeq v1.7.0 with the same options
as indicated above. Due to the overall low number of intronic reads compared
to exonic ones, library size factors were not inferred from intron counts, but
instead the previously calculated ones were re-used. Differentially used introns
with a positive log2 fold-change (FDR < 0.01, Benjamini & Hochberg p-value
correction) were defined as retained introns (RI), whilst those that did not fulfil
this criteria were classified as non-retained (NRI). The most extreme intron
retention events (i.e. top 200) were selected for subsequent analyses based on
fold-change information.
Splice site strength was calculated with the MaxEntScan algorithm [Yeo and Burge,
2004], and motif enrichment analysis was performed with Homer v2 [Heinz et al.,
2010] using a common set of background sequences. The calculation of differential
GC content in introns vs. adjacent exons was performed similarly to Amit et al.
[2012], by discarding 20 nucleotides (nt) from each end of the introns and 3 nt
from the exons in order to account for splicing signals.
112
Chapter 4. The regulation of splicing by core spliceosomal factors
Gene and transcript-level analyses
A gene was defined as expressed if it is detected above 1 FPKM in any given sam-
ple. Gene Ontology enrichment analyses were performed with DAVID [Huang
et al., 2009a,b] and WebGestalt [Wang et al., 2013], using an adjusted p-value
threshold of 0.01 (Benjamini & Yekutieli correction [Benjamini and Yekutieli,
2001]) and the set of expressed genes as a background (n = 13,216). Differential
gene expression was assessed with DESeq v1.10.1 [Anders and Huber, 2010],
and significant changes were evaluated following the criteria previously used for
the analysis of differential splicing (FDR < 0.01, Benjamini & Hochberg correction).
Transcript expression estimates were calculated with MISO v0.4.1 [Katz et al.,
2010], using the annotation from Ensembl 66.
Co-transcriptional splicing ratio
Co-transcriptional splicing was evaluated by comparing intronic expression levels
for the first vs. last introns of each transcript. Given the potential complexity intro-
duced by alternative splicing events, this analysis was limited to genes which had
only one transcript annotated in Ensembl 66. Genes with only one intron, those
shorter than 300 bp and those that overlap with other genes in either strand were
also discarded. Altogether, such filtering led to a final set of 2,366 genes (i.e. tran-
scripts) that could be considered in the analysis. Following the identification of
the first and last introns of each transcript in the study set, a co-transcriptional
splicing ratio was calculated using the formula detailed in Figure 4.9a.
113

Chapter 5
The impact of splicing at the
protein level
While mRNAs contain the necessary information for the synthesis of proteins,
their mere existence does not ensure the production of functional protein
products. Hence, many studies have focused on establishing links between the
transcriptome and the proteome, with most of the emphasis placed at the gene
level (e.g. Fu et al. [2009]). In the present chapter, I investigate the possibility
of recapitulating differential splicing events at the protein level, by referring to
proteomics data for the same control and PRPF8 knock-down samples as studied
in Chapter 4. As a first step, I assess the feasibility of such integration by focusing
on the most extreme changes in splicing. I then proceed to evaluate the correlation
between the fold-change estimates obtained from the RNA-seq and proteomics
experiments, whilst expanding the analysis to a larger set of genes. Finally, I
investigate whether such correlation can be improved by incorporating extra
information on transcript relative abundances from the RNA-seq experiment,
and I explore the extent to which differences in gene expression levels could be
influencing the obtained results.
The findings described here constitute the first results from an ongoing collab-
orative project. All the computational analyses have been performed by myself,
under the supervision of Dr. John Marioni from EMBL-EBI. All the experimental
work has been carried out by Dr. Vi Wickramasinghe from the Hutchison Re-
search Institute, and the proteomics data has been generated and analysed by Dr.
Yansheng Liu from ETH Zürich.
115
Chapter 5. The impact of splicing at the protein level
Publications derived from this chapter
• This is an ongoing project.
116
Chapter 5. The impact of splicing at the protein level
5.1 Introduction
Establishing a correlation between the transcriptome and the proteome is not
straightforward, with initial genome-wide estimates of ⇠40% correlation obtained
from the quantification of gene expression levels with microarrays [Maier et al.,
2009; de Sousa Abreu et al., 2009]. Following the introduction of RNA-seq, this
technology was proven to correlate better with protein levels [Fu et al., 2009],
leading to higher estimates of 56-64% correlation (Li et al. [2014]; Lundberg et al.
[2010]; Nagaraj et al. [2011]). Nonetheless, these findings still imply that 40% of
the variation in protein abundance cannot be attributed to mRNA expression,
and both biological processes (i.e. post-transcriptional regulation, translational
efficiency, mRNA and protein turnover...) and limitations inherent to the tech-
nologies used have been proposed as an explanation for such observation [Haider
and Pal, 2013; Lundberg et al., 2010].
On the other hand, when taking into account the diversity of the transcriptome in
terms of alternatively spliced transcripts, the reconciliation of mRNA and protein
levels becomes even more difficult. There have been several efforts to detect
proteins derived from alternatively spliced transcripts, which rely primarily on
the detection of peptides that uniquely support annotated isoforms (e.g. Blakeley
et al. [2010]; Ezkurdia et al. [2012]; Leoni et al. [2011]) or novel exon-exon
junctions (e.g. Xing et al. [2011]; Zhou et al. [2010]). Further approaches are
based on the incorporation of expression data, including evidence from RNA-seq
experiments, to prioritise the set of junctions to consider when interpreting the
proteomics datasets and hence reduce mapping noise [Gloria et al., 2012; Ning
and Nesvizhskii, 2010; Sheynkman et al., 2013; Tanner et al., 2007]. However, in
spite of these attempts, understanding the prevalence of alternative splicing at
the protein level in a genome-wide scale still remains a challenge. Similarly, it
is not easy to predict the extent to which the detected differences in alternative
splicing across two given conditions might ultimately lead to differences in
protein expression levels.
In the present chapter, I describe a collaborative effort aimed at assessing the
impact of differential splicing events at the protein level. This is achieved by
integrating fold-changes obtained by RNA-seq and SWATH-MS, a new proteomics
method that enables the quantification of a large fraction of the proteome with
117
Chapter 5. The impact of splicing at the protein level
high accuracy [Liu et al., 2013]. I introduce here a novel pipeline for the integration
of these two types of data, which relies not only on the information from peptides
that map uniquely to one transcript isoform, but also on that from the ones that
map ubiquitously to several transcripts of the same gene. This is achieved by
using knowledge from the RNA-seq experiment in order to guide the peptide
assignments, and enables to expand the set of usable data. By executing such
pipeline on the PRPF8 knock-down vs. control samples studied in Chapter 4, I
am able to show that a large fraction of the most extreme changes in splicing can
be recapitulated at the protein level (i.e. 9 out of 11 switch events that overlap
with the proteomics dataset are supported by peptide evidence). Moreover, when
expanding such analysis to the larger set of differentially used transcripts, I predict
a 43% correlation between mRNA and protein fold-changes, and I observe that
such correlation increases up to 78% when focusing on differential splicing events
that affect major transcripts. Finally, I show that the estimated correlation remains
stable after discarding differentially expressed genes, which strongly supports the
idea that protein expression levels can be regulated by changes in transcript usage.
5.2 Results
In order to attempt proteomics validation of the splicing changes predicted from
RNA-seq data, our collaborators generated a SWATH-MS dataset that comprises
both PRPF8 knock-down (KD) and control samples (3 biological replicates in each
condition; CAL51 cell lines), thus matching the conditions studied in Chapter 4.
Specifically, in the previous chapter, mapping to the genome enabled the study of
differential splicing events from an exon-centric perspective, as well as the identi-
fication of cases of intron retention. Here such data are re-analysed in a transcript-
centric fashion to facilitate the integration with the proteomics data (see Methods).
The results from such analyses, together with those obtained from the SWATH-MS
dataset, are summarised in Table 5.1.
118
Chapter 5. The impact of splicing at the protein level
RNA-seq
differential transcript usage 668 transcripts (540 genes)
switch events 142 genes
differential gene expression 1,607 genes
SWATH-MS
initial number 16,779 peptides
filter 1: remove peptides that map to >1 genes 14,648 peptides (2,800 genes)
peptides that map to only one transcript 2,964 peptides (856 genes)
filter 2: subset based on fold-change significance 2,901 peptides (1,275 genes)
peptides that map to only one transcript 537 peptides (297 genes)
Table 5.1| Summary of the results obtained from the RNA-seq and SWATH-
MS datasets. Details on the analysis workflow used have been described in the
Methods section.
5.2.1 Extreme changes in splicing can be detected at the protein level
Amongst the genes detected to undergo differential splicing (n = 540 genes with
differential transcript usage, DTU; see Methods), an interesting set are those cases
in which the identity of major transcripts changes across conditions, previously
introduced in this thesis as switch events (SE). Such events constitute the most
extreme and prevalent changes in splicing, and thus constitute a good set to test
whether changes in transcript isoform levels can be recapitulated at the protein
scale. In this context, peptides that map simultaneously to both major transcripts
of a given switch contain ambiguous information, but those that map to either of
the major transcripts provide valuable evidence for the validation of such event
(Figure 5.1).
119
Chapter 5. The impact of splicing at the protein level
INTEGRATION
RNA-seq
differential transcript usage/
switch event
SWATH-MS
peptide 1
overlap
peptide 2
peptide 3
assignment
ambiguouspeptide 3
peptide 2
peptide 1
consider peptides that are unique to each event
Figure 5.1| Integration of the RNA-seq and SWATH-MS data. The RNA-seq
and SWATH-MS data were first analysed separately as detailed in the Meth-
ods section. Integration of the results of such analyses was then performed
by focusing on the set of peptides that map uniquely to the differential splic-
ing events under consideration (i.e. differential transcript usage/switch events;
all events/events that affect major transcripts). Finally, the correlation between
transcript and peptide fold-changes was evaluated. In this plot, protein coding
transcripts are represented by coloured lines, and red boxes correspond to the
aligned peptides.
Following the logic described above, the integration of the results from the RNA-
seq and SWATH-MS datasets revealed a set of 27 switched genes with enough pep-
tide evidence available to attempt validation (see Methods). In this set, 57% of the
detected peptides were consistent with the predicted SE, and a total of 14 events
could be classified as supported by the proteomics data (Table 5.2). Notably, the
validation rate increased drastically when further filtering the peptide data based
on the significance of the calculated fold-changes: in this scenario, ⇠90% of the
peptides with a significant fold-change supported the predicted SE, and 9 out of
11 events could be successfully recapitulated (Table 5.3 and Figure 5.2). For ex-
ample, for TMPO, a protein involved in the structural organisation of the nuclear
envelope, a SE between two protein coding transcripts was strongly supported by
several peptides (Figure 5.2). Similarly, two different peptides that support each of
the transcripts from the switch were identified for HNRNPK, an hnRNP believed
120
Chapter 5. The impact of splicing at the protein level
to play a role in cell cycle progression (Figure 5.2). In both cases, evidence for
alternative protein isoforms had already been reported [Ezkurdia et al., 2012].
gene name
number of peptides is the
that support that do not support switch event
the switch the switch validated?
TMPO 8 0 Y
EPS8L2 5 0 Y
TSTA3 3 0 Y
NUP98 3 1 Y
FXR1 3 2 Y
ACOT9 3 0 Y
HNRNPK 3 1 Y
PFN2 2 1 Y
C9orf5 1 0 Y
XPNPEP1 1 0 Y
HELLS 1 0 Y
HP1BP3 1 0 Y
MOV10 1 0 Y
CPNE1 1 0 Y
NOP56 2 3 N
UQCRC1 0 6 N
AP1B1 0 1 N
CSNK2A1 0 2 N
GOSR2 0 1 N
ABCB6 0 1 N
COMMD4 0 2 N
BRD4 0 1 N
FAM98B 0 3 N
GMPPB 0 2 N
SSNA1 0 1 N
ASRGL1 2 2 A
PDLIM7 1 1 A
total number 41 31of peptides
Table 5.2| Validation of switch events. For each of the 27 genes with usable pep-
tide evidence, the available peptides are classified as either supportive or contrary
to the observed switch event (SE). Specifically, a given peptide is considered to
support the SE if it suggests an increase in the abundance of the protein isoform
expected in PRPF8 KDs, or a decrease in the one expected in controls. Y: yes; N:
no; A: ambiguous.
121
Chapter 5. The impact of splicing at the protein level
gene name
number of peptides is the
that support that do not support switch event
the switch the switch validated?
TMPO 4 0 Y
TSTA3 3 0 Y
EPS8L2 3 0 Y
HNRNPK 2 0 Y
PFN2 1 0 Y
C9orf5 1 0 Y
HELLS 1 0 Y
ACOT9 1 0 Y
MOV10 1 0 Y
AP1B1 0 1 N
CSNK2A1 0 1 N
total number 17 2of peptides
Table 5.3| Validation of switch events when filtering by peptide fold-change
significance. When subsetting peptides based on their fold-change significance,
the vast majority support the trend marked by the RNA-seq data. None of the
genes listed here are detected as differentially expressed, with the exception of
HELLS. Y: yes; N: no; A: ambiguous.
Figure 5.2| Examples of switch events with supporting peptide evidence.
Fold-changes represent changes in expression levels for PRPF8 KD vs. control
samples (yellow: down-regulated in PRPF8 KD; blue: up-regulated in PRPF8
KD). Barplots feature information from the two transcripts involved in the switch
(RNA-seq), as well as from the overlapping peptides (SWATH-MS; Table 5.3). Ta-
bles include further details on the median and the standard deviation of the es-
timates obtained with each technology (FPKMs for RNA-seq, peptide intensities
for SWATH-MS). For each gene, genomic maps for the highlighted transcripts
are depicted in the bottom panels. C: control; KD: PRPF8 KD.
122
Chapter 5. The impact of splicing at the protein level
TM
PO
TS
TA
3
HN
RN
PK
m
ed
ia
n 
C
SD
 C
m
ed
ia
n 
KD
SD
 K
D
lo
g2
 fo
ld
-
ch
an
ge
56
.5
2
4.
96
14
.3
0
1.
69
-1
.9
8
31
.7
4
2.
18
73
.8
3
2.
88
1.
22
23
66
2.
08
13
79
.5
2
42
78
8.
79
40
32
.1
8
0.
85
10
44
6.
97
15
82
.2
9
20
10
6.
24
31
2.
26
0.
94
68
29
.0
3
71
7.
78
97
87
.0
3
72
3.
80
0.
52
84
93
.1
9
10
36
.1
7
15
53
2.
29
16
82
.8
7
0.
87
lo
g2
 fo
ld
-c
ha
ng
e
ï

ï

ï







RNA-seq SWATH-MS
pr
ot
ei
n 
co
di
ng
 - 
EN
ST
00
00
05
56
02
9
pr
ot
ei
n 
co
di
ng
 - 
EN
ST
00
00
02
66
73
2
to
ta
l a
nn
ot
at
ed
 tr
an
sc
rip
ts
: 1
3
m
ed
ia
n 
C
SD
 C
m
ed
ia
n 
KD
SD
 K
D
lo
g2
 fo
ld
-
ch
an
ge
30
.8
8
0.
38
4.
79
2.
45
-2
.6
4
1.
19
0.
43
6.
7
1.
14
2.
50
86
43
.5
2
45
5.
58
44
26
.6
1
38
2.
21
-0
.9
7
28
00
9.
28
19
23
.3
2
17
18
2.
17
13
00
.0
8
-0
.7
1
77
46
.5
13
56
.6
7
56
77
.6
8
12
2.
86
-0
.4
5
ï
ï



lo
g2
 fo
ld
-c
ha
ng
e
RNA-seq SWATH-MS
RNA-seq SWATH-MS
pr
ot
ei
n 
co
di
ng
 - 
EN
ST
00
00
04
25
75
3
pr
ot
ei
n 
co
di
ng
 - 
EN
ST
00
00
05
24
71
9
to
ta
l a
nn
ot
at
ed
 tr
an
sc
rip
ts
: 1
6
EP
S8
L2

ï
ï
ï

lo
g2
 fo
ld
-c
ha
ng
e
17
.3
6
1.
36
0.
11
0.
36
-7
.2
4
1.
47
1.
53
9.
75
1.
62
2.
73
18
28
0.
15
20
55
.4
6
12
52
3.
59
12
81
.4
5
-0
.5
5
15
72
5.
20
32
5.
63
92
13
.4
8
64
4.
50
-0
.7
7
26
49
6.
55
11
20
.9
4
13
87
2.
36
17
6.
14
-0
.9
3
RNA-seq SWATH-MS
m
ed
ia
n 
C
SD
 C
m
ed
ia
n 
KD
SD
 K
D
lo
g2
 fo
ld
-
ch
an
ge
pr
ot
ei
n 
co
di
ng
 - 
EN
ST
00
00
03
18
56
2
pr
ot
ei
n 
co
di
ng
 - 
EN
ST
00
00
05
31
34
8
to
ta
l a
nn
ot
at
ed
 tr
an
sc
rip
ts
: 3
2
23
5.
19
4.
57
10
9.
59
4.
52
-1
.1
0
13
3.
63
5.
25
15
3.
65
1.
70
0.
20
27
52
17
.9
9
11
16
1.
66
20
91
73
.7
0
50
59
.5
4
-0
.4
0
39
54
8.
98
42
62
.1
7
64
74
0.
16
84
25
.9
3
0.
71
ï

ï





lo
g2
 fo
ld
-c
ha
ng
e
m
ed
ia
n 
C
SD
 C
m
ed
ia
n 
KD
SD
 K
D
lo
g2
 fo
ld
-
ch
an
ge
pr
ot
ei
n 
co
di
ng
 - 
EN
ST
00
00
03
60
38
4
pr
ot
ei
n 
co
di
ng
 - 
EN
ST
00
00
04
57
15
6
to
ta
l a
nn
ot
at
ed
 tr
an
sc
rip
ts
: 1
5
123
Chapter 5. The impact of splicing at the protein level
5.2.2 RNA-seq fold-change estimates correlate with those obtained
from SWATH-MS
Next, I investigated whether the integration strategy introduced in the previous
section can be extended to the general validation of DTU events. More specifically,
I evaluated the impact of the predicted changes in splicing at the protein level
on the basis of the correlation between transcript and peptide fold-changes. In
this context, a positive correlation would indicate that the predicted changes in
splicing contribute to determining the detected fluctuations in protein levels,
whilst a lack of correlation would support the contrary.
Quantitative integration of the fold-changes from the two technologies used in
this chapter can be easily achieved by focusing on the peptides that map to only
one transcript. Given the size of the proteomics dataset analysed here, it was
possible to identify a small number of peptides that map uniquely to transcripts
involved in DTU events (n = 48 peptides from 21 genes), from which a Spear-
man correlation coefficient of 0.41 was obtained when comparing fold-changes
(p-value = 3.67·103; Table 5.4). Further focusing on peptides with a significant
fold-change led to a bigger decrease in the amount of available information (n =
15 peptides from 6 genes), which was not enough to obtain a correlation estimate.
Hence, it is evident that even though uniquely mapping peptides constitute a
valuable source of information, they represent only a minority of the cases, and
therefore most of the data are lost when following such integration approach.
Alternatively, I evaluated a novel analysis strategy that enables incorporating in-
formation from the larger set of ubiquitous peptides. Briefly, such strategy re-
lies on the use of the fold-change information from those peptides that overlap
uniquely with each given DTU event (i.e. peptides that support more than one
event are not considered for analysis, similarly to the approach followed in the
previous section for the validation of SE; see Figure 5.1). This is based on the as-
sumption that changes in protein expression levels will be preceded by a change
in transcript expression levels. Implementation of such analysis approach resulted
in a Spearman correlation coefficient of 0.31 when taking into account all the avail-
able peptides (n = 336 from 80 genes; p-value = 9.45·109; Table 5.4), and a corre-
lation of 0.43 when focusing on those with a significant fold-change (n = 101 from
40 genes; p-value = 9.04·106; Table 5.4).
124
Chapter 5. The impact of splicing at the protein level
5.2.3 The correlation estimates can be improved by incorporating infor-
mation on major transcripts
Given the observation that the correlation estimates improve when filtering out
peptides with non-significant fold-changes, I next evaluated whether similar ef-
fects could be obtained by adopting an analogous approach for the RNA-seq data.
Such strategy consisted on focusing on those DTU events that involve major tran-
scripts, and discarding those that affect lowly abundant ones. Notably, this re-
sulted in an increase of the observed correlation coefficients, with a value of 0.48
when taking into account all the peptides (n = 270 peptides from 54 genes; p-value
< 2.20·1016; Table 5.4) and 0.78 when focusing on peptides with a significant fold-
change (n = 66 peptides from 27 genes; p-value = 1.47·1014; Table 5.4).
all DTU events
number number rS p-valueof peptides of genes
peptides that map 48 21 0.41 3.67E-03to only one transcript
all peptides 336 80 0.31 9.45E-09
subset based 101 40 0.43 9.04E-06on fold-change significance
DTU events in major transcripts
number number rS p-valueof peptides of genes
all peptides 270 54 0.48 <2.20E-16
subset based 66 27 0.78 1.47E-14on fold-change significance
Table 5.4| Validation of differential transcript usage events. Peptide fold-
changes (FCs) were correlated with RNA-seq FCs for those transcripts involved
in differential transcript usage events (DTU; see Figure E.1 for the correspond-
ing scatter plots). Two different integration strategies are evaluated here: ei-
ther taking into account all the DTU events (top) or only those that affect major
transcripts (bottom). Incorporating information on major transcripts leads to an
increase in the observed correlation coefficients.
125
Chapter 5. The impact of splicing at the protein level
5.2.4 The detected correlation is not driven by differences in gene ex-
pression levels
The results obtained in the previous sections indicate that there is a correlation be-
tween the fold-changes predicted by RNA-seq and those obtained by SWATH-MS,
thus supporting the hypothesis that the detected changes in protein abundance
could arise from changes in transcript usage. Nonetheless, DTU is not isolated
from differential gene expression, and it is possible that the observed agreement
is primarily driven by overall differences in gene expression levels. In order to
dissect the impact of such events, I repeated the above analyses by focusing on the
DTU events from those genes that are not differentially expressed (n = 457 genes;
see Methods). In general, the results obtained are consistent with those previously
reported, with comparable correlation coefficients (see Table 5.4). Similarly, only
one of the genes predicted to undergo a SE was also detected as differentially ex-
pressed (i.e. HELLS; see Table 5.3). Overall, these results indicate that differences
in gene expression levels are not the main cause behind the observed correlation.
all DTU events (excluding DGE)
number number rS p-valueof peptides of genes
peptides that map 47 20 0.40 5.81E-03to only one transcript
all peptides 283 70 0.37 1.24E-10
subset based 84 35 0.55 6.85E-08on fold-change significance
DTU events in major transcripts (excluding DGE)
number number rS p-valueof peptides of genes
all peptides 245 47 0.48 7.37E-16
subset based 61 24 0.80 1.13E-14on fold-change significance
Table 5.5| Validation of differential transcript usage events after excluding
differentially expressed genes. The analysis strategy followed is analogous to
that described in Table 5.3, and the corresponding scatter plots can be found in
Figure E.2. The observed correlation is not biased by changes in gene expression
levels.
126
Chapter 5. The impact of splicing at the protein level
5.3 Discussion
In the present chapter, I describe a novel analysis approach for the validation of
differential splicing events at the protein level. Importantly, the findings reported
here rely on the use of SWATH-MS, a novel proteomics method that has not yet
been applied to the study of protein isoforms [Gillet et al., 2012]. Specifically, and
unlike targeted proteomics [Domon and Aebersold, 2010], peptide quantification
with SWATH-MS is performed in a data-independent fashion, hence enabling
the detection of a larger number of peptides. Furthermore, this technology also
offers higher reproducibility rates than traditional shotgun methods [Domon and
Aebersold, 2010], as demonstrated by the fact that SWATH-MS results correlate
closely with those obtained from targeted proteomics experiments (i.e. 98%
correlation; Liu et al. [2013]).
Previous attempts to integrate transcriptomics and proteomics information at
the transcript level have focused on the detection of protein isoforms, rather
than in the validation of predicted changes in splicing (e.g. Blakeley et al. [2010];
Ezkurdia et al. [2012]; Leoni et al. [2011]). Such integration has been limited by
the use of uniquely mapping peptides, and further studies have relied on the
existence of a matched RNA-seq dataset to refine the identification of the set of
expressed proteins in a given sample (e.g. Sheynkman et al. [2013]). The approach
introduced in this chapter relies on a similar strategy; however it differs from
the aforementioned studies not only in the main goal of the analysis, but also in
the amount of usable peptide information. By assuming that changes in protein
expression are more likely to arise from those transcripts for which mRNA levels
have already been detected to fluctuate, I show here that it is possible to retain
information from ubiquitous peptides, i.e. those that map to more than one tran-
script within a given gene. Following filtering of instances of differential splicing
predicted from the RNA-seq data based on peptide availability, it was possible
to validate a large proportion of switch events (i.e. 9 out of 11 events), hence
indicating that extreme changes in splicing can be recapitulated at the protein
level. Subsequently, the application of this analysis strategy to the entire set of
differentially used transcripts led to a 43% correlation between the RNA-seq and
SWATH-MS fold-changes. Further prioritisation of peptide assignments based on
information about major transcripts resulted in an improved correlation of 78%,
which exceeds the one reported in previous studies [Li et al., 2014; Lundberg
127
Chapter 5. The impact of splicing at the protein level
et al., 2010; Nagaraj et al., 2011]. Altogether, these findings evidence that the
transcript relative abundances derived from RNA-seq experiments can assist
in the interpretation of proteomics data, and support the idea that differential
splicing events in minor transcripts correspond to subtle changes that do not have
a strong impact on protein levels.
Reasons for the lack of correlation amongst the observed mRNA and protein
levels most likely include the effects of a variety of cellular processes that are not
taken into account in the experiments performed here (e.g. post-transcriptional
regulation, translational efficiency, mRNA and protein turnover...), as well as limi-
tations inherent to the technologies used. Regarding the former, it is possible that
the importance of the previously mentioned processes has been underestimated,
and that transcription has a less decisive role in shaping mRNA expression levels
than expected. However, recent observations by Li et al. [2014] strongly contradict
this scenario, suggesting that mRNA levels could explain ⇠81% of the variance in
protein levels, and indicating that further work is needed to improve the current
understanding of the remaining sources of variation that might affect protein
levels. On the other hand, proteomics studies show that observing a protein is
unlikely unless there are at least a certain number of RNA molecules per cell
[Ramakrishnan et al., 2009]. In this context, it is difficult to distinguish between
transcriptional noise and insufficient sensitivity of the methods used, and it is
expected that further technological advances will result in a higher throughput,
lower detection limits, and better correlation estimates. Finally, protein folding,
localisation and post-translational modifications all have prominent roles in
establishing the function of proteins, which cannot be assessed from their mere
detection. Thus, trying to predict which changes in mRNA levels will eventually
have a functional impact vs. what proportion can be attributed to transcriptional
and splicing noise constitutes a much more complicated question.
Overall, even though the results described here are mostly exploratory, they show
that it is possible to attempt validation of the changes in splicing predicted from
RNA-seq data by relying on information from SWATH-MS experiments. Future
steps of this ongoing project will include the automation of the described analysis
workflow, and the development of a testing approach to expand the validation
of differential splicing beyond switch events. Eventually, all this information will
be used to revisit the biological insights described in Chapter 4, and to further
128
Chapter 5. The impact of splicing at the protein level
understand the consequences of knocking-down PRPF8. In the longer term, the
described strategy can be extended in order to take into consideration not only
the information derived from major transcripts, but that from all the expressed
transcripts in a given sample. This would allow for a more accurate evaluation
of the true extent of alternative splicing at the protein level, and would help in
assessing the coding potential of minor and novel transcripts. However, such
advantages would come at the cost of increasing the complexity of peptide assign-
ment. This could be compensated by having access to larger amounts of data, and
it is expected that future technological advances will allow for an even more high
throughput integration.
5.4 Computational methods
All the computational analyses described in this chapter have been carried out by myself,
unless otherwise indicated in the text. The experimental work has been performed by Dr. Vi
Wickramasinghe and Dr. Yansheng Liu.
Analysis of the RNA-seq data
The RNA-seq dataset used here corresponds to the one analysed in the previous
chapter of the thesis, and includes 3 control samples and 4 PRPF8 KDs (CAL51
cells; see Chapter 4 - Methods and Table A.3). Raw reads were directly mapped
to the transcriptome with Bowtie v0.12.7 [Langmead et al., 2009], using Ensembl
v66 as a reference [Flicek et al., 2012]. Following the estimation of transcript
expression levels with MMSEQ v1.0.7 [Turro et al., 2011], its companion tool
MMDIFF [Turro et al., 2014] was used to identify both differentially expressed
genes (n = 1,607 genes; Table 5.1) and differentially used transcripts (n = 668
transcripts from 540 genes; Table 5.1). A posterior probability of 0.9 was used
as the significance threshold in both analyses. Switch events within the set of
genes identified to undergo differential transcript usage were identified with
SwitchSeq (Gonzàlez-Porta and Brazma [2014]; n = 150 genes). Switch events that
involved major transcripts with identical protein sequences were removed from
the analyses (n = 8; final set of switched genes = 142; Table 5.1).
RNA-seq fold-changes were calculated from the transcript-level expression esti-
mates. For each transcript, the fold-change represents the median transcript ex-
129
Chapter 5. The impact of splicing at the protein level
pression in PRPF8 KD vs. control samples.
Analysis of the SWATH-MS data
Protein expression levels were assessed with SWATH-MS [Liu et al., 2013] for
a set of 3 control and 3 PRPF8 KD samples. An initial set of 16,779 peptides
were detected and mapped against all the protein coding transcripts annotated
in Ensembl v66, including those with a nonsense-mediated decay biotype.
Removal of peptides that mapped to more than one gene led to a set of 14,648
peptides (corresponding to 2,800 genes), which was used for downstream analysis
(Table 5.1).
Raw peptide intensities were first quantile-normalised in order to enable compar-
ison across samples. For each peptide, the observed intensities across the bio-
logical replicates in each condition were summarised by using the median, and a
fold-change was obtained by dividing the value obtained for PRPF8 KD vs. control
samples. Fold-change significance was assessed with a t-test, and a p-value of 0.1
was used as the significance threshold.
Integration of RNA-seq and SWATH-MS fold-change estimates
The fold-changes derived from these two technologies were integrated as de-
scribed in Figure 5.1. Spearman correlation was used to evaluate the relationship
between transcript and peptide fold-changes, as suggested by Maier et al. [2009].
Information on gene functions was retrieved from Genecards [Safran et al., 2010],
unless otherwise indicated in the text.
130
Chapter 6
Conclusions
Splicing is only one of the many processes that shape the final set of RNA
molecules present in eukaryote cells; nonetheless, it emerges as the most promi-
nent mechanism for message diversification. Recently, the introduction of RNA
sequencing has facilitated the study of this mechanism, and has allowed for a thor-
ough characterisation of transcriptome composition. In this thesis, I have focused
on the application of this technology to the study of human transcript diversity, as
well as its potential impact on the protein repertoire.
Most protein coding genes express one dominant transcript
in a given condition
High throughput sequencing technologies have contributed significantly to the
efforts for the identification and annotation of expressed loci within the human
genome, as illustrated by the ENCODE and GENCODE projects, respectively [EN-
CODE Project Consortium et al., 2012; Harrow et al., 2012]. These have resulted
in high quality annotations, characterised by the existence of a diversity of gene
types and transcript isoforms. However, despite these efforts, the contribution of
each of the annotated transcripts to the overall transcriptome diversity in a given
sample had remained largely uncharacterised. The findings derived from Chapter
2 provide a first answer to this question, by showing that gene expression tends
to be dominated by one transcript in a given condition. In the future, evaluation
of major transcript expression patterns across conditions and species will not only
allow for the characterisation of splicing programs, but might also contribute to
improving the existing annotation. For example, information about recurrent ma-
131
Chapter 6. Conclusions
jor and minor transcripts could be displayed along with transcript structures in
order to inform about the prevalence of the annotated features, and major tran-
script dominance could be used to identify sets of genes that are prone to express
several transcripts simultaneously. Furthermore, assessing the impact of unnan-
otated features will be of particular interest: even though novel transcripts have
been reported to be expressed at lower levels (e.g. Djebali et al. [2012]), they might
still represent the most abundant transcript in a specific set of genes. Eventually, a
bigger challenge will be to understand the effect of major and minor transcripts at
the protein level, and to evaluate possible functional differences among alternative
transcripts. Integrative analysis of RNA-seq and proteomics data, together with a
better understanding of the structural and evolutionary properties that might dif-
ferentiate a given set of transcripts, will be essential towards accomplishing these
goals.
Alterations in splicing are widespread in kidney cancer, even
though extreme changes are rare
The applications of RNA sequencing go far beyond the improvement of the ex-
isting annotation. In a clinical context, such technology opens the door for an
unprecedented characterisation of the changes in transcriptome composition that
might govern a specific disease, offering new possibilities for biomarker discov-
ery and drug design [Costa et al., 2013]. For example, the findings reported in
Chapter 3 show that the splicing process is broadly altered in ccRCC; however,
analysis of major transcript expression patterns evidences the subtlety of most of
the detected alterations. In the future, it will be interesting to investigate whether
similar results are observed for other cancer types. Specifically, identification of
sets of genes that commonly undergo extreme changes in splicing without being
differentially expressed is likely to result in novel findings, since these effects have
been largely ignored in the past. Consistency in the detected switch events across
tumour types could provide extra evidence for a functional role of the involved
transcripts, whilst high variability in major transcripts would be supportive of a
general dysregulation of splicing programs. Lastly, given enough sample size,
exploration of the relationship between switch events and mutation status at the
patient level might provide evidence for the existence of cis-regulated events, and
correlation of alterations in splicing with clinical variables might prove useful for
132
Chapter 6. Conclusions
biomarker discovery and tumour stratification.
PRPF8 abundance dictates the patterns of alternative and
constitutive mRNA splicing
The impact of alternative transcript products in disease can be further understood
by gaining a deeper understanding of the role of the many proteins involved in
the splicing reaction. In Chapter 4, I show that fluctuations in the concentration
of PRPF8 can result in alterations of the rate of splicing, which in turn can cause
changes in splice site choice and lead to alternative splicing events. Even though
the chain of events investigated in this chapter corresponds to a somehow artificial
scenario, changes in the concentration and function of spliceosomal components
are known to occur in development and cell cycle progression [Lane et al., 2013;
Park et al., 2004], and have also been reported across tissues and in disease con-
texts [Grosso et al., 2008; Matera and Wang, 2014]. PRPF8 mutations are found in
autosomal dominant retinitis pigmentosa [Tanackovic et al., 2011], and it remains
to be seen whether those also affect the kinetic competition equilibrium between
weak and strong splice sites, just as observed when this gene is downregulated.
Ultimately, knowledge on the mechanistic basis of the splicing reaction is essen-
tial to predict the effects of perturbations of the system, an exercise regarded as
deciphering the splicing code [Barash et al., 2010].
Extreme changes in splicing can be detected at the protein
level
Proteins derived from alternative splicing products have been detected in a vari-
ety of studies and contexts (e.g. Ezkurdia et al. [2012]; Leoni et al. [2011]). The
main challenge in these analyses has been the limited number of peptides that
can be used to uniquely identify relevant proteins, and past studies have relied
on the use of RNA-seq datasets to create sample-specific databases that would
optimise peptide assignment (e.g. Sheynkman et al. [2013]). In Chapter 5, I eval-
uated a similar approach with the goal of validating differential splicing events
across two conditions of the same experiment. The findings derived from the
first exploratory analyses indicate that it is possible to retain information from
ubiquitously mapping peptides, by incorporating previous knowledge on major
133
Chapter 6. Conclusions
transcripts. Moreover, they suggest that the changes in splicing detected at the
transcriptomic level are likely to have an effect in the protein repertoire. In the
future, expanding the analysis workflow to consider the abundances of all the ex-
pressed transcripts in a given sample would allow for the inspection of the coding
potential of minor transcripts and unnanotated isoforms. Altogether, the knowl-
edge derived from such integrative analyses will ultimately help in understanding
the impact of alternative splicing at the protein level.
Concluding remarks
Previous to the establishment of RNA sequencing, microarrays constituted the
only option for the genome-wide analysis of expression levels. Despite not
being initially intended for the study of alternative transcripts, adequate probe
design strategies allowed for the detection of alternative splicing events [Lee and
Roy, 2004]. However, compared to sequencing, such an approach offers more
limited throughput and a lower dynamic range in terms of expression levels of
the targeted molecules [Zhao et al., 2014]. On the other hand, there are also
challenges linked to the use of RNA sequencing for the characterisation of the
transcriptome, the main ones associated with the short read length [Hooper,
2014]. In this context, transcript reconstruction remains a difficult, yet not
impossible task. While exon-centric analysis approaches have been devised in
order to circumvent the existing limitations, further technological developments
will be crucial to overcome them. Indeed, during the timespan of this PhD,
significant advances have already been achieved in terms of read length, evolving
from 50 nucleotides a few years ago to more than 150 bases nowadays. The past
years have also seen the rise of Pacific Biosciences and nanopore sequencing as
alternatives to Illumina platforms for the study of full-length transcripts [Sharon
et al., 2013]. Hence, it is a matter of time for transcript-level analyses to become
central to any transcriptomics workflow.
Overall, RNA sequencing is only one of the many applications of high throughput
sequencing. Novel techniques such as single cell sequencing are being established,
and the routine application of this technology in the clinical sector is becoming
closer to a reality. Furthermore, the rapid decrease in the costs of sequencing,
together with the continuous increase in the amount of data obtained from each
run, will contribute even further to the democratisation of sequencing [Shendure
134
Chapter 6. Conclusions
and Ji, 2008]. Altogether, these facts evidence that the era of data production
is just at its beginning, and that high throughput sequencing will continue to
revolutionise many fields of basic and applied research.
135

Full list of publications
Gonzàlez-Porta, M., Calvo, M., Sammeth, M. & Guigó, R. Estimation of alternative
splicing variability in human populations. Genome Research 22, 528–38 (2012).
Roux, J., Gonzàlez-Porta, M. & Robinson-Rechavi, M. Comparative analysis
of human and mouse expression data illuminates tissue-specific evolutionary
patterns of miRNAs. Nucleic Acids Research 40, 5890–900 (2012).
Gonzàlez-Porta, M., Frankish, A., Rung, J., Harrow, J. & Brazma, A. Transcriptome
analysis of human tissues and cell lines reveals one dominant transcript per gene.
Genome Biology 14, R70 (2013).
Lappalainen, T. et al. Transcriptome and genome sequencing uncovers functional
variation in humans. Nature 501, 506–11 (2013).
’t Hoen, P. et al. Reproducibility of high-throughput mRNA and small RNA
sequencing across laboratories. Nature Biotechnology 31, 1015–22 (2013).
Gonzàlez-Porta, M. & Brazma, A. Identification, annotation and visualisation of
extreme changes in splicing from rna-seq experiments with switchseq. bioRxiv
(2014). 10.1101/005967.
(submitted) Scelo, G.?, Riazalhosseini, Y.?, Greger, L.?, Letourneau, L.?, Gonzàlez-
Porta, M.? et al. Whole-genome sequencing reveals variation in the genomic
landscape of clear cell Renal Cell Carcinoma in Europe.
(submitted) Wickramasinghe, V.?, Gonzàlez-Porta, M.? et al. PRPF8 abundance
dictates the patterns of alternative and constitutive messenger RNA splicing.
(submitted) Ferreira, P.?, Monlong, J.?, Gonzàlez-Porta, M.?, Barann, M.? et al.
Small changes in the big picture: genetic fine-tuning of transcriptome variation
?shared first authors
137
across human populations.
(submitted) Ballester, B. et al. Multi-species, multi-transcription factor binding
atlas highlights the conserved control of tissue-specific pathways.
(in preparation) Diaz-Muñoz, M.D. et al. HuR-dependent regulation of mRNA
splicing is essential for B cell antibody response.
138
Appendix A
RNA-seq datasets used in the
thesis
139
Ta
bl
e
A
.1
|
R
N
A
-s
eq
da
ta
us
ed
in
C
ha
pt
er
2.
A
ll
sa
m
pl
es
ar
e
po
ly
A
-s
el
ec
te
d.
Pa
ir
ed
re
ad
s
ar
e
co
un
te
d
on
ce
.
A
rr
ay
Ex
pr
es
s
EN
A
EN
A
So
ur
ce
Sa
m
pl
e
Se
qu
en
ci
ng
pl
at
fo
rm
R
ea
d
Se
qu
en
ci
ng
de
pt
h
ex
pe
ri
m
en
ti
d
ex
pe
ri
m
en
ti
d
sa
m
pl
e
id
de
sc
ri
pt
io
n
le
ng
th
(b
p)
(M
re
ad
s)
E-
M
TA
B-
51
3
ER
P0
00
54
6
ER
R
03
08
72
Bo
dy
M
ap
2.
0
th
yr
oi
d
Ill
um
in
a
H
iS
eq
20
00
50
x2
81
.9
1
E-
M
TA
B-
51
3
ER
P0
00
54
6
ER
R
03
08
73
Bo
dy
M
ap
2.
0
te
st
es
Ill
um
in
a
H
iS
eq
20
00
50
x2
81
.8
4
E-
M
TA
B-
51
3
ER
P0
00
54
6
ER
R
03
08
74
Bo
dy
M
ap
2.
0
ov
ar
y
Ill
um
in
a
H
iS
eq
20
00
50
x2
80
.9
5
E-
M
TA
B-
51
3
ER
P0
00
54
6
ER
R
03
08
75
Bo
dy
M
ap
2.
0
w
hi
te
bl
oo
d
ce
ll
Ill
um
in
a
H
iS
eq
20
00
50
x2
81
.2
2
E-
M
TA
B-
51
3
ER
P0
00
54
6
ER
R
03
08
76
Bo
dy
M
ap
2.
0
sk
el
et
al
m
us
cl
e
Ill
um
in
a
H
iS
eq
20
00
50
x2
82
.1
1
E-
M
TA
B-
51
3
ER
P0
00
54
6
ER
R
03
08
77
Bo
dy
M
ap
2.
0
pr
os
ta
te
Ill
um
in
a
H
iS
eq
20
00
50
x2
82
.3
3
E-
M
TA
B-
51
3
ER
P0
00
54
6
ER
R
03
08
78
Bo
dy
M
ap
2.
0
ly
m
ph
no
de
Ill
um
in
a
H
iS
eq
20
00
50
x2
82
.0
8
E-
M
TA
B-
51
3
ER
P0
00
54
6
ER
R
03
08
79
Bo
dy
M
ap
2.
0
lu
ng
Ill
um
in
a
H
iS
eq
20
00
50
x2
79
.3
0
E-
M
TA
B-
51
3
ER
P0
00
54
6
ER
R
03
08
80
Bo
dy
M
ap
2.
0
ad
ip
os
e
Ill
um
in
a
H
iS
eq
20
00
50
x2
77
.3
0
E-
M
TA
B-
51
3
ER
P0
00
54
6
ER
R
03
08
81
Bo
dy
M
ap
2.
0
ad
re
na
l
Ill
um
in
a
H
iS
eq
20
00
50
x2
74
.4
7
E-
M
TA
B-
51
3
ER
P0
00
54
6
ER
R
03
08
82
Bo
dy
M
ap
2.
0
br
ai
n
Ill
um
in
a
H
iS
eq
20
00
50
x2
73
.5
1
E-
M
TA
B-
51
3
ER
P0
00
54
6
ER
R
03
08
83
Bo
dy
M
ap
2.
0
br
ea
st
Ill
um
in
a
H
iS
eq
20
00
50
x2
75
.8
6
E-
M
TA
B-
51
3
ER
P0
00
54
6
ER
R
03
08
84
Bo
dy
M
ap
2.
0
co
lo
n
Ill
um
in
a
H
iS
eq
20
00
50
x2
82
.4
4
E-
M
TA
B-
51
3
ER
P0
00
54
6
ER
R
03
08
85
Bo
dy
M
ap
2.
0
ki
nd
ey
Ill
um
in
a
H
iS
eq
20
00
50
x2
80
.4
0
E-
M
TA
B-
51
3
ER
P0
00
54
6
ER
R
03
08
86
Bo
dy
M
ap
2.
0
he
ar
t
Ill
um
in
a
H
iS
eq
20
00
50
x2
82
.9
2
E-
M
TA
B-
51
3
ER
P0
00
54
6
ER
R
03
08
87
Bo
dy
M
ap
2.
0
liv
er
Ill
um
in
a
H
iS
eq
20
00
50
x2
80
.0
5
E-
G
EO
D
-2
62
84
SR
P0
07
46
1
SR
R
30
78
97
EN
C
O
D
E
ce
ll_
gm
12
87
8
Ill
um
in
a
G
A
II
70
x2
32
.2
1
E-
G
EO
D
-2
62
84
SR
P0
07
46
1
SR
R
30
78
98
EN
C
O
D
E
ce
ll_
gm
12
87
8
Ill
um
in
a
G
A
II
70
x2
47
.0
0
E-
G
EO
D
-2
62
84
SR
P0
07
46
1
SR
R
31
53
30
EN
C
O
D
E
ce
ll_
he
la
-s
3
Ill
um
in
a
G
A
II
70
x2
36
.9
4
E-
G
EO
D
-2
62
84
SR
P0
07
46
1
SR
R
31
53
31
EN
C
O
D
E
ce
ll_
he
la
-s
3
Ill
um
in
a
G
A
II
70
x2
37
.4
5
E-
G
EO
D
-2
62
84
SR
P0
07
46
1
SR
R
30
79
26
EN
C
O
D
E
ce
ll_
he
pg
2
Ill
um
in
a
G
A
II
70
x2
41
.1
4
E-
G
EO
D
-2
62
84
SR
P0
07
46
1
SR
R
30
79
27
EN
C
O
D
E
ce
ll_
he
pg
2
Ill
um
in
a
G
A
II
70
x2
41
.1
6
140
E-
G
EO
D
-2
62
84
SR
P0
07
46
1
SR
R
30
79
05
EN
C
O
D
E
ce
ll_
hu
ve
c
Ill
um
in
a
G
A
II
70
x2
28
.6
0
E-
G
EO
D
-2
62
84
SR
P0
07
46
1
SR
R
30
79
06
EN
C
O
D
E
ce
ll_
hu
ve
c
Ill
um
in
a
G
A
II
70
x2
35
.8
7
E-
G
EO
D
-2
62
84
SR
P0
07
46
1
SR
R
31
53
36
EN
C
O
D
E
ce
ll_
k5
62
Ill
um
in
a
G
A
II
70
x2
35
.7
5
E-
G
EO
D
-2
62
84
SR
P0
07
46
1
SR
R
31
53
37
EN
C
O
D
E
ce
ll_
k5
62
Ill
um
in
a
G
A
II
70
x2
38
.6
7
E-
G
EO
D
-2
62
84
SR
P0
07
46
1
SR
R
30
78
99
EN
C
O
D
E
cy
to
so
l_
gm
12
87
8
Ill
um
in
a
G
A
II
70
x2
42
.6
9
E-
G
EO
D
-2
62
84
SR
P0
07
46
1
SR
R
30
79
00
EN
C
O
D
E
cy
to
so
l_
gm
12
87
8
Ill
um
in
a
G
A
II
70
x2
36
.5
8
E-
G
EO
D
-2
62
84
SR
P0
07
46
1
SR
R
31
53
34
EN
C
O
D
E
cy
to
so
l_
he
la
-s
3
Ill
um
in
a
G
A
II
70
x2
39
.4
6
E-
G
EO
D
-2
62
84
SR
P0
07
46
1
SR
R
31
53
35
EN
C
O
D
E
cy
to
so
l_
he
la
-s
3
Ill
um
in
a
G
A
II
70
x2
33
.6
3
E-
G
EO
D
-2
62
84
SR
P0
07
46
1
SR
R
30
79
28
EN
C
O
D
E
cy
to
so
l_
he
pg
2
Ill
um
in
a
G
A
II
70
x2
37
.9
5
E-
G
EO
D
-2
62
84
SR
P0
07
46
1
SR
R
30
79
29
EN
C
O
D
E
cy
to
so
l_
he
pg
2
Ill
um
in
a
G
A
II
70
x2
37
.3
9
E-
G
EO
D
-2
62
84
SR
P0
07
46
1
SR
R
30
79
17
EN
C
O
D
E
cy
to
so
l_
hu
ve
c
Ill
um
in
a
G
A
II
70
x2
40
.5
9
E-
G
EO
D
-2
62
84
SR
P0
07
46
1
SR
R
30
79
18
EN
C
O
D
E
cy
to
so
l_
hu
ve
c
Ill
um
in
a
G
A
II
70
x2
37
.0
8
E-
G
EO
D
-2
62
84
SR
P0
07
46
1
SR
R
38
76
61
EN
C
O
D
E
cy
to
so
l_
k5
62
Ill
um
in
a
G
A
II
70
x2
12
4.
83
E-
G
EO
D
-2
62
84
SR
P0
07
46
1
SR
R
38
76
62
EN
C
O
D
E
cy
to
so
l_
k5
62
Ill
um
in
a
G
A
II
70
x2
88
.4
5
E-
G
EO
D
-2
62
84
SR
P0
07
46
1
SR
R
31
52
97
EN
C
O
D
E
nu
cl
eu
s_
gm
12
87
8
Ill
um
in
a
G
A
II
70
x2
42
.0
4
E-
G
EO
D
-2
62
84
SR
P0
07
46
1
SR
R
31
52
98
EN
C
O
D
E
nu
cl
eu
s_
gm
12
87
8
Ill
um
in
a
G
A
II
70
x2
38
.9
8
E-
G
EO
D
-2
62
84
SR
P0
07
46
1
SR
R
31
53
32
EN
C
O
D
E
nu
cl
eu
s_
he
la
-s
3
Ill
um
in
a
G
A
II
70
x2
24
.6
7
E-
G
EO
D
-2
62
84
SR
P0
07
46
1
SR
R
31
53
33
EN
C
O
D
E
nu
cl
eu
s_
he
la
-s
3
Ill
um
in
a
G
A
II
70
x2
36
.5
1
E-
G
EO
D
-2
62
84
SR
P0
07
46
1
SR
R
30
79
15
EN
C
O
D
E
nu
cl
eu
s_
he
pg
2
Ill
um
in
a
G
A
II
70
x2
34
.1
8
E-
G
EO
D
-2
62
84
SR
P0
07
46
1
SR
R
30
79
16
EN
C
O
D
E
nu
cl
eu
s_
he
pg
2
Ill
um
in
a
G
A
II
70
x2
28
.5
3
E-
G
EO
D
-2
62
84
SR
P0
07
46
1
SR
R
30
79
09
EN
C
O
D
E
nu
cl
eu
s_
hu
ve
c
Ill
um
in
a
G
A
II
70
x2
38
.8
3
E-
G
EO
D
-2
62
84
SR
P0
07
46
1
SR
R
30
79
10
EN
C
O
D
E
nu
cl
eu
s_
hu
ve
c
Ill
um
in
a
G
A
II
70
x2
37
.6
9
E-
G
EO
D
-2
62
84
SR
P0
07
46
1
SR
R
31
52
99
EN
C
O
D
E
nu
cl
eu
s_
k5
62
Ill
um
in
a
G
A
II
70
x2
38
.2
3
E-
G
EO
D
-2
62
84
SR
P0
07
46
1
SR
R
31
53
00
EN
C
O
D
E
nu
cl
eu
s_
k5
62
Ill
um
in
a
G
A
II
70
x2
35
.6
8
141
Ta
bl
e
A
.2
|
R
N
A
-s
eq
da
ta
us
ed
in
C
ha
pt
er
3.
A
ll
sa
m
pl
es
ar
e
po
ly
A
-s
el
ec
te
d.
Pa
ir
ed
re
ad
s
ar
e
co
un
te
d
on
ce
.
Sa
m
pl
e
id
Sa
m
pl
e
de
sc
ri
pt
io
n
Se
qu
en
ci
ng
pl
at
fo
rm
R
ea
d
le
ng
th
(b
p)
Se
qu
en
ci
ng
de
pt
h
(M
re
ad
s)
B0
0E
U
N
U
.B
C
08
M
3A
C
X
X
.6
ce
ll
lin
e
Ill
um
in
a
H
iS
eq
20
00
10
0x
2
44
.7
9
B0
0E
U
N
Z
.B
C
08
T1
A
C
X
X
.2
ce
ll
lin
e
Ill
um
in
a
H
iS
eq
20
00
10
0x
2
72
.8
4
B0
0E
U
N
X
.B
D
0C
6H
A
C
X
X
.6
ce
ll
lin
e
Ill
um
in
a
H
iS
eq
20
00
10
0x
2
36
.0
4
B0
0E
U
N
V.
BC
08
T1
A
C
X
X
.2
ce
ll
lin
e
Ill
um
in
a
H
iS
eq
20
00
10
0x
2
48
.2
1
B0
0E
U
N
Y.
C
11
LH
A
C
X
X
.1
ce
ll
lin
e
Ill
um
in
a
H
iS
eq
20
00
10
0x
2
83
.3
0
B0
0E
U
N
W
.A
C
08
1R
A
C
X
X
.2
ce
ll
lin
e
Ill
um
in
a
H
iS
eq
20
00
10
0x
2
66
.7
4
B0
0E
4G
T.
A
C
01
U
U
A
BX
X
.8
no
rm
al
Ill
um
in
a
H
iS
eq
20
00
10
0x
2
87
.9
6
B0
0E
4G
Y.
A
C
07
U
W
A
C
X
X
.3
no
rm
al
Ill
um
in
a
H
iS
eq
20
00
10
0x
2
51
.0
7
B0
0E
4G
Z
.A
C
07
U
W
A
C
X
X
.2
no
rm
al
Ill
um
in
a
H
iS
eq
20
00
10
0x
2
50
.7
7
B0
0E
4H
2.
BC
07
B5
A
C
X
X
.6
no
rm
al
Ill
um
in
a
H
iS
eq
20
00
10
0x
2
46
.9
1
B0
0E
4H
3.
BC
07
B5
A
C
X
X
.5
no
rm
al
Ill
um
in
a
H
iS
eq
20
00
10
0x
2
48
.8
6
B0
0E
4H
4.
BC
07
B5
A
C
X
X
.4
no
rm
al
Ill
um
in
a
H
iS
eq
20
00
10
0x
2
46
.1
1
B0
0E
4H
5.
BC
07
B5
A
C
X
X
.6
no
rm
al
Ill
um
in
a
H
iS
eq
20
00
10
0x
2
47
.9
9
B0
0E
4H
7.
BC
07
B5
A
C
X
X
.5
no
rm
al
Ill
um
in
a
H
iS
eq
20
00
10
0x
2
45
.0
4
B0
0E
4H
9.
BC
07
B5
A
C
X
X
.4
no
rm
al
Ill
um
in
a
H
iS
eq
20
00
10
0x
2
40
.3
8
B0
0E
4H
O
.A
C
07
U
W
A
C
X
X
.1
no
rm
al
Ill
um
in
a
H
iS
eq
20
00
10
0x
2
46
.7
4
B0
0E
4H
P.
BC
07
B5
A
C
X
X
.7
no
rm
al
Ill
um
in
a
H
iS
eq
20
00
10
0x
2
43
.6
1
B0
0E
4H
Q
.A
C
07
U
W
A
C
X
X
.3
no
rm
al
Ill
um
in
a
H
iS
eq
20
00
10
0x
2
41
.4
8
B0
0E
4H
R
.A
C
07
U
W
A
C
X
X
.2
no
rm
al
Ill
um
in
a
H
iS
eq
20
00
10
0x
2
56
.6
6
B0
0E
4H
T.
A
C
07
U
W
A
C
X
X
.1
no
rm
al
Ill
um
in
a
H
iS
eq
20
00
10
0x
2
45
.2
0
B0
0E
4H
V.
BC
07
B5
A
C
X
X
.7
no
rm
al
Ill
um
in
a
H
iS
eq
20
00
10
0x
2
35
.2
4
B0
0E
4I
7.
BC
0G
G
FA
C
X
X
.3
no
rm
al
Ill
um
in
a
H
iS
eq
20
00
10
0x
2
23
.8
6
B0
0E
4I
9.
A
D
0P
8H
A
C
X
X
.1
no
rm
al
Ill
um
in
a
H
iS
eq
20
00
10
0x
2
45
.2
2
142
B0
0E
4I
A
.A
D
0P
8H
A
C
X
X
.2
no
rm
al
Ill
um
in
a
H
iS
eq
20
00
10
0x
2
83
.2
5
B0
0E
4I
B.
A
C
0M
L7
A
C
X
X
.1
no
rm
al
Ill
um
in
a
H
iS
eq
20
00
10
0x
2
38
.8
9
B0
0E
4I
D
.B
C
0G
G
FA
C
X
X
.3
no
rm
al
Ill
um
in
a
H
iS
eq
20
00
10
0x
2
62
.5
9
B0
0E
4I
F.
A
D
0P
8H
A
C
X
X
.1
no
rm
al
Ill
um
in
a
H
iS
eq
20
00
10
0x
2
73
.9
6
B0
0E
4I
G
.A
D
0P
8H
A
C
X
X
.2
no
rm
al
Ill
um
in
a
H
iS
eq
20
00
10
0x
2
45
.1
8
B0
0E
4I
H
.A
C
0M
L7
A
C
X
X
.1
no
rm
al
Ill
um
in
a
H
iS
eq
20
00
10
0x
2
92
.3
8
B0
0E
4I
V.
BC
0G
G
FA
C
X
X
.4
no
rm
al
Ill
um
in
a
H
iS
eq
20
00
10
0x
2
52
.4
7
B0
0E
4I
W
.B
C
0G
G
FA
C
X
X
.4
no
rm
al
Ill
um
in
a
H
iS
eq
20
00
10
0x
2
35
.6
1
B0
0E
4J
H
.B
C
0G
G
FA
C
X
X
.2
no
rm
al
Ill
um
in
a
H
iS
eq
20
00
10
0x
2
43
.3
8
B0
0E
X
Z
C
.B
C
0M
LY
A
C
X
X
.4
no
rm
al
Ill
um
in
a
H
iS
eq
20
00
10
0x
2
41
.1
7
B0
0E
X
Z
D
.B
C
0M
LY
A
C
X
X
.3
no
rm
al
Ill
um
in
a
H
iS
eq
20
00
10
0x
2
34
.8
6
B0
0E
X
Z
V.
BC
0M
LY
A
C
X
X
.3
no
rm
al
Ill
um
in
a
H
iS
eq
20
00
10
0x
2
61
.7
6
B0
0E
X
Z
X
.B
C
0M
LY
A
C
X
X
.3
no
rm
al
Ill
um
in
a
H
iS
eq
20
00
10
0x
2
37
.2
3
B0
0E
X
Z
Y.
BC
0M
LY
A
C
X
X
.4
no
rm
al
Ill
um
in
a
H
iS
eq
20
00
10
0x
2
54
.0
7
B0
0E
Y
0G
.B
C
0M
LY
A
C
X
X
.5
no
rm
al
Ill
um
in
a
H
iS
eq
20
00
10
0x
2
39
.0
3
B0
0E
Y
0H
.B
C
0M
LY
A
C
X
X
.4
no
rm
al
Ill
um
in
a
H
iS
eq
20
00
10
0x
2
44
.7
1
B0
0E
Y
0I
.C
0M
P5
A
C
X
X
.3
no
rm
al
Ill
um
in
a
H
iS
eq
20
00
10
0x
2
40
.2
7
B0
0E
Y
0Q
.A
C
0R
M
C
A
C
X
X
.1
no
rm
al
Ill
um
in
a
H
iS
eq
20
00
10
0x
2
31
.4
8
B0
0E
Y
0S
.A
C
0M
FT
A
C
X
X
.1
no
rm
al
Ill
um
in
a
H
iS
eq
20
00
10
0x
2
41
.4
0
B0
0E
Y
0T
.C
0M
P5
A
C
X
X
.5
no
rm
al
Ill
um
in
a
H
iS
eq
20
00
10
0x
2
48
.0
4
B0
0E
Y
0U
.A
C
0M
FT
A
C
X
X
.2
no
rm
al
Ill
um
in
a
H
iS
eq
20
00
10
0x
2
51
.2
7
B0
0E
Y
0V
.A
C
0M
FT
A
C
X
X
.1
no
rm
al
Ill
um
in
a
H
iS
eq
20
00
10
0x
2
46
.2
9
B0
0E
Y
0Y
.C
0M
P5
A
C
X
X
.3
no
rm
al
Ill
um
in
a
H
iS
eq
20
00
10
0x
2
62
.1
1
B0
0E
Y
17
.C
0M
P5
A
C
X
X
.4
no
rm
al
Ill
um
in
a
H
iS
eq
20
00
10
0x
2
46
.5
3
B0
0E
Y
1D
.C
0L
M
0A
C
X
X
.6
no
rm
al
Ill
um
in
a
H
iS
eq
20
00
10
0x
2
52
.4
5
B0
0E
Y
1E
.C
0L
M
0A
C
X
X
.6
no
rm
al
Ill
um
in
a
H
iS
eq
20
00
10
0x
2
32
.1
7
B0
0E
Y
1G
.A
C
0R
M
C
A
C
X
X
.4
no
rm
al
Ill
um
in
a
H
iS
eq
20
00
10
0x
2
27
.9
7
143
B0
0F
3D
Y.
D
16
U
7A
C
X
X
.4
no
rm
al
Ill
um
in
a
H
iS
eq
20
00
10
0x
2
45
.2
1
B0
0E
4H
A
.A
C
07
U
W
A
C
X
X
.1
tu
m
or
Ill
um
in
a
H
iS
eq
20
00
10
0x
2
40
.9
6
B0
0E
4H
B.
BC
07
B5
A
C
X
X
.7
tu
m
or
Ill
um
in
a
H
iS
eq
20
00
10
0x
2
60
.3
9
B0
0E
4H
E.
BC
07
B5
A
C
X
X
.6
tu
m
or
Ill
um
in
a
H
iS
eq
20
00
10
0x
2
32
.9
5
B0
0E
4H
F.
BC
07
B5
A
C
X
X
.5
tu
m
or
Ill
um
in
a
H
iS
eq
20
00
10
0x
2
34
.6
1
B0
0E
4H
G
.B
C
07
B5
A
C
X
X
.4
tu
m
or
Ill
um
in
a
H
iS
eq
20
00
10
0x
2
40
.3
0
B0
0E
4H
H
.B
C
07
B5
A
C
X
X
.6
tu
m
or
Ill
um
in
a
H
iS
eq
20
00
10
0x
2
32
.8
1
B0
0E
4H
J.B
C
07
B5
A
C
X
X
.5
tu
m
or
Ill
um
in
a
H
iS
eq
20
00
10
0x
2
33
.8
6
B0
0E
4H
L.
BC
07
B5
A
C
X
X
.4
tu
m
or
Ill
um
in
a
H
iS
eq
20
00
10
0x
2
32
.2
4
B0
0E
4H
Y.
A
C
08
1R
A
C
X
X
.2
tu
m
or
Ill
um
in
a
H
iS
eq
20
00
10
0x
2
54
.1
4
B0
0E
4H
Z
.B
C
08
T1
A
C
X
X
.2
tu
m
or
Ill
um
in
a
H
iS
eq
20
00
10
0x
2
90
.4
4
B0
0E
4I
0.
BC
08
M
3A
C
X
X
.6
tu
m
or
Ill
um
in
a
H
iS
eq
20
00
10
0x
2
78
.4
6
B0
0E
4I
1.
BC
08
M
3A
C
X
X
.6
tu
m
or
Ill
um
in
a
H
iS
eq
20
00
10
0x
2
82
.1
1
B0
0E
4I
3.
A
C
07
U
W
A
C
X
X
.2
tu
m
or
Ill
um
in
a
H
iS
eq
20
00
10
0x
2
39
.7
1
B0
0E
4I
5.
A
C
07
U
W
A
C
X
X
.2
tu
m
or
Ill
um
in
a
H
iS
eq
20
00
10
0x
2
44
.1
0
B0
0E
4I
J.B
C
0G
G
FA
C
X
X
.3
tu
m
or
Ill
um
in
a
H
iS
eq
20
00
10
0x
2
43
.8
8
B0
0E
4I
L.
A
D
0P
8H
A
C
X
X
.1
tu
m
or
Ill
um
in
a
H
iS
eq
20
00
10
0x
2
48
.9
9
B0
0E
4I
M
.A
D
0P
8H
A
C
X
X
.2
tu
m
or
Ill
um
in
a
H
iS
eq
20
00
10
0x
2
44
.6
6
B0
0E
4I
N
.A
C
0M
L7
A
C
X
X
.1
tu
m
or
Ill
um
in
a
H
iS
eq
20
00
10
0x
2
36
.6
3
B0
0E
4I
P.
BC
0G
G
FA
C
X
X
.3
tu
m
or
Ill
um
in
a
H
iS
eq
20
00
10
0x
2
49
.2
3
B0
0E
4I
R
.A
D
0P
8H
A
C
X
X
.1
tu
m
or
Ill
um
in
a
H
iS
eq
20
00
10
0x
2
42
.3
8
B0
0E
4I
S.
A
D
0P
8H
A
C
X
X
.2
tu
m
or
Ill
um
in
a
H
iS
eq
20
00
10
0x
2
48
.9
9
B0
0E
4I
T.
A
C
0M
L7
A
C
X
X
.1
tu
m
or
Ill
um
in
a
H
iS
eq
20
00
10
0x
2
28
.9
9
B0
0E
4J
1.
BC
0G
G
FA
C
X
X
.5
tu
m
or
Ill
um
in
a
H
iS
eq
20
00
10
0x
2
58
.1
2
B0
0E
4J
4.
BC
0G
G
FA
C
X
X
.4
tu
m
or
Ill
um
in
a
H
iS
eq
20
00
10
0x
2
63
.6
0
B0
0E
4J
5.
BC
0G
G
FA
C
X
X
.2
tu
m
or
Ill
um
in
a
H
iS
eq
20
00
10
0x
2
32
.9
2
B0
0E
4J
7.
C
0L
M
0A
C
X
X
.6
tu
m
or
Ill
um
in
a
H
iS
eq
20
00
10
0x
2
82
.4
5
144
B0
0E
4J
B.
BC
0M
LY
A
C
X
X
.2
tu
m
or
Ill
um
in
a
H
iS
eq
20
00
10
0x
2
43
.7
6
B0
0E
X
Z
H
.B
C
0M
LY
A
C
X
X
.2
tu
m
or
Ill
um
in
a
H
iS
eq
20
00
10
0x
2
33
.2
7
B0
0E
X
Z
I.B
C
0M
LY
A
C
X
X
.1
tu
m
or
Ill
um
in
a
H
iS
eq
20
00
10
0x
2
22
.4
0
B0
0E
X
Z
R
.C
0M
P5
A
C
X
X
.3
tu
m
or
Ill
um
in
a
H
iS
eq
20
00
10
0x
2
38
.6
8
B0
0E
X
Z
U
.B
C
0M
LY
A
C
X
X
.5
tu
m
or
Ill
um
in
a
H
iS
eq
20
00
10
0x
2
39
.1
6
B0
0E
Y
02
.B
C
0M
LY
A
C
X
X
.5
tu
m
or
Ill
um
in
a
H
iS
eq
20
00
10
0x
2
47
.2
0
B0
0E
Y
0J
.C
0M
P5
A
C
X
X
.5
tu
m
or
Ill
um
in
a
H
iS
eq
20
00
10
0x
2
57
.8
5
B0
0E
Y
0M
.C
0M
P5
A
C
X
X
.4
tu
m
or
Ill
um
in
a
H
iS
eq
20
00
10
0x
2
49
.4
5
B0
0E
Y
0N
.C
0M
P5
A
C
X
X
.4
tu
m
or
Ill
um
in
a
H
iS
eq
20
00
10
0x
2
42
.1
8
B0
0E
Y
0O
.C
0M
P5
A
C
X
X
.3
tu
m
or
Ill
um
in
a
H
iS
eq
20
00
10
0x
2
39
.7
5
B0
0E
Y
0R
.A
C
0M
FT
A
C
X
X
.2
tu
m
or
Ill
um
in
a
H
iS
eq
20
00
10
0x
2
55
.8
0
B0
0E
Y
0Z
.A
C
0R
M
C
A
C
X
X
.3
tu
m
or
Ill
um
in
a
H
iS
eq
20
00
10
0x
2
32
.1
3
B0
0E
Y
16
.C
0M
P5
A
C
X
X
.5
tu
m
or
Ill
um
in
a
H
iS
eq
20
00
10
0x
2
38
.6
5
B0
0E
Y
18
.A
C
0M
FT
A
C
X
X
.1
tu
m
or
Ill
um
in
a
H
iS
eq
20
00
10
0x
2
42
.8
1
B0
0E
Y
19
.C
0M
P5
A
C
X
X
.5
tu
m
or
Ill
um
in
a
H
iS
eq
20
00
10
0x
2
36
.9
4
B0
0E
Y
1A
.C
0M
P5
A
C
X
X
.4
tu
m
or
Ill
um
in
a
H
iS
eq
20
00
10
0x
2
47
.4
2
B0
0E
Y
1B
.A
C
0M
FT
A
C
X
X
.2
tu
m
or
Ill
um
in
a
H
iS
eq
20
00
10
0x
2
34
.2
3
B0
0E
Y
1K
.C
0L
M
0A
C
X
X
.3
tu
m
or
Ill
um
in
a
H
iS
eq
20
00
10
0x
2
29
.5
8
B0
0E
Y
1L
.C
0L
M
0A
C
X
X
.7
tu
m
or
Ill
um
in
a
H
iS
eq
20
00
10
0x
2
37
.0
1
145
Ta
bl
e
A
.3
|
R
N
A
-s
eq
da
ta
us
ed
in
C
ha
pt
er
s
4
an
d
5.
A
ll
sa
m
pl
es
ar
e
po
ly
A
-
se
le
ct
ed
.P
ai
re
d
re
ad
s
ar
e
co
un
te
d
on
ce
.
Sa
m
pl
e
id
Sa
m
pl
e
de
sc
ri
pt
io
n
Se
qu
en
ci
ng
pl
at
fo
rm
R
ea
d
le
ng
th
(b
p)
Se
qu
en
ci
ng
de
pt
h
(M
re
ad
s)
C
1
C
A
L5
1
co
nt
ro
l
Ill
um
in
a
H
iS
eq
20
00
10
0x
2
18
.9
7
C
2
C
A
L5
1
co
nt
ro
l
Ill
um
in
a
H
iS
eq
20
00
10
0x
2
26
.1
9
C
3
C
A
L5
1
co
nt
ro
l
Ill
um
in
a
H
iS
eq
20
00
10
0x
2
29
.2
4
K
D
1
C
A
L5
1
PR
PF
8
K
D
Ill
um
in
a
H
iS
eq
20
00
10
0x
2
24
.6
6
K
D
2
C
A
L5
1
PR
PF
8
K
D
Ill
um
in
a
H
iS
eq
20
00
10
0x
2
18
.4
2
K
D
3
C
A
L5
1
PR
PF
8
K
D
Ill
um
in
a
H
iS
eq
20
00
10
0x
2
28
.8
0
K
D
4
C
A
L5
1
PR
PF
8
K
D
Ill
um
in
a
H
iS
eq
20
00
10
0x
2
28
.9
5
146
Appendix B
Supplementary Material for
Chapter 2
147
0.
0
0.
2
0.
4
0.
6
0.
8
1.
0
0.
0
0.
2
0.
4
0.
6
0.
8
1.
0
0.
0
0.
2
0.
4
0.
6
0.
8
1.
0
0.
0
0.
2
0.
4
0.
6
0.
8
1.
0
transcript position in ranking
tr
an
sc
rip
t r
el
at
iv
e 
ab
un
da
nc
e
1 FPKM 5 FPKM 10 FPKM
BM
 PE
EN
CO
D
E CELL
EN
CO
D
E CYTO
SO
L
EN
CO
D
E  N
U
CLEU
S
1 2 3 4 5 +5 1 2 3 4 5 +5 1 2 3 4 5 +5
tr
an
sc
rip
t r
el
at
iv
e 
ab
un
da
nc
e
1 2 3 4 5 +5 1 2 3 4 5 +5 1 2 3 4 5 +5 1 2 3 4 5 +5 1 2 3 4 5 +5
0.
0
0.
2
0.
4
0.
6
0.
8
1.
0
transcript position in ranking
2 transcripts 3 transcripts 5 transcripts +5 transcripts4 transcripts
(a)
(b)
Figure B.1| Transcript relative abundances across all the studied datasets.
(a) Relative abundance of the subset of transcripts in each position of the ranking for all
datasets, including different expression thresholds. Only genes with more than one
annotated transcript are represented here.
(b) Relative abundance of the subset of transcripts in each position of the ranking for the
BM dataset, separating genes by the number of annotated transcripts. The expression
threshold used here is 1 FPKM.
148
1 FPKM 5 FPKM 10 FPKM
BM
 PE
EN
CO
D
E CELL
EN
CO
D
E CYTO
SO
L
EN
CO
D
E  N
U
CLEU
S
0.
0
0.
2
0.
4
0.
6
0.
8
1.
0
0.
0
0.
2
0.
4
0.
6
0.
8
1.
0
0.
0
0.
2
0.
4
0.
6
0.
8
1.
0
0.
0
0.
2
0.
4
0.
6
0.
8
1.
0
2 3 4 5 +5 2 3 4 5 +5 2 3 4 5 +5
number of annotated transcripts
m
aj
or
 tr
an
sc
rip
t d
om
in
an
ce
Figure B.2| Major transcript dominance across all the studied datasets. Major
transcript dominance is measured by the ratio of expression between the major
transcript and the second most abundant one.
149
term padj
protein transport 6.85E-08
Transcription 1.46E-06
transcription regulation 7.27E-06
establishment of protein localization 1.73E-05
protein transport 2.19E-05
small GTPase mediated signal transduction 2.65E-05
respiratory chain 8.53E-05
transcription 2.34E-04
ubl conjugation pathway 2.50E-04
oxidative phosphorylation 4.36E-04
cellular macromolecule catabolic process 9.00E-04
macromolecule catabolic process 1.21E-03
electron transport 2.51E-03
mitochondrial ATP synthesis coupled electron transport 2.96E-03
regulation of transcription 3.11E-03
GTPase activity 3.36E-03
respiratory chain 3.44E-03
intracellular transport 6.93E-03
Oxidative phosphorylation 2.34E-02
Table B.1| GO enrichment analysis for recurrent 5-fold dominant genes.
term padj
regulation of translational initiation 2.23E-03
mrna processing 3.83E-03
translation initiation factor activity 4.25E-03
mrna splicing 6.20E-03
RNA splicing 6.63E-03
RNA binding 7.65E-03
posttranscriptional regulation of gene expression 1.83E-02
translation regulation 2.28E-02
regulation of translation 2.31E-02
mRNA processing 2.51E-02
mRNA metabolic process 2.95E-02
RNA processing 3.39E-02
spliceosome 3.50E-02
Table B.2| GO enrichment analysis for genes that tolerate splicing.
150
transcript position in the ranking
1 2 3 4 5 >5
ENCODE CELL
1 FPKM 71.36 14.06 5.62 2.91 1.74 4.31
5 FPKM 75.29 12.57 4.86 2.45 1.43 3.40
10 FPKM 80.14 10.52 3.90 1.91 1.08 2.45
ENCODE CYTOSOL
1 FPKM 81.05 10.54 3.40 1.63 0.95 2.44
5 FPKM 83.88 9.39 2.86 1.32 0.74 1.81
10 FPKM 86.77 8.04 2.30 1.02 0.55 1.31
ENCODE NUCLEUS
1 FPKM 63.49 16.20 7.53 4.15 2.54 6.09
5 FPKM 66.79 14.96 6.87 3.74 2.27 5.36
10 FPKM 71.71 12.97 5.89 3.17 1.90 4.36
Table B.3| mRNA pool estimates for the cell line dataset.
expressed genes genes with a dominant major transcript2-fold dominance 5-fold dominance
ENCODE CELL
1 FPKM 9770 6550 66.83 4061 41.29
5 FPKM 4952 3475 71.51 2334 49.01
10 FPKM 2787 2029 74.67 1449 54.73
ENCODE CYTOSOL
1 FPKM 9281 6139 66.29 3619 39.16
5 FPKM 5076 3590 71.31 2352 47.09
10 FPKM 3041 2259 74.81 1601 53.39
ENCODE NUCLEUS
1 FPKM 9923 5936 60.32 3182 32.62
5 FPKM 5796 3620 63.70 2085 37.53
10 FPKM 3516 2268 65.85 1391 41.34
Table B.4| Number of dominant transcripts in the cell line dataset.
151
term padj
ribosomal protein 1.48E-04
protein biosynthesis 1.87E-04
ribosomal subunit 3.07E-04
structural constituent of ribosome 3.20E-04
ribosome 3.23E-04
translation 1.99E-03
ribosome 3.23E-03
small ribosomal subunit 2.03E-02
transit peptide 3.18E-02
mRNA transport 3.68E-02
Table B.5| GO enrichment analysis for genes with a major retained intron
both in the nucleus and the cytosol.
Gene id Transcript id HGNC symbol Reannotated biotype
ENSG00000029364 ENST00000031146 SLC39A9 protein coding
ENSG00000090054 ENST00000486910 SPTLC1 removed (artefact)
ENSG00000097033 ENST00000370558 SH3GLB1 removed (merged)
ENSG00000101150 ENST00000474176 TPD52L2 retained intron
ENSG00000103549 ENST00000562110 RNF40 protein coding (NMD)
ENSG00000112159 ENST00000487831 MDN1 protein coding (NMD)
ENSG00000114999 ENST00000460450 TTL processed transcript
ENSG00000116288 ENST00000469225 PARK7 protein coding
ENSG00000122696 ENST00000496760 SLC25A51 lncRNA
ENSG00000130703 ENST00000471817 OSBPL2 retained intron
ENSG00000136908 ENST00000495270 DPM2 retained intron
ENSG00000136935 ENST00000475407 GOLGA1 protein coding (NMD)
ENSG00000143149 ENST00000463610 ALDH9A1 processed transcript
ENSG00000148396 ENST00000467838 SEC16A retained intron
ENSG00000149823 ENST00000527646 VPS51 procesed transcript
ENSG00000156502 ENST00000497254 SUPV3L1 retained intron
ENSG00000162775 ENST00000487146 RBM15 protein coding
ENSG00000166582 ENST00000472570 CENPV retained intron
ENSG00000173812 ENST00000310837 EIF1 possible protein coding
ENSG00000185305 ENST00000502271 ARL15 possible protein coding
ENSG00000198917 ENST00000467582 C9orf114 protein coding
ENSG00000213995 ENST00000470164 CARKD protein coding or NMD
Table B.6| Re-annotation of major processed transcripts.
152
Appendix C
Supplementary Material for
Chapter 3
153
+... ......
gene_id transcript_id sample1 sample2 ...
expression levels
(TPMs, RPKMs/FPKMs, etc.)
annotation
switchseq -t get_data
INPUT
OUTPUT
switchseq -t get_switch
SwitchSeq + tviz
self-contained html (+ txt, json)
high resolution plots
Figure C.1| SwitchSeq analysis workflow. Further information on installation
and execution instructions can be found on the project website (https://github.
com/mgonzalezporta/SwitchSeq).
154
155
LAMC1
LAMA5
TNC
THBS2
COL4A2
GNGT2
IL3RA
JAK3
GNB4
YWHAH
COL3A1
CREB5
THBS3
GNG11
CSF3R
NR4A1
COL4A3
FGFR3
RBL2
LPAR2
CSF1
GNB5
PRLR
NGFR
AKT2
IFNAR2
PIK3R3
IL7R
COL6A2
GNB1
PPP2R5B
COL1A1
PRKCZ
IL4R
ITGB7
KDR
ANGPT2
EIF4EBP1
TLR2
CDKN1A
ITGA4
FGFR2
MET
IL2RG
COL5A2
COL1A2
TSC1
CCND2
FLT1
SGK2
NGF
YWHAZ
EGF
MTCP1
G6PC
CSF1R
SGK3
CCND3
PCK1
COL4A1
FN1
OSMR
PDGFD
ITGA5
MLST8
MYC
SYK
EGFR
FGF11
PCK2
VWF
LAMA4
CCND1
PGF
DDIT4
FGF1
PTK2
VEGFA(b)
156
IRS1
FGFR4
GYS1
AKT3
YWHAE
COL4A5
BCL2L11
ITGB3
MDM2
ITGA7
SPP1
ITGAV
THBS4
IL6
PDGFB
COL27A1
ITGA9
MAP2K2
GHR
IL2RA
FGF7
ITGB8
PKN1
INSR
GNG7
F2R
KIT
ITGB1
PTEN
THBS1
LAMC3
IL2RB
ITGA3
LAMC2
ITGA1
ITGA11
KITLG
COL5A1
MAPK1
COL4A4
FLT4
PIK3R1
EFNA3
LPAR6
PIK3AP1
PDGFRA
GNG2
NOS3
PIK3R5
COL5A3
GNB2
MAPK3
TEK
ATF2
EFNA1
MTOR
HGF
PDGFA
LAMA1
LAMB3
LAMB1
PIK3CD
EFNA5
CREB3L3
PDGFRB
ITGB6
EPHA2
SOS1
PKN3
IL7
ITGB4
RPS6KB2
TLR4
COL6A3
FGF2
EIF4E
VEGFC
BCL2
157
GNG12
FGF9
CCNE1
PPP2R2B
BRCA1
CHUK
CREB1
CREB3L2
PRKAA1
FGFR1
CDK4
BAD
TSC2
CDK2
KRAS
TNXB
BCL2L1
YWHAQ
NFKB1
PPP2CB
GRB2
YWHAG
HSP90AB1
NRAS
CDKN1B
EFNA4
FGF13
FASLG
GNG4
PDPK1
MAP2K1
IFNAR1
CHAD
PPP2R5A
FIGF
ITGA2
VEGFB
AKT1
LAMA2
CREB3L4
PHLPP1
PKN2
CREB3
LPAR1
FGF12
CDK6
RPTOR
PRKAA2
HSP90AA1
FOXO3
PPP2R3A
COL11A1
TP53
IGF1
LPAR5
JAK1
OSM
GNG10
EPO
IL6R
LAMA3
PPP2R2A
ANGPT1
JAK2
RXRA
IGF1R
CREB3L1
ITGB5
158
Figure C.2| Summary of detected alterations in the PI(3)K-AKT-MTOR sig-
nalling pathway.
(a) The PI(3)K-AKT-MTOR signalling pathway. Genes which have been identified
as altered in any of the three considered analyses (i.e. switch events, gene expres-
sion changes and confirmed somatic mutations) have been highlighted in red.
(b) Patient-specific landscape of alterations for genes in the PI(3)K-AKT-MTOR sig-
nalling pathway. Each gene is represented by three tracks as described in Fig-
ure 3.7, and only those previously highlighted in a have been included. Genes
have been sorted based on the total number of alterations of any kind.
159
160
THBS1
LAMC3
TLN2
ITGA3
LAMC2
ITGA1
ITGA11
COL5A1
MAPK1
ZYX
COL4A4
FLT4
PIK3R1
PDGFRA
PIK3R5
COL5A3
MAPK3
HGF
PDGFA
LAMA1
LAMB3
LAMB1
PIK3CD
PDGFRB
SHC2
MAPK8
ITGB6
SOS1
ITGB4
PPP1R12B
COL6A3
VEGFC
BCL2
LAMC1
LAMA5
TNC
PARVG
THBS2
COL4A2
PAK1
RAC3
COL3A1
FLNB
THBS3
SHC1
FLNA
BRAF
RAPGEF1
COL4A3
ACTN1
AKT2
PIK3R3
COL6A2
MYL5
COL1A1
ITGB7
KDR
MAPK10
MAPK9
ITGA4
CAPN2
SRC
MET
COL5A2
COL1A2
ACTN4
CCND2
FLT1
EGF
DIAPH1
PPP1R12A
CCND3
RAC2
COL4A1
VAV1
FN1
PDGFD
ITGA5
BIRC3
EGFR
BCAR1
VWF
CAV2
LAMA4
CCND1
PGF
FYN
CAV1
PTK2
VEGFA(b)
161
DOCK1
PPP1CC
MYL12A
PPP1CA
ARHGAP5
MYL9
CTNNB1
PXN
BAD
TNXB
PAK2
ILK
GRB2
ACTG1
PDPK1
PPP1CB
MAP2K1
CHAD
CDC42
ROCK1
FIGF
ARHGAP35
ITGA2
VEGFB
AKT1
LAMA2
VASP
VAV2
MYLK
PRKCB
ACTB
COL11A1
IGF1
LAMA3
FLNC
MYLK4
IGF1R
ITGB5
VAV3
AKT3
TLN1
COL4A5
ITGB3
PIP5K1C
PARVB
ITGA7
SPP1
ITGAV
THBS4
JUN
PDGFB
COL27A1
ERBB2
ITGA9
ITGB8
ITGB1
PTEN
162
Figure C.3| Summary of detected alterations in the focal adhesion pathway.
(a) The focal adhesion pathway. Genes which have been identified as altered in any
of the three considered analyses (i.e. switch events, gene expression changes and
confirmed somatic mutations) have been highlighted in red.
(b) Patient-specific landscape of alterations for genes in the focal adhesion pathway. Each
gene is represented by three tracks as described in Figure 3.7, and only those
previously highlighted in a have been included. Genes have been sorted based
on the total number of alterations of any kind.
pathway padj
Metabolic pathways 2.25E-127
Pathways in cancer 2.16E-049
Cell cycle 1.79E-034
MAPK signaling pathway 1.91E-028
Ubiquitin mediated proteolysis 1.32E-027
Regulation of actin cytoskeleton 5.06E-027
RNA transport 1.13E-026
Spliceosome 3.87E-026
Wnt signaling pathway 2.97E-023
Renal cell carcinoma 1.72E-019
Apoptosis 2.27E-019
Focal adhesion 4.78E-018
Adherens junction 2.29E-017
ErbB signaling pathway 7.36E-017
TGF-beta signaling pathway 3.40E-016
p53 signaling pathway 1.17E-015
mRNA surveillance pathway 1.84E-013
Table C.1| Pathway enrichment analysis for differentially spliced genes in cell
lines vs. tumour samples.
163

Appendix D
Supplementary Material for
Chapter 4
165
IB: PRPF8
IB:  Actin
C KDkDa
250
   50
(b)(a)
Figure D.1| Validation of PRPF8 down-regulation. Cal51 cells were treated
with either PRPF8 siRNAs (PRPF8 knock-down; KD), or all-stars siRNAs (con-
trol; C).
(a) qRT-PCR results. PRPF8 expression levels were efficiently down-regulated (p-
value < 0.001).
(b) Western blot validation. The knock-down can also be recapitulated at the pro-
tein level.
Figure provided by Dr. Vi Wickramasinghe.
166
COMPLEX A COMPLEX B
*
*
* * *
*
*
COMPLEX B COMPLEX C
*
*
*
Figure D.2| Cell cycle analysis following knock-down of splicing factors from
several spliceosomal complexes. Compared to other components of the A and
C-complexes of the spliceosome, knock-down (KD) of those from complex B gen-
erally results in the accumulation of cells in mitosis. More specifically, KD of
PRPF8 has the biggest effects on the measured phenotype. Complex A: not sig-
nificant (n.s.) except for U2AF2 p < 0.05. Complex B: PRPF8 p < 0.001; BRR2,
EFTUD2 p < 0.01; PRPF31 n.s.; PRPF38A p<0.05; PRPF4 n.s.; CDC5L p<0.01;
PRPF19 p<0.05; CDC40 n.s. Complex C: n.s.
Figure provided by Dr. Vi Wickramasinghe.
167
normalised cDNA counts
control
PRPF8 KD
(b)
-100 -50 0 50 100
0
50
0
10
00
15
00
20
00
25
00
-100 -50 0 50 100
0
50
0
10
00
15
00
20
00
25
00
su
m
 o
f n
or
m
al
ise
d 
cD
N
A 
co
un
ts
su
m
 o
f n
or
m
al
ise
d 
cD
N
A 
co
un
ts
position relative to junction position relative to junction
CONTROL PRPF8 KD(a)
0.0 0.1 0.2 0.3 0.4 0.5 0.6
Figure D.3| Results from Sm iCLIP experiments for the 3’ splice site. Plots to
be interpreted as in Figure 4.3.
(a) RNA maps for control and PRPF8 KD samples.
(b) Changes in the distribution of normalised cDNA counts between control and PRPF8
KD samples.
168
0 20 40 60 80 100
control
normalised intron expression
PRPF8
KD
fk
fl
Figure D.4| Differences in intron expression between first and last introns.
Intron counts were normalised to take into account intron length and both counts
and length of adjacent internal exons (see Chapter 4 - Methods). In general, last
introns ( fl) have higher expression levels than the ones located towards the 5’
end of the same transcript ( fk).
169

Appendix E
Supplementary Material for
Chapter 5
171
−6 −4 −2 0 2
−1
.0
−0
.5
0.
0
0.
5
1.
0
1.
5
FC RNA−seq
FC
 S
W
AT
H
DTU events in major trascripts
all peptides
DTU events in major trascripts
subset based on FC significance
−6 −4 −2 0
−1
.0
−0
.5
0.
0
0.
5
1.
0
1.
5
FC RNA−seq
FC
 S
W
AT
H
all DTU events
all peptides
all DTU events
subset based on FC significance
−6 −4 −2 0 2 4
−2
−1
0
1
2
3
FC RNA−seq
FC
 S
W
AT
H
−6 −4 −2 0 2 4
−1
.0
−0
.5
0.
0
0.
5
1.
0
FC RNA−seq
FC
 S
W
AT
H
all DTU events
peptides that map to only one transcript
−3 −2 −1 0 1 2
−1
.0
−0
.5
0.
0
0.
5
1.
0
FC RNA−seq
FC
 S
W
AT
H
Figure E.1| RNA-seq and SWATH-MS fold-change estimates for the differen-
tially used transcripts with peptide evidence.
172
all DTU events (excluding DGE)
peptides that map to only one transcript
−3 −2 −1 0 1 2
−1
.0
−0
.5
0.
0
0.
5
1.
0
FC RNA−seq
FC
 S
W
AT
H
all DTU events (excluding DGE)
all peptides
all DTU events (excluding DGE)
subset based on FC significance
−6 −4 −2 0 2 4
−1
0
1
2
3
FC RNA−seq
FC
 S
W
AT
H
−6 −4 −2 0 2 4
−1
.0
−0
.5
0.
0
0.
5
1.
0
FC RNA−seq
FC
 S
W
AT
H
DTU events in major trascripts (excluding DGE)
all peptides
−6 −4 −2 0 2
−1
.0
−0
.5
0.
0
0.
5
1.
0
1.
5
FC RNA−seq
FC
 S
W
AT
H
DTU events in major trascripts (excluding DGE)
subset based on FC significance
−6 −4 −2 0
−1
.0
−0
.5
0.
0
0.
5
1.
0
1.
5
FC RNA−seq
FC
 S
W
AT
H
Figure E.2| RNA-seq and SWATH-MS fold-change estimates for the differ-
entially used transcripts with peptide evidence after excluding differentially
expressed genes.
173

References
Alberts B, Johnson A, Lewis J, Raff M, Roberts K, and Walter P. 2002. Molecular
Biology of the Cell. Garland, 4th edition. 2, 5, 6, 7
Alexander RD, Innocente SA, Barrass JD, and Beggs JD. 2010. Splicing-dependent
RNA polymerase pausing in yeast. Molecular Cell 40: 582–93. 112
Amit M, Donyo M, Hollander D, Goren A, Kim E, Gelfman S, Galit L, Burstein
D, Schwartz S, Postolsky B, et al.. 2012. Differential GC content between exons
and introns establishes distinct strategies of splice-site recognition. Cell Reports
1: 543–556. 96, 111, 114
Anders S and Huber W. 2010. Differential expression analysis for sequence count
data. Genome Biology 11: R106. 30, 115
Anders S, Pyl PT, and Huber W. 2014. HTSeq: A python framework to work with
high-throughput sequencing data. bioRxiv. 10.1101/002824. 21, 22, 89
Anders S, Reyes A, and Huber W. 2012. Detecting differential usage of exons from
RNA-seq data. Genome Research 22: 2008–2017. 21, 33, 58, 78, 89, 113
Andrews S. 2010. Fastqc: A quality control tool for high throughput sequence
data. http://www.bioinformatics.babraham.ac.uk/projects/fastqc/. 19
Ardehali M and Lis J. 2009. Tracking rates of transcription and splicing in vivo.
Nature Structural & Molecular Biology 16: 1123–1124. 110
Ast G. 2004. How did alternative splicing evolve? Nature Reviews Genetics 5:
773–782. 13
Audenet F, Yates D, Géraldine C, Cussenot O, and Rouprêt M. 2012. Genetic
pathways involved in carcinogenesis of clear cell renal cell carcinoma: genomics
towards personalized medicine. BJU International 109: 1864–1870. 87
Auman J and Howard M. 2010. Colorectal cancer cell lines lack the molecular
heterogeneity of clinical colorectal tumors. Clinical Colorectal Cancer 9: 40–47.
82, 87
175
Barash Y, Calarco J, Gao W, Pan Q, Wang X, Shai O, Blencowe B, and Frey B. 2010.
Deciphering the splicing code. Nature 465: 53–59. 93, 135
Barbosa-Morais NL, Irimia M, Pan Q, Xiong HY, Gueroussov S, Lee LJ, Slobode-
niuc V, Kutter C, Watt S, Colak R, et al.. 2012. The evolutionary landscape of
alternative splicing in vertebrate species. Science (New York, N.Y.) 338: 1587–1593.
55
Beaumont MA and Rannala B. 2004. The Bayesian revolution in genetics. Nature
Reviews Genetics 5: 251–61. 23
Benjamini Y and Hochberg Y. 1995. Controlling the false discovery rate: a practical
and powerful approach to multiple testing. J R Stat Soc Ser B 57: 289–300. 59,
89, 90, 114
Benjamini Y and Speed TP. 2012. Summarizing and correcting the GC content bias
in high-throughput sequencing. Nucleic Acids Research 40: e72. 17
Benjamini Y and Yekutieli D. 2001. The control of the false discovery rate in
multiple testing under dependency. Ann. Statist. 29: 1165–1188. 89, 115
Bentley DL. 2014. Coupling mRNA processing with transcription in time and
space. Nature Reviews Genetics 15: 163–75. 14, 111
Bentley DR, Balasubramanian S, Swerdlow HP, Smith GP, Milton J, Brown CG,
Hall KP, Evers DJ, Barnes CL, Bignell HR, et al.. 2008. Accurate whole human
genome sequencing using reversible terminator chemistry. Nature 456: 53–9. 17
Berget S, Moore C, and Sharp P. 1977. Spliced segments at the 5’ terminus of
adenovirus 2 late mRNA. Proceedings of the National Academy of Sciences of the
United States of America 74: 3171–3175. 6
Blakeley P, Siepen J, Lawless C, and Hubbard S. 2010. Investigating protein iso-
forms via proteomics: a feasibility study. Proteomics 10: 1127–1140. 119, 129
Briese M, Rot G, Wang Z, König J, Sim AY, Levitt M, Gorup C, Zupan B, Saieva L,
Pellizzoni L, et al.., Genome-wide analysis of spliceosomal protein-rna interac-
tions and their atp-dependent remodeling. In preparation. 97, 99, 112
Brown S, Stoilov P, and Xing Y. 2012. Chromatin and epigenetic regulation of
pre-mRNA processing. Human Molecular Genetics 21: R90–R96. 13
Brugarolas J. 2013. PBRM1 and BAP1 as novel targets for renal cell carcinoma.
Cancer Journal 19: 324–332. 65, 66
Buckley P, Lee M, Sul J, Miyashiro K, Bell T, Fisher S, Kim J, and Eberwine J. 2011.
Cytoplasmic intron sequence-retaining transcripts can be dendritically targeted
via ID element retrotransposons. Neuron 69: 877–884. 54
176
Buljan M, Chalancon G, Eustermann S, Wagner G, Fuxreiter M, Bateman A, and
Babu M. 2012. Tissue-specific splicing of disordered segments that embed bind-
ing motifs rewires protein interaction networks. Molecular Cell 46: 871–883. 11
Carrillo Oesterreich F, Preibisch S, and Neugebauer K. 2010. Global analysis of
nascent RNA reveals transcriptional pausing in terminal exons. Molecular Cell
40: 571–581. 112
Chow L, Gelinas R, Broker T, and Roberts R. 1977. An amazing sequence arrange-
ment at the 5’ ends of adenovirus 2 messenger RNA. Cell 12: 1–8. 6
Chow WH, Dong LM, and Devesa SS. 2010. Epidemiology and risk factors for
kidney cancer. Nature Reviews Urology 7: 245–57. 65
Cock P, Fields C, Goto N, Heuer M, and Rice P. 2010. The sanger FASTQ file format
for sequences with quality scores, and the Solexa/Illumina FASTQ variants.
Nucleic Acids Research 38: 1767–1771. 17
Costa V, Aprile M, Esposito R, and Ciccodicola A. 2013. RNA-Seq and human
complex diseases: recent accomplishments and future perspectives. European
Journal of Human Genetics 21: 134–42. 134
Cox TR and Erler JT. 2011. Remodeling and homeostasis of the extracellular ma-
trix: implications for fibrotic diseases and cancer. Disease models & mechanisms 4:
165–178. 86
Crick F. 1970. Central dogma of molecular biology. Nature 227: 561–563. 1
Cusnir M and Cavalcante L. 2012. Inter-tumor heterogeneity. Human Vaccines &
Immunotherapeutics 8: 1143–1145. 86
Darnell J. 2013. Reflections on the history of pre-mRNA processing and highlights
of current knowledge: a unified picture. RNA 19: 443–460. 4
Das S and Vikalo H. 2013. Base calling for high-throughput short-read sequencing:
dynamic programming solutions. BMC Bioinformatics 14: 129. 17
De Conti L, Baralle M, and Buratti E. 2013. Exon and intron definition in pre-
mRNA splicing. Wiley Interdisciplinary Reviews 4: 49–60. 12
Dees N, Zhang Q, Kandoth C, Wendl M, Schierding W, Koboldt D, Mooney T,
Callaway M, Dooling D, Mardis E, et al.. 2012. MuSiC: identifying mutational
significance in cancer genomes. Genome Research 22: 1589–1598. 91
van Dijk E, Jaszczyszyn Y, and Thermes C. 2014. Library preparation methods for
next-generation sequencing: Tone down the bias. Experimental Cell Research 322:
12–20. 15
177
Djebali S, Davis C, Merkel A, Dobin A, Lassmann T, Mortazavi A, Tanzer A, La-
garde J, Lin W, Schlesinger F, et al.. 2012. Landscape of transcription in human
cells. Nature 489: 101–108. 39, 53, 134
Domon B and Aebersold R. 2010. Options and considerations when selecting a
quantitative proteomics strategy. Nature Biotechnology 28: 710–721. 129
Ellis J, Miriam B, Colak R, Irimia M, Kim T, Calarco J, Wang X, Pan Q, Dave O,
Kim P, et al.. 2012. Tissue-specific alternative splicing remodels protein-protein
interaction networks. Molecular Cell 46: 884–892. 11
Emig D, Salomonis N, Baumbach J, Lengauer T, Conklin B, and Albrecht M. 2010.
AltAnalyze and DomainGraph: analyzing and visualizing exon expression data.
Nucleic Acids Research 38: W755–W762. 11
ENCODE Project Consortium, Bernstein B, Birney E, Dunham I, Green E, Gunter
C, and Snyder M. 2012. An integrated encyclopedia of DNA elements in the
human genome. Nature 489: 57–74. 1, 56, 133
Engstrom P, Steijger T, Sipos B, Grant G, Kahles A, Consortium R, Alioto T, Behr
J, Bertone P, Bohnert R, et al.. 2013. Systematic evaluation of spliced alignment
programs for RNA-seq data. Nature Methods 10: 1185–1191. 53
Ertel A, Verghese A, Byers S, Ochs M, and Tozeren A. 2006. Pathway-specific dif-
ferences between tumor cell lines and normal and tumor tissue cells. Molecular
Cancer 5. 82, 87
Ezkurdia I, del Pozo A, Frankish A, Rodriguez J, Harrow J, Ashman K, Valencia
A, and Tress M. 2012. Comparative proteomics reveals a significant bias toward
alternative protein isoforms with conserved structure and function. Molecular
Biology and Evolution 29: 2265–2283. 119, 123, 129, 135
Fica S, Tuttle N, Novak T, Li N, Lu J, Koodathingal P, Dai Q, Staley J, and Piccirilli
J. 2013. RNA catalyses nuclear pre-mRNA splicing. Nature 503: 229–234. 7
Flicek P, Amode M, Barrell D, Beal K, Brent S, Denise C, Clapham P, Coates G,
Fairley S, Fitzgerald S, et al.. 2012. Ensembl 2012. Nucleic Acids Research 40:
D84–D90. 39, 57, 75, 89, 113, 131
Floris M, Raimondo D, Leoni G, Orsini M, Marcatili P, and Tramontano A. 2011.
MAISTAS: a tool for automatic structural evaluation of alternative splicing prod-
ucts. Bioinformatics 27: 1625–1629. 11, 75
Fu X, Fu N, Guo S, Yan Z, Xu Y, Hu H, Menzel C, Chen W, Li Y, Zeng R, et al..
2009. Estimating accuracy of RNA-Seq and microarrays with proteomics. BMC
Genomics 10: 161. 117, 119
178
Fuda N, Ardehali M, and Lis J. 2009. Defining mechanisms that regulate RNA
polymerase II transcription in vivo. Nature 461: 186–192. 3
Galej W, Oubridge C, Newman A, and Nagai K. 2013. Crystal structure of prp8
reveals active site cavity of the spliceosome. Nature 493: 638–643. 95, 110
Gerlinger M, Rowan A, Horswell S, Larkin J, Endesfelder D, Gronroos E, Mar-
tinez P, Matthews N, Stewart A, Tarpey P, et al.. 2012. Intratumor heterogeneity
and branched evolution revealed by multiregion sequencing. The New England
Journal of Medicine 366: 883–892. 86
Gillet LC, Navarro P, Tate S, Röst H, Selevsek N, Reiter L, Bonner R, and Aebersold
R. 2012. Targeted data extraction of the MS/MS spectra generated by data-
independent acquisition: a new concept for consistent and accurate proteome
analysis. Molecular & Cellular Proteomics 11: O111.016717. 129
Gloria L, Covey P, Bedinger P, Mueller L, Thannhauser T, Zhang S, Fei Z, Gio-
vannoni J, and Rose J. 2012. Enabling proteomic studies with RNA-Seq: the
proteome of tomato pollen as a test case. Proteomics 12: 761–774. 119
Goldman N, Bertone P, Chen S, Dessimoz C, Emily L, Sipos B, and Birney E.
2013. Towards practical, high-capacity, low-maintenance information storage in
synthesized DNA. Nature 494: 77–80. 1
Gonzàlez-Porta M, Calvo M, Sammeth M, and Guigó R. 2012. Estimation of alter-
native splicing variability in human populations. Genome Research 22: 528–538.
89
Gonzàlez-Porta M and Brazma A. 2014. Identification, annotation and visualisa-
tion of extreme changes in splicing from rna-seq experiments with switchseq.
bioRxiv. 10.1101/005967. 78, 131
Grabherr M, Haas B, Yassour M, Levin J, Thompson D, Amit I, Adiconis X, Fan L,
Raychowdhury R, Zeng Q, et al.. 2011. Full-length transcriptome assembly from
RNA-Seq data without a reference genome. Nature Biotechnology 29: 644–652. 21
Grainger R and Beggs J. 2005. Prp8 protein: at the heart of the spliceosome. RNA
11: 533–557. 95, 110
Griebel T, Zacher B, Ribeca P, Raineri E, Lacroix V, Guigó R, and Sammeth M.
2012. Modelling and simulating generic RNA-Seq experiments with the flux
simulator. Nucleic Acids Research 40: 10073–10083. 40, 57
Grosso A, Gomes A, Nuno B, Caldeira S, Thorne N, Grech G, von Lindern M, and
Maria C. 2008. Tissue-specific splicing factor gene expression signatures. Nucleic
Acids Research 36: 4823–4832. 135
179
Hagen RM and Ladomery MR. 2012. Role of splice variants in the metastatic
progression of prostate cancer. Biochemical Society Transactions 40: 870–4. 63, 65
Haider S and Pal R. 2013. Integrated analysis of transcriptomic and proteomic
data. Current Genomics 14: 91–110. 119
Hansen KD, Brenner SE, and Dudoit S. 2010. Biases in Illumina transcriptome
sequencing caused by random hexamer priming. Nucleic Acids Research 38: e131.
17
Hardcastle TJ and Kelly KA. 2010. baySeq: empirical Bayesian methods for identi-
fying differential expression in sequence count data. BMC Bioinformatics 11: 422.
33
Harrow J, Frankish A, Gonzalez J, Tapanari E, Diekhans M, Kokocinski F, Aken
B, Barrell D, Zadissa A, Searle S, et al.. 2012. GENCODE: the reference human
genome annotation for the ENCODE project. Genome Research 22: 1760–1774. 11,
57, 133
Hebenstreit D, Fang M, Gu M, Charoensawan V, van Oudenaarden A, and Teich-
mann SA. 2011. RNA sequencing reveals two major classes of gene expression
levels in metazoan cells. Molecular Systems Biology 7: 497. 41, 58
Heinz S, Benner C, Spann N, Bertolino E, Lin Y, Laslo P, Cheng J, Murre C, Singh
H, and Glass C. 2010. Simple combinations of lineage-determining transcrip-
tion factors prime cis-regulatory elements required for macrophage and b cell
identities. Molecular Cell 38: 576–589. 114
Hemberg M, Prabakaran S, Chauhan R, Winter D, Tweedie-Cullen R, Dittrich C,
Hong E, Gunawardena J, Steen H, Steen J, et al.., Peptides from rnas classi-
fied as non-coding. Poster at the "Systems Biology: Global regulation of gene
expression meeting", CSHL. 53
Hofmann J, Husedzinovic A, and Gruss O. 2010. The function of spliceosome
components in open mitosis. Nucleus 1: 447–459. 95
Hooper JE. 2014. A survey of software for genome-wide discovery of differential
splicing in RNA-Seq data. Human Genomics 8: 3. 136
Hoskins A and Moore M. 2012. The spliceosome: a flexible, reversible macro-
molecular machine. Trends in Biochemical Sciences 37: 179–188. 95
House A and Lynch K. 2008. Regulation of alternative splicing: more than just the
ABCs. The Journal of Biological Chemistry 283: 1217–1221. 39, 111
Huang DW, Sherman B, and Lempicki R. 2009a. Bioinformatics enrichment tools:
paths toward the comprehensive functional analysis of large gene lists. Nucleic
Acids Research 37: 1–13. 59, 89, 115
180
Huang DW, Sherman B, and Lempicki R. 2009b. Systematic and integrative anal-
ysis of large gene lists using DAVID bioinformatics resources. Nature Protocols
4: 44–57. 59, 89, 115
International Human Genome Sequencing Consortium. 2001. Initial sequencing
and analysis of the human genome. Nature 409: 860–921. 1
International Human Genome Sequencing Consortium. 2004. Finishing the eu-
chromatic sequence of the human genome. Nature 431: 931–945. 1
Jonasch E, Futreal P, Davis I, Bailey S, Kim W, Brugarolas J, Giaccia A, Kurban G,
Pause A, Frydman J, et al.. 2012. State of the science: an update on renal cell
carcinoma. Molecular Cancer Research 10: 859–880. 65, 84
Kaida D, Schneider-Poetsch T, and Yoshida M. 2012. Splicing in oncogenesis and
tumor suppression. Cancer Science 103: 1611–6. 63, 65
Kalsotra A and Cooper TA. 2011. Functional consequences of developmentally
regulated alternative splicing. Nature Reviews Genetics 12: 715–29. 11
Kanehisa M and Goto S. 2000. KEGG: kyoto encyclopedia of genes and genomes.
Nucleic Acids Research 28: 27–30. 91
Kanehisa M, Goto S, Sato Y, Kawashima M, Furumichi M, and Tanabe M. 2014.
Data, information, knowledge and principle: back to metabolism in KEGG. Nu-
cleic Acids Research 42: D199–D205. 91
Katz Y, Wang E, Airoldi E, and Burge C. 2010. Analysis and design of RNA
sequencing experiments for identifying isoform regulation. Nature Methods 7:
1009–1015. 23, 40, 58, 90, 115
Keren H, Galit L, and Ast G. 2010. Alternative splicing and evolution: diversifica-
tion, exon definition and function. Nature Reviews Genetics 11: 345–355. 10, 11,
54, 65
Kim D, Pertea G, Trapnell C, Pimentel H, Kelley R, and Salzberg SL. 2013.
TopHat2: accurate alignment of transcriptomes in the presence of insertions,
deletions and gene fusions. Genome Biology 14: R36. 18, 19, 20, 26, 89
Kornblihtt A, Schor I, Alló M, Dujardin G, Petrillo E, and Muñoz M. 2013. Alter-
native splicing: a pivotal step between eukaryotic transcription and translation.
Nature Reviews Molecular Cell Biology 14: 153–165. 10
Kuehner J, Pearson E, and Moore C. 2011. Unravelling the means to an end: RNA
polymerase II transcription termination. Nature Reviews Molecular Cell Biology
12: 283–294. 4
Kurmangaliyev YZ and Gelfand MS. 2008. Computational analysis of splicing
errors and mutations in human transcripts. BMC Genomics 9: 13. 54
181
Kwak H and Lis JT. 2013. Control of transcriptional elongation. Annual Review of
Genetics 47: 483–508. 4
König J, Zarnack K, Rot G, Curk T, Kayikci M, Zupan B, Turner D, Luscombe
N, and Ule J. 2010. iCLIP reveals the function of hnRNP particles in splicing
at individual nucleotide resolution. Nature Structural & Molecular Biology 17:
909–915. 97, 112
Ladomery M. 2013. Aberrant alternative splicing is another hallmark of cancer.
International Journal of Cell Biology 2013: 463786. 14
Lane KR, Yu Y, Lackey PE, Chen X, Marzluff WF, and Cook JG. 2013. Cell cycle-
regulated protein abundance changes in synchronously proliferating HeLa cells
include regulation of pre-mRNA splicing proteins. PloS One 8: e58456. 135
Langmead B and Salzberg S. 2012. Fast gapped-read alignment with bowtie 2.
Nature Methods 9: 357–359. 19
Langmead B, Trapnell C, Pop M, and Salzberg SL. 2009. Ultrafast and memory-
efficient alignment of short DNA sequences to the human genome. Genome
Biology 10: R25. 57, 131
Lasda E and Blumenthal T. 2011. Trans-splicing. Wiley Interdisciplinary Reviews 2:
417–434. 12
Lee C and Roy M. 2004. Analysis of alternative splicing with microarrays: suc-
cesses and challenges. Genome Biology 5: 231. 136
Leoni G, Le Pera L, Ferrè F, Raimondo D, and Tramontano A. 2011. Coding po-
tential of the products of alternative splicing in human. Genome Biology 12. 119,
129, 135
Levental KR, Yu H, Kass L, Lakins JN, Egeblad M, Erler JT, Fong SF, Csiszar K, Gi-
accia A, Weninger W, et al.. 2009. Matrix crosslinking forces tumor progression
by enhancing integrin signaling. Cell 139: 891–906. 86
Levin J, Yassour M, Adiconis X, Nusbaum C, Thompson D, Friedman N, Gnirke
A, and Regev A. 2010. Comprehensive comparative analysis of strand-specific
RNA sequencing methods. Nature Methods 7: 709–715. 18
Li H, Handsaker B, Wysoker A, Fennell T, Ruan J, Homer N, Marth G, Abecasis
G, Durbin R, and Subgroup GPDP. 2009. The sequence Alignment/Map format
and SAMtools. Bioinformatics 25: 2078–2079. 20
Li J, Jiang H, and Wong WH. 2010. Modeling non-uniformity in short-read rates
in RNA-Seq data. Genome Biology 11: R50. 29
182
Li JJ, Bickel PJ, and Biggin MD. 2014. System wide analyses have underestimated
protein abundances and the importance of transcription in mammals. PeerJ 2:
e270. 54, 119, 129, 130
Li X, Zhang W, Xu T, Ramsey J, Zhang L, Hill R, Hansen KC, Hesselberth JR, and
Zhao R. 2013. Comprehensive in vivo RNA-binding site analyses reveal a role
of Prp8 in spliceosomal assembly. Nucleic Acids Research 41: 3805–18. 95, 98, 111
Li Y, Bor Y, Misawa Y, Xue Y, Rekosh D, and Hammarskjold M. 2006. An intron
with a constitutive transport element is retained in a tap messenger RNA. Nature
443: 234–237. 54
Linehan W, Srinivasan R, and Schmidt L. 2010. The genetic basis of kidney cancer:
a metabolic disease. Nature Reviews Urology 7: 277–285. 66, 86
Liu Y, Huttenhain R, Surinova S, Gillet L, Mouritsen J, Brunner R, Navarro P,
and Aebersold R. 2013. Quantitative measurements of n-linked glycoproteins in
human plasma by SWATH-MS. Proteomics 13: 1247–1256. 120, 129, 132
Love MI, Huber W, and Anders S. 2014. Moderated estimation of fold change and
dispersion for RNA-Seq data with DESeq2. bioRxiv. 10.1101/002832. 21, 28, 30
Lu P, Weaver VM, and Werb Z. 2012. The extracellular matrix: a dynamic niche in
cancer progression. The Journal of cell biology 196: 395–406. 86
Lukk M, Kapushesky M, Nikkila J, Parkinson H, Goncalves A, Huber W, Ukko-
nen E, and Brazma A. 2010. A global map of human gene expression. Nature
Biotechnology 28: 322–324. 39, 88
Lundberg E, Fagerberg L, Klevebring D, Matic I, Geiger T, Cox J, Algenas C, Lun-
deberg J, Mann M, and Uhlen M. 2010. Defining the transcriptome and pro-
teome in three functionally different human cell lines. Molecular Systems Biology
6: 450. 54, 119, 129
Maas S. 2012. Posttranscriptional recoding by RNA editing. Advances in protein
chemistry and structural biology 86: 193–224. 12
Maier T, Guell M, and Serrano L. 2009. Correlation of mRNA and protein in
complex biological samples. FEBS letters 583: 3966–3973. 119, 132
Malone JH and Oliver B. 2011. Microarrays, deep sequencing and the true measure
of the transcriptome. BMC Biology 9: 34. 14
Mardis ER. 2013. Next-generation sequencing platforms. Annual Review of Analyt-
ical Chemistry 6: 287–303. 14, 16, 18
Marinov G, Williams B, Ken M, Schroth G, Gertz J, Myers R, and Wold B. 2014.
From single-cell to cell-pool transcriptomes: stochasticity in gene expression
and RNA splicing. Genome Research 24: 496–510. 54
183
Marioni J, Mason C, Mane S, Stephens M, and Gilad Y. 2008. RNA-seq: an assess-
ment of technical reproducibility and comparison with gene expression arrays.
Genome Research 18: 1509–1517. 30
Martin J and Wang Z. 2011. Next-generation transcriptome assembly. Nature
Reviews Genetics 12: 671–682. 21
Martinez P, Birkbak N, Gerlinger M, Nicholas M, Burrell R, Rowan A, Joshi T,
Fisher R, Larkin J, Szallasi Z, et al.. 2013. Parallel evolution of tumour subclones
mimics diversity between tumours. The Journal of Pathology 230: 356–364. 86
Matera A and Wang Z. 2014. A day in the life of the spliceosome. Nature Reviews
Molecular Cell Biology 15: 108–121. 7, 8, 12, 13, 95, 135
McGlincy N and Smith C. 2008. Alternative splicing resulting in nonsense-
mediated mRNA decay: what is the meaning of nonsense? Trends in Biochemical
Sciences 33: 385–393. 11, 54, 65, 87
Melamud E and Moult J. 2009. Stochastic noise in splicing machinery. Nucleic
Acids Research 37: 4873–4886. 11, 55
Merkin J, Russell C, Chen P, and Burge C. 2012. Evolutionary dynamics of gene
and isoform regulation in mammalian tissues. Science 338: 1593–1599. 11, 55
Metzker M. 2010. Sequencing technologies - the next generation. Nature Reviews
Genetics 11: 31–46. 17
Michal C, Regina G, Siegfried Z, Andersen C, Thorsen K, Ørntoft T, Mu D, and
Karni R. 2013. The splicing factor SRSF6 is amplified and is an oncoprotein
in lung and colon cancers. The Journal of Pathology 229: 630–639. 86
Minoche AE, Dohm JC, and Himmelbauer H. 2011. Evaluation of genomic high-
throughput sequencing data generated on Illumina HiSeq and genome analyzer
systems. Genome Biology 12: R112. 19
Mortazavi A, Williams B, Kenneth M, Schaeffer L, and Wold B. 2008. Mapping
and quantifying mammalian transcriptomes by RNA-Seq. Nature Methods 5:
621–628. 14, 22, 27
Nagaraj N, Wisniewski J, Geiger T, Cox J, Kircher M, Kelso J, Paabo S, and Mann
M. 2011. Deep proteome and transcriptome mapping of a human cancer cell
line. Molecular Systems Biology 7: 548. 54, 119, 130
Neumann B, Walter T, Hériché J, Bulkescher J, Erfle H, Conrad C, Rogers P, Poser
I, Held M, Liebel U, et al.. 2010. Phenotypic profiling of the human genome by
time-lapse microscopy reveals cell division genes. Nature 464: 721–727. 95, 110
Nilsen T. 2003. The spliceosome: the most complex macromolecular machine in
the cell? BioEssays 25: 1147–1149. 6
184
Nilsen T and Graveley B. 2010. Expansion of the eukaryotic proteome by alterna-
tive splicing. Nature 463: 457–463. 10, 54, 55, 56, 65
Ning K and Nesvizhskii AI. 2010. The utility of mass spectrometry-based pro-
teomic data for validation of novel alternative splice forms reconstructed from
RNA-Seq data: a preliminary assessment. BMC Bioinformatics 11 Suppl 11: S14.
119
Padgett R. 2012. New connections between splicing and human disease. Trends in
Genetics 28: 147–154. 14, 63, 65
Pan Q, Shai O, Lee L, Frey B, and Blencowe B. 2008. Deep surveying of alternative
splicing complexity in the human transcriptome by high-throughput sequenc-
ing. Nature Genetics 40: 1413–1415. 10, 37, 39
Parenteau J, Durand M, Morin G, Gagnon J, Lucier J, Wellinger R, Chabot B, and
Elela S. 2011. Introns within ribosomal protein genes regulate the production
and function of yeast ribosomes. Cell 147: 320–331. 54
Park JW, Parisky K, Celotto AM, Reenan RA, and Graveley BR. 2004. Identification
of alternative splicing regulators by RNA interference in Drosophila. Proceedings
of the National Academy of Sciences 101: 15974–9. 135
Paule MR and White RJ. 2000. Survey and summary: transcription by RNA poly-
merases I and III. Nucleic Acids Research 28: 1283–98. 3
Pedrotti S and Cooper TA. 2014. In Brief: (mis)splicing in disease. The Journal of
Pathology 233: 1–3. 63, 65
Pérez-Ortín JE, Alepuz P, Chávez S, and Choder M. 2013. Eukaryotic mRNA decay:
methodologies, pathways, and links to other stages of gene expression. Journal
of Molecular Biology 425: 3750–75. 5, 54
Plass M and Eyras E. 2014. Approaches to link RNA secondary structures with
splicing regulation. Methods in Molecular Biology 1126: 341–356. 13
Porrua O and Libri D. 2013. RNA quality control in the nucleus: the angels’ share
of RNA. Biochimica et Biophysica Acta 1829: 604–611. 5, 54
Query CC and Konarska MM. 2004. Suppression of multiple substrate mutations
by spliceosomal prp8 alleles suggests functional correlations with ribosomal am-
biguity mutants. Molecular Cell 14: 343–54. 112
Quinlan AR and Hall IM. 2010. BEDTools: a flexible suite of utilities for comparing
genomic features. Bioinformatics 26: 841–2. 113
Ramakrishnan SR, Vogel C, Prince JT, Li Z, Penalva LO, Myers M, Marcotte EM,
Miranker DP, and Wang R. 2009. Integrating shotgun proteomics and mRNA
185
expression data to improve protein identification. Bioinformatics 25: 1397–403.
55, 130
Ramskold D, Wang ET, Burge CB, and Sandberg R. 2009. An abundance of ubiq-
uitously expressed genes revealed by tissue transcriptome sequence data. PLoS
Computational Biology 5: e1000598. 58
Roberts A, Trapnell C, Donaghey J, Rinn JL, and Pachter L. 2011. Improving RNA-
Seq expression estimates by correcting for fragment bias. Genome Biology 12:
R22. 29
Robertson MP and Joyce GF. 2012. The origins of the RNA world. Cold Spring
Harbor Perspectives in Biology 4. pii: a003608. 7
Robinson M, Davis M, and Smyth G. 2010. edgeR: a bioconductor package for
differential expression analysis of digital gene expression data. Bioinformatics
26: 139–140. 30, 33, 89
Robinson MD and Oshlack A. 2010. A scaling normalization method for differen-
tial expression analysis of RNA-seq data. Genome Biology 11: R25. 28
Rodriguez J, Maietta P, Ezkurdia I, Pietrelli A, Wesselink J, Lopez G, Valencia
A, and Tress M. 2013. APPRIS: annotation of principal and alternative splice
isoforms. Nucleic Acids Research 41: D110–D117. 11, 40, 41, 75, 87
Roux J and Robinson-Rechavi M. 2011. Age-dependent gain of alternative splice
forms and biased duplication explain the relation between splicing and dupli-
cation. Genome research 21: 357–363. 55
Safran M, Dalah I, Alexander J, Rosen N, Iny Stein T, Shmoish M, Nativ N, Bahir I,
Doniger T, Krug H, et al.. 2010. GeneCards Version 3: the human gene integrator.
Database 2010: baq020. 89, 132
Sakabe NJ and de Souza SJ. 2007. Sequence features responsible for intron reten-
tion in human. BMC Genomics 8: 59. 96, 111
Saltzman A, Pan Q, and Blencowe B. 2011. Regulation of alternative splicing by
the core spliceosomal machinery. Genes & Development 25: 373–384. 14, 86, 111,
112
Sato Y, Yoshizato T, Shiraishi Y, Maekawa S, Okuno Y, Kamura T, Shimamura T,
Aiko S, Nagae G, Suzuki H, et al.. 2013. Integrated molecular analysis of clear-
cell renal cell carcinoma. Nature Genetics 45: 860–867. 65, 78, 79, 84, 85, 86
Scelo G, Riazalhosseini Y, Greger L, Letourneau L, Gonzàlez-Porta M, et al..,
Whole-genome sequencing reveals variation in the genomic landscape of clear
cell renal cell carcinoma in europe. Under review in Nature Communications.
65, 84, 85
186
Schmieder R and Edwards R. 2011. Quality control and preprocessing of metage-
nomic datasets. Bioinformatics 27: 863–4. 89
Schoenberg D and Maquat L. 2012. Regulation of cytoplasmic mRNA decay. Na-
ture Reviews Genetics 13: 246–259. 6
Schwanhausser B, Busse D, Li N, Dittmar G, Schuchhardt J, Wolf J, Chen W, and
Selbach M. 2011. Global quantification of mammalian gene expression control.
Nature 473: 337–342. 2
Shalek AK, Satija R, Adiconis X, Gertner RS, Gaublomme JT, Raychowdhury R,
Schwartz S, Yosef N, Malboeuf C, Lu D, et al.. 2013. Single-cell transcriptomics
reveals bimodality in expression and splicing in immune cells. Nature 498: 236–
40. 54
Sharon D, Tilgner H, Grubert F, and Snyder M. 2013. A single-molecule long-read
survey of the human transcriptome. Nature Biotechnology 31: 1009–1014. 136
Shendure J and Ji H. 2008. Next-generation DNA sequencing. Nature Biotechnology
26: 1135–45. 136
Sheynkman G, Shortreed M, Frey B, and Smith L. 2013. Discovery and mass spec-
trometric analysis of novel splice-junction peptides using RNA-Seq. Molecular
& Cellular Proteomics 12: 2341–2353. 119, 129, 135
Shibata M and Shen M. 2013. The roots of cancer: stem cells and the basis for
tumor heterogeneity. BioEssays 35: 253–260. 86
Shionyu M, Takahashi Ki, and Go M. 2012. AS-EAST: a functional annotation tool
for putative proteins encoded by alternatively spliced transcripts. Bioinformatics
28: 2076–2077. 11
Shiroguchi K, Jia T, Sims P, and Xie X. 2012. Digital RNA sequencing mini-
mizes sequence-dependent bias and amplification noise with optimized single-
molecule barcodes. Proceedings of the National Academy of Sciences of the United
States of America 109: 1347–1352. 18
Singh J and Padgett R. 2009. Rates of in situ transcription and splicing in large
human genes. Nature Structural & Molecular Biology 16: 1128–1133. 110
Smith D, Query C, and Konarska M. 2008. "Nought may endure but mutability":
spliceosome dynamics and the regulation of splicing. Molecular Cell 30: 657–666.
39
de Sousa Abreu R, Penalva L, Marcotte E, and Vogel C. 2009. Global signatures of
protein and mRNA expression levels. Molecular bioSystems 5: 1512–1526. 119
187
van Staveren WCG, Solís DYW, Hébrant A, Detours V, Dumont JE, and Maenhaut
C. 2009. Human cancer cell lines: Experimental models for cancer cells in situ?
For cancer stem cells? Biochimica et Biophysica Acta 1795: 92–103. 82, 87
Steijger T, Abril J, Engstrom P, Kokocinski F, Consortium R, Abril J, Akerman M,
Alioto T, Ambrosini G, Antonarakis S, et al.. 2013. Assessment of transcript
reconstruction methods for RNA-seq. Nature Methods 10: 1177–1184. 53, 58
Tanackovic G, Ransijn A, Thibault P, Abou Elela S, Klinck R, Berson E, Chabot
B, and Rivolta C. 2011. PRPF mutations are associated with generalized de-
fects in spliceosome formation and pre-mRNA splicing in patients with retinitis
pigmentosa. Human Molecular Genetics 20: 2116–2130. 95, 135
Taneri B, Snyder B, and Gaasterland T. 2011. Distribution of alternatively spliced
transcript isoforms within human and mouse transcriptomes. J OMICS Res 1:
1–5. 39
Tang J, Lee J, Hou M, Wang C, Chen C, Huang H, and Chang H. 2013. Alternative
splicing for diseases, cancers, drugs, and databases. The Scientific World Journal
2013: 703568. 88
Tanner S, Shen Z, Ng J, Florea L, Guigó R, Briggs S, and Bafna V. 2007. Improving
gene annotation using peptide mass spectrometry. Genome Research 17: 231–239.
119
Tazi J, Bakkour N, and Stamm S. 2009. Alternative splicing and disease. Biochimica
et Biophysica Acta 1792: 14–26. 14, 65
TCGA Network. 2013. Comprehensive molecular characterisation of clear cell
renal cell carcinoma. Nature 499: 43–49. 65, 78, 79, 84, 85, 86
Tilgner H, Knowles D, Johnson R, Davis C, Chakrabortty S, Djebali S, Curado Ja,
Snyder M, Gingeras T, and Guigó R. 2012. Deep sequencing of subcellular RNA
fractions shows splicing to be predominantly co-transcriptional in the human
genome but inefficient for lncRNAs. Genome Research 22: 1616–1625. 13, 106, 111
Trapnell C, Hendrickson D, Sauvageau M, Goff L, Rinn J, and Pachter L. 2013.
Differential analysis of gene regulation at transcript resolution with RNA-seq.
Nature Biotechnology 31: 46–53. 33
Trapnell C, Pachter L, and Salzberg S. 2009. TopHat: discovering splice junctions
with RNA-Seq. Bioinformatics 25: 1105–1111. 57, 113
Trapnell C, Williams B, Pertea G, Mortazavi A, Kwan G, van Baren M, Salzberg S,
Wold B, and Pachter L. 2010. Transcript assembly and quantification by RNA-
Seq reveals unannotated transcripts and isoform switching during cell differen-
tiation. Nature Biotechnology 28: 511–515. 11, 18, 22, 23, 24, 26, 29, 33, 40, 58,
59
188
Tress M, Wesselink J, Frankish A, López G, Goldman N, Loytynoja A, Massingham
T, Pardi F, Whelan S, Harrow J, et al.. 2008a. Determination and validation of
principal gene products. Bioinformatics 24: 11–17. 56
Tress ML, Bodenmiller B, Aebersold R, and Valencia A. 2008b. Proteomics studies
confirm the presence of alternative protein isoforms on a large scale. Genome
Biology 9: R162. 10
Turro E, Astle WJ, and Tavaré S. 2014. Flexible analysis of RNA-seq data using
mixed effects models. Bioinformatics 30: 180–8. 33, 34, 78, 131
Turro E, Su SY, Gonçalves A, Coin LJM, Richardson S, and Lewin A. 2011. Haplo-
type and isoform specific expression estimation using multi-mapping RNA-seq
reads. Genome Biology 12: R13. 18, 22, 23, 24, 29, 33, 40, 58, 131
Turunen J, Niemea E, Verma B, and Frilander M. 2013. The significant other:
splicing by the minor spliceosome. Wiley Interdisciplinary Reviews 4: 61–76. 7
UniProt Consortium. 2012. Reorganizing the protein space at the universal protein
resource (UniProt). Nucleic Acids Research 40: D71–D75. 46, 56
Valletti A, Gigante M, Palumbo O, Carella M, Divella C, Sbisà E, Tullo A, Picardi E,
D’Erchia AM, Battaglia M, et al.. 2013. Genome-wide analysis of differentially
expressed genes and splicing isoforms in clear cell renal cell carcinoma. PloS
One 8: e78452. 66, 84
Vaquerizas J, Kummerfeld S, Teichmann S, and Luscombe N. 2009. A census of
human transcription factors: function, expression and evolution. Nature Reviews
Genetics 10: 252–263. 4
Venter JC, Adams MD, Myers EW, Li PW, Mural RJ, Sutton GG, Smith HO, Yandell
M, Evans CA, Holt RA, et al.. 2001. The sequence of the human genome. Science
291: 1304–1351. 1
Villalba A, Coll O, and Gebauer F. 2011. Cytoplasmic polyadenylation and trans-
lational control. Current Opinion in Genetics & Development 21: 452–457. 6
Vogel C and Marcotte E. 2012. Insights into the regulation of protein abundance
from proteomic and transcriptomic analyses. Nature Reviews Genetics 13: 227–
232. 2, 41, 54, 58
Volkmann G and Mootz H. 2013. Recent progress in intein research: from mech-
anism to directed evolution and applications. Cellular and molecular life sciences
70: 1185–1206. 12
Waks Z, Klein AM, and Silver PA. 2011. Cell-to-cell variability of alternative RNA
splicing. Molecular Systems Biology 7: 506. 39
189
Wang ET, Sandberg R, Luo S, Khrebtukova I, Zhang L, Mayr C, Kingsmore SF,
Schroth GP, and Burge CB. 2008. Alternative isoform regulation in human tissue
transcriptomes. Nature 456: 470–476. 10, 11, 37, 39
Wang J, Duncan D, Shi Z, and Zhang B. 2013. WEB-based GEne SeT AnaLysis
Toolkit (WebGestalt): update 2013. Nucleic Acids Research 41: W77–83. 59, 89,
115
Wang X, Wu Z, and Zhang X. 2010. Isoform abundance inference provides a more
accurate estimation of gene expression levels in RNA-seq. Journal of Bioinformat-
ics and Computational Biology 8 Suppl 1: 177–192. 25
Wang Z and Burge C. 2008. Splicing regulation: from a parts list of regulatory
elements to an integrated splicing code. RNA 14: 802–813. 111
Wang Z, Gerstein M, and Snyder M. 2009. RNA-Seq: a revolutionary tool for
transcriptomics. Nature Reviews Genetics 10: 57–63. 2, 14
Wickramasinghe V, Gonzàlez-Porta M, et al.., Prpf8 abundance dictates the pat-
terns of alternative and constitutive messenger rna splicing. Submitted. 112
Will CL and Luhrmann R. 2011. Spliceosome structure and function. Cold Spring
Harbor Perspectives in Biology 3. pii: a003707. 7, 10
Wong J, Ritchie W, Ebner O, Selbach M, Wong J, Huang Y, Gao D, Pinello N,
Gonzalez M, Baidya K, et al.. 2013a. Orchestrated intron retention regulates
normal granulocyte differentiation. Cell 154: 583–595. 54
Wong KH, Jin Y, and Moqtaderi Z. 2013b. Multiplex Illumina sequencing using
DNA barcoding. Current Protocols in Molecular Biology / edited by Frederick M.
Ausubel et al. Chapter 7: Unit 7.11. 18
Wu P, Phan J, and Wang M. 2013. Assessing the impact of human genome annota-
tion choice on RNA-seq expression estimates. BMC Bioinformatics 14 Suppl 11.
57
Xing X, Li Q, Sun H, Fu X, Zhan F, Huang X, Li J, Chen C, Shyr Y, Zeng R, et al..
2011. The discovery of novel protein-coding features in mouse genome based
on mass spectrometry data. Genomics 98: 343–351. 119
Yan L, Yang M, Guo H, Yang L, Wu J, Li R, Liu P, Lian Y, Zheng X, Yan J, et al..
2013. Single-cell RNA-Seq profiling of human preimplantation embryos and
embryonic stem cells. Nature Structural & Molecular Biology 20: 1131–1139. 54
Yap K, Lim Z, Khandelia P, Friedman B, and Makeyev E. 2012. Coordinated regula-
tion of neuronal mRNA steady-state levels through developmentally controlled
intron retention. Genes & Development 26: 1209–1223. 11, 54, 65, 87
190
Yeo G and Burge C. 2004. Maximum entropy modeling of short sequence motifs
with applications to RNA splicing signals. Journal of Computational Biology 11:
377–394. 114
Yu Y, Maroney P, Denker J, Zhang X, Dybkov O, Luhrmann R, Jankowsky E,
Chasin L, and Nilsen T. 2008. Dynamic regulation of alternative splicing by
silencers that modulate 5’ splice site competition. Cell 135: 1224–1236. 112
Zhao Q, Caballero O, Davis I, Jonasch E, Tamboli P, Yung W, Weinstein J, for
TCGA research network KS, Strausberg R, and Yao J. 2013. Tumor-specific iso-
form switch of the fibroblast growth factor receptor 2 underlies the mesenchy-
mal and malignant phenotypes of clear cell renal cell carcinomas. Clinical Cancer
Research 19: 2460–2472. 66, 75, 84
Zhao S, Fung-Leung W, Bittner A, Ngo K, Liu X, and Zhang S. 2014. Comparison
of RNA-Seq and microarray in transcriptome profiling of activated T cells. PLoS
One 9: e78644. 136
Zhou A, Zhang F, and Chen JY. 2010. PEPPI: a peptidomic database of human
protein isoforms for proteomics experiments. BMC Bioinformatics 11 Suppl 6:
S7. 119
Zhuang F, Fuchs RT, and Robb GB. 2012. Small RNA expression profiling by
high-throughput sequencing: implications of enzymatic manipulation. Journal
of Nucleic Acids 2012: 360358. 16
191