Major depressive disorder (MDD) is a highly prevalent and disabling condition with a complex pathophysiology that has not been fully elucidated to date. While the socioeconomic burden of the disease is significant, many individuals remain undiagnosed or misdiagnosed. This is largely because the current diagnostic approach that relies on clinical evaluations of signs and symptoms can be subjective, and time and resources tend to be rather limited in primary care where the majority seek help for depression. Therefore, there is a significant and pressing need for an objective, reliable and readily accessible diagnostic test to enable earlier and more accurate diagnosis of MDD. In particular, as individuals experiencing subthreshold levels of depressive symptoms have an increased risk of developing MDD, it would be clinically relevant for such a diagnostic test to be able to identify depressed patients and/or individuals with high risks of incident MDD among symptomatic individuals.

This thesis sought to develop risk prediction models that could potentially be utilised within a clinical setting to facilitate earlier and more accurate diagnosis of MDD. Such models were used to obtain probability estimates of the investigated individuals having or developing MDD based on their blood-based proteomic profiles and other characteristics, including sociodemographic and lifestyle factors. A targeted mass spectrometry approach was used to measure the abundances of a panel of peptides representing proteins, many of which have been previously associated with psychiatric disorders. Biomarkers were investigated in serum samples, which are widely used for blood-based biomarker discovery, as well as in dried blood spot samples, which are relatively novel in the field and carry several advantages. Importantly, this thesis focused on adopting appropriate statistical methods to ensure that the diagnostic predictions made by the models were accurate and reproducible, by addressing problems of model overfitting and model selection uncertainty. A particularly significant aspect of this was the development and application of a multimodel-based approach combining feature extraction and model averaging, which resulted in improved model predictive performance and generalisability.

Diagnostic prediction models based on serum proteomic, sociodemographic/lifestyle and clinical data were shown to be able to differentiate between subthreshold symptomatic individuals who developed and did not develop MDD. Additionally, diagnostic prediction models based on dried blood spot proteomic and digital mental health assessment data were shown to be able to identify currently depressed patients without an existing MDD diagnosis as well as currently not depressed patients with an existing MDD diagnosis among subthreshold symptomatic individuals. These results clearly demonstrate the potential of such prediction models to be used as an aid to the diagnosis of MDD in clinical practice, especially within the primary care setting. Moreover, MDD was found to be associated with several blood-based proteomic biomarkers, which mainly represented an immune/inflammatory profile, as well as with various other patient features, most notably body mass index and childhood trauma. Although further investigations are needed, these associations reveal disturbances in the stress response pathways involving the hypothalamic-pituitary-adrenal axis in the pathophysiology of depression.