Search

Understanding the etiology of complex disease remains a challenge in biology. In recent years there has been an explosion in biological data, this study investigates machine learning and network analysis methods as tools to aid candidate disease gene prioritisation, specifically relating to hypertension and cardiovascular disease. This thesis comprises four sets of analyses: Firstly, non synonymous single nucleotide polymorphisms (nsSNPs) were analysed in terms of sequence and structure based properties using a classifier to provide a model for predicting deleterious nsSNPs. The degree of sequence conservation at the nsSNP position was found to be the single best attribute but other sequence and structural attributes in combination were also useful. Predictions for nsSNPs within Ensembl have been made publicly available. Secondly, predicting protein function for proteins with an absence of experimental data or lack of clear similarity to a sequence of known function was addressed. Protein domain attributes based on physicochemical and predicted structural characteristics of the sequence were used as input to classifiers for predicting membership of large and diverse protein superfamiles from the SCOP database. An enrichment method was investigated that involved adding domains to the training dataset that are currently absent from SCOP. This analysis resulted in improved classifier accuracy, optimised classifiers achieved 66.3% for single domain proteins and 55.6% when including domains from multi domain proteins. The domains from superfamilies with low sequence similarity, share global sequence properties enabling applications to be developed which compliment profile methods for detecting distant sequence relationships. Thirdly, a topological analysis of the human protein interactome was performed. The results were combined with functional annotation and sequence based properties to build models for predicting hypertension associated proteins. The study found that predicted hypertension related proteins are not generally associated with network hubs and do not exhibit high clustering coefficients. Despite this, they tend to be closer and better connected to other hypertension proteins on the interaction network than would be expected by chance. Classifiers that combined PPI network, amino acid sequence and functional properties produced a range of precision and recall scores according to the applied 3 weights. Finally, interactome properties of proteins implicated in cardiovascular disease and cancer were studied. The analysis quantified the influential (central) nature of each protein and defined characteristics of functional modules and pathways in which the disease proteins reside. Such proteins were found to be enriched 2 fold within proteins that are influential (p<0.05) in the interactome. Additionally, they cluster in large, complex, highly connected communities, acting as interfaces between multiple processes more often than expected. An approach to prioritising disease candidates based on this analysis was proposed. Each analyses can provide some new insights into the effort to identify novel disease related proteins for cardiovascular disease.

Depression is a prevalent and severe mental health disorder that is one of the leading causes of disability worldwide. It can cause various physical and psychological problems, leading to loss of productivity, increased social burden, and even suicide. The current diagnosis of depression relies on skilled clinicians and self-reported questionnaires, which have limitations including subjective recall bias and loss of day-to-day fluctuation information. As a result, the majority of individuals with depression did not receive timely and effective treatment. Therefore, there is a need for more effective auxiliary techniques for recognizing and monitoring depression. With the development and widespread use of sensors, mobile technology provides a cost-effective and convenient means for gathering individuals' behavioral data related to depression symptoms. Several past studies have attempted to monitor depression using mobile phones and wearable devices. However, the majority of these studies were conducted on relatively small and homogeneous cohorts with short follow-up periods, which may have limited the generalizability of their findings. Furthermore, the impact of participant attrition and engagement, the direction of relationships over time, and individual differences need further exploration. To address these limitations, this thesis extracts a variety of behavioral features from multiple data streams of mobile phone and wearable data and explores their associations with depression symptom severity using a large, longitudinal, multi-center data set. Specifically, Chapter 1 provides an overview of the background of depression, motivations for using mobile technology for depression monitoring, and existing related studies. Chapter 2 performs a novel investigation into long-term participant retention and engagement from a European longitudinal observational program, the RADAR-MDD study, which is used throughout the whole thesis. A significantly higher participant retention rate is found in the RADAR-MDD study than in previous remote digital health studies. According to the data-driven method, lower participant engagement is found to be associated with higher depression symptom severity, younger age, and longer questionnaire response/completion time in the study app. Finally, the strategies for increasing participant engagement in future digital health research are also discussed in this chapter. Next, the associations between depression symptom severity and various categories of behaviors are explored separately in the following chapters: sleep (Chapter 3), sociability as measured by Bluetooth device counts (Chapter 4), mobility (Chapter 5), daily walking (Chapter 6), and circadian rhythms (Chapter 7). These associations are examined using multilevel models that incorporate demographics as between-participant covariates. A number of significant associations between behavioral characteristics and depression symptom severity are found in these chapters. For example, higher depression severity is significantly associated with worse sleep, lower sociability, lower mobility, slower cadence of daily walking, and weaker circadian rhythmicity. Notably, the longitudinal association between mobility and depression over time is assessed using dynamic structural equation models in Chapter 5. Changes in several mobility features are found to significantly affect subsequent changes in depression severity. Furthermore, daily-life gait patterns are found to provide extra information for recognizing depression relative to laboratory gait patterns in Chapter 6. Taken together, the findings in this thesis demonstrate that depression is closely associated with individuals' daily-life behaviors, which can be captured by mobile technology in real-world settings. Despite challenges of data quality and participant attrition, the evidence may provide support for the development of future clinical tools to passively monitor mental health status and trajectory with minimal burden on the participant.

Background: Clozapine is the only evidence-based medication for treatment-resistant schizophrenia. However, clozapine is severely under-prescribed mainly because of clozapine-induced agranulocytosis, an adverse drug reaction of clozapine that occurs in 0.4% of clozapinetreated patients. Since there are no clinical predictors for clozapine-induced agranulocytosis, all clozapine users are required to be under strict blood test monitoring throughout the duration of their clozapine treatment. Aims of this thesis: To use data from electronic health records to test the following hypothesis: (i) Predictors for clozapine-induced agranulocytosis can be investigated using the results of the analysed electronics health records data published in (Iqbal et al., 2020); (ii) The frequency density of clozapine blood test results that indicate clozapineinduced agranulocytosis risk changes with clozapine treatment time; (iii) Clozapine-treated patients are at an increased risk of COVID-19 infection; (iv) Clozapine-treated patients are at an increased risk of severe outcomes of COVID-19 infection. Methods: All data were extracted from Clinical Record Interactive Search (CRIS), the de-identified electronic health records of South London and Maudsley NHS Foundation Trust (SLAM). All data extraction was performed using SQL. All analysis was performed using either R, python or STATA. The statistical methods used in this thesis are logistic regression, survival analysis and Cox proportional hazard models. Results: We found that the data from (Iqbal et al., 2020) was not informative for investigating predictors for clozapine-induced agranulocytosis, thus this study did not bear any significant results. However, it helped us to realise that the next step was to study the patterns of clozapine blood monitoring data. From studying the patterns in clozapine blood monitoring results, we showed that the highest risk of clozapine-induced agranulocytosis is in the early months of treatments. The Kaplan-Meier survival curve and the incidence rates analysis showed that 75% of blood test results that indicated clozapine-induced agranulocytosis risk occurred within the first 6 months of clozapine treatment. At the onset of the COVID-19 pandemic, we investigated the associations between clozapine treatment and increased risk of COVID-19. We found that clozapine-treated patients had an increased risk of COVID-19 compared with those who were on other antipsychotic medication (unadjusted HR = 2.62, 95% CI 1.73 - 3.96), which was attenuated after adjusting for potential confounders, including clinical contact (adjusted HR=1.76, 95% CI 1.14 - 2.72). We followed up on the previous study to investigate the associations between clozapine treatment and increased risk of severe outcomes of COVID-19, namely COVID-related hospitalisation, COVID-related intensive care treatment, and death. We found that even though clozapine treatment appears to increase the risk of COVID-19 infection, it does not increase the risk of the severe outcomes of COVID-19. Conclusion: In conclusion, electronic health records are a valuable resource for studying clozapine health outcomes. In particular, the CRIS data is a very informative resource for answering research questions related to mental health disorders.

Whilst the pathological hallmarks of Alzheimer's Disease (AD) are well established, there continues to exist a critical lack in understanding of the mechanisms which give rise to the disease. Existing biomarkers from both brain and cerebral spinal fluid (CSF) demonstrate clinical utility, but these are limited by cost and invasiveness, and as such, offer limited accessibility both for treating and tracking AD at scale. Exposures like low educational attainment are also consistently associated with AD, but without understanding the biological pathways through which such relationships arise, the extent to which this knowledge can inform treatment progress is negligible. Blood metabolites are small molecules which are easily accessible and reflective of the interplay between biology and the wider environment. Consisting of molecules such as lipids and amino acids, they also hold biological relevance to AD. Identifying metabolites with aetiological relevance to AD and disentangling how these associate with exposures along the AD causal pathway could, therefore, offer unique insights into underlying mechanisms of the disease. With this in mind, this thesis aimed to elucidate blood metabolites with mechanistic relevance to AD and integrate knowledge from both genetic and cognition-based exposures, to further understand interlinking pathways into the disease. Study one began by utilising cross-trait linkage disequilibrium score regression (LDSC) to compare phenotyping approaches across both AD and blood metabolite genomic data to establish appropriate measures for use in downstream analyses. Here, the trade-off between specificity and sample size was explored. For AD, strict phenotyping based on clinical AD status, and minimal phenotyping based on parental AD were compared. For metabolites, quantification via small sampled high precision methodology (mass-spectrometry) versus a larger sampled, lower precision method (nuclear magnetic resonance spectroscopy) were also compared. Strict phenotyping was favoured for AD, whilst larger sample sizes were favoured for metabolites. Study two then explored causal relationships between blood metabolites and clinical AD through use of Mendelian randomisation (MR). Using knowledge from existing associations with midlife cognition, 19 candidate blood metabolites were selected to investigate whether mid-life associations translate through to later AD risk. Univariable MR was used to assess the bi-directional causal association between each metabolite and AD in-turn, and Bayesian methodology interrogated metabolite combinations, which may together be on the causal pathway to AD. Glycoprotein Acetyls demonstrated a risk-increasing causal association with AD, whilst a number of high density lipoproteins (HDLs), particularly Free Cholesterol in Extra Large HDLs (XL.HDL.FC), showed protective causal associations. Study three extended MR causal analyses, but made use of cross-trait polygenic risk scoring (PRS) to prioritise AD-specific candidate metabolites, of which 34 were selected. Cognition and educational attainment were also introduced to disentangle independent and mediated relationships. Univariable MR was first used to interrogate bidirectional causal relationships between metabolites, education, cognition, and AD in-turn, and multivariable MR was then used to disentangle independent versus mediating mechanisms. Glutamine and XL.HDL.FC showed evidence of protective causal effects on AD, as did educational attainment and cognition. No evidence of metabolites mediating the effect of either education or cognition on AD was found, though cognition fully mediated the effect of educational attainment. Finally, study four harnessed the use of longitudinal data to further understand possible mediating roles of blood metabolites in the relationship between educational attainment and AD, using AD endophenotypes in-place of clinical diagnosis. 118 longitudinal metabolites and four longitudinal AD endophenotypes (hippocampal volume, FDG-PET, MMSE scores, and plasma P-tau181) were utilised, and both data reduction and latent growth curve modelling analysed mediation at the single analyte and group-metabolite level. No robust evidence of metabolite mediators were found. Taken together, results from this thesis offer novel insights into blood metabolites with causal relevance to AD and highlight the importance of measurement consideration in optimising power to detect relationships. No robust evidence was found for a mediating role of blood metabolites in the context of cognitive factors and AD. However, a number of metabolites are offered as direct causal candidates for future study, and methodological pipelines pave the way for wider mediating relationships to be investigated.

This thesis examines the longitudinal changes in cognition for patients with schizophrenia using free text from medical records. To this end, we introduce a unified framework to extract, annotate, classify and analyse cognitive impairments from unstructured text, evaluate symptom trajectories and measure their association with socio-demographic factors and clinical outcomes. The framework was further extended to any type of symptom that can be defined using a list of keywords, and allows easy implementation and deployment within the Clinical Record Interactive Search (CRIS) system, which provides researchers with regulated and secure access to anonymised information from clinical records. A standardized approach to extract and annotate portions of unstructured text relevant to cognitive impairments was developed in conjunction with clinicians and researchers. This annotated dataset was then used to train text classification algorithms, in order to separate affirmed versus irrelevant or negated mentions of cognitive symptoms. An extensive comparative study, looking at existing text classification methods within the biomedical as well as general domains was conducted on both public and internal datasets. The results showed that transformer-based approaches, which are the current state of the art for many natural language processing tasks, outperform other methods in terms of accuracy, ease of implementation and scalability, particularly when trained on a combination of general and medical data. This text classification model was subsequently used to derive cognitive score time series from the free text of medical records. This "digital signature" of cognitive changes was in turn validated against scores obtained from clinically administered tests, confirming the accuracy and reliability of the model. Symptom trajectories were then evaluated using mixed linear models, again comparing the results obtained with the transformer model against standardized instruments. Both approaches demonstrated similar rates of change, indicating a gradual cognitive decline with age, which is attenuated by certain socio-demographic factors such as education, employment or marital status. The transformer-based model highlighted a strong association between education and cognition, showing that certain cognitive impairments, specifically attention and social cognition, were more likely to be reported early for patients with a higher education level. The relationship between cognition and clinical outcomes was also analysed, indicating that cognitive problems are correlated with adverse outcomes. This supports the findings in the literature that these symptoms account for much of the disability associated with schizophrenia. Finally, the text classification framework was tested and generalized to cover other symptoms and patient groups, allowing the development of a standardized set of tools that were then deployed within health research settings. This formed the basis for other research, notably COVID-related projects, which involved extracting mentions of anxiety and violent behaviour from the free text of clinical records, paving the way to further clinical applications. The contribution of this research is both methodological and practical. The use of a novel symptom extraction, classification and analysis framework demonstrates that cognitive impairments can be reliably harvested from the free text of medical records using deep learning models. The framework shows that these impairments are common in patients with schizophrenia and are correlated with adverse clinical outcomes. It provides a scalable and adaptable means of conducting research using large, unstructured datasets, typical of the vast amount of data routinely collected in clinical records. Such automated tools can be utilized to detect early impairments, screen individuals and identify those who would benefit from more comprehensive assessments, and ultimately support real-time clinical decision making.

Background: In recent years, wearable devices applied to seizure detection have progressively become available. The goal of this thesis was to provide insight into the role of wearable non-EEG seizure detection devices in epilepsy, with a specific focus on two aspects often disregarded by the current literature: focal onset seizures and patients' direct experiences. Methods: This thesis makes use of methods that can be grouped into four major categories: qualitative data (online survey, focus group); review and analysis of the literature (including narrative review of the literature, systematic review, meta-analysis, meta-regression); experimental study (digital semiology of focal seizures with motor manifestations; acceptability of the device; assessment of seizure-related risk); quantitative data analysis. Results: The results are summarised in 6 main points: 1. People with epilepsy, their caregivers and healthcare providers are interested in the use of preferably small, commonly used, multimodal wearable devices to acquire information in the attempt to mitigate seizure unpredictability; 2. Current devices do not exactly match patients' needs and mainly focus on tonic-clonic seizure detection; 3. Wearable devices should incorporate HR signals to early detect focal onset seizures; 4. Focal onset seizures with motor manifestations exhibit a common digital semiology and evolution pattern characterized by early cardiac manifestations followed by motor phenomena and final electrodermal activity response; 5. Patients acceptability of wearable devices is good as well as self-management; 6. Wearable devices can help at identifying potentially life-threating post-ictal states such as prolonged immobility. Conclusion: We anticipate that focal onset seizure detection will open new avenues to improve both the safety and treatment of people with epilepsy, with our imagination being the only limiting factor.

The use of Electronic Health Records (EHRs) in recording the details of patient interactions with healthcare services has generated large amounts of data with great potential for secondary usage in research. However, although the vast information available offers opportunities to improve care by learning from similar patients in parallel situations, there are great challenges in extracting correct and contextually meaningful knowledge due to the free-text, unstandardised and uncertainty-ridden form of clinical text. The focus of the presented work has been on detecting concepts related to Adverse Drug Events (ADEs) from the EHR using Natural Language Processing (NLP) tools to transform the unstructured text into semantically meaningful annotated knowledge. Specifically, this thesis explored the potential of NLP to identify ADEs from mental health EHRs in order to understand how drugs are working in real-world settings, to complement the current body of knowledge from clinical trials. Four studies were performed on the EHRs of the South London and Maudsley (SLAM) NHS Foundation Trust, with some analyses further performed on two other large psychiatric NHS Trusts: Camden & Islington (C&I) NHS Foundation Trust and the Oxford Health (Oxford) NHS Foundation Trust. The first study presented means to identify ADEs within an EHR, with a use case in identifying patients who have experienced Extra-Pyramidal Side Effects (EPSEs) at any point and achieved an overall 0.85 precision and 0.86 recall. The second study focused on anchoring ADEs to a point in time and achieved 0.89 precision and 0.86 recall in SLAM and 0.84 precision and 0.87 recall in C&I, contributing to the third study, which built a complete view of the patient medication and Adverse Drugs Reaction (ADR) profile. These methods were applied to study the side effect profile of Clozapine, a potent antipsychotic, in the three large mental health hospitals.

The increase in life expectancy has profoundly increased the ageing population, which, unfortunately, is also accompanied by a rise in age-related disorders, including Alzheimer’s Disease (AD). The most common form of dementia is AD, which was first described over a century ago, however, to date, there still exists a lack of understanding the molecular changes specific to the disease, a clinically established robust blood-based biomarker for accurate disease diagnosis and a lack of treatments. This thesis begins by investigating microarray gene expression profiling from Asymptomatic AD (AsymAD) human brains, who were clinically free from dementia; however, upon autopsy, they were observed to consist of hallmark AD pathology. A significant increase of transcriptomic activity in the frontal cortex (FC) brain region of AsymAD subjects was detected, suggesting fundamental changes in AD may initially begin within the FC brain region prior to symptoms of AD. In addition, overactivation of the brain “glutamate-glutamine cycle” and disruption to the brain energy pathways in both AsymAD and AD subjects were identified, suggesting these may be the earliest biological pathways disrupted in the disease, providing potential targets for early disease intervention. Secondly, existing and novel microarray gene expression studies of human AD brains were integrated into the largest known AD meta-analysis to date and is the first to incorporate numerous non-AD neurological disorders to identify AD-specific molecular changes. Seven genes were observed to be specifically and consistently perturbed across AD brains, with SPCS1 gene expression pattern found to replicate in RNA-Seq data. The cerebellum brain region is often regarded to be free from hallmark AD pathology and was incorporated into the analysis as a secondary control to identify an additional nineteen genes that may be involved with hallmark AD pathology. Furthermore, biological processes often reported as disrupted in AD were observed to be tissue-specific, and viral components were found to be specifically enriched across AD brains. Thirdly, an automated transcriptomic based drug repositioning pipeline was developed to query the reprocessed Connectivity Map to identify candidate compounds for disease intervention. Drug repositioning the AsymAD gene expression profile identified several candidate compounds that are already FDA approved for the treatment of AD and cognitive impairment, suggesting these compounds may be effective in the early stage of the disease. Drug repositioning the AD gene expression profile identified an anti-biotic compound for disease intervention. Finally, a machine learning approach was used to identify a blood-based 28 gene expression profile, which is enriched for “herpes simplex infection”, and can distinguish AD from Parkinson’s Disease, Multiple Sclerosis, Amyotrophic Lateral Sclerosis, Bipolar Disorder, Schizophrenia, Coronary Artery Disease, Rheumatoid Arthritis, Chronic Obstructive Pulmonary Disease, and cognitively healthy subjects with 66.3% PPV and 90.6% NPV. Overall, the work undertaken in this thesis provides new insight into the molecular changes occurring in both the asymptomatic and symptomatic phase of the disease, demonstrates a framework for a possible blood-based transcriptomic diagnosis test, provides new potential therapeutic targets, identifies candidate compounds that require further investigation for disease intervention and provides new avenues for future AD-related research.

Ageing is associated with changes in cellular and molecular processes including the alteration of stem cell pools. In particular, biological and functional changes observed in ageing neural stem cells (NSCs), are linked to age-related cognitive decline. Recently, the systemic environment has been shown to alter both NSC regulation and age-related cognitive decline. Interestingly, a well-documented and naturally occurring way of altering the composition of the systemic environment is through diet and nutrition. Studies have found an overabundance of nutrients to be detrimental for human and animal health; the presence of specific nutrients as well as the overall increase in calorie or protein intake was shown to overstimulate conserved molecular pathways and to reduce lifespan. Conversely, dietary restriction was found to be the most efficient way of extending an organism's lifespan. In this study, we examined nutrient sensing pathways in relation to their function in NSCs, ageing and cognition. We focus on the Sirtuin, mTOR and Insulin / Insulin like growth factor- 1 pathways and employ both in vitro and epidemiological methods to assess their contribution to age-related phenotypes. Using a human hippocampal progenitor cell line (HPC) we modelled ageing through treatment with ageing human serum and via pharmacological interventions combined with increased passage number. A semiautomated imaging platform was used for the immunocytochemical quantification of NSC proliferation, differentiation, damage and apoptosis. Further to this, candidate gene selection was preformed through literature search and validated by qPCR to measure gene expression alterations within our ageing models. Results from the in vitro experiments, were used to inform the selection of 9 genes belonging to nutrient-sensing pathways as candidate genes for a role in NSC regulation. Next, we investigated the effects of lifestyle and candidate-gene genotype on cognition using 1633 participants from the adult longitudinal population-based TwinsUK cohort. We report that treatment with human serum is able to induce changes in the HPC model that relate to both the serum-donor's hippocampal volumes and their cognitive performance. In addition, we report interesting associations between candidate-gene expression and NSC regulation. Our data also revealed increasing passage number causes significant alterations in nutrient-sensing genes and emphasised FOXO3A, NAMPT, PTEN, GRB10 and mTOR as interesting candidates for further ageing research. Finally, epidemiological analysis showed a significant effect of lifestyle on cognition and highlighted associations between SNPs in SIRT1 and ABTB1 and cognitive performance. The results outlined in this thesis support an important role for nutrient-sensing pathways in human NSCs ageing. In addition, we provide support for the use and further development of novel in vitro models to investigate NSCs ageing and to validate existing pathways, uncover new targets and test novel therapies.

This thesis is an exploration of a problem that exists between cutting edge Natural Language Processing (NLP) methodologies and their real world exploitation in clinical research. I detail the development and validation of a range of NLP methodologies on clinical records, with a specific focus on the case of the symptomatology of Serious Mental Illness (SMI). This publication based thesis covers five main themes: Pre-work to describe the eld of NLP within the context of clinical data. The proposition, development and evaluation of the TextHunter desktop application, a suite of high-throughput tools to overcome bottlenecks in the development of NLP applications. The application of the tools to the novel domain of SMI symptomatology, enabling the development of language models for 46 symptom concepts with a median F1 score of 0.87, and enabling the pro ling of symptom distribution amongst 7 962 patients, based on discharge summaries. A knowledge discovery project using artificial neural networks and clustering techniques, to identify real world patterns of symptom depiction in clinical free text. Here, I demonstrate a granularity and diversity of vocabulary beyond what is described in standard clinical terminologies. A commentary on the realities of text analytics in the NHS, and the development of a software architecture 'CogStack' to address these. This culminated in the establishment of the Clinical Analytics Platform at King's College Hospital.

Treatment-resistant schizophrenia is diagnosed after unsuccessful trials of two or more antipsychotics. Clozapine has demonstrated efficacy in treatment-resistant patients but requires extensive monitoring to attain therapeutic plasma concentrations and prevent adverse effects. As a result, there is great clinical demand for characterization of intra-individual and population level variation in the progression of weight gain and for the attainment of therapeutically active clozapine plasma concentrations. Identifying such factors could stratify patients into homogeneous groups to receive targeted treatment interventions to improve patient adherence and prevent the risk of relapse or toxicity related adverse events. This thesis utilizes two cohorts; a clinical trial population of individuals with psychosis on a variety of medications and a large UK-based cohort of clozapine treated patients derived from therapeutic monitoring records. These cohorts are used to explore the contributory factors to psychotropic medication induced weight gain, with a focus on gene expression, genetics, and clinical factors. Using a variety of modelling approaches, this thesis identified associations between demographic factors and antipsychotic induced weight gain, and proposes a profile of individuals most at risk of antipsychotic weight gain. The factors influencing clozapine plasma concentrations are also outlined.

Psychosis is a complex condition that features in many psychiatric disorders, and significantly affects the quality of life for both patients and family members. As part of the Genetics and Psychosis (GAP) study, this thesis presents one of the largest blood gene expression datasets on first-episode psychosis patients to date. This work aimed to characterise the blood-based biological perturbations in psychosis and to investigate the predictive ability of gene expression data. Firstly changes in expression, between healthy controls and first-episode psychosis patients was explored, to identify genes associated with psychosis. I identified hundreds of differentially expressed genes and found associations to pathways involved in transcription, oxidative stress and viral replication. Secondly, network approaches were used to construct modules of genes based on co-expression. I identified modules correlated to the severity of positive symptoms, and enrich-ment for pathways associated with the stress response and multiple brain regions. Thirdly regularised generalised linear models with bootstrapping were used to generate predictions based on combinations of gene expression, genetic and demographic data. The highest performance was found for models incorporating gene expression data, with minimal improvement using additional data. Prediction accuracy for identifying psychosis samples was found to increase with severity of positive symptoms in schizophrenia samples, but not in other psychoses. Finally, machine learning methods were used on public schizophrenia gene expression data to build a variety of predictive models. These models were tested on the Genetics and Psychosis (GAP) gene expression data. The results show increased predictive performance on samples with a schizophrenia diagnosis, compared to other types of psychosis. Overall the thesis presents work analysing a novel gene expression dataset. The results suggest that blood gene expression signatures are more predictive for positive symptoms in schizophrenia than for other psychoses. This work also highlights expression differences in innate immune pathways and the stress response.

Dementia progression is characterised by a lengthy pre-symptomatic phase, where pathology may be accumulating, followed by a more rapid decline evidenced by onset of clinical symptoms and eventually functional impairment. The rate of decline in symptoms from cognitive impairment to dementia greatly varies between individuals, complicating prognosis and the assessment of much needed disease-modifying drugs. As a result there is a huge demand for greater understanding of the between-subject variability in progression and a need to understand biomarkers and risk factors for predicting future cognitive decline. In cohorts derived from two of the largest NHS foundation trust mental health service providers in the UK and multiple Alzheimer’s cohorts, this thesis explores methods for modelling disease progression and investigates the relationship between blood based proteins, genetic variants, health indicators and potential repurposing medications with cognitive decline. Through applying approaches that tackle three areas that require consideration for modelling of cognitive decline and disease progression, this thesis identified associations of antidepressant medication and psychotic symptoms associated with faster cognitive decline in dementia, in two separate cohorts.

The past two decades have seen an unprecedented increase in the amount of data collected on a daily basis. While consumer-orientated internet services, such as Google and Twitter, take up a large proportion of the former, advances in mobile technology and diagnostic tools have allowed biomedical applications to reach a similar scale in variety, velocity and volume. In the past, what used to be sparse biomedical data is now complemented with feature-rich, time-dependent information. As a result, the field is forced to consider novel approaches for extracting actionable information from time-series data sets, guaranteeing reliability and scalability of associated services as well as ensuring a high-degree of compliance from data sources. This thesis focuses on providing an in-depth evaluation of time-series analyses, scalable IT infrastructure and strategies for improved user engagement. In order to reflect the breadth and broadness of biomedical applications, the topics were distilled into two distinct studies in the fields of remote symptom detection, and large scale patient monitoring.

The availability of exome sequence data for thousands of cancer samples has enabled the investigation of the sequence-level mutations that contribute to cancer. There is a need for strategies to analyse sequence data to gain new biological and clinical insights. This thesis investigates the use of machine learning and network-based methods to identify the mutated genes associated with important clinical features and cancer types, and to aid candidate gene prioritisation in colorectal cancer, and rheumatoid arthritis. Firstly, tumour/normal exome sequence data was analysed to identify the mutated genes associated with cancer grade and cancer stage across and within three adenocarcinomas. Tumour grading is an important prognostic indicator which is based upon subjective assessment by pathologists, and is not standardised across cancer types. Despite this, this study found that protein coding mutations within TP53 were indicative of high grade status across three adenocarcinomas once adjusted for age, gender, stage, and tumour type. Secondly, Random Forest models were used to identify the mutations that discriminate each of five high-order cancer types. Based on this work a Random Forest approach was used to investigate whether exome sequence data could be used to assign cancers to their tissue of origin without prior knowledge, for future use as a classifier for cancers of unknown primary origin. Finally, a network-based method to perform candidate disease gene prioritisation called ‘k-pseudo cliques analysis’ was developed. The method identifies sets of highly interacting proteins that are enriched for low gene-level p-values. In tests, the identified gene sets outperformed a univariate test for general cancer gene enrichment. As part of the final chapter a network-based method called ‘Region Growing Analysis’ was used to perform candidate disease gene prioritisation of rheumatoid arthritis genome-wide association study data. The findings and methods developed in this thesis can provide insights to the genetic correlates of cancer phenotypes and suggest new candidate disease genes.

The aim of the research in this thesis was to discover and validate blood biomarkers of early Alzheimer's Disease (AD). Existing and novel datasets from cohort studies were used for discovery and to attempt validation of previously reported biomarkers. For example, this thesis presents the first study to investigate associations between brain amyloid and blood metabolites. Further, this thesis presents the first study to combine more than one modality of blood biomarker in AD research and the first study to use a Bayesian methodology in this field. This thesis begins by aiming to validate candidate protein markers of brain amyloid burden in a novel proteomics dataset. Secondly, pathway-based methods are used to investigate the use of gene expression measurements as a potential biomarker of AD diagnosis. In the fourth chapter I generated a novel metabolomics dataset to investigate associations between blood metabolites and brain amyloid burden. A panel is found that predicts dichotomized amyloid burden with reasonable accuracy. The accuracy is improved by the inclusion of a candidate protein in the model. The fifth chapter of this thesis is focused on the use of a Bayesian methodology to predict measurements of amyloid using a variety of omics data. The Bayesian methodology allows incorporation of historical information by placing informative priors on demographic variables. No improvement is seen over demographics alone. The final chapter of this thesis aims to predict amyloid and tau burden using a polygenic risk score and levels of tau in blood. I have also considered a combined amyloid and tau pathology endpoint. The blood markers considered here do not improve predictive ability over demographics alone. Much of the work in this thesis highlights the importance of demographic factors in the diagnosis of early AD. The metabolite discovery work shows an improvement in predictive ability over demographics alone and warrants further investigation and replication. The other chapters of this thesis highlight that (in the settings investigated so far) blood measurements add minimal information above demographics alone.

Whilst antidepressants are widely prescribed, there is a large degree of variation between patients in terms of treatment outcomes. Furthermore, the mechanisms by which these drugs exert their effects remain unclear. In this thesis, genetic biomarkers of antidepressant outcomes have been explored, in order to better understand the molecular mechanisms underpinning effective antidepressant treatment. The research presented here use data from the GENDEP project, which is a large pharmacogenetic study of depressed patients receiving antidepressant treatment. Firstly, the pharmacological underpinnings of antidepressant-associated side effects were used to categorise these side effects and conduct a candidate gene analysis. Whilst a significant association between variation within the HTR2C gene and serotonergic side effects was found, the observation was not replicated in a second sample. Secondly, the role of variability in drug metabolism rates in treatment outcomes was investigated. Examining genotypic information on the cytochrome P450 enzymes, no associations with treatment response, side effects or study discontinuation were observed. Furthermore, serum concentrations of antidepressant were unrelated to treatment response or overall burden of side effects, predicting only a minority of specific side effects. Thirdly, transcriptomic changes with drug administration were explored in relation to treatment efficacy. Two genes were identified where changes in expression levels were significantly associated with treatment response amongst patients taking nortriptyline. Furthermore, using a network-based approach, changes in gene expression across one module of coexpressed genes showed significant correlation with symptom improvement; this biological network generalised across different antidepressant medications. Finally, genomic and transcriptomic data were combined, in an examination of the genetic control of gene expression. This analysis then was used to gain an insight into the molecular processes that link genotype to phenotype. The evidence presented within this thesis, when considered in combination with existing literature, highlights that antidepressant efficacy is a complex trait, influenced by many genes of small effect. Nevertheless, by layering together different levels of information, we can begin to dissect the molecular mechanisms involved in antidepressant action.