Autoimmune disease arises from dysregulation of the immune system, leading to its attack on the body's own tissues and organs. The clinical heterogeneity of these diseases arises from several sources, such as genetic predisposition, environmental triggers, and aberrant immune responses. One emerging area of interest is the role of transposable elements (TEs) in autoimmune disease pathogenesis because these self-nucleic acids can be mistakenly detected as foreign, which can trigger a chronic immune reaction.There is growing appreciation for the role of TEs in systemic lupus erythematosus (SLE) and studies have found differentially expressed TEs in SLE patients, which suggests a link between TE activity and disease mechanisms. Our work investigated TE expression in four immune cell types from SLE patients, revealing cell-specific and SLE subphenotype-specific differentially expressed TEs, with additional cell-type-specific TE associations in different ancestry groups. TE expression was also associated with host gene expression involved in antiviral and immune responses, supporting the hypothesis that TEs could activate the innate immune system and contribute to chronic inflammation and autoimmunity. This study underscores the importance of TEs in SLE heterogeneity and highlights the need for further exploration of TE expression in normal immune cells and functional studies to understand their role in SLE pathogenesis. Future work to study whether antiretroviral drugs could reduce expression of TEs and mitigate SLE symptoms is warranted, given the potential involvement of TEs in autoimmune disease pathogenesis

Many neurobehavioral traits are highly heritable, yet specific genes underlying them are difficult to identify. Copy number variants – large, often multi-gene, deletions and duplications - are one category of mutation that drives traits including autism spectrum disorder (ASD), bipolar disorder, schizophrenia, obesity, and intellectual disability. A 600kb 30-gene region on chromosome 16p11.2, for example, is strongly associated with ASD when duplicated or deleted, while deletions of a 3mb 60-gene region on chromosome 22q11.2 are a well-known genetic cause of schizophrenia. Under the assumption that a subset of these 30 or 60 genes contribute to associated traits, either individually or in combination, we investigate the genic architecture of CNV association with neurobehavioral traits.

We think about CNVs’ impact on traits in the context of gene expression: individuals that have genetic duplications have increased gene expression (1.5x of normal), while those with deletions have decreased gene expression (0.5x of normal). Among non-carriers, expression levels vary as well, although typically to a smaller extent. This gives rise to a hypothesis: expression variation of the genes in a CNV region may be associated with the CNV-associated (or related) traits in non-carriers. Under this hypothesis, we can separate out the association of the expression of specific CNV genes with specific traits, which cannot be done in a CNV carrier where all genes are upregulated or downregulated.

We studied large populations of non-carriers with genetic data (and corresponding controls, where applicable) for five neurobehavioral traits: ASD, bipolar disorder, schizophrenia, BMI (as a proxy for obesity), and IQ (as a proxy for intellectual disability). We took an expression prediction approach which allowed us to convert GWAS-style data into imputed expression-level data which can be used for association analyses. We studied the association of individual CNV genes, pairs of CNV genes, and all genes in the region to these five traits. This study design was also used to assess association between CNV gene expression variation and clinical traits within a large biobank with genotypic and clinical information.

The second chapter of this dissertation focuses on individual genes within GWAS datasets and clinical biobanks, and the third focuses on the extension of this approach to combinations of genes and the entire region. In support of our hypothesis, we were able to detect individual genes at 16p11.2 associated with neurobehavioral traits, most notably INO80E, significantly associated with schizophrenia and BMI, and nominally associated with bipolar disorder. Using the biobank, we found additional genes associated with related clinical traits including psychosis and mood disorders, with an overall over-representation of mental disorders in CNV gene associated phenotypes. We then found that variance in traits was better explained by pairs of CNV genes in nearly all instances, including those where we had identified single-gene associations. The regionwide prediction was associated with BMI and IQ at both 16p11.2 and 22q11.2, but not with any neuropsychiatric trait. The importance of the combinatorial contributions of genes did not extend to matched control regions for the same traits. In sum, our studies provide insight into the transcription-based action of CNV genes, identify potential candidate genes for further study, describe combinatorial patterns of CNV gene impacts on neurobehavior, and demonstrate the utility of integrating genetic, clinical, and transcriptomic data for in silico analyses.

Real world data (RWD), data from various sources other than clinical trials, is increasingly being integrated into the research setting. In particular, electronic health records (EHRs), which serve as a clinical record to document a patient’s medical history as well as support administrative functions, have been an invaluable resource rich with patient data. Here we present three projects spanning four chapters where EHRs, in combination with clinical trials and pharmacokinetic and pharmacodynamic (PKPD) modelling, were used to extend and complement studies and findings in the laboratory focusing on transporter-mediated drug interactions. Transporter-mediated drug interactions have the potential to influence both drug efficacy as well as toxicity. During the clinical development of the Janus Kinase 2 (JAK2) inhibitor fedratinib, several patients developed symptoms similar to Wernicke’s encephalopathy, a life-threating disease caused by Vitamin B1 (thiamine) deficiency; subsequent in vitro studies showed that fedratinib is a potent inhibitor of ThTR-2. Motivated by this drug-nutrient interaction (DNI) observed in the fedratinib trial, we investigated if commonly used prescription drugs can inhibit ThTR-2. Using a multifaceted approach, we started with an in vitro high-throughput screen which was further complemented by quantitative structure activity relationship (QSAR) modelling and real world data. Our comprehensive analysis suggested that several marketed drugs inhibit ThTR-2 and may contribute to thiamine deficiency, especially in at-risk populations. In order to further explore the impact of these potential inhibitors in humans, we designed and conducted a clinical study in healthy volunteers. Interestingly, we observed that thiamine concentrations were higher when co-administered with trimethoprim, one of the potent, clinically relevant inhibitors identified in our screen. The maximum concentration achieved (Cmax) and area under the curve from 0 to 24 hours (AUC0-24) were 2.7- and 4.6-fold higher in the combination arm, respectively. We hypothesized that trimethoprim may inhibit OCT1, a hepatic uptake transporter, in addition to ThTR-2, which was supported using EHR data by comparing laboratory values of endogenous OCT1 biomarkers in patients prescribed trimethoprim versus patients not prescribed trimethoprim. Next, we shifted our focus to pharmacogenomics, that is, genetic factors that affect drug response. Response to allopurinol, the first line treatment for gout, is highly variable; the reduced function variant BCRP p.Q141K has been associated with poor response to allopurinol. Thus, we aimed to characterize the relationship between BCRP p.Q141K, allopurinol/oxypurinol, and serum uric acid (SUA) levels by performing a clinical trial, building a PKPD model, and mining EHRs. Our clinical study found that p.Q141K associated with longer half-life of oxypurinol and our PKPD model found that gender affected oxypurinol volume of distribution while BCRP genotype and kidney function were significant covariates for baseline SUA levels. Additionally, using RWD, we found that drugs that were clinical inhibitors of BCRP associated with increased SUA levels, suggesting the potential of these drugs to cause hyperuricemia. Finally, given the ongoing COVID19 pandemic, we conducted extensive in vitro experiments aimed at predicting the potential for 25 small molecule drugs in clinical trials for COVID19 to cause transporter-mediated drug-drug interactions (DDIs). We found that 21 of the drugs were predicted to cause a clinically relevant DDI, and we were able to provide preliminary validation of these in vitro findings using EHR data, including a database representing nearly 120,000 COVID19 patients.

Collectively, my dissertation research demonstrates how the integration of benchwork, clinical trials, and real world data provides us a new approach to translational research, bridging findings from the laboratory to patients.

Cancer is among the leading causes of mortality worldwide and the number of cancer-related deaths is expected to rise to 16.4 million by 2040. Given the wealth of publicly available cancer data that has been generated over the past few decades, it is now possible to investigate cancer at an unprecedented scale using computational approaches. This body of work covers three projects that leverage human tumor samples to evaluate cancer models, predict cancer therapeutics, and investigate the prognostic value of infiltrating B cell repertoires. In the first project, we compared cell lines from the Cancer Cell Line Encyclopedia to primary tumor samples from the Cancer Genome Atlas (TCGA) to evaluate how well each cell lines represents its primary tumors. We predicted subtype classifications for individual cell lines and proposed a new pan-cancer cell line panel with the most representative cell lines across 22 tumor types to facilitate pan-cancer studies. In the second project, we applied a computational drug repositioning approach to identify compounds to sensitize drug resistant breast cancers using patient samples from the I-SPY2 TRIAL and the Connectivity Map drug perturbation dataset. We identified a drug hit, fulvestrant, which we validated experimentally and found that it increased drug response in a paclitaxel-resistant breast cancer cell line. In the third project, we extracted B cell repertoires from TCGA RNA-seq samples and performed diversity and network analysis. We then evaluated the prognostic value of these repertoire features and identified significant associations with survival in a subset of tumor types. Together, this research demonstrates how computational methods can leverage publicly available datasets to extract new insights into cancer biology

Alzheimer's disease (AD) is a multifactorial neurodegenerative condition characterized by heterogeneous molecular alterations across various brain cell types, posing significant challenges for the development of effective treatments. To address this complexity, we undertook two complementary projects aimed at advancing precision medicine approaches for AD, each building on the other to create a robust foundation for targeted therapies.The first project serves as a proof of concept for an innovative drug discovery strategy, utilizing a network correction approach grounded in direct human evidence and real-world data. By integrating diverse datasets, including single-cell human transcriptomics, drug perturbations, and electronic clinical records, we identified the combination of letrozole and irinotecan as potential therapeutics designed to correct gene expression alterations across multiple cell types implicated in AD. Rigorous validation in AD mouse models demonstrated that this combination therapy, targeting both neurons and glial cells, significantly ameliorated memory deficits and other AD-related pathologies, outperforming the single-drug treatments targeting either neurons or glial cells alone. The success of this project underscores the potential of cell-type-directed network-correcting therapy, demonstrating that targeting the transcriptomic landscape at a cell-type-specific level may offer a more efficacious approach to treating AD. Building on this foundation, the second project was designed to expand the knowledge foundation for precision medicine by comprehensively characterizing the molecular influences of major AD risk factors, such as age, apolipoprotein E4 (APOE4), and sex using AD mouse models. By analyzing single-nucleus RNA-sequencing (snRNA-seq) data from the hippocampus of human APOE4 and APOE3 knock-in (KI) female and male mice across different ages, we identified significant variations in cell type abundance and gene expression patterns, particularly driven by sex differences and the interplay between age and APOE genotype. This detailed molecular profiling not only enriches our understanding of the disease but also provides a valuable dataset for future applications of the network correction method. Specifically, it enables the mapping of distinct risk profiles and the identification of tailored therapeutic interventions for individuals based on their unique transcriptomic signatures. Together, these two projects are highly complementary: the first project validates the efficacy of network correction therapy, while the second project supplies the essential molecular data needed to understand disease heterogeneity, enabling the application of this therapeutic approach in a precision medicine framework. By integrating these insights, we can more accurately tailor treatments to individuals with distinct risk profiles, thereby advancing the development of personalized therapies for AD.

Alzheimer’s Disease (AD) is a devastating neurodegenerative disorder that is difficult to study and treat despite decades of progress. This is due to disease heterogeneity, lack of precise phenotyping, and limited understanding of molecular mechanisms underlying clinical manifestations. Electronic medical records (EMR) are emerging as a real-world dataset with abundance of longitudinal human data across diagnoses, medications, and measurements with opportunity to derive insights without predefined selection criteria or limitations in scope. Recent developments of integrative heterogeneous graph databases that combine knowledge across omics relationships provide a means to further identify molecular hypotheses underlying complex clinical phenotypes. We performed deep phenotyping to characterize AD and sex differences in the EMR against a control cohort, and identified sex and AD associated comorbidities, medication use, and lab values. Extending this work to apply machine learning, we utilize clinical information to predict AD onset and identify prioritized genes via knowledge networks (e.g., APOE, ACTB, IL6) and genetic colocalization analysis (e.g., MS4A6A with osteoporosis). Our findings suggest that AD onset risk can be predicted based on clinical data and that there are sex-specific relationships in AD including musculoskeletal disorders among females with AD and neurological or sensory disorders among males with AD. Extensions to knowledge networks and molecular datasets further prioritize genes depending on an individual’s comorbid conditions. By leveraging clinical data to identify hypotheses for complex disease, we can further make steps towards better understanding molecular mechanisms and advance personalized treatment approaches in AD.

Alzheimer’s disease (AD) is the leading cause of dementia worldwide, and no effective therapies are available. The multifactorial etiology and pathophysiological complexity of AD cause patient heterogeneity and pose challenges for drug development, with almost all efforts to target AD-related pathways having failed in human trials. Although apolipoprotein (apo) E4 is the major genetic risk factor for AD—60–80% of patients have at least one APOE4 allele and ~70% of homozygotes develop AD by age 85—it has not been actively considered in drug target stratification and development for AD. Here, we used an apoE-genotype-specific drug repositioning approach to screen for drugs to treat apoE4-related AD. From a meta-analysis of 610 human temporal lobar samples from public databases, we established apoE-genotype-specific transcriptomic signatures of AD and applied them to a validated Connectivity Map (CMap) database containing transcriptomic perturbation signatures of 1300 existing drugs to identify those capable of perturbing an entire gene-expression network away from the apoE-genotype-driven disease state towards a normal state. The loop-diuretic bumetanide was the top predicted drug candidate for apoE4/4 AD. Treating aged apoE4 knock-in (apoE4-KI) mice with bumetanide rescued cognitive and neuronal plasticity deficits, warranting further efficacy tests in AD clinical trials. This study highlights the power of combining precision medicine, computational drug repositioning, and targeting network alterations in developing new therapies for AD and other neurodegenerative disorders.

Objectives

Methods

Results

Conclusion