The human heart is one of the first organs to form during embryogenesis. By week 3, heart is a linear tube; then, between week 3 and 8, it undergoes a complex looping process to form heart chambers. From week 8 up until birth, the heart undergoes a period of rapid growth and expansion as the various cardiovascular lineages mature to support the developing fetus. Defects in this process lead to congenital heart defects. Therefore, understanding heart development at the molecular level will give insight into the mechanisms leading to the disease, and principles can further be applied for re- generative medicine purposes. In addition to genetic factors, lifestyle also significantly contributes to maintaining a healthy heart; therefore, lowering the risk factors that contribute to heart diseases such as obesity is vital. This thesis investigates computational and statistical approaches that help understand the influences of molecular, genetic, and lifestyle components in maintaining a healthy heart.Single-cell sequencing has provided a unique opportunity to study the heart at cellular resolution. Recent advances in single-cell sequencing allow scientists to sequence different modalities, such as proteomics, spatial transcriptomics, methylation, and chromatin accessibility at the level of individual cells. These advances offer a unique opportunity to develop computational methods for integrative single-cell data analysis to benefit from the information in multiple modalities. In the first chapter of this thesis, I present Single-cell Integrative Gene Regulatory Network (sci-GRN) inference framework that uses single-cell transcriptomics and epigenomics to infer gene regulatory networks (GRNs), and I demonstrate its application to the human heart development. Here, I first describe how each omic is processed independently to cluster the cells. Next, I describe how we annotated cell types from scRNA-seq data using known marker genes. Third, I present how I used the scRNA-seq data to infer the expression and cell type of each ATAC cell. Lastly, I present how sci-GRN uses the integrated data to build a gene regulatory network. In addition to genetic factors, lifestyle also significantly contributes to maintaining a healthy heart; therefore, lowering the risk factors contributing to heart diseases is vital. Studies have shown that obesity is associated with an increase in cardiovascular diseases. Therefore, it is crucial to tackling obesity systematically. Enara Health is a medical company based in San Mateo, California that offers a unique hybrid digital and in-person weight loss program. By combining mobile technology with healthcare visits, the program provides patients with personalized physician-driven medical weight loss programs to tackle obesity[2]. In the second chapter of my thesis, I present our study on analyzing one of the cardiovascular diseases” risk factors, obesity. I demonstrate how my statistical analysis implies the significant role of anti-obesity medication initiation and duration on weight loss in obese patients enrolled in Enara’s weight loss program. In the final chapter, I further analyze the impact of Enara’s weight-loss pro- gram mobile application (mHealth) on weight loss in obese patients. The mobile application under study allows for examining and comparing self-monitoring, feedback, support, and educational features within the same mHealth program, which may shed insight into which features and usage are better for sustained weight loss.[2]

Since response to therapy can differ greatly between cancer patients, a precision medicine approach to treating cancer based on the uniqueness of patient tumors could greatly improve response rate and quality of life. High-throughput assays provide the means to probe multiple types of genomic alterations across a patient’s cancer genome. By leveraging prior knowledge about genetic pathways, I have created tools to address the challenges of tackling large multi-dimensional datasets to make biological sense of the data. I developed PATHMARK to identify clusters of genes that are dysregulated together forming networks that offer insights into disease mechanisms and treatment strategies. PATHMARK utilizes conventional univariate differential analysis with a filter on pathway interactions to identify sub-networks that are significantly more connected than by chance. I developed PARADIGM-SHIFT to predict the functional impact of mutations detected from whole-exome sequencing data. PARADIGM-SHIFT analyzes the inferred activities of a network surrounding a mutated gene, comparing the levels between mutant and wild-type samples. The approach predicts if mutations are likely to be neutral, gain-of-function, or loss-of-function. I demonstrate how inferences about mutations in novel genes and in non-coding regions can be gleaned from models trained on known coding mutations. The predictions form “molecular machines” that link events together based on shared pathway alteration. With the growing number of available datasets and computational tools, it has become increasingly important to make analyses easily accessible and reproducible. To support these ideals, I have developed my tools to be compatible within the Galaxy system, which has enabled collaborators to apply my tools to analyze their data.

In this work, I examine the use of low-level feature classifiers to enhance the performance of clinically relevant gene expression classifiers. In other machine learning domains, notably machine vision, recognizing higher level or more abstract concepts by using lower level or more concrete feature classifiers is a common motif. For example, a classifier to recognize images of “parties” might usefully first have sub-classifiers to recognize concrete elements commonly found in parties: indoors, people, party hats, cake, candles, banners, confetti, etc. The output of these low level feature classifiers can be supplied to an overall “party” classifier to assign a label. Image pixels are far removed from an abstract concept like “party,” so training classifiers for these more concretely defined sub-concepts is a way to bridge this semantic gap. Moreover, there may be many more training examples available for the lower level concepts (e.g. people) than the target (e.g. party), allowing one to utilize the robustness that comes from plentiful data. I hypothesize that there is a similar semantic gap between gene expression levels and clinically relevant gene expression classification targets, such as survival prognosis or drug sensitivity for tumor types, and that these clinically relevant classification tasks can benefit from decomposition into classifiers for more concrete concepts, such as tissue type, chromatin state, mutations, and gene essentiality. I will present a series of experiments that show modest but real improvements from the use of low-level feature classifiers in several gene expression prediction tasks. I will also present a series of tools developed for this work including wekaMine (a large-scale model selection and feature-building pipeline), viewtab (a “big data” spreadsheet), csvsql (a program to allow arbitrary SQL queries on csv/tab files), and SamplePsychic (a web application to apply a suite of feature classifiers to gene expression samples and explore the results).

The cellular components of tumors and their microenvironment play pivotal roles intumor progression, patient survival, and the response to cancer treatments. In my doctoral thesis, I describe a new way to extract transcriptional signatures from gene expression data of tumor components and microenvironments and access their influence on cancer patients. Tumor immune infiltration has been studied for years for its high correlation with patients’ survival. Many immune therapies are dependent on the detection and quantification of cell types present in bulk tumors. Bulk tumor microenvironment deconvolution has been largely limited by the low number of cell-type signatures. Leveraging cell type signatures derived from scRNA-seq data provides a broader range of cell types in the detection and quantification of tumor infiltration, therefore helping in developing cancer immunotherapy and targeted cancer therapies. I developed a new method called scBeacon, a novel tool that derives cell-type signatures by integrating and clustering multiple scRNA-seq datasets to extract signatures from public data consortiums while minimizing batch effects. I derived a comprehensive set of human cell-type signatures from Single Cell Expression Atlas and performed TCGA bulk tumors deconvolution analysis using the cell-type signature profile. These cell type estimates enable the detection of a pan-cancer high-risk sample group that is not detected by traditional gene expression analysis. Cancers are traditionally classified into types and subtypes by the organ and cell of origin. However, there are tumor samples that consist of a mixture of cancer subtypes which raise challenges in characterizing the subtype profile for mixture samples. Inspired by the scBeacon deconvolution analysis, I used a similar approach to detect and quantify the subtypes in testicular germ cell cancer mixture samples. In addition, I used single-cell RNA-seq signatures to characterize the major cell types and their differential states in testicular germ cell cancer samples. For glioblastoma, I used predefined cell state marker genes to deconvolute bulk glioblastoma tumors using a hierarchical deconvolution approach. It proved using hierarchical deconvolution addresses the nature of cell type differentiation, which could give a finer resolution for deconvolution, especially when some rare cell types come from a subpopulation of a very similar cell type. This approach is particularly useful for brain tissue deconvolution because of the complexity of cell type lineages in brain development.

Cancer is an ideal target for personal genomics-based medicine that uses high-throughput genome assays such as DNA sequencing, RNA sequencing, and expression analysis (collectively called omics); however, researchers and physicians are overwhelmed by the quantities of big data from these assays and cannot interpret this information accurately without specialized tools. To address this problem, I have created software methods and tools called OCCAM (OmiC data Cancer Analytic Model) and DIPSC (Differential Pathway Signature Correlation) for automatically extracting knowledge from this data and turning it into an actionable knowledge base called the activitome. An activitome signature measures a mutation's effect on the cellular molecular pathway. As well, activitome signatures can also be computed for clinical phenotypes. By comparing the vectors of activitome signatures of different mutations and clinical outcomes, intrinsic relationships between these events may be uncovered. OCCAM identifies activitome signatures that can be used to guide the development and application of therapies. DIPSC overcomes the confounding problem of correlating multiple activitome signatures from the same set of samples. In addition, to support the collection of this big data, I have developed MedBook, a federated distributed social network designed for a medical research and decision support system. OCCAM and DIPSC are two of the many apps that will operate inside of MedBook. MedBook extends the Galaxy system with a signature database, an end-user oriented application platform, a rich data medical knowledge- publishing model, and the Biomedical Evidence Graph (BMEG). The goal of MedBook is to improve the outcomes by learning from every patient.

Advancements in sequencing technology have led to an influx of cancer genomics data, transforming cancer research into a field limited by data interpretation rather than acquisition. Machine learning methods that can make use of this wealth of data are desperately needed. Similarly, patient stratification is a critical task in cancer diagnosis and treatment. While stratification approaches using various biomarkers for patient-to-patient comparisons have been successful in elucidating previously unseen subtypes, the potential of many other sparse but rich genotype and phenotype data (e.g. tumor images) remains untapped.

To this end, I present two methods. The first uses social network analysis techniques to extract subtypes from sparse data. The second is a semi-supervised multiview learning framework that integrates both prior knowledge and a variety of genomic data to predict outcomes in cancer. Crucially, this method accommodates samples for which we have different data types, paving the way for integration of data from past studies.

I apply these methods to several cancer datasets. Of note, I show that TCGA--defined molecular subtypes of glioblastoma are independent of both tumor location and volume, and that both the imaging and genomic data provide important perspectives of the disease. Analysis of a large drug sensitivity database identifies an epigenetic effect from chromatin modifiers that lends sensitivity to Panobinostat. Multiview learning, the second method I developed, also outperforms other methods in predicting sensitivity in all of the study drugs. In this dissertation I begin with unsupervised single--platform analysis, then combine multiple platforms, and finally analyze many data platforms using semi--supervised analysis.

Regulation of gene expression is critical for all life processes. RNAseq allows for the investigation of the expression state and the inference of gene regulatory net- works (GRNs). The GRN reveals mechanisms of disease and development both at an organismal and cellular level. Constructing GRNs from expression data is done by de- tecting statistical relationships over a series of samples. However, these relationships are noisy and do not take into account many important variables. To explore one such limitation, I implemented dual RNAseq to investigate the regulatory mechanisms that occur between both host and virus transcriptomes in SARS-CoV-2 infected samples (Chapter 1). To address the temporal aspect of GRNs, I developed a statistical frame- work for detecting relationships between transcription factors and their targets from inferred trajectory progressions (”pseudotime”) in single-cells (Chapter 2). Finally, I developed an interpretable, time-aware deep learning transformer for modeling disease progression and demonstrate its ability to identify gene programs that promote breast cancer resistance post-treatment (Chapter 3).

Background

Results

Conclusion

Background

Methods

Results

Conclusions