This dissertation presents a novel data mining algorithm identifying molecular signatures, called attractor metagenes, from large biological data sets. It also presents a computational model for combining such signatures to create prognostic biomarkers. Using the algorithm on multiple cancer data sets, we identified three such gene co-expression signatures that are present in nearly identical form in different tumor types representing biomolecular events in cancer, namely mitotic chromosomal instability, mesenchymal transition, and lymphocyte infiltration. A comprehensive experimental investigation using mouse xenograft models on the mesenchymal transition attractor metagene showed that the signature was expressed in the human cancer cells, but not in the mouse stroma. The attractor metagenes were used to build the winning model of a breast cancer prognosis challenge. When applied on larger data sets from 12 different cancer types from The Cancer Genome Atlas Pan-Cancer project, the algorithm identified additional pan-cancer molecular signatures, some of which involve methylation sites, microRNA expression, and protein activity.

The emergence of high throughput technologies is enabling the characterization of cancer genomes at unprecedented resolution and scale. However, such data suffer from the typical limitations of observational studies, which are frequently challenged by their inability to differentiate between causality and correlation. Recently, several datasets of genome-wide functional assays performed on tumor cell lines have become available. Given the ability of these assays to interrogate cancer genomes for the function of each individual gene, these data can provide vital cues to identify causal events and, with them, novel drug targets. Unfortunately, current analytical methods have been unable to overcome the challenges posed by these assays, which include poor signal to noise ratio and wide-spread off-target effects. Given the largely orthogonal strengths and weaknesses of descriptive analysis of genetic and genomic observational data from cancer genomes and genome-wide functional screening, I hypothesized that integrating the two data types into unified computational models would significantly increase the power of the biological analysis. In this dissertation I use integrative approaches to tackle two crucial problems in cancer research: the identification of driver genes and the discovery of tumor lethalities. I use the resulting methods to study breast cancer, the second most common form of this disease. The first part of the dissertation focuses on the analysis of regions of copy number alteration for the identification of driver genes. I first describe how a simple integrative method enabled the identification of BIN3, a novel driver of metastasis in breast cancer. I then describe Helios, an unsupervised method for the identification of driver genes in regions of SCNA that integrates different data sources into a single probabilistic score. Applying Helios to breast cancer data identified a set of candidate drivers highly enriched with known drivers (p-value < e-14). In vitro validation of 12 novel candidates predicted by Helios found 10 conferred enhanced anchorage independent growth, demonstrating Helios's exquisite sensitivity and specificity. I further provide an extensive characterization of RSF-1, a driver identified by Helios whose amplification correlates with poor prognosis, which displayed increased tumorigenesis and metastasis in mouse models. The second part of this dissertation addresses the problem of identifying tumor vulnerabilities using genome-wide shRNA screens across tumor cell lines. I approach this endeavor using a novel integrative method that employs different biomarkers of cellular state to facilitate the identification of clusters of hairpins with similar phenotype. When applied to breast cancer data, the method not only recapitulates the main subtypes and lethalities associated to this malignancy, but also identifies several novel putative lethalities. Taken together, this research demonstrates the importance of the computational integration of genome-wide functional and observational data in cancer research, providing novel approaches that yield important insights into the biology of the disease.

Cancer research, like many areas of science, is adapting to a new era characterized by increasing quantity, quality, and diversity of observational data. An example of the advances, and the resulting challenges, is represented by The Cancer Genome Atlas, an enormous public effort that has provided genomic profiles of hundreds of tumors of each of the most common solid cancer types. Alongside this resource is a host of other data and knowledge, including gene interaction databases, Mendelian disease causal variants, and electronic health records spanning many millions of patients. Thus, a current challenge is how best to integrate these data to discover mechanisms of oncogenesis and cancer progression. Ultimately, this could enable genomics-based prediction of an individual patient's outcome and targeted therapies, a goal termed precision medicine. In this thesis, I develop novel approaches that examine patterns in populations of cancer patients to identify key genetic changes and suggest likely roles of these driver genes in the diseases. In the first section I show how genomics can lead to the identification of driver alterations in melanoma. The most recurrent genetic mutations are often in important cancer driver genes: in a newly sequenced melanoma cohort, recurrent inactivating mutations point to an exciting new melanoma candidate tumor suppressor, FBXW7, with therapeutic implications. But each tumor is unique, underlining the fact that recurrence will never capture all relevant mutations responsible for the disease. Tumors are a result of random events that must collaborate to endow a cell with all of the invasive and immortal properties of a cancer. Some combinations of events are lethal to a developing tumor, while other combinations are simply not preferentially selected. In order to discover these complex patterns, I develop a method based on the joint entropy of a set of genes, called GAMToC. Using GAMToC, I identify sets of recurrently altered genes with a strongly non-random joint pattern of co-occurrence and mutual exclusivity. Then, I extend this method as a means of identifying novel genes with a role in cancer, by virtue of their non-random pattern of alteration. Insights into the roles of these novel drivers can come from their most strongly co-selected partners. In the final section of the main text, I develop the use of cancer comorbidity, or increased cancer risk, as a novel data source for understanding cancer. The recent availability of clinical records spanning a large percentage of the American population has enabled discovery of many cancer comorbidities. Although most cancers arise as a result of somatic mutations accumulating over a patient's lifespan, mutations present at birth could predispose some rare populations to increased cancer risk. Mendelian disease phenotype provides strong insight into the genotype of an afflicted individual. Thus, if Mendelian diseases with cancer comorbidity can be shown to have specific defects in processes that are important in the development of that cancer, statistical comorbidity could provide a new a resource for prioritizing Mendelian disease genes as novel cancer related genes. For this purpose, I integrate clinical comorbidity, Mendelian disease causal variants, and somatic genomic profiles of thousands of cancers. I demonstrate that comorbidity indeed is associated with significant genetic similarity between Mendelian diseases and the cancers these patients are predisposed to, suggesting highly interesting and plausible new candidate cancer genes. While cancer may be the result of a series of selected random events, patterns of incidence across large populations, as measured by genomics or by other phenotypes, contain much non-random signal yet to be mined.

Adult neural stem cells are specialized astrocytes that generate neurons in restricted regions of the mammalian brain. The largest neurogenic region is the ventricular-subventricular zone, which lines the lateral ventricles and generates olfactory bulb neurons. Stem cell astrocytes give rise to new neurons in both homeostatic and regenerative conditions, suggesting that they can potentially be harnessed for regenerating the brain after injury, stroke, or neurodegenerative disease. Previous work has shown that stem cell astrocytes exist in both quiescent and activated states, but due to a lack of markers, it was not feasible to purify them. Using a novel fluorescence activated cell sorting (FACS) strategy that allows quiescent neural stem cells (qNSCs) and activated neural stem cells (aNSCs) to be purified for the first time, we performed transcriptome profiling to illuminate the molecular pathways active in each population. This analysis revealed that qNSCs are enriched in signaling pathways, especially G-protein coupled receptors, as well as for adhesion molecules, which facilitate interactions with the niche. qNSCs and aNSCs utilize different metabolic pathways. qNSCs are enriched for lipid and glycolytic metabolism, while aNSCs are enriched for DNA, RNA, and protein metabolism. Many receptors and ligands are reciprocally distributed between qNSCs and aNSCs, suggesting that they may regulate each other. Finally, comparison of the transcriptomes of qNSCs and aNSCs with their counterparts in other organs revealed that pathways underlying stem cell quiescence are shared across diverse tissues. A key step in recruiting adult neural stem cells for brain repair is to define the molecular pathways regulating their switch from a quiescent to an activated state. MicroRNAs are small non-coding RNAs that simultaneously target hundreds of mRNAs for degradation and translational repression. MicroRNAs have been implicated in stem cell self-renewal and differentiation. However, their role in adult neural stem cell activation is unknown. We performed miRNA profiling of FACS-purified quiescent and activated adult neural stem cells to define their miRNA signatures. Bioinformatic analysis identified the miR-17~92 cluster as highly upregulated in activated (actively dividing) stem cells in comparison to their quiescent counterparts. Conditional deletion of the miR-17~92 cluster in FACS purified neural stem cells in vitro reduced adult neural stem cell activation, proliferation, and self-renewal. In addition, miR-17~92 deletion led to a selective decrease in neuronal differentiation. Using an in vivo conditional deletion model, we showed that loss of miR-17~92 led to an increase in the proportion of GFAP+ cells and decrease in MCM2+ cells, as well as decreased neurogenesis. Finally, I identify Sphingosine 1 phosphate receptor 1 (S1pr1) as a computationally predicted target of the miR-17~92 cluster. S1pr1 is highly enriched in quiescent neural stem cells. Treatment of quiescent neural stem cells with S1P, the ligand for S1PR1, reduced their activation and proliferation. In vivo deletion of miR-17~92 lead to an increase in S1PR1+ cells, even among MCM2+ cells. Together, these data reveal that the miR-17~92 cluster is a key regulator of adult neural stem cell activation from the quiescent state and subsequent proliferation.

The central theme in this thesis is Empirical Bayes. It starts off with application of Bayes and Empirical Bayes methods to deoxyribonucleic acid fingerprinting. Different Bayes factors are obtained and an alternative Bayes factor using the method of Savage is studied both for normal and non- normal priors. It then moves on to deeper methodological aspects of Empirical Bayes theory. A 1983 conjecture by Carl Morris on the parametric empirical Bayes prediction intervals for the normal regression model is studied and an improvement suggested. Carlin and Louis’ (1996) parametric empirical Bayes prediction interval for the same model is also dealt with analytically while their approach had been primarily numerical. It is seen that both of these intervals have the same coverage probability up to a certain order of approximation and they have the same expected length up to the same order of approximation. Both the intervals are equal tailed up to the same order of approximation. Then the corrected proof of an important published result by Datta, Ghosh and Mukerjee (2000) is provided using first principles of probability matching. This result is relevant to our work on parametric empirical Bayes prediction intervals.

Genome-wide association studies (GWAS) have been tremendously successful in identifying genetic variants associated with complex diseases, such as rheumatoid arthritis (RA). However, the majority of these associations lie outside traditional protein coding regions and do not necessarily represent the causal effect. Therefore, the challenges post-GWAS are to identify causal variants, link them to target genes and explore the functional mechanisms involved in disease. The aim of the work presented here is to use high level bioinformatics to help address these challenges. There is now an increasing amount of experimental data generated by several large consortia with the aim of characterising the non-coding regions of the human genome, which has the ability to refine and prioritise genetic associations. However, whilst being publicly available, manually mining and utilising it to full effect can be prohibitive. I developed an automated tool, ASSIMILATOR, which quickly and effectively facilitated the mining and rapid interpretation of this data, inferring the likely functional consequence of variants and informing further investigation. This was used in a large extended GWAS in RA which assessed the functional impact of associated variants at the 22q12 locus, showing evidence that they could affect gene regulation. Environmental factors, such as vitamin D, can also affect gene regulation, increasing the risk of disease but are generally not incorporated into most GWAS. Vitamin D deficiency is common in RA and can regulate genes through vitamin D response elements (VDREs). I interrogated a large, publicly available VDRE ChIP-Seq dataset using a permutation testing approach to test for VDRE enrichment in RA loci. This study was the first comprehensive analysis of VDREs and RA associated variants and showed that they are enriched for VDREs, suggesting an involvement of vitamin D in RA.Indeed, evidence suggests that disease associated variants effect gene regulation through enhancer elements. These can act over large distances through physical interactions. A newly developed technique, Capture Hi-C, was used to identify regions of the genome which physically interact with associated variants for four autoimmune diseases. This study showed the complex physical interactions between genetic elements, which could be mediated by regions associated with disease. This work is pivotal in fully characterising genetic associations and determining their effect on disease. Further work has re-defined the 6q23 locus, a region associated with multiple diseases, resulting in a major re-evaluation of the likely causal gene in RA from TNFAIP3 to IL20RA, a druggable target, illustrating the huge potential of this research. Furthermore, it has been used to study the genetic associations unique to multiple sclerosis in the same region, showing chromatin interactions which support previously implicated genes and identify novel candidates. This could help improve our understanding and treatment of the disease. Bioinformatics is fundamental to fully exploit new and existing datasets and has made many positive impacts on our understanding of complex disease. This empowers researchers to fully explore disease aetiology and to further the discovery of new therapies.

INTRODUCTION: Recent epidemiological studies have shown that Diesel Engine Exhaust (DEE) exposure is associated with lung cancer. Well recognized exposures, such as smoking, have long been known to cause lung cancer, and the mechanisms by which the disease occurs have been closely investigated. However, there is very little information regarding the mechanisms by which chronic DEE exposure leads to a disease outcome. It has also been shown that transcriptomic changes in the deeper portions of the airway may be detectable in the more proximal parts. The goal of this study was to assess transcriptomic alterations in the nasal epithelium of DEE exposed factory workers to better understand the physiologic effects of DEE and how chronic exposure may lead to disease. METHODS: Nasal epithelium brushings were obtained from 41 subjects who work in a factory with DEE exposure, and 38 comparable control subjects who work in factories without DEE exposure. The median Elemental Carbon (EC) levels of exposed individuals was 60.7g/m3, with a range of 17.2-105.4 g/m3, while the median of EC levels of unexposed controls was 10.87g/m3, with a range of 9.89-12.55g/m3. RNA was isolated from nasal epithelial cells, and profiled for gene expression using Affymetrix Human Gene 1.0ST microarray chips. Linear modeling was used to detect differential expression between DEE exposure and controls. Pathway enrichment in differentially expressed genes was assessed using EnrichR. GSEA provided comparisons between the genes known to be differentially expressed due to smoking, and the genes that were found in our data to be differentially expressed due to smoking or DEE. A linear modeling approach was further used to investigate the effects of the interaction between smoking status and DEE exposure, and boxplot analysis was used to explore the interaction effect. RESULTS: We found 225 genes whose expression is associated with DEE exposure at FDR q < 0.25, after adjusting for smoking status. Within this set of genes, we observed increased expression of genes involved in the oxidative stress response, cell cycle, and protein modification, as well as genes associated with the AhR pathway and the Nrf2-mediated xenobiotic metabolism response. Additionally, decreased expression of genes involved in transmembrane transport, such as CFTR and the solute carrier family genes was also found. Furthermore, we discovered 8 genes at FDR q < 0.25 that have altered expression due to the interaction of DEE and smoking status, suggesting a synergistic relationship between the effects of these exposures on some aspects of the physiological response. For these genes, the effects of DEE were generally more dramatic in never smokers. CONCLUSIONS: The transcriptomic alterations we identified may help provide insight into the underlying mechanisms of DEE carcinogenicity. The relationship between cigarette smoke exposure and DEE exposure may provide more information about how chronic DEE exposure leads to lung cancer and other respiratory diseases.

While commonly thought of as a waste product, food scraps and residuals represent an important opportunity for energy and nutrient recapture within the food system. As demands on production continue to increase, conservation of these valuable resources has become a priority area. In the wake of new legislation in Vermont, Act 148, the Universal Recycling Law, the fate of microbial species in food waste, scraps and residuals is increasingly important. The presence of antimicrobial resistance genes in all types of foods calls for an increased need to estimate risk of antibiotic resistance transfer and maintenance across all segments of food production and distribution systems, from farm to fork. Specifically, the fate of antibiotic resistance genes (ARGs) in these co-mingled food wastes has not been sufficiently characterized; as legislative programs increase in popularity, surveillance of these materials is pressing and should be documented to assess the risk and potential measures for mitigation and management as we approach commercial scales of implementation Previous studies have relied on a combination of targeted techniques, such as 16S rRNA sequencing and qPCR on a specific subset of ARGs; however, these may not cover the full extent of resistance or microorganisms of concern in any given sample. As sequencing technologies improve and costs continue to drop, more comprehensive tools, such as shotgun metagenomic sequencing, can be applied to these problems for both surveillance and novel gene discovery. In this study, we leveraged the increased screening power of the Illumina HiSeq and shotgun metagenomic sequencing to identify and characterize ARGs, microbial communities, and associated virulence factors of food scraps, on-farm composts, and several consumer products. Isolates were also screened for antibiotic resistance to demonstrate the functionality of ARGs identified. The resistome, microbiome, and virulence genes were characterized in all samples. Fifty unique ARGs were identified that spanned 8 major drug classes. Most frequently found were genes related to aminoglycoside, macrolide, and tetracycline resistance. Additionally, 54 distinct virulence factors and 495 bacterial species were identified. Virulence factors were present across the farm setting and mainly included gene transfer mechanisms, while bacteria clustered distinctly into site and farm, as well as separate on farm niches. The relationship between these categories was also assessed by both Pearson correlation and co-inertia analysis, with the most significant relationship being between ARGs and virulence factors (P = 0.05, RV = 0.67). While limited in this study, these patterns reinforce the finding that spread of antibiotic resistance genes may be dependent on the virulence factors present enabling transfer, rather than total microbial community composition.

Thesis (Ph. D.)--Washington State University, December 2007. / Includes bibliographical references (p. 103-114).

Thesis (Ph. D.)--University of Hong Kong, 2007. / Title proper from title frame. Also available in printed format.