Deciphering transcriptional patterns of gene regulation : a computational approach

With rapid advancements in sequencing technology, we now have the ability to sequence the entire human genome, and to quantify expression of tens of thousands of genes from hundreds of individuals. This provides an extraordinary opportunity to learn phenotype relevant genomic patterns that can improve our understanding of molecular and cellular processes underlying a trait. The high dimensional nature of genomic data presents a range of computational and statistical challenges. This dissertation presents a compilation of projects that were driven by the motivation to efficiently capture gene regulatory patterns in the human transcriptome, while addressing statistical and computational challenges that accompany this data. We attempt to address two major difficulties in this domain: a) artifacts and noise in transcriptomic data, and b) limited statistical power. First, we present our work on investigating the effect of artifactual variation in gene expression data and its impact on trans-eQTL discovery. Here we performed an in-depth analysis of diverse pre-recorded covariates and latent confounders to understand their contribution to heterogeneity in gene expression measurements. Next, we discovered 673 trans-eQTLs across 16 human tissues using v6 data from the Genotype Tissue Expression (GTEx) project. Finally, we characterized two trait-associated trans-eQTLs; one in Skeletal Muscle and another in Thyroid. Second, we present a principal component based residualization method to correct gene expression measurements prior to reconstruction of co-expression networks. In this work, we demonstrated theoretically, in simulation, and empirically, that principal component correction of gene expression measurements prior to network inference can reduce false positive edges. Using data from the GTEx project in multiple tissues, we showed that this approach reduced false discoveries beyond correcting only for known confounders. Third, we present a multi-study integration approach to identify universal transcriptional patterns underlying epithelial to mesenchymal transition (EMT) across different cancer types. With informed statistical analysis and functional validation, we identified consensus ranked universal EMT genes. This gene list consisted of a) known EMT genes, b) genes studied in a subset of carcinomas, unknown in prostate cancer, and c) novel unknown EMT and cancer genes such as C1orf116. Finally we present methods to integrate co-expression signals across multiple human RNA-seq data to reconstruct networks with increased power. First, we considered multiple aggregation strategies to build context-agnostic networks using data from recount2. These networks captured ubiquitous patterns of gene co-expression shared across tissues and cell types. Next, we briefly describe a hierarchical mixture model groupNet that leverages signal from multiple datasets to learn the structure of a Gaussian Markov random field (GRMF) to build context-specific co-expression networks