Statistical methods for the analysis of multi-omics and multi-study datasets

The generation of multiple, large scale genomic datasets has become increasingly common in biological and biomedical research. These datasets can be collected for a common set of samples over multiple assays, known as a multi-omics dataset, or for the same assay over multiple collections of samples, known as a multi-study dataset. The growth of multi-omics datasets has given rise to many methods that identify sources of common variation across data types. However, the unsupervised nature of these methods makes it difficult to evaluate their performance. We propose Multi-Omics VIsualization of Estimated contributions (MOVIE), to evaluate the extent of overfitting of multi-omics methods. MOVIE utilizes a cross-validation approach to identify method stability and provides a visualization of the overfitting. Plotting the contributions of one data type against another produces contribution plots, where contributions are calculated for each subject and each data type from the results of each multi-omics method. The usefulness of MOVIE is demonstrated through evaluating the performance of multi-omics methods on large and small-sample experimental datasets and identifying overfitting in a permuted null dataset. We also propose a statistical model for comparing RNA splicing patterns from an independent experimental dataset to a reference dataset containing a large number of samples. The relative proportion of RNA isoforms expressed for a given gene has been associated with disease states in cancer, retinal diseases, and neurological disorders. Examination of relative isoform proportions can help determine biological mechanisms, but such analyses often require a per-gene investigation of splicing patterns. Leveraging large public datasets produced by genomic consortia as a reference, one can compare splicing patterns in a dataset of interest with those of a reference panel in which samples are divided into distinct groups, such as tissue of origin, or disease status. Our proposed model ACTOR, A latent Dirichlet model to Compare expressed isoform proportions TO a Reference panel, uses a variational Bayes procedure to estimate posterior distributions for the group membership of one or more samples. Using the Genotype-Tissue Expression (GTEx) project as a reference dataset, we evaluate ACTOR on simulated and real RNA-seq datasets to determine tissue-type classifications of genes. In multi-omics analyses, variable sample quality of one assay can drastically alter the results. Current methods account for this heuristically by removing samples based on a semi-arbitrary cut point of an external quality score. Leveraging an external sample quality score, we propose Sample Quality Weighted Canonical Correlation Analysis (SQWCCA) as an extension to Sparse Canonical Correlation Analysis. SQWCCA calculates sample weights based on an external sample quality score which improve the weighted correlation between the two assays. We evaluated SQWCCA through simulations and a dataset of samples from Crohn's disease patients and other patients who do not have inflammatory bowel disease. In simulations, SQWCCA was able to identify poor quality samples, while avoiding the unnecessary removal of samples when using a non-informative quality score. In the real dataset, SQWCCA identified six samples of low quality using the Transcription Start Site (TSS) enrichment score as the external quality score.Doctor of Philosoph