Advances in omics technologies have led to an abundance of comprehensive biomolecular information of biological systems, down to single-cell resolution. With omics data, biologists can gain a deeper understanding of the complex-hierarchical networks that constitute an organism. To this end, deep learning methods are often applied to assist in discovering meaningful patterns and relationships from omics data. Though deep learning methods can offer high performance on many complex tasks, some challenges arise with omics-based tasks: (1) Omics data are often high-dimensional with low-sample size and/or high levels of sparsity, with complex dependency structures between and within omics data types. (2) There is an imbalance of annotated data across different species and environments. These difficulties make desirable the integration of omics data across different modalities, group samples, and platforms, as well as environments and species. This thesis examines, builds and implements approaches to address these challenges through Integrative Learning techniques, which I use as a general term to encompass techniques that incorporate multiple sources of related data for improved learning (e.g., transfer learning, multi-task learning, and multi-modal data integration). In this work I highlight and address these challenges in different omics-based tasks. In addition to applied methods development, I provide probabilistic and mathematical frameworks that underpin many of these applied problems in omics analysis. Lastly, I showcase some of my more current experiments in deconvolution. Though these experiments are prelimary, often through toy examples, they demonstrate some possible future directions I want to consider.
Read
