UC San Diego

The completion of the Human Genome Project in 2003 provided a nearly complete map of human DNA and marked the beginning of modern genomics. Today, we have genome sequences for almost a million individuals and have identified thousands of genetic variants associated with common diseases, most of which lie in the non-coding genome. Despite major progress in biochemically and functionally annotating the genome across diverse cell types, the mechanisms linking non-coding genetic variation to phenotype remain poorly understood.In this thesis, I use a form of predictive modeling called deep learning to analyze large-scale genomics datasets, with the goal of uncovering how non-coding DNA influences gene regulation across cell types and states. While deep learning has transformed the fields of language and vision, its application to genomics poses distinct challenges. To address these, I first present EUGENe (Elucidating the Utility of Genomic Elements with Neural Nets), a software ecosystem that enables the training, evaluation, and interpretation of deep learning models on genomics data. EUGENe includes several modular tools designed to be scalable, adaptable, and accessible to a wide range of users, and represents an important step in democratizing the use of deep learning in genomics.I then apply EUGENe to two biological systems. First, I analyze a high-throughput reporter assay of synthetic DNA to investigate how transcription factor binding site syntax shapes enhancer activity. I develop a model that predicts enhancer function from binding site syntax alone, assess its interpretability, and use it to design synthetic enhancers. Second, I apply EUGENe to single-cell chromatin accessibility profiles from stimulated, stem cell–derived pancreatic organoids. I train accurate sequence-to-profile models across 15 cell states, interpret learned regulatory features, and use the models to predict the effects of non-coding variants, revealing mechanisms underlying environment-induced changes in β-cell gene regulation.Together, this thesis demonstrates the power of interpretable predictive models for decoding the logic of gene regulation, and provides frameworks that can be broadly applied to future functional genomics studies.