Learning Regulatory DNA-Sequence Code Of Epigenetic Events Using Deep Neural Networks

Epigenetic events, such as DNA methylation and histone modifications, arise from a complex interplay among genomic sequence, chromatin-remodeling factors, and environmental cues. These regulatory mechanisms can induce changes in gene expression without altering the underlying DNA sequence, playing critical roles in development, disease, and cellular differentiation. Among these events, DNA methylation is frequently profiled using bisulfite sequencing (e.g., whole-genome bisulfite sequencing [WGBS], reduced representation bisulfite sequencing [RRBS]). However, predictive modeling of epigenetic states—including methylation patterns and regulatory variant effects—remains challenging due to data sparsity, label noise, and limited uncertainty estimation in current deep learning approaches. This dissertation addresses these issues by introducing a suite of data-centric and uncertainty-aware deep learning frameworks. First, I conduct a systematic evaluation to quantify how data quality, sample size, and label noise affect model performance. Second, I propose a transfer learning method to impute sparse methylation profiles to improve data quality. Third, I develop a Monte Carlo dropout–based pipeline that quantifies uncertainty for non-coding genetic variant effect predictions, aiding in the prioritization of potentially regulatory variants in tissue-specific contexts. Collectively, these contributions advance scalable, interpretable, and reliable computational approaches for epigenomic data analysis, paving the way for improved understanding and practical utilization of epigenetic events in developmental and disease settings.