Deciphering the Dynamic Regulatory Architecture of the Genome Using Interpretable Deep Learning Models

Functional genomics experiments profiling genome-wide regulatory state have revealed millions of putative regulatory elements in diverse cell states. These massive datasets have spurred the development of powerful predictive models called Deep Neural Networks (DNNs) that can accurately map DNA sequence to associated cell-type specific molecular phenotypes such as transcription factor (TF) binding, chromatin accessibility and gene expression. Beyond high prediction accuracy, the primary appeal of DNNs is that they are capable of learning predictive sequence features and their non-linear interactions directly from raw DNA sequence without any prior assumptions. Hence, interpreting these purported black box models could reveal novel insights into the combinatorial regulatory code. I will present efficient interpretation engines for extracting predictive and biological meaningful patterns from integrative deep learning models of regulatory DNA. I will show how we can use interpretable deep learning models to obtain new insights into the DNA sequence affinity of TFs, infer high-resolution point binding events of TFs, reveal epistatic motif interactions in cis-regulatory sequence grammars, unravel dynamic regulatory drivers of cellular differentiation and interpret non-coding regulatory genetic variants associated with disease.