Decoding the sequence basis of genome regulation

About This Grant

Project Summary Deciphering how DNA sequences encode regulatory activities provides a new lens for dissecting the underlying molecular mechanisms of genome regulation. However, although deep learning approaches have significantly advanced our ability to predict chromatin accessibility, transcription factor binding, 3D genome architecture, and transcription from DNA sequence, current models remain limited in their ability to deliver mechanistic insights, cannot effectively capture long-range sequence dependencies, and do not reflect the heterogeneous and dynamic nature of molecular-level genome regulation. Motivated by our observation that simple explainable models can simultaneously contribute to improved biological understanding and predictive power— indicating that we are still in a data-limited regime—this proposal aims to address these challenges by developing a new generation of sequence-based AI frameworks focusing on biology-informed simplicity, explainability, and mechanistic insights without compromising on performance. Our approach is built on three key pillars: (i) We will develop explainable sequence models that discover simple rules for diverse regulatory functions. (ii) We will develop novel architectures to accurately predict long-range regulatory interactions directly from sequence, and generate genome-wide regulatory interaction maps across cell types. (iii) We will pioneer "sequence-to-distribution" generative models to capture the dynamic, single-molecule nature of chromatin regulation, unveiling its inherent heterogeneity and the sequence determinants of molecular cooperativity. Achieving these goals promises both improved predictions and deeper mechanistic insights, ultimately contributing toward making genomic regulation an "open book" where every single base’s contribution can be explained.