Programmatic Design and Editing of Cis-Regulatory Elements Using Deep Learning - Inferring Transcription Factor Binding from Chromatin Accessibility

Attend this research seminar from the Broad Institute's Models, Inference and Algorithms series featuring two presentations on computational approaches to understanding and designing gene regulatory elements. Learn about Ledidi, a computational method that transforms the discrete task of designing genomic edits into a continuous optimization problem, enabling precise control over transcription factor binding, chromatin accessibility, transcription, and enhancer activity across multiple species. Discover how this approach can design cell type-specific enhancers that not only induce specificity but also quantitatively control regulatory strength, with some designed enhancers showing greater activity than naturally occurring ones. Explore the challenges and advances in transcription factor binding prediction using deep learning, from early activation-based approaches in DeepBind to modern attribution frameworks. Understand how base-pair-level contribution scores derived from chromatin accessibility models can reveal interpretable sequence features consistent with transcription factor binding, and how these inferred binding sites enable prediction of transcription factor occupancy across diverse cell types. Gain insights into the development of scalable and interpretable frameworks for decoding gene regulation, addressing fundamental questions about what deep neural networks learn and why they make particular predictions in the context of regulatory genomics.