秀色直播

Dissecting gene regulation syntax using glass-box machine learning

Selin Jessa, Stanford University
Tuesday April 29, 12-1pm
Zoom Link:听
In Person: 550 Sherbrooke, Room 189

Abstract:听Most cells in the body share an identical genome, but gene regulation during development ensures that distinct cell types activate appropriate gene programs in the right place and time. This process is directed by transcription factors (TFs), which recognize and bind short DNA sequences in non-coding genomic elements, and drive cell type-specific gene expression programs. Yet, the map of TFs that bind the genome in each cell type remains incomplete. Furthermore, the syntax of binding sites鈥攖heir composition, orientation, and spacing鈥攃ontributes to cell type-specific regulation, just as word organization in a sentence impacts its meaning. However, we lack a systematic understanding of how this syntax mediates combinatorial activity of TFs.

I will introduce a "glass-box" deep learning strategy to study TF activities during human development. TFs typically bind physically accessible or uncompacted DNA, and in turn, they often promote local DNA accessibility. We thus train convolutional neural networks to predict DNA accessibility using local sequence alone. We use a model interpretation algorithm to extract the sequence features which are predictive of DNA accessibility, which generally reflect TF binding site motifs. We apply this strategy to nearly 200 human fetal cell types to define the TF motifs relevant in human development. Finally, we use these trained models to perform in silico experiments to identify synergistic TF motif pairs and systematically dissect the role of binding site syntax. Altogether, our approach reveals novel motif rules which are encoded in DNA and mediate the combinatorial, context-specific regulation that governs development.