Peter Koo: Interpreting and Designing Regulatory DNA with Deep Learning

Speaker:
Peter Koo (CSHL, US)

Title:
Interpreting and Designing Regulatory DNA with Deep Learning

Abstract:
Genomic AI models have achieved strong performance in predicting functional genomics data, suggesting they capture key features underlying gene regulation. However, their black-box nature makes it challenging to determine whether they truly learn meaningful biological mechanisms. In this talk, I will present a suite of perturbation-based interpretation methods that leverage in silico mutations to identify the sequence features and combinatorial interactions that drive regulatory activity. These methods support interpretation across multiple scales—from individual nucleotides to entire cis-regulatory elements—revealing complex dependencies and generalizable cis-regulatory principles. In the second part of the talk, I will introduce DNA Discrete Diffusion (D3), a generative AI framework based on discrete diffusion for designing regulatory DNA sequences with tunable activity levels. D3 enables cell type-specific sequence generation, captures diverse regulatory patterns, and enhances the performance of predictive models when used to augment training data in low-data regimes. Together, these approaches demonstrate how predictive and generative AI can advance biological discovery by deepening our understanding of gene regulation and enabling the machine-guided design of regulatory sequences.

Lecture Series:
SysBio Lecture Series: AI for Systems Medicine