Decoding the Mechanistic Impact of Genetic Variation on Regulatory Sequences with Deep Learning

Evan Seitz, David M. McCandlish, Justin Kinney, Peter K Koo

Track: Machine learning: computational method and/or computational results

Nature Biotechnology: Yes

Keywords: Deep learning, cis-regulatory code, genetic variation, transcription factor binding, regulatory sequences, mechanism-based attribution, evolutionary pathways, synthetic sequence design, functional genomics

TL;DR: SEAM is an AI-driven tool that systematically maps how genetic mutations reshape regulatory mechanisms, revealing novel sequence reprogramming pathways and advancing functional genomics.

Abstract: Non-coding DNA encodes complex cis-regulatory mechanisms that govern gene expression by orchestrating transcription factor binding within specific sequence contexts. While deep learning has advanced our understanding of these mechanisms, how genetic variation reconfigures them remains an open challenge. Here, we introduce SEAM, an AI-driven tool that systematically investigates how mutations reshape regulatory mechanisms. By mapping sequences into a mechanism space and clustering them based on shared features, SEAM reveals how specific mutations can reprogram regulatory DNA, driving mechanistic and functional diversity. SEAM highlights the remarkable evolvability of human regulatory elements, disentangles transcription factor-specific effects from broader sequence context, and provides a powerful framework for decoding the cis-regulatory code. By enabling systematic, unbiased exploration of reprogrammable mechanisms, SEAM illuminates evolutionary pathways and informs the rational design of synthetic sequences with tailored functions.

Anonymization: This submission has been anonymized for double-blind review via the removal of identifying information such as names, affiliations, and identifying URLs.

Presenter: ~Evan_Seitz1

Format: Yes, the presenting author will attend in person if this work is accepted to the workshop.

Funding: No, the presenting author of this submission does *not* fall under ICLR’s funding aims, or has sufficient alternate funding.

Submission Number: 80