뉴스&공지 KAIST Graduate School of Engineering Biology

Toward interpretable and Generalizable AI for Regulatory Genomics

Sequence-to-function models are transforming functional genomics by learning predictive relationships between regulatory DNA sequence and molecular phenotypes such as chromatin accessibility, transcription factor binding, and gene expression. These models have expanded our ability to read the regulatory genome, prioritize functional variants, design synthetic regulatory sequences, and formulate mechanistic hypotheses at scale. However, their biological utility depends on whether their predictions can be interpreted, tested, and generalized beyond the contexts represented in training data. 

In this talk, I will discuss recent progress in interpretable deep learning for regulatory genomics, with an emphasis on methods for extracting cis-regulatory mechanisms from model predictions and attribution maps. I will also examine current limitations in generalization, experimental grounding, and model evaluation, and outline directions toward more reliable AI systems for understanding and engineering regulatory DNA.