Toward generalizable and interpretable AI in regulatory genomics.

Deciphering how DNA sequence encodes gene regulation remains a central challenge in biology. Advances in machine learning and functional genomics have enabled sequence-to-function (seq2func) models that predict molecular regulatory readouts directly from DNA sequence, supporting variant effect prediction, mechanistic interpretation and regulatory sequence design. Despite strong performance on held-out genomic regions, generalization across genetic variation and cellular contexts remains inconsistent. In this Review, we examine how model architectures, training data and prediction tasks shape model behavior. We also synthesize how interpretability methods and evaluation practices have elucidated cis-regulatory organization and highlighted systematic failure modes, clarifying why strong predictive accuracy can fail to translate into robust regulatory understanding. Thus, we suggest that progress requires reframing seq2func models as continually refined systems, in which targeted perturbation experiments, systematic evaluation and iterative model updates are tightly coupled through artificial intelligence-experiment feedback loops, enabling self-improving models that progressively deepen mechanistic understanding and more reliably support biological discovery.