Toward generalizable and interpretable AI in regulatory genomics

Area of Science:

• Genomics
• Computational Biology
• Molecular Biology

Background:

• Deciphering DNA's role in gene regulation is a key biological challenge.
• Machine learning and functional genomics have led to sequence-to-function (seq2func) models for predicting regulatory readouts from DNA sequence.
• These models aid in variant effect prediction, mechanistic interpretation, and regulatory sequence design.

Purpose of the Study:

• To review how model architectures, training data, and prediction tasks influence seq2func model behavior.
• To synthesize how interpretability methods and evaluation practices reveal cis-regulatory organization and model limitations.
• To propose a framework for advancing seq2func models through AI-experiment feedback loops.

Main Methods:

• Review of existing literature on seq2func models, including their architectures, training data, and prediction tasks.
• Synthesis of interpretability methods and evaluation practices applied to seq2func models.
• Analysis of model generalization across genetic variation and cellular contexts.

Main Results:

• Seq2func models show strong performance on held-out regions but inconsistent generalization.
• Interpretability methods have elucidated cis-regulatory organization but also highlighted systematic failure modes.
• Current practices show that high predictive accuracy does not always ensure robust regulatory understanding.

Conclusions:

• Progress in seq2func modeling necessitates reframing them as continually refined systems.
• Tightly coupling targeted experiments, systematic evaluation, and iterative model updates via AI-experiment feedback loops is crucial.
• This approach enables self-improving models for deeper mechanistic understanding and more reliable biological discovery.