Few-shot regulatory sequence design with a genomic language model

Prompting Evo 2 with tens of examples enables generative in-context learning to produce novel, diverse regulatory DNA sequences without additional training, including cell-type-specific enhancers in human cell lines.

• Why it matters: Generative biology needs ways to transfer knowledge from natural sequences and experiments into functional designs, but few-shot adaptation without retraining remains a key gap.
• What they did: The study prompts Evo 2 with tens of regulatory examples, analyzes mechanisms for induction-like genomic in-context learning, and further boosts conditioning using reinforcement learning with sequence-to-function rewards.
• The result: It generates enhancers driving cell-type-specific gene expression in three human cell lines, and in vivo designs label layer 5 extratelencephalic neurons in mouse cortex with high specificity and stronger activity than existing enhancers.

AI summary of the abstract · may contain errors · not medical advice · read the paper.