Paper2Agent converts 74 of 100 biology papers into AI agents

Paper2Agent, a framework from a Stanford-led team, converted 74 of 100 computational biology papers into working AI agents that answer natural-language questions by executing the papers' own code. The system generated 599 candidate tools, of which 593 passed automated validation, at roughly $13-14 and 45 minutes per paper. The work appears in Nature on September 16, 2026, authored by Jiacheng Miao, Joe R. Davis, Yaohui Zhang, Jonathan K. Pritchard and James Zou.

On a tutorial-derived benchmark, Paper2Agent's agents hit 91.2% accuracy versus 80.3% for approaches that give a model direct repository access. The team's AlphaGenome agent, built as 22 tools, scored 98.7% on tutorial questions and 100% on novel genomic queries, while running 1.9× faster than manual use of the underlying repository. Five researchers in our tracked expert set circulated the paper within its first days of publication.

The authors also show three of these paper-agents chained together, converging on GPR137 as the likely causal gene behind psoriasis-associated variant rs887314 by integrating AlphaGenome computational predictions with MPRA-coupled scCRISPRi perturbation data. The paper reports "significant concordance" with independent experimental evidence.

About 26% of the sampled papers could not be converted at all, mostly because of missing codebases, incomplete documentation, or unresolvable environments. The authors argue this failure rate "may itself serve as a practical measure of reproducibility," and propose journals establish an "agent availability" section alongside existing data and code requirements.