Paper2Agent turns 74 of 100 biology papers into AI agents

The authors of a Nature paper published September 16 generated an AI agent from the AlphaGenome codebase in about 45 minutes on a personal laptop, for $14. The agent shipped with 22 tools; all passed automated validation.

Parametrised in plain English, Paper2Agent scans a paper's linked repository, runs the tutorial code, and wraps the working parts as Model Context Protocol servers. "Paper2Agent transforms research output from passive artefacts into active systems that accelerate use and discovery," write Jiacheng Miao, James Zou and colleagues.

On queries drawn from the AlphaGenome tutorial the agent scored 98.7 ± 1.3%; on novel queries it scored 100.0 ± 0.0%. Handing Claude the same repository directly produced 82.7 ± 3.4% and 78.7 ± 4.4%. Scaled up, "74 were successfully agentified, yielding 599 proposed tools, of which 593 passed automated validation" across a set of 100 computational biology papers.

The more interesting result is what happens when several of these agents work in concert. Three of them, built from separate publications, cross-referenced each other and flagged GPR137 as the probable causal gene at a psoriasis-associated variant. "Only GPR137 knockdown showed significant concordance with the CRE perturbation signature under stimulated conditions," the authors report, at Spearman correlation 0.613 and P = 3.79 × 10−3.

Tools are locked once they pass validation. "This design mitigates the risk of 'code hallucination', where executing inaccurate LLM-generated code could lead to incorrect scientific results," the paper says. By the following week five researchers on our Who's Who tracker had shared the link.