Paper2Agent turns 74 of 100 biology papers into AI agents

Paper2Agent, a framework described in a new Nature paper by Jiacheng Miao, Joe R. Davis, Yaohui Zhang, Jonathan K. Pritchard and James Zou, turned 74 of 100 computational biology papers into interactive AI agents that answer questions and run analyses on demand. Across 300 tutorial-derived benchmark questions, the agents scored 91.2 ± 1.6% accuracy, outperforming Claude Code pointed at the same repositories.

The system, the authors write, shifts 'research output from a document or codebase encoding knowledge to a knowledgeable entity capable of execution and dialogue.' The pipeline downloads a paper's codebase, sets up the environment, extracts tools from tutorials, validates them against reference outputs, and packages the result as a Model Context Protocol server that chat agents can call.

The AlphaGenome case study was generated in about 45 minutes for $14 on a personal laptop, produced 22 tools, and hit 98.7 ± 1.3% accuracy on 15 tutorial-derived queries. In the same benchmark, Claude with direct repo access scored 82.7% and Biomni scored 37.3%. A Scanpy agent produced 7 tools in about 45 minutes for $13, with outputs the authors say 'match those produced by human researchers.'

Of 599 proposed tools, 593 validated, meaning the framework's own tests locked them against reference outputs before shipping. The paper also chains three paper-agents to identify GPR137 as a causal gene for psoriasis; the agents, the authors write, 'integrated computational predictions with independent experimental validation.'

Five researchers we track have shared the paper's link so far.