Paper2Agent converts 74 of 100 biology papers into AI agents

Paper2Agent, described in a Nature paper published on September 16, turns a computational biology paper into an AI agent that exposes its methods, data, and code as callable tools. On a test batch of 100 papers, 74 were successfully agentified, and 593 of 599 generated tools passed validation.

The authors write that "when a paper describes a new computational method, substantial technical barriers often remain before the method can be used on new data." Paper2Agent produces what they call "virtual corresponding authors": Model Context Protocol servers that wrap a manuscript's code and workflows so a user can query them in natural language.

On genomics, the AlphaGenome agent generated 22 validated tools in about 45 minutes for $14, and scored 98.7 ± 1.3% accuracy on tutorial queries and 100.0% on novel queries. That beat Claude given direct repository access, which came in at 82.7% and 78.7%, and cut median runtime by 1.9 to 3.1 times. The Scanpy agent, for single-cell analysis, produced 7 validated tools in the same time for $13 and reproduced human researcher results across four datasets.

The team also had agents talk to each other. AlphaGenome, CRISPR screening and Perturb-seq agents converged on GPR137 as the causal gene for a psoriasis-associated variant, rs887314, with validation showing "significant concordance with the CRE perturbation signature under stimulated conditions." Five researchers we track posted the paper's link the week it went up.

Not every paper converts. Common failure modes are missing executable code, incomplete documentation and non-generalizable scripts. The authors argue that agentification success itself can serve as "a practical measure of reproducibility."