Paper2Agent turns 74 of 100 biology papers into AI agents

Paper2Agent, a framework described in Nature by Jiacheng Miao, James Zou and colleagues, converts research papers into AI agents that act as "virtual corresponding authors," exposing the manuscript, code, datasets and workflows behind a natural-language interface. Across 100 computational biology papers the pipeline agentified 74, producing 599 tools of which 593 passed automated validation.

The authors demonstrate the approach on three published methods: AlphaGenome for genomic variant interpretation, Scanpy for single-cell analysis and TISSUE for spatial transcriptomics. Building the AlphaGenome agent yielded 22 MCP tools in about 45 minutes at a cost of roughly US$14 on a personal laptop; the Scanpy build ran about US$13. On tutorial-derived queries the AlphaGenome agent scored 98.7 ± 1.3% accuracy and hit 100.0 ± 0.0% on novel queries, against 82.7 ± 3.4% for a "Claude + Repo" baseline and 37.3 ± 4.0% for Biomni.

"Paper2Agent reimagines research dissemination by turning static papers into active AI agents," the authors write. "Each agent serves as an interactive expert on the corresponding paper." As a proof of scientific utility they report that multiple paper agents worked together to prioritize GPR137 as a causal gene for psoriasis.

The system is built on the Model Context Protocol and orchestrated through Claude Code's agent SDK, with sub-agents that handle environment setup, tutorial discovery, execution, tool extraction and test verification. On a broader 300-question benchmark the framework averaged 91.2 ± 1.6% accuracy, query runtime dropped 1.9× to 3.8× relative to baseline methods, and per-query cost sat near US$0.20 versus US$0.38 for the Claude + Repo comparison.

The full evaluation spans 100 biology papers, 26 data and discovery papers and 10 non-biology papers, 136 total. Five researchers we follow have shared the link since it appeared last week.