Paper2Agent converts research papers into interactive AI agents

Paper2Agent, a framework introduced in Nature, converted 74 computational biology papers into AI agents, generating 593 validated tools from 599 proposed. Each agent packages a paper's manuscript, code, datasets and workflows into what the authors call a "virtual corresponding author, exposing its manuscript, supplementary materials, datasets, code and workflows as active, agent-native knowledge rather than static text."

Jiacheng Miao, Joe R. Davis, Yaohui Zhang, Jonathan K. Pritchard and James Zou describe a pipeline that locates a paper's repository, sets up its environment, scans its tutorials, executes them, extracts working functions, and assembles them into a Model Context Protocol server. Building the AlphaGenome agent cost about US$14 and produced 22 validated tools; the Scanpy single-cell library produced 7. Processing time was roughly 45 minutes per paper.

On 300 tutorial questions the framework scored 91.2% ± 1.6% accuracy against a Claude + Repo baseline of 80.3% ± 2.3%. Novel AlphaGenome queries scored 100.0% ± 0.0% and open-ended questions 82.7% ± 2.4%. The authors report cutting "median runtime by 1.9× and 3.1× relative to Claude + Repo and Biomni, respectively."

"MCP is a standardized protocol that allows structured APIs and tools to be exposed in a way that is directly accessible to LLMs and agent frameworks," the authors write. In one case study, the AlphaGenome agent prioritised *SORT1* as a causal gene for an LDL cholesterol variant, validated against GTEx eQTL data at P = 1.1 × 10⁻⁶⁵.

The evaluation leans biological: 100 computational biology papers, 26 data-focused papers, and 10 non-biology papers. The larger claim reaches past individual reuse: "Once scientific knowledge is encoded in active agents rather than static artefacts, the potential extends beyond individual use." Five researchers we track in our Who's Who shared it.