Paper2Agent turns a scientific paper, its code and its data into a validated AI agent that applies the paper’s method to new data. In a test using the AlphaGenome paper, it outperformed competing AI agents on accuracy and independently proposed a different candidate gene for a cholesterol-associated variant
Paper2Agent turns a scientific paper, its code and its data into a validated AI agent that applies the paper’s method to new data. In a test using the AlphaGenome paper, it outperformed competing AI agents on accuracy and independently proposed a different candidate gene for a cholesterol-associated variant
On September 16, Nature published a paper by James Zou and colleagues at Stanford University describing Paper2Agent. It uses a paper’s text, code and data to build an agent that answers questions about the work and collaborates with agents built from other papers.
Applying a paper’s method almost always requires overcoming technical barriers: finding the code repository, setting up the software environment and understanding its functions. These steps are also necessary when using AlphaGenome, a model developed by DeepMind to predict how DNA mutations affect genes.
Paper2Agent finds the paper’s code, sets up the environment, runs the authors’ tutorials and converts the steps into functions, retaining only those that reproduce the published numerical results and figures. This process reduces the risk of “code hallucination”, in which code produces a plausible but incorrect answer. It then makes these functions available through MCP, a standard protocol for connecting agents to external software.
Using the AlphaGenome paper, the pipeline built 22 working tools in 45 minutes at a cost of $14, without human intervention. On the most difficult questions, the agent answered correctly in 82,7% of cases, compared with 56,7% for Claude Code (Anthropic’s coding agent) with access to the repository and 72,2% for Biomni, a specialized biomedical agent. On routine questions, Paper2Agent was 2–4 times faster than both.
The authors asked the agent why a DNA variant discussed in the AlphaGenome paper was associated with “bad” cholesterol. The agent identified SORT1, the gene encoding sortilin, which transports cholesterol out of the liver. This differed from the paper’s identification of CELSR2 and PSRC1. The DNA variant affects the activity of all three genes, and both hypotheses were independently confirmed. According to Zou, the difference illustrates how Paper2Agent can help researchers reassess published conclusions without conducting a new experiment.
A more complex task involved agents built from three papers: one on AlphaGenome, one on how DNA regions affect genes and one on T-cells. Together, they searched for a gene within a DNA region associated with psoriasis risk. The AlphaGenome agent identified GPR137 as a candidate. The agents proposed ten ways to test this, and the researcher chose one: comparing gene activity after disabling the DNA region with gene activity after disabling GPR137. The similarity was significant only in activated T-cells, the cells that trigger psoriasis. The agents and the researcher developed this procedure together; it appeared in none of the three papers.
To test the method beyond a single favorable example, the authors applied it without modification to another 136 papers spanning computational biology, econometrics and game theory. They successfully built agents for 74 of 100 biology papers; incomplete code or insufficient data prevented them from building agents for the others. Outside biology, accuracy was 98,1%.
Paper2Agent’s code is openly available on GitHub, and the software can be installed as a skill for AI coding agents. Users can also build an agent for a paper on the Paper2Agent website. Paper2Agent follows The Virtual Lab, Zou’s earlier work on nanobodies against SARS-CoV-2. Its purpose is to make a paper’s method reusable, whereas the earlier project focused on proposing new hypotheses.
People still decide which hypotheses to test and how to interpret the results. Zou says that turning static papers into a living source of knowledge “could help us reimagine what knowledge will look like in the future”. The authors expect papers to include an “agent availability” section alongside their data and code availability sections.