A new artificial-intelligence tool quickly transforms a research paper into a bespoke AI agent that serves as a “virtual corresponding author”. The agent can respond immediately to questions about the paper, providing a convenient way for scientists to grasp advances in unfamiliar fields. It can also apply the paper’s methods to a fresh data set and even collaborate autonomously with agents for papers from other scientific disciplines.
The ability of this tool, called Paper2Agent, to convert static papers into dynamic sources of information “can help us to reimagine what knowledge looks like in the future”, says James Zou, a computer scientist at Stanford University in California and co-author of the paper, which was published today in Nature1.
‘Living’ papers
Agents are AI assistants that can reason and carry out complex tasks. The Paper2Agent tool starts by accessing a paper’s main text, code, data sets and other elements. The information is deposited onto a digital platform called an MCP server. Then a team of AI agents autonomously builds tools that can apply the paper’s methods to fresh data and places those tools on the server as well.
Scientists can then connect to the server using a large language model (LLM) of their choice. This creates a paper-specific agent that can scientists can interact with in plain, conversational language.
Can AI review the scientific literature — and figure out what it all means?
Zou and his team tested the technology on the paper2 that introduced AlphaGenome, an AI model that predicts the properties of DNA sequences, such as their effect on gene expression. Paper2Agent autonomously created an agent for the AlphaGenome paper in about 45 minutes, and the required computing power cost US$14. The agent passed its initial test with flying colours, answering genetics questions with near-perfect accuracy. The agent also bested the scores of other top biomedical AI agents that were given access to the same paper and asked the same questions. Among the outscored agents was Biomni, a tool developed by academic researchers that draws from dozens of databases but scored far lower than the AlphaGenome agent. Zou says the paper agent’s success comes from its mastery of AlphaGenome’s tools and abilities.
Next, the team asked the agent to determine why a single change to a DNA ‘letter’ in a genetic sequence is associated with ‘bad’ cholesterol. The authors asked the agent to identify the precise gene that could explain this link. It identified a gene — a different causal gene to the one pinpointed in the original AlphaGenome paper. Zou says that AlphaGenome’s data on genetic variants support both hypotheses. The discrepancy, he adds, highlights a strength of Paper2Agent: scientists can use the tool to re-evaluate published conclusions without designing entirely new experiments.
Strengths and weaknesses
