Paper2Agent Turns Scientific Papers into AI Agents for Research
AI agents representing papers could foster cross-disciplinary collaboration.
Credit: Getty
New artificial intelligence tools can quickly transform research papers into bespoke AI agents that act as “virtual corresponding authors”. These agents can respond immediately to questions about papers, helping scientists keep up with advances in unfamiliar fields. They can also apply a paper’s methods to new datasets and collaborate autonomously with paper agents from other scientific disciplines.
The tool, called Paper2Agent, transforms static papers into dynamic information sources. It “helps us reimagine the future of knowledge”, said James Zou, a computer scientist at Stanford University in California and co-author of the paper. Nature1
How Paper2Agent creates a “living” research paper
AI agents are assistants that can reason and perform complex tasks. Paper2Agent begins by accessing a paper’s text, code, datasets and other elements. It stores this information on a digital platform called an MCP server.
A team of AI agents then autonomously builds tools that can apply the paper’s techniques to new data. Those tools are also placed on the server.
Scientists can connect to the server using a large language model (LLM) of their choice. This creates a paper-specific agent that scientists can interact with in plain conversational language.
Can AI review scientific literature and figure out what it means?
Testing Paper2Agent with AlphaGenome
Zou and his team tested the technology on a paper about AlphaGenome.2 AlphaGenome is an AI model that predicts properties of DNA sequences, including their effects on gene expression.
Paper2Agent autonomously created an agent for the AlphaGenome paper in about 45 minutes. The process required US$14 of computing power.
The agent passed its first test with near-perfect accuracy, answering genetics questions about the paper. It outscored other leading biomedical AI agents that accessed the same papers and were asked the same questions. One of the agents that scored better was Biomni, a tool developed by academic researchers.
Although Biomni uses dozens of databases, it scored much lower than the AlphaGenome agent. Zou said Paper2Agent’s success lies in its mastery of AlphaGenome’s tools and capabilities.
Using AI agents to reassess scientific conclusions
The team then asked the agents to investigate why a single change to a DNA “letter” in a gene sequence was associated with “bad” cholesterol. The researchers asked the agents to identify the exact genes that could explain this association.
The agent identified a different causative gene from the one identified in the original AlphaGenome paper. Zou says that AlphaGenome’s data on genetic variation supports both hypotheses.
This contradiction highlights one of Paper2Agent’s strengths. Scientists could use the tool to reevaluate published conclusions without having to design entirely new experiments.
Advantages and disadvantages
Source: www.nature.com