Researchers have introduced an open-source framework named Paper2Agent. The tool transforms academic reports into interactive AI agents that users can converse with. By providing a paper alongside its codebase, data, or supplementary materials, the system extracts core workflows. It then creates a tested and runnable toolkit for other datasets.
Google offers NotebookLM for uploading documents and chatting about their content. However, Paper2Agent goes further by running the actual methods described in the text. James Zou, a computer scientist at Stanford, notes that knowledge should be dynamic and interactive rather than static. This shift aims to improve reproducibility and spark new discoveries.
Zou and his team detailed the tool on 16 September in Nature. They tested Paper2Agent across statistics, econometrics, and astrophysics. Their proof-of-concept focus centered on computational biology. The team began with AlphaGenome, which is a deep-learning model predicting how DNA mutations affect gene regulation.
Feeding documentation and code into Paper2Agent took about 45 minutes on a personal laptop. The automated process cost less than 15 US dollars in computing power and produced 22 functional tools. A testing agent validated the tools by running sub-tools against reference results. The agent attempted up to six fixes per function when errors occurred before dropping broken tools.
The validated tools were packaged into a Model Context Protocol server and connected to Claude Code. This setup created an AlphaGenome agent that takes plain English questions, runs analyses, and returns results and visualizations. The agent outperformed standard Claude and a specialist AI co-scientist tool called Biomni during testing.
The researchers also linked multiple paper agents together, including AlphaGenome and papers on inherited DNA variants and gene silencing. Prompted to investigate psoriasis, the connected agents identified a gene called GPR137 as a likely causal factor. Human-selected validation confirmed that silencing GPR137 produced gene activity changes similar to the disease variant.
Dongping Chen from the University of Maryland calls the approach compelling. Olivier Elemento from Weill Cornell Medicine views the method as a major advance for the publication process. Artur Skowroński from VirtusLab suggests the tool could bring papers to life in classrooms.
Zou and his colleagues tested 100 computational biology papers with the system, and 26 failed due to incomplete code or missing documentation. Zou views these failures as a feature because they expose errors and missing information. Agentification acts as a certificate showing that work is well documented.
Zou and his team published a multi-agent drug development platform called Virtual Biotech in Science and released a Paper2Agent version of that study simultaneously. They also fed the Paper2Agent manuscript into the tool itself. That resulting agent now operates at paper2agent.ai.
