The Pulse
Paper2Agent Turns 74 of 100 Research Papers Into AI Agents
Nature reports that Paper2Agent successfully agentified 74 of 100 computational biology papers, converting their methods, code and data into interactive systems connected through Model Context Protocol servers.
AI.info Team ·
Paper2Agent successfully converted 74 of 100 computational biology papers into working AI agents, turning published methods into software that can answer questions, rerun analyses and process new data. The Stanford-led system also produced 599 proposed tools from those papers, with 593 passing automated validation.
The results, published in Nature on September 16, 2026, point to a practical use for AI agents in scientific publishing: packaging a paper’s manuscript, code, datasets and workflows into an executable system rather than leaving them as disconnected research materials.
Paper2Agent does not treat a paper as a document to summarize. It builds a Model Context Protocol, or MCP, server around the paper’s methods and resources, then connects that server to an AI agent. Users can ask for an analysis in ordinary language while the agent calls validated tools from the underlying research code.
From a PDF to a working research system
Computational papers often require readers to find a code repository, install dependencies, identify the right inputs, reconstruct undocumented steps and adapt scripts before they can use the published method. Paper2Agent automates much of that work through a six-stage pipeline.
The system locates the paper’s codebase, creates an isolated software environment, scans for tutorials, executes example workflows, extracts reusable functions and tests those functions against the original outputs. The resulting tools are assembled into an MCP server that can be hosted remotely and connected to compatible AI agents.
Each server contains executable tools, static resources and prompts. Tools perform operations such as scoring a genetic variant or processing single-cell data. Resources hold the manuscript, supplementary material, code, datasets and figures. Prompts encode multi-step procedures so an agent can follow the intended order of operations instead of improvising a workflow from scratch.
The researchers used Claude Code to implement the system and Claude Sonnet 4 as the downstream agent in their case studies. Paper2Agent does not require the final agent to retain direct access to the original repository during evaluation, which allowed the team to test whether the generated tools preserved the paper’s functionality.
AlphaGenome exposes the performance gap
The strongest individual results came from an agent built around AlphaGenome, a model for predicting how DNA variants affect gene regulation. Paper2Agent generated 22 MCP tools for AlphaGenome in about 45 minutes at a reported cost of $14 on a personal laptop. All 22 passed automated validation.
In benchmark tests, the AlphaGenome agent reached 98.7% accuracy on 15 tutorial-derived queries and 100% on 15 new queries across five independent runs. Claude Code with direct repository access scored 82.7% and 78.7% on the same categories, while Biomni scored 37.3% and 56%.
On 30 open-ended queries requiring multiple tools and biological interpretation, Paper2Agent reached 82.7% accuracy. Claude Code with repository access scored 56.7%, and Biomni scored 72.2%. The Paper2Agent system also reduced median runtime compared with both alternatives.
The tests covered tasks such as estimating the effect of a DNA variant on gene expression, splicing or chromatin accessibility in particular tissues and cell types. Each generated tool included a traceable reference to the original source code, giving users a way to inspect where an operation came from.
Scanpy shows how agents can run a full workflow
A second case study used Scanpy, a widely used package for single-cell transcriptomics. Paper2Agent generated seven tools for preprocessing and clustering in about 45 minutes at a reported cost of $13, with every tool passing validation.
The resulting agent could perform quality control, normalize data, select features, reduce dimensionality, construct a cell-neighbor graph, cluster cells and identify cell types. The researchers encoded the sequence in MCP prompts, allowing users to provide a dataset path without manually specifying every step.
When tested on four public single-cell datasets outside the Scanpy repository, the agent produced outputs that matched human-run analyses in cell and gene counts after quality control and in the top marker genes identified for each cluster. Tests on seven diverse datasets showed that the agent adjusted parameters based on the characteristics of each dataset.
Paper2Agent also created an agent for TISSUE, a method for spatial transcriptomics. The authors used the result to show how a paper agent could provide uncertainty-aware analysis while preserving access to the source method and its supporting data.
The 26 papers that could not become agents
Paper2Agent’s large-scale evaluation also exposed the condition of scientific software. Among the 100 computational biology papers, 26 could not be successfully agentified. The reported causes included missing executable code, missing data or model files, dependency and environment failures, and scripts that did not generalize beyond their original examples.
The researchers describe that failure rate as a possible reproducibility signal. A paper that cannot be installed, executed and tested may still communicate a scientific result, but its method is harder for others to reuse. The system’s validation loop can repair some inherited bugs, yet it cannot supply missing research artefacts or resolve every undocumented assumption.
Across 300 tutorial-derived questions, Paper2Agent with Sonnet 4 achieved 91.2% accuracy, compared with 80.3% for Claude Code working directly with the paper and repository. Each Paper2Agent query cost about 20 cents and took a median of 1.6 minutes, compared with 38 cents and 4.3 minutes for the direct-repository approach.
The evaluation extended beyond biology. Across 42 execution-based tasks drawn from 10 papers in areas including AI, statistics, econometrics, game theory and astrophysics, Paper2Agent reached 98.1% accuracy. For 26 data- and discovery-focused papers, its resource layer achieved 89% accuracy on 100 synthesis questions while costing less and running faster than a browser-use baseline.
Three paper agents test a psoriasis hypothesis
The researchers then connected separate paper agents to study a psoriasis-associated genetic variant. AlphaGenome predicted that GPR137 was the top affected gene at the rs887314 locus in CD4-positive T cells, assigning it an RNA-sequencing quantile score of 0.997.
An AI co-scientist cross-referenced that prediction with two other agents: one representing a screen that measured gene-expression changes after regulatory-element perturbation, and another representing a Perturb-seq study of gene knockdowns in human CD4-positive T cells.
The system proposed ten validation strategies. A human researcher selected signature-correlation analysis, which compared the expression pattern caused by regulatory-element perturbation with patterns produced by knocking down candidate genes. GPR137 was the only top-ranked candidate showing significant agreement under stimulated conditions, with Spearman correlations of 0.613 and 0.630 in two stimulation settings.
The result supports GPR137 as a probable causal gene at the locus, but the paper does not present the finding as an experimentally established mechanism. The correlation was not significant in resting cells, and the authors describe the role as activation-dependent.
Why human judgment still matters
Paper2Agent can generate hypotheses, propose tests and execute analyses, but the researchers place responsibility for choosing scientific directions and evaluating evidence with human investigators. They describe benchmark accuracy mainly as a measure of faithful execution, not proof that an agent’s open-ended interpretation is scientifically correct.
That distinction matters because the system can make published work easier to reuse without deciding whether a result is biologically meaningful. A paper agent may reproduce an analysis faithfully and still inherit limitations in the original data, design or interpretation.
The authors propose that journals could eventually add an “agent availability” section alongside existing data and code statements. Such a section would identify whether a paper has an interactive agent, what the agent can execute and which dependencies require maintenance.
Paper2Agent makes the publication process more demanding as well as more accessible. Agents depend on functioning repositories, documented environments and maintained upstream software. They also introduce security, intellectual-property and attribution questions that do not arise in the same way from a static PDF. The study’s clearest result is therefore both technical and editorial: of 100 computational biology papers tested, 26 lacked the ingredients needed to become reliable working agents.