Modern biology generates massive datasets, especially from technologies like single cell RNA sequencing, which measures gene expression in individual cells. While these datasets contain valuable information, analyzing them fully can be challenging and often depends on the experience and creativity of the researcher.

Researchers at Stanford University have developed a new AI system called CellVoyager to help explore these complex datasets more effectively. Instead of simply following instructions, CellVoyager can independently analyze data, suggest new directions, and generate hypotheses based on what has already been studied.

Schematic of the CellVoyager agentic framework

A. Overall schematic of CellVoyager (1) The manuscript text, associated scRNA-seq data, and a Python environment specification are given as input. The biological background and prior analyses performed in the manuscript are extracted and a Jupyter notebook is initialized. (2) Exploration blueprints (consisting of a hypothesis and a stepwise plan) are generated. The agent self-critiques this blueprint, incorporating any suggested improvements. (3) Each hypothesis is explored for T steps. Each step involves generating and executing code, with up to F steps of bug fixing, replanning based on the code execution output, and self-critiquing the new plan. This process can be repeated for N distinct hypotheses, with the agent being given past hypotheses it tested to prevent it from repeating the same analyses. A final report is generated that summarizes noteworthy findings from the N exploration trajectories. B. Overview of CellVoyager analysis and interpretation module. The self-critiqued plan contains python code for the next step of the analysis. This code is executed and the outputs from running the code, along with the contents of past code cells, are fed into a vision-language model (VLM) for interpretation. The VLM outputs a summary of the outputs, suggested future directions based on promising results, and possible ways to improve the analysis; these inform how the agent decides to choose the next step in the analysis and CellVoyager updates its analysis plan accordingly. C. Basic analysis block of CellVoyager in a Jupyter notebook, consisting of a description of the analysis step, the code to carry out the analysis, the outputs and figures, and the interpretation of the outputs and figures. D. An example of an analysis block of CellVoyager, with code omitted for brevity.

Traditional AI tools in biology typically rely on user input to decide what analyses to run. This means that important patterns can be missed if researchers do not think to look for them. CellVoyager addresses this limitation by reviewing both the dataset and previous analyses, then proposing new experiments or interpretations that build on existing work.

The system operates within a computational environment similar to a Jupyter notebook, where it can test ideas, run analyses, and refine its approach. It uses large language models to guide its reasoning and decision making, allowing it to connect biological context with computational methods.

To evaluate its performance, the researchers tested CellVoyager using a benchmark dataset built from 50 previously published single cell RNA sequencing studies. The system was asked to predict which analyses researchers would perform based only on background information. It outperformed other AI models, including GPT-4o, by a significant margin.

More importantly, CellVoyager was able to generate new biological insights. In one example, it identified that CD8 positive T cells in COVID-19 may be more prone to a specific type of inflammatory cell death called pyroptosis. In another case, it discovered a link between increased variability in gene expression and aging in a specific region of the brain.

These findings were reviewed by the original study authors, who rated many of the AI generated hypotheses as both scientifically valid and interesting. This suggests that AI systems like CellVoyager can complement human expertise by uncovering patterns that might otherwise go unnoticed.

The broader implication is that AI driven tools could help researchers extract more value from existing datasets without requiring entirely new experiments. As the volume of biological data continues to grow, approaches like this may become essential for accelerating discovery.

By combining artificial intelligence with single cell RNA sequencing, CellVoyager represents a step toward more autonomous and collaborative data analysis in computational biology.

Availability – Code for running the agent on case studies and for CellBench evaluation is available in the CellVoyager GitHub repository: https://github.com/zou-group/CellVoyager.