Mine biological datasets for new leads with thousands of agents
OpenAI describes running thousands of AI agents over complex biological data to uncover new research leads.
Done with a custom-built or unnamed agent
- What they did
- OpenAI, PerturbAI and Scripps Research split analysis of an existing whole-brain CRISPR atlas of mouse brain cells across thousands of GPT-Rosalind agents. Each agent took one gene disruption, analyzed the cellular responses, searched the literature and proposed mechanisms. Their reports were merged into a searchable reasoning atlas. Blinded human experts then judged the outputs.
- How it went
- The run took 32 minutes, against an estimated months to a year for a person. Experts preferred its reports over a single agent's 75% of the time, and it surfaced leads on Tsc1/Tsc2 and DCTN1-linked cholesterol. These leads still need lab testing.
- Worth knowing
- A single agent given the same workload cited only 11 references, against more than 700 for the distributed agents, so splitting work per perturbation mattered.
Try it yourself with Claude Code
I have a dataset of [type of biological data, e.g. gene expression CSVs] in [folder path]. Write and run analysis scripts to look for patterns, outliers, or correlations that could be new research leads about [your research question]. Finish with a ranked list of the five most promising leads, each with the evidence behind it, the method used, and caveats, and save the scripts so I can rerun them.
Want help getting this working? Join the Vibe Coding Academy ↗ Run by Alex Finn, who also runs this site.
Discussion on HN · Oct 21, 2026
Related use cases
Find a Linux vulnerability hidden for 23 years
An account of Claude Code uncovering a long-standing vulnerability in Linux.
Generate parametric CAD models from a prompt
An open-source text-to-CAD tool that produces 3D models from text and image references by generating CAD as code.
Control and calibrate a robotic hand with OpenClaw
An OpenClaw agent running Claude drove a 16-joint printed hand, checked its actions with a USB camera, calibrated firmware and narrated in Telegram.
Reverse engineer a classic game ROM via Ghidra MCP
The author connected Claude to Ghidra through MCP to analyze and rename functions in an Atari River Raid ROM.
Open recent protein-ligand structures in PDB
A researcher had Comet find the three latest GSK3B protein-ligand complexes in the PDB and open each in Ligand Interaction view.
Orchestrate a Spot robot through multi-step tasks
Gemini Robotics ER 2 drives a Boston Dynamics Spot, watching video to confirm tasks like tying a trash bag are done before moving on.
Get 5 like this every morning
The best things people got an AI agent to do, each with the prompt to try it.