OpenAI, PerturbAI and Scripps Research used thousands of GPT-Rosalind-powered agents to investigate existing mouse-brain experimental data. Their October 21 announcement describes a reasoning atlas connecting gene disruptions, cellular effects, scientific literature and possible biological mechanisms.
Divide the analysis into focused investigations
Each agent investigated a gene disruption, examined cellular responses, searched the literature and proposed explanations. Combining the reports gave scientists a way to compare related genes and identify unexpected responses that deserved closer attention.
The work analyzes an existing whole-brain CRISPR atlas. It should therefore be understood as a method for interpreting experimental evidence and generating follow-up questions, rather than a report that agents conducted new laboratory experiments or proved a treatment effective.
Compare the distributed approach with a single agent
OpenAI reports a comparison that kept the model and resources fixed as the workload increased. At the largest workload, distributed agents cited more than 700 references, while the single agent cited 11. In blinded comparisons, human experts preferred the distributed-agent reports 75% of the time.
Those are reported measures of the analyses and their expert assessment. More references or a preferred report does not independently establish that a proposed biological mechanism is correct. The announcement says further experiments are needed to test the leads.
Follow the hypothesis back to the observation
One example concerns DCTN1 and cholesterol regulation. The agents connected an observed change after gene disruption with previous research and proposed questions about cholesterol distribution within neurons. Another concerns different responses to Tsc1 and Tsc2 disruption across brain-cell groups.
For researchers exploring the reasoning atlas, the useful next step is to inspect the underlying observation, the cited literature and the proposed experimental test separately. The contribution is a way to examine more possible connections in a large dataset. Its value for disease biology depends on what subsequent experiments confirm, revise or reject.