Nearly 2,000 GPT-Rosalind agents processed mouse-brain results from 1,947 gene disruptions in 32 minutes, giving OpenAI, PerturbAI and Scripps Research leads for further lab work.
Scientists used CRISPR to disrupt 1,947 genes linked to disease, then used single-nucleus RNA sequencing to measure activity inside cell nuclei. This Perturb-seq method combines gene edits with RNA readings from about 7.7 million cells. GPT-Rosalind compared those measurements with scientific literature only after the experiments were completed.
The system grouped gene alterations, cell responses and related studies into a searchable atlas. Researchers estimated that a person reviewing comparable results gene by gene might need several months to a year; they did not time a human analyst doing the same work. The largest workload comparison produced more than 700 references from the distributed setup and 11 from a single agent.
Experts assessed reports without knowing which system produced them and chose the distributed agents’ output in 75% of evaluations. This preference score does not establish whether the biological explanations are accurate.
A DCTN1 disruption, involving a gene linked to Perry syndrome, coincided with stronger activity among genes responsible for cholesterol synthesis. GPT-Rosalind raised cholesterol transport as a possible mechanism for follow-up experiments, but the preprint offers a hypothesis rather than an established drug target.
Sarah Teichmann, who is a founding co-chair of the Human Cell Atlas and holds a stem cell medicine chair at the University of Cambridge, said AI could help researchers “turn unexpected observations into new questions to test.”
The planned drug-target effort will draw on PerturbAI’s in-vivo data, with OpenAI as a partner. OpenAI acquired Orchestra Bio on Oct. 20; the company develops software that helps coordinate drug-development programs, as described in an earlier EconoTimes report on the deal.
