Researchers at OpenAI, PerturbAI and Scripps Research used AI to turn existing mouse-brain results into leads for laboratory tests in 32 minutes, but the preprint does not show a wider gain in scientific productivity.
OpenAI’s Oct. 21 announcement described the collaboration and its timing. The bioRxiv preprint specifies that 1,946 GPT-Rosalind agents handled the analysis, which connected gene-disruption results with scientific literature.
A separate preprint on the underlying atlas covers 1,947 disease-associated gene disruptions across more than 7.7 million cells in postnatal mouse brains. Those figures count different things: gene targets in the atlas and AI agents in the later analysis.
One lead involved Dctn1 disruption and increased expression of genes involved in cholesterol production. The researchers raised a possible connection between cholesterol transport and neurodegeneration, but said experiments are needed to test the idea.
The authors project that a person working through the perturbations and literature one by one would need months to a year. That estimate uses assumed time for reviewing each perturbation and paper; it was not measured in a timed comparison with human researchers. In the largest workload, the parallel system cited more than 700 sources, while a single-agent setup cited 11.
The preprint reports a blinded evaluation of 20 matched report pairs. Across 60 expert judgments, reviewers preferred the distributed-agent interpretations in 45, favored the single-agent versions in 10 and tied in five. Those preferences compare the reports; they do not establish whether the biological explanations were accurate.
The work addresses the challenge of reviewing large experiments and connecting results to scientific literature. Sarah Teichmann, chair in stem cell medicine at the University of Cambridge and founding co-chair of the Human Cell Atlas, said AI tools could help researchers search large datasets and choose findings for follow-up.
The productivity question has a longer history. A 2020 American Economic Review study by Nicholas Bloom, Charles I. Jones, John Van Reenen and Michael Webb found that research effort rose while productivity fell across several industries and products. The authors estimated that maintaining the early 1970s rate of chip-density growth required more than 18 times as many researchers.
In an Oct. 29 essay, Noah Smith argued that AI could expand the effective research workforce. He said it is too early to judge AI’s eventual effect on total factor productivity. The mouse-brain analysis measures processing time and offers questions for further study; it does not measure a change in scientific output.
