This week, PerturbAI, OpenAI, and their collaborators shared results from a year-long effort aimed at developing an artificial intelligence (AI) reasoning framework that helps scientists interpret large experimental datasets, connect results with existing research, and generate testable hypotheses. Details are provided in a new bioAxriv preprint titled “Distributed AI agents enable comprehensive discovery of biological insights at scale.”
Besides the scientists at PerturbAI and OpenAI, other authors listed on the paper are scientists affiliated with Scripps Research Institute, University of California, San Diego, and Howard Hughes Medical Institute. In the paper, they explained the results of applying the AI framework to data from PerturbAI’s whole-brain atlas, published earlier this year. Specifically, the team developed thousands of agents to analyze cellular responses, search the scientific literature, and connect experimental results with existing biological knowledge. The system then synthesized the findings into a reasoning atlas that linked genes, cell responses, details from existing literature, and possible biological mechanisms.
When applied to the whole-brain atlas, “thousands of parallel agents thoroughly examined the dataset in minutes” and without access to the source manuscript “recovered its conclusions, predicted held-out perturbation effects, and identified perturbation relationships with shared or opposing outcomes across effect strengths,” the scientists wrote. In fact, the system completed the analysis in 32 minutes. Furthermore, “distributed agents scaled literature coverage with workload, retrieving 100s times more references than [a] single-agent mode.”
To evaluate the quality of the agents’ interpretations, biologists reviewed their reports for novelty, evidential support, follow-up priority, and confidence level. According to numbers reported in the preprint, in blinded comparisons against reports generated by single AI agents, human experts preferred the reasoning-agents’ interpretations 75% of the time or 45 of 60 judgements. Single agent interpretations were preferred in 10 or 16.7%, and five were ties (8.3%).
The work builds on a year-long collaboration that is now expanding into joint drug-target discovery that will leverage PerturbAI’s proprietary causal biological data. Using PerturbAI’s causal in vivo data, PerturbAI and OpenAI will now apply their AI reasoning frameworks to discover new drug targets. “Biomedical experiments are generating data faster than scientists can interpret it, meaning promising new discoveries are likely to remain hidden,” said Grace Zheng, PhD, CEO and co-founder of PerturbAI. “By combining large-scale causal data generated directly in living systems with frontier AI, our partnership with OpenAI is making it possible to investigate more biological data, faster and with greater depth, and to make target discovery more systematic.”
“By combining PerturbAI’s experimental data with OpenAI’s reasoning capabilities, we can help scientists uncover new biological insights and identify promising leads for drug discovery,” added Joy Jiao, head of life sciences at OpenAI. “The scale, quality, and speed of PerturbAI’s discovery engine make them an ideal partner, and this is just the beginning of what we can achieve together. Running nearly 2,000 agents in parallel shows how we can scale biological interpretation alongside high-throughput experimentation, helping researchers turn those findings into better medicines.”