PerturbAI and OpenAI launch AI reasoning agents for biomedical discovery

Scaling biological interpretation

PerturbAI and OpenAI have revealed a new artificial intelligence reasoning framework designed to help scientists navigate massive biological datasets. The project, which involved a year-long collaboration, aims to help researchers connect experimental results with existing scientific literature. By automating the synthesis of these findings, the system can generate new, testable hypotheses for biological research.

The team developed a system that utilizes thousands of individual AI agents to perform specific tasks simultaneously. These agents are tasked with analyzing cellular responses and searching for relevant information in the scientific literature. This distributed approach allows the framework to scale its coverage of research papers much more effectively than a single AI agent could manage alone.

The goal of this framework is to address the growing bottleneck in modern science where data is being produced faster than it can be interpreted. Grace Zheng, CEO and co-founder of PerturbAI, noted that experiments are generating data faster than scientists can interpret it. The partnership seeks to make the discovery of new biological insights more systematic and efficient.

Testing the reasoning atlas

To test the capabilities of this framework, the researchers applied it to a whole-brain atlas published by PerturbAI earlier this year. The system was asked to analyze the dataset without having access to the original manuscript. It successfully recovered the study's conclusions and predicted effects on perturbations that were held out from the initial analysis.

The speed of the system was a notable finding in the study. The scientists reported that thousands of parallel agents were able to thoroughly examine the entire dataset in just 32 minutes. This level of processing allows researchers to identify relationships with shared or opposing outcomes across different effect strengths in a very short timeframe.

The framework also synthesized these findings into what the researchers call a reasoning atlas. This output links specific genes and cell responses to details found in existing literature. By connecting these dots, the AI provides a map of possible biological mechanisms that helps scientists understand the underlying causes of biological phenomena.

Human expert evaluations

The researchers conducted a blinded comparison to determine how well the AI's interpretations held up against human standards. Biologists reviewed the reports generated by the reasoning-agents based on several criteria, including novelty, confidence levels, and the priority for follow-up research. They compared these results against reports produced by single AI agents to see which provided better insights.

The results showed a clear preference for the multi-agent reasoning framework. Human experts preferred the interpretations of the reasoning-agents 75% of the time, which accounted for 45 out of 60 total judgments. In contrast, the single-agent interpretations were preferred only 16.7% of the time, while the remaining 8.3% of the judgments resulted in a tie.

These findings suggest that distributing the workload across thousands of agents improves the quality of the output. Joy Jiao, head of life sciences at OpenAI, stated that running nearly 2,000 agents in parallel shows how the team can scale biological interpretation alongside high-throughput experimentation. This approach helps researchers turn complex findings into actionable leads for medicine.

Expanding into drug discovery

The collaboration between PerturbAI and OpenAI is now moving into a new phase focused on drug-target discovery. The companies plan to leverage PerturbAI’s proprietary causal biological data to identify new targets for pharmaceutical development. Causal data is particularly valuable because it helps establish a direct relationship between a specific intervention and a biological outcome.

By combining this experimental data with OpenAI's reasoning capabilities, the teams hope to uncover new biological insights that were previously inaccessible. The framework will be used to investigate more biological data with greater depth and speed. This systematic approach aims to streamline the early stages of drug development, where identifying the right target is a primary challenge.

The researchers believe that this integration of frontier AI and large-scale causal data is a significant step forward for the field. The scale and speed of the discovery engine are intended to help scientists move from raw data to better medicines. This partnership represents an ongoing effort to use distributed AI to solve some of the most complex problems in biotechnology.