PerturbAI, OpenAI mine brain CRISPR data for drug discovery leads

PerturbAI and its partner OpenAI revealed on Wednesday a new approach that uses thousands of AI agents working at the same time to study complex biological data and find possible targets for new drugs.

"Finding new drug targets still depends largely on chance," said PerturbAI CEO and co-founder Grace Zheng. She explained that companies often decide where to invest before scientists have enough evidence to understand how possible drug targets work.

Working with Scripps Research, PerturbAI and OpenAI used nearly 2000 AI agents powered by GPT-Rosalind, OpenAI's reasoning model for biological and chemical research, to study a huge dataset of mouse brain cells. The agents finished their analysis in just 32 minutes. Researchers estimated that people doing the same work could take months to a year.

The analysis used PerturbAI's atlas of CRISPR experiments across the whole brain, which was released when the startup came out of stealth in March. The dataset shows how disrupting genes affects different types of cells across the mouse brain.

Each agent examined which genes became more or less active in different types of brain cells. It then searched published studies for evidence that could explain these changes. The agents wrote reports describing possible biological processes behind the changes and cited research to support their ideas. The reports were combined into a searchable database so scientists could compare what happens when different genes are disrupted.

OpenAI said an important test of the agents' usefulness is whether their findings can suggest new directions for research. One promising lead involved DCNT1, a gene linked to Perry syndrome, a rare brain disease that causes symptoms similar to those of Parkinson's disease.

In the whole-brain CRISPR atlas, disrupting DCNT1 increased the activity of genes involved in making cholesterol. The agents linked this finding to research suggesting that when cells cannot move cholesterol properly, they may respond by making more of it. This led to questions for future experiments: Does disrupting DCNT1 change how cholesterol is spread inside nerve cells? And could reversing these changes help protect the cells?

Researchers compared 20 pairs of reports, with three experts assessing each pair. Across the 60 judgments they made, the experts preferred the explanations produced by groups of AI agents over those produced by a single agent in 45 cases.

The findings are described in a preprint, a research paper that has not yet been reviewed by other experts. The results still need to be tested through experiments.

Since then, PerturbAI has also expanded its genetic atlases to cover metabolic diseases and ageing. Wednesday's announcement came one day after OpenAI acquired Orchestra Bio. Orchestra Bio brings AI agents designed to help drugmakers track research studies and assess how setbacks might affect drug development schedules.