OpenAI Foundation Commits Over $125 Million to Medical AI Data, Including a Grant to Buy Failed Biotech Companies' Archives from Bankruptcy Auctions
OpenAI Foundation Commits Over $125 Million to Medical AI Data, Including a Grant to Buy Failed Biotech Companies' Archives from Bankruptcy Auctions
On September 15, OpenAI Foundation, the nonprofit parent organization of OpenAI, announced its second scientific program, Public Data for Health, and distributed over $125 million in its first tranche to research teams and nonprofits. Among the recipients are the CTD Commons archival project at 1Day Sooner, the OpenADMET project at the University of California, San Francisco (UCSF), and the University of North Carolina at Chapel Hill.
Back in March, the foundation had only announced open medical data as one of its future directions, without naming a single project or dollar amount. Now that announcement has materialized into concrete grants.
In mathematics, AI models already generate new verifiable knowledge without collecting new data. In biology, a model cannot look back and observe an event that has already occurred, or measure something that no one has ever measured directly. "Everyone agrees that data is the main bottleneck for successfully applying AI in biology," says Morgan Levine, former vice president of computation at the longevity company Altos Labs.
The foundation selects data by three criteria: the data must link different levels of biology within a single dataset, rescue information that is about to disappear permanently, or measure the phenomenon directly rather than through a cheap surrogate. The first criterion applies to the OpenADMET grant at UCSF: ninety percent of drug candidates fail in clinical trials, and a common reason is the difficulty of predicting how a compound distributes through the body and whether it crosses the blood-brain barrier, the filter that determines whether a drug reaches the brain. The team will link the molecular profiles of tens of thousands of compounds with the structures of transporter proteins and a model of this barrier. OpenADMET aims to repeat the story of AlphaFold2, which predicted protein structures thanks to the open Protein Data Bank and the CASP competition. This is already the project's second grant in three weeks: in August, the scientific infrastructure funder Astera Institute, through its Radial project, funded the development of code and rules for its blind trials, which will now be used to verify the accuracy of the new ADMET data.
The most unusual grant rescues data that is about to disappear permanently. The idea was proposed by genomics specialist Ruxandra Teslo: large pharmaceutical companies accumulate decades of toxicology, manufacturing, and FDA correspondence archives, while startups have no such archive because FDA dossiers are not published. "The process itself remains practically a black box, especially for small biotech companies, which are the ones creating the innovations," Teslo says. She found a loophole: a bankruptcy court has the authority to sell rights to such documents as an ordinary company asset. The idea is now being carried out by 1Day Sooner, a group advocating for clinical trial volunteers, under the leadership of Josh Morrison, as part of the CTD Commons project. "Every FDA submission represents the labor and sacrifices of researchers, reviewers, and trial participants. Even when these efforts end in failure, the valuable knowledge they generate should not go to waste," Morrison says. The organization already has three datasets, two of which were donated by Lumen Bioscience, but two other attempts to purchase bankrupt companies' archives this year fell through. The CTD Commons data will become an open resource for all researchers.
The third recipient, the University of North Carolina at Chapel Hill, will focus on direct measurement. Personalized cancer vaccines select targets based on tumor DNA sequencing, an indirect way of guessing which proteins will appear on the tumor's surface. The university will directly measure these proteins and T-cell responses for hundreds of tumors and make the data available to all researchers.
OpenAI Foundation owns 26% of OpenAI's equity, and at the company's discussed valuation of $1 trillion, it would hold a stake worth approximately $250 billion, which exceeds the total assets of the Gates Foundation. The foundation only began systematic grantmaking this year and expects to distribute $1 billion by year's end. It has directed part of that capital toward data: the raw material without which models cannot learn to discover drugs.