The OpenAI Foundation is committing more than $125 million to create and preserve biological and medical datasets, betting that better real-world data (rather than larger models alone) will be essential for the next wave of AI-driven drug and disease research.
The new program, called Public Data for Health, will fund universities and nonprofits producing datasets that individual laboratories or companies may lack the money or incentive to create. The resulting resources are intended to be broadly accessible to researchers rather than reserved exclusively for OpenAI’s own models. Human datasets must also preserve privacy and consent.
One project, OpenADMET, will generate experimental data to improve predictions of how drug candidates are absorbed, distributed and transported through the body. Researchers will profile tens of thousands of compounds, connect those measurements with transporter-protein structures and functional assays, and test selected molecules against models of the human blood-brain barrier.
Another initiative, CTD Commons, tackles data that already exists but frequently disappears. It plans to recover regulatory dossiers from abandoned or failed biotech drug programs, including animal toxicology, manufacturing information and exchanges with the US Food and Drug Administration. Much of this material never appears in scientific publications, even though failures can contain valuable information for future development.
At the University of North Carolina, a generative immunotherapy project will instead create new multimodal data for personalized cancer vaccines by directly measuring proteins on tumor cells and patients’ immune responses.
For biomedical AI, creating better training and evaluation data may therefore become as important as improving the models themselves.