OpenAI Foundation launches Data for Public Health to fund high-quality scientific datasets that advance AI in medicine, including a $40 million grant for cancer vaccine data at UNC and support for OpenAdmet’s drug-effect prediction challenges.

Researchers aim to accelerate personalized cancer vaccine design, focusing on improving target selection and assessing formulations’ ability to elicit tumor-specific immune responses in triple-negative breast cancer.

Led by immunologist Benjamin Vincent and computational biologist Alex Rubinsteyn, the project will gather and analyze de-identified tumor and immune-cell data from three biobanks to train AI models for better tumor antigen discovery and vaccine optimization.

CTD Commons will preserve and organize regulatory and drug-development knowledge to prevent valuable information from being lost.

Leaders acknowledge regulatory and safety concerns surrounding AI, including calls to slow progress to ensure safety measures keep pace.

OpenADMET builds on existing datasets and precedents to address universal drug-discovery challenges and reduce failure rates in development.

The program prioritizes data that is connected, scarce, or direct, guiding datasets that are multimodal, hard-to-measure, or closest to patient outcomes.

Three data strategies emphasize connected data, scarce data, and direct data to maximize relevance for drug discovery and clinical outcomes.

Key partners include Scale AI for data structuring, Palantir Foundry for secure data management, and Snowflake for cross-border data sharing.

Accessibility and privacy are central: datasets should be as accessible as possible while protecting privacy and consent under Foundation oversight.

Regulatory dialogue, scientific measurements, and drug knowledge in common documents could dramatically enhance AI’s ability to assist in drug development and regulatory navigation.

Beyond healthcare, AI-data acquisition trends include major moves to acquire corporate data from failed companies, signaling a broader “land grab” for training data.