OpenAI’s restructuring last year created not only a new corporate structure for the company behind ChatGPT but also a potentially enormous new philanthropic force, the OpenAI Foundation. Now the Foundation is putting $125 million behind a bet on what increasingly powerful artificial intelligence will need to accelerate medical discovery: better data.

The new initiative, Public Data for Health, will fund nonprofits and universities building, preserving, and opening high-quality scientific datasets for researchers. Initial grantees include OpenADMET, CTD Commons, and the University of North Carolina, with projects spanning drug discovery, regulatory knowledge, and personalized cancer vaccines.

“The importance of data is only going to go up as AI progresses and as intelligence becomes more abundant and more used within the life sciences and within academia,” Jacob Trefethen, head of Life Sciences & Curing Diseases at the OpenAI Foundation, told Inside Precision Medicine.

The premise is straightforward: increasingly capable AI models may be able to find patterns across enormous quantities of biological information, but what they can discover ultimately depends on the observations available to them. “One of the great complements to AI progress is actually more observations of the world,” Trefethen said. “Data is, in that sense, the foundational input to discovery.”

Open data, with limits

Accessibility is a central principle of the program. “The shared thread, I think, is accessibility,” Trefethen said. “Making sure that the data that we fund is as accessible as possible to as many researchers as possible is going to be a real big focus.”

But accessibility does not necessarily mean putting every dataset freely on the internet. Trefethen said the appropriate model will depend on the information involved. Some scientific datasets, such as measurements of protein dynamics, could potentially be released openly as soon as they are generated. Human data presents a different challenge.

The Foundation says resulting datasets will be made as broadly accessible as possible while protecting privacy and individual consent. Trefethen said the Foundation generally follows the expertise and best practices of the universities and research centers handling patient information while retaining oversight through its role as funder.

The grantees “follow best practices around any data that relates to patients,” Trefethen said, with the goal of making information “as accessible as you can make it to researchers at other institutions so that the scientific community can make the most progress on behalf of patients … Given we’re making a grant, we play some governance role in ensuring that good practices are followed,” he added.

Cancer vaccine data centers

One of the first projects will support the University of North Carolina in establishing an initiative focused on generative immunotherapy, including data that could improve personalized cancer vaccines. Such vaccines can use information from an individual patient’s tumor to identify neoantigens—abnormal proteins that could be recognized by the immune system—and design a vaccine intended to direct T cells against the cancer.

Machine learning is already important to that process. “A lot of what’s enabled progress in that field so far is machine learning,” Trefethen said. “It is predicting from the sequence of a tumor what is going to be presented on the outside of the tumor, and how will the patient’s T cells respond.”

But sequence data alone does not necessarily reveal what is actually happening biologically. The UNC project aims to connect several layers of information: sequencing tumors, measuring what proteins are actually displayed by tumor cells, and determining how patients’ immune systems respond. “By creating that linked multimodal data, we think that we’ll be able to move forward the whole field,” Trefethen said.

The project illustrates one of three broad categories the Foundation is prioritizing: connected data, which combines different layers of biology to provide a more complete picture. The other two are scarce data, particularly valuable information that is difficult to obtain or at risk of disappearing, and direct data, measurements of biological and clinical states that are as close as possible to phenomena that matter for patients.

Drug development’s data problem

Another grant will support OpenADMET, an effort involving researchers at the University of California, San Francisco, aimed at improving predictions of absorption, distribution, metabolism, excretion, and toxicity—ADMET properties that can help determine whether an experimental medicine ultimately succeeds or fails. “They’re going to collect really high-quality data related to ADMET properties, and they’re going to run these blind challenges, and then anyone can enter,” Trefethen said.

The approach has precedents in computational biology. Community challenges such as the Critical Assessment of Structure Prediction, or CASP, provided common tests against which competing computational approaches could be evaluated, helping drive advances in protein-structure prediction.

Trefethen said an earlier competition by the group behind OpenADMET attracted more than 300 entrants. “Picking the problems that could unlock progress for patients across the board and then targeting collecting high-quality data is a way that we can help the field as a whole,” he said.

CTD Commons, another initial grantee, will focus on preserving and organizing regulatory and drug-development knowledge, extending the initiative beyond newly generated experimental measurements to information that could otherwise remain fragmented or be lost.

What data is worth $125M?

The Foundation is deliberately keeping its definition of valuable biomedical data broad. “We’re trying to balance a desire to stay broad and learn from many different research communities [about] what’s most needed to drive progress in their corner of biology and the life sciences,” Trefethen said.

One consideration is scarcity. If an important biological or epidemiological phenomenon is poorly measured, creating a rigorous dataset around it could produce a larger scientific return than generating additional observations in an already data-rich field. “If you can create a high-quality dataset related to it that everyone can learn from, that’s going to be more attractive to us,” Trefethen said.

The Foundation is also interested in something even further upstream: technologies capable of generating biological observations that scientists currently cannot readily make. “What aspects of molecular biology are not well imaged right now?” Trefethen asked. “What aspects of the spatial aspects of biology are not well measured?”

That could include new instruments, imaging approaches, or experimental methods that make previously inaccessible biological phenomena measurable. “Fundamentally, progress requires observation,” Trefethen said. “We do get stuck at the equipment that we’re using.”

A new kind of science funder

Public Data for Health is part of a much larger expansion of the OpenAI Foundation’s philanthropy. Trefethen said the organization intends to award more than $1 billion in grants across its programs during its first year and expects that amount to increase.

“We’re trying to scale up our grantmaking as quickly as is compatible with high-quality, rigorous decision-making,” he said.

Grants will generally run for about two years, sometimes three, with formal assessments quarterly or every six months depending on their size. Trefethen said the Foundation also intends to maintain informal contact with grantees throughout their projects.

The Foundation is still early in determining exactly how broad its life sciences portfolio will become. Trefethen said it wants to learn from researchers rather than prematurely deciding which areas of biology are most deserving of support.

The scale of that ambition also makes the unusual organization behind the money relevant. OpenAI began in December 2015 as a nonprofit research laboratory. After a series of restructurings, the nonprofit—now the OpenAI Foundation—controls OpenAI Group PBC, appoints the company’s directors, and holds a substantial equity stake in the commercial enterprise.

The arrangement inverts the familiar relationship between a corporate foundation and its benefactor: the nonprofit is not controlled by the company; it controls the company. Its financial resources are also connected to OpenAI Group’s success through its equity ownership.

Trefethen said the Foundation conducts its philanthropic activities and makes grantmaking decisions independently of OpenAI’s commercial operations. Researchers receiving its grants are not required to use OpenAI technology. “This is an independent philanthropic endeavor, and our grantees can use any AI models they want—not just OpenAI models,” Trefethen said.

That distinction matters for Public Data for Health. The Foundation says the program is not designed to generate proprietary datasets for OpenAI models. Instead, outside research institutions will create scientific resources intended to be broadly accessible, including to researchers using competing AI systems.

But as the Foundation’s grantmaking grows, so will the significance of those choices. At more than $1 billion in planned first-year grants across its programs, the organization is entering philanthropy with resources capable of influencing which scientific datasets get built, which experimental technologies advance, and potentially which problems become tractable for AI.

For a funder whose wealth and governance remain intertwined with one of the companies competing to lead the AI industry, the consequences will extend beyond how much money it gives away. They will also depend on what scientific infrastructure that money creates—and who ultimately gets to use it.