AI has changed the speed and scale at which scientific discovery happens, but progress still depends on the data available for training a model. When sufficient data exists to reflect the underlying reality of a problem accurately and at scale, AI can compress work that once took years into months or days. The clearest example is the Protein Data Bank, the open archive of experimentally determined protein structures that researchers have deposited and curated since 1971. AlphaFold was possible because that archive existed in standardized, openly accessible form long before anyone knew what would be done with it. The Sloan Digital Sky Survey is another example: one carefully designed data collection effort became one of the most widely used resources in the history of astronomy. In both cases the dataset came first and the discoveries followed.
It is likely that most of the low-hanging fruit of readily available scientific data has already been tapped, and further advances in particular disciplines will require intentional effort. Deciding where and how to invest in that data is a major obstacle. Foundational datasets can be expensive and slow to build, and the incentives for scientists to develop data resources for broad public benefit do not always exist. Funders willing to underwrite dataset creation have no reliable way to determine which dataset, for which scientific problem, would justify the cost, and no mechanism for reaching that determination jointly with other funders so that the resulting commitment is large enough to matter. Without such a mechanism, dataset creation will remain ad hoc: a few well-positioned proposals get built while many high-value opportunities are never pursued.
Identifying and Building the Datasets for AI-Driven Science will test a structured, problem-first process that is run with prospective funders at the table. SeedAI will convene a time-bounded working group of roughly 15 members, balanced between the domain scientists who would produce the data and the AI researchers who would use it, alongside funders and data infrastructure and governance experts. The group works in four steps:
Evidence. SeedAI will conduct semi-structured interviews with domain scientists, AI researchers, funders, and data infrastructure and governance experts; a landscape scan of existing datasets and consortia, prior dataset solicitations, agency roadmaps, and workshop reports; and a public call for resource questions.
Ranking. The group develops each shortlisted candidate into a one-page specification naming the dataset and the cellular or experimental context it covers, the testable metric, the scaling hypothesis, a pilot design that could be run in months, a decision threshold for advancing to a full build, and an estimate of pilot and full-scale cost. Members score every candidate independently against the published rubric on scientific importance, feasibility, and payoff relative to effort, deliberate where scores diverge most, and score a second time. The output is a ranked portfolio of three to five candidate datasets.
Review. Reviewers who did not take part in scoring, drawn from both prospective funders and expert scientists, judge in writing whether each candidate’s metric is measurable, its scaling hypothesis falsifiable, its pilot executable at the stated cost, and the ranking defensible. A portfolio that does not pass is revised once and resubmitted.
Solicitation. The group converts the top-ranked candidate, or the top two, into a solicitation for scaling pilots. It will define the scope and resource question, eligibility, award size and duration, evaluation criteria, and the review process. Eligibility will be written so that consortia and institutions beyond the largest research universities can compete. It will set requirements for data standards, metadata, provenance, and open release, along with an attribution protocol that gives the scientists who produce the data durable credit. The group will also document what a full dataset build would require.
A final hybrid roundtable with funders, agency staff, and coalition partners reviews both deliverables and identifies a host and pathway for the solicitation. The solicitation and a policy playbook will then be published openly, with the rubric, ranked portfolio, recorded dissents, and review summaries released alongside them. A landscape summary of data gaps in regulatory genomics may also be published. Interviews are anonymized, and findings are reported by role and never by name.
Why regulatory genomics?
Regulatory genomics is the study of how non-coding DNA, such as enhancers and promoters, control gene expression. It is a strong first test case for three reasons. First, it is an area where AI has most clearly shown that data availability determines progress. For instance, sequence-to-function models trained on resources such as ENCODE and GTEx can now predict many regulatory signals directly from DNA sequence. Second, its assays are mature and scale predictably, so scaling pilots can be run in months at costs in the hundreds of thousands of dollars. Third, its payoff is direct. Most disease-associated genetic variants fall within non-coding regions. Models that explain how these variants act would help identify human-relevant drug targets and determine which ones advance to trials, the stage where most of the cost and failure in drug development occurs.
Where the ideas come from
The process is inspired by the framework Peter Koo of Cold Spring Harbor Laboratory described in “Before the Next PDB: Can a Field Converge Before It Builds?” (2026). Koo argues that fields fail to build foundational datasets because no institutional mechanism forces them to narrow their options before large-scale capital is committed. He proposes a three-stage process. In the first, a field defines a specific “resource question,” the single data type that would most advance it. In the second, a constrained convening, whose only permitted output is a ranked portfolio, produces three to five candidate datasets, each with a testable metric, a scaling hypothesis, a falsifiable pilot design, and a numerical decision threshold. In the third, scaling pilots costing hundreds of thousands to a few million dollars measure whether the metric actually improves as data accumulates, before anyone commits the tens or hundreds of millions a full build requires. This project carries out the first two stages and writes the solicitation for the third.
Several organizations have begun to treat the scientific data gap as a first-order problem:
Renaissance Philanthropy, working with the UK Department for Science, Innovation and Technology, ran an open competition for proposals to build new AI-for-science datasets. It drew 79 proposals across five priority domains and named ten winners.
Google.org announced a $20 million AI for Science fund in January 2026, with an emphasis on shared datasets and open science.
Focused research organizations, time-limited nonprofits funded to build a specific scientific resource, have begun to target datasets explicitly. Bind Research, the first FRO formed in the UK, is generating experimental data on how small molecules bind disordered proteins.
Identifying and Building the Datasets for AI-Driven Science will add to these efforts by producing a broadly applicable process for deciding, across candidates and across fields, where the next dollar of dataset investment should go, or a mechanism through which several funders can reach that decision together. It includes prospective funders as cornerstones of the effort, so the solicitation is pre-qualified as actionable. And it captures the process layer, so others can repeat the approach rather than start over.
Who’s involved
Identifying and Building the Datasets for AI-Driven Science is led by SeedAI with support from the Alfred P. Sloan Foundation, Astera Institute, and Renaissance Philanthropy. The working group will be drawn from the Accelerate Science Now coalition, which SeedAI convenes, and from the wider field, including core facilities, regional universities, and industry laboratories, because data-generation expertise is distributed across those settings.
Get involved
Submit a resource question by December 9, 2026. In no more than 500 words, describe the dataset that would most advance predictive models in regulatory genomics, how you would measure improvement, and why more data should help. For anything else, contact Trisha Tucholski at trisha@seedai.org.