AI in precision medicine is skipping a step: Checking its data
By Patrick Short, PhD
Most discussion of AI in precision medicine centers on what models can do, such as nominating drug targets or interpreting variants. Much less attention goes to the condition every one of those outputs depends on. The underlying genomic and clinical data has to be consented, standardized, and governed well enough to support the conclusion a model draws from it.
For trial sponsors, that condition sits upstream of every decision a model informs. A target chosen from skewed association data carries that flaw into protocol design and analysis. A cohort selected with a score that underperforms in some ancestries carries it into recruitment.
Three recent guests on The Genetics Podcast approached this prerequisite from different directions. Dr. James Zou of Stanford described what AI can do today. Dr. Luke Pilling of the University of Exeter, who works with UK Biobank, walked through what the resource still lacks, from primary care records to clinical codes that can be trusted. Tania Simoncelli, who worked with the ACLU on the gene patent challenge, discussed what happens when data governance gets it wrong.
Taken together, their framings point to one argument. The trustworthiness of a model's output is largely set when the data is collected, years before anyone trains on it.
What AI in precision medicine can do depends on data it was never designed to fix
In Episode 261, James Zou, Associate Professor of Biomedical Data Science at Stanford, described a virtual biotech built from AI scientist agents. The conversation covered how the single-cell evidence behind a drug target relates to its chances of success in trials. Zou also said the agents learn from published evidence, which skews toward successful trials, and that failed and proprietary trial data is the next gap for the virtual biotech.
That gap is instructive. An agent can reason only over the evidence it can reach, and proprietary trial data is held by sponsors under study-specific consent terms and formats. Zou's published work examines the same weakness from the evaluation side. A 2021 Nature Medicine comment reviewing 130 FDA-approved medical AI devices, co-authored by Zou, described evaluation limits that can mask vulnerabilities once devices reach patients.
Regulators have reached a similar position on inputs. The FDA's January 2025 draft guidance on AI for drug and biological product decisions warns that variable dataset quality, size, and representativeness may introduce bias. It recommends that data be fit for use, meaning relevant and reliable, with reliability defined as accurate, complete, and traceable. It also asks sponsors to document how data were collected, processed, annotated, and stored.
The guidance is a draft with nonbinding recommendations. It applies to AI that supports evidence on safety, effectiveness, or quality, excludes drug discovery uses, and places consent outside its scope.
The central challenges of AI in precision medicine therefore sit in the data layer. They include unrepresentative cohorts, inconsistent clinical coding, consent that does not cover new uses, and weak control over who can access what. Representativeness is the easiest of these to measure.
Representative, standardized data is still the exception
UK Biobank is the reference case for scale. Its consortium has reported whole-genome sequencing of 490,640 participants in a cohort of about 500,000. UK Biobank's summary of that work notes that most disease associations appeared in people of European ancestries. Strong or novel signals were also found in African and Asian ancestries.
On Episode 262 of The Genetics Podcast, Luke Pilling, Senior Lecturer in Genetic Epidemiology at the University of Exeter, described what the resource still lacks for many research questions. Our recap of that conversation lists primary care records, iron biomarkers, proteomics, cleaner clinical codes, and more diverse populations among the open gaps. Pilling also said clinical codes need to be checked with clinicians before anyone trusts them, and that Our Future Health participants consented for recontact, which podcast host Patrick Short, co-founder and CEO of Sano, pointed to as an advantage of building in recontactability from day one.
The ancestry gap extends across genetic research. A review reports that the GWAS Diversity Monitor put European-ancestry participants at 94.48% as of September 2024. That cumulative count can include repeated use of the same cohorts. The same review reports that the GWAS Catalog, counted differently, reached 86.5% European-ancestry participants in 2023.
Data quality matters for AI in precision medicine because a model inherits the composition and coding of the cohort it learns from. Experts have argued that clinical use of current polygenic risk scores risks widening health disparities unless genetic studies diversify beyond European-ancestry populations. Sano has covered how polygenic risk scores are built and applied and what ancestry diversity requires of study design.
Standards for describing data use are developing in parallel. The Global Alliance for Genomics and Health (GA4GH) publishes the Data Use Ontology (DUO), a machine-readable vocabulary designed to represent a dataset's allowable uses consistently. A 2026 European review describes DUO translating consent clauses into structured terms that can be matched against a researcher's stated purpose.
These sources describe design intent and do not report adoption levels. Scale and standards answer part of the trust question. The terms of access answer the rest.
Shared data only helps a model when its consent and access terms are clear
Sharing data well depends on the conditions attached to access. UK Biobank's own model shows how those conditions are set. Its access policy makes data available to eligible academic, charity, government, and commercial researchers for health-related research in the public interest.
The terms have changed over time. UK Biobank states that it considered applications from insurance companies until January 2025 and no longer approves them for direct data access. Since 2024, access has run through its research platform by default, where earlier researchers could download some data types.
Consent scope is the harder issue. In a peer-reviewed critique, Barn argues that participants give broad consent to an area of research. UK Biobank relies on legitimate interests and public interest as its lawful bases under UK GDPR. Barn suggests participants may not have understood the full range of lawful uses.
Governance has also been tested by events. A Nature World View reported that in April, de-identified data for UK Biobank's 500,000 participants appeared in listings on an Alibaba-owned e-commerce platform. UK Biobank believes no data were sold, and their oversight committee report states that its research analysis platform was suspended on April 23, 2026.
For a sponsor using a model trained on a shared resource, these events reach the output directly. If a dataset's consent basis is contested or its access rules change mid-program, the evidence behind a model-derived target gets harder to defend. Those consequences tend to surface late.
Governance failures show up years after the data is collected
Tania Simoncelli, VP of Translational Impact and Engagement at Biohub and formerly at the Chan Zuckerberg Initiative, has seen this pattern across a long career in genetics policy. On Episode 266, Simoncelli said trials fail because “we're not getting quite the right patients at the right time.” Her TED profile notes her ACLU work on the gene patent challenge and her role as Special Assistant to FDA Commissioner Margaret Hamburg. The same profile notes that Nature named her one of ten people who mattered in 2013 for that work.
That challenge, AMP v. Myriad, shows how control over genetic information can shape access for years. The Supreme Court's 2013 opinion records that Myriad's patent enforcement had left it as the only entity providing BRCA testing. The Court held that a naturally occurring DNA segment is not patent eligible merely because it has been isolated, while cDNA is patent eligible.
Consumer genetics produced a more recent example. In 2025, 23andMe's assets, including customer genetic data, were sold through bankruptcy to TTAM Research Institute, a nonprofit, for $305 million. TTAM committed to honor 23andMe's existing privacy policies, including customers' ability to delete their data and opt out of research. The sale put a question in front of every genomic program: whether data collected under one set of consent terms can move to a new owner under those same terms.
Both cases show that terms set at collection, covering ownership, permitted uses, and transfer, determine what can legitimately be done with data much later. Model training is one of those later uses, and sponsors can design for it at the start.
Consent-first data collection is a design decision for trial sponsors
Genomic data intended for AI and secondary research should be consented with a defined scope participants can understand. That consent should be recorded in a form systems can check and paired with permission to recontact participants as new uses arise. For a sponsor, this translates into decisions made before the first participant enrolls:
- Consent language that names secondary research and model development explicitly
- Recontact permission, so participants can be asked about new uses directly
- Standardized phenotype and clinical coding at capture, so records can be pooled across sites
- Ancestry composition tracked during enrollment, with recruitment adjusted when the cohort skews
- Documented provenance for collection, processing, annotation, and storage, in line with the FDA draft
In programs supported by Sano, genetic data collection is consent-first by design, and the platform is ISO 27001 certified and HIPAA and GDPR compliant. Participants are recontactable for future studies, so a sponsor can return to participants who opted in to recontact for a later program. Sano's prescreening and digital consent workflows capture that consent at the point of entry. Sano also runs the recruitment, genetic testing, genetic counseling, and long-term participant engagement around that consent, so the same participants stay connected from enrollment through follow-up. Participant data is used only for purposes participants agree to and is not shared without explicit permission, and every study runs under appropriate ethics board review.
These choices connect across the trial lifecycle. Consent scope set during prescreening determines which analyses are possible at readout. Recontact permission determines whether the same cohort can support the next indication. Coding decisions at capture determine whether the data can be pooled with other sources.
What sponsors should ask before trusting a model's output
Zou’s point about proprietary trial data has a practical consequence. As agentic systems begin to draw on sponsor trial archives, today's consent and coding decisions will set the limits of what those systems can use. Data collected under narrow, unrecorded, or non-recontactable consent may be scientifically rich and still unavailable for model development.
Sponsors evaluating an AI-derived target or stratification can put a short set of questions to the data behind it. What did participants consent to, and can they be recontacted? How were clinical codes standardized, what is the ancestry mix, and who has had access under what terms?
Answers that cannot be documented are a program risk, whatever the model's benchmark performance. The same questions apply to the data a sponsor is collecting now, which will become someone's training set later.
To discuss consent-first genetic data collection for an upcoming program, get in touch.