
Something important is happening in biomedical data infrastructure. In September 2026, the US National Institutes of Health issued a Request for Information concerning technologies for its Data COUNTS programme — Collect Data Once, Use Numerous Times — intended to make health data reusable for biomedical research, public health and regulatory decision-making. NIH has attached an important qualification to that ambition: the data must not merely be accessible, it must be provenanced. NIH describes an architecture intended to extract and de-identify electronic health-record information while preserving full data provenance and lineage, with the stated aim of supporting rapid and reproducible research across institutional boundaries. That language should not be underestimated.
At almost exactly the same time, a related discussion has been taking place in Europe. Speaking to the European Parliament’s Committee on Public Health, imec programme director and fellow Liesbet Lagae argued that Europe requires infrastructure capable of producing and connecting the biological data necessary for AI-enabled medicine. Clinical records alone are insufficient; models intended to understand disease or predict biological responses also require experimental information about DNA, proteins, cells, tissues and their responses to interventions. Both observations expose a problem considerably larger than data storage: if biological evidence is going to be collected in one place, processed somewhere else, combined with information from other institutions and eventually used to train artificial-intelligence systems, how do we preserve the history of that evidence? That question should matter particularly to Britain, because the UK is already building much of the infrastructure in which this problem will arise.
The UK’s Life Sciences Sector Plan contains an ambition that would have seemed extraordinary only a decade ago. Government intends to create a Health Data Research Service backed by up to £600 million of investment — up to £500 million from government and £100 million from Wellcome, based at the Wellcome Genome Campus — with the aim of giving researchers a more unified route into population-scale health information and ultimately bringing together clinical, pathology, radiology and genomic data in forms suitable for artificial-intelligence research. By 2030, the ambition extends considerably further: Our Future Health is intended to encompass up to five million consented participants, linking primary and secondary healthcare information with genomic data and biological samples. UK Biobank is expanding its proteomic measurements to cover roughly 500,000 participants alongside other multiomic information, and Genomics England’s agreement with Illumina covers up to 300,000 whole genome equivalents over the next five years, with an option to extend that to 500,000, while the NHS Genomic Medicine Service’s National Genomics Interoperability programme is developing a Unified Genomic Record intended to connect genomic information with relevant clinical and diagnostic data. The UK’s AI for Science Strategy similarly identifies linked pathology, radiology and genomic datasets as a foundation for model development, biomarker identification, disease modelling and improved preclinical prediction.
Taken together, these are not simply databases. They are the beginnings of an enormous national biomedical information system, and that distinction matters. A database answers what data we have. An evidence system must also answer where it came from, what happened to it, and why somebody should trust it.
This is what makes the American Data COUNTS initiative particularly interesting. NIH does not describe provenance as an incidental metadata feature; the programme explicitly aims to preserve full data provenance and lineage. Its proposed real-world data ecosystem encompasses healthcare organisations, pharmacies, claims and billing information, clinical data, environmental observations, sensors, wearables, social determinants of health and mortality information, and the resulting information may subsequently contribute to precision medicine, drug development, longitudinal research, regulatory decisions and public-health policy. That creates a fundamental evidential problem.
Imagine a relatively simple clinical observation: a blood result begins inside a hospital laboratory information system. It may subsequently be extracted from the originating system, mapped to a common terminology, de-identified, transformed into another schema, subjected to quality-control rules, combined with information from other healthcare providers, linked to genomic information, included within a research cohort, processed through an analytical pipeline, and ultimately used to train or validate an artificial-intelligence model. The value may remain recognisably the same. Its evidential context does not. Every transformation introduces questions — who performed it, which software version was used, which mapping rules applied, which reference database existed at that moment, which quality-control thresholds were applied, whether the original record was subsequently corrected, whether the transformation was deterministic, whether a human decision was involved, which organisation was responsible, which consent or governance framework authorised that particular use, and whether the resulting claim can be reconstructed five years later. Those questions are not peripheral to trustworthy AI. They are prerequisites for it.
The recent imec proposal approaches the challenge from another direction. Europe possesses clinical information through hospitals, health systems and emerging mechanisms such as the European Health Data Space, but training sophisticated biological models requires something different: experimental biology. A model intended to predict cellular behaviour may require genomic measurements, protein interactions, tissue responses, toxicity experiments and observations generated using sophisticated laboratory instrumentation. Imec’s argument is therefore that Europe should consider infrastructure combining experimental capabilities, measurement technologies and artificial intelligence — an attractive proposition that creates an important secondary requirement. If experiments continuously produce information that feeds biological models, and those models subsequently influence which experiments should be performed next, provenance becomes part of the scientific architecture itself. The sequence is no longer simply experiment leading to result; it becomes sample, instrument, measurement, processing, dataset, model, prediction, experiment, new measurement, updated model — potentially thousands or millions of times. Scientific evidence becomes recursive, and without persistent provenance, reconstructing that process becomes progressively more difficult.
The United States is asking how reusable healthcare information can retain provenance. Europe is asking how biological information can be generated at sufficient scale and quality for AI-driven medicine. Britain is attempting both. The UK’s emerging ecosystem potentially connects NHS clinical records to pathology, imaging, genomic sequencing, biobanks, research cohorts, multiomics, clinical trials, artificial intelligence, regulatory evidence and clinical care. This is an extraordinary scientific opportunity. It is also an extraordinary provenance problem.
The difficulty is not necessarily that individual systems lack audit trails — many have excellent ones. A sequencing laboratory may maintain detailed records; a hospital may retain an excellent clinical audit history; a biobank may have rigorous sample-management procedures; a research platform may preserve computational workflows; an AI environment may record model versions and training runs. The problem appears between them. An audit trail inside System A tells us what happened inside System A; an audit trail inside System B tells us what happened inside System B. Neither necessarily provides a persistent evidence chain describing what happened when information moved from A to B, was transformed in C, incorporated into D and eventually contributed to a conclusion produced by E. This is the interoperability problem viewed from the perspective of evidence rather than data.
The distinction is important. Consider a genomic variant: technically, it may be straightforward to transmit — chromosome 17, position X, reference allele Y, alternate allele Z. But that observation has a history. It originated from a biological sample, collected under particular circumstances. DNA was extracted; a library was prepared; an instrument generated raw reads; software processed those reads; reads were aligned against a particular reference assembly; a specific variant-calling pipeline was used; quality thresholds were applied; annotations were drawn from databases existing at a particular moment; a laboratory or researcher interpreted the result, and perhaps that interpretation subsequently changed. The variant is data. Everything surrounding it is evidence. AI systems trained on biomedical information increasingly consume the former while depending implicitly on the latter, and that dependence needs to become explicit.
There is a temptation to assume that sufficiently large datasets overcome imperfections within individual records. Scale certainly provides statistical advantages, but it does not eliminate provenance requirements — indeed, artificial intelligence makes them more important. Suppose a biomedical model produces an unexpected association, and researchers discover that it can be traced disproportionately to observations originating from several laboratories. They need to know whether this represents biology or infrastructure. Perhaps those laboratories used a particular instrument; perhaps they used an earlier reference genome; perhaps a reagent lot introduced an artefact; perhaps an extraction method affected the measurement; perhaps a coding convention inside one health system systematically differed from another; perhaps historical diagnostic practices produced a population bias. The model sees correlations. Scientists must determine whether those correlations represent nature or process, and without lineage, separating the two can become extraordinarily difficult.
There is another reason Britain should care. High-quality biological data is becoming strategic infrastructure, and the UK already possesses extraordinary assets because of the NHS, Genomics England, UK Biobank, Our Future Health and decades of publicly supported biomedical research. AI increases their potential value, but value does not arise simply from possessing information — it arises from being able to demonstrate its quality. A dataset with demonstrable origin, processing history, quality controls and governance status is potentially considerably more valuable scientifically than a larger dataset whose history cannot reliably be reconstructed. Provenance therefore should not be viewed merely as a compliance cost. It is part of the infrastructure that turns biological data into a trustworthy scientific asset.
Imec’s intervention in Brussels raised questions from MEPs about who controls biological data and models, how intellectual property should operate and how errors within training information should be handled. These questions apply equally to Britain. The UK government’s Life Sciences Sector Plan explicitly seeks to make British health and genomic assets attractive to international research and investment, which is economically rational — but greater movement of data across organisational boundaries increases the importance of being able to distinguish who owns the data, who controls access, who created a derived dataset, which transformations were performed, which rights travelled with it and which did not, and which evidence contributed to a commercial or scientific output. Provenance cannot resolve every intellectual-property question, but it can make those questions considerably easier to answer.
This does not require replacing the systems Britain is already building — quite the opposite. The NHS should continue operating clinical systems; Genomics England should operate genomic infrastructure; UK Biobank should manage its research resource; laboratories should retain their LIMS; researchers should use specialist analysis environments; the Health Data Research Service should provide secure routes to health information. The architectural opportunity is to establish an additional layer capable of recording verifiable evidence events across those boundaries: sample collected, laboratory accession created, sequencing performed, raw dataset generated, QC completed, pipeline executed, derived genomic dataset created, dataset deposited, research access authorised, dataset transformed, cohort assembled, AI training dataset created, model trained, model version validated, scientific or clinical conclusion generated. The objective is not necessarily to place the underlying biological information into another database. Instead, each significant event can produce a cryptographically verifiable record describing the evidence state at that moment. The underlying information can remain where governance, privacy, security and operational requirements dictate. The evidence chain records what existed, where it came from, what happened to it, and who asserted that event.
This distinction sits at the centre of the architecture we are developing at The BioChain, which we set out at greater length in Who Owns the Evidence Chain? and applied to the specific case of imec’s own European proposal in Europe Needs a Biological Data Infrastructure. BioChain is not intended to replace LIMS, electronic health records, genomic repositories, research environments or enterprise data platforms; it is intended to sit across them. The proposition is relatively simple: data should remain within the systems best suited to manage it, while evidence about that data should remain independently verifiable as it moves between them. A BioChain record can therefore represent an evidence event without requiring sensitive underlying biological information to be publicly exposed — a distinction that matters particularly in healthcare and genomics. A cryptographic commitment can demonstrate that a particular dataset or document existed in a particular state without requiring the evidence itself to be placed onto a public ledger, and later, an authorised party possessing the underlying information can verify it against that record. The objective is not radical transparency. It is verifiable continuity.
There is an opportunity here that extends beyond BioChain. Britain should consider whether provenance infrastructure deserves to become an explicit component of its AI-ready health-data strategy. A relatively modest demonstrator could test the concept: take a de-identified or synthetic workflow representing biological sample, sequencing, genomic processing, research repository, linked phenotype, AI dataset, model output; allow each participating system to remain independent; then construct a persistent evidence chain across the complete workflow. Change a source file. Change an annotation database. Update a pipeline. Withdraw an input. Retrain the model. Correct a clinical observation. The demonstrator should then answer a deceptively simple question: can we determine exactly which downstream scientific outputs were affected? If the answer can be produced reliably across heterogeneous systems, provenance stops being an archival function. It becomes an operational scientific capability.
Data engineering already understands lineage. Modern platforms routinely record where tables originated, which transformations occurred and which downstream datasets depend on them. Biomedical science needs to extend the principle, because a biological observation carries more than computational dependencies — it carries scientific ones. Its credibility may depend on a specimen, a person or organism, a laboratory, an instrument, a protocol, a reagent, a reference material, a software pipeline, a database version, an interpretation, a consent, an institution, and ultimately a chain of human and machine decisions. That is not merely data lineage. It is evidence lineage.
The UK has several characteristics that make it particularly suitable for developing this approach. It possesses a national health system capable of generating population-scale longitudinal information; internationally important genomic infrastructure; large research cohorts; mature biobanking capabilities; significant biomedical research institutions; and heavy investment in artificial intelligence and high-performance computing. Government has explicitly identified linked clinical, pathology, imaging and genomic information as strategic infrastructure for the next generation of life-science research. Few countries possess all of these simultaneously, which gives Britain an opportunity not simply to create one of the world’s largest AI-ready biomedical datasets, but to create one of the world’s most verifiable biomedical evidence environments. Those are not the same achievement.
The NIH phrase is compelling: collect data once, use numerous times. It describes exactly where biomedical science is heading. A blood sample collected today may contribute to clinical care; its associated information may later support epidemiology; its genomic data may contribute to a research cohort; that cohort may support drug discovery; the resulting dataset may train an artificial-intelligence system; that model may generate a hypothesis; the hypothesis may influence another experiment; and the resulting evidence may eventually contribute to a regulatory or clinical decision. We should want that reuse — it extracts considerably greater scientific value from evidence that is expensive, and sometimes impossible, to reproduce. But reuse changes the obligation. The more frequently evidence travels, the more important its history becomes. The emerging American model recognises this through explicit requirements for provenance and lineage. The European debate recognises the need to connect biological experimentation with AI infrastructure. Britain is already constructing many of the components required to do both. The missing question is whether we are connecting them merely as data, or preserving them as evidence — a distinction that could determine how trustworthy the next generation of biomedical AI ultimately becomes. Collect once. Use many times. Verify every time. That may be the infrastructure principle AI-enabled biology now requires.
Metadata is part of the solution, but metadata and verifiable provenance are not identical. Metadata can describe where information originated; a provenance architecture should additionally make it possible to establish whether the relevant evidence, metadata or processing record has subsequently changed, and to reconstruct the sequence of events connecting one evidence state with another.
No. Sensitive health or genomic information should remain within appropriately governed systems. The relevant concept is recording verifiable evidence events — potentially including hashes, signatures, identifiers, timestamps and relationships — without exposing the underlying sensitive information.
Yes, as do genomic laboratories, repositories and research environments. The challenge arises when evidence moves between independent systems and organisations. The requirement is not another audit log inside every application — it is continuity across the complete evidence chain.
AI systems can combine information from thousands of sources and propagate their outputs into subsequent models, analyses and decisions. If an upstream dataset is later discovered to contain an error, researchers need to determine where it travelled and what it influenced — which requires machine-readable lineage across the complete workflow.
No. Regulatory traceability is one benefit, but provenance also supports reproducibility, quality control, scientific investigation, intellectual-property management, model validation and the economic value of research datasets.
The immediate requirement is not another enormous national programme. It is experimentation — a cross-institutional demonstrator connecting genomic, clinical, research and AI workflows could establish which provenance information needs to survive organisational boundaries, and test whether an independent evidence layer can preserve it without centralising the underlying sensitive data.
If your organisation generates, manages, analyses or governs biological evidence and would be interested in participating in a UK or European provenance demonstrator, The BioChain would welcome the conversation.
Get in touch
Regulatory developments, technical notes and platform news — sent occasionally, straight to your inbox.