Europe is beginning to confront one of the central constraints on AI-driven drug discovery: not simply computing power, but access to sufficiently rich biological data. Building the infrastructure to generate that data is essential. But if the resulting evidence is to move between laboratories, companies, models and regulators, Europe will need to solve a second problem at the same time — how the evidence itself remains traceable as it does.
Artificial intelligence has changed the economics of biological discovery. Models can increasingly operate across chemical structures, genomic sequences, protein interactions, cellular phenotypes and clinical observations, with an ambition to identify promising therapeutic targets earlier, predict failures before expensive experimental programmes begin, design new biological entities, and shorten the journey from hypothesis to medicine. But there is a fundamental constraint underneath that ambition: an AI system can only learn from the evidence available to it, and in biology, generating that evidence remains extraordinarily difficult.
That constraint was at the centre of an intervention made to the European Parliament’s Committee on Public Health (SANT) in Brussels on 8 September 2026. Liesbet Lagae — co-founder and Program Director of Life Science Technologies at the Belgian research and innovation organisation imec, and a researcher with more than 165 peer-reviewed papers and 25 patents to her name — argued that Europe’s investment in artificial intelligence and health data should be accompanied by infrastructure capable of producing experimental biological data at scale. Her case was that hospital data alone is not enough: it needs to be complemented with experimental information about how cells, proteins and tissues actually function, generated through tools such as DNA sequencing, neuronal-activity recording, miniaturised toxicity-testing systems and DNA synthesis. That experimental evidence, she noted, currently sits scattered across academic laboratories, hospitals and pharmaceutical companies rather than being connected into anything resembling a shared infrastructure.
The distinction Lagae drew matters. Clinical datasets tell us what happened to patients; experimental biological datasets can help explain why. Genomics, proteomics, cellular measurements, tissue models and other experimental systems provide information about the mechanisms through which disease develops and potential therapies act, and AI-driven medicine will increasingly require both. MEPs pressed her on exactly the questions this raises: Ondřej Dostál asked about the implications for patents and how value would be shared between the infrastructure, the developers of the models built on it, and the organisations providing the underlying data; András Kulja asked how the initiative would manage distortions in the data and balance the drive to accelerate research against the need to properly evaluate long-term effects. No European budget, implementation timeline or special authorisation procedure for AI-designed drugs was established during the discussion — this was a proposal being tested in front of its future regulators, not yet a settled programme.
Europe therefore faces an infrastructure question, but underneath it lies another one. Imagine a future European biological data infrastructure operating exactly as intended: a tissue sample is collected, a laboratory performs an assay, an instrument produces measurements, software processes the raw signal, quality-control procedures remove problematic observations, and the resulting dataset is deposited into the infrastructure. Another organisation harmonises it with data generated elsewhere. A biotechnology company obtains authorised access, and the information becomes part of a training dataset. A model identifies a biological relationship; researchers use that prediction to prioritise a therapeutic candidate; the candidate ultimately contributes to a regulatory submission.
At the beginning of that sequence is an observation. At the end may be a decision affecting patients. Between them are potentially hundreds of transformations. The important question is therefore not simply where the data is stored, but whether we can reconstruct what happened to the evidence along the way — a distinction that will only become more important as biological data infrastructures grow.
Europe has become very good at discussing data infrastructure: repositories, federated architectures, interoperability, common data spaces, APIs, compute resources and access controls. All of that is necessary, but none of it automatically creates provenance. A repository can tell us that a dataset exists. An access-control system can tell us who is permitted to retrieve it. A laboratory information system can document laboratory processes. A workflow platform can record computational jobs. A model registry can identify a model. Yet a scientific conclusion may depend on evidence that has travelled through all of them, and no individual system necessarily holds the complete history.
Call this the inter-system provenance problem: the evidence chain crosses organisational and technological boundaries while the systems responsible for recording it remain local. This is not an argument for replacing those systems — quite the opposite. The laboratory system should remain responsible for laboratory operations, the repository for storing research data, the hospital system for clinical records, and the AI platform for models and computation. What is missing is a means of preserving the relationship between the evidence held across all of them.
Scientific evidence has always had provenance of a kind. Researchers record experimental methods, laboratories maintain notebooks, instruments produce metadata, and publications describe analytical procedures. But AI changes the scale involved. A conventional scientific analysis may involve hundreds or thousands of observations; a biomedical foundation model may eventually depend on millions or billions. The resulting evidence chain becomes too large to reconstruct manually.
Consider a single training dataset assembled from fifty institutions. Some contributing datasets may have been updated, others corrected; some records may have been excluded after quality-control failures; consent conditions may differ; analytical pipelines may have changed; reference genomes may have been updated; annotations may have been revised; and the model itself may have gone through multiple training runs. The apparently simple question — which evidence produced this result? — rapidly becomes surprisingly difficult to answer.
This is why the proposal discussed at the European Parliament deserves attention beyond its immediate policy details. Europe has an opportunity to think about evidence architecture while the biological infrastructure itself is still being conceived; retrofitting provenance afterwards would be considerably harder. A European biological data infrastructure could therefore be designed around several principles.
Persistent identity. Samples, datasets, workflows, analytical outputs and models should have persistent identifiers — not simply filenames, and not database row numbers meaningful only within one system, but identities that survive movement between systems.
Transformation provenance. When evidence changes, the relationship between the old and new state should be recorded. If dataset B was generated from dataset A through pipeline X, that relationship matters; if dataset C excluded observations from B following quality control, that matters too. The scientific record is not simply C — it is A → B → C, with the transformation between each state preserved.
Authority. Provenance must record not merely what happened but who or what authorised it — a laboratory, a researcher, an institution, an automated workflow, a validated analytical pipeline, and eventually, perhaps, an autonomous scientific agent.
Time. Evidence changes: annotations improve, databases are corrected, models are retrained. Scientific provenance must therefore be temporal. The relevant question may not be what the database says today, but what evidence was available when a particular decision was made — a distinction that can be critical in regulatory science and outbreak investigation.
Verifiability. Perhaps most importantly, provenance should not depend entirely on trusting the organisation currently holding the database. A durable evidence architecture should make it possible to establish that a particular object existed in a particular state, and that subsequent changes to it can be identified. That does not mean placing sensitive biological data onto a public blockchain, which in many circumstances would be inappropriate; it means separating the evidence itself from cryptographic evidence about the evidence.
That distinction is important. Genomic datasets can be enormous, clinical information can be highly sensitive, and commercial research data may constitute valuable intellectual property, so moving that information into another infrastructure merely to establish provenance would often make little sense. Instead, the underlying information can remain where it belongs: a hospital retains its clinical records, a sequencing centre retains its genomic files, a research repository retains its deposited datasets, and a pharmaceutical company retains its proprietary experimental information.
What can travel between them is a much smaller evidence record — an identifier, a cryptographic digest, a timestamp, the identity of the responsible authority, the relationship to preceding evidence, relevant workflow metadata, and, where appropriate, references to policies or permissions. The result is not a biological database. It is an evidence graph describing biological data.
The European biological infrastructure envisaged by imec would necessarily involve organisations with very different interests: universities that want to publish, biotechnology companies that want to innovate, pharmaceutical companies that need defensible development pipelines, hospitals that must protect patients, technology companies that want to develop models, regulators that require evidence, and patients and citizens who expect appropriate governance. These actors do not need to trust one another completely, but they do need to collaborate, and provenance provides a mechanism through which that collaboration can occur without requiring every participant to surrender control of its underlying systems. An organisation can retain custody of its data while still contributing verifiable evidence about that data to a wider research ecosystem — a distinction that could become increasingly important as European biomedical AI becomes more federated.
The questions raised during the European Parliament discussion — control of data, intellectual property, safety — are essential ones. But provenance introduces another dimension. Imagine that five institutions contribute data used to develop a therapeutic model, a sixth organisation develops the model, a seventh validates its predictions, and an eighth develops a resulting therapeutic candidate. Who contributed what? Which dataset influenced which model, and which version of it, under what licence? Which transformations occurred before the data entered the model, and which experimental result validated the prediction? These are intellectual-property questions, but they are also questions of evidence. A sufficiently mature provenance infrastructure could therefore support not only scientific reproducibility but attribution, licensing, auditing and, eventually, new mechanisms for allocating value across collaborative research networks.
There is a further reason this matters. AI systems increasingly mediate the relationship between observation and scientific conclusion, which means regulators may increasingly need to inspect not merely experimental results but the computational lineage through which conclusions were produced. A future regulatory evidence package might therefore need to trace something resembling sample → assay → raw measurement → processed dataset → training corpus → model version → inference → validation experiment → candidate. Each arrow represents a transformation; each transformation represents a potential source of uncertainty; and each, ideally, should be reconstructable.
The FAIR principles — making data Findable, Accessible, Interoperable and Reusable — transformed scientific data management. The next challenge may be to make scientific evidence verifiable. A dataset can satisfy many FAIR principles while its subsequent history remains difficult to reconstruct, and AI exposes that weakness because models can absorb evidence from thousands of sources while obscuring the relationship between individual observations and downstream conclusions. The scientific infrastructure of the next decade may therefore require an additional property: lineage — not merely knowing where information is, but knowing how it became what it is.
At The BioChain, we think about this as the evidence layer. The BioChain is not intended to replace laboratory information management systems, electronic health records, genomic repositories, data lakes or AI platforms — those systems perform important, specialised functions. The evidence layer sits between them, preserving verifiable relationships between scientific objects as they cross organisational and technological boundaries. The original data can remain in existing systems; what The BioChain records is evidence about evidence — what existed, when it existed, who asserted it, what preceded it, what subsequently changed, and whether that history can still be independently verified. We set out this thinking in more depth in Who Owns the Evidence Chain?, and the architecture becomes particularly relevant precisely where no single organisation controls the entire scientific process — which is exactly the kind of research environment Europe is now contemplating.
We return to this question, and to the United States and United Kingdom initiatives now moving in parallel, in Europe Is Building the Infrastructure for AI Biology. It Should Build the Evidence Chain With It..
The next generation of biological research infrastructure will almost certainly produce extraordinary volumes of data. Those data will train models; those models will generate hypotheses; those hypotheses will direct experiments; those experiments will generate more data. Biology is becoming a computational feedback loop, and the integrity of that loop will depend not only on how much information Europe can generate, but on whether we can establish where that information came from. For a related, more concrete illustration of the problem — what happens when the underlying measurements feeding a model are not themselves comparable across laboratories — see a related, more concrete illustration of the problem.
For how the same question plays out in Britain’s own AI-ready health-data build-out, see The Missing Layer in Britain’s AI-Ready Health Data Infrastructure.
The question therefore should not simply be whether Europe can build a biological data infrastructure for AI. It should also be whether Europe can build an infrastructure in which the evidence remains traceable as humans and machines repeatedly transform it — because the most valuable scientific dataset in the world becomes considerably less valuable when nobody can reliably reconstruct how it came to exist.
No. The underlying architectural problem is provenance. Distributed ledgers are one possible component of a provenance architecture, but they should not be confused with the objective itself. Sensitive biological and clinical information generally should not be written directly to public ledgers.
No. A provenance layer should complement existing infrastructure rather than replace it. Repositories, LIMS platforms, electronic health records and analytical systems should remain authoritative within their respective domains.
Metadata is essential but is often stored within individual systems. The harder problem arises when evidence crosses between systems and institutions; preserving those relationships requires persistent identity and interoperable provenance.
AI can combine evidence at a scale impossible for human researchers, which magnifies both the opportunity and the provenance problem. A conclusion may ultimately depend on thousands of datasets and computational transformations.
Potentially. Reliable records of contribution and derivation could improve attribution and provide evidence relevant to licensing or ownership disputes. Provenance itself, however, does not determine legal ownership.
Treat provenance as part of the infrastructure design rather than a compliance feature added afterwards. If Europe’s biological data infrastructure is meant to support decades of AI-driven science, evidence lineage should be considered foundational from the outset.
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.