The Gene and Protein Databases Behind Medical AI Get €65 Million From the EU

Where the €65 million goes

The package was confirmed after an amendment to the Horizon Europe work programme. It is divided between three named projects and one open call, as set out in the European Commission's Horizon Europe work programmes.

UniProt receives €30 million. The database holds protein sequences and records of what those proteins do. It is maintained by a consortium that includes EMBL-EBI in the United Kingdom and the SIB Swiss Institute of Bioinformatics in Switzerland. Sequence and structure models are trained on collections of this kind, and their output is checked against them.

GENCODE receives €19 million. The project produces annotation of the human genome: which regions code for proteins, where gene boundaries sit, which transcript variants exist. The Centre for Genomic Regulation in Barcelona is one of the groups involved. The annotation itself is browsable at gencodegenes.org.

Europe PMC receives €8 million. It is an open repository of life science literature, with abstracts and full texts that people can search and machines can process. It sits at europepmc.org.

The remaining €8 million is reserved for a competitive call. Other repositories holding life science data or knowledge can apply. No recipients have been named.

Why the funding model was changed

Databases are infrastructure, not experiments with an end date. For years they were funded through short project grants that had to be renewed, and the teams behind them spent part of every cycle applying for the next one. That is the model the Commission describes as fragmented, and the new money is structured at EU level instead, with continuity of service as the stated aim.

The practical risk is not hard to describe. A repository that stops updating does not vanish. Its entries stay online and slowly stop matching new literature, new protein records and new genome assemblies, while every tool that links to it inherits the gap. Rebuilding an abandoned repository is expensive, and parts of the work cannot be rebuilt at all.

Where AI enters the picture

Machine learning in biology stands on reference data. Protein models learn from sequence collections of the UniProt kind. Genome interpretation tools use annotation of the GENCODE kind. Language models that answer scientific questions are trained and evaluated on text of the Europe PMC kind. When the underlying data is incomplete, the model learns from an incomplete picture, and correcting that after training is neither quick nor cheap.

Since 2 August 2025, providers of general-purpose AI models in the EU have been bound by the transparency duties of Article 53 of the AI Act, supervised directly by the EU AI Office. In practice they must publish a summary of the content used for training. Documenting a public, versioned scientific dataset is manageable. Documenting a pile of scraped pages is not. Public research infrastructure therefore supplies something beyond raw material: it supplies training data that can be described.

The high-risk obligations of the same act were pushed back under the 2026 simplification package known as the AI Omnibus: high-risk rules now apply in 2027 and 2028, with Annex III systems and Annex I systems on separate application dates rather than one single deadline. That timetable concerns regulated systems such as medical devices and hiring tools, not the reference databases funded this month.

One detail worth keeping straight: this money goes to reference data and published literature, not to hospital records. Personal health data sits under a different set of rules, and nothing in the announcement changes how patient information is handled.

Open to the world, paid for in Europe

All three funded services are open access, and the Commission frames the €65 million as protection for that arrangement. Pharmaceutical companies, biotech startups and AI developers use these repositories on the same terms as universities, and the user base is global. A good share of it sits outside Europe.

Two of the recipients are themselves outside the EU. EMBL-EBI is in Hinxton, England, and SIB is in Switzerland. Their participation in European research infrastructure is not new, and the amounts announced this month continue an existing arrangement rather than starting one. Still, part of the €65 million ends up in institutions on non-EU soil, which is a detail that tends to surface whenever Brussels funds research capacity.

What the package does not settle

Sixty-five million euro is a serious commitment for services whose running costs are counted in millions rather than billions. It is not a permanent settlement. The announcement sets amounts for this round, includes €8 million for an open competition, and says nothing about what happens once the funded period ends.

Reference databases also grow. More sequences arrive, more papers are published, more genome assemblies are released. Storage and curation costs follow the volume, not the calendar, and curation is still done by people with doctoral degrees. Whether €65 million keeps pace with that growth over several years is a question the announcement does not answer. What is clear today is the part still undecided: a competitive call worth €8 million, open to repositories that have not yet been chosen.