EU report urges stronger governance for biological AI models

New Joint Research Centre research based on a dataset of 480 biological AI models finds that progress is strongest in data-rich areas such as protein structure, function annotation and molecular design. Protein-focused systems have benefited from long-running research communities and curated resources including the Protein Data Bank, UniProt and European Research Infrastructures such as the European Molecular Biology Laboratory.

Fields with less abundant and less standardised data, including single-cell biology, remain less developed despite clinical relevance in areas such as tumour characterisation and predicting responses to immunotherapy. The report says strong benchmark performance does not necessarily mean a model is ready for clinical or industrial deployment, noting that models including AlphaFold and ESM3 are domain-mature but remain at low-to-mid TRL.

The analysis identifies a “maturity paradox” in which scientific sophistication outpaces validation across the full innovation pipeline, including governance and real-world integration. It also flags risks where publicly available models could be misused for pathogen design or toxin engineering.

Europe has strong scientific capacity and high-performance computing resources through EuroHPC JU and AI Factories, but collaboration and data governance remain uneven. Academia participates in the development of 85% of surveyed models, industry participates in nearly 40%, and only 17% of industry-only developed models release training code.