EU Joint Research Centre · Health

Biological AI models: new paradigms to leverage the languages of life

A new JRC report, "Artificial Intelligence for Biology: Capabilities, Readiness, and Policy Implications," assesses the current state of AI models applied to biological data. While AI is rapidly transforming biological research, particularly in protein-centric applications, areas like single-cell biology lag due to data limitations. The EU has a strong scientific base but needs improved collaboration, data governance, and readiness assessments for real-world deployment.

AI models for biological data, including DNA, RNA, and proteins, are advancing quickly, outpacing existing governance frameworks. Protein-centric AI models, used in drug design and discovery, are highly advanced due to extensive research and curated data from repositories like the Protein Data Bank and UniProt, supported by European research infrastructures.

In contrast, areas like single-cell biology, relevant for predicting immunotherapy response in tumors, are less developed due to limited and non-standardized data. The report introduces a framework to assess deployment readiness by combining domain-specific maturity with technology readiness levels (TRL).

High-profile models like AlphaFold and ESM3, while domain-mature, are at low-to-mid TRL, meaning they are not yet certified for clinical or industrial use. This "maturity paradox" highlights insufficient validation of the innovation pipeline, posing risks like biosecurity concerns if publicly available models are misused for applications such as pathogen design.

Building biological AI models relies on training data, computational infrastructure, and collaboration. Training data is unevenly distributed, with fragmented repositories hindering reproducibility. While the US and EU host most training datasets, industry often has greater access to computing resources than academia.

Collaboration patterns show academia's significant involvement, but industry's increasing activity may lead to less disclosure of proprietary models and datasets. Intra-EU collaboration is limited compared to collaborations with the US, China, and the UK, which dominate model development.

The report recommends the EU broaden support for emerging research topics, strengthen biological data infrastructure through better coordination, support European biological AI foundation models as public goods, and develop frameworks to assess both scientific maturity and technology readiness with clearer regulatory pathways.

AI-samenvatting op basis van de bron.