Pharma Consortium’s Secret Data Boosts AI Protein Models

Key points

• A consortium of pharmaceutical companies, including AbbVie and Astex, formed the AI Structural Biology (AISB) Network to pool proprietary data.
• The group fine-tuned the open-source OpenFold3 model on over 20,000 proprietary protein structures, specifically focusing on protein-ligand interactions.
• The resulting model predicted more than half of a test set of 1,056 structures with high accuracy, outperforming public-data-only models and individual company silos.
• Public databases like the Protein Data Bank (PDB) lack sufficient examples of drug-like molecules interacting with proteins, limiting current AI models.
• The study, reported in a blog post, is not yet peer-reviewed, and the model is not publicly available, though the team plans to submit a paper to a journal.

Background

AlphaFold, which won the 2024 Nobel Prize in Chemistry, revolutionized protein structure prediction using public data from the Protein Data Bank (PDB). However, newer models like AlphaFold 3 struggle with 'co-folding'—predicting how proteins interact with drugs—because the PDB contains very few examples of such interactions. This data gap has driven pharmaceutical companies to explore their internal, proprietary datasets, which are estimated to be larger than public repositories but remain locked in company vaults due to trade secrets.

Why it matters

This development suggests that the future of AI in drug discovery may depend on collaborative data sharing among private companies rather than relying solely on public repositories. By pooling proprietary data, the industry can create more accurate models for predicting drug interactions, potentially accelerating the development of new therapies and addressing the limitations of current open-source tools.

What to watch

The AISB Network plans to submit a peer-reviewed paper detailing their findings. Additionally, efforts like the UK-funded OpenBind project are working to release new public protein structures to address data gaps. The success of this private consortium may influence future policies on data sharing and the development of open-source AI models in the pharmaceutical industry.