Summary
A pharmaceutical consortium fine-tuned an OpenFold3-based model on 20,167 private protein–ligand structures. It outperformed comparable models trained on public data in a test of 1,056 held-out structures, although the study has not yet been peer-reviewed.
A consortium of pharmaceutical companies has reported that private drug-discovery data can substantially improve artificial-intelligence models that predict how proteins bind to potential medicines. The group fine-tuned an OpenFold3-based system on 20,167 proprietary protein structures and tested it against 1,056 structures held back from training.
The consortium model classified more than half of the held-out structures as meeting the study’s high-accuracy standard. The publicly available version of OpenFold3 reached the same standard on about one-third of the structures, while the open-source Boltz-2 model reached it on around 40%.
The work is described in a blog post rather than a peer-reviewed paper, and the resulting model is not publicly available. The consortium plans to submit a paper describing the study to a peer-reviewed journal.
Why protein–drug data matters
Proteins are long chains of amino acids that fold into three-dimensional shapes. Their shapes and surfaces determine how they interact with other molecules, including drugs. A model that can predict these structures may help researchers assess potential medicines before carrying out some of the most time-consuming laboratory work.
The Protein Data Bank (PDB), an open repository containing more than 200,000 experimentally determined protein structures, provided the main training foundation for AlphaFold 2. Its large collection helped AlphaFold 2 predict individual protein structures with high accuracy.
Later systems, including AlphaFold 3 and its open-source successors, also attempt to predict interactions between proteins and other molecules. This is a harder task because the public database contains relatively few experimentally determined examples of proteins bound to drug-like molecules—perhaps about 10,000, according to the source.
The missing examples are especially important for drug discovery. Research cited by the source has found that the accuracy of protein–molecule, or “co-folding”, models can fall sharply when they encounter molecules that differ substantially from those represented in their training data.
How the consortium trained the model
The AI Structural Biology Network was formed last year by AbbVie, Astex and other pharmaceutical companies to test whether their internal structural data could address this gap. The participating companies supplied structures generated during drug-development programmes using methods such as X-ray crystallography and cryo-electron microscopy.
The network began with OpenFold3, which had previously been trained only on PDB data. It then fine-tuned the model using 20,167 structures of proteins bound to potential drugs, also called ligands. Fine-tuning means adapting an existing trained model to additional examples rather than training a new system entirely from the beginning.
The data were provided in a way intended to keep each company’s proprietary structures private. The evaluation used 1,056 protein–ligand structures that had been excluded from training, allowing the consortium to test how the adapted model handled examples it had not seen.
The consortium model also outperformed systems trained on the individual datasets of single companies. John Karanicolas of AbbVie said this result points to an advantage from pooling information across organisations: each company’s internal dataset may cover only part of the range of protein–drug interactions needed for robust prediction.
A larger data question for drug discovery
The result supports a broader argument that the next gains in protein-structure AI may depend not only on better architectures but also on access to more experimentally determined structures involving drugs. Pharmaceutical companies may hold thousands of such structures that were generated for proprietary programmes and never deposited in public repositories.
The total size of these private collections is unknown. A separate project, OpenBind, has released hundreds of new protein structures and is developing thousands more, with support of up to £8 million (US$10.8 million) in UK government funding.
For now, the AISB result is an early consortium report rather than a completed peer-reviewed study. Its central finding is that adding private protein–ligand structures improved performance in the consortium’s held-out test; broader evaluation will depend on the planned paper and access to the model or comparable public data.