Five pharma companies fine-tuned a protein-ligand structure prediction model on proprietary data and improved accuracy on a held-out test set

Five pharma companies fine-tuned a protein-ligand structure prediction model on proprietary data and improved accuracy on a held-out test set

On 14 September, Nature reported the results of the AI Structural Biology (AISB) network: five companies jointly fine-tuned OpenFold3 Preview 2, an open model that predicts the three-dimensional structure of a protein together with a bound molecule, using 20,167 proprietary structures of such complexes. For validation, they held out 1,056 structures that were not used during training.

For a drug to work, its molecule must occupy a specific site on the surface of the target protein. OpenFold3 builds a three-dimensional model of such a pair from the protein sequence and a description of the molecule. The structures most useful for this task often remain inside pharmaceutical programs, because they involve undisclosed molecules and target proteins.

In March, OpenFold released the weights, code, and data for OpenFold3, so the base model can be reproduced and fine-tuned. Protein-ligand complex structures collected within pharmaceutical programs, however, still belong to the companies.

Each company kept its structures on its own servers and fine-tuned its own copy of the model. The network then merged the updates into a shared version; this approach is known as federated learning.

The report from network participant Apheris compared how accurately the models reproduce the relative positions of protein and ligand. On the 1,056 held-out structures, the fine-tuned version reached the specified accuracy threshold in 52.1% of cases, compared to 35.6% for the base OpenFold3 Preview 2 and 40.9% for Boltz-2. A separate test measured how precisely the ligand is placed within the protein's binding pocket: 46.8%, versus 28.9% for the base OpenFold3 and 36.5% for Boltz-2.

The test structures also came from the same five companies: each held out 5% of its data, and the projects were split between training and validation. The AISB version outperformed models that individual companies had fine-tuned on their own data alone. The model trained on a pooled proprietary set of structures showed an advantage in this evaluation; the structures themselves and the weights of the new model remain with the network participants.