Pharmaceutical companies pool 20,000 confidential structures, and folding model doubles its accuracy

Trained on proprietary data, a modified OpenFold3 accurately predicts more than half of the protein-ligand complexes tested, compared to a third for the public version.

A consortium of pharmaceutical companies has re-entered OpenFold3, AlphaFold 3's open source replica, out of over 20,000 protein structures from their own archives — data that has never been transferred to public databases.

On a test set of 1,056 protein-ligand structures set aside before training, the fed model predicts more than half of the cases with high accuracy, where the public version of OpenFold3 achieves this for only one third. The difference suggests that the ceiling for current folding models is less related to their architecture than to the nature of the available data: public structures over-represent certain protein families and types of complexes.

Two reservations accompany the announcement, relayed by Nature on September 14th: the works are described in a blog post and have not been read by peers, and the model obtained is not made public. Experience shows above all the strategic value of data accumulated for decades in private laboratories.