Private pharma data pushes AI to 52.1%

AI models learn from examples. The more relevant examples they see, the better they get. This is as true for predicting how drug molecules interact with proteins as it is for recognising faces or translating text. The problem is that some of the most relevant examples in drug design sit locked inside pharmaceutical company databases, unseen by any public AI system.

A consortium of five firms, AbbVie, Astex Pharmaceuticals, Bristol Myers Squibb, Johnson & Johnson and Takeda, says it found a way to train on that private data collectively without any company revealing its files. On the consortium's own benchmark, the resulting model performed substantially better than its publicly trained starting point. But neither the model, the test data nor a peer-reviewed manuscript has been released (Nature).

Teaching a model without pooling the data

The base model was OpenFold3 Preview 2, an open-source system that predicts how proteins and small drug-like molecules arrange themselves together in three dimensions. Think of it as predicting both the shape of a lock and how a key fits inside it (OpenFold3 repository).

Instead of merging all five companies' structures into one database, each firm trained a copy of the model on its own data locally. Only the numerical adjustments the model learned during training were shared back with a central coordinator, not the underlying molecular records. Berlin-based Apheris managed this coordination using NVIDIA's open federated-learning framework, while Mohammed AlQuraishi's laboratory at Columbia University contributed model expertise. The consortium reports using 40 NVIDIA H100 GPUs across five separate corporate environments (Apheris, Nature).

The raw structures stayed behind corporate firewalls. The consortium has not, however, published formal evidence that the shared model updates could not reveal information about the training data, a known theoretical risk in federated learning.

What the private structures added

The consortium trained on 20,167 proprietary protein-drug complex structures and tested on a separate private set of 1,056 structures (Apheris). Its headline measure asked how often the model's predicted drug-protein arrangement closely matched the experimentally observed one. The federated model achieved high-quality scores on 52.1% of test cases, compared with 35.6% for the original OpenFold3 and 40.9% for Boltz-2, another leading open model (Apheris, Nature).

A control model trained on updated public data alone did not match these gains, which the consortium interprets as evidence that its private structures supplied predictive information not captured by publicly available records (Apheris).

Why might that be? The public Protein Data Bank is vast but reflects which proteins academic labs chose to study. Pharmaceutical collections, the consortium and outside commentators suggest, tend to contain methodical series of one target paired with many related drug candidates, recording exactly the subtle binding variations that matter for design (Nature, C&EN). The consortium also reports that the federated model outperformed every single-company model, consistent with the idea that combining data across firms helped.

What outsiders cannot yet verify

No manuscript has been submitted. The model's learned numerical settings, the benchmark structures and the individual predictions are all unavailable to outside researchers. No confidence intervals or statistical tests have been published. The benchmark was constructed and evaluated by the same organisations that built the system (Nature).

This matters in part because of a known pitfall. Work by the PoseBusters group showed that predictions from other AI structure models can look geometrically close to a reference structure while placing atoms in physically impossible arrangements, such as bonds at angles that don't occur in real molecules (PoseBusters). Whether AISB's predictions pass those checks is unknown.

The benchmark itself measures how well the model reconstructs known experimental arrangements. It does not measure whether this helps identify compounds that actually bind to their target in a living system or prove safe. An accurate predicted pose is useful to a chemist deciding which compounds to investigate next, but the consortium has not reported any downstream drug-design outcome.

The ten-company MELLODDY project previously showed federated learning could improve molecular property predictions across pharmaceutical competitors (MELLODDY). AISB extends that principle into three-dimensional structure prediction. The next step the field needs from this consortium is a published manuscript and enough inspectable data for independent researchers to test whether the reported gains hold outside consortium-controlled evaluation.