Archive · 2 Oct 2026
Drug Rivals Train Shared AI Without Sharing Molecular Data
Five drug companies wanted to teach an AI about medicines, but they did not want to hand their private information to one another.
They used a method called federated learning instead.
Each company trained its own copy of the AI using its own molecular data.
The companies shared only improvements to the AI, not the confidential molecule files.
Their combined model was better at predicting how drug molecules fit with proteins than the public model it started from.
It made strong predictions for 52.1% of more than 1,000 test molecules, compared with 35.6% for the public model.
This could help researchers avoid some failed experiments.
However, the test was conducted internally, and a good prediction is still far from creating an approved drug.
AbbVie, Astex Pharmaceuticals, Bristol Myers Squibb, Johnson & Johnson and Takeda jointly trained AISB-1-Fed with Columbia University researchers.
The companies used more than 20,000 proprietary molecular structures without transferring any structure files between organizations.
Federated learning combined model improvements produced inside each company while keeping the underlying data private.
On a test of more than 1,000 private structures, AISB-1-Fed made high-quality binding predictions for 52.1%, compared with 35.6% for the public model.
The result is promising but comes from an internal benchmark and does not demonstrate that the system can produce an approved medicine.
- Who
- AbbVie, Astex Pharmaceuticals, Bristol Myers Squibb, Johnson & Johnson, Takeda and a Columbia University laboratory.
- What
- They jointly trained the AISB-1-Fed AI model using federated learning without sharing proprietary molecular structures.
- When
- The collaboration was completed in under ten weeks.
- Where
- Each company trained the model inside its own systems; the article does not identify a single shared training location.
- Why
- To improve drug-protein binding predictions while protecting commercially sensitive molecular data.
Promise of Federated Learning
Caveats and Limitations
Data collaboration
Promise of Federated Learning
The project shows that rival companies can jointly improve an AI model without exposing their confidential molecular structures.
Caveats and Limitations
Federated learning is more difficult to operate than ordinary centralized training, which may limit how widely it is adopted.
Drug-discovery impact
Promise of Federated Learning
Better predictions of how molecules bind to proteins could help researchers reduce laboratory dead ends.
Caveats and Limitations
The result is based on an internal benchmark, has not yet been independently scrutinized in peer-reviewed literature, and does not amount to an approved drug.
- AI model
- AISB-1-Fed
- Participants
- AbbVie, Astex Pharmaceuticals, Bristol Myers Squibb, Johnson & Johnson, Takeda and Columbia University researchers
- Training data
- More than 20,000 proprietary molecular structures
- Test set
- More than 1,000 private structures
- AISB-1-Fed result
- High-quality binding predictions for 52.1% of test structures
- Public model result
- High-quality binding predictions for 35.6% of test structures
- Training time
- Under ten weeks