Five pharma rivals trained a shared AI on 20,000 private structures

Five competing pharma companies used federated learning to train a shared model on 20,000 private structures, with no company seeing another's data. It outperformed every public protein-ligand co-folding benchmark, lifting high-quality predictions from 36% to 52%.

• Trained on 20,000 private structures without sharing data between companies
• Outperformed every public protein-ligand co-folding benchmark
• High-quality predictions rose from 36% to 52%
• First successful AI collaboration among rival pharma firms