AI-generated from publicly available materials.Five major pharmaceutical companies have successfully collaborated to train a cutting-edge AI model on over 20,000 proprietary drug-discovery structures, achieving superior performance compared to publicly available benchmarks. This initiative, led by the AI Structural Biology (AISB) Network, marks a significant advancement in the use of federated learning for drug discovery.
The federated model, named AISB-1-Fed, was developed using structural data from AbbVie, Astex Pharmaceuticals, Bristol Myers Squibb, Johnson & Johnson, and Takeda, in partnership with Columbia University's lab. This collaborative effort resulted in a model that surpassed previous benchmarks, achieving 52.1% high-quality predictions for protein-ligand interfaces, a notable increase from the 35.6% of the baseline model, OpenFold3 Preview 2. The consortium highlighted this achievement as a clear indication that federated fine-tuning can lead to significant improvements in co-folding accuracy, rather than marginal gains seen in past attempts.
The initiative addresses a critical challenge in drug discovery: the scarcity of relevant structural data. While public databases contain numerous protein structures, only a small fraction pertains to drug interactions, which are often closely guarded by pharmaceutical companies. The AISB Network was formed to leverage the collective data of its members, effectively tripling the available training data for AI models in this field.
This project exemplifies the potential of federated learning, allowing companies to collaborate without compromising sensitive data. The technical framework ensured that only model parameters were shared, safeguarding proprietary information while enhancing model performance. The success of AISB-1-Fed not only paves the way for future advancements in drug discovery but also demonstrates the power of collaborative efforts in overcoming data constraints in the pharmaceutical industry.