AI-generated from publicly available materials.Four major pharmaceutical companies—AbbVie, AstraZeneca, Bristol Myers Squibb, and Johnson & Johnson—are collaborating on a new initiative to enhance drug discovery through AI-driven predictions of molecular interactions. This project, named AISB Bind, is a significant step in leveraging federated learning to optimize drug development processes.
The AISB Bind initiative, announced by the AI Structural Biology (AISB) Network and supported by Apheris and the AlQuraishi Lab at Columbia University, aims to train machine learning models using data from these companies without compromising sensitive information. The training is set to commence this month, with the first models anticipated by early 2027. Apheris co-founder Robin Roehm emphasized the potential of combining diverse data from private repositories to create superior predictive models.
This collaboration builds on previous efforts, particularly the Federated OpenFold3 project, where the companies successfully improved the accuracy of protein binding predictions by utilizing proprietary data while ensuring that no sensitive information left their environments. The resulting model, AISB-1-Fed, demonstrated significant improvements in prediction accuracy compared to its predecessors, showcasing the value of shared data in refining AI models.
Despite the promising advancements, Roehm noted that challenges remain. The diversity of training data is crucial, as each company possesses a unique dataset shaped by its historical focus. The initiative aims to address this by pooling data to enhance the model's generalizability. Bristol Myers Squibb's involvement highlights the industry's recognition of the limitations of isolated data and the potential benefits of collaborative efforts in AI-driven drug discovery.