RecursionDrug discovery

A clinical-stage biotechnology company manages 50 petabytes of biological and chemical data to industrialize the drug discovery process.

Traditional physics-based simulations required weeks to screen compound libraries, while earlier AI models could not accurately predict how strongly a drug would bind to its target. This technological gap forced scientists to rely on slow, expensive wet lab experiments to validate potential candidates.

The team trained the Boltz-2 foundation model on a DGX H100 SuperPOD to simultaneously predict protein structures and binding affinities. Custom GPU kernels were developed to accelerate mathematical operations and optimize memory usage during the training process. The model deploys as a containerized microservice, enabling researchers to integrate high-speed inference directly into screening workflows.