Case Study: Recursion accelerates open-source protein co-folding and binding-affinity predictions with NVIDIA AI

Recursion accelerates protein co-folding 2x–3x with NVIDIA

The customer, Recursion, partnered with NVIDIA to tackle the challenge of accelerating AI-driven drug discovery. They needed to build and scale a sophisticated biomolecular foundation model that could predict both protein complex structures and their binding affinities, a process that was traditionally slow and computationally expensive.

NVIDIA provided the solution with its DGX SuperPOD infrastructure, which powered Recursion's BioHive-2 supercomputer for training the model. NVIDIA also developed custom cuEquivariance kernels to accelerate the model's performance. The result was Boltz-2, which predicts structures and affinities with high accuracy in about 20 seconds on a single A100 GPU, a process that is 1,000 times faster than previous physics-based methods and significantly accelerates drug discovery workflows.

Recursion