How Recursion and MIT Built Boltz-2 to Cut Drug Discovery Wet Lab Work 40% with NVIDIA

Recursion is a US-based biotechnology company applying computational and AI methods to drug discovery, operating at the intersection of machine learning and experimental biology. MIT’s Jameel Clinic and Computer Science and AI Laboratory (CSAIL) are research partners focused on applying AI to medicine. Their collaboration targeted one of the central bottlenecks in structure-guided drug discovery: the fact that predicting protein-ligand binding structures and estimating binding affinities has historically required separate computational pipelines, each consuming significant GPU time and wet lab validation resources.

Physics-based methods for binding affinity estimation traditionally took weeks to process large compound libraries. The computational cost of running these predictions at the scale required for modern drug screening programs created a fundamental throughput constraint: teams could only evaluate a fraction of candidate compounds in the time available, forcing heavy reliance on expensive, time-consuming wet lab experiments to validate predictions that should have been pre-filtered computationally.

Working on NVIDIA’s BioHive-2 supercomputer — a 63-node DGX H100 system — MIT and Recursion trained Boltz-2, an open-source foundation model that unified protein complex co-folding and binding affinity prediction into a single architecture. NVIDIA cuEquivariance kernels accelerated the custom equivariant operations required for accurate molecular geometry modeling, delivering up to 5x speedup on triangle operations versus standard implementations. NVIDIA NIM packaging enables production deployment of Boltz-2 as an enterprise microservice with 2x–3x inference speedup, and the model is available as Boltz-2 NIM for teams deploying at scale.

The results validated the approach across multiple benchmarks. Boltz-2 ranked first on CASP16 affinity prediction data. It achieves 0.62–0.66 Pearson correlation on affinity predictions and processes individual ligand-protein pairs in approximately 20 GPU-seconds on an A100. At the application level, the operational transformation is significant: as Recursion CTO Ben Mabey described it, “With AI in the loop today, we can get 80% of the value with 40% of the wet lab work.” That ratio — maintained prediction quality while substantially reducing experimental burden — is the metric that matters for drug discovery pipelines operating under time and cost pressure.

Boltz-2 is open-source and available for broad adoption across the drug discovery community. Its combination of structure prediction and affinity estimation in a single model, trained on a purpose-built supercomputing cluster and packaged for enterprise deployment via NVIDIA NIM, establishes a new baseline for what computational drug discovery infrastructure can deliver before wet lab experiments begin.