The race to accelerate drug discovery has found a powerful ally.
Boltz-2, a next-generation biomolecular foundation model developed by MIT’s Jameel Clinic and CSAIL in collaboration with Utah-based biotech startup Recursion, the AI model is carrying out drug discovery at supercomputer speed.
Unveiled two months ago, the AI model unifies complex-structure prediction,
binding-affinity estimation in a single package, thereby helping researchers screen vast molecular libraries in hours instead of weeks.
Boltz-2 compresses research and development cycles that once consumed immense time and cost, speeding up the whole process.
Drug discovery at supercomputer speed
At its core, Boltz-2 approaches the accuracy of physics-based free-energy perturbation (FEP) simulations, a gold standard for predicting how tightly molecules bind, while being up to 1,000 times faster.
On Recursion’s BioHive-2 supercomputer—ranked among the world’s most powerful—Boltz-2 can process millions of ligand-protein pairs in parallel, returning binding results in about 20 seconds per pair on a single NVIDIA A100 GPU.
This speed translates into a sea change for pharmaceutical research. Instead of laboriously testing compounds one by one in wet labs or waiting weeks for physics simulations to finish, scientists can now triage molecules digitally at supercomputer scale.
In benchmarks, Boltz-2 outperformed traditional docking methods and prior machine-learning approaches, even doubling average precision on high-throughput screens like MF-PCBA.
Engineering scale for real-world impact
Much of Boltz-2’s leap forward rests on engineering as much as biology.
Recursion’s BioHive-2 supercomputer, built on NVIDIA’s DGX SuperPOD with H100 Tensor Core GPUs, was optimized to handle the massive training data—over three million assay-labeled examples—needed to teach the model both pose and potency in a single forward pass.
NVIDIA engineers also played a critical role, profiling the model to eliminate computational bottlenecks.
They introduced custom cuEquivariance kernels that accelerated key “triangle” operations, cutting training and inference costs by up to 3x.
These optimizations, now available for other developers, not only reduced memory usage but made Boltz-2 viable as a production-scale tool for drugmakers.
For enterprises, the model is also packaged as Boltz-2 NIM, a production-ready microservice that ingests protein, RNA, DNA, or ligand sequences and returns 3D structures and affinity predictions with enterprise support.
The NIM deployment path offers biopharma companies higher throughput and lower compute spend, a critical edge in the high-stakes race to develop new medicines.
Boltz-2 was formally released on June 6 as an open-source model under an MIT license, with its code, weights, and training pipeline available for both academic and commercial use.
Trained on Recursion’s BioHive-2 and accelerated by NVIDIA, the model achieves near-FEP accuracy at speeds up to 1,000× faster, marking one of the most significant advances yet in AI-enabled drug discovery.
Released under an MIT license, Boltz‑2 can be retrained, fine-tuned, and deployed freely, while enterprise-scale inference is supported through the NVIDIA Boltz‑2 NIM under an NVIDIA AI Enterprise license.