AI Breakthrough: How Boltz-2 Is Accelerating Drug Discovery by 1000x

Scientists at the Massachusetts Institute of Technology, in collaboration with the biotechnology startup Recursion and engineers from NVIDIA, have unveiled Boltz-2—a groundbreaking biomolecular AI system designed to radically accelerate drug discovery. Released under the MIT license in June, this model integrates the prediction of three-dimensional structures with the assessment of molecular binding affinities into a single tool. As a result, the evaluation of millions of compounds now takes hours rather than weeks.

The principal advantage of Boltz-2 lies in its accuracy, which approaches that of free-energy perturbation (FEP) physical simulations—long regarded as the gold standard for measuring molecular binding strength—while operating up to a thousand times faster. Running on Recursion’s BioHive-2 supercomputer, powered by the NVIDIA DGX SuperPOD with H100 Tensor Core GPUs, the model can analyze millions of ligand–protein pairs simultaneously, delivering results in just 20 seconds per GPU. For the pharmaceutical sector, this represents a transformative breakthrough: instead of laborious laboratory experiments or weeks-long computational delays, researchers can instantly eliminate unsuitable molecules and focus on promising candidates.

In benchmarking tests, Boltz-2 outperformed classical molecular docking methods and prior machine learning approaches, achieving double the accuracy in large-scale screening tasks, including MF-PCBA. This leap was made possible through engineering innovations: NVIDIA’s specialists removed computational bottlenecks by introducing proprietary cuEquivariance kernels, which accelerated key triangular operations. These optimizations reduced both training and inference costs by up to threefold while lowering memory consumption.

The model was trained on more than three million biological assay samples, enabling Boltz-2 to simultaneously capture molecular spatial configurations and activity. In practice, this translates into faster, more precise identification of potential drug candidates at the earliest stages of research.

For industry use, a dedicated version—Boltz-2 NIM—has been developed. This microservice accepts sequences of proteins, RNA, DNA, or ligands and returns their three-dimensional structures along with interaction predictions. Integrated with NVIDIA AI Enterprise, the service allows pharmaceutical companies to significantly cut computational costs while maintaining research efficiency at scale.

Because the model is released under the open-source MIT license, it can be freely used, adapted, and retrained. Combined with powerful computing clusters such as BioHive-2, Boltz-2 emerges as a tool capable of dramatically accelerating the creation of new therapeutics. Amid the fierce race for innovative treatments, the project is already being hailed as one of the most pivotal advances in applying AI to molecular biology.