Boltz-2

Overview

Boltz-2 is an open-source model developed by researchers at MIT CSAIL (Barzilay lab) in collaboration with Recursion Pharmaceuticals. Released in June 2025, it extends Boltz-1 by combining biomolecular structure prediction with direct binding affinity estimation for small-molecule complexes — making it the first openly available model to tackle both tasks within a single unified framework. This represents a meaningful shift in computational drug discovery, where structure prediction and affinity scoring have historically been handled by separate, disconnected pipelines.

The central innovation of Boltz-2 is its ability to predict the 3D binding pose of a small molecule within a protein active site while simultaneously estimating the relative binding free energy of that complex. On standard benchmarking datasets, Boltz-2 achieves a Spearman correlation of approximately 0.6 with FEP+ (free energy perturbation plus), which is considered the gold-standard computational method for binding affinity ranking. Critically, it delivers these predictions in seconds on a single GPU, compared to the hours or days typically required by physics-based free energy approaches running on high-performance computing clusters.

By releasing model weights, training code, and inference utilities under permissive open-source licensing, Boltz-2 lowers the barrier to high-quality affinity prediction for academic labs and smaller biotech organizations that lack access to commercial FEP software or large compute budgets.

Key Features

• Joint Structure and Affinity Prediction: Simultaneously predicts the 3D binding pose of a small molecule within a protein active site and estimates relative binding free energy, unifying what previously required two separate modeling pipelines.
• FEP+ Correlation: Achieves approximately 0.6 Spearman rank correlation with FEP+ on standard relative binding free energy benchmarks, approaching physics-based accuracy from a machine learning model.
• Single-GPU Inference: Completes predictions in seconds per complex on a single consumer or workstation GPU, enabling rapid virtual screening without access to HPC infrastructure.
• Open Weights and Code: Released under permissive open-source licensing with publicly available model weights, training code, and inference utilities via the jwohlwend/boltz GitHub repository.
• Boltz-1 Foundation: Inherits Boltz-1's competitive structure prediction capabilities for proteins, nucleic acids, and small molecules, with affinity prediction added as an additional output head.

Technical Details

Boltz-2 extends the Boltz-1 diffusion-based architecture with a binding affinity prediction module. The core structure prediction backbone uses a denoising diffusion process to generate 3D atomic coordinates for protein-ligand complexes. A pairwise representation module encodes inter-atomic relationships between protein residues and ligand atoms, providing the geometric context required for both accurate pose prediction and affinity estimation. A learned regression head then operates on these complex-level representations to estimate relative binding free energies without requiring explicit thermodynamic cycle calculations.

Training is multi-task: the model is supervised simultaneously on structure prediction objectives using Protein Data Bank co-crystal structures and on affinity prediction objectives using experimental binding data (IC50, Kd, Ki values) drawn from ChEMBL and related databases. FEP benchmark datasets of curated congeneric series are used for evaluation. On structure prediction, Boltz-2 retains competitive performance on the PoseBusters benchmark for protein-ligand pose quality. On affinity prediction, the approximately 0.6 Spearman correlation with FEP+ is notable given that it is achieved at a tiny fraction of the computational cost; the practical ceiling for FEP+ itself on well-parameterized systems is around 0.8 Spearman correlation.

Applications

Boltz-2 is positioned primarily for structure-based drug discovery workflows. In lead optimization, it can rapidly rank congeneric compound series by predicted binding affinity to guide synthesis prioritization, replacing or augmenting expensive FEP+ campaigns in early discovery. It is also well-suited to virtual screening, where large libraries of candidate compounds are scored against a target in a single pass combining pose prediction and affinity estimation. Medicinal chemists can use Boltz-2 as a fast pre-filter before committing to full relative binding free energy (RBFE) calculations, substantially reducing the number of expensive simulations required. The ability to visualize predicted binding poses alongside affinity estimates also supports structure-guided design decisions about which scaffold positions to modify.

Impact

Boltz-2 represents a significant step toward democratizing binding affinity prediction, a capability that has historically been gated behind expensive commercial software (Schrodinger FEP+) and substantial computational resources. By making a model approaching FEP+ accuracy openly available, it enables academic labs and resource-constrained teams to conduct quantitative affinity-driven lead optimization. Key limitations remain: performance is strongest for congeneric series sharing a common scaffold, absolute affinity prediction and ranking across chemically diverse sets are more challenging, and the model predicts a single low-energy conformation without fully accounting for protein flexibility or allosteric effects. As of its release, the technical paper remains a preprint and has not yet undergone formal peer review. Nonetheless, Boltz-2 continues the trajectory established by Boltz-1 of making state-of-the-art computational biology tools openly accessible, and its unified structure-plus-affinity framework is likely to influence subsequent model designs in the field.