Boltz

Overview

Boltz is a family of open-source models and a supporting codebase for predicting biomolecular interactions, including complex structures and ligand binding affinities. The project provides two major model releases documented in technical reports: Boltz-1 (first release) and Boltz-2 (later release), with Boltz-2 emphasizing joint modeling of structure and binding affinity. The authors report that Boltz-1 approached AlphaFold3 accuracy, and Boltz-2 advances affinity prediction accuracy toward that of physics-based free-energy perturbation (FEP) methods while running orders of magnitude faster.

Key features

• Structure and affinity prediction: Boltz jointly models complex 3D structures and binding affinity metrics so it can be used both for predicting physical conformations and for virtual screening/lead optimization.
• Two affinity outputs: affinity_probability_binary (binder vs decoy probability, useful for hit discovery) and affinity_pred_value (a continuous estimate reported as log10(IC50) for comparing binders and small modifications during optimization).
• Open-source with weights: All code and pretrained weights are provided under the MIT license for academic and commercial use.
• Performance goals: The project claims structural accuracy approaching AlphaFold3 and affinity predictions approaching FEP accuracy while running ~1000x faster than FEP, making large-scale in silico screening practical.
• GPU acceleration and hardware support: Recommended to run on NVIDIA GPUs (uses NVIDIA cuEquivariance kernels when available); community forks add support for other hardware (e.g., Tenstorrent).

Installation & usage

• Install from PyPI (recommended): pip install boltz[cuda] -U (omit [cuda] for CPU-only installs).
• Install from source for daily updates: clone the repository and pip install -e .[cuda].
• Command-line inference: boltz predict input_path --use_msa_server where input_path is a YAML file or directory describing molecules and requested predictions.
• Prediction outputs include structural data and affinity fields; the README and docs provide detailed prediction and input format instructions.

Training & evaluation

• The repository contains training and evaluation pipelines (Boltz-1 training code is available; updated Boltz-2 training/evaluation code is noted as coming soon in the README).
• The team plans to provide evaluation scripts and prediction sets comparing Boltz-1, Boltz-2, Chai-1, and AlphaFold3 on benchmark datasets, and affinity evaluations on FEP+ and other test sets.

Licensing & citation

• License: MIT — code and weights can be used in academic and commercial projects.
• Citation: The README provides BibTeX entries for Boltz-1 (Wohlwend et al., 2024) and Boltz-2 (Passaro et al., 2025) technical reports hosted on bioRxiv.

Who maintains it

• The repository is maintained by Jeremy Wohlwend (GitHub: jwohlwend) and collaborators; core author lists for the papers include Saro Passaro, Gabriele Corso, Jeremy Wohlwend, Regina Barzilay, Tommi Jaakkola, Noah Getz and others.

When to use

• Boltz is appropriate for researchers and practitioners working on protein–ligand complex modeling, virtual screening, hit-to-lead optimization, and anyone who needs faster approximate alternatives to expensive physics-based affinity calculations for early-stage drug discovery.

Notes & caveats

• The README emphasizes using GPU acceleration for practical speed; CPU-only runs are significantly slower.
• As with any ML-based affinity predictor, users should validate predictions experimentally and be mindful of dataset and domain limitations described in the technical reports.