Boltz is an official repository for biomolecular interaction prediction models, written in Python and designed for drug discovery and protein structure analysis. The project provides two main model versions: Boltz-1, the first fully open source model to approach AlphaFold3 accuracy, and Boltz-2, a newer biomolecular foundation model that jointly models complex structures and binding affinities. Boltz-2 represents a significant advancement by being the first deep learning model to approach the accuracy of physics-based free-energy perturbation methods while executing approximately 1000 times faster, making accurate computational screening practical for early-stage drug discovery applications.
The repository is classified by GitGenius as a CLI development project with command-line tools featuring subcommand support and option flags. It demonstrates a lightweight design philosophy with zero-dependency architecture. The codebase is released under the MIT license, making it freely available for both academic and commercial uses. Installation is straightforward through PyPI or directly from GitHub, with support for both GPU and CPU-only hardware, though GPU versions offer significantly better performance.
The primary inference workflow involves running the boltz predict command on YAML files or directories of YAML files for batched processing. Users can specify biomolecules to model and properties to predict through these input formats. The project provides two distinct binding affinity predictions: affinity_pred_value and affinity_probability_binary. The affinity_probability_binary field, ranging from 0 to 1, represents the predicted probability that a ligand is a binder and is designed for hit-discovery stages to detect binders from decoys. The affinity_pred_value aims to measure specific affinity differences between binders and how these change with molecular modifications, making it suitable for ligand optimization stages such as hit-to-lead and lead-optimization phases. This value reports binding affinity as log10(IC50) derived from IC50 measurements in micromolar units.
The repository shows active maintenance and community engagement.
The project leverages NVIDIA cuEquivariance kernels for acceleration on recent NVIDIA GPUs and also runs on Tenstorrent hardware through a community fork. Authentication to MSA servers is supported through multiple credential provision methods. The repository documentation indicates that evaluation code and training code for Boltz-2 are coming soon, along with evaluation scripts and structural predictions for comparison with other models including Chai-1 and AlphaFold3 on test benchmark datasets. The project actively welcomes external contributions and maintains a Slack channel for community discussion and collaboration around model advancements and applications.
