Foundry is a central repository maintained by the Rosetta Commons that provides unified tooling and infrastructure for training and deploying biomolecular foundation models focused on protein design and structure prediction. The repository serves as a shared platform for multiple complementary models including RFdiffusion3 (RFD3) for generative protein design, RosettaFold3 (RF3) for structure prediction, and ProteinMPNN/LigandMPNN for inverse folding tasks. All models within Foundry rely on AtomWorks, a unified framework for manipulating and processing biomolecular structures, ensuring consistency across training and inference pipelines.
The repository is classified across multiple domains including computational biology, bioinformatics, workflow management, data sharing, tool platforms, reproducibility, containerization, and cloud computing. This broad classification reflects Foundry's role as both a scientific research platform and a production-ready tool infrastructure. The codebase is written primarily in Python and maintains a documented homepage at rosettacommons.github.io/foundry.
RFdiffusion3 is an all-atom generative model capable of designing protein structures under complex constraints, with an extension called RFdiffusion3NA that adds nucleic acid design capabilities. RosettaFold3 functions as a structure prediction neural network that bridges the gap between closed-source alternatives and open-source tools, supporting prediction of protein-DNA complexes and other biomolecular structures. ProteinMPNN and LigandMPNN are lightweight inverse-folding models designed for sequence design on fixed backbones under constrained conditions.
The repository demonstrates active maintenance and community engagement. GitGenius tracking shows a median issue and pull request response latency of 6.7 hours across 179 items, with a mean latency of 64.8 hours. The most frequently discussed topics center on RFdiffusion3 (56 labeled issues), general questions (58 labeled issues), and bug reports (50 labeled issues). The core development team includes rclune with 252 tracked events, Ubiquinone-dot with 136 events, and r-krishna with 54 events. The repository shares overlapping contributors with related projects including google-deepmind/alphafold3, jwohlwend/boltz, and rosettacommons/rfdiffusion, indicating integration within a broader ecosystem of protein modeling tools.
Foundry provides multiple installation pathways including standard pip installation, specialized Intel XPU support, and Apple Silicon MPS acceleration for macOS users. The repository includes an official Docker image maintained by Rosetta Commons, available both as a full image with pre-installed model weights and as a slim variant for custom configurations. Users can download model checkpoints through a command-line interface with support for base models, all available models, or individual model selection. The repository includes interactive tutorials via Google Colab and comprehensive Jupyter notebook examples demonstrating end-to-end protein design pipelines.
The codebase maintains strict dependency flow from foundry to atomworks, with a modular architecture allowing individual models to be developed and installed independently. New models can be added as separate packages within the models directory with their own pyproject.toml files. The repository enforces code formatting standards through pre-commit hooks running ruff format, ensuring consistency across contributions. Development setup supports editable installations of both foundry and individual models, enabling simultaneous work on shared utilities and model-specific implementations.
