Foundry is a central repository and toolkit for training and deploying biomolecular foundation models, with shared infrastructure for protein design, inverse folding, and structure prediction tasks.
Foundry addresses the need for unified tooling to work with multiple classes of protein design models by providing a cohesive platform built on AtomWorks, a framework for manipulating biomolecular structures. The toolkit supports RFdiffusion3 for generative protein design under constraints, ProteinMPNN and LigandMPNN for inverse folding, and RF3 for protein folding. Users can run inference and training through a command-line interface and Python API, with model weights managed through a registry system that defaults to a local checkpoint directory.
The project suits research teams and practitioners working on protein design pipelines who need to combine multiple models in a single workflow. Foundry provides official Docker images with pre-installed weights for reproducible deployments, and supports inference on Apple Silicon via a community fork with MPS acceleration, Intel XPU devices, and standard GPU setups. A Google Colab notebook demonstrates an end-to-end design pipeline. The tool is particularly valuable for those seeking to avoid integrating multiple separate model repositories and instead work within a unified interface.
The project maintains a substantial base of external adopters who report issues and feature requests, with maintainers responding to new issues and pull requests within a day.
