RosettaCommons/foundry

What it solves

Foundry provides a unified infrastructure for the design, folding, and sequence optimization of proteins and nucleic acids. It streamlines the process of using and training various biomolecular models, reducing the friction associated with managing different model architectures and their specific requirements.

How it works

Foundry acts as a tooling layer that integrates multiple specialized models and relies on AtomWorks for biomolecular structure processing and featurization. It supports several key model classes:

• Generative Design: RFdiffusion3 (RFD3) and RFdiffusion3NA (RFD3NA) create all-atom protein and nucleic acid structures under complex constraints.
• Structure Prediction: RosettaFold3 (RF3) predicts protein structures to provide an open-source alternative to closed-source models.
• Inverse Folding: ProteinMPNN and LigandMPNN design diverse sequences for specific backbones.
• Optimization: RFOptimization combines RF3 gradients, Boltz-2 prediction, and MPNN redesign for sequence optimization.

Who it’s for

It is designed for researchers and developers in computational biology and protein engineering who need a standardized environment to run inference or train models for biomolecular design.

Highlights

• Unified Tooling: A single framework to manage RFD3, RF3, and MPNN variants.
• Broad Hardware Support: Compatible with NVIDIA GPUs, Intel XPU, and Apple Silicon (MPS).
• Integrated Ecosystem: Deeply integrated with AtomWorks for structure I/O and preprocessing.
• Flexible Deployment: Available via pip, Docker images, and Google Colab tutorials.