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, removing the need to manage multiple disparate tools for different stages of the protein design pipeline.

How it works

Foundry acts as a central hub for several specialized AI models:

• RFdiffusion3 (RFD3): An all-atom generative model for designing protein structures under complex constraints.
• RFdiffusion3NA (RFD3NA): An extension of RFD3 that adds the ability to design nucleic acid structures.
• RosettaFold3 (RF3): A structure prediction neural network used for protein folding.
• ProteinMPNN / LigandMPNN: Lightweight inverse-folding models used to design diverse sequences for specific protein backbones.

All these models rely on AtomWorks, a framework for manipulating and processing biomolecular structures, to handle training and inference data.

Who it’s for

This project is intended for researchers and developers in computational biology and protein engineering who need to design new proteins or predict their structures.

Highlights

• Unified Tooling: Combines generative design, structure prediction, and sequence design into one package.
• Broad Hardware Support: Compatible with NVIDIA GPUs, Intel XPU, and Apple Silicon (MPS).
• ** idée a Docker Image**: Provides official Docker images for easier deployment and weights management.
• Extensible Architecture: Allows developers to add new models as independent packages within the system.