RFdiffusion is an open source Python-based method for protein structure generation developed by Rosetta Commons. The repository implements a diffusion model approach to protein design that can operate with or without conditional information such as motifs or target structures. The method addresses a diverse range of protein design challenges including motif scaffolding, unconditional protein generation, symmetric unconditional generation with support for cyclic, dihedral, and tetrahedral symmetries, symmetric motif scaffolding, binder design, and design diversification through partial diffusion sampling.
The repository is classified across multiple computational biology and machine learning domains including protein design, diffusion models, generative AI, protein structure prediction, deep learning, de novo design, molecular design, protein engineering, structure generation, and computational biology. The codebase is primarily written in Python and leverages NVIDIA's SE(3)-Transformer implementation as a core dependency for equivariant neural network operations on protein structures.
Installation and setup are designed to be accessible to users without extensive computational infrastructure. The repository provides multiple access pathways including a Google Colab notebook for cloud-based execution, an official Docker image maintained by Rosetta Commons, and local installation via conda environments. Local setup is estimated to take less than thirty minutes on standard desktop computers, though users must customize CUDA and PyTorch versions in the provided environment file to match their specific GPU hardware and drivers.
The primary execution interface is through a hydra-based configuration system via the scripts/run_inference.py script. This approach allows users to specify inference parameters through command-line arguments while maintaining sensible defaults derived directly from model checkpoints, ensuring that inference behavior matches training conditions by default. The contig mapping system enables flexible specification of protein design tasks, allowing users to define length ranges for regions to be generated and to anchor specific motifs from input PDB files at designated positions.
Question-type issues represent the most common tracked label category.
The method significantly outperforms earlier approaches like Constrained Hallucination and RFjoint Inpainting for motif scaffolding tasks. Users can control inference behavior through numerous tunable parameters including diffusion trajectory count, temperature settings, and auxiliary potentials. The repository includes example scripts and pre-packaged scaffold files for protein-protein interaction design tasks, along with comprehensive documentation maintained through a dedicated Google Sites resource. Output files are structured to facilitate downstream analysis and validation of generated protein designs.
