microsoft/foldingdiff

What it solves

FoldingDiff is designed to generate novel protein backbone structures. It addresses the challenge of creating new protein shapes that can be used for protein design, providing a way to sample diverse and physically plausible protein folds without needing a pre-existing sequence.

How it works

The project implements a diffusion model based on the BERT architecture. It is trained on the CATH dataset of protein structures, specifically learning to generate the internal angles of the protein backbone. During the sampling process, the model reverses the diffusion process to "fold" a structure from noise into a coherent protein backbone. These sampled angles are then converted into Cartesian coordinates to produce PDB files.

Who it’s for

This tool is for researchers in computational biology and protein engineering who want to generate new protein backbones for subsequent inverse folding (designing amino acid sequences that fit the structure) and structural prediction.

Highlights

• Diffusion-based generation: Uses a diffusion model to generate novel protein backbones from noise.
• Integrated evaluation: Includes scripts to calculate TM-scores and evaluate "designability" using inverse folding models like ProteinMPNN and ESM-IF1.
• Visualizations: Provides tools to generate GIFs of the protein folding process using PyMOL.
• Pre-trained weights: Offers weights trained on the CATH dataset available via HuggingFace.