kyegomez/Open-AF3

What it solves

This project is an open-source PyTorch implementation of AlphaFold 3, designed to predict the 3D structures of biomolecular interactions, including proteins, nucleic acids, and ligands.

How it works

The model replaces the Evoformer with a "Pairformer" (consisting of 48 blocks) that operates on pair and single representations. It uses a Genetic Diffusion module that operates directly on raw atomic coordinates, training the network to denoise noised atomic coordinates to predict true coordinates. To reduce hallucinations, it employs a cross-distillation method using AlphaFold Multimer v2.3 predictions.

Who it’s for

Researchers and developers in computational biology and structural biology who want an open-source alternative to the AlphaFold 3 architecture for biomolecular structure prediction.

Highlights

• Broad Prediction Capabilities: Predicts structures from input polymer sequences, residue modifications, and ligand SMILES.
• Diffusion-Based Generation: Uses a diffusion rollout procedure to generate final atomic positions from random noise.
• Confidence Measures: Regresses atom-level and pairwise errors to predict pLDDT and predicted aligned error (PAE) matrices.
• PyTorch Implementation: Provides a modular implementation of the core AlphaFold 3 architecture in PyTorch.