AlphaFold 3 in PyTorch, Debugged by the Internet

What it does

This is a from-scratch PyTorch port of Google DeepMind’s AlphaFold 3, the transformer-and-diffusion model that predicts 3D structures of proteins, nucleic acids, and small molecules. The repository exposes both atom-level and higher-level Alphafold3Input APIs for training and sampling, and it ships with scripts to curate and filter the massive PDB dataset required for training.

The interesting bit

The project reads like a public peer-review session: dozens of contributors have patched specific mismatches against the paper, from distogram atom indices to diffusion module inconsistencies. That makes it less a single-author rewrite and more a distributed bug bash against the supplementary materials.

Key highlights

• Full model architecture with pairformer stacks, MSA modules, template embedders, and an elucidated atom diffusion decoder
• Helper classes like Alphafold3Input that abstract away raw tensor construction for standard molecular inputs
• PDB dataset preparation and clustering scripts for assembling training data from raw mmCIF files
• Confidence head, clash penalty ranking, and sample ranking logic for evaluating predictions
• A Gradio frontend interface for interactive use

Caveats

• MSA and template parsing are only at initial support level
• Data preparation demands hundreds of gigabytes of PDB archives and multiple preprocessing stages
• The README notes a long trail of paper-to-code mismatches that required community fixes, so edge-case correctness is still being actively validated

Verdict

Grab this if you need a hackable, PyTorch-native AlphaFold 3 for research or custom training pipelines. Skip it if you want a polished, batteries-included inference service with pre-packaged weights.