aqlaboratory/openfold-3

A fully open source biomolecular structure prediction model based on AlphaFold3

What it solves

OpenFold3-preview provides an open-source, bitwise reproduction of DeepMind's AlphaFold3. It allows researchers and developers to predict the 3D structures of biomolecules—including proteins, RNA, DNA, and small molecules—without relying on proprietary software, while offering a framework for long-term maintenance and community-driven development.

How it works

The model replicates the input features and architecture of AlphaFold3. It utilizes specialized kernels for acceleration via cuEquivariance and DeepSpeed4Science to improve inference efficiency. The system supports Multiple Sequence Alignment (MSA) generation through the ColabFold server or standard tools like JackHMMER and hhblits, and it can distribute predictions across multiple GPUs to handle large-scale jobs.

Who it’s for

It is designed for computational biologists, biochemists, and AI researchers who need to predict biomolecular structures for academic or commercial use and want an open-source alternative to AlphaFold3.

Highlights

  • Broad Modality Support: Predicts structures for standard and non-canonical proteins, RNA, DNA chains, and small molecules.
  • High Performance: Matches AlphaFold3's performance on monomeric RNA structures and performs competitively on protein-protein and protein-ligand complexes.
  • Hardware Optimization: Includes kernel acceleration and a low-memory mode for constrained GPU resources.
  • Scalable Inference: Supports multi-query jobs and distributed predictions across multiple GPUs.

Related

  • Project
    bytedance/Protenix

    Protenix is an open‑source deep‑learning system for high‑accuracy protein (and RNA‑protein) structure prediction. It builds on AlphaFold 3, offers several model sizes (including a fast Mini version), supports MSA/template/RNA inputs, and adds constraint‑aware folding. The package is installable via pip, includes a CLI, and a public web server. Benchmarks claim it outperforms AlphaFold 3 on antibody‑antigen complexes. All code is Apache 2.0; only the newest v2 weights are proprietary.

  • Project
    kyegomez/Open-AF3

    An open-source PyTorch implementation of AlphaFold 3 for predicting the 3D structures of biomolecular interactions, including proteins, nucleic acids, and ligands.

  • Project
    google-deepmind/alphafold3

    AlphaFold 3 is an AI system for predicting the 3D structures of biomolecular interactions, helping researchers model how proteins and other molecules interact.

  • Project
    sokrypton/ColabFold

    A tool that makes protein folding predictions accessible via Google Colab and local installations, integrating multiple models like AlphaFold2 and AlphaFold3 with a public MSA server.

  • Project
    lucidrains/alphafold3-pytorch

    A PyTorch implementation of AlphaFold 3 for predicting the 3D structures of biomolecular interactions, including comprehensive tools for PDB dataset curation.