OpenFold3

OpenFold3 is the OpenFold Consortium's fully open-source reproduction of AlphaFold3, extending the consortium's earlier AlphaFold2 reimplementation OpenFold to the all-atom, multi-modal regime. Where the original OpenFold predicted single-chain and homo-oligomeric protein structures, OpenFold3 predicts standard and non-canonical protein, RNA, and DNA chains, small molecules and ligands, and their complexes within a single framework — the same biomolecular scope introduced by AlphaFold3.

The project's central motivation is access. AlphaFold3's weights are released only for limited non-commercial academic use, leaving a gap for industry and independent researchers who need to run, fine-tune, and build on an AlphaFold3-class model. OpenFold3 closes that gap: it is released under the permissive Apache 2.0 license, and the consortium ships the weights, code, and datasets needed to retrain it from scratch — what it describes as the only functionally reproducible AlphaFold3 reconstruction near parity with the original. Its first preview reached that parity on monomeric RNA structures.

OpenFold3 was built by the AlQuraishi Lab at Columbia University together with the OpenFold Consortium, with co-leadership from the Bioresilience Program at Lawrence Livermore National Laboratory and the Steinegger Lab at Seoul National University, plus industry and pharmaceutical members. The preview was released on 28 October 2025, followed by a second prerelease (OpenFold3-preview2) that added updated weights, updated training and inference code, and the full public release of training and assessment data.

#Key Features

• All-atom biomolecular scope: Predicts protein, RNA, and DNA chains (canonical and non-canonical), small molecules, and complexes such as protein-protein and protein-ligand assemblies from a single model.
• Permissive open license: Released under Apache 2.0 for both academic and commercial use, in contrast to AlphaFold3's non-commercial weight release.
• Fully reproducible and trainable: Ships training, inference, and evaluation code together with all datasets needed for training and assessment, making it a functionally reproducible AlphaFold3 that can be retrained from scratch.
• Native nucleic-acid modeling: Bootstraps its own RNA self-distillation set from clustered Rfam families, reaching AlphaFold3 parity on monomeric RNA structures at first preview.
• Accelerated inference: Uses cuEquivariance and DeepSpeed4Science kernels with multi-GPU distributed prediction and memory-constrained inference modes, plus MSA pipelines via the ColabFold server or JackHMMER/hhblits and structure templates for protein monomers.

#Technical Details

OpenFold3 reproduces the input features and architecture described in the AlphaFold3 publication, pairing a representation trunk with a diffusion module that generates raw atom coordinates. Training draws on roughly 245,000 structures and alignments from the RCSB Protein Data Bank, augmented by distillation sets reproduced from the AlphaFold3 recipe: a long-monomer set of about 13 million MGnify sequences (length >= 200 residues), a 400,000-monomer short set (length < 200), an AlphaFold2-predicted set covering unresolved PDB segments, and roughly 125,000 OpenFold3-predicted RNA monomers derived from clustered Rfam. This entire training and assessment corpus is released on the AWS Registry of Open Data under CC BY 4.0. The second preview was trained for 155,000 steps across three stages on 256 NVIDIA H100 GPUs. The team evaluates OpenFold3 on FoldBench (monomeric and multimeric, on data held out relative to the AlphaFold3 training cutoff), CASP16 protein and RNA monomers and protein-protein complexes, the Runs N' Poses protein-ligand benchmark, and AlphaFold3 antibody-antigen sets, against Protenix, Boltz-1, Boltz-2, and Chai-1. Performance is comparable to AlphaFold3 on protein monomers and protein-RNA interfaces and narrows the gap on protein-ligand poses, while AlphaFold3 retains a lead on antibody-antigen complexes and scales better with MSA seed count.

#Applications

OpenFold3 targets structure-based drug discovery, where predicting protein-ligand poses and protein-nucleic acid interactions guides hit identification and lead optimization, and it serves protein engineers modeling multi-chain complexes and designed interfaces. Because it is Apache 2.0 and trainable from scratch on public data, commercial teams can deploy and adapt it without the licensing constraints of AlphaFold3. The model also anchors a federated fine-tuning effort in which pharmaceutical members refine OpenFold3 on proprietary structural data through a secure, privacy-preserving collaboration.

#Impact

OpenFold3 provides the biomolecular-structure community with an AlphaFold3-class model that is genuinely open for both research and industry, extending the OpenFold lineage from AlphaFold2 to all-atom complex prediction and demonstrating that AlphaFold3-level performance is reproducible from public data and open tooling. By releasing weights, code, and the complete training corpus, it lowers the barrier to retraining, ablation, and downstream method development in ways closed releases cannot. Its principal caveats are maturity and documentation status: the current artifacts are preview/prerelease versions documented by a technical report rather than a peer-reviewed paper, only final preview weights are distributed (no intermediate checkpoints) and only through a HuggingFace repository that gates downloads behind a registration form, and RNA and disordered-region prediction remain hard problems where accuracy still trails the best specialist tools.