This repository extends OpenFold3-preview with local, GPU-selectable web workflows for protein-complex visualization, antibody–antigen structure prediction, and antibody affinity maturation:
- OpenFold3 Studio accepts multiple protein chains and colors every chain distinctly in an interactive 3Dmol.js structure viewer.
- Antibody–Antigen Studio accepts light-chain, heavy-chain, and antigen sequences, applies ANARCII IMGT numbering, and highlights CDRs in both the sequence and predicted structure.
- Affinity Studio predicts independent single-substitution binding ΔΔG values and displays one selectable substitution heatmap for each of CDRH1–3 and CDRL1–3.
The standalone affinity model fuses pinned ESM-2 650M sequence embeddings with mutation-centered Gudhi/R-TDA topology and a frozen PyTorch conversion of the original topology CNN. Its deterministic S641 rebuild reached Pearson 0.9099 in fixed ten-fold validation and 0.8375 in leave-one-complex-out validation. These are research-reproduction results rather than a leakage-free benchmark, because the archived CNN training set is unavailable. See the Affinity Studio guide for setup, validation, licensing, and interpretation details.
OpenFold3-preview is a biomolecular structure prediction model aiming to be a bitwise reproduction of DeepMind's AlphaFold3, developed by the AlQuraishi Lab at Columbia University and the OpenFold consortium. This research preview is intended to gather community feedback and allow developers to start building on top of the OpenFold ecosystem. The OpenFold project is committed to long-term maintenance and open source support, and our repository is freely available for academic and commercial use under the Apache 2.0 license.
For our reproduction of AlphaFold2, please refer to the original OpenFold repository.
A technical description of our most recent model version, OpenFold3-preview2, is available here.
Please visit our portal and full documentation for instructions on running model training and inference.
We provide the full training data for reproducing OpenFold3-preview, including our reproduction of the MGnify-based 13M-sequence distillation dataset described in the AlphaFold3 paper. For more details about dataset access, please refer to our portal.
OpenFold3-preview replicates the input features described in the AlphaFold3 publication, as well as batch job support and efficient kernel-accelerated inference.
A summary of our supported features includes:
- Structure prediction of standard and non-canonical protein, RNA, and DNA chains, and small molecules
- Pipelines for generating MSAs using the ColabFold server or using JackHMMER / hhblits following the AlphaFold3 protocol
- Structure templates for protein monomers
- Kernel acceleration through cuEquivariance and DeepSpeed4Science kernels - more details here
- Support for multi-query jobs with distributed predictions across multiple GPUs
- Custom settings for memory constrained GPU resources
- Training data processing and model training
Make your first predictions with OpenFold3-preview in a few easy steps:
- Install OpenFold3-preview using our pip package
pip install openfold3 - Setup your installation of OpenFold3-preview and download model parameters:
setup_openfold- Run your first prediction using the ColabFold MSA server with the
run_openfoldbinary. It may also be necessary to configure your environment variables depending on your system.
run_openfold predict --query_json=examples/example_inference_inputs/query_ubiquitin.jsonMore information on how to customize your inference prediction can be found at our documentation home at https://openfold-3.readthedocs.io/en/latest/. More examples for inputs and outputs can be found in our HuggingFace examples.
OpenFold3 Studio provides a local web interface for multi-chain protein prediction and interactive mmCIF visualization. Install the optional web dependencies and start the server:
pip install -e '.[web]'
openfold3-webOpen http://127.0.0.1:8000, add one sequence row for each protein chain, and select Predict structure. GPU predictions are serialized, their live logs are shown in the browser, and completed models are colored by chain by default. The viewer can also color by pLDDT confidence, switch representations, select among diffusion samples, and download the mmCIF and confidence metrics.
Prediction inputs, logs, and results are saved under
openfold3_web_outputs/. Use openfold3-web --help to select another output
directory, host, or port. See the web interface guide
for details.
A separate antibody interface accepts one light chain, one heavy chain, and one protein antigen. It uses IMGT numbering to highlight CDRs in both the sequence and predicted structure, and lets you choose the physical GPU for every job. Install it in Python 3.11 or newer and start it on port 8001:
pip install -e '.[antibody-web]'
openfold3-antibody-webOpen http://127.0.0.1:8001, select a GPU, or load the bundled 1VFB antibody–lysozyme example. See the antibody interface guide for IMGT, GPU-queue, visualization, and output details.
Affinity Studio is a third, independent app for structure-based single-mutation ΔΔG prediction and IMGT-guided antibody maturation. It combines VMD/Jackal structure preparation, mutation-centered Gudhi H0 and R/TDA H1/H2 persistence features, the frozen original topology CNN, and the pinned ESM-2 650M sequence encoder. It ranks all 19 substitutions at every CDR residue and keeps each chain visually distinct with the BSD-licensed open-source 3Dmol.js viewer. Create its isolated environment and start it on port 8002:
./scripts/setup_affinity_environment.sh
conda activate openfold3-affinity
openfold3-affinity setup-tools \
--vmd-archive /path/to/vmd-2.0.0.bin.LINUXAMD64.tar.gz \
--jackal-dir /path/to/jackal_64bit
openfold3-affinity doctor
openfold3-affinity-web --host 127.0.0.1 --port 8002The setup script verifies and installs the trained S641 model bundled at
affinity_model_artifacts/openfold3-affinity-s641-esm2-gudhi-revision-cnn-v2.tar.gz.
Its SHA-256 is
e542542d2e5f2e67406bc4256367f65ddd791075c80c4dd8d99f007303347e3e.
The archive contains the fitted regressor, frozen topology encoder,
normalization data, metrics, and provenance, so users do not need the M3 or
seq2topo reference repositories.
If openfold3_affinity_outputs/egfr_antibody_maturation/ contains the completed
example report, the page opens that result automatically as a read-only demo;
no inference job or GPU allocation is started. The validated reference report
contains all 874 independent CDR substitutions with no failed results. Eleven
have predicted ΔΔG below zero; the lowest is L:G30L at IMGT position 36 in
CDRL1 (-0.456 kcal/mol). These predictions are computational hypotheses and
require experimental confirmation.
VMD, Jackal, and ESM-2 weights remain user-supplied and are not stored in Git. The pinned ESM-2 checkpoint is downloaded to the PyTorch cache on first use and verified before inference. See the Affinity Studio guide for tool setup, model installation, CLI examples, training, licensing, and safe server shutdown.
OpenFold3-preview-2 results:
Performance of OF3p2 and other models on a diverse set of benchmarks:
For more details on inferences procedures and benchmarking methods, please refer to the OpenFold3 Preview 2 whitepaper
OpenFold3-preview results:
OpenFold3-preview performs competitively with the state of the art in open source biomolecular structure prediction, while being the only model to match AlphaFold3's performance on monomeric RNA structures.
Performance of OF3p and other models on a diverse set of benchmarks:
- Protein and RNA monomers from CASP16 and Ludaic et al
- Protein-protein complexes from CASP16 and FoldBench
- Protein-ligand complexes from the Runs and Poses
For more details on inferences procedures and benchmarking methods, please refer to the OpenFold3 Preview whitepaper.
The final OpenFold3 model is still in development, and we are actively working on the following features:
- Full parity on all modalities with AlphaFold3
- Workflows for training on custom non-PDB data
If you encounter problems using OpenFold3-preview, feel free to create an issue! We also welcome pull requests from the community.
In addition, we offer a public Slack channel for discussions and questions around OpenFold3.
If you use OpenFold3-preview in your research, please cite the following:
@software{openfold3-preview,
title = {OpenFold3-preview},
author = {{The OpenFold3 Team}},
year = {2025},
version = {0.4.2},
doi = {10.5281/zenodo.19001000},
url = {https://github.com/aqlaboratory/openfold-3},
abstract = {OpenFold3-preview is a biomolecular structure prediction model aiming to be a bitwise reproduction of DeepMind's AlphaFold3, developed by the AlQuraishi Lab at Columbia University and the OpenFold consortium.}
}
Any work that cites OpenFold3-preview should also cite AlphaFold3.