OpenFold3-preview

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.

Parameters available in OpenFold3

Each parameter set is compatible with a specific range of OpenFold3 releases. See OpenFold3 Parameters in our documentation for checkpoint file names and details on how parameters are resolved at inference time.

OpenBind-0

The default parameters shipped with OpenFold3 >=0.5.0 are OpenBind-0, which is trained using the forthcoming OpenFold3 architecture with a June 2025 training cutoff. See the announcement blog post for benchmarks and details. We reccomend using the OpenBind-0 model over OpenFold3-preview2 for everyday usage.

If you want to keep using the OpenFold3-preview2 weights, install an older release of OpenFold3 - see how do I use an older set of parameters in our documentation.

Technical report

A technical description of the OpenFold3-preview2 model version is available here. Note that the OpenFold3-preview2 weights require an older release of OpenFold3.

Documentation

Please visit our portal and full documentation for instructions on running model training and inference.

Datasets

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.

Features

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

Quick-Start for Inference

Make your first predictions with OpenFold3-preview in a few easy steps:

1. Install OpenFold3-preview using our pip package

pip install openfold3 

2. Setup your installation of OpenFold3-preview and download model parameters:

setup_openfold

3. Run your first prediction using the ColabFold MSA server with the run_openfold binary. 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.json

More 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.

以上为 README 的开头部分。其余内容请阅读原文。 查看完整 README ↗