OneScience-Group/AlphaFold3

<p align="center"> <strong> <span style="font-size: 30px;">AlphaFold3</span> </strong> </p>

Model Introduction

AlphaFold 3 is a biomolecular structure prediction model developed by Google DeepMind and Isomorphic Labs. It predicts the three-dimensional structures and interactions of proteins, DNA, RNA, small-molecule ligands, and their complexes.

Paper: Accurate structure prediction of biomolecular interactions with AlphaFold 3 https://www.nature.com/articles/s41586-024-07487-w

Model Description

AlphaFold 3 uses a Pairformer and a diffusion model to predict biomolecular complex structures. This model package provides a JAX / Flax inference implementation and database-search scripts, together with the accompanying Hugging Face dataset OneScience-Group/AlphaFold3_dataset.

Use Cases

Usage Instructions

1. OneCode

Try one-click AI4S development in the OneCode online environment:

Try one-click AI4S development

2. Manual Installation

Hardware Requirements

• —GPU or DCU is recommended.
• —A CPU can be used for import checks and lightweight configuration tests; full training and inference will be slow.
• —DCU users must install DTK in advance. DTK 25.04.2 or later is recommended, or a OneScience-recommended version matching the current cluster.

Environment Check

• —NVIDIA GPU:

nvidia-smi

• —Hygon DCU:

hy-smi

Download the Model Package

hf download OneScience-Group/AlphaFold3 --local-dir ./AlphaFold3
cd AlphaFold3

Install the Runtime Environment

DCU Environment

# Activate DTK and CONDA first
conda create -n onescience311 python=3.11 -y
conda activate onescience311
# uv installation supported
pip install onescience[bio-dcu] -i http://mirrors.onescience.ai:3141/pypi/simple/  --trusted-host mirrors.onescience.ai

After installation, return to the model package directory:

cd ./AlphaFold3

If the AlphaFold 3 C++ extension and runtime data files have not yet been built in the current environment, run:

python -m onescience.flax_model.alphafold3.build_extension
python -m onescience.flax_models.alphafold3.build_data

Training and Inference Data Overview

The OneScience community has uploaded the complete data required for AlphaFold 3 inference and database searches to Hugging Face: OneScience-Group/AlphaFold3_dataset. This model package does not include a training entry point; the dataset is primarily used to construct MSA / template features and perform database searches before inference.

hf download OneScience-Group/AlphaFold3_dataset \
  --repo-type dataset \
  --local-dir ./data/alphafold3

Model Weights

Model weights will be available soon.

Preparing Weights

Place the AlphaFold 3 model weights in the following directory, or specify them via environment variables:

weight/
  AlphaFold3/
    ...

Default lookup order:

• —ALPHAFOLD3_MODEL_DIR
• —${ONESCIENCE_MODELS_DIR}/AlphaFold3
• —weight/AlphaFold3

Example:

export ALPHAFOLD3_MODEL_DIR=/path/to/AlphaFold3

Direct Inference

When the input JSON already contains features such as MSA and templates, you can run inference directly:

bash scripts/infer.sh

Equivalent Python command example:

python scripts/run_alphafold.py \
  --json_path inputs/7r6r_data.json \
  --model_dir weight/AlphaFold3 \
  --output_dir outputs \
  --run_data_pipeline=false \
  --flash_attention_implementation=triton

Output is written to outputs/ and includes the top-ranked structure, structures generated from different seed / sample combinations, a CSV file containing ranking scores, and a copy of the input JSON.

Jackhmmer / Nhmmer Database Search

When the input JSON contains only sequences and requires a local database search, use:

bash scripts/infer_jackhmmer.sh

Common environment variables:

export ALPHAFOLD3_DATASET_ROOT=/path/to/alphafold3
export ALPHAFOLD3_MODEL_DIR=/path/to/AlphaFold3
export ALPHAFOLD3_JSON_PATH=inputs/t1119_search.json
export ALPHAFOLD3_OUTPUT_DIR=outputs
export ALPHAFOLD3_RUN_INFERENCE=false

ALPHAFOLD3_DATASET_ROOT is expected to contain directories such as public_databases/, jackhmmer_split/, and mmseqsDB/.

MMseqs Database Search

If the runtime environment provides the MMseqs executable and the required MMseqs databases, use:

bash scripts/infer_mmseqs.sh

Common environment variables:

export ALPHAFOLD3_MMSEQS_HOME=/path/to/mmseqs
export ALPHAFOLD3_DATASET_ROOT=/path/to/alphafold3
export ALPHAFOLD3_MMSEQS_DB_DIR=/path/to/alphafold3/mmseqsDB
export ALPHAFOLD3_RUN_INFERENCE=false

To proceed with inference after the database search, set ALPHAFOLD3_RUN_INFERENCE to true and ensure that the weights directory is available.

Data Format

AlphaFold 3 inputs are provided in JSON format with the following basic structure:

{
  "dialect": "alphafold3",
  "version": 1,
  "name": "example",
  "sequences": [
    {
      "protein": {
        "id": "A",
        "sequence": "..."
      }
    }
  ],
  "modelSeeds": [100],
  "bondedAtomPairs": null,
  "userCCD": null
}

This repository provides two examples:

• —inputs/7r6r_data.json: contains sequence, MSA, and template information; suitable for direct inference.
• —inputs/t1119_search.json: contains only sequences; suitable for database-search pipeline validation.

It is recommended to download the full Hugging Face dataset OneScience-Group/AlphaFold3_dataset to data/alphafold3/ under the model package. The relative directory structure expected by the database-search pipeline is as follows:

data/
  alphafold3/
    public_databases/
      mmcif_files/
      pdb_seqres_2022_09_28.fasta
      ...
    jackhmmer_split/
      bfd-first_non_consensus_sequences.fasta@64
      mgy_clusters_2022_05.fa@512
      uniprot_cluster_annot_2021_04.fa@256
      uniref90_2022_05.fa@128
    mmseqsDB/
      small_bfd_db
      mgnify_db
      uniprot_cluster_annot_db
      uniref90_db

Verification

Static import check:

python tests/check_import_boundaries.py

OneScience Official Information

Citation & License

• —This repository is adapted from the open-source AlphaFold 3 model to support DCUs.
• —The AlphaFold 3 source code is licensed under CC BY-NC-SA 4.0; model parameters are subject to separate usage terms.
• —For scientific use, please cite the original paper: Accurate structure prediction of biomolecular interactions with AlphaFold 3.