SiteAF3: Accurate Site-specific Folding via Conditional Diffusion Based on Alphafold3

Introduction

SiteAF3 is a universal molecular interaction analysis platform based on AlphaFold3, supporting structure prediction and analysis for various receptor-ligand type combinations.

SiteAF3 is published on PNAS!

We notice that AlphaFold3 has some updates, which may affect the use of SiteAF3. If you want to use latest version of AlphaFold3, known issues include:

1. Change all the ".cached_ccd()" in ./src/embeddings/embed_cond.py and aptamer/SiteAF3/src/diffusion/run_cond_Diff.py into ".Ccd()".

Or you can use the corresponding AlphaFold3 version:

git clone https://github.com/google-deepmind/alphafold3.git
cd alphafold3
git checkout 7a4a2f7

🚨 v1.1.0 Update Supporting --use_pocket_masked_af3_msa_for_embedding, enhancing binding pocket locating. See the performance in Table 1.

Core Features

🎯 Supported Molecule Types

Receptor Types:

• protein: Protein

Ligand Types:

• protein: Protein
• nucleic: Nucleic acid (DNA/RNA)
• small_molecule: Small molecule ligand / ions

📊 Supported Molecule Combinations

Receptor Type	Ligand Type	Example Usage
protein	nucleic	Protein-RNA/DNA complex
protein	small_molecule	Protein-drug complex
protein	protein	Protein-peptide interaction

🔧 Main Tools

1. Hotspot structure generation (generate_hotspot.py)
2. Pocket structure generation (generate_pocket.py)
3. Inference scripts (run_SiteAF3.py)

Installation

Dependencies

Replace /PATH/TO/alphafold3/src/alphafold3/model/model.py with ./AF3_code/model.py before intalling AlphaFold3.

# Installing AlphaFold3 dependencies
# Please follow the official AlphaFold3 documentation for installation
conda activate your_AF3_env_name
conda env update --file environment.yml

Quick Start

1. Generate hotspot and pocket files

You can use the scripts here, or you can use PyMOL to build your own files.

python generate_hotspot.py \
    --input_pdb /PATH/TO/pdb_file \
    --receptor_type protein \
    --receptor_chains A B (OPTIONAL) \
    --ligand_type nucleic \
    --ligand_chains C (OPTIONAL) \

python generate_pocket.py \
    --input_pdb /PATH/TO/pdb_file \
    --receptor_type protein \
    --receptor_chains (OPTIONAL) \
    --ligand_type small_molecule \
    --ligand_chains (OPTIONAL) \
    --hotspot_cutoff 8.0 \
    --pocket_cutoff 10.0

2. View chain information

python generate_hotspot.py --input_pdb input.pdb --list_chains

3. Structure prediction

Hotspot files are required if enabling --use_hotspot_msa_for_embedding

python run_SiteAF3.py \
    --config_file /PATH/TO/JSON_file \
    --output_dir /PATH/TO/output_dir \
    --use_pocket_msa_for_embedding (OPTIONAL) \
    --use_hotspot_msa_for_embedding (OPTIONAL) \
    --use_af3_msa_for_embedding (OPTIONAL) \
    --use_pocket_masked_af3_msa_for_embedding  (v1.1.0 OPTIONAL) \
    --verbose

SiteAF3 Main Program Configuration

In the JSON configuration file for run_SiteAF3.py, See more JSON cases in test_input folder.

{
    "name": "my_complex",
    "receptor": [{
        "rec_struct_path": "input.pdb",
        "fixed_chain_id": ["A", "B"],
        "hotspot_path": "/PATH/TO/input_hotspot.pdb",
        "pocket_path": "/PATH/TO/input_pocket.pdb"
    }],
    "ligand": [{
        "type": "rna",
        "chain_id": "C",
        "sequence": "..."
    }],
    "modelSeeds": [42, 123, 456]
}

Advanced Features

Intelligent Chain Classification

The program automatically detects chain types based on:

1. Atom check: Search for characteristic atoms (CA atoms→protein, P atoms→nucleic acid)
2. Residue ratio: Count ratio of standard amino acids and nucleic acid residues
3. Small molecule identification: Exclude water molecules and ions, identify ligand molecules
4. Atom count: Small molecules typically have 1-200 heavy atoms
5. Element composition: Molecules containing non-standard elements

Representative Atom Selection

Automatically selects based on molecule type:

• Protein: CA atoms
• Nucleic acid: P atoms or C4' atoms
• Small molecule: Geometric center

Output Files

File Structure

output_dir/
├── seed_{1}
│   ├── confidences.json
│   ├── summary_confidences.cif
│   ├── info.json
│   ├── model.cif
│   ├── ranking_scores.csv
│   ├── sample_{0}
│   │   │── model.cif
│   │   ├── confidences.json
│   │   ├── summary_confidences.cif
│   ├── sample_{1}           
│   └── ......      
├── seed_{2} 
└── ......  

Testing and Validation

Result summary

Prediction accuracy across biomolecular complexes

Troubleshooting

Use the --verbose parameter to get detailed debugging information.

Version 1.1.1 - Bug Fix and Broader CCD Support

8/31/2025

• Fix the bug of baseline model in V1.1.0
• Support CCD when using multiple ligands as input

Version 1.1.0 - Universal Molecule Type Support

7/12/2025

Supporting --use_pocket_masked_af3_msa_for_embedding, enhancing binding pocket locating.

Version 1.0.0 - Universal Molecule Type Support

7/6/2025

Features:

• 🎯 Support for multiple receptor-ligand combinations
• 🔧 Compatible to Alphafold3
• 🧬 Intelligent molecular recognition system
• 📁 Flexible configuration system
• 🛠️ Parallel processing support

Contributing

Issues and Pull Requests are welcome!

License

This project is licensed under the MIT License - see the LICENSE file for details.

Citation

If you use SiteAF3 in your research, please cite:

@article{doi:10.1073/pnas.2521048122,
journal = {Proceedings of the National Academy of Sciences},
doi = {10.1073/pnas.2521048122},
issn = {0027-8424},
number = {44},
publisher = {Proceedings of the National Academy of Sciences},
title = {Accurate site-specific folding via conditional diffusion based on AlphaFold3},
volume = {122},
author = {Tang, Haocheng and Wang, Junmei},
note = {[Online; accessed 2025-11-06]},
date = {2025-10-30},
year = {2025},
month = {10},
day = {30},
}

@software{siteaf3,
  title={SiteAF3: Universal Molecular Interaction Analysis Platform},
  author={Haocheng Tang},
  year={2025},
  url={https://github.com/HaCTang/SiteAF3}
}

---