DMCB-GIST/alphadta

AlphaDTA AF3 Embeddings, IGN Inputs & Interaction Graphs

This repository provides data resources for AlphaDTA, including:

1. 1.Precomputed AlphaDTA reproduction data for existing LP-PDBbind / PDBCleanSplit experiments.
2. 2.PDBbind2020-wide AF3 embeddings and IGN inputs for users who want to define custom splits and generate interaction graphs.

Please read `OUTPUT_TERMS_OF_USE.md` before using this repository.

AlphaFold 3 Output Terms of Use

The AlphaFold 3-derived outputs provided in this repository are subject to the AlphaFold 3 Output Terms of Use.

By using these outputs, you agree to the applicable AlphaFold 3 Output Terms of Use.

Please read:

• —`OUTPUT_TERMS_OF_USE.md`
• —`LEGALLY_BINDING_TERMS_OF_USE.txt`

The AlphaFold 3 outputs distributed in this repository have been processed for use with AlphaDTA. In particular, the released embeddings are pocket-filtered, reordered, sharded, and compressed rather than being distributed as the original AlphaFold 3 output directories.

Commercial use is not allowed under the included AlphaFold 3 Output Terms of Use.

Repository Layout

.
├── precomputed_alphadta/
│   ├── embeddings/
│   ├── csv/
│   └── interaction_graph/
│
├── pdbbind2020/
│   ├── csv/
│   ├── embeddings/
│   └── ign_input/
│       ├── crystal/
│       └── af3/
│
├── README.md
└── OUTPUT_TERMS_OF_USE.md

1. precomputed_alphadta/

This folder contains precomputed files used to reproduce the original AlphaDTA experiments.

precomputed_alphadta/embeddings/

AF3 embeddings stored as sharded, compressed archives (emb-*.tar.zst) per subset:

• —lp_only/ — PDB IDs only in LP-PDBbind.
• —shared/ — PDB IDs shared between LP-PDBbind and PDBCleanSplit.
• —pdbcleansplit_only/ — PDB IDs only in PDBCleanSplit.
• —bdb2020+/
• —egfr/
• —mpro/
• —casf2016/

Each subset folder contains:

emb-*.tar.zst
metadata/shards_pdbids.json

metadata/shards_pdbids.json maps each shard filename to the PDB IDs contained in that shard.

precomputed_alphadta/csv/

CSV files describing the predefined splits and benchmarks used in AlphaDTA experiments.

precomputed_alphadta/csv/
├── cleansplit/
│   ├── casf2016.csv
│   └── train-validation.csv
└── lp_pdbbind/
    ├── train.csv
    ├── valid.csv
    ├── test.csv
    ├── bdb2020+.csv
    ├── egfr.csv
    └── mpro.csv

Each CSV contains:

pdbid
protein_length
ligand_length
total_length
pK

precomputed_alphadta/interaction_graph/

Precomputed interaction graphs used for AlphaDTA reproduction.

These graphs correspond to the predefined LP-PDBbind / PDBCleanSplit splits and benchmark sets.

2. pdbbind2020/

This folder is intended for users who want to create their own train/validation/test splits from PDBbind2020.

Unlike precomputed_alphadta/, this folder does not provide one fixed graph split. Instead, it provides PDBbind2020-wide inputs that can be regrouped into user-defined splits.

pdbbind2020/
├── csv/
│   └── pdbbind2020.csv
├── embeddings/
│   ├── emb-000000.tar.zst
│   ├── emb-000001.tar.zst
│   └── metadata/
│       └── shards_pdbids.json
└── ign_input/
    ├── crystal/
    │   ├── ign-crystal-000000.tar.zst
    │   └── metadata/
    │       └── shards_pdbids.json
    └── af3/
        ├── ign-af3-000000.tar.zst
        └── metadata/
            └── shards_pdbids.json

pdbbind2020/csv/pdbbind2020.csv

This CSV contains one row per PDB ID available in the PDBbind2020 embedding release.

Columns:

pdbid
protein_length
ligand_length
total_length
pK

pdbbind2020/embeddings/

This folder contains AF3 embedding archives for PDBbind2020 entries.

pdbbind2020/embeddings/
├── emb-000000.tar.zst
├── emb-000001.tar.zst
└── metadata/
    └── shards_pdbids.json

metadata/shards_pdbids.json maps each embedding shard to the PDB IDs contained in that shard.

pdbbind2020/ign_input/

This folder contains IGN preprocessing inputs for PDBbind2020.

Two structure sources are provided:

pdbbind2020/ign_input/crystal/

IGN inputs generated from experimentally determined PDBbind crystal structures.

pdbbind2020/ign_input/af3/

IGN inputs generated from AlphaFold 3 predicted complex structures.

Each folder contains:

*.tar.zst
metadata/shards_pdbids.json

Generating Custom Interaction Graphs from pdbbind2020/ign_input/

The released pdbbind2020/ign_input/ files are intended for users who want to define custom PDBbind2020 splits and generate interaction graphs without rerunning the full AF3/CIF preprocessing pipeline.

Two IGN input sources are provided:

pdbbind2020/ign_input/crystal/
pdbbind2020/ign_input/af3/

• —crystal/: IGN inputs generated from experimentally determined PDBbind crystal structures.
• —af3/: IGN inputs generated from AlphaFold 3 predicted complex structures.

A typical workflow is:

1. 1.Download and extract the desired IGN input archive.

For example:

   mkdir -p /path/to/extracted_ign_af3
   tar -I zstd -xf pdbbind2020/ign_input/af3/ign-af3-000000.tar.zst \
       -C /path/to/extracted_ign_af3

1. 1.Create a custom split CSV.

The CSV should contain at least:

   pdbid
   pK

For example:

   pdbid,pK
   1a1e,7.20
   2xyz,6.45

1. 1.Create a dataset directory for the custom split.

   mkdir -p /path/to/custom_dataset/processed_structure/ign_input

1. 1.Copy or symlink the selected IGN input entries into:

   /path/to/custom_dataset/processed_structure/ign_input/

You can create custom_split.csv by selecting rows from:

   pdbbind2020/csv/pdbbind2020.csv

For example, if split_pdbids.txt contains one PDB ID per line:

   while read pdbid; do
       if [ -e "/path/to/extracted_ign_af3/${pdbid}" ]; then
           cp -a "/path/to/extracted_ign_af3/${pdbid}" \
                 /path/to/custom_dataset/processed_structure/ign_input/
       elif [ -e "/path/to/extracted_ign_af3/${pdbid}_ligand" ]; then
           cp -a "/path/to/extracted_ign_af3/${pdbid}_ligand" \
                 /path/to/custom_dataset/processed_structure/ign_input/
       else
           echo "Missing IGN input: ${pdbid}"
       fi
   done < split_pdbids.txt

The preprocessing code supports IGN input entries as either files or directories, as long as their names match the PDB IDs.

1. 1.Generate interaction graphs directly from the prepared IGN inputs:

   python preprocess/preprocess_structure.py \
       --dataset_dir /path/to/custom_dataset \
       --label_csv /path/to/custom_split.csv \
       --start_from ign_input \
       --num_process 12 \
       --verbose

The generated graphs will be written to:

/path/to/custom_dataset/processed_structure/graph_ls
/path/to/custom_dataset/processed_structure/graph_dic

This workflow allows users to create multiple custom train/validation/test splits from the same released PDBbind2020 IGN inputs. Users only need to regroup or symlink the desired PDB IDs and rerun graph generation with --start_from ign_input.

For the preprocessing code, see:

https://github.com/DMCB-GIST/AlphaDTA

The relevant script is:

preprocess/preprocess_structure.py

Important: Pocket-filtered AF3 Embeddings

The provided AF3 embeddings are not full-sequence embeddings.

Instead, AlphaDTA uses pocket-related tokens to align the embedding input with interaction-graph inputs and focus on binding-relevant regions.

Token ordering and selection:

• —AlphaFold 3 embeddings follow the input token order.
• —The AlphaDTA input convention is:

• —protein sequence first,
• —ligand SMILES second.
• —The original AF3 embedding token order is therefore:

• —[protein residues; ligand heavy atoms].
• —Pocket residues are selected as protein residues whose Cα atoms are within 10 Å of any ligand heavy atom.
• —The retained tokens are:

• —all selected pocket residues,
• —all ligand heavy atoms.
• —The final embedding token order is:

• —[pocket residues; ligand heavy atoms].

AF3 Embedding Format (.pt)

Each {pdbid}.pt file contains at least:

single: shape (T, 384)
pair:   shape (T, T, 128)

where T = # pocket residues + # ligand heavy atoms.

Compression & Sharding Strategy

Files are grouped into size-based shards and compressed as tar.zst.

• —Target shard size: approximately 20 GB per shard.
• —Actual shard sizes may vary.
• —Metadata files map shard names to the PDB IDs contained in each shard.

Example compression pipeline:

tar -C <SOURCE_DIR> -T <LIST_FILE> -cf - | zstd -3 -T<NUM_THREADS> -o <OUTPUT>.tar.zst

Citations

If you use this repository in academic work, please cite the relevant papers:

@article{abramson2024accurate,
  title={Accurate structure prediction of biomolecular interactions with AlphaFold 3},
  author={Abramson, Josh and Adler, Jonas and Dunger, Jack and Evans, Richard and Green, Tim and Pritzel, Alexander and Ronneberger, Olaf and Willmore, Lindsay and Ballard, Andrew J and Bambrick, Joshua and others},
  journal={Nature},
  volume={630},
  number={8016},
  pages={493--500},
  year={2024},
  publisher={Nature Publishing Group UK London}
}

@article{wang2004pdbbind,
  title={The PDBbind database: Collection of binding affinities for protein-ligand complexes with known three-dimensional structures},
  author={Wang, Renxiao and Fang, Xueliang and Lu, Yipin and Wang, Shaomeng},
  journal={Journal of medicinal chemistry},
  volume={47},
  number={12},
  pages={2977--2980},
  year={2004},
  publisher={ACS Publications}
}

@article{li2026leak,
  title={Leak Proof PDBBind: A Reorganized Data Set of Protein--Ligand Complexes for More Generalizable Binding Affinity Prediction},
  author={Li, Jie and Guan, Xingyi and Zhang, Oufan and Sun, Kunyang and Wang, Yingze and Bagni, Dorian and Head-Gordon, Teresa},
  journal={The Journal of Physical Chemistry B},
  volume={130},
  number={2},
  pages={730--740},
  year={2026},
  publisher={ACS Publications}
}

@article{graber2025resolving,
  title={Resolving data bias improves generalization in binding affinity prediction},
  author={Graber, David and Stockinger, Peter and Meyer, Fabian and Mishra, Siddhartha and Horn, Claus and Buller, Rebecca},
  journal={Nature Machine Intelligence},
  pages={1--13},
  year={2025},
  publisher={Nature Publishing Group UK London}
}