karina-zadorozhny/ATOMICA

Dataset Card: Atomica Molecular Interactions | Sequence Data

Summary

A dataset of atomic‑scale molecular interaction interfaces, ready for ML workflows. Each row contains up to five interacting sequences (or SMILES), their modalities, an example ID, and the original data split.

Citation

@article{
  Fang2025ATOMICA,
  author = {Fang, Ada and Zhang, Zaixi and Zhou, Andrew and Zitnik, Marinka},
  title = {ATOMICA: Learning Universal Representations of Intermolecular Interactions},
  year = {2025},
  journal = {bioRxiv},
  doi = {10.1101/2025.04.02.646906},
}

Raw files from ATOMICA Havard Dataverse

Available in: karina-zadorozhny/ATOMICA/raw/

Interaction Type	Count
small_molecule-small_molecule	1,767,710 (86% self-pairs)
protein-protein	124,541
protein-small_molecule	119,017
protein-ion	74,514
protein-peptide	8,475
rna-small_molecule	5,185
protein-rna	3,511
protein-dna	2,750

Retrieved using:

IDENTIFIERS = {
    "protein-peptide": "11033993",
    "protein-rna": "11033983",
    "protein-protein": "11033984",
    "rna-small_molecule": "11033985",
    "protein-small_molecule": "11033996",
    "protein-ion": "11033989",
    "protein-dna": "11033982",
    "small_molecule-small_molecule": "11033997",
}

for fname, identifier in IDENTIFIERS.items():
    pooch.retrieve(
                f"https://dataverse.harvard.edu/api/access/datafile/{identifier}",
                fname=f"{fname}.csv",
                known_hash=None,
                path=".",
                progressbar=True,
            )

Processed Files

Imporant note:

• Only contains sequences! This processed version does not contain any structural information unlike the original ATOMICA dataset.
• Only contains a subset of avaialable interactions
• See extra filtering steps applied to each interaction modality

Interaction Type	Count
small_molecule-small_molecule	246,728
protein-protein	60,624
protein-small_molecule	32,120
protein-dna	2,087
protein-rna	1,785

The data is stored as Parquet files partitioned by:

• split: original data split (e.g. train, validation, test)

Each file contains columns:

• id – unique interaction identifier
• sequence1…sequence5 – up to five extracted sequences or SMILES (strings, null if unused)
• modality1…modality5 – corresponding modalities (amino_acid, nucleotide, smiles)

| interaction_type                  | sequence1                                                    | modality1    | sequence2                                                   | modality2   | sequence3                                                   | modality3   | sequence4 | modality4 | sequence5 | modality5 | split |
|-----------------------------------|--------------------------------------------------------------|--------------|-------------------------------------------------------------|-------------|-------------------------------------------------------------|-------------|-----------|-----------|-----------|-----------|-------|
| protein-dna                       | MAGVKNSIIWFRKGLRLHDNPALLEACKDAKHVYPVFVLDPHFLQQ...           | amino_acid   | CAGCGGTTGCCGTG                                             | nucleotide  | CACGGCAACCGCTG                                             | nucleotide  | None      | None      | None      | None      | test  |
| protein-protein                   | GPLGSPEFGRPGWVIGVNPDIGGAIAVLSPDGSSQVFDNPFVHIVV...           | amino_acid   | GPLGSPEFGRPGWVIGVNPDIGGAIAVLSPDGSSQVFDNPFVHIVV...           | amino_acid  | None                                                        | None        | None      | None      | None      | None      | test  |
| protein-rna                       | MHHHHHHENLYFQGSGMAGSVGLALCGQTLVVRGGSRFLATSIASS...           | amino_acid   | MAAETRNVAGAEAPPPQKRYYRQRAHSNPMADHTLRYPVKPEEMDW...           | amino_acid  | GCCCGGAUAGCUCAGUCGGUAGAGCAUCAGACUUUUAAUCUGAGGG...           | nucleotide  | None      | None      | None      | None      | test  |
| protein-small_molecule            | MTSNNLPTVLESIVEGRRGHLEEIRARIAHVDVDALPKSTRSLFDS...           | amino_acid   | O=C(O)c1ccccc1NC[C@@H](O)[C@H](O)[C@H](O)COP(=...           | smiles      | None                                                        | None        | None      | None      | None      | None      | test  |
| small_molecule-small_molecule     | c1ccccc1                                                    | smiles       | Cc1ccc([NH+]=P(C)(C)c2ccccc2)c(c1)S(=O)(=O)[O-]             | smiles      | None                                                        | None        | None      | None      | None      | None      | test  |

Processing Overview

1. Raw data retrieval

   • Download CSVs per interaction type from Harvard Dataverse via file IDs.

2. Sequence & modality extraction

   • Protein–nucleotide / protein–RNA
     • Fetch all molecule entries for each PDB ID from PDBe REST API.
     • Extract sequence and molecule_type.
   • Protein–protein
     • Identify the two chains in each interface ID from id • Fetch PDB entries and select sequences for those chains.
   • Protein–small molecule / protein–ion
     • Extract protein sequences as above.
     • Fetch ligand SMILES from PDBe API, canonicalize via RDKit.
   • Small molecule–small molecule
     • Split the raw ID string into two SMILES, drop self‑pairs (since we're only using sequence information)

3. Data cleaning & formatting

   • Pad or truncate to exactly five sequence/modality slots.
   • Map raw molecule types to {amino_acid, nucleotide, smiles}.
   • Random train,test,val split stratified on interaction type.