Learning Universal Representations of Intermolecular Interactions

Authors

• Ada Fang
• Michael Desgagné
• Zaixi Zhang
• Andrew Zhou
• Joseph Loscalzo
• Bradley L. Pentelute
• Marinka Zitnik

Preprint | Project Website

ATOMICA is a geometric AI model that learns universal representations of molecular interactions at an atomic scale. The model is pretrained on 2,037,972 molecular interaction interfaces from the Protein Data Bank and Cambridge Structural Database, this includes protein-small molecule, protein-ion, small molecule-small molecule, protein-protein, protein-peptide, protein-RNA, protein-DNA, and nucleic acid-small molecule complexes. Embeddings of ATOMICA can be generated with the open source model weights and code to be used for various downstream tasks. In the paper, we demonstrate the utility of ATOMICA embeddings for studying the human interfaceome network with ATOMICANets and for annotating ions and small molecules to proteins in the dark proteome.

🚀 Installation and Setup

ATOMICA requires PyTorch with CUDA support. Please refer to the installation instructions in setup which provides instructions for setting up with uv or mamba/conda.

⚡ Quick Start

Generate embeddings from list of PDB files with ATOMICA model in just a few lines. See the tutorial at tutorials/1_get_embeddings for more details.

⭐ Other setup

Optional steps, only required if you plan on training your own ATOMICA model.

Download All Processed Datasets

The data for pretraining and downstream analyses is hosted at Harvard Dataverse.

We provide the following datasets:

• Processed CSD and QBioLiP (based on PDB) interaction complex graphs for pretraining
• Processed protein interfaces of human proteome binding sites to ion, small molecule, lipid, nucleic acid, and protein modalities
• Processed protein interfaces of dark proteome binding sites to ion and small molecules

Download All Model Checkpoints

Model checkpoints are provided on Hugging Face. The following models are available:

• ATOMICA model
• Pretrained ATOMICA-Interface model
• Finetuned ATOMICA-Ligand prediction models for the following ligands:
  • metal ions: Ca, Co, Cu, Fe, K, Mg, Mn, Na, Zn
  • small molecules: ADP, ATP, GTP, GDP, FAD, NAD, NAP, NDP, HEM, HEC, CIT, CLA

Training / Finetuning your own ATOMICA model

Training scripts for pretraining ATOMICA and finetuning ATOMICA-Interface and ATOMICA-Ligand are provided in scripts/.

🌱 Tutorials

Get embeddings from ATOMICA model

Refer to the tutorial at tutorials/1_get_embeddings for more details.

Inference with ATOMICA-Ligand

Refer to the jupyter notebook at tutorials/2_atomica_ligand for an example of how to use the model for dark proteome ligand predictions.

Other tutorials

Coming soon!

💡 Questions

For questions, please leave a GitHub issue or contact Ada Fang at ada_fang@g.harvard.edu.

⚖️ License

The code in this package is licensed under the MIT License.

📜 Citation

If you use ATOMICA in your research, please cite the following preprint:

@article{fang2025atomica,
  title={Learning Universal Representations of Intermolecular Interactions with ATOMICA},
  author={Fang, Ada and Desgagné, Michael and Zhang, Zaixi and Zhou, Andrew and Loscalzo, Joseph, and Pentelute, Bradley L and Zitnik, Marinka},
  journal={In Review},
  url={https://www.biorxiv.org/content/10.1101/2025.04.02.646906},
  year={2025}
}