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

1. Download the Repository

Clone the GitHub Repository:

git clone https://github.com/mims-harvard/ATOMICA
cd ATOMICA

To include the model checkpoints and datasets (hosted as a Git submodule on HuggingFace), initialize and update the submodule:

git submodule init
git submodule update

Alternatively, you can clone the repository with submodules in one command:

git clone --recurse-submodules https://github.com/mims-harvard/ATOMICA
cd ATOMICA

2. Set Up Environment

We use uv for dependency management. Install dependencies with:

uv sync

Note: Some features (molecule visualization) require system-level X11 libraries. Most operations (PDB download, conversion, embedding) work without these. See docs/SYSTEM_DEPENDENCIES.md for details.

3. (optional) Download 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

4. Model Checkpoints

Model checkpoints are available via the downloads submodule (hosted on Hugging Face). If you initialized the submodule in step 1, the checkpoints are already available in the downloads/ directory. The following models are included:

• ATOMICA pretrained model (downloads/ATOMICA_checkpoints/pretrain/)
• Pretrained ATOMICA-Interface model (downloads/ATOMICA_checkpoints/prot_interface/)
• Finetuned ATOMICA-Ligand prediction models for the following ligands (downloads/ATOMICA_checkpoints/ligand/):
  • 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

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

🌱 Tutorials

Inference with ATOMICA-Ligand

Refer to the jupyter notebook at tutorials/atomica_ligand/example_run_atomica_ligand.ipynb for an example of how to use the model for binder prediction.

Explore ATOMICANets

Refer to the jupyter notebook at tutorials/atomica_net/example_atomica_net.ipynb

Embedding your own structures

Make sure you have initialized the Git submodule to access the ATOMICA model weights and config files (see Installation step 1). The model files will be available in the downloads/ATOMICA_checkpoints/ directory.

For embedding biomolecular complexes: process .pdb files with data/process_pdbs.py and embed with get_embeddings.py. See the tutorial for data processing at data/README.md here and the examples at data/example.

For embedding protein-(ion/small molecule/lipid/nucleic acid/protein) interfaces: first predict (ion/small molecule/lipid/nucleic acid/protein) binding sites with PeSTo, second process the PeSTo output .pdb files with data/process_PeSTo_results.py, finally embed with get_embeddings.py.

💡 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}
}