Learning Universal Representations of Intermolecular Interactions with ATOMICA.

Area of Science:

• Computational Biology
• Biophysics
• Machine Learning

Background:

• Molecular interactions are fundamental to biological processes.
• Current machine learning models often isolate molecules or focus on single interaction types, limiting generalization.
• A systematic approach to modeling diverse biomolecular interactions is needed.

Purpose of the Study:

• Introduce ATOMICA, a geometric deep learning model for learning atomic-scale representations of intermolecular interfaces.
• Enable generalization across diverse biomolecular modalities (small molecules, ions, amino acids, nucleic acids).
• Systematically annotate molecular interactions and the dark proteome.

Main Methods:

• Developed ATOMICA, a geometric deep learning model utilizing a self-supervised denoising and masking objective.
• Trained on over 2 million interaction complexes to generate hierarchical embeddings (atoms, chemical blocks, interfaces).
• Applied ATOMICA to construct five modality-specific interfaceome networks (ATOMICA Nets).

Main Results:

• ATOMICA generalizes across molecular classes, recovering shared physicochemical features without supervision.
• The model's latent space captures similarities across interaction types and improves with data.
• Constructed ATOMICA Nets connecting proteins based on interaction similarity, identifying disease pathways and predicting disease-associated proteins.
• Annotated 2,646 uncharacterized ligand-binding sites in the dark proteome, including potential zinc finger motifs.

Conclusions:

• ATOMICA provides a unified framework for learning and representing intermolecular interfaces across diverse biomolecules.
• The model facilitates systematic annotation of molecular interactions, aiding in disease pathway identification and dark proteome characterization.
• ATOMICA advances the understanding of molecular interactions and their role in biological processes and disease.