Learning Universal Representations of Intermolecular Interactions with ATOMICA

The paper introduces ATOMICA, a geometric deep learning model trained on over two million complexes to generate universal, multiscale atomic representations of intermolecular interfaces across five molecular modalities, demonstrating superior performance in structure-function benchmarks and successfully predicting functional ligands for previously uncharacterized "dark" protein pockets.

The Gist

Imagine the human body and the natural world as a massive, bustling city. In this city, every biological process—from digesting your lunch to fighting a virus—is a conversation between different "citizens" (molecules). Sometimes a protein talks to a small drug molecule, sometimes a protein shakes hands with another protein, and sometimes a metal ion joins the chat.

For a long time, scientists trying to understand these conversations had a problem: they were building separate dictionaries for every type of conversation. They had one dictionary for protein-to-protein talks, another for protein-to-drug chats, and a third for RNA interactions. If you wanted to understand a new type of conversation, you had to build a whole new dictionary from scratch.

Enter ATOMICA.

Think of ATOMICA as a universal translator or a master diplomat that has learned the "language of touch" for all types of molecular citizens at once. Instead of learning separate languages, it learns the underlying grammar of how things fit together in 3D space.

Here is how it works, broken down into simple concepts:

1. The "Lego" Approach (The Architecture)

Most models look at molecules like a string of beads (a sequence). ATOMICA looks at them like 3D Lego structures.

• The Atoms: It sees the individual plastic bricks (atoms).
• The Blocks: It groups those bricks into meaningful chunks, like "amino acid bricks" for proteins or "chemical motif bricks" for drugs.
• The Interface: It focuses specifically on the interface—the exact spot where two molecules touch. It's like a diplomat who doesn't care about the whole country, but only about the specific handshake happening at the border.

2. The "Gym" Training (The Learning)

To become this expert, ATOMICA didn't just read books; it went to a massive gym with over 2 million different molecular complexes (a mix of proteins, drugs, DNA, lipids, and metal ions).

• The Workout: The trainers (scientists) would take a complex, shake it up, rotate it, or hide a piece of it (masking), and ask ATOMICA to guess what the original shape and missing piece were.
• The Result: By doing this millions of times, ATOMICA learned the "physics of fit." It learned that certain shapes and chemical charges naturally attract each other, regardless of whether they are made of protein or plastic.

3. Why This is a Big Deal (The Superpowers)

Because ATOMICA learned from everything at once, it has some cool superpowers:

• The "Low-Data" Hero: Imagine you are trying to learn a rare language that only has 50 examples. A normal student would fail. But ATOMICA, having learned 2 million examples of other languages, can look at those 50 examples and say, "Ah, I've seen this pattern before in a different context!" It uses what it knows about common interactions to understand rare ones.
• The "Dark Proteome" Detective: There are millions of proteins in our bodies that scientists have no idea what they do. They are like "dark matter" in the universe. ATOMICA looked at the 3D shape of these mysterious proteins and said, "This pocket looks exactly like a place that holds a heme (a red blood cell helper)."
  • The Proof: The team took 5 of these predictions, built the proteins in a lab, and tested them. Five out of five actually grabbed the heme, just like ATOMICA predicted. It found a needle in a haystack without ever seeing the needle before.
• The "Drug Hunter": If you have a protein you want to stop (like a cancer cell), you need a drug that fits its "handshake" spot. ATOMICA can look at a drug and say, "This looks like it fits that protein's handshake," even if the drug and protein have never met before.

4. The "Invisible Handshake" (Cross-Modality)

One of the most magical things ATOMICA does is realize that a drug and a protein can look very similar in the "language of touch."

• Imagine a thief (a drug) trying to break into a house (a protein) by mimicking the key (a natural protein partner).
• ATOMICA can look at the thief and the natural key and say, "These two look like they belong in the same lock." This helps scientists find new drugs that can block bad interactions by mimicking the good ones.

Summary

ATOMICA is like a master architect who has studied every building, bridge, and house in the world. Because it understands the fundamental rules of how bricks fit together, it can now look at a blueprint for a building it has never seen before and instantly know:

1. What kind of room it is.
2. What furniture (drugs/ions) fits inside it.
3. How to fix it if it's broken.

It moves us from "guessing" how molecules interact to "knowing" based on a deep, universal understanding of the 3D world.

Technical Summary

1. Problem Statement

Molecular interactions are fundamental to biological processes, yet current representation learning models suffer from significant limitations:

• Modality Silos: Most models focus on single entities (e.g., protein language models) or specific interaction pairs (e.g., protein-ligand or protein-protein), failing to learn a unified representation across diverse molecular classes.
• Lack of Generalization: Models trained on specific interaction types often cannot transfer knowledge to other modalities (e.g., from protein-small molecule to protein-RNA), especially in "low-data" regimes where structural data is scarce.
• Missing Interface-Centric Learning: Existing foundation models often prioritize sequence or isolated structure over the geometric and chemical context of the interface where interactions occur.

The authors propose ATOMICA, a geometric deep learning model designed to learn universal, atomic-scale representations of intermolecular interfaces across five distinct molecular modalities.

2. Methodology

A. Dataset Curation

ATOMICA is trained on a massive, multimodal dataset of 2,037,972 interaction complexes derived from:

• Cambridge Structural Database (CSD): ~1.77 million small-molecule crystal pairs.
• Protein Data Bank (PDB) & Q-BioLiP: ~338,000 biological complexes.
• Modalities Covered: Proteins, small molecules, metal ions, lipids, and nucleic acids (DNA/RNA).
• Interaction Types: Eight distinct types, including protein-protein, protein-ligand, protein-RNA, protein-ion, etc.
• Interface Definition: Atoms within an 8 Å distance of the partner molecule are included to capture the local molecular context.

B. Model Architecture

ATOMICA is a hierarchical, all-atom geometric graph neural network utilizing SE(3)-equivariant tensor field networks.

• Hierarchical Graph Representation:
  • Atom Level: Nodes represent atoms (element type + 3D coordinates).
  • Block Level: Atoms are grouped into chemically meaningful blocks (e.g., amino acids, nucleotides, functional motifs). Small molecules are tokenized into 290 common chemical motifs using a graph-based Byte Pair Encoding algorithm.
  • Edges: Defined as intramolecular (within a molecule) and intermolecular (between molecules) connections based on $k$-nearest neighbors in Euclidean space.
• Message Passing: Uses SE(3)-equivariant layers to ensure the model's output is invariant to rotation and translation, passing messages over both intra- and intermolecular edges.
• Multi-Scale Embeddings: Generates embeddings at three levels:
  1. Atom-level ($h_{atom}$)
  2. Block-level ($h_{block}$)
  3. Graph/Interface-level ($h_{graph}$, $h_{interface}$)

C. Self-Supervised Pretraining Objectives

The model is pretrained using a combination of denoising and masked prediction tasks to learn geometric and chemical features without labeled interaction data:

1. Denoising (Denoising Score Matching): The model reconstructs the original interaction complex graph after it is perturbed by:
   • Rigid SE(3) transformations (rotation/translation) of one molecule.
   • Random torsion angle rotations of rotatable bonds.
   • Goal: Learn relative spatial relationships and local chemical context rather than absolute coordinates.
2. Masked Block Identity Prediction: Randomly masks 10% of chemical blocks at the interface and predicts their identity.
   • Goal: Learn the chemical compatibility and functional roles of specific blocks within an interface.

3. Key Contributions

1. Unified Multimodal Representation: ATOMICA is the first model to learn a shared embedding space for five distinct molecular modalities (proteins, small molecules, ions, lipids, nucleic acids) and eight interaction types, enabling cross-modality transfer.
2. Scaling Laws for Interactions: Demonstrates that pretraining on diverse interaction modalities significantly improves performance on low-data interfaces (e.g., protein-DNA/RNA) compared to models trained on single modality pairs.
3. Cross-Modality Interface Comparison: Proves that the shared latent space allows for meaningful comparison between orthosteric inhibitors and native binding partners, even across different molecular types (e.g., protein-peptide vs. protein-inhibitor).
4. Dark Proteome Annotation: Successfully applies the model to predict ligand identities (ions and cofactors) for "dark" protein pockets (proteins with unknown function), leading to experimental validation.

4. Results

A. Benchmark Performance

• RNA Structure-Function (RNAGlib): ATOMICA outperformed existing RNA structure encoders (gRNAde, RNAglib) and RNA language models (RiNALMo, RNA-FM) across four tasks:
  • Protein-binding site prediction (AUPRC +0.118 over best baseline).
  • RNA functional annotation (GO terms) (F1 macro +0.105).
  • Small-molecule binding site prediction.
  • RNA pocket ligand identification.
• Protein Pocket Ligand Classification (MaSIF):
  • Surpassed specialized pocket encoders (MASIF) by +0.096 F1 macro.
  • Achieved performance comparable to massive protein language models (e.g., ProstT5 with 600M+ parameters) despite having only 7.5M parameters.

B. Zero-Shot Analysis & Interpretability

• ATOMICA Score: A metric derived by masking a block and measuring the change in graph embedding. It successfully identified residues involved in non-covalent contacts (hydrogen bonds, hydrophobic, aromatic) with higher precision than ESM-2 (3B).
• Latent Space Structure: UMAP and PCA visualizations confirmed that the latent space groups molecules by chemical similarity (e.g., separating metals from non-metals, amino acids from nucleotides) and reflects physicochemical properties.

C. Cross-Modality Retrieval (PPI Inhibitors)

• Tested on orthosteric PPI inhibitors (small molecules blocking protein-protein interfaces).
• Finding: Inhibitor embeddings in the ATOMICA space were significantly more similar to the native peptide/protein interface blocks than to random surface patches.
• Metric: Fold Change@10 showed that top-ranked inhibitor-block pairs localized spatially to the native binding site in 78% of protein-peptide complexes and 100% of protein-protein complexes tested.

D. Experimental Validation (Dark Proteome)

• Prediction: Applied ATOMICA-Ligand to 2,646 dark protein pockets, predicting metal ions and cofactors.
• Validation: Five predicted heme-binding proteins were synthesized and tested via UV-Vis spectroscopy.
  • Result: All five candidates showed experimental evidence of heme binding (red-shifted Soret band).
  • Significance: Validated predictions included proteins lacking canonical heme-binding motifs (e.g., CXXCH), demonstrating the model's ability to learn from 3D geometry rather than just sequence motifs.

5. Significance and Impact

• Paradigm Shift: Moves the field from entity-centric models (learning about proteins or drugs in isolation) to interaction-centric models, recognizing that function arises from the interface.
• Data Efficiency: Demonstrates that multimodal pretraining acts as a powerful regularizer, boosting performance in data-scarce domains (like protein-nucleic acid interactions) by leveraging knowledge from data-rich domains.
• Drug Discovery & Functional Annotation: Provides a scalable tool for annotating the "dark proteome" (uncharacterized proteins) and suggesting putative ligands, accelerating the discovery of new drug targets and cofactors.
• Generalizability: The shared embedding space enables novel applications, such as using small-molecule inhibitors to probe protein-protein interaction surfaces, bridging the gap between structural biology and chemical biology.

In conclusion, ATOMICA establishes that a single, geometrically grounded model can learn universal representations of intermolecular interactions, outperforming specialized models and large language models in specific structural tasks while offering a unified framework for understanding biomolecular complexity.