AI Model Catalog | Microsoft Foundry Models

Atomica

Version: 2

Mims Harvard•Last updated July 2025

ATOMICA is a hierarchical geometric deep learning model trained on over 2.1 million molecular interaction interfaces. It represents interaction complexes using an all-atom graph structure, where nodes correspond to atoms or grouped chemical blocks, and edges reflect both intra- and intermolecular spatial relationships. The model uses SE(3)-equivariant message passing to ensure that learned embeddings are invariant to rotations and translations of molecular structures. The architecture produces embeddings at multiple scales—atom, block, and graph—that capture fine-grained structural detail and broader functional motifs. Project Website The model has been pretrained on 2,105,703 molecular interaction interfaces from the Protein Data Bank and Cambridge Structural Database, spanning multiple interaction types including protein-small molecule, protein-ion, small molecule-small molecule, protein-protein, protein-peptide, protein-RNA, protein-DNA, and nucleic acid-small molecule complexes. The pretraining strategy involves denoising transformations (rotation, translation, torsion) and masked block-type prediction, enabling the model to learn chemically grounded, transferable features. ATOMICA supports downstream tasks via plug-and-play adaptation with task-specific heads, including binding site prediction and protein interface fingerprinting. Project Website | GitHub

Model Architecture

ATOMICA employs a hierarchical geometric deep learning architecture with the following key components:

An all-atom graph representation of molecular complexes where nodes correspond to atoms or chemical blocks
SE(3)-equivariant message passing to ensure rotational and translational invariance
Multi-scale embeddings at atom, block, and graph levels
Pretraining via denoising transformations and masked block-type prediction

Sample inputs and outputs (for real time inference)

Single PDB Input:

data = {
  "input_data": {
    "columns": ["pdb_data"],
    "index": [0],
    "data": [
      ["https://path/to/local/protein.jsonl.gz"]
    ]
  }
}

Multiple PDB Input:

data = {
  "input_data": {
    "columns": ["pdb_data"],
    "index": [0, 1],
    "data": [
      ["https://path/to/local/protein1.jsonl.gz"],
      ["https://path/to/local/protein2.jsonl.gz"]
    ]
  }
}

Output Sample:

{
  "predictions": [
    {
      "graph_embedding": [0.123456, -0.234567, ...],
      "block_embedding": [[0.345678, -0.456789, ...], ...],
      "atom_embedding": [[0.567890, -0.678901, ...], ...]
    }
  ]
}

Output Processing Example:

def process_predictions(predictions, is_batch=False):
    """Process predictions in a consistent way."""
    if not predictions:
        print("No predictions found")
        return

    # Extract predictions from response
    if isinstance(predictions, dict) and 'predictions' in predictions:
        predictions = predictions['predictions']

    # Handle both single and batch predictions
    if is_batch:
        if isinstance(predictions, list):
            # Multiple predictions
            for i, pred in enumerate(predictions, 1):
                print(f"\nPDB {i}:")
                _display_embeddings(pred)
        elif isinstance(predictions, dict):
            # Single prediction in batch format
            _display_embeddings(predictions)
        else:
            print(f"Error: Unexpected predictions type: {type(predictions)}")
    else:
        # Single prediction format
        if isinstance(predictions, list) and len(predictions) > 0:
            # Single prediction in list
            _display_embeddings(predictions[0])
        elif isinstance(predictions, dict):
            # Direct dictionary format
            _display_embeddings(predictions)
        else:
            print(f"Error: Invalid prediction format: {type(predictions)}")

def _display_embeddings(pred):
    """Display embeddings in a concise format."""
    if not isinstance(pred, dict):
        print(f"Error: Invalid prediction format: {type(pred)}")
        return
        
    for embed_type in ["graph_embedding", "block_embedding", "atom_embedding"]:
        if embed_type in pred:
            embedding = pred[embed_type]
            if isinstance(embedding, list):
                arr = np.array(embedding)
                print(f"\n{embed_type}:")
                print(f"Shape: {arr.shape}")
                print(f"Mean: {arr.mean():.6f}, Std: {arr.std():.6f}")
                print(f"Range: [{arr.min():.6f}, {arr.max():.6f}]")
        else:
            print(f"\nWarning: {embed_type} not found in predictions")
            print("Available keys:", list(pred.keys()) if isinstance(pred, dict) else "Not a dictionary")

Data and Resource Specification for Deployment

Supported Data Input Format

Input Format: The model accepts molecular structure data in PDB format, specifically processed as compressed JSONL files (.jsonl.gz). Each file contains molecular interaction complex graphs with atomic coordinates and structural information.
Input Methods: The model supports:
- URLs pointing to remote .jsonl.gz files
- Base64-encoded molecular structure data
Output Format: The model generates multi-scale molecular embeddings in three hierarchical levels:
- Graph embedding: Overall molecular complex representation
- Block embedding: Chemical block-level features
- Atom embedding: Individual atomic-level representations
Data Sources and Technical Details: For comprehensive information about training datasets, model architecture, and validation results, refer to the official ATOMICA repository

Model Specifications

LicenseMit

Last UpdatedJuly 2025

PublisherMims Harvard

Languages1 Language

Quick Start