CHEESE: AI-Based Tools for Accelerated Drug Discovery

Charge Types

Training

• 70,000 DFT calculations on diverse drug-like chemotypes
• QM level: ωB97X-D/def2-svp (dispersion-corrected DFT with split-valence basis)
• Validated on an out-of-distribution scaffold-split test set (Morgan Tanimoto < 0.3)

Model Architecture

Two parallel Geometric Transformers (one for ESP, one for RESP), each consisting of:

• transformer.pt (~41 MB) — a geometry-aware transformer that takes atomic features + 3D coordinates and outputs learned representations
• projection.pt (~2.7 KB) — a final linear layer mapping features to one charge value per atom

Built on the en-transformer package (equivariant/geometric transformer).

Input Features

• Atomic numbers (integer per atom)
• 3D Cartesian coordinates (from MMFF-optimized conformer)
• Bond matrix (symmetric adjacency matrix with integer-encoded bond types, 13 states)

Inference Pipeline

How It Works

• A set of low-energy 3D conformers is generated for both the query and the probe (retrieved) molecule.
• Each probe conformer is aligned onto each query conformer using the selected alignment method.
• The best-scoring pairs are returned with their RMSD, energy, and alignment score.
• You can interactively switch between conformer pairs in the 3D viewer to inspect the overlap.

Alignment Methods

MCS (Maximum Common Substructure)

Identifies the largest shared substructure between the query and probe molecules, then aligns conformers by superimposing the matched atoms.

• Finds the maximum common substructure (MCS) between two molecules using RDKit's FMCS algorithm
• Generates 3D conformer ensembles for both the query and probe molecules
• Aligns each probe conformer onto each query conformer using the MCS atom mapping
• Reports RMSD computed over the matched atom pairs
• Best suited when molecules share a clear common scaffold

MCS Constrained

Uses the MCS core as a rigid constraint during conformer generation, then aligns on the constrained atoms. This produces tighter overlaps on the shared scaffold.

• Identifies the MCS core the same way as the standard MCS method
• Generates probe conformers with the core atoms constrained to match the reference geometry (ConstrainedEmbed)
• Non-core atoms are energy-minimised while the core stays fixed
• Yields lower RMSD on the core but may differ more in peripheral regions
• Useful when preserving the binding-pose geometry of the shared scaffold is critical

Open3DAlign

An unsupervised 3D alignment method that optimises the overlap of molecular volumes and force-field atom types without requiring a predefined atom mapping.

• Uses the Open3DAlign algorithm implemented in RDKit (GetO3A)
• Automatically determines the optimal atom-atom mapping based on MMFF94 atom types
• Maximises a score that combines steric (shape) and electrostatic overlap
• Does not require a common substructure — works for structurally diverse molecule pairs
• Returns an alignment score and RMSD for each conformer pair

When to Use Which Method?

Methodology & Results

• Searched 100 randomly selected inhibitors using our tool
• Found at least one new inhibitor for 14 of the query inhibitors
• Discovered 5 new inhibitors from a single query inhibitor
• Some newly found inhibitors were highly similar in 3D Shape and Electrostatic similarity

Benefits

• Complete data privacy — no transmission over internet
• Deploy on AWS, your own servers, or dedicated instances
• Full CHEESE functionality with custom database indexing
• Train models on your own proprietary data

Authentication

API calls require an API key. To obtain one:

1. Go to cheese.deepmedchem.com and sign in
2. Click "Generate API Key" in the top right corner
3. Copy the API key and include it in request headers

Verify the API is running and accessible.

response = requests.get(MY_URL + "/test", headers=headers)
# {"message": "Health check successful !!"}

curl -X GET 'https://api.cheese.deepmedchem.com/test' \
  -H 'Accept: application/json' \
  -H 'X-API-Key: YOUR_API_KEY'

Returns the list of searchable molecular databases.

response = requests.get(MY_URL + "/available_dbs", headers=headers)

curl -X GET 'https://api.cheese.deepmedchem.com/available_dbs' \
  -H 'Accept: application/json' \
  -H 'X-API-Key: YOUR_API_KEY'

Returns random molecules in SMILES format from available databases.

params = {"n_mols": 5}
response = requests.get(MY_URL + "/random_molecule",
    headers=headers, params=params)

curl -X GET 'https://api.cheese.deepmedchem.com/random_molecule?n_mols=5' \
  -H 'Accept: application/json' \
  -H 'X-API-Key: YOUR_API_KEY'

Basic molecular similarity search. Returns a list of SMILES strings and database IDs of similar molecules.

params = {
    "search_input": "CC(=O)Oc1ccccc1C(=O)O",
    "search_type": "espsim_shape",
    "n_neighbors": 5,
    "search_quality": "fast",
    "db_names": ["ENAMINE-REAL"],
}

response = requests.get(MY_URL + "/molsearch_simple",
    params=params, headers=headers)

curl -X GET 'https://api.cheese.deepmedchem.com/molsearch_simple?search_input=CC(%3DO)Oc1ccccc1C(%3DO)O&search_type=espsim_shape&n_neighbors=5&search_quality=fast&db_names=ENAMINE-REAL' \
  -H 'Accept: application/json' \
  -H 'X-API-Key: YOUR_API_KEY'

Search for multiple query molecules at once. Returns results keyed by input SMILES.

params = {
    "search_input": ["CC1=CN(C)N=C1", "CNC1=CC=CC=C1"],
    "search_type": "espsim_shape",
    "n_neighbors": 5,
    "search_quality": "fast",
    "db_names": "ZINC15",
}

response = requests.get(MY_URL + "/molsearch_array",
    params=params, headers=headers)

curl -X GET 'https://api.cheese.deepmedchem.com/molsearch_array?search_input=CC1%3DCN(C)N%3DC1&search_input=CNC1%3DCC%3DCC%3DC1&search_type=espsim_shape&n_neighbors=5&search_quality=fast&db_names=ZINC15' \
  -H 'Accept: application/json' \
  -H 'X-API-Key: YOUR_API_KEY'

Search multiple molecules in batch or centroid mode for consensus results.

params = {
    "search_input": ["SMILES1", "SMILES2", "SMILES3"],
    "search_type": "espsim_shape",
    "n_neighbors": 10,
    "search_mode": "batch",  # or "centroid"
}

response = requests.get(MY_URL + "/batch_search",
    params=params, headers=headers)

curl -X GET 'https://api.cheese.deepmedchem.com/batch_search?search_input=SMILES1&search_input=SMILES2&search_input=SMILES3&search_type=espsim_shape&n_neighbors=10&search_mode=batch' \
  -H 'Accept: application/json' \
  -H 'X-API-Key: YOUR_API_KEY'

Filter out PAINS, enforce scaffold hops, and sort results by Morgan Tanimoto similarity.

params = {
    "search_input": "CC(=O)Oc1ccccc1C(=O)O",
    "search_type": "espsim_shape",
    "n_neighbors": 10,
    "search_quality": "fast",
    "db_names": ["ZINC15"],
    "descriptors": True, "properties": True,
    "filter_molecules": True,
    "order_molecules": True,
    "filtering": ["PAINS", "Murcko scaffold hop"],
    "ordering": ["Morgan Tanimoto"],
}

response = requests.get(MY_URL + "/molsearch",
    params=params, headers=headers)

curl -X GET 'https://api.cheese.deepmedchem.com/molsearch?search_input=CC(%3DO)Oc1ccccc1C(%3DO)O&search_type=espsim_shape&n_neighbors=10&search_quality=fast&db_names=ZINC15&descriptors=true&properties=true&filter_molecules=true&order_molecules=true&filtering=PAINS&filtering=Murcko+scaffold+hop&ordering=Morgan+Tanimoto' \
  -H 'Accept: application/json' \
  -H 'X-API-Key: YOUR_API_KEY'

Note: Filtering may reduce the number of results. Increase n_neighbors to compensate.