CHEESE similarity search

POST /api/v2/search_cheese uses the same constructible-product acquisition family as default search, then exactly reranks assembled products with a CHEESE model. Use it when 3D shape or electrostatic similarity is the scientific objective; it has different interpretation and latency from ECFP4 search.

curl --request POST 'https://api.deepmedchem.com/api/v2/search_cheese' \
  --header "x-api-key: $DMC_API_KEY" \
  --header 'content-type: application/json' \
  --data '{
    "query_smiles": "CCO",
    "database_id": "enamine-real-v5a",
    "scorer": "shape",
    "limit": 20
  }'

scorer is required and is either shape or esp (electrostatic). The response identifies the model version, exact metric name, proposal score, final score, and timing. Model/conformer/charge assumptions are part of that pinned model release; do not compare numbers from different model versions as if they were calibrated identically.

Set shortlist_multiplier to 0 to disable candidate over-fetch. The endpoint then proposes exactly limit products and still scores those assembled products with the requested CHEESE model. It cannot disable CHEESE scoring itself without ceasing to be a CHEESE search.

shape_hits = dmc.search_cheese(
    "CCO",
    database="enamine-real-v5a",
    scorer="shape",
    limit=20,
)

CHEESE work has a slower guardrail than default Morgan search and separate capacity. A timeout does not imply no neighbors exist. More than one query always uses a selection_batch run.