addie-models

ADMET prediction service for the NovoMCP open computational chemistry engine. Serves 31 base ADMET endpoints plus a 22-model Therapeutics Data Commons (TDC) state-of-the-art overlay — CYP inhibition/substrate, clearance, hepatotoxicity (DILI), cardiotoxicity (hERG + DICTrank), Ames, Tox21 nuclear-receptor/stress-response panels, permeability, solubility, and more.

Weights are published separately on Hugging Face: NovoMCP/addie-models (~510 MiB, MIT).

Quickstart

Build and run from source — the recommended path; no dependency on a prebuilt image. Weights download from Hugging Face on first boot (no cloud credentials):

docker build -t addie-models .
docker run -p 8025:8025 addie-models
# first boot downloads the weights (~510 MiB), then serves on :8025

Or pull the prebuilt image:

docker run -p 8025:8025 ghcr.io/novomcp/addie-models:latest

Or run without Docker:

pip install -r requirements.txt   # Python 3.11 recommended
python3 main.py

Then:

curl -s http://localhost:8025/health
# {"status":"healthy","models_loaded":31, ...}

curl -s -X POST http://localhost:8025/addie/process \
  -H 'Content-Type: application/json' \
  -d '{"molecules":[{"id":"aspirin","smiles":"CC(=O)Oc1ccccc1C(=O)O"}]}'

Endpoints

Method	Path	Purpose
GET	/health	Liveness + models_loaded count
GET	/models	List loaded model endpoints
POST	/addie/process	Predict ADMET for a batch of molecules (up to 100)

Configuration

Env var	Default	Effect
STORAGE_BACKEND	HF	Where weights come from: HF (Hugging Face) | S3 | AZURE
HF_MODEL_REPO	NovoMCP/addie-models	Hugging Face weights repo (when STORAGE_BACKEND=HF)
PORT	8025	HTTP port
MODEL_PREFIX	production/	Base-model key prefix within the weights store
MODEL_BUCKET	—	S3 bucket (only when STORAGE_BACKEND=S3)

Model coverage

• Base ADMET (31): binding affinity, 4 cardiotoxicity time-windows, 5 CYP450 inhibition, 7 nuclear-receptor (Tox21), 5 stress-response (Tox21), 9 toxicity (Ames, carcinogenicity, clinical/developmental/reproductive/respiratory toxicity, eye corrosion/irritation, hepatotoxicity).
• TDC SOTA overlay (22): CYP (Veith) + substrate, clearance (hepatocyte/microsome), DILI, hERG, Caco-2, HIA, bioavailability, lipophilicity, solubility, BBB, PPBR, VDss, half-life, LD50, P-gp substrate. Five endpoints are 5-seed ensembles that meet or beat the published TDC leaderboard.

See the model card for per-endpoint training data and attribution.

Testing

# with the service running locally
ADDIE_URL=http://localhost:8025 python3 tests/test_model_coverage.py

License

• Code: Apache-2.0 (see LICENSE).
• Model weights: MIT, with attribution to training-data sources — see the model card and its NOTICE.