NovoMCP addie-models — ADMET prediction weights

Pre-trained model weights for the addie-models ADMET service, part of the NovoMCP open computational chemistry engine. These weights power predict_admet, get_molecule_profile, and batch_profile when you self-host the addie-models compute service.

• Service code: https://github.com/NovoMCP/addie-models
• Engine: https://github.com/NovoMCP/novomcp
• License: weights MIT; see Attribution for training-data credits.

What's here

53 model heads across 31 base ADMET endpoints + a TDC state-of-the-art overlay. ~510 MiB.

Group	Heads	Format
Base ADMET	31 (production/) — 4 cardiotoxicity time-windows, 5 CYP450 inhibition, 7 nuclear-receptor, 5 stress-response, 9 toxicity, binding affinity	CatBoost model.tar.gz
TDC SOTA overlay	22 (models/) — CYP (Veith), CYP substrate, clearance, DILI, hERG, Ames, Caco-2, HIA, bioavailability, lipophilicity, solubility, BBB, PPBR, VDss, half-life, LD50, P-gp substrate	CatBoost .cbm (5-seed ensembles for winners), Chemprop .ckpt, TorchScript .pt
Calibrators	per-endpoint calibrator.json, DICTrank + DILI de-flatten calibration	JSON

The five SOTA winners (cyp2d6_veith, cyp3a4_veith, cyp3a4_substrate_carbonmangels, clearance_hepatocyte_az, dili) are 5-seed ensembles that meet or beat the published TDC leaderboard on held-out test splits.

Usage

These weights are consumed by the addie-models service, which downloads them automatically:

docker run -e STORAGE_BACKEND=HF -p 8025:8025 ghcr.io/novomcp/addie-models:latest
# on first boot the service pulls this repo, then serves /addie/process

Or point an existing checkout at it:

export STORAGE_BACKEND=HF
export HF_MODEL_REPO=NovoMCP/addie-models
python3 main.py

Directly with huggingface_hub:

from huggingface_hub import snapshot_download
local_dir = snapshot_download("NovoMCP/addie-models")
# base heads under local_dir/production/, TDC overlay under local_dir/models/

Attribution & data sources

Weights are released under MIT. They were trained on the datasets below; attribution to these sources is required when redistributing. Model weights are not a copyright derivative of the underlying factual measurements, but we credit every source:

• Tox21 (NIH/NCATS/EPA/FDA) — nuclear-receptor + stress-response heads. Public domain.
• PubChem BioAssay / Veith et al. 2009 — CYP450 inhibition heads. Public domain.
• FDA DICTrank (Liu et al. 2023) and DILIrank (Xu et al. 2015) — cardiotoxicity + DILI heads. Public.
• ChEMBL / AstraZeneca ADME datasets — ppbr_az, clearance_hepatocyte_az, clearance_microsome_az, lipophilicity_astrazeneca. Source data licensed CC-BY-SA 3.0; credited here per that license.
• AqSolDB (Sorkun et al. 2019) — solubility_aqsoldb. Source data CC-BY 4.0; credited here per that license.
• Academic measurement sets — Wang (Caco-2, 2016), Hou (HIA, 2007), Ma (bioavailability, 2008), Martins (BBB, 2012), Lombardo (VDss, 2016), Obach (half-life, 2008), Zhu (LD50, 2009), Carbon-Mangels (CYP substrate, 2011), Broccatelli/Wang (P-gp).
• Therapeutics Data Commons (Huang et al. 2021) — benchmark splits for the overlay.

See NOTICE for the full citation list.

Citation

@software{novomcp_addie_models,
  title  = {NovoMCP addie-models: ADMET prediction weights},
  author = {Harrison, Ari and contributors},
  url    = {https://github.com/NovoMCP/addie-models},
  year   = {2026}
}