Skip to content

novomcp-properties

Trained ML models for physicochemical properties: pKa, aqueous solubility, and bond dissociation energy (BDE). CPU-only. Model weights load from Hugging Face on first start — no cloud credentials required.

Pre-reqs

  • Docker
  • ~4 GB RAM
  • No GPU
  • Network access on first boot (weights download from Hugging Face)
  • For the charge-based pKa routes (sulfonamides / aromatic N–H): a running novomcp-qm service supplying per-atom charges. Without it, those routes are unavailable and pKa is served by the general model only.

Deploy

docker run -d \
  --name novomcp-properties \
  -p 8030:8030 \
  --restart unless-stopped \
  ghcr.io/novomcp/novomcp-properties:latest

Wire into the engine

export NOVOMCP_PROPERTIES_URL=http://localhost:8030
export NOVOMCP_QM_URL=http://localhost:8031   # required for the charge-based pKa routes

Verify

curl -s http://localhost:8030/health
# {"status":"healthy","service":"novomcp-properties","version":"1.0.0","port":8030,
#  "predictors":{"pka":{"backend":"rdkit-empirical","ready":false,"weights_loaded":false,"empirical_only":true},
#                "solubility":{"backend":"chemprop-aqsoldb","ready":true},
#                "bde":{"backend":"alfabet","ready":true}},
#  "ready":"2/3"}
# pka shows ready:false until you opt in to the NonCommercial weights (HF_PKA_MODEL_REPO); solubility + bde are ready.

Tools that light up

  • predict_pka, acidic/basic ionization constants
  • predict_solubility, LogS (log molar aqueous solubility) with temperature dependence
  • predict_bde, bond dissociation energies for radical chemistry

Env vars

Var Default Purpose
PORT 8030 HTTP listen port
STORAGE_BACKEND HF Weights backend: HF | LOCAL | S3
HF_MODEL_REPO NovoMCP/novomcp-properties Hugging Face weights repo (permissive: solubility)
HF_PKA_MODEL_REPO – NonCommercial pKa weights repo, opt-in (e.g. NovoMCP/novomcp-pka)
NOVOMCP_QM_URL – novomcp-qm endpoint for per-atom charges (charge-based pKa routes)
BATCH_SIZE 64 Molecules per batch

Notes

  • pKa weights are NonCommercial / ShareAlike. The pKa model is trained primarily on the IUPAC Dissociation Constants (CC-BY-NC-4.0) plus ChEMBL (CC-BY-SA 3.0), so its weights ship separately under CC-BY-NC-SA-4.0 (NovoMCP/novomcp-pka) and are opt-in: set HF_PKA_MODEL_REPO for non-commercial use. Left unset, the pKa endpoints return 503 (solubility and BDE are unaffected). The service code is Apache-2.0; only the pKa weights carry the NonCommercial / ShareAlike terms.
  • pKa model: a routed ensemble — a per-atom-charge specialist for sulfonamides / aromatic N–H, and a general model for everything else; each route reports an uncertainty estimate. Benchmarked on SAMPL7.
  • Solubility model: pre-trained on AqSolDB, fine-tuned on BigSolDB with temperature as an input feature.
  • BDE model: alfabet pretrained network.
  • Solubility and BDE outputs are screening-grade ML predictions: dependable for ranking within a comparable series, not as absolute experimental values.
  • If weights can't be loaded, the affected predictor reports unavailable and its endpoints return 503 rather than serving a silent fallback.
  • All three are stateless, safe to scale horizontally behind a load balancer for high-throughput screening.