Delower's picture
Upload folder using huggingface_hub
ea77201 verified
|
Raw History Blame Contribute Delete
2.55 kB
# hERG ExtraTrees (comprehensive features) — DEPLOYMENT REFITS
**DEPLOYMENT REFIT — trained on ALL labeled data, including the regulatory/external sets
(hERG-70, new-compounds 967, FDA panel 36, REGFAIL).** This is the third hERG deployment
model, added after the Phase-5 finish chain (2026-11-16): the comprehensive-features
ExtraTrees is the **best single (non-fusion) model** of the locked-pool Phase-5 benchmark,
refit on the complete labeled corpus for production screening use:
- classification: 14,441 compounds (`cls_alldata.csv`)
- regression: 10,370 compounds (`reg_alldata.csv`)
Hard gate: every external/regulatory compound was verified present in the actual training
files AND in the fitted training matrices by InChIKey connectivity before these bundles
were staged (hERG-70 462/462, new-compounds 967/967, FDA 36/36, REGFAIL 7/7, hERG-60
249/249, Test-1 43/43 cls / 17/43 reg; counts in `validation_results.json`).
**Do not validate these models on the absorbed sets** — scores there are in-sample by
construction. Honest external benchmarks of this architecture (locked-pool Phase-5
refits, external sets locked out) are listed in `config.json` /
`validation_results.json` under `honest_locked_benchmark_scores`: classification
scaffold refit hERG-70 ACC 0.760 / MCC **0.543** (best single model), FDA 4 errors,
REGFAIL Tier-1 4/5; regression hERG-70 RMSE 0.838.
## Contents
- `models/` — 4 bundles (cls/reg x scaffold/random; per task both protocols wrap the same
fitted model: identical combo/hyperparameters + all-data training make the protocol
distinction vanish). Headline = scaffold.
- `predict_herg.py` — standalone inference (SMILES -> blocker probability / pIC50).
- `feature_extraction.py` — the exact Phase-5 feature blocks.
- `config.json`, `validation_results.json`, `regulatory_inference_sample.csv`
(in-sample demonstration predictions on REGFAIL compounds, labelled as such).
Recipe (identical to the locked-pool finals): ExtraTrees, 300 trees, seed 42.
Feature blocks — classification: Morgan + MACCS + descriptors + physchem-extra +
hERG + pharmacophore blocks (2,499 features after preprocessing); regression:
Morgan + descriptors + hERG + pharmacophore blocks. Classification cutoff:
10 uM (blocker = pIC50 >= 5.0; call = prob >= 0.5).
Sibling deployment repos (same training files, same gate): Delower/herg-xgboost-deployment,
Delower/herg-cardiogenai-deployment. The three models are used together for DPP-4 lead
screening: a lead flagged by all three is the highest-concern class.