Kidney toxicity β€” Arch5 feature-block ensemble deployment model

Binary classifier for kidney toxicity (nephrotoxicity) concern: the probability that a compound belongs to the toxic class. This is the deployment refit of the kidney programme's Arch5 feature-block ensemble, one of the two winners selected on the locked regulatory benchmark. It is deployed as a pair with the Arch3 deployment model (Delower/kidney-arch3-bagging-deployment); a compound flagged by both models is the highest-concern readout.

Recipe

The locked-wave winning recipe, reused unchanged with no re-tuning (tuned hyperparameters taken verbatim from the locked arm). Three feature-block components, blended with validation-MCC-proportional weights re-derived by the arm's own rule on an internal seed-42 scaffold split of the deployment pool; final components and preprocessing then refit on the full pool:

  • A β€” LightGBM on Morgan fingerprint bits (radius 2, 1024): num_leaves=63, learning_rate=0.06, n_estimators=300, min_child_samples=20, subsample=1.0, colsample_bytree=0.8, reg_lambda=1.0. Weight 0.1473.
  • B β€” ExtraTrees (600 trees, max_features=0.5, class_weight="balanced") on prepared RDKit descriptors (211 of 217 kept by the arm's DescPrep) + MACCS keys (167) + Morgan count fingerprint (1024) β€” 1,402 columns. Weight 0.3739.
  • C β€” LightGBM on ChemBERTa-77M embeddings (768; seyonec/ChemBERTa-zinc-base-v1, mean-pooled, pretrained encoder only β€” benchmark compounds were inference-only for the encoder)
    • prepared physchem descriptors (13): num_leaves=31, learning_rate=0.06, n_estimators=600, min_child_samples=10, subsample=0.8, colsample_bytree=1.0, reg_lambda=5.0. Weight 0.4788.

Flag threshold: probability β‰₯ 0.5 (the programme convention).

Deployment training set

All labeled kidney data the programme holds, assembled by InChIKey connectivity dedup of the standardized parent with benchmark labels governing conflicts: 1,805 unique compounds (915 positive / 890 negative) = corrected regulatory benchmarks (169 kept) β†’ v1 benchmarks (0 new β€” the corrected set is a superset) β†’ internal test (327) β†’ enriched training pool (1,309) β†’ locked training set (0 new β€” fully covered). 1,307 duplicate rows dropped; 0 label conflicts. In short: locked training + internal test + enriched additions + every regulatory benchmark compound absorbed.

Hard gate (passed inside the training scripts before any training): every regulatory benchmark compound verified compound by compound inside the actual training matrix β€” corrected benchmark: withdrawn 16/16, clinical failures 7/7, PASS 146/146; original v1 benchmark: withdrawn 12/12, clinical failures 5/5, PASS 146/146; missing = 0 in every set. One exception, stated explicitly: each withdrawn file contains one further substance (17th corrected / 13th v1) with no resolvable structure β€” it cannot enter any feature matrix, was left unscored by every arm in every wave, and is excluded from the gate denominators for that reason. See gate_verification.json.

Honest benchmark scores β€” benchmark version only

The deployment refit absorbed the regulatory benchmark and is never scored on it. The frozen claim for this recipe comes from the locked-wave benchmark model (benchmark locked out of training, lockout verified 0; v1 pooled basis, 163 scored compounds: 12 withdrawn + 5 clinical failures + 146 PASS):

Metric Value
Pooled regulatory AUROC 0.701
Pooled balanced accuracy 0.521
Withdrawn drugs caught 12/12
Clinical failures caught 5/5
PASS compounds correct 6/146 (0.041)

Read the PASS row honestly: the benchmark version ranks toxic above clean well (AUROC 0.701) but its probabilities sit almost entirely above the 0.5 line β€” as a standalone 0.5-threshold screen it is nearly uninformative on clean compounds. It is a ranking model first; the dual flag with Arch3 is the decision readout.

Deployment screen results

Applied to the DPP-4 programme's potent leads (flag = prob β‰₯ 0.5):

  • 1,113 workbook leads β€” scored by both models 1,113/1,113: this model flagged 668 (60.0%), Arch3 flagged 626 (56.2%), flagged by both 419, by either 875, pass both 238.
  • 4,898 full potent union β€” scored by both models 4,898/4,898: this model flagged 3,219 (65.7%), Arch3 flagged 3,041 (62.1%), flagged by both 2,195, by either 4,065, pass both 833.

Flag rates of 56–66% are this programme's known high-sensitivity / low-specificity profile, stated plainly: every kidney arm trades specificity for withdrawn-drug recall, so a single-model flag is low-information on its own; the dual-flag column is the readout.

Files

  • model.joblib β€” joblib bundle: the three fitted components (LightGBM / ExtraTrees / LightGBM), the ensemble weights, and the two fitted DescPrep preprocessing objects (descriptors and physchem). Unpickling requires deployment_common.py (shipped here) to be importable.
  • config.json β€” recipe, feature blocks, training composition, gate counts, benchmark scores.
  • deployment_common.py β€” the programme's shared featurization module (fingerprints, descriptors, DescPrep).
  • predict.py β€” prediction entry point (predict(smiles) / predict_many([...])); featurizes exactly as the deployment screen did, including fresh ChemBERTa-77M embeddings.
  • validate.py + examples.json β€” runnable validation: reproduces this model's probabilities on 8 screened DPP-4 leads (tolerance 1e-6; the model is deterministic). Verified locally against this exact layout before upload: max deviation 0.
  • gate_verification.json, training_manifest.json β€” hard-gate result and full training manifest.
  • absorbed_compounds.csv β€” the 1,805-compound deployment training set (the absorbed data, for audit).

Limitations

Requires scikit-learn, lightgbm, rdkit, joblib, numpy, torch, transformers (component C downloads the ChemBERTa-77M encoder from the HF hub on first use). Probabilities are screening scores, not calibrated clinical risks.

Downloads last month
-
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support