DPP-4 pIC50 β€” ExtraTrees (DEPLOYMENT REFIT)

DEPLOYMENT REFIT β€” trained on ALL labeled data including the regulatory set (absorbed 2026-11-14 by Hossain's decision). Benchmark metrics do NOT apply to this version; it has not been and must not be evaluated on the absorbed sets. The benchmark version remains Delower/dpp4-pic50-extra-trees.

What this is

The benchmark ExtraTrees pipeline (LazyPredict ColumnTransformer: SimpleImputer(mean) + StandardScaler over feature_0..feature_1029 β†’ ExtraTreesRegressor, 100 trees, random_state 42; features = Morgan fingerprint radius 2, 1024 bits + 6 physchem descriptors: MolWt, MolLogP, TPSA, NumHDonors, NumHAcceptors, NumRotatableBonds) refit with identical features and hyperparameters on 100% of the labeled data.

Recipe fidelity was verified before refitting: a clone of the benchmark pipeline fit on the benchmark's own training split (8,068 rows; train_test_split test_size=0.20, random_state 42, remainder split 50/50) reproduced the benchmark's published held-out test RMSE of 0.444906 exactly (difference 0.00e+00). That reproduction was a recipe check on the benchmark's own split β€” it is not a metric for this refit.

Training data

  • Benchmark version: 8,068 training rows (train split only).
  • This refit: 10,096 rows = 10,085 curated DPP-4 rows (combined_curated_smiles.csv, all splits reunited) + 11 regulatory (FDA) gliptins appended with their experimental wet-lab pIC50 values from DPP4_FDA_regulatory_master.csv.
  • Absorbed regulatory compounds (11): Linagliptin (pIC50 9.85), Teneligliptin (9.40), Evogliptin (9.18), Saxagliptin (8.82), Trelagliptin (8.89), Omarigliptin (8.80), Anagliptin (8.42), Alogliptin (8.12), Gemigliptin (7.80), Sitagliptin (7.74), Vildagliptin (7.47).
  • Gosogliptin was excluded: it has no experimental IC50/pIC50 in the regulatory master table.
  • Hard-gate verification (Hossain's rule): all 11 regulatory compounds were verified present in the actual training file by full InChIKey β€” 11/11 present, 0 missing (see training_manifest.csv and gate_verification.json). The fitted trees' root sample count is 10,096, matching the training file.

Important caveat (verified 2026-11-14)

Identity checks during this build showed that the curated DPP-4 set already contained all 11 gliptins by connectivity β€” and 9 of the 11 by full InChIKey (Linagliptin 1 row, Teneligliptin 1, Evogliptin 1, Omarigliptin 1, Anagliptin 2, Alogliptin 1, Gemigliptin 1, Sitagliptin 1, Vildagliptin 1; Saxagliptin and Trelagliptin connectivity-only, 5 and 1 stereovariant rows) β€” via its ChEMBL/parquet sources, carrying different assay values than the regulatory wet-lab values. Absorption therefore adds the 11 regulatory master rows with their wet-lab pIC50 values; it does not introduce previously unseen compounds. Per-drug counts are in training_summary.json and gate_verification.json.

Files

  • model.pkl β€” refit sklearn pipeline (54.7 MB)
  • infer.py β€” featurisation + prediction helpers (includes the _RemainderColsList unpickling shim for newer sklearn)
  • config.json β€” model configuration
  • training_summary.json β€” row counts, absorbed list, smoke check
  • absorbed_compounds.csv β€” the 11 regulatory rows absorbed
  • training_manifest.csv β€” the exact 10,096-row training file (SMILES, pIC50, source, InChIKey)
  • gate_verification.json β€” hard-gate InChIKey verification result: PASS

Smoke check (training compounds only)

Predictions on three curated training rows were finite and sane (pIC50 6.92 vs exp 6.36; 4.22 vs 4.22; 7.11 vs 7.11). No accuracy on the absorbed regulatory set is reported anywhere for this version, by design.

Downloads last month
34
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support