{ "project": "hERG toxicity prediction - Phase 5 XGBoost (Approach A finals)", "version": "phase5-locked-pool-finals-2026-11-13", "models": { "cls_scaffold": { "file": "models/xgboostA_cls_base+physchem_extra+morgan_counts+ecfp6+fcfp4+herg+pharm+morgan1024+morgan1024_counts_scaffold.pkl", "task": "classification", "protocol": "scaffold", "headline": true }, "cls_random": { "file": "models/xgboostA_cls_base+physchem_extra+morgan_counts+ecfp6+fcfp4+herg+pharm+morgan1024+morgan1024_counts_random.pkl", "task": "classification", "protocol": "random" }, "reg_scaffold": { "file": "models/xgboostA_reg_base+morgan_counts+ecfp6+fcfp4+herg+pharm_scaffold.pkl", "task": "regression", "protocol": "scaffold", "headline": true }, "reg_random": { "file": "models/xgboostA_reg_base+morgan_counts+ecfp6+fcfp4+herg+pharm_random.pkl", "task": "regression", "protocol": "random" } }, "feature_blocks": { "classification": [ "morgan", "maccs", "desc", "physchem_extra", "morgan_counts", "ecfp6", "fcfp4", "herg", "pharm", "morgan1024", "morgan1024_counts" ], "regression": [ "morgan", "desc", "morgan_counts", "ecfp6", "fcfp4", "herg", "pharm" ], "excluded": "CDK/PaDEL block - dropped after the PaDEL build stalled; identical for all four models", "block_kinds": { "bits_raw": [ "morgan", "maccs", "ecfp6", "fcfp4", "morgan1024" ], "continuous_prepped": [ "desc", "physchem_extra", "herg" ], "counts_log1p_prepped": [ "morgan_counts", "pharm", "morgan1024_counts (classification only)" ] }, "block_definitions": "feature_extraction.py (RDKit only, end-to-end from SMILES)" }, "xgboost_params": { "n_estimators": 900, "learning_rate": 0.03, "max_depth": 8, "subsample": 0.8, "colsample_bytree": 0.8, "tree_method": "hist", "n_jobs": 2, "random_state": 42 }, "preprocessing": "Per continuous/count block, a ContPrep fitted on the protocol train+val keys ONLY: drop columns with variance <=1e-12 or >=5% NaN (judged on the fit fold), median-impute with fit-fold medians, standardise with fit-fold mean/SD (+1e-9). Count blocks are log1p-transformed (after clipping negatives to 0) before the same steps. Fitted states are stored inside each model pickle under 'preps'; predict_herg.py applies them exactly.", "thresholds": { "primary_cutoff_uM": 10, "primary_rule": "blocker if IC50 < 10 uM, i.e. pIC50 >= 5.0", "cls_call_rule": "prob_blocker >= 0.5", "reg_derived_call_rule": "pred_pIC50 >= 5.0", "secondary_cutoff_uM": 1 }, "training_data": { "classification_pool_compounds": 12702, "regression_pool_compounds": 8657, "sources": "ChEMBL + BindingDB + PubChem functional hERG IC50 (regression: functional IC50 only; Ki never pooled into the primary regression), canonical-deduplicated", "lockout": "1,738 unique compounds of every evaluation set (internal test splits, hERG-70, hERG-60, new-compounds slice, FDA 36 panel, regulatory-failure set) removed by InChIKey connectivity block before training; post-lock overlap verified = 0", "fit": "headline finals refit on protocol train+val; preprocessing fit on the same keys only" }, "pickle_format": "dict {model: XGBClassifier/XGBRegressor, task, blocks, params, preps, protocol, trained_on}" }