Paired CYP1A2/3A4-HepG2 phenotyping supports machine-learning classification of literature-curated hepatotoxicity categories.

Limited CYP1A2/3A4 activity in wild-type HepG2 cells can obscure hepatotoxic responses that require metabolic activation. We established matched CYP1A2/3A4-co-expressing and empty-vector (EV) HepG2 cells and profiled 81 compounds with six high-content imaging readouts and three extracellular injury biomarkers. Absolute responses and paired CYP/EV differences were encoded as 54 compound-level features for random forest (RF) classification of three literature-curated study labels: CYP1A2/3A4 metabolic activation-associated hepatotoxic compounds (MAHC), CYP1A2/3A4-independent hepatotoxic compounds (IDHC), and non-hepatotoxic compounds (NHC). In 62 model-development compounds (26 MAHC, 12 IDHC, 24 NHC), the primary nested analysis yielded macro-F1 values of 0.959 for three-class classification and 0.911 for MAHC-IDHC; across 10 nested-cross-validation runs, means were 0.945 ± 0.014 and 0.920 ± 0.030. Feature-set ablation showed that paired CYP/EV information was nonredundant: neither EV-only (macro-F1 0.572) nor CYP-only (0.367) reproduced MAHC-IDHC discrimination, which the differential features alone retained. The final RF classified all 19 same-laboratory held-out compounds under common Cmax-based dosing (accuracy 1.00; 95% CI, 0.82-1.00), although only three were IDHC. Zebrafish CYP modulation produced responses consistent with the assigned labels in the three compounds tested. These data support matched CYP/EV phenotyping for classifying these literature-curated categories.