Reliability-aware molecular screening for drug-induced liver injury in predictive toxicology under chemical scaffold shift
Abstract
IntroductionDrug-induced liver injury (DILI) remains a challenging endpoint in predictive toxicology because molecular datasets are limited, chemical scaffolds shift across data sources, and probability scores do not necessarily indicate prediction reliability. This study aimed to develop a reliability-aware molecular screening framework for DILI prediction under chemical scaffold shift.MethodsWe developed CC-SCAC-AFP, which combines an AttentiveFP ranking backbone with coverage-constrained split conformal prediction and adaptive novelty-conditioned calibration. The framework was evaluated on the Therapeutics Data Commons DILI task using leakage-audited Murcko-scaffold partitions, five seeds for model comparison, and 20 four-way scaffold splits for reliability analysis. DILIst and DILIrank 2.0 were reserved for strict external evaluation.ResultsAttentiveFP achieved an internal mean AUROC of 0.862 (standard deviation [SD], 0.023). Across 20 reliability splits at α = 0.10, CC-SCAC-AFP achieved empirical coverage of 0.949 (SD, 0.050), compared with 0.937 (SD, 0.054) for global split conformal prediction. The corresponding mean prediction-set sizes were 1.604 and 1.521, respectively, indicating that the additional coverage was largely attributable to increased conservatism rather than improved efficiency at matched set size. On external tests, AUROC was 0.598 for both DILIst and DILIrank 2.0. Coverage was 0.891 and 0.801, respectively, below the nominal 0.90 target. The DILIrank 2.0 Brier score of 0.348 was worse than that of a constant predictor based on the observed prevalence (0.249).DiscussionCC-SCAC-AFP provides a chemically indexed, set-valued reliability screening framework for early DILI molecular prioritization and makes the coverage–efficiency trade-off explicit. Two-label prediction sets identify compounds requiring further review. Because source-domain calibration did not preserve the nominal coverage target on the external datasets, recalibration in the target domain is necessary before applying the framework to a new chemical library.
Related
- Revealing the toxicological effects and mechanisms of tris(1-chloro-2-propyl) phosphate exposure on osteoarthritis based on network toxicology, machine learning, and molecular docking
- Investigating potential molecular mechanisms linking 6PPD-quinone exposure to Intervertebral Disc Degeneration: A network toxicology, molecular docking, and molecular dynamics simulation study
- Exploring the mechanism of 6PPD/6PPD-Q-induced allergic rhinitis based on network toxicology, molecular docking, and molecular dynamic simulation
- Network toxicology, molecular docking and in vitro validation reveal the molecular mechanisms of bisphenol A-induced MASLD
- Network toxicology, molecular docking and molecular dynamics simulations for bisphenol A neurotoxicity in depression pathogenesis
BibTeX
@article{jin2026reliabilityaware,
title = {{Reliability-aware molecular screening for drug-induced liver injury in predictive toxicology under chemical scaffold shift}},
author = {Zhi Jin and Hua Sun and Yiting Zhou and Qiyu Dong},
journal = {Frontiers in Pharmacology},
year = {2026},
month = nov,
doi = {10.3389/fphar.2026.1976307},
url = {https://www.frontiersin.org/articles/10.3389/fphar.2026.1976307},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - Reliability-aware molecular screening for drug-induced liver injury in predictive toxicology under chemical scaffold shift
AU - Zhi Jin
AU - Hua Sun
AU - Yiting Zhou
AU - Qiyu Dong
PY - 2026
DA - 2026/11/02
JO - Frontiers in Pharmacology
DO - 10.3389/fphar.2026.1976307
UR - https://www.frontiersin.org/articles/10.3389/fphar.2026.1976307
AB - IntroductionDrug-induced liver injury (DILI) remains a challenging endpoint in predictive toxicology because molecular datasets are limited, chemical scaffolds shift across data sources, and probability scores do not necessarily indicate prediction reliability. This study aimed to develop a reliability-aware molecular screening framework for DILI prediction under chemical scaffold shift.MethodsWe developed CC-SCAC-AFP, which combines an AttentiveFP ranking backbone with coverage-constrained split conformal prediction and adaptive novelty-conditioned calibration. The framework was evaluated on the Therapeutics Data Commons DILI task using leakage-audited Murcko-scaffold partitions, five seeds for model comparison, and 20 four-way scaffold splits for reliability analysis. DILIst and DILIrank 2.0 were reserved for strict external evaluation.ResultsAttentiveFP achieved an internal mean AUROC of 0.862 (standard deviation [SD], 0.023). Across 20 reliability splits at α = 0.10, CC-SCAC-AFP achieved empirical coverage of 0.949 (SD, 0.050), compared with 0.937 (SD, 0.054) for global split conformal prediction. The corresponding mean prediction-set sizes were 1.604 and 1.521, respectively, indicating that the additional coverage was largely attributable to increased conservatism rather than improved efficiency at matched set size. On external tests, AUROC was 0.598 for both DILIst and DILIrank 2.0. Coverage was 0.891 and 0.801, respectively, below the nominal 0.90 target. The DILIrank 2.0 Brier score of 0.348 was worse than that of a constant predictor based on the observed prevalence (0.249).DiscussionCC-SCAC-AFP provides a chemically indexed, set-valued reliability screening framework for early DILI molecular prioritization and makes the coverage–efficiency trade-off explicit. Two-label prediction sets identify compounds requiring further review. Because source-domain calibration did not preserve the nominal coverage target on the external datasets, recalibration in the target domain is necessary before applying the framework to a new chemical library.
SN - 1663-9812
ER -