Comparative evaluation and selection of optimal QSAR-based machine learning model for liver toxicity prediction

Area of Science:

• Computational toxicology
• Medicinal chemistry
• Pharmacology

Background:

• Drug-induced liver toxicity is a major cause of acute liver failure.
• Current toxicity testing methods are costly, slow, and not scalable.
• Need for robust, interpretable computational tools for early hepatotoxicity prediction.

Purpose of the Study:

• To develop and validate a machine learning model for predicting drug-induced liver toxicity from molecular structure.
• To benchmark seven machine learning algorithms for hepatotoxicity prediction.
• To identify key structural features driving hepatotoxicity predictions.

Main Methods:

• Compiled a multi-source dataset of 6,219 compounds.
• Benchmarked seven machine learning algorithms using a unified preprocessing pipeline.
• Employed RDKIT for molecular feature calculation and Grid Search/Randomized Search for hyperparameter optimization.
• Utilized SHAP and LIME for model interpretability and applicability domain analysis.

Main Results:

• Achieved a test set ROC-AUC of 0.728, recall of 0.706, and F1-score of 0.700.
• External validation yielded an ROC-AUC of 0.741, demonstrating good generalization.
• Identified ECFP4 fingerprints and VSA-type descriptors as key predictors of hepatotoxicity.
• Model correctly identifies ~71% of hepatotoxic compounds before experimental testing.

Conclusions:

• Developed a reproducible, interpretable, and externally validated computational tool for early-stage hepatotoxicity screening.
• The model supports efficient and cost-effective prioritization of drug candidates.
• Identified structural features provide mechanistic insights into DILI mechanisms like lipophilic accumulation and CYP inhibition.