Identifying Molecular Toxicity Patterns Through Interpretable Learning of Chemical Properties

Abstract

The prediction of molecular toxicity is becoming a more and more important requirement, and computational methods are becoming more and more powerful in their ability to describe complex relationships between molecular structure and biological activity. This study presents an interpretable machine-learning framework to extract the toxicity-related molecular patterns from the Tox21 dataset based on the physicochemical and structural properties. Logistic Regression, Support Vector Machine, Random Forest, and XGBoost were tested on 12 nuclear receptor and stress-response endpoints with the use of molecular descriptors and Morgan fingerprints as complementary representations. Random Forest had the highest overall performance with the highest values of ROC-AUC = 0.852, PR-AUC = 0.526, F1 score = 0.664, balanced accuracy = 0.793 and Matthews correlation coefficient = 0.579. NR-AR-LBD (ROC-AUC = 0.91), SR-MMP (0.89) and NR-AhR (0.88) showed highest endpoint specific performance. The hybrid representation (physicochemical descriptors + Morgan fingerprints) performed better as a whole than either of the two feature sets used. LogP, molecular weight, topological polar surface area, aromatic ring count and molecular refractivity were found to be the most influential predictors by SHAP analysis. Performing cross-endpoints analysis, a different physicochemical signature was also identified between the nuclear receptor pathway and the stress-response pathway. The results indicate that interpretable machine learning can help go beyond just prediction and discover nonlinear, chemically meaningful and pathway-specific patterns. The proposed framework provides a clear foundation for computational toxicology, chemical prioritization and mechanism-oriented toxicity screening.