HumanToxPredict: An interpretable data-driven framework for human oral TDLo prediction and toxicity assessment of pharmaceuticals and bioactive compounds.

Abstract

Predicting the human toxic dose low (TDLo) is critically important for pharmaceutical safety assessment and regulatory toxicology. A global computational toxicology framework that integrates toxicity data for children, infants, men, and women was developed using indicator-variable-based demographic encoding. Initially, Partial Least Squares (PLS)-based quantitative structure-activity relationship (QSAR) and quantitative read-across structure-activity relationship (q-RASAR) models were developed using interpretable molecular descriptors and similarity-based descriptors. A structure-level group-wise partitioning strategy was adopted. The q-RASAR model demonstrated improved predictive performance compared with the corresponding QSAR model. Additionally, Random Forest (RF) and Support Vector Machine (SVM) models were explored to investigate non-linear relationships. To further improve practical applicability, classification models were also developed, with RF showing the strongest overall performance among the evaluated models based on multi-metric analysis. Model interpretation suggests that pTDLo values are influenced by substituted nitrogen environments, topological arrangements, and similarity-driven toxicological information. The final model was further applied to 3660 investigational drugs, with most compounds falling within the applicability domain. We also developed a user-friendly software tool, HumanToxPredict, to enable rapid toxicity prediction. Overall, the present study introduces an interpretable, population-integrated human toxicity prediction framework that integrates q-RASAR, machine learning, and large-scale chemical screening.