Development of machine learning-based multi-task quantitative structure-activity relationship models for predicting toxicities in six human organ systems.

Traditional toxicity assessment relies heavily on animal testing, particularly for chemicals lacking toxicity data. This study developed machine learning (ML)-driven quantitative structure-activity relationship (QSAR) models to predict human organ-specific toxicities, including cardiotoxicity, developmental toxicity, hepatotoxicity, neurotoxicity, nephrotoxicity, and reproductive toxicity. We collected in vivo data for 2,389 chemicals and Tox21 high-throughput screening data for 1,746 chemicals, resulting in 1,743 chemicals with matched datasets. Eighty-eight ML-based QSAR models were developed using three feature scenarios: (1) Tox21 data alone, (2) molecular descriptors alone, and (3) combined features. Five descriptor types and four ML algorithms (random forests, decision trees, support vector machines, and deep neural network [DNN]) were applied, with and without chi-square-based feature selection. Performance was evaluated using nested cross-validation and five metrics (recall, precision, balanced accuracy, F1 score, and ROC-AUC). DNN models in Scenario 2 performed best for developmental and neurotoxicity, while those in Scenario 3 outperformed others for the remaining toxicities. ROC-AUC values approached 0.8 across endpoints, and models without feature selection generally performed better. SHAP and contribution maps enhanced interpretability, highlighting key structural features of toxicity. This study demonstrates the potential of ML-assisted QSAR models for accurate multi-organ toxicity prediction, supporting drug development and chemical risk assessment.