Developing a multi-task quantitative structure-activity relationship (QSAR) model for predicting the toxicity...

E-cigarette emissions contain a complex mixture of active ingredients, solvents, flavors, thermal degradation products, and contaminants. Despite recent efforts, comprehensive toxicity data remain unavailable for the majority of these compounds, posing a significant challenge for hazard identification and chemical prioritization. To address this gap, we developed interpretable machine learning (ML)-based quantitative structure-activity relationship (QSAR) models to predict e-cigarette toxicity for five endpoints: eye irritation, skin irritation, skin sensitization, carcinogenicity, and genotoxicity. We implemented both single- and multi-task learning strategies, using a single-task QSAR model as baseline benchmarks, to systematically evaluate the added value of multi-task architectures for jointly predicting related toxicity endpoints. The developed multi-task QSAR models achieved moderate to high performance for eye and skin toxicity (AUC ≥ 0.69, Accuracy ≥ 61 %) and carcinogenicity and genotoxicity (AUC ≥ 0.66, Accuracy ≥ 61 %). Certain endpoints, such as eye irritation, improved substantially compared to the best-performing single-task model (AUC +13 %, Accuracy +20 %), while others, like carcinogenicity, saw reduced performance (AUC -13 %, Accuracy -5 %). Feature importance analysis revealed that nitrogen-containing groups (e.g., amines and amides) were influential predictors of eye and skin toxicity, whereas cancer-related endpoints were highly dependent on the electrostatic and electronegative properties of compounds. Overall, this study demonstrates how multi-task QSAR modeling can be used not only to predict the toxicity of understudied e-cigarette constituents but also to elucidate relationships among toxicity endpoints, thereby supporting model-informed chemical hazard evaluation for e-cigarette products.