Accurate toxicity assessment is essential for chemical safety, but experimental testing is costly, slow, and ethically constrained, motivating the adoption of computational approaches such as QSAR. However, many QSAR models still rely on a single descriptor type or algorithm and relatively small, endpoint‑specific datasets, which limits robustness, interpretability, and broader applicability. To overcome these limitations, this study develops an interpretable multi‑modality consensus QSAR framework that integrates multiple molecular representations with both machine learning and deep learning to predict eight mechanistically diverse toxicity endpoints-skin sensitization, respiratory toxicity, AMES mutagenicity, hepatotoxicity, developmental toxicity, cardiotoxicity, drug‑induced nephrotoxicity, and neurotoxicity-across 30,160 unique compounds. Models for each descriptor-algorithm combination were optimized using 10-fold cross-validation, and top-performing models were combined into weighted multi-modality consensus predictors based on cross-validated AUC weights. Across all endpoints, multi-modality consensus models consistently achieved moderate to excellent performance on unseen and external sets (AUC 0.80-0.99, BACC 0.76-0.90). DeLong's test confirmed that multi-modality consensus models outperformed the best individual models, with significant AUC improvements (p < 0.05) in 7 of 8 endpoints. Chemical space and applicability domain analyses demonstrated broad coverage of diverse compounds, with in-domain predictions consistently achieving higher and more stable performance, while SHAP analysis highlighted global features and mechanistic-level motifs, supporting the model's biological plausibility and reliability. Overall, this multi-modality consensus framework provides a reliability aware and interpretable approach for broad spectrum toxicity prediction and multi-endpoint chemical safety assessment.
Nursyafi et al. (Mon,) studied this question.