Computational in silico methods offer a powerful alternative to animal-based toxicity testing, which remains time-consuming, expensive, and ethically challenging. In practice, QSAR and read-across (RA) are among the most widely explored approaches, yet both suffer from important limitations, including end point-specific modeling, restricted applicability domains, nonquantitative or subjective analogue selection, and sensitivity to biases in chemical space coverage. To address these limitations, this study introduces ToxCML, a large-scale hybrid multifeature consensus quantitative RA Structure-Activity Relationship (mfCoQ-RASAR) platform that unifies consensus QSAR and similarity-based consensus RA into a weight-optimized workflow for predicting 18 toxicity end points across 54,601 unique chemicals. The framework combines multiple molecular representations (MACCS, Morgan, atom pair fingerprints, RDKit fingerprints, and physicochemical descriptors) with machine-learning-based QSAR and k-NN RA models and incorporates tiered applicability-domain analysis and chemical-space mapping that indicate broadly overlapping training-test distributions and in-domain coverage generally exceeding 95% across end points. Across all end points and evaluation settings involving unseen test sets or external validation sets, the hybrid mfCoQ-RASAR models achieve strong discrimination and accuracy (AUC approximately 0.86-0.99; BACC 0.73-0.98), with a consistent performance hierarchy in which mfCoQ-RASAR provides the highest or near-highest performance, consensus QSAR remains highly competitive, and consensus RA, while weaker, still offers informative predictive and analogue-based discriminatory capability. These results indicate that our framework delivers reliable, chemically well-contextualized, and broadly applicable multiend point toxicity predictions for unseen compounds and may support large-scale toxicity screening, hazard prioritization, and efforts to reduce animal testing in regulatory and industrial settings. The ToxCML platform is accessible through a public web server available at http://cardiosim.metaheart.kr:8080/.
Nursyafi et al. (Mon,) studied this question.