The current research on quantitative structure-activity relationships (QSAR) of hazardous properties of compounds has significant progresses by utilizing intricate descriptors. Herein, to address the difficulties in descriptor selection and combination as well as the lack of interpretability of models, based on multi-endpoint toxicological data, an adaptive and interpretable modeling architecture is proposed to assist consensus prediction for hazardous properties of industrial chemicals. This architecture can adaptively adjust the modeling strategy according to the balanced degree of available data. After that, an integrated approach based on the concept of consensus is employed to improve the accuracy of predictions. Additionally, a novel interpretability method of feature perturbation is utilized to identify the active sites of compounds for computer-aided mechanism research. The statistical results demonstrate the outstanding performance and versatility of this structure, while the suggested interpretability strategy aids in the comprehension mechanism and enhances the dependability of the established model.