Abstract Animal experiments traditionally identify sensitizers in cosmetic materials. However, with growing concerns over animal ethics and bans on such experiments globally, alternative methods like machine learning are gaining prominence for their efficiency and cost‐effectiveness. In this study, to develop a robust sensitizer detector model, we first constructed benchmark data sets using data from previous studies and a public database, then 589 sensitizers and 831 nonsensitizers were collected. In addition, a graph‐based autoencoder and Mondrian conformal prediction (MCP) were combined to build a robust sensitizer detector, iSKIN. In the independent test set, the Matthews correlation coefficient (MCC) and the area under the receiver operating characteristic curve (ROCAUC) values of the iSKIN model without MCP were 0.472 and 0.804, respectively, which are higher than those of the three baseline models. When setting the significance level in MCP at 0.7, the MCC and ROCAUC values of iSKIN could achieve 0.753 and 0.927, respectively. Regrouping experiments proved that the MCP method is robust in the improvement of model performance. Through key structure analysis, seven key substructures in sensitizers were identified to guide cosmetic material design. Notably, long chains with halogen atoms and phenyl groups with two chlorine atoms at ortho‐positions were potential sensitizers. Finally, a user‐friendly web tool ( http://www.iskin.work/ ) of the iSKIN model was deployed to be used by other researchers. In summary, the proposed iSKIN model has achieved state‐of‐the‐art performance so far, which can contribute to the safety evaluation of cosmetic raw materials and provide a reference for the chemical structure design of these materials.
