Prediction of the skin sensitising potential and potency of compounds via mechanism-based binary and ternary classification models.

Skin sensitisation, one of the most frequent forms of human immune toxicity, is authenticated to be a significant endpoint in the field of drug discovery and cosmetics. Due to the drawbacks of traditional animal testing methods, in silico methods have advanced to study skin sensitisation. In this study, mechanism-based binary and ternary classification models were constructed with a comprehensive data set. 1007 compounds were collected to develop five series of local and global models based on mechanisms. In each series, compounds were classified into five groups according to EC3 values, and applied as training sets, test sets and external validation sets. For each of the five series, 81 binary classification models and 81 ternary classification models were acquired via 9 molecular fingerprints and 9 machine learning methods using a novel KNIME workflow. Meanwhile, the applicability domains for the best 10 models were figured out to certify the rationality of prediction effect. In addition, 8 toxic substructures probably causing skin sensitisation were identified to speculate whether a compound is a skin sensitiser. The mechanism-based prediction models and the toxic substructures can be applied to predict the skin sensitising potential and potency of compounds.