Komedojenite ve cilt irritasyon potensiyeli olan kozmetik bileşenlerinin in silico methodlar ile değerlendirilmesi

Comedogenicity and skin irritation are the common adverse reactions to cosmetic ingredients. While comedogenicity is the potential of a ingredient that induce formation of comedones which are the initial lesion for acne. Skin irritation is the formation of reversible damage to the skin after the application of a chemical for up to 4 hours. Before cosmetic testing on animals was prohibited in 2013, comedogenicity was tested on rabbits ears. However, there has been no full replacement of animal testing on cosmetics. For this reason, in this thesis study, we proposed to develop two different QSAR classification model to predict the comedogenic and skin irritaticy potential of cosmetic ingredients by using different machine learning algorithms and types of molecular descriptors. Our dataset was obtained from the literature, composed of 121 cosmetic ingredients mainly fatty acids, fatty alcohols and their derivatives and pigments, tested on rabbits for both comedogenicity and irritation. The total of 4837 molecular descriptors were calculated by the means of four different software. Descriptors for models were selected via WEKA by comparing correlation-based and learning-based (via decision tree algorithm) method. Modelling studies were performed with WEKA using various classifiers. The 10-fold cross-validaton (CV) was used to evaluate the performance of model, also predictivity of the model was evaluated on test set. All models were compared using classification accuracy, AUC, MCC and F-score, and then, the best model was chosen by this way. The result of QSAR modelling with random forest classifiers are promising for both comedogenicity and skin irritancy prediction. While for comedogenicity prediction, Alva model with 3 descriptor gave successful results with 86.96% accuracy for the 10-fold CV model and 79.31% accuracy for the test set. Skin irritation model, Mold2 with eight descriptors gave a successful performance with 85.75% accuracy for 10-fold CV model and 82.75% accuracy for the test set. In conclusion, this stuy provided an animal-friendly, rapidly perform and inexpensive two different classification model for the safety evaluation of cosmetics. Further studies are needed to develop their performance and to extend their chemical domain. Keywords: in silico, comedogenic, machine learning, skin irritation, QSAR, cosmetics