Abstract
In order to develop a prediction system for the safety of chemicals, many attempts have been made by examining quantitative structure-activity relationships (QSAR). The results, however, were not always satisfactory enough in view of predictability when it is assumed that they are used in actual situations. In the present study, therefore, we have attempted to develop an in silico prediction system enabling the risk assessment of cosmetic raw materials by combining a molecular orbital calculation method and an artificial neural network system. Human patch test data on 161 samples were collected from a past publication and experiment results in our laboratory. Molecular weight, polarizabilityα, polarizabilityγ, dipole moment, and ionization potential were obtained from molecular orbital calculations as descriptors to predict the skin irritation. In addition, concentration and exposure time were added as descriptors. A neural network system was employed for the analysis. Consequently, by using leave-one-out cross-validation methods, it was shown that the neural network model can predict the positive rate in a human patch test with reasonable accuracy (root mean square error was 0.352). The above results suggest that their combinational use will enable us not only to predict the toxicological potential of cosmetic raw materials but also to make the risk assessment possible.
