In silico Genotoxicity Prediction by Similarity Search and Machine Learning Algorithm: Optimization and Validation of the Method for High Energetic Materials

The European regulation REACh (Registration, Evaluation, Authorization, and restriction of Chemicals) has placed responsibility on the industry to manage the risk from chemicals since 2006. In order to ensure a high level of protection of human health and environment, toxicity prediction methods are now a widely used tool for regulatory decision making and selection of leads in new substances design. These in silico methods are an alternative to traditional in vitro and in vivo testing methods, which are laborious, time-consuming, highly expensive, and even involve animal welfare issues. Many computational methods have been employed to predict the toxicity profile of substances, but they are mostly adapted to pharmaceutical molecules and not to High Energetic Materials (HEMs).In line with these restrictions, ArianeGroup set up a collaborative project with the French CNRS to develop optimized tools for the prediction of HEM properties, such as genotoxicity. Several in silico methods can be used to predict the properties of molecules, such as QSAR, Local QSAR or Machine Learning. We already demonstrated that using Local QSAR allows for better predictions with a good reliability [1].We therefore developed a genotoxicity prediction tool based on the structural similarity search coupled with a supervised machine learning algorithm. This tool is composed of 3 predictive models: the Ames test, the Chromosomal Aberration test and the Mouse Lymphoma Assay. The aim of this paper is to evaluate the performance of these models to predict the genotoxicity of HEMs. We also present the methodology we applied to build these models and to optimize their performances. The dimensional reduction of the training set and the hyperparameters tuning of the different algorithms showed a performance acceleration and a significant reduction of the overfitting, which caused a decline in the generalization capacity of the predictive models. The performance of the predictive models was evaluated on a test set of HEMs and compared to the results of other prediction softwares.