No abstract

Various toxicological profiles, such as genotoxic potential, need to be studied in drug discovery processes and submitted to the drug regulatory authorities for drug safety evaluation. As part of the effort for developing low cost and efficient adverse drug reaction testing tools, several statistical learning methods have been used for developing genotoxicity prediction systems with an accuracy of up to 73.8% for genotoxic (GT+) and 92.8% for nongenotoxic (GT-) agents. These systems have been developed and tested by using less than 400 known GT+ and GTagents, which is significantly less in number and diversity than the 860 GT+ and GT-agents known at present. There is a need to examine if a similar level of accuracy can be achieved for the more diverse set of molecules and to evaluate other statistical learning methods not yet applied to genotoxicity prediction. This work is intended for testing several statistical learning methods by using 860 GT+ and GT-agents, which include support vector machines (SVM), probabilistic neural network (PNN), k-nearest neighbor (k-NN), and C4.5 decision tree (DT). A feature selection method, recursive feature elimination, is used for selecting molecular descriptors relevant to genotoxicity study. The overall accuracies of SVM, k-NN, and PNN are comparable to and those of DT lower than the results from earlier studies, with SVM giving the highest accuracies of 77.8% for GT+ and 92.7% for GT-agents. Our study suggests that statistical learning methods, particularly SVM, k-NN, and PNN, are useful for facilitating the prediction of genotoxic potential of a diverse set of molecules.

Toxicity of various compounds has been measured in many studies by their toxic effects against Tetrahymena pyriformis. Efforts have also been made to use computational quantitative structure-activity relationship (QSAR) and statistical learning methods (SLMs) for predicting Tetrahymena pyriformis toxicity (TPT) at impressive accuracies. Because of the diversity of compounds and toxicity mechanisms, it is desirable to explore additional methods and to examine if these methods are applicable to more diverse sets of compounds. We tested several SLMs (logistic regression, C4.5 decision tree, k-nearest neighbor, probabilistic neural network, support vector machines) for their capability in predicting TPT by using 1129 compounds (841 TPT and 288 non-TPT agents) which are more diverse than those in other studies. A feature selection method was used for improving prediction performance and selecting molecular descriptors responsible for distinguishing TPT and non-TPT agents. The prediction accuracies are 86.9% approximately 94.2% for TPT and 71.2% approximately 87.5% for non-TPT agents based on 5-fold cross-validation studies, which are comparable to some of earlier studies despite the use of more diverse sets of compounds. The selected molecular descriptors are consistent with those used in other studies and experimental findings. These suggest that SLMs are useful for predicting TPT potential of diverse sets of compounds and for characterizing the molecular descriptors associated with TPT.