Prediction of Carcinogenicity of Diverse Chemical Substances by Support Vector Machine and Neural Network

Abstract

Among chemicals existing in environments, very few are identified as carcinogens. But from the viewpoint of the animal protection, it becomes a social, urgent problem to develop a technology to predict the carcinogenicity of the chemicals from the structures. Up to now, a lot of researches have been done, but it is extremely difficult to predict the carcinogenicity of non-congeners with various chemical structures. The large scale reliable data of the carcinogenicity of non-congener chemicals were collected from IARC, EU, EPA, NTP, ACGIH, and JSOH databases and evaluated into four ranks. For 846 chemicals so collected, 69 descriptors were generated by using Fujitsu CAChe ProjectLeader from their three dimension molecular structures. Relationships between carcinogenic ranks and descriptors were analyzed using support vector machine (SVM) and neural network (ANN) techniques. ANN showed a satisfactory predictability for screening test using animals, but took much computing time. On the other hand, SVM could predict within short time, but its predictability power, inferior to ANN, was considered to be improved by introducing topological descriptors. The prediction of the carcinogenicity of large-scale non-congener chemicals of this research is an extremely complex problem, so the high ability of ANN was proven. On the other hand, SVM has also several advantages out of ANN, so it is expected SVM and ANN are properly used for various nonlinear problems.