Evaluation of Kernels Applied in Support Vector Machines in the Data Analysis of Organochlorines Exposure in Study of Biomarkers

Abstract

One of the biggest concerns in metabolomic data analysis is adopting good predictive models in cases of asymmetric databases and few samples. As a result, more studies should evaluate statistical methods that can produce adequate analyses despite those limitations. Support vector machines (SVMs) can be an alternative. This study assesses the behavior of SVMs and the application of different kernels in order to propose a strategy for biomarker recognition in cases of organochlorine exposure. The database used in this study comes from an original experiment previously published, in which a HepG2 cell line was exposed to four different organochlorine pesticides, a mixture, and a control. The database consists of 153 identified and 246 unidentified metabolites. First, a partial least squares discriminant analysis (PLS-DA) was applied to the database to obtain the top ten candidate biomarker metabolites. Subsequently, an SVMs algorithm was implemented using four different kernels: linear, sigmoid, polynomial, and radial basis function (RBF). Afterward, a methodology based on recursive feature elimination (SVM-RFE) was adopted in order to obtain the top ten metabolites. Finally, PLS-DA and SVMs were compared. It was demonstrated that SVMs have a good predictive power when they are trained with few samples. The sigmoid and linear kernels with hard and soft margin showed the best performance. Moreover, the SVM-RFE strategy can rapidly identify candidate biomarker metabolites using a limited number of training samples.