Support Vector Machines for Biomarkers Detection in in vitro and in vivo Experiments of Organochlorines Exposure

Abstract

Metabolomic studies generate large amounts of data, whose complexity increases if they are derived from in vivo experiments. As a result, analysis methods highly used in metabolomics, such as Partial Least Squares Discriminant Analysis (PLS-DA), can have particular difficulties with this type of data. However, there is evidence that indicates that Support Vector Machines (SVMs) can better deal with complex data. On the other hand, chronic exposure to organochlorines is a public health problem. It has been associated with diseases such as cancer. Therefore, its identification is relevant to reduce their impact on human health. This study explores the performance of SVMs in classifying metabolic profiles and identifying relevant metabolites in studies of exposure to organochlorines. For this purpose, two experiments were conducted: in the first one, organochlorine exposure was evaluated in HepG2 cells; and, in the second one, it was evaluated in serum samples of agricultural workers exposed to pesticides. The performance of SVMs was compared with that of PLS-DA. Four kernel functions were assessed in SVMs, and the accuracy of both methods was evaluated using a k-fold cross-validation test. In order to identify the most relevant metabolites, Recursive Feature Elimination (RFE) was used in SVMs and Variable Importance in Projection (VIP) in PLS-DA. The results show that SVMs exhibit a higher percentage of accuracy with fewer training samples and better performance in classifying the samples from the exposed agricultural workers. Finally, a workflow based on SVMs for the identification of biomarkers in samples with high biological complexity is proposed.