Application of a Deep Neural Network to Metabolomics Studies and Its Performance in Determining Important Variables

Deep neural networks (DNNs), which are kinds of the machine learning approaches, are powerful tools for analyzing big sets of data derived from biological and environmental systems. However, DNNs are not applicable to metabolomics studies because they have difficulty in identifying contribution factors, e.g., biomarkers, in constructed classification and regression models. In this paper, we describe an improved DNN-based analytical approach that incorporates an importance estimation for each variable using a mean decrease accuracy (MDA) calculation, which is based on a permutation algorithm; this approach is called DNN-MDA. The performance of the DNN-MDA approach was evaluated using a data set of metabolic profiles derived from yellowfin goby that lived in various rivers throughout Japan. Its performance was compared with that of conventional multivariate and machine learning methods, and the DNN-MDA approach was found to have the best classification accuracy (97.8%) among the examined methods. In addition to this, the DNN-MDA approach facilitated the identification of important variables such as trimethylamine N-oxide, inosinic acid, and glycine, which were characteristic metabolites that contributed to the discrimination of the geographical differences between fish caught in the Kanto region and those caught in other regions. As a result, the DNN-MDA approach is a useful and powerful tool for determining the geographical origins of specimens and identifying their biomarkers in metabolomics studies that are conducted in biological and environmental systems.