Classifying Schizophrenia Subjects by Fusing Networks From Single-Nucleotide Polymorphisms, DNA Methylation, and Functional Magnetic Resonance Imaging Data

To comprehensively employ complementary information from multiple types of data for better disease diagnosis, in this study, we applied a network fusion-based approach to three types of data including genetic, epigenetic, and neuroimaging data from a study of schizophrenia patients (SZ). A network is a map of interactions, which is helpful for investigating the connectivity of components or links between subunits. We exploit the potential of using networks as features for discriminating SZ from healthy controls. We first constructed a single network for each type of data. Then we built four fused networks by network fusion method: three fused networks for each combination of two types of data and one fused network for all of three data sets. Based on the local consistency of network, we can predict the group of SZ subjects with unknown labels. The group prediction method was applied to test the power of network-based features and the performance was evaluated by a tenfold cross-validation. The results show that the prediction accuracy is the highest when applying our prediction method to the fused network derived from three data types among all seven tested networks. As a conclusion, when making a diagnosis or predicting the labels of SZ subjects, we recommend more approaches that attempt to comprehensively utilize the multiple types data that are often available.