Manifold Learning Analysis for Allele-Skewed DNA Modification SNPs for Psychiatric Disorders

Abstract

Bipolar disorder (BPD) and schizophrenia (SCZ) are two severe worldwide psychiatric disorders. Identifying genetic components contributing to both disorders will provide meaningful insights into their pathogenesis and widely-existed misdiagnosis. In this study, we employ allele-skewed DNA modification (ASM-SNP) data to investigate the two psychiatric disorders via state-of-the-art manifold learning, data-driven feature selection, and novel pathway analysis. We propose a novel manifold learning analysis for ASM-SNP data of bipolar disorder and schizophrenia based on a data-driven feature selection algorithm: nonnegative singular value approximation (NSVA). Our results indicate that t-distributed stochastic neighbor embedding (t-SNE) outperforms its peers in distinguishing psychiatric disorder samples from normal ones in both visualization and phenotype classification. It achieves the best phenotype diagnosis results with the average AUC 0.95 by using only about 20% top-ranked SNPs. Furthermore, our results from manifold learning along with support vector machine analysis suggest that the possible non-separability of SCZ and BPD in genetics. We also validate that SCZ and BPD both share the same or similar genetic variations from pathway analysis. This study indicates the inevitable misdiagnosis issue between BPD and SCZ from a machine learning and systems biology approach. The result sheds light on the existing psychiatry research to reexamine the existing behavior-based classification for BPD and SCZ. To the best of our knowledge, this study is the first comprehensive investigation of BPD and SCZ in bioinformatics.