Predicting Features of Human Mental Disorders through Methylation Profile and Machine Learning Models

Abstract

Listen

Translate article

In this research, two types of mental disorders were studied: Major Depression Disorder (Depression) and Bipolar disorder (BP). We train the data on both binary/multi-class classification models through machine learning approaches and fuzzy clustering using statistical analysis as a probabilistic reference for the potential users to identify whether or not they potentially have certain types of mental disorders. In order to achieve our goal, we first gather the ICD-10 MDD 58 Depression and Control samples (GSE41826) and 72 BP and Control subjects (GSE129428). All groups, for both case and control groups, are collected by Infinium Methylation EPIC Bead Chip, where each sample consists of around 440,000 target CpGs over 23 chromosomes. we applied statistical procedures and identified differentially methylated loci (DMLs) that have the most impact on diagnosis through DNA methylation. Furthermore, we use Principal Component Analysis (PCA) to project the selected 6000 CpGs into 50 Principal Components (PCs). Based on the top three PCs, we then build and test several binary/multi classification models. These advances can provide insights for the development of new methods that can diagnose multiple mental diseases at one time. we construct a Phonogram, denoting the overlapping CpG islands that the two diseases have in common from the predicting features for the diagnosis of diseases. Also, disease ontology analysis was conducted on our data to build up these relationships through web-based DO Analysis tools, Visualization, and Integrated Discovery (DAVID). With the accuracy rates returned from the above models, we can prove the meaningfulness of our models in practical use and relate these features of the models with other diseases that have already been proven to have relationships with certain types of genes referenced by CpG islands.