Transparent Machine Learning for Multi-Omics Analysis of Mental Disorders

Schizophrenia and bipolar disorder are two severe mental disorders that affect more than 65 million individuals worldwide. The aim of thisproject was to find co-prediction mechanisms for genes associated with schizophrenia and bipolar disorder using a multi-omics data set and a transparent machine learning approach. The overall purpose of theproject was to further understand the biological mechanisms of these complex disorders. In this work, publicly available multi-omics data collected from post-mortem brain tissue were used. The omics types included were gene expression, DNA methylation, and SNP array data. The data consisted of samples from individuals with schizophrenia, bipolar disorder, and healthy controls. Individuals with schizophrenia or bipolar disorder were considered as a combined CASE class. Using machine learning techniques, a multi-omics pipeline was developedto integrate these data in a manner such that all types were adequately represented. A feature selection was performed on methylation and SNP data, where the most important sites were estimated and mapped to their corresponding genes. Next, those genes were intersected with the gene expression data, and another feature selection was performed on the gene expression data. The most important genes were used to develop an interpretable rule-based model with an accuracy of 88%. The model wasthen visualized as a network. The graph highlighted genes that may be of biological importance, including CACNG8, RTN4, TERT, OSBPL8, and ANTXR1. Moreover, strong co-predictions were found, most notable between CNKSR4 and KDM4C in CASE samples. However, further investigations would need to be performed in order to prove that these are real biological interactions. Through the methods developed and the results found in this project, we hope to shed new light towards analyzing multi-omics data as well as to reveal more about the underlying mechanisms of psychiatric disorders