Machine learning reduced gene/non-coding RNA features that classify Schizophrenia patients accurately and highlight insightful gene clusters
AbstractSchizophrenia (SCZ) is a chronic and severely disabling neurodevelopmental disorder that affects people worldwide. RNA-seq has been a powerful method to detect the differentially expressed genes/non-coding RNAs in patients; however, due to overfitting problems differentially expressed targets (DETs) cannot be used properly as biomarkers. In this study, dorsolateral prefrontal cortex (dlpfc) RNA-seq data from 254 individuals’ was obtained from the CommonMind consortium and analyzed with machine learning methods, including random forest, forward feature selection (ffs), and factor analysis, to reduce the numbers of gene/non-coding RNA feature vectors to overcome overfitting problem and explore involved functional clusters. In 2-fold shuffle testing, the average predictive accuracy for SCZ patients was 67% based on coding genes, and the 96% based on long non-coding RNAs (lncRNAs). Coding genes were further clustered into 14 factors and lncRNAs were clustered into 45 factors to represent the underlying features. The largest contribution factor for coding genes contains number of genes critical in neurodevelopment and previously reported in relation with various brain disorders. Genomic loci of lncRNAs were more insightful, enriched for genes critical in synapse function (p=7.3E-3), cell junction (p=0.017), neuron differentiation (p=8.3E-3), phosphorylation (8.2E-4), and involving the Wnt signaling pathway (p=0.029). Taken together, machine learning is a powerful algorithm to reduce functional biomarkers in SCZ patients. The lncRNAs capture the characteristics of SCZ tissue more accurately than mRNA as the formers regulate every level of gene expression, not limited to mRNA levels.