Identification of Diagnostic Markers for Major Depressive Disorder Using Machine Learning Methods

Abstract

Background Major depressive disorder (MDD) is a global health challenge that impacts the quality of patients’ lives severely. The disorder can manifest in many forms with different combinations of symptoms, which makes its clinical diagnosis difficult. Robust biomarkers are greatly needed to improve diagnosis and to understand the etiology of the disease. The main purpose of this study was to create a predictive model for MDD diagnosis based on peripheral blood transcriptomes. Materials and Methods We collected nine RNA expression datasets for MDD patients and healthy samples from the Gene Expression Omnibus database. After a series of quality control and heterogeneity tests, 302 samples from six studies were deemed suitable for the study. R package “MetaOmics” was applied for systematic meta-analysis of genome-wide expression data. Receiver operating characteristic (ROC) curve analysis was used to evaluate the diagnostic effectiveness of individual genes. To obtain a better diagnostic model, we also adopted the support vector machine (SVM), random forest (RF), k-nearest neighbors (kNN), and naive Bayesian (NB) tools for modeling, with the RF method being used for feature selection. Results Our analysis revealed six differentially expressed genes (AKR1C3, ARG1, KLRB1, MAFG, TPST1, and WWC3) with a false discovery rate (FDR) < 0.05 between MDD patients and control subjects. We then evaluated the diagnostic ability of these genes individually. With single gene prediction, we achieved a corresponding area under the curve (AUC) value of 0.63 ± 0.04, 0.67 ± 0.07, 0.70 ± 0.11, 0.64 ± 0.08, 0.68 ± 0.07, and 0.62 ± 0.09, respectively, for these genes. Next, we constructed the classifiers of SVM, RF, kNN, and NB with an AUC of 0.84 ± 0.09, 0.81 ± 0.10, 0.73 ± 0.11, and 0.83 ± 0.09, respectively, in validation datasets, suggesting that the SVM classifier might be superior for constructing an MDD diagnostic model. The final SVM classifier including 70 feature genes was capable of distinguishing MDD samples from healthy controls and yielded an AUC of 0.78 in an independent dataset. Conclusion This study provides new insights into potential biomarkers through meta-analysis of GEO data. Constructing different machine learning models based on these biomarkers could be a valuable approach for diagnosing MDD in clinical practice.