Serum Protein Glycopatterns as Biomarkers for Machine Learning Diagnosis of Major Depressive Disorder

The diagnosis of major depressive disorder (MDD) currently relies on subjective clinical assessment, underscoring the need for objective biomarkers. Glycosylation, a common post-translational modification, is involved in neuroinflammation and immune regulation, both implicated in MDD pathogenesis. This study aimed to identify serum protein glycopatterns as biomarkers for MDD diagnosis and severity stratification. Serum samples from 150 individuals, including healthy volunteers (HV, n = 38), mild-to-moderate MDD (M-MDD, n = 52), and severe MDD (S-MDD, n = 60), were analyzed using lectin microarrays. Thirty-six lectins showed significantly altered binding signals that indicating widespread glycosylation changes between HV, M-MDD, and S-MDD. These features were used to train seven machine-learning models, among which K-nearest neighbours (KNN) performed best, achieving 95.3% ± 2.7% accuracy and an AUC (Micro) of 0.990 ± 0.010 in a nested 5-fold cross-validation framework. In a blinded cohort (n = 71), the KNN model showed accuracies of 0.831, 0.662, and 0.831 for MDD, M-MDD, and S-MDD, respectively. SHAP analysis identified key lectins contributing to classification. These results demonstrated that serum glycopatterns combined with machine learning provided a promising objective tool for MDD diagnosis and severity assessment.