Metabolic biomarkers of treatment-resistant schizophrenia and cognitive correlates: insights from a machine learning.

Early identification of Treatment-resistant schizophrenia (TRS) remains a challenge. Metabolomics offers a promising strategy for identifying biomarkers and uncovering the metabolic alterations underlying TRS. Machine learning (ML), increasingly applied in omics research, enables the development of predictive models from complex datasets. This study aimed to develop a metabolomics-based ML model for TRS identification and to examine its association with cognitive impairments. 112 TRS patients, 232 non-TRS patients, and 145 sex- and age-matched healthy controls were recruited in this study. Psychiatric symptoms were assessed using the Positive and Negative Syndrome Scale (PANSS), and cognitive function was evaluated with the Repeatable Battery for the Assessment of Neuropsychological Status (RBANS) and the Toronto Alexithymia Scale (TAS-26). Fasting blood samples were collected from all participants and analyzed by UHPLC-HRMS, enabling the quantification and profiling of 699 metabolites. The resulting data were then subjected to orthogonal partial least squares discriminant analysis (OPLS-DA), and a predictive model was constructed using a random forest algorithm, which was validated through 5 × 5 nested cross-validation with inner-loop LASSO feature selection. From 699 metabolites, 58 metabolites exhibited significant changes (VIP > 1.5, FDR p < 0.05). 5 × 5 nested cross-validation combining an inner-loop LASSO feature selection model yielded a pooled out-of-fold AUC of 0.958 (95% CI (DeLong): 0.940-0.977) for TRS classification. Further correlation analyses revealed distinct associations between the top five predictive metabolites and cognitive domains in TRS versus non-TRS patients. Gamma-glutamylthreonine (gamma-Glu-Thr) showed a significant negative correlation with visuospatial cognition in TRS patients (r = -0.25, p = 0.009). In contrast, it showed a positive correlation in the non-TRS group (r = 0.16, p = 0.015). However, these associations did not remain significant after FDR correction. This study established an RF-metabolomic model for identifying TRS and provided specific metabolic signatures associated with cognitive dysfunction.