Machine learning-guided identification and experimental validation of novel transcriptomic biomarkers for schizophrenia

Background Schizophrenia (SCZ) is a severe neurodevelopmental psychiatric disorder characterized by marked clinical heterogeneity, making objective diagnosis challenging. This study aimed to identify characteristic genes associated with SCZ using machine learning approaches and to evaluate their diagnostic potential while assessing model robustness and cross-dataset generalizability. Methods Five SCZ-related gene expression datasets were obtained from the Gene Expression Omnibus (GEO) database. Four datasets were integrated as a discovery cohort, whereas GSE62191 was retained for independent external validation. Batch effects were assessed before and after ComBat correction using principal component analysis and PERMANOVA. Differentially expressed genes (DEGs) were identified, followed by Gene Ontology and Kyoto Encyclopedia of Genes and Genomes enrichment analyses. Random Forest (RF) analysis was used to identify characteristic genes, and an Artificial Neural Network (ANN)-based molecular diagnostic model was constructed. Additional analyses included leakage-controlled repeated cross-validation, feature-selection stability, permutation importance, leave-one-dataset-out validation, comparison of binary and continuous-expression representations, and age- and sex-adjusted sensitivity analyses. Eight representative genes were further validated by RT-qPCR in peripheral blood mononuclear cells from 20 schizophrenia patients and 20 healthy controls. Results A total of 1,893 DEGs were identified, including 809 upregulated and 1,084 downregulated genes. RF analysis identified 20 characteristic genes associated with SCZ. Functional analyses suggested involvement in inflammation, autophagy, oxidative stress, and metabolic dysregulation. Batch-associated variation was markedly reduced after correction (PERMANOVA R², 0.980 to 0.033). The original ANN achieved AUCs of 0.975, 0.659, and 0.696 in the training, validation, and test cohorts, respectively, whereas leakage-controlled repeated cross-validation yielded a mean AUC of 0.695 ± 0.032. Leave-one-dataset-out AUCs ranged from 0.614 to 0.818, indicating study-level heterogeneity. The molecular diagnostic model achieved AUC values of 0.782 in the training cohort and 0.721 in the external validation cohort. Continuous-expression modeling showed similar cross-validated performance but better external performance than the binary representation. RT-qPCR confirmed the differential expression of BID , ABCG2 , NEUROD6 , CRHBP , BAG3 , TOLLIP , HSD11B1 , and ETV5 , consistent with the transcriptomic findings. Conclusions This study identified candidate transcriptomic biomarkers associated with SCZ and established a machine learning-based diagnostic framework. However, leakage-controlled and study-level validation demonstrated only modest and heterogeneous predictive performance, indicating that the identified gene panel should be considered exploratory rather than a clinically established diagnostic signature. Experimental validation supported the biological and translational relevance of the identified biomarkers, which require further confirmation in larger, independent, multicenter cohorts.