Development and multi-cohort evaluation of a 7-gene machine learning predictive signature for major depressive disorder: integrating systemic lupus erythematosus-associated immune signatures

Systemic lupus erythematosus (SLE) and major depressive disorder (MDD) share overlapping immune-inflammatory pathophysiology. This study aimed to identify shared genes dysregulated in both diseases, and construct an exploratory machine learning signature for general MDD based on these cross-disease transcripts. We integrated four public transcriptomic datasets (containing six sub-cohorts) from the GEO database for comprehensive bioinformatic analyses. A total of 107 unique combinations of 12 machine learning algorithms were systematically evaluated for feature selection and hyperparameter tuning, followed by exploratory evaluation in three non-training cohorts used in model selection. Functional enrichment, ssGSEA immune-related signature score analysis, and DGIdb drug-gene interaction prediction were additionally conducted to explore biological mechanisms and hypothesis-generating drug-gene associations. Twenty-nine shared differentially expressed genes were identified between the SLE and MDD cohorts, enriched in immune and inflammatory pathways. The Stepglm[both]+glmBoost integrated model, screened via systematic comparison of all algorithm combinations, further narrowed down to seven exploratory predictive genes (KLRB1, IFI27, CEACAM6, HP, C9orf66, WDR41, TSHZ3). This gene set achieved moderate discriminatory performance in the training cohort and three post-selection exploratory evaluation cohorts (training AUC = 0.792; exploratory evaluation AUCs = 0.675, 0.609, and 0.775) with variable calibration across cohorts. Notably, only CEACAM6 among the seven genes remained significant under FDR correction (|log2FC| > 0.25, adjusted p < 0.05), underscoring the exploratory nature of the signature. The 29-gene overlap and the seven-gene signature were derived using nominal p- value exploratory screening and should therefore be regarded as hypothesis-generating rather than statistically robust or biologically established markers. ssGSEA signature score analysis revealed distinct but overlapping immune-related patterns, and DGIdb analysis identified 8 drug-gene interaction pairs involving three of the exploratory predictive genes. Using shared immune genes of SLE and MDD, we built an MDD multi-gene predictive model through 107 machine learning approaches and screened 7 predictive genes with substantial contributions to model performance. These genes offer potential clues for further investigation into the shared immune mechanisms between SLE and MDD. Not applicable. This study utilized publicly available transcriptomic datasets from the GEO database and did not involve prospective clinical trial registration.