HOXC4 as a machine learning-derived immune biomarker for predicting treatment and prognosis in prostate cancer patients

Background: Prostate cancer (PCa) remains a prevalent male malignancy, imposing significant health burdens and socio-economic challenges globally. While early diagnosis is critical for effective intervention, current clinical strategies lack precision in risk stratification and treatment response prediction. This study aimed to identify key machine learning-derived biomarkers for PCa and to comprehensively evaluate the diagnostic, prognostic, and immune-related significance of HOXC4.

Methods: In this study, we integrated 258 samples from four Gene Expression Omnibus (GEO) datasets to identify differentially expressed genes (DEGs) between tumor and normal tissues. Robust feature selection was performed using the intersection of least absolute shrinkage and selection operator (LASSO) regression, random forest, and support vector machine recursive feature elimination (SVM-RFE) algorithms. Ten machine learning models were constructed, with the optimal model interpreted using SHapley Additive exPlanations (SHAP) analysis. These findings were further integrated with gene set enrichment analysis (GSEA), CIBERSORT-based immune infiltration analysis, independent external validation in the Memorial Sloan Kettering Cancer Center (MSKCC) cohort, and pan-cancer analysis.

Results: A total of 123 DEGs were identified, from which five key feature genes (RAB17, FAM107A, COL9A1, HOXC4, and TRPM4) were selected. Among the ten models, Partial Least Squares (PLS) exhibited superior diagnostic performance, with an area under the receiver operating characteristic curve (AUC) of 0.977. SHAP interpretation highlighted HOXC4 as a positive risk-associated feature and prognostic candidate. High HOXC4 expression was associated with poor progression-free survival (P=0.02), and this prognostic association was further validated in the independent MSKCC cohort. GSEA indicated that HOXC4-high phenotypes were enriched in cellular energy metabolism and protein processing pathways, whereas HOXC4-low phenotypes showed immune-related enrichment patterns. Immune infiltration analysis suggested that high HOXC4 expression was associated with an immunosuppressive tumor microenvironment characterized by increased regulatory T cell and M2 macrophage infiltration. HOXC4 also showed significant correlations with several immune checkpoint molecules, including CD40 and SIGLEC-15. Furthermore, pan-cancer analysis validated the heterogeneous expression patterns of HOXC4 across diverse tumor types.

Conclusions: Identified via interpretable machine learning, HOXC4 serves as a promising diagnostic and prognostic biomarker linked to immune suppression and metabolic reprogramming. This study highlights the potential of combining machine learning-driven feature selection with biological validation to optimize personalized management in PCa.