Daniel Liu, Jimmy Kuo, Yao-Chong Wu, Shang-Yu Hsieh, An Rong Huang, Chorng-Horng Lin
Neuroblastoma (NB) exhibits profound clinical and molecular heterogeneity, highlighting the need for reproducible biomarkers for disease stratification. This study aimed to identify robust stage-associated gene signatures through integrative machine learning (ML) and cross-platform validation. Four independent NB microarray datasets were integrated to perform a multi-cohort meta-analysis. Differentially expressed genes (DEGs) were identified using Linear Models for Microarray Data (limma), followed by feature selection through three orthogonal ML models - Random Forest (RF), eXtreme Gradient Boosting (XGBoost), and Neural Network (NN). Overlapping markers among methods were used to define a consensus gene set, which was then compared with gene expression data from the Therapeutically Applicable Research to Generate Effective Treatments - Neuroblastoma (TARGET-NBL) cohort for clinical relevance assessment. We identified a seven-gene consensus set (BIRC5, CCDC127, AW473883, BRD7, TCF7L2, GABARAPL1, and CD1E) through ML, within which TCF7L2 emerged as a critical intersection between statistical and ML frameworks. While the large-scale 632-DEG set provided the broadest prognostic coverage, the refined seven-gene signature maintained robust clinical relevance. Notably, three mapped core genes (CD1E, GABARAPL1, and TCF7L2) were independently validated as significant favorable survival-associated markers (p < 0.05) in the TARGET cohort, even after rigorous clinical adjustment for age and COG risk status. By bridging transcriptomic discovery with cross-platform validation, this study establishes a high-confidence consensus gene signature for NB. Our findings suggest that while high-dimensional signatures offer extensive predictive power, the consensus-based core genes - particularly TCF7L2 - represent reproducible candidate biomarkers associated with disease progression and patient outcome.