Predicting neuroblastoma prognosis using machine learning analysis of metabolism-related gene network perturbation

Neuroblastomas (NB) are highly heterogeneous pediatric extracranial solid tumors in children with variable epigenetic, biological, and clinical characteristics. However, predictive models that can accurately classify patient risks and predict prognoses are currently limited. We analyzed a metabolism-related genes network perturbation using machine learning algorithms to construct a model to assess the risk and prognosis of patients with NB. Metabolism-related gene expression data for patients with NB were obtained from the Gene Expression Omnibus (GEO) (GSE49710, N = 498), ArrayExpress (E-MTAB-8248, N = 228), and TARGET (TARGET-NBL, N = 150) databases. An individual-specific gene interaction perturbation network was constructed using the Reactome Pathway Database. Unsupervised clusters and principal components analysis were analyzed using the R package “Consensus Cluster Plus”. Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) enrichment analyses were conducted using the R package “Cluster Profiler”. Gene set variation analysis (GSVA) and single-sample gene set enrichment analysis (ssGSEA) were performed to evaluate immune cell infiltration in patients with NB using the HALLMARK database. Ten machine learning methods (Lasso, Enet, plsRcox, CoxBoost, StepCox, gradient boosting machine [GBM], Ridge, random survival forests [RSF], survival-support vector machine [SVM], and super principal component [PC]) were used with 110 machine learning algorithms to screen for a metabolism-related signature to predict NB prognosis in GSE49710 and other cohorts (E-MTAB-8248 and GSE76427). Finally, Immunohistochemical staining (IHC) was performed to validate the UCK2 expression levels in human tissues. Gene network perturbation analysis of 948 metabolism-related genes revealed distinct discriminability, and patients with NB were classified into three differentiated subtypes. Kaplan-Meier survival analysis revealed that the prognoses were best for patients with subtype C3, followed by those with subtypes C1 and C2. Correlation analysis of clinical information indicated that subtypes C2 and C3 were associated with higher and lower percentages of high-degree malignancies, respectively, whereas C3 subtype showed a lower percentage of high-degree malignancies. The KRAS and myogenesis pathways were upregulated, and the levels of MYC targets were downregulated in patients with subtype C3; those with C2 subtype exhibited opposite trends. Patients with C3- and C2-subtypes exhibited immune-activated and immune-suppressed phenotypes, respectively. The combination of the StepCox [forward] and RSF algorithms provided the most accurate prognostic predictions for patients with NB. The importance score was highest for UCK2 among all subtypes. IHC staining further confirmed that UCK2 expression was substantially higher in the tissues of patients with NB than in those of controls. The machine learning-based prognostic prediction model that analyzes metabolism-related gene interaction perturbation networks supports the development of personalized management strategies for patients with NB.