Improved prognostic survival models for pediatric medulloblastoma using high dimensional gene expression data.
AI interpretation is pending for this paper.
Open original publication →What the AI sees
Not AI summarized yet.
Research significance
Pending deeper interpretation.
Source abstract
Genetic, epigenetic, and transcriptomic analyses have stratified medulloblastoma (MB) into four canonical subgroups of Wingless Type (WNT), Sonic Hedgehog (SHH), and Group 3 and Group 4, with distinct patient profiles and prognoses. Recent classification strategies have also considered combining Group 3 and Group 4 tumors into a Non-WNT/Non-SHH subgroup to account for biological overlap and heterogeneity. Using high-dimensional gene expression data from 487 pediatric and young adult patients and over twenty-one thousand transcripts, this study explores which genes can improve prognostic accuracy for survival while accounting for molecular stratification, histological subtype, key oncogenic drivers (MYC and MYCN amplification), and established clinical covariates, including age group (< 3 vs. 3-21 years) and metastatic status. We then develop a multi-stage framework for identifying prognostic genes and evaluating modern survival modeling strategies. In the first stage, gene screening was performed using Benjamini-Hochberg adjusted Cox regression across false discovery rate (FDR) thresholds from 1% to 6%, with the number of retained genes increasing from 15 at 1% to 146 at 6% FDR. In the second stage, multiple survival models were evaluated, including LASSO, Elastic Net, Ridge regression, SCAD, MCP, PCA-Cox, and Random Survival Forests, using ten-fold cross-validation with the Integrated Brier Score as the primary calibration metric and the concordance index as a secondary discrimination measure. Although Ridge regression achieved the lowest prediction error at higher FDR thresholds, it did not perform variable selection and retained large gene sets, limiting interpretability. In contrast, the 6% FDR Elastic Net model provided an optimal balance between predictive accuracy and model sparsity while reducing the gene set from 146 to 49 genes, yielding an interpretable final multivariable model. Gene-level effects from the final Elastic Net-penalized Cox model revealed a clear prognostic gradient. Genes associated with poorer survival included FKBP4, CSNK2A2, GPC4, GATA3, NPY, LYPD1, CLCA4, and BNC2, which have been implicated in tumor progression, signaling pathways, and immune-related processes, whereas genes associated with improved survival included ZNF774, COX10, FBLIM1, and UNC13C, reflecting roles in cellular regulation and protective biological processes. These findings demonstrate that combining FDR-based screening with Elastic Net-penalized Cox modeling yields a robust, parsimonious, and biologically meaningful prognostic framework for medulloblastoma, achieving strong predictive performance while maintaining interpretability in high-dimensional genomic settings.