Immunobiology. 2026 May 12;231(3):153187. doi: 10.1016/j.imbio.2026.153187. Online ahead of print.
ABSTRACT
Despite the proliferation of prognostic gene signatures for glioma, clinical translation remains stalled by poor reproducibility and overfitting. In this study, we address this stability crisis by developing a robust “Dual-Signature Framework” using stability selection-a rigorous resampling method-rather than standard regression. Analyzing RNA-seq data from 1351 patients across the TCGA (n = 694) and CGGA (n = 657) cohorts, we constructed two distinct models. The primary 20-gene “Data-Driven” signature achieved superior predictive accuracy (C-index: 0.7392), significantly outperforming 14 published benchmark models and the current best single-gene predictor (HOXA5). In parallel, we derived a 7-gene “Biology-Driven” signature (including HOXA5, CHI3L1, MMP14) that retained 98% of the predictive power (C-index: 0.7252) while prioritizing mechanistic interpretability. Both models successfully stratified patients into distinct risk groups with high statistical significance (Log-rank p < 0.001) in external validation. Comprehensive subgroup analyses across 19 clinical and molecular subgroups demonstrated robust performance (C-index range: 0.59-0.85), with extended calibration analysis confirming excellent probability estimation (Brier score 0.20 for 5-year predictions). By integrating stability-driven feature selection with biological pathway constraints, this study provides a reproducible, high-performance alternative to unstable “black box” models, offering a translation-ready tool for personalized glioma risk assessment.
PMID:42134029 | DOI:10.1016/j.imbio.2026.153187
