A Transcriptome-Wide Analysis Nominates DIO3OS and C9orf139 as Survival-Associated Long Non-Coding RNAs in Glioblastoma

Abstract

Background: Long non-coding RNAs (lncRNAs) are widely proposed as determinants of temozolomide response in glioblastoma (GBM), but many nominations rest on analyst-derived survival endpoints, screens unadjusted for O6-methylguanine-DNA methyltransferase (MGMT) promoter methylation, and models evaluated with the same data used to select them. We asked which lncRNA–outcome associations remain when endpoint definition, confounder adjustment, and external validation are addressed explicitly. Methods: We analyzed 94 IDH-wildtype, temozolomide-treated primary GBM patients from The Cancer Genome Atlas (TCGA), using the rule-based progression-free interval (PFI) of the TCGA Pan-Cancer Clinical Data Resource and overall survival (OS) as separate endpoints. A joint transcriptome-wide discovery for OS was performed across TCGA and the identically processed CPTAC-3 cohort (n = 282, 209 deaths), the latter unselected for IDH status because IDH annotation is unavailable for CPTAC-3, with adjustment for MGMT promoter methylation status, cellular composition, and expression subtype. Nominated transcripts were tested for external validation in two independent CGGA cohorts, and incremental discrimination was assessed as the optimism-corrected change in C-index (ΔC-index), benchmarked against the gain expected from a transcriptome-wide search alone. Results: No lncRNA was associated with PFI at a transcriptome-wide false-discovery rate (FDR) below 0.05. The joint OS discovery nominated DIO3OS, the only transcript reaching transcriptome-wide significance (hazard ratio 1.38 per standard deviation, 95% confidence interval 1.20–1.58, FDR 0.022), and C9orf139 (FDR 0.065); both were independently significant in each discovery cohort and robust to adjustment for MGMT status, cellular composition, and expression subtype. Both, however, were null with reversed point estimates in the two CGGA cohorts, were unassociated with PFI, and added only modest optimism-corrected discrimination (ΔC-index 0.07–0.11), close to what a transcriptome-wide search alone generates. Replication was thus confined to two cohorts sharing a processing pipeline. Conclusions: We nominate DIO3OS and C9orf139 as candidate survival-associated lncRNAs in GBM, requiring prospective evaluation in cohorts with complete IDH annotation and progression endpoints, and show how endpoint definition, selection optimism, and external validation determine which lncRNA associations survive scrutiny.