Evaluating a Glioma Transcriptomic Signature Against a Clinical Reference Model and a Random-Signature Null Distribution: A Leakage-Controlled Internal Audit and a Survey of the Field

Vollständiger Abstract

Worum geht es in dieser Arbeit?

Objective: The glioma prognostic literature contains a large number of transcriptomic risk signatures, yet whether these signatures add measurable information beyond a clinical model containing grade and molecular markers is rarely tested. We pursued three objectives jointly: to evaluate a leakage-controlled signature in terms of both discrimination and calibration; to test it against a clinical reference model and against a null distribution of random signatures; and to quantify the reporting practice of the field. Methods: The CGGA mRNAseq_693 cohort (n = 404, 209 deaths) served as development, the CGGA mRNAseq_325 cohort (n = 222, 138 deaths) as independent internal validation, and the TCGA lower-grade glioma and glioblastoma cohorts (n = 664, 247 deaths) as external validation. A univariate Cox score test was applied to 17,544 genes with false discovery rate control by the Benjamini–Hochberg procedure, and an elastic-net penalised Cox model was fitted on the top 200 candidate genes. Every model-building operation-imputation, scaling, candidate pool, and penalty selection was confined to the development cohort, and the final model was locked. Three arms were compared: clinical only, transcriptomic only, and combined. Calibration was quantified by the integrated calibration index derived from a smoothed calibration curve. In addition, the abstract-level reporting content of 515 glioma signature records indexed in Web of Science was analysed. Results: The signature made no measurable contribution beyond the clinical reference model. In the independent internal validation cohort the clinical model reached a concordance index of 0.800 (95% CI 0.767–0.831) and the signature 0.801 (0.769–0.832); the paired bootstrap difference was indistinguishable from zero (Δ = +0.001; 95% CI −0.032 to +0.035), and restricting the clinical model to variables known at diagnosis did not change this (Δ = +0.008; −0.024 to +0.043). Against a null distribution of 1000 random 20-gene sets drawn from the same candidate pool, the signature exceeded the null internally (p = 0.015) but was indistinguishable from it in external validation (p = 0.154); 99.2% of random sets reached a concordance index above 0.75 and 82.2% above 0.80 in the external cohort (Figure 3). Discrimination fell markedly within the IDH-wildtype (0.629) and WHO grade IV (0.606) strata, and the risk score correlated at 0.722 with a proliferation metagene despite containing no canonical proliferation gene. The concordance indices themselves—0.801 internally and 0.824 externally for the 17-gene subset available in TCGA, with an integrated calibration index of 0.045 at 36 months—are therefore best read as an illustration of the problem rather than as evidence of clinical utility: they sit squarely inside the range that random gene sets reach in the same data, and they fall in the range routinely presented as successful in the published literature. In the literature survey, 8.7% of the 515 records mentioned calibration, 4.5% decision curve analysis, and only 1.0% any comparison against a clinical reference model; none reported all three. Conclusions: A signature developed under a leakage-controlled protocol and well calibrated was nevertheless indistinguishable from the appropriate references on two of the three criteria we propose. Most of its discrimination rests on the IDH and grade axis that a broad range of prognostic gene sets can capture. Concordance indices reported in the glioma signature literature cannot be interpreted without a clinical reference model and a random-signature null distribution.