Related Experiment Videos
Cross-Platform External Validation of a Locked 12-Gene Prognostic Panel in Lung Adenocarcinoma Benchmarked Against
1Department of Medical Laboratories, College of Applied Medical Sciences, Qassim University, Buraydah 51452, Saudi Arabia.
Background/Objectives:
A prognostic gene panel is clinically useful only if its risk ranking survives transfer across measurement platforms and adds information to pathological stage. We evaluated the cross-platform transportability of a 12-gene panel selected in TCGA-LUAD and fixed before external scoring, separating panel performance from model choice and benchmarking it against stage.
Methods:
Selection used 504 TCGA patients and genes measurable on all validation platforms. Five molecular survival learners received identical gene inputs; age-and-sex and stage-based Cox models served as clinical comparators. External evaluation included 1066 patients in GSE72094, GSE68465 and GSE31210. Discrimination, paired model differences, prognostic value beyond stage, calibration and selection stability were assessed.
Results:
SurvivalPFN achieved Harrell C-indices of 0.631 (95% CI 0.572-0.689), 0.583 (0.541-0.624) and 0.665 (0.581-0.750), respectively, and regularized Cox regression on the same genes performed comparably (paired differences 0.005-0.011). The panel score was associated with survival independently of stage, with hazard ratios per standard deviation of 1.36 (95% CI 1.12-1.65) in GSE72094 and 1.56 (1.11-2.17) in GSE31210. Stage alone, however, discriminated at least as well as the panel alone in every cohort. Adding the panel to stage gave the highest C-indices observed in these cohorts, 0.663 and 0.718, against 0.634 and 0.678 for stage alone, but neither gain was statistically significant. Panel-defined risk groups separated overall survival in all three cohorts (hazard ratios 2.39, 1.43 and 3.65). Predicted 3-year risk matched observed risk in both primary cohorts (observed-to-predicted ratios 1.02 and 1.00), whereas TCGA-derived conformal correction improved calibration only in GSE68465. Results were consistent across learners and in sensitivity analyses of cohort overlap, histology and gene expression level.
Conclusions:
The fixed panel carried prognostic information that transported across RNA-sequencing and microarray platforms and complemented pathological stage, but it did not outperform stage. Its prospective evaluation should therefore be as an adjunct to stage, with local recalibration.