Work

Independent prognostic value of a 4-gene transcriptomic signature in lung adenocarcinoma across five external cohorts

Year: 2026
Type: preprint
Abstract: Background. Transcriptomic prognostic signatures for lung adenocarcinoma are frequently reported with in-sample discrimination and without independent external validation, and are prone to feature-selection leakage and undisclosed model selection. We asked whether a pre-specified, leakage-free signature adds prognostic value beyond standard-of-care staging, and whether it transports to independent cohorts. Methods. In 504 TCGA lung adenocarcinoma patients (182 deaths), eight survival models were evaluated by repeated 5-fold nested cross-validation, with feature selection strictly inside folds and all metrics computed out-of-fold. Five classical and tree-based learners formed a confirmatory tier eligible for deployment; three neural models formed a pre-declared exploratory tier; a rank-average ensemble was the pre-specified primary model. A 4-gene signature (DKK1, MS4A1, CCL20, and one long non-coding RNA) was locked and combined with stage, age and sex in a cross-validated integrated model, then applied unchanged to five independent cohorts (1,254 patients, 475 events) and pooled by meta-analysis. Discrimination, calibration, decision-curve analysis, and reclassification were reported with 95% confidence intervals. Results. Internal discrimination was modest (confirmatory-tier Harrell's concordance 0.607–0.623; deployed Cox 0.613, 0.596–0.627); two exploratory neural models did not exceed 0.60. The signature was independently prognostic after adjustment for stage, age and sex (hazard ratio 1.34 per standard deviation, 1.15–1.56; likelihood-ratio p = 2 x 10 -4 ), robust to multiple imputation and to exclusion of squamous-confounded samples. Combining the molecular score with staging raised discrimination from 0.645 to 0.676, a leakage-free gain of 0.031. Proportional hazards were violated for the risk-group comparison, where high-risk patients lost 293 days (169–418) of restricted mean survival over 5 years. External discrimination was homogeneous but modest (pooled Harrell's concordance 0.606, 0.578–0.634; I 2 ), competitive with a published 14-gene signature (0.616) at one-third the gene count. All 25 pre-specified reporting-guideline items were met. Conclusions. A pre-specified, leakage-free 4-gene signature carries independent, modest prognostic value beyond stage, age and sex, and its combination with staging (Harrell's concordance 0.676) is the predictor we advance. Molecular-signature discrimination alone (approximately 0.60–0.62) is not clinically actionable, and external transport, though homogeneous, is modest and driven partly by one weak, event-dominant cohort. These are prognostic, not predictive, associations from a single retrospective cohort; prospective validation is required before clinical application.
Language: English
Subfield: Oncology
Field: Medicine
Open Access status: green
OpenAlex ID: W7172415111
Cites: 69
Cited by:
Related to: 10