A public-data audit of transcriptomic prognostic transportability in R-CHOP-treated diffuse large B-cell lymphoma

Gene-expression biomarkers for diffuse large B-cell lymphoma (DLBCL) are often transported across retrospective cohorts that differ in treatment, assay platform, and clinical annotation. Using REMARK as the reporting framework, we audited prognostic associations with fixed scoring rules. Four Gene Expression Omnibus cohorts were harmonized for survival analysis. Treatment context was assessed after outcome-blind transcriptional state discovery. A directly mapped five-gene historical benchmark was evaluated with continuous Cox models in four R-CHOP cohorts. A 76-gene program selected in GSE31312 and GSE10846 served as a case study of two-cohort discovery, with GSE87371 as the only full-coverage independent evaluation and GSE32918 as a cross-platform sensitivity analysis. Coverage ablation, technical replicates, matched random programs, meta-analysis, and clinical-increment analyses assessed attenuation. The survival analyses included 470, 232, 87, and 140 patients. State composition was strongly associated with CHOP versus R-CHOP context in GSE10846 (Cramer’s V = 0.986). The historical score’s unadjusted random-effects hazard ratio (HR) was 1.35 (95% confidence interval 1.04–1.75); the prediction interval crossed the null (0.77–2.37). Common age/sex adjustment yielded HR 1.27 (1.05–1.54), with a prediction interval of 0.94–1.72. The 76-gene program was not independently validated in GSE87371 (unadjusted HR 1.23, 0.87–1.75; IPI/COO-adjusted HR 0.88, 0.56–1.40; age/sex-adjusted HR 1.09, 0.76–1.55). Its effect exceeded most matched random programs, but cross-cohort attenuation was pronounced and estimated clinical increment was negligible. With only 32 validation events, modest effects remain compatible with the data. Public DLBCL transcriptomic associations can be cohort-specific despite fixed scoring rules. Independent full-coverage evaluation and transparent treatment and platform audits should precede claims of general prognostic utility.