This study evaluated whether transcriptomic representations provide incremental predictive value for immunotherapy response in advanced urothelial carcinoma beyond clinical variables and established tumor biomarkers. Using data from 298 patients in the IMvigor210 trial of atezolizumab, the authors compared knowledge-guided tumor microenvironment gene-set features with data-driven transcriptome-wide gene selection. They report measurable AUPRC gains for knowledge-guided features over clinical models and demonstrate instability in data-driven gene selection despite pathway-level convergence.
The primary cohort comprised 298 patients with advanced urothelial carcinoma from the IMvigor210 trial, a clinical study of atezolizumab (PD-L1 inhibitor). An external validation cohort was obtained from the public Gene Expression Omnibus dataset GSE176307. The authors state that only de-identified, publicly available human data were used and that IMvigor210 data came from the IMvigor210CoreBiologies resource.
Two broad approaches to representing transcriptomic information were compared:
Knowledge-guided features: pre-specified gene-set representations focused on the tumor microenvironment and biologically annotated pathways.
Data-driven selection: transcriptome-wide selection of individual genes based on predictive modeling without a priori biological grouping.
The intent was to quantify incremental predictive value when adding transcriptomic representations to models containing clinical variables alone and to models that also included standard or research-enriched molecular biomarkers.
Performance was primarily reported using area under the precision-recall curve (AUPRC). Key reported findings include:
Knowledge-guided tumor microenvironment gene-set features improved AUPRC by 0.058 relative to a model containing only clinical variables.
When standard biomarkers (tumor mutational burden, TMB, and PD-L1) were included, the additional gain from knowledge-guided transcriptomic features was smaller: +0.011.
With research-enriched biomarkers added (neoantigen burden and immune phenotype), the incremental AUPRC gain from knowledge-guided transcriptomic features was +0.016.
Under the clinical-plus-biomarker baseline, the data-driven transcriptome-wide gene selection representation yielded an AUPRC 0.150 lower than the knowledge-guided representation.
These results indicate that biologically informed, gene-set–level transcriptomic representations produced greater predictive improvement than unstable, data-driven gene lists, particularly when added to clinical-only models.
The authors evaluated stability of genes selected by the data-driven approach across cross-validation folds. They report a mean Jaccard similarity of 0.227, indicating low overlap of selected genes between folds and therefore poor stability of individual-gene signatures.
Despite this instability at the gene level, pathway analysis found repeated enrichment of the IL-6/JAK/STAT3 signaling Hallmark pathway across the data-driven results. The recurrence of this pathway suggests that, although specific genes varied, there may be pathway-level biological convergence underlying associations with immunotherapy response.
External evaluation in the GSE176307 cohort replicated the overall pattern observed in the primary analysis. In that cohort, knowledge-guided features improved AUPRC by 0.128 over clinical variables alone and by 0.019 over the clinical-plus-biomarker baseline. The authors present these external results as supportive of the main conclusion that biologically informed transcriptomic representations can add predictive value beyond established clinical and tumor biomarkers.
In this analysis of advanced urothelial carcinoma patients treated with atezolizumab, knowledge-guided transcriptomic representations of the tumor microenvironment provided measurable incremental predictive value for immunotherapy response over clinical variables and modest additional gains when established biomarkers were already included. By contrast, data-driven transcriptome-wide gene selection produced unstable gene lists and inferior predictive performance under the clinical-plus-biomarker baseline, although pathway-level signals such as IL-6/JAK/STAT3 emerged recurrently.
The authors interpret these findings to support the use of biologically informed transcriptomic representations when molecular profiling is used to augment prediction beyond existing biomarkers. The results emphasize the trade-off between interpretability and reproducibility in transcriptomic modeling: gene-set or pathway-based features may offer more consistent, clinically actionable signals than unstable single-gene signatures.
The authors declared no competing interests. They confirmed adherence to relevant ethical guidelines and noted that all data used were publicly available and de-identified. Primary data sources named were the IMvigor210CoreBiologies resource for IMvigor210 and the Gene Expression Omnibus accession GSE176307 for external validation. The manuscript states that all data produced in the work are contained in the manuscript.