Glycosylation signature to predict response to immune checkpoint blockade in lung cancer.

Abstract

272 Background: Immune checkpoint blockade (ICB) has improved long-term survival in non-small cell lung cancer (NSCLC); however, most patients do not achieve durable benefit. Current biomarkers, including PD-L1 immunohistochemistry and tumor mutation burden, provide limited predictive accuracy. Although transcriptomic signatures of immune infiltration and exclusion correlate with response, they largely focus on the tumor microenvironment. We hypothesized that integrating cancer-intrinsic transcriptional programs with immune contextures would better capture tumor-immune interactions and improve predictions of ICB response. Methods: We assembled transcriptomic and clinical data of 818 tumors across nine ICB-treated NSCLC cohorts, including an in-house sequenced cohort of 51 late stage patients, generating one the largest response-annotated whole-transcriptome NSCLC datasets to date. We systematically identified NSCLC-specific transcriptional programs associated with ICB response, and characterized their clinical and genomic correlates in > 1,100 TCGA NSCLC tumors. We developed a multivariate predictive model integrating cancer-intrinsic and immune features, and evaluated performance across five independent validation cohorts, benchmarking against established biomarkers. Results: We uncovered a glycosylation-associated transcriptional program linked to ICB resistance, characterized by low tumor mutation burden, reduced PD-L1 expression, and decreased T-cell infiltration, consistent with an immune-cold tumor state. Tumors with high glycosylation additionally upregulated alternative inhibitory checkpoint ligands, suggesting reliance on non-PD-L1 immune evasion mechanisms to maintain immune exclusion. Integrating glycosylation and immune infiltration features, we developed IOSelect-lung , which robustly stratified ICB outcomes. Across five independent test cohorts, IOSelect-lung achieved a mean AUC of 0.77, significantly outperforming existing signatures (AUC 0.39-0.65). The model was predictive in the ICB arm but not in the chemotherapy control arm, indicating treatment-specific predictive value rather than general prognostic association. In joint analyses incorporating clinicopathological and hematological variables, IOSelect-lung demonstrated predictive power beyond established clinical and molecular biomarkers. Conclusions: A cancer-intrinsic glycosylation program defines an immune-cold, ICB-resistant subset of NSCLC. We introduced IOSelect-lung , a parsimonious model that reflects NSCLC-specific tumor biology and provides treatment specific predictive value beyond current biomarkers. These findings highlight glycosylation pathways and alternative checkpoints as potential therapeutic vulnerabilities, and support the development of cancer-type-specific biomarkers to improve patient stratification for ICB.