Immune checkpoint inhibitors (ICIs) show limited response rates in lung adenocarcinoma (LUAD). Mechanisms of immunotherapy resistance are incompletely understood, and robust predictive biomarkers are needed to improve patient selection and guide combination strategies.
The authors combined single-cell transcriptomics, multiple bulk RNA sequencing cohorts, and spatial transcriptomics to systematically identify an immunotherapy resistance–related gene signature (IRRG) and to build a prognostic risk score. Validation used six GEO datasets and previously published immunotherapy cohorts (OAK, POPLAR, and NG). Data sources included public repositories (Zenodo, Genome Sequence Archive, UCSC Xena, and NCBI GEO). Ethical approval for human tumor specimen collection was obtained from the Institutional Review Board of Sun Yat-sen Memorial Hospital (Approval No.: [SYSKY-2026-615-01]) and written informed consent was obtained from participants.
Single-cell RNA sequencing identified a malignant epithelial subpopulation designated Cluster 0 that was significantly enriched in patients who did not respond to neoadjuvant immunotherapy (stable disease). This cluster was defined by activation of proliferative transcriptional programs, specifically MYC Targets, E2F Targets, and G2M Checkpoint pathways, combined with a suppressed interferon response. Marker genes from this Cluster 0 predicted poorer prognosis across five independent cohorts, indicating that the transcriptional program of this malignant subset relates to adverse clinical outcomes.
Using marker genes associated with immunotherapy resistance, the investigators developed a SuperPC-based IRRG score to stratify prognosis. The IRRG score achieved robust prognostic separation in all six GEO validation cohorts analyzed and was reported to outperform 50 previously published signatures. High IRRG was associated with an immunosuppressive tumor microenvironment characterized by reduced infiltration of CD8+ T cells, natural killer (NK) cells, and tumor-infiltrating lymphocytes (TILs). These associations link the transcriptomic signature to a quantitatively and qualitatively altered immune milieu.
Within the IRRG, PSMB5 emerged as the central hub gene. PSMB5 showed the strongest adverse prognostic impact reported in the OAK cohort (hazard ratio, HR = 1.36) and in the TCGA cohort (HR = 1.54). Additionally, PSMB5 expression correlated negatively with CD8+ T cell infiltration (reported correlation r = –0.22). High PSMB5 expression consistently predicted worse overall survival (OS) and progression-free survival (PFS) across the OAK, POPLAR, and NG immunotherapy cohorts, supporting its role as a predictive marker for ICI outcomes.
Spatial transcriptomic analysis localized high PSMB5 expression to tumor-dense regions in patients with stable disease. Complementary multiplex immunofluorescence studies demonstrated spatial exclusion of CD8+ T cells from areas of high PSMB5 expression. Together these spatial data indicate that PSMB5 is associated not only with an immunosuppressive transcriptional program but also with physical segregation of effector T cells from tumor regions, a plausible mechanism of immune evasion and resistance to ICIs.
The IRRG score provides a transcriptomic classifier that robustly predicts prognosis across multiple cohorts and appears to outperform many existing signatures. Its hub gene, PSMB5, links prognostic risk to both quantitative decreases in CD8+ T cell infiltration and spatial exclusion of these cells within tumor tissue. These observations nominate PSMB5 as a potential predictive biomarker for immunotherapy response in LUAD and as a candidate target for combination strategies with PD-1 blockade to overcome immune exclusion. The source reports these findings across multiple independent cohorts, including TCGA and immunotherapy datasets OAK, POPLAR, and NG.
The study makes data sources explicit: single-cell and spatial data from the neoadjuvant immunotherapy cohort are available via Zenodo (DOI provided in the source); the HRA001130 dataset for PSMB5-stratified single-cell analysis is available in the Genome Sequence Archive (GSA). TCGA-LUAD transcriptomic and somatic copy number variation data were obtained via UCSC Xena with GTEx normal lung tissue integrated through the UCSC Xena GTEx/TCGA pipeline. Raw expression matrices for six GEO validation cohorts (GSE13213, GSE26939, GSE29016, GSE30219, GSE31210, and GSE42127) were downloaded from NCBI GEO. Data from OAK, POPLAR, and NG immunotherapy cohorts were obtained from previously published studies. The authors declare no competing interests and report that appropriate IRB approval and patient consent procedures were followed.
Notes on limitations and reporting
Details such as the full gene list composing the IRRG, algorithmic parameters for the SuperPC model, and numeric performance comparisons with the 50 published signatures are reported in the source but are not reproduced in full here. Readers should consult the original publication and supplementary materials for exact methods, full gene lists, and cohort-level statistics.