Answers are grounded in this paper and cite their sources.
While immune checkpoint inhibitors (ICIs) are adopted as standard therapy for advanced non-small cell lung cancer (NSCLC), genetic determinants of response heterogeneity remain elusive [1]. As most hematopoietic lineages undergo dynamic changes during tumor pathogenesis and immunotherapy [2], elucidating how germline variants modulate their transcriptomes is critical. Expression quantitative trait loci (eQTL) analysis, especially integrated with single-cell RNA sequencing (scRNA-seq), enables gene regulation mapping at single-cell resolution [3, 4]. Detailed methodologies are described in the Supplementary Materials. To investigate how germline variants shape immune gene regulation during ICI treatment, we performed single-cell-eQTL (sc-eQTL) analysis and transcriptomic network profiling. Peripheral blood mononuclear cells (PBMCs) were collected from 73 NSCLC patients treated with anti-programmed cell death protein-1 (PD-1) or programmed death-ligand 1 (PD-L1) therapy, at both baseline and 1-5 weeks post-treatment (Figure 1A, Supplementary Table S1). By integrating scRNA-seq with SNP array data, we analyzed cell-type-resolved sc-eQTLs and gene networks (Figure 1A-B). The sc-eQTL-regulated transcriptomic network explains ICI outcomes in NSCLC patients. (A) Experimental and analytical schemes. Using sc-eQTL mapping and gene network analysis in PBMCs of 73 ICI-treated NSCLC patients at baseline and treatment, we evaluated lung cancer- and treatment-specific eQTLs, transcriptomic network, clinical relevance based on GNN-refined gene sets, and the location of eQTLs within the gene network. (B) Single-cell immune landscape and treatment-related composition of this cohort. Left: UMAP visualization of 559,714 immune cells revealed eight immune clusters. Right: Cell-type proportions varied by treatment (BL: baseline vs. TRT: treatment) and clinical response (DCB vs. NCB). (C) Distribution of sc-eQTLs across cell types and conditions. The upset plot shows eQTL counts by sharing pattern. In the dot plot, “Shared” indicates eQTLs shared across two or more clusters; “Cell type-specific” denotes cluster-restricted eQTLs, with black dots indicating each specific cluster. Bar plots display treatment-specific distributions. Notably, CD8+ T cells and monocytes contributed substantial unique eGenes. (D) Representative treatment-specific eQTLs. Treatment-specific eQTLs were identified by comparing genotype-dependent gene expression across baseline and treatment conditions (see Supplementary Figure S2C for additional details). Boxplots show genotype-dependent expression of TNF (monocyte), TNFRSF1A, and PRF1 (CD8+ T cells) across treatment conditions. Each association met lfsr < 0.05 (mashr); posterior β values are indicated. (E) Identification of lung cancer-specific eGenes. A Venn diagram shows the overlap of eGenes with GTEx v8 and OneK1K blood eQTL datasets. Seven hundred two eGenes (4.6%) were exclusive to our cohort, indicating disease-specific transcriptional regulation. eQTLs identified exclusively in our study are highlighted in red. (F) Identification of ICI response-explaining co-expression module. Volcano plot of module activity differences (DCB at baseline as reference) across WGCNA modules. Modules were constructed from individual immune clusters using WGCNA; The CD8-brown module showed significant differential activity (|log2FC| > 1, FDR < 0.05) at baseline between response groups. (G) Regulatory relevance of the CD8-brown module. Correlation analysis of the CD8-brown module with the EOMES and TBX21 regulons in CD8+ T cells. The color scale represents the density of cell counts. (H) Enrichment of CD8-brown module activity in tumor microenvironment CD8+ T cells. In the tumor and metastatic lymph node scRNA-seq dataset from 18 NSCLC patients, we identified NK/T cells and projected our CD8+ T-specific WGCNA network upon this dataset. Violin plots show elevated CD8-brown module activity in effector CD8+ T subtypes (CD8+ TEM, CD8+ TEMRA) from NCB tumors. ****P < 0.0001, Wilcoxon rank-sum test. (I) Clinical relevance of the fraction CD8-brown+ CD8+ T cells. Kaplan-Meier curves for OS and PFS stratified by the fraction of CD8+ T cells with high CD8-brown activity (>75%; CD8-brown core gene score over 0.36 was defined as high CD8-brown activity). Patients with higher fractions of CD8-brown+ CD8+ T cells (Ratio-high in the table below) had significantly worse outcomes (OS: P = 0.035, PFS: P = 0.018). The tables below indicate the number of patients in each group. (J) eGenes show higher centrality in CD8+ T cells. Violin plots compare rank-normalized WGCNA module connectivity (kME) between eGenes and non-eGenes in CD8+ T cells at baseline and after ICI treatment eGenes were more centrally located (P < 0.001). (K) eQTL effect size by module membership. Effect sizes (|β|) of eQTLs are smaller for genes assigned to WGCNA modules, suggesting functional constraints on highly connected genes. P < 0.001. (L) Inverse correlation between module centrality and eQTL effect size. Boxplots across kME rank quintiles show decreasing eQTL effect size with increasing centrality in both baseline and treatment. Significant trends (P < 0.001, linear regression) indicate constrained genetic regulation of core network genes. eQTL significance was defined using local false sign rate (lfsr < 0.05) from mashr. Group comparisons were assessed with Wilcoxon rank-sum or linear regression. Multivariate survival models were adjusted for clinical covariates. Abbreviations: BL, baseline; CD8+ TCM, CD8+ central memory T cell; CD8+ TEM, effector memory CD8+ T cells; CD8+ TEMRA, CD8+ terminally differentiated effector memory T cells with expressing CD45RA; DCB, durable clinical benefit; DC, dendritic cell; eGenes, eQTL-regulated genes; EOMES, eomesodermin; eQTL, expression quantitative trait loci; FC, fold change; FDR, false discovery rate; GNN: graph neural network; GRN, gene regulatory network; GTEx, Genotype-Tissue Expression project; ICI, immune checkpoint inhibitor; kME, module eigengene; lfsr, local false sign rate; LN, lymph node; NCB, non-durable clinical benefit; NK, natural killer cell; NSCLC, non-small cell lung cancer; OneK1K, OneK1K sc-eQTL study; OS, overall survival; PBMC, peripheral blood mononuclear cells; PFS: progression-free survival; PRF1, perforin 1; sc-eQTL, single-cell eQTL; TBX21, t-box transcription factor 21; TNF, tumor necrosis factor; TNFRSF1A, TNF receptor 1A; TRT, treatment; UMAP, uniform manifold approximation and projection; WGCNA, weighted gene co-expression network analysis. After quality control and pseudobulk aggregation, we identified 9,147 eQTL pairs—expression-regulating SNPs (eSNPs) linked to 3,616 blood expression-regulated genes (eGenes)—across eight immune cell clusters and treatment conditions (Figure 1B-C, Supplementary Figure S1A, Supplementary Table S2). Consistent with previous studies [3, 4], eGene counts correlated with cell abundance, and eSNPs were enriched in regulatory elements (Supplementary Figure S1B-D). Multiadaptive shrinkage [5] revealed distinct cell-type- and treatment-dependent regulation, including 245 treatment-specific eQTLs (Supplementary Figure S2A-D, Supplementary Tables S3-S4). For instance, tumor necrosis factor (TNF) was regulated in monocytes post-treatment (posterior β = 1.17), while TNF receptor 1A (TNFRSF1A) was baseline-regulated in CD8+ T cells (β = 1.10), indicating genetic variants may shape immune gene expression during ICI therapy (Figure 1D). Additional examples include key cytotoxic mediators perforin 1 (PRF1) and granzyme B (GZMB) in baseline CD8+ T cells (Figure 1D, Supplementary Figure S2E). To validate our findings, we conducted two complementary analyses. First, colocalization analyses with genome-wide association study (GWAS) loci for autoimmune and blood traits showed overlaps (PP.H4 > 0.6), suggesting possible shared regulatory mechanisms (Supplementary Figure S3, Supplementary Tables S5-S6). Second, comparison with external eQTL studies showed our study-specific eQTLs, hereafter referred to as lung cancer-specific eQTLs, were enriched in cancer- and immune response-related pathways (Figure 1E, Supplementary Figure S4A), reflecting chronic immune adaptation. Nevertheless, most regulatory patterns seen in healthy donors from the 1M-scBloodNL study [3] were preserved (Supplementary Figure S4B, Supplementary Tables S7-S8). Beyond eQTL mapping, we used weighted gene co-expression network analysis (WGCNA) to build co-expression networks for each major immune cell type [6], enabling comparison of cell-type-specific regulation and transcriptomes (Supplementary Tables S9-S10). Among the identified modules, the CD8-brown module showed elevated baseline activity in the non-durable clinical benefits (NCB) group (Figure 1F, Supplementary Figure S5A). This module was cytotoxic-gene-rich (PRF1, apolipoprotein B mRNA-editing enzyme, catalytic subunit 3G [APOBEC3G], and GZMB) and highly expressed in the NCB group, particularly in differentiated CD8⁺ T subclusters (Supplementary Figure S5B-D). Its activity was supported by external blood datasets from healthy individuals, tumor patients, and ICI-treated tumor datasets (Supplementary Figure S5E-F). To explore potential regulators, we applied single-cell regulatory network inference and clustering (SCENIC) [7], identifying eomesodermin (EOMES) and t-box transcription factor 21 (TBX21) as putative regulators linked to CD8+ T cell differentiation and exhaustion [8]. Their regulon activity aligned with both the CD8-brown module (Figure 1G, Supplementary Figure S6A-B), and the abundance of effector memory CD8+ T (CD8+ TEM) cells (Supplementary Figure S6C-D). Core genes, including EOMES, TBX21, and interleukin-2 receptor, beta subunit (IL2RB), were up-regulated in NCB compared to durable clinical benefit (DCB) group (Supplementary Figure S6E). Notably, TBX21 and EOMES were not eGenes in our analysis. Building on these findings, we employed a graph neural network (GNN)-based framework to refine the CD8-brown module by integrating protein-protein interaction (PPI) networks, gene ontology (GO) annotations, and co-expression data (Supplementary Figure S7A-B). The refined subnetwork captured a more coherent CD8⁺ T cell differentiation program and showed stronger topological connectivity than the original WGCNA-derived gene set in Search Tool for the Retrieval of Interacting Genes/Proteins (STRING; node density: 0.333 vs 0.183; Supplementary Figure S7C-D, Supplementary Table S11) [9]. Pathway enrichment for canonical pathways in CD8+ T cells supported this refinement (Supplementary Figure S7E-F). This module was well-conserved across external scRNA-seq datasets, including the tumor microenvironment (TME) of ICI-treated NSCLC patients (Supplementary Figure S8) [10] and blood from cancer patients (Supplementary Figure S9). Its activation was significantly enhanced in non-responders, particularly in differentiated CD8+ T cell subclusters like CD8+ TEM (Figure 1H, Supplementary Figures S8-S9, Supplementary Table S12), suggesting that this systemic cytotoxic signature is recapitulated in the tumor microenvironment and may reflect ICI response-associated immune status. Given its strong association with ICI response, we evaluated whether the refined core genes from the CD8-brown module could explain survival outcomes. We derived a core gene score from 15 GNN-prioritized CD8-brown genes with high network centrality (Supplementary Methods, Supplementary Figure S7, Supplementary Table S13), which strongly correlated with module activity (Supplementary Figure S10A). Using grid search and bootstrapping (Supplementary Figure S10B-C), we determined a threshold to classify patients: those with >75% of CD8+ T cells with high CD8-brown core gene score (>0.36) displayed significantly shorter overall survival (OS, P = 0.035) and progression-free survival (PFS, P = 0.018; Figure 1I, Supplementary Figure S10B-C). The survival association remained significant in a multivariate Cox model adjusting for clinical covariates (Supplementary Figure S10D). Overall, our analysis suggests that the TBX21-EOMES regulatory axis may drive the CD8-brown module, representing CD8+ T differentiation with cytotoxic eGenes like PRF1 and GZMB, which in turn stratify poor responders to ICI. Finally, to evaluate how germline variants influence the immune network associated with ICI response, we examined the distribution of eQTLs within co-expression and regulatory networks (Figure 1J-L, Supplementary Figure S11A-C). eGenes were enriched in central network positions (Supplementary Figure S11B); however, no single eQTL–including PRF1, GZMB or TNFRSF1A–exerted a dominant regulatory effect on the CD8-brown module. Instead, genes with higher centrality or WGCNA-assignment were linked to smaller eQTL effect sizes (Figure 1K-L), suggesting functional constraints on key immune differentiation pathways. Despite these insights, several limitations exist. First, dichotomizing responses into DCB and NCB may obscure heterogeneous outcomes, while the treatment-induced transcriptomic changes appeared smaller than inter-sample heterogeneity (Supplementary Figure S12). Second, although monocyte-specific eQTL were detected and monocytes-ICI interactions are well-known, we excluded them from downstream analysis due to minimal contribution to ICI-associated co-expression modules (Supplementary Figure S13). Third, our focus on cell-type- and condition-specific eQTLs and co-expression networks may overlook intercellular interactions and trans effects influencing immune function. Lastly, although the CD8-brown core gene score stratified survival in our cohort (c-index = 0.68-0.71 for PFS/OS, Supplementary Figure S10D), its predictive value requires validation in independent datasets. Taken together, our study identified 3,616 blood- and 702 lung cancer-specific eGenes via sc-eQTL mapping and revealed a transcriptomic module (CD8-brown) associated with CD8+ T cell differentiation and non-responder for ICI response. The limited influence of eQTLs on co-expression networks suggests functional constraints in the immune transcriptome. By linking genetic variation to cytotoxic network activity and clinical outcomes, our analysis provides a framework to understand systemic immunity under ICI treatment in metastatic NSCLC. Future research involving larger cohorts and experimental validation of key regulatory variants—such as those affecting PRF1 and GZMB—could clarify causal mechanisms and refine personalized ICI strategies. Hyungtai Sim: data curation, formal analysis, validation, investigation, writing-original draft. Geun-Ho Park: data curation, formal analysis, investigation. Woong-Yang Park: investigation. Se-Hoon Lee: conceptualization, investigation, writing-original draft. Murim Choi: conceptualization, investigation, writing-original draft. We appreciate the participation of the patients in this study. The authors declare no competing interests. This work was supported in part by grants from the Korean Research Foundation (NRF-2021R1A2C3014067, NRF-RS-2023-00207857 to Murim Choi, and NRF-2020R1A2C3006535, NRF-RS-2025-00519956 to Se-Hoon Lee), the Korea Health Technology R&D Project through the Korea Health Industry Development Institute (HR20C0025 to Se-Hoon Lee), and Future Medicine 20*30 Project of the Samsung Medical Center (SMX1230041 to Se-Hoon Lee). This study investigated human samples and was given permission by the Samsung Medical Center Institutional Review Board (No. 2018-04-048, 2022-01-094). All participants underwent an informed consent process before being enrolled in the study. The expression data in a matrix format and metadata for scRNA-seq have been deposited at K-BDS under the accession number KAP241028. Additional datasets analyzed in this manuscript are publicly available as follows: the Gene Expression Omnibus (GEO; www.ncbi.nlm.nih.gov/geo/) under accession numbers GSE200996 (the HNSCC dataset), GSE123814 (the TNBC dataset), and GSE205335 (the NSCLC dataset) and the AIDA Phase 1 dataset from CellxGene accessible at https://cellxgene.cziscience.com/collections/ced320a1-29f3-47c1-a735-513c7084d508. All other data supporting the findings of this study, including derived datasets, are provided in the manuscript, its supporting files, or are available in the GitHub repository: https://github.com/snu-mchoi-lab/LungCancer_ICI_sceQTL_2024 Please note: The publisher is not responsible for the content or functionality of any supporting information supplied by the authors. Any queries (other than missing content) should be directed to the corresponding author for the article.
No takes yet. Share an insight, caveat, or question.
Sim et al. (2025) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: