Single-cell transcriptomic profiles of LUAD samples were downloaded from the GSE117570 dataset, which contained single-cell sequencing data of four non-small cell lung cancer (NSCLC) tissues and four paired normal tissues (Table 1). Of the eight tissue samples, three pairs of LUAD tissues and paired normal tissues were the objects of analysis in this study after excluding one pair of lung squamous cell carcinoma and matched normal tissues. The RNA-seq data of LUAD were extracted from Gene Expression Omnibus (GEO) and The Cancer Genome Atlas (TCGA) (Table 2). The samples of the two databases were processed in the same way, and only those with complete survival time and survival time shorter than 10 years were retained. The TCGA database included 491 samples of LUAD and 59 paracancerous tissues. Four GEO datasets analyzed were GSE30219 (n = 238), GSE31210 (n = 226), GSE19188 (n = 78), and GSE50081 (n = 126).

R package SeuratR package (version 3.6.3, https://satijalab.org/seurat/) [18] was adopted to process and analyze the scRNA-seq data. Mitochondrial genes were filtered using PercentageFeatureSet function, and cells with mitochondrial genes <15% and gene number <250 were included in the analysis. The FindVariableFeatures function was used to filter highly variable genes after data normalization. Then, the batches were removed using FindIntegrationAnchors function, the scRNA-seq data were integrated by the IntegrateData method and scaled by the ScaleData function, followed by using principal component analysis (PCA) to reduce dimensionality (dim = 30). The cells on the first 50 principal components were clustered by the FindNeighbors and main cell clusters were classified by FindClusters functions. The markers for each cell cluster were screened under the criteria of logfc = 0.35, Minpct = 0.35, and corrected p < 0.05.

T cell clusters were identified according to CD45 (PTPRC) and CD3 (CD3D, CD3E, CD3G) [19] of T cells, and then CD8+ T cells were screened based on CD8+ (CD8A and CD8B). The same clustering method was used to divide CD8+ T cells into subgroups at Resolution = 0.1. Gene set variation analysis score was calculated based on the expression data for CD8+ T cells in different subgroups of normal samples and LUAD samples in TCGA [20], and then the samples were grouped by the median score of each CD8+ T cell subgroup. Survival analyses in both groups were conducted using the “survival” package, and the prognosis of each CD8+ T cell subgroup was evaluated using the Kaplan-Meier (KM) method.

A co-expression network was developed using the “WGCNA” package [21] based on the LUAD expression profile data of TCGA. The matrix was generated by calculating the distance between genes using the Pearson correlation coefficient. To ensure a scale-free nature of the network, soft threshold was determined by selecting the smallest network. Hierarchical clustering was performed based on topological overlap matrix (TOM), which was converted from adjacency matrix. Dynamic tree cut was performed, with a minimum number of 70 genes in a module, and then eigengenes in each gene module were calculated. Finally, cluster analysis was performed to combine modules close to each other into a new module.

The Pearson correlation test was employed to analyze the relationship between each module and prognostic CD8+ T cell subgroups so as to determine the modules with the highest positive correlation. Genes with genetic significance (GS) > 0.8 and module membership (MM) > 0.6 in each module were imported into the “survival” package to perform univariate COX regression analysis. Under the threshold of p < 0.05, prognostic genes for LUAD were filtered.

Clustering analysis was conducted by running the “pam” algorithm of the “ConsensusClusterPlus” package. The optimal number of clusters in the range of set k (2–10) was determined by extracting 80% of the LUAD samples in TCGA during 500 bootstrap resampling. Distribution differences of different clinical characteristics among the clusters were analyzed by Chi-square test.

‘Maftools’ was used to visualize and analyze the top 20 most frequently mutated genes in subtypes in TCGA. Following the study of Thorsson et al. [22], the scores of fraction altered, homologous recombination defects (HRDs), tumor mutation burden (TMB), and number of segments were calculated and compared by the Kruskal-Wallis Test.

Transcriptional profiles of each subgroup in TCGA-LUAD were used for pathway enrichment analysis of the Kyoto Encyclopedia of Genes and Genomes (KEGG) for Gene Set Enrichment Analysis (GSEA). Significant KEGG pathways were selected under false discovery rate (FDR) < 0.05. The scores of 10 carcinogenic signaling pathways (nuclear respiratory factor 1 [Nrf1], transforming growth factor-β [TGF-β] signaling, Notch, p53, β-catenin/Wnt, cell cycle, Myc, Hippo, phosphatidylinositol 3-kinase/protein kinase B [PI3K/Akt], and receptor tyrosine kinase [RTK]-RAS) were calculated [23] and compared using Kruskal-Wallis Test.

The TME characteristics of each subgroup were analyzed using ESTIMATE [24] and CIBERSORT [25] algorithms. The stromal score was calculated based on the gene expression profile of each subgroup in the R package “ESTIMATE”. The levels of 22 infiltrating immune cells were calculated by CIBERSORT using the gene signature matrix. Gene sets of immune-related KEGG pathways were identified, and their expression in each subgroup was analyzed.

Differential expression between each subgroup and the other two subgroups in the TCGA-LUAD dataset was analyzed using the R package “limma”. DEGs screened under the thresholds of FDR < 0.05 and | log2 (Fold Change) | > log2 (1.5) were integrated to further select prognostically significant genes using the coxph function of the “survival” package (p < 0.0001). Then, a risk score system for evaluating LUAD prognosis was developed according to the coefficient of the genes selected by LASSO in the R package “glmnet” with 10-fold cross-validation. The formula of Risk score = ∑ Coefficien_mRNAi × Expression _mRNAi was employed to calculate the risk score for each case in the TCGA-LUAD dataset and four independent GEO datasets (GSE30219, GSE31210, GSE19188, GSE50081). Data were normalized by Z-score, with 0 as the boundary to classify samples into risk groups. Survival differences between risk groups were reflected in the KM curves. The area under the curve (AUC) of the receiver operating characteristics curve was measured to assess the prediction accuracy of the risk score system.

Rooney et al. [26] designed a quantitative method for measuring immune cytolytic activity (CYT) based on the average of Granzyme A and Perforin expression levels. Using this method, the CYT scores of LUAD samples in TCGA were determined. Tumor Immune Dysfunction and Exclusion scores [27] of the risk groups in TCGA-LUAD were calculated. A tumor-reactive T cell signature (TRS) for indicating tumor reactivity [28] was introduced to further calculate the TRS score of the LUAD samples in TCGA.

Independent prognostic variables for LUAD were selected from age, gender, risk score and American Joint Committee on Cancer (AJCC) stage by performing univariate and multivariate Cox regression analysis using the “survival” package. Next, the independent prognostic factors were integrated into the “rms” package in the R program to develop a nomogram. Calibration curve, ROC curve, and decision curve analysis (DCA) were employed to assess the accuracy and clinical benefit of the nomogram.

Total RNA from BEAS-2B, H1299, and A549 was extracted using TRIzol reagent (Thermo Fisher, Waltham, MA, USA). Using the HiScript II SuperMix (Vazyme, Nanjing, Jiangsu province, China), cDNA was obtained from the RNA (500 ng). Subsequently, qRT-PCR was conducted with SYBR Green Master Mix in ABI 7500 System (Thermo Fisher, Waltham, MA, USA) under the conditions of 45 cycles for 10 min at 94°C, for 10 s at 94°C, and for 45 s at 60°C. The internal reference was GAPDH. See Table 3 for the primer sequences for target genes.

BEAS-2B (BNCC359274), H1299 (BNCC100268), and A549 (BNCC337696) cells (Beijing Bena Biotechnology Co., Beijing, China) were commercially purchased and cultured in DEME F-12 medium. of the negative control (NC), CD200R1 siRNA, and CDCP1 siRNA (Sagon, China) were transfected utilizing Lipofectamine 2000 (Invitrogen, Carlsbad, California, USA). The siRNAs, GGCTACTGTTGATTTTGACTATC (CD200R1-si) and AGCAACATTACAGTTCTCATAAA (CDCP1-si) were the target sequences for CD200R1 siRNA and CDCP1 siRNA, respectively. An anti-mouse CD8 microbeads isolation kit (Miltenyi, Bergesch Gladbach, Germany) was used to extract splenic CD8+ T cells from two euthanized mice with a magnetic separator.

The invasion of BEAS-2B, H1299, and A549 cell lines was analyzed by performing a Transwell assay. The cells (5 × 10⁴) were inoculated to chambers coated with Matrigel (BD Biosciences, New York, USA). Complete Dulbecco’s modified Eagles medium and serum-free medium were added to the lower and the upper layers, respectively. The migrating or invading cells were fixed using 4% paraformaldehyde. The cells were stained with 0.1% crystalline violet after 24-h incubation. A light microscope was used to count cell numbers.

We used a cytotoxicity assay kit (Roche, Basel, Switzerland) for LDH detection. After co-culturing the H1299 and A549 cell lines with CD8+ T cells for 24 h, 50 μL of each sample was transferred into a 96-well plate and mixed with an equal amount of reagent mixtures. The intensity of red color in the colorimetric assay was determined at 490-nm wavelength after incubation at 37°C (Thermo Fisher, Waltham, MA, USA) to calculate the number of damaged/dead cells.

Data were analyzed using the R program and GraphPad Prism. The clinical characteristics of each molecular subtype were compared using the chi-square test, and the biological characteristics were compared using the Kruskal-Wallis test. The KM curve and ROC curve were plotted using “survival” and “survivalROC” packages, respectively. Unless otherwise specified, a default p < 0.05 was considered statistically significant.

A total of 7484 cells from three pairs of LUAD tissues and paired normal tissues were included after quality control (Suppl. Fig. S1A). After eliminating the batch effect, cells in the samples were distributed unevenly (Suppl. Fig. S1B, Suppl. Fig. S1C). A total of 20 cell subsets were identified by clustering. Clusters 0 and 6 were considered the main T cell distribution subsets as they specifically expressed CD45 and CD3 (CD3D, CD3E, CD3G) (Suppl. Fig. S1D). CD8+ T cells were screened based on CD8 expression from all the identified T cells were extracted and divided into two clusters (Suppl. Fig. S1E, Fig. 1A). CD8+ T cells only existed in two samples, and the proportions of clusters CD8+ 0 and CD8+ 1 in the two samples were noticeably different (Fig. 1B). Expression analysis showed that the top 10 high-expressed gene markers in CD8+ T cell cluster 0 had a low expression in CD8+ T cell cluster 1, while the top 10 high-expressed gene markers in CD8+ T cell cluster 1 had a low expression in CD8+ T cell cluster 0 (Fig. 1C). KEGG enrichment analysis showed that the gene markers of the two CD8+ T cell clusters were significantly enriched in apoptosis and salmonella infection, and CD8+ T cell cluster 0 was also significantly correlated with natural killer cell mediated cytotoxicity, viral myocarditis, graft vs. host disease, allograft rejection and human cytomegalovirus infection (Fig. 1D). In LUAD samples, the content of CD8+ T cell cluster 0 was noticeably lower but that of CD8+ T cell cluster 1 was significantly higher (Fig. 1E).

A high score of CD8+ T cell cluster 0 was related to a noticeably better OS of the samples (Fig. 1F). However, no significant difference in the OS was detected between samples grouped based on cluster CD8+ 1 (Fig. 1G). In other words, CD8+ 0 was the most related to LUAD prognosis.

WGCNA was conducted to cluster genes of LUAD samples in TCGA into modules and hierarchical clustering was performed to detect the outliers (Fig. 2A). When the soft threshold was 10, the correlation noise was noticeably reduced and the network was scale-free (Fig. 2B). A total of 12 modules were dissected by clustering tree (Fig. 2C). The brown module containing 710 genes was significantly correlated with CD8+ 0 (Fig. 2D). A significant correlation between GS and MM could be observed from the scatter diagram of these genes (Fig. 2E), with 106 genes under the criteria of GS > 0.6 and MM > 0.8, of which 42 were significantly linked with the prognosis of LUAD (Suppl. Table S1).

According to the cumulative distribution function (CDF) curve, the delta area of CDF, and the consistency matrix, the classification was stable when there were three clusters. Therefore, the samples were divided into three subtypes (clusters 1, 2, and 3) (Figs. 3A–3C). TCGA samples in the three clusters showed significant differences in the 5-year survival rate. Specifically, the 5-year survival rate of cluster 1 was the lowest but the highest in cluster 3, and that of cluster 2 was intermediate between clusters 1 and 3 (Fig. 3D). In GSE30219, significant prognostic differences were observed between the three clusters, with the optimal prognosis in cluster 3 and the worst in cluster 1 (Fig. 3E).

The chi-square test was used to confirm the proportion of different clinical phenotypes in clusters 1, 2, and 3. T stage distribution and survival state of the three clusters were significantly different. A difference in clinical phenotypic distribution between cluster 3 and the other two clusters was observed. Compared with clusters 1 and cluster 2, cluster 3 (T1 stage, N0 stage, stage I) had a higher proportion of female samples at early stages (Fig. 4A). Among all three clusters, TTN, RYR2, and ZFHX4 had the highest mutation in cluster 1 and the lowest in cluster 3 (Fig. 4B). HRDs, fraction altered, number of segments and TMB scores were used to evaluate the mutation status in LUAD. We found that the mutation score of the first three factors was the lowest in cluster 3 and the highest in cluster 1 but TMB score in cluster 1 was significantly higher than in cluster 3 (Fig. 4C).

Enriched biological pathways in TCGA were identified by GSEA. A total of 18 biological pathways in cluster 1 were significantly inhibited, 17 of these 18 pathways were highly activated in cluster 3, and 23 pathways were activated in cluster 2. Many of these pathways significantly associated with the three clusters were modulated immune responses, such as complement, inflammatory response, interferon alpha/gamma response, allograft rejection, etc. (Fig. 5A). Among the 10 major biological pathways regulating cancer progression, Hippo, Myc, Notch, Nrf1 and RTK-RAS and TGF β signaling showed significantly different ssGSEA scores among the three molecular subtypes (Fig. 5B).

The stromal score, immune cell infiltration, immune score, immune checkpoints, and signal molecules in immune pathways among the three clusters were calculated. CIBERSORT was used to calculate and compare the scores of 22 immune cells of each cluster. Significant differences in helper follicular T cells, memory B cells, naive B cells, CD8+ T cells, plasma cells, delta gamma T cells, resting/activated mast cells, M1 macrophage, resting/activated memory CD4 T cells, M2 macrophage, resting/activated dendritic cells, and activated NK cells were observed among the three clusters (Fig. 6A). The levels of immune and stromal scores were cluster 1 < cluster 2 < cluster 3 (Fig. 6B). Specifically, the number of immune checkpoints in cluster 1 was significantly lower than in cluster 2, and the content of immune checkpoints in these two clusters was significantly lower than those in cluster 3 (Fig. 6C). Moreover, the expression of signal molecules in the immune pathway was the highest in clusters 3 and the lowest in cluster 2 (Fig. 6D).

A total of 1833 DEGs, 18 DEGs, and 1264 DEGs were screened from cluster 1 and non-cluster 1, cluster 2 and non-cluster 2, cluster 3 and non-cluster 3 in TCGA-LUAD datasets, respectively. After integrating all these DEGs, we obtained 1915 genes. Univariate COX regression analysis on these 1915 genes selected 23 genes with prognostic effects (Fig. 7A). The optimal λ was determined by Lasso, and a prognostic model (Figs. 7B, 7C) was successfully devised using 11 genes. The risk score formula was Risk score = −0.391 × CD200R1 − 0.141 × CLEC17A − 0.076 × ZC3H12D − 0.049 × GNG7 − 0.038 × SNX30 + 0.229 × CDCP1 + 0.05 × NEIL3 + 0.05 × IGF2BP1 + 0.081 × RHOV + 0.006 × ABCC2 + 0.073 × KRT81. Normalized risk score was calculated for each case in the LUAD sample set in TCGA as the training cohort. We observed that a greater death risk was positively correlated with a higher risk score (Fig. 7D). The survival curve revealed the significant difference between the two risk groups; specifically, the low-risk patients showed a significantly improved OS than high-risk patients (Fig. 7E), with 1-, 3-, 5-year AUC of ROC curve of 0.72, 0.73 and 0.69, respectively (Fig. 7F). In all the four cohorts (GSE30219, GSE31210, GSE19188, and GSE50081), the KM curve demonstrated that low-risk samples of had greater survival advantages than the high-risk samples (Figs. 7G, 7I–7K). According to the AUC of the ROC curve, the prognostic model showed a stable and accurate prediction of OS for the samples in GSE30219 (Fig. 7H). In the GSE31210 and GSE50081 cohorts, the AUC for 1-, 3- and 5-year survival prediction was higher than 0.7 (Figs. 7L, 7N). For another validation cohort of GSE19188, the AUC for 1-year, the respective 3- and 5-year survival prediction were 0.77, 0.68, and 0.65 (Fig. 7M).

The relationship between different clinical features and the risk score in TCGA, where the risk score showed gender differences, was analyzed. Samples with high T stage, N stage, and AJCC stage tended to have a higher risk score, and male patients had a significantly higher risk score than females. Samples with high T stage, N stage, and AJCC stage tended to have a high risk-score. The risk score of samples at the T3 stage was significantly higher than those at T1 and T2 stages. Compared with N0 stage samples, the risk score of samples at N1 and N2 stage was significantly higher. Samples at stage IV and stage III had a significantly higher risk score than those at stage I. The risk score was also related to smoking history (Fig. 8A). Sankey diagram displayed that the samples in the high-risk group were mainly from clusters 1 and 2, and that most of the samples in cluster 3 were high-risk samples (Fig. 8B).

ESTIMATE and Spearman correlation analysis showed a significantly negatively related association between risk score and immune score in the TCGA cohort (Fig. 9A). Low-risk group had a significantly higher immune score than the high-risk group (Fig. 9B). As for the three immune checkpoints CTLA4, PD-L1 and PD-1, the expression of CTLA-4 and PD-1 was significantly downregulated in high-risk group (Fig. 9C), and the TIDE score of anti-PD1 or anti-CTLA4 treatment was significantly higher in high-risk groups (Fig. 9D). Samples with a low-risk score showed greater tumor immunoreactivity as cytolytic activity and TRS score were obviously high (Figs. 9E, 9F).

We also obtained IC50 values from the Genomics of Drug Sensitivity in Cancer (GDSC) for chemotherapy drugs and targeted therapeutic drugs such as Cisplatin, Erlotinib, Rapamycin, Sorafenib, MG−132, AZ628, VX−680, and Saracatinib. It was found that the high-risk patients were closely correlated with the sensitivity of all of these drugs, while the low-risk patients were closely related to the resistance to these drugs (Fig. 9G).

According to Cox regression analysis (Figs. 10A, 10B), the risk score and AJCC stage were independent indicators for predicting the OS of LUAD and were combined to create a nomogram (Fig. 10C). The calibration curve demonstrated that the nomogram accurately predicted 1-, 3-, and 5-year OS (Fig. 10D). The DCA results showed that the nomogram had the greatest net benefit, indicating that the nomogram was the most practical indicator among other variables (Fig. 10E). The tROC and C-index also supported that the nomogram had the strongest prediction ability (Fig. 10F, Suppl. Fig. S2).

We also conducted PCR to detect the mRNA expression of CD200R1, CLEC17A, ZC3H12D, GNG7, SNX30, CDCP1, NEIL3, IGF2BP1, RHOV, ABCC2, KRT81 in BEAS-2B, H1299 and A549 cell lines. In H1299 and A549 cell lines. The results showed that the expression of CD200R1, CLEC17A, ZC3H12D, GNG7 and SNX30 was significantly downregulated (Figs. 11A–11E), but that of CDCP1, NEIL3, IGF2BP1, RHOV, ABCC2 and KRT81 was upregulated (Figs. 11F–11K). Subsequently, the invasion of H1299 and A549 cell lines was detected after the interference with CD200R1 and CDCP1. Inhibition of CD200R1 significantly increased the viability of H1299 and A549 cell lines (Figs. 11L–11N). However, the viability of H1299 and A549 cell lines was significantly reduced after suppressing CDCP1 (Figs. 11O–11Q). These findings verified the reliability of the risk score model.

Similarly, the cell viability of H1299 and A549 cell lines was increased by inhibiting CD200R1 expression but reduced after inhibiting CDCP1 expression (Figs. 12A and 12B). To elucidate the effects of CD200R1 and CDCP1 expression on CD8+ T cells, we co-cultured tumor cell lines and CD8+ T lymphocytes using Transwell assay. The effects of knocking down CD200R1 and CDCP1 on tumor cell killing of CD8+ T cells in H1299 and A549 cell lines were analyzed. As shown in Figs. 12C–12E, knocking down CD200R1 inhibited the tumor-killing effect of CD8+ T cells, whereas knocking down CDCP1 enhanced the killing ability of CD8+ T cells. In addition, we examined the expression of surface markers CD69 and CD25 in CD8+ T cells after 24-h of co-culturing. CD69 is widely recognized as an activation marker for T cells, and its expression level reflects the killing capacity of CD8+ T cells. CD25 is an activation marker for regulatory T cells that reflects the immunosuppressive function of T cells. When CD200R1 was inhibited in H1299 and A549 cell lines, the results showed that the expression level of CD69 on the surface of CD8+ T cells was downregulated, but that of CD25 was upregulated. In contrast, when CDCP1 was inhibited in the H1299 and A549 cell lines, the level of CD69 on the surface of CD8+ T cells was increased, whereas CD25 expression was decreased (Figs. 12F–12J).