Diverse regulated cell death modes predict the immune microenvironment and drug sensitivity in lung adenocarcinoma

2.1 Data acquisition

The RCD-related genes comprise the key regulatory genes of 15 cell death modes, as mentioned earlier. Gene lists were constructed from the KEGG database, MSigDB database, and manual input through literature retrieval (Tang et al., 2019; Zou et al., 2022). The workflow of this study is summarized in Figure 1. TCGA RNA-seq data and clinical phenotype data from LUAD samples were downloaded from the University of California Santa Cruz (UCSC) Xena database (https://xenabrowser.net/hub/). Gene expression data from normal lung tissues were downloaded from the Genotype-Tissue Expression (GTEx) portal database (https://www.gtexportal.org/). The “combat” R package was used to remove batch effects. The “org.Hs.eg.db” R package was applied to conduct annotations of ensemble IDs. In addition, we downloaded data from the Gene Expression Omnibus (GEO) repository (https://www.ncbi.nlm.nih.gov/geo/) using the “GEOquery” R package, and the accession IDs were GSE40791, GSE31210, and GSE30219. The probes were annotated using the “AnnoProbe” R package. We used the gene expression data from the Kaplan‒Meier Plotter database (https://kmplot.com/analysis/) for gene signature validation. We also downloaded masked somatic mutation data from UCSC Xena.

2.2 Identification and depiction of prognosis-related RCD genes

Raw transcriptome count data from 585 patients from TCGA and 578 healthy donors from GTEx were included. Subsequently, the “Deseq. 2” R package was employed to screen out DEGs between LUAD and normal lung samples according to the adj. p < 0.05 and |log2FC | > 1.5. We used XIANTAO Scholar online tools (https://www.xiantao.love/) to identify enriched biological processes and pathways based on the DEGs. The “maftools” R package (Mayakonda et al., 2018) was used to explore masked somatic mutations among TCGA LUAD patients.

2.3 Construction of the prognostic RCD gene signature

β(i) represents the risk coefficient of the gene, and Exp(i) refers to the gene expression levels. The coefficients were retained to 3 decimal places. According to the best cutoff value determined by the “surv_cutpoint” function in the “survminer” package, we divided LUAD patients into low and high RCD score groups. Then, we conducted clinical relevance analyses between the collected clinical information (age, sex, T stage, N stage, M stage, and stage) and RCD scores.

2.4 Unsupervised clustering

We applied the “ConsensusClusterPlus” R package to complete consensus clustering analysis to identify the potential subtypes of LUAD. The parameters of “ConsensusClusterPlus” were selected as follows: “maxK” was “8,” “clusterAlg” was “pam,” and “distance” was “pearson.”

2.5 Nomogram

Clinical information (age, sex, T stage, N stage, M stage, and stage) and RCD scores were integrated to build the prognostic nomogram model by using the “rms” package. The nomogram plot was drawn by the “regplot” package. The C-index and calibration curve established by the “rcorrcens” and “calibrate” functions were used to evaluate the efficacy. Receiver operator characteristic (ROC) analyses were performed by the “timeROC” package, and the AUC value was used to indicate model performance.

2.6 Tumor microenvironment

We used the “immunedeconv” R package to conduct tumor microenvironment analyses based on the RNA expression data of all genes. The parameter was selected as “abis” to complete a 17-immune-cell deconvolution. The parameter was selected as the “estimate” to predict the stromal score, immune score, tumor purity, and ESTIMATE score of LUAD patients. The parameter “xcell” was selected to perform enrichment analysis for 36 immune and stromal cells as well as to calculate the immune score, stromal score, and microenvironment score. The parameter was selected as the “timer” to evaluate the abundance of six major immune cell types.

2.7 Drug sensitivity

For the analyses of chemotherapy drug sensitivity, the “oncoPredict” package (Maeser et al., 2021) was applied. The training data of the “calcPhenotype” function were selected as “GDSC2” (Genomics of Drug Sensitivity in Cancer 2 datasets). The parameter “batchCorrect” was selected as “eb.”

For the prediction of immunotherapy efficacy, we used two methods, the Tumor Immune Dysfunction and Exclusion (TIDE) algorithm and the “easier” R package. We uploaded the expression matrix in our study to the TIDE online website (http://tide.dfci.harvard.edu/) to obtain the TIDE score and predicted immunotherapy responses of each patient. The “easier” package (Lapuente-Santana et al., 2021) was used to compute immune cell fractions and evaluate pathway and transcription factor activities to obtain immune checkpoint inhibitor (ICI) response scores.

2.8 Survival

We applied the “survival” and “survminer” R packages to perform univariate and multivariate Cox survival analyses. Survival curves were described by Kaplan‒Meier plots and compared with the log-rank test.

2.9 Statistical analysis

We used R software (v.4.2.2) to perform all analyses. Student's t test or the Wilcoxon test was used to analyze differences between the two groups. The Kruskal‒Wallis test was used to compare differences among multiple groups. p < 0.05 was considered statistically significant.

3.1 Study workflow

For the clinical cohorts in this study, we retrospectively included 585 patients from TCGA (524 LUAD and 61 normal lung), 578 healthy donors from GTEx, 307 samples from GSE30219 (Rousseaux et al., 2013) (293 tumors and 14 normal lung), 246 samples from GSE31210 (Okayama et al., 2012) (226 tumors and 20 normal lung), and 194 patients from GSE40791 (Zhang et al., 2012) (100 tumors and 94 normal lung). In addition, 15 RCD modes with 1350 related genes were integrated (Supporting Information: Table 1). Altogether, 580 apoptosis genes, 367 autophagy genes, 220 LCD genes, 121 anoikis genes, 101 necroptosis genes, 88 ferroptosis genes, 52 pyroptosis genes, 9 parthanatos genes, 22 mitochondrial transmembrane permeability-driven necrosis genes, 19 cuproptosis genes, 17 disulfidptosis genes, 7 alkaliptosis genes, 15 entois genes, 8 netotic cell death genes, and 5 oxeiptosis genes were identified. The workflow of this study is shown in Figure 1.

3.2 Multiomics patterns of RCD-related genes in LUAD patients

By comparing LUAD tissues in TCGA with normal lung tissues in GTEx cohorts, we identified 486 differentially expressed genes (DEGs). Among all DEGs, 318 were upregulated and 168 genes were downregulated in LUAD patients (Supporting Information: Table 2). Gene expression heatmaps showed the scaled RNA levels of DEGs (Figure 2a). A principal component analysis (PCA) plot showed the clustering effect of LUAD and normal lungs based on DEGs (Figure 2b). The volcano plot and the differential ranking chart of the DEGs are shown in Figure 2c,d. In addition, Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) enrichment analyses indicated that these DEGs are involved in apoptosis and autophagy signaling pathways (Figure 2e,f). The genome variations in RCD-related genes were also assessed in the TCGA-LUAD cohort. Our results indicated that approximately 89.59% (542/605) of LUAD patients possessed genetic mutations. The top 20 mutations were displayed; among them, TP53 had the highest mutation frequency (51%), KRAS possessed a 26% mutation frequency, KEAP1 possessed a 17% mutation frequency, HMCN1 possessed a 16% mutation frequency, EGFR possessed a 14% mutation frequency, NLRP3 possessed a 12% mutation frequency, and others ranged from 7% to 11% (Figure 2g).

3.3 Construction of a prognostic gene signature based on RCD DEGs

To construct an RCD gene signature for LUAD patients, univariate analysis was leveraged to generally screen out prognosis-related genes. In the TCGA-LUAD cohort, 124 genes were obtained that met the cutoff of p < 0.05. We identified differences between LUAD and normal lung tissues based on GSE40791 gene expression profiling by array (Supporting Information: Figure S1A). The intersection of the DEGs in the TCGA and GSE40791 cohorts included 47 genes (MMP1, FGA, MELK, CTSV, HMGA1, ERO1A, BIK, PMAIP1, CDK5R1, CD27, MIF, ANGPTL4, KRT8, KRT18, ENO1, PPIF, SFN, HSPD1, PIM2, ITGA5, CTSH, HGF, SORT1, BMP5, CTSW, DRAM1, PIK3R1, MSX1, CX3CL1, TLR4, SESN1, ITPRIP, PTGDS, DAPK2, HSPB8, CTSG, IL33, NLRC4, LRRK2, FGR, ITGA8, KL, PECAM1, DNASE2B, FOXF1, LAMP3, and ADRB2). Then, a 17-gene signature was constructed by LASSO regression machine learning (Figure 3a,b). Among the 17 genes, 10 were derived from apoptosis, 3 were derived from autophagy, 3 were from anoikis, 2 were from LCD, 1 was from mitochondrial transmembrane permeability-driven necrosis, 1 was derived from pyroptosis, and 1 was from parthanatos. We next investigated the functional correlation of these genes. GO and KEGG enrichment analyses indicated that they are involved in apoptosis and autophagy signaling pathways and are related to lysosome function (Supporting Information: Figure S1B). We applied the Wilcoxon test to compare the RNA levels of each gene between LUAD and normal lung samples and performed Kaplan–Meier analysis to identify their specific influence on survival, indicating that all the model genes were significantly different between the two groups (p < 0.05) and exerted a significant influence on OS (p < 0.05) (Supporting Information: Figure S2).

Our 17-gene model exported and calculated the RCD score of each patient by the following formula. RCD scores = (0.236 × ERO1A exp.) + (0.0216 × ANGPTL4 exp.) + (0.071 × KRT8 exp.) + (0.0616 × KRT18 exp.) + (−0.0196 × BMP5 exp.) + (0.179 × CDK5R1 exp.) + (0.039 × HSPD1 exp.) + (−0.003 × HGF exp.) + (−0.149 × SORT1 exp.) + (−0.082 × KL exp.) + (0.013 × MIF exp.) + (0.135 × MSX1 exp.) + (−0.136 × NLRC4 exp.) + (−0.121 × PIM2 exp.) + (−0.035 × CX3CL1 exp.) + (0.211 × ITPRIP exp.) + (−0.051 × CTSW exp.). Based on the best cutoff value identified by the surv_cutpoint function (Figure 3c), we divided LUAD patients from TCGA into high- and low-RCD groups, which were selected as a training data set. The RCD scores of patients were significantly correlated with clinical characteristics, including survival outcome (alive or dead), pathologic T stage (T1–T4), pathologic N stage (N0–N3), and tumor clinical stage (I–IV) (Figure 3d,e). The top 12 biological processes, molecular functions and cellular components of the 17 genes are shown in a circle plot (Figure 3f). The protein‒protein interaction network results based on 17 model genes and their top 20 related genes are shown in Figure 3g.

3.4 Internal and external validation of the RCD gene signature

To confirm the prognostic value of the RCD gene model, we next compared the survival of LUAD patients with high and low RCD scores in multiple clinical cohorts. Our results indicated that LUAD patients with high RCD scores had a shorter OS than those with low RCD scores (Figure 4a). Subsequently, we used KM-plotter data, GSE31210, and GSE30219, as external validation cohorts. According to the 17-gene signature, 1144 KM plotter patients were divided into high and low groups, and survival outcomes were compared. The results showed that patients with low RCD scores had better OS and PFS (Figure 4b). Furthermore, according to the 17-gene LASSO signature prediction model, the RCD scores of 226 LUAD patients in the GSE31210 cohort and 293 LUAD patients in the GSE30219 cohort were calculated, and patients were divided into high and low groups. High RCD scores resulted in poor survival times, including OS and PFS (Figure 4c–f).

3.5 Consensus clustering of 47 differentially expressed RCD genes

To explore the underlying subtypes of LUAD, 47 RCD-related genes were included to employ consensus clustering analyses. We revealed that when LUAD patients were divided into two clusters (k = 2), the differences between clusters were most obvious, indicating that all TCGA LUAD patients could be well classified into two subgroups (Figure 5a). The OS outcomes of patients between the two clusters showed an obvious difference, in which cluster 2 was related to a more favorable prognosis, whereas cluster 1 was associated with a poorer prognosis (p < 0.0001, Figure 5b). Consistently, the RCD scores of patients in cluster 1 were significantly higher than those in cluster 2 (p < 0.001, Figure 5c). Similar results were found in GSE31210 (Figure 5d,e). In the GSE30219 cohort, LUAD patients in cluster 1 had a better prognosis (both OS and PFS) than those in cluster 2 (Figure 5f,g). Moreover, the columnar alluvial diagrams of the TCGA and GSE31210 cohorts indicated that the majority of patients in cluster 2 had a low RCD score, whereas GSE30219 had the opposite result (Figure 5h). The cluster with better prognosis had more patients with low RCD scores, suggesting that molecular clusters were significantly associated with RCD scores.

3.6 Establishment and validation of the nomogram based on the RCD risk score

Multivariate Cox regression analyses based on RCD scores and clinical features were performed to determine whether the RCD score could be an independent prognostic biomarker. Our results in the TCGA cohort indicated that the RCD score was an independent predictor for OS (HR = 3.90, 95% CI: 2.64–5.60, p < 0.001, Figure 6a). Then, we used the above variables to create a nomogram survival model for TCGA LUAD patients to estimate the 1-, 3-, and 5-year OS rates (Figure 6b). The C-index value of the nomogram model was 0.75. Calibration curves are shown in Figure 6c. Moreover, we conducted Kaplan–Meier analyses in the TCGA-LUAD cohort to compare survival differences between groups with different nomogram scores (Figure 6d). Furthermore, we validated the prognostic value of the nomogram score in the GSE30219 cohort and observed that LUAD patients with higher nomogram scores had shorter OS (p < 0.0001, Figure 6e) and PFS (p < 0.0001, Figure 6f). Additionally, we calculated the area under the ROC curve (AUC) values to assess nomogram performance in the TCGA and GSE30219 cohorts and demonstrated excellent performance of the nomogram model in predicting the survival of LUAD patients (Figure 6g). Finally, we explored the relationship between nomogram scores and RCD scores and revealed that they were positively correlated (Figure 6h) and that the RCD scores of patients with high nomogram scores were significantly higher than those of patients with low nomogram scores (Figure 6i).

3.7 Decipherment of the tumor immune microenvironment based on the RCD signature

Furthermore, we performed tumor microenvironment analyses to identify potential differences in tumor-immune features between two different RCD groups as well as the two nomogram groups. First, we identified whether there were differences in immune scores, tumor purity, stroma scores, and microenvironment scores. Violin plots of the ESTIMATE and XCELL microenvironment scores indicated that tumors with low RCD and nomogram scores might possess stronger immune activities (Figure 7a,b). Then, we used various deconvolution algorithms to acquire the enrichment scores of multiple immune cells. Bar plots showed that the abundance of multiple immune cells was significantly different between the high- and low-RCD groups (Figure 7c), as well as between the high- and low-nomogram groups (Figure 7d). The correlation heatmap indicated that both RCD scores and nomogram scores were negatively correlated with the enrichment of most immune cells (Figure 7e). Strikingly, three antigen-presenting cells (APCs), including dendritic cells, B cells, and macrophages, were all significantly higher in the RCD-low group, with a significant negative correlation with RCD scores (Figure 7f), suggesting that the 17-gene signature may affect antigen presentation function to influence prognosis.

3.8 Prediction of chemotherapy and immunotherapy sensitivity based on the RCD signature

To further elucidate the relationship between the RCD signature and drug sensitivity, we predicted each drug's half maximal inhibitory concentration (IC50) value in LUAD samples by the oncoPredict R package according to the gene expression or drug response data of the GDSC2 data set. The landscape of the correlation between drug sensitivities and RCD scores as well as nomogram scores is shown in Figure 8a. The results identified drugs with significant differences in therapeutic responses between the high and low RCD groups. We found that the IC50 values of eight drugs, including doramapimod, axitinib, GSK269962A, JQ1, ribociclib, AZD6482, SB216763, and TAF1 5496, were significantly higher in the RCD high group than in the RCD low group (Figure 8b, Supporting Information: Figure S3A), suggesting that LUAD patients with high RCD scores were more resistant to these chemotherapeutic drugs.

To explore ICI immunotherapy sensitivity, we performed TIDE and Easier immunotherapy response assessments in each LUAD patient. In TIDE explorations, we revealed that the TIDE scores were higher in the RCD low group, with a significant negative correlation (Figure 8c, Supporting Information: Figure S3B). In Easier evaluations, we integrated molecular pathway activity (Figure 8d), transcription factor activity (Figure 8e), and immune cell fractions (Figure 8f, Supporting Information: Figure S3C) to obtain ICI response scores. We revealed significant differences in the immune response between the RCD score high and low groups (Figure 8d–g). Easier and TIDE immunotherapy prediction results indicated that LUAD patients with low RCD scores were more likely to respond to immunotherapy, while LUAD patients with high RCD scores may not benefit from immunotherapy (Figure 8c,h), suggesting that high RCD activities are significantly correlated with low immunotherapy sensitivity.