Introduction

Lung cancer (LC) is one of the most common malignant tumors in humans and has the highest mortality rate among all cancers1,2. Smoking remains the main cause of LC, and in addition, genetic and dietary risk factors also play an important role in the pathogenesis of LC3. According to the histopathological features of LC, it can be divided into non-small cell lung cancer (NSCLC) and small cell lung cancer (SCLC), while LUAD, the main histological type of LC belongs to NSCLC4,5. LUAD mainly originates from bronchial mucinous glands and can occur in the bronchioles or central airways. And LUAD possesses extensive inter- and intra-tumoral histological heterogeneity, and it contains four subtypes which are adenocarcinoma in situ (AIS), minimally invasive adenocarcinoma (MIA), Invasive adenocarcinoma and Invasive adenocarcinoma variant6,7. In the treatment of LUAD, despite the significant advances in therapeutic treatments for LUAD, most patients still suffer from drug‐resistance or insensitivity to current regimens, and the 5-year overall survival (OS) of LUAD remains below 20%8,9. Herein, it is imperative to improve the existing intervention measures for LUAD for the prognosis of patients.

Cell death is a fundamental physiological process in all living organisms, occurring every day in billions of cells in the human body. It plays a significant role in embryonic development, organ maintenance, aging, and the coordination of immune responses and autoimmunity10. The Nomenclature of Cell Death Committee (NCCD) has meticulously formulated and elucidated the concept of cell death by considering morphological, biochemical, and functional perspectives. This comprehensive analysis has resulted in the categorization of cell death into two distinct modalities: accidental cell death (ACD) and PCD11. ACD is an uncontrolled process of cell death triggered by accidental stimulation of injury which exceed the regulatory capacity of cells. PCD, also referred to as regulated cell death (RCD), refers to the spontaneous and orderly death of cells under the control of genes in order to maintain the stability of the internal environment. The presently recognized forms of PCD predominantly encompass apoptosis, autophagy-dependent cell death, necroptosis, parthanatos, entotic cell death, netotic cell death, lysosome-dependent cell death (LCD), ferroptosis, cuproptosis, pyroptosis, oxeiptosis, and alkaliptosis12. Apoptosis is a form of cellular suicide characterized by several morphological features including cell shrinkage, chromatin condensation, membrane blebbing, deoxyribonucleic acid fragmentation and the formation of apoptotic bodies13. On autophagy-dependent cell death, autophagy, a metabolic cellular mechanism, degrades proteins and/or organelles by lysosomes as a cytoprotective response. Under conditions of nutritional deprivation or cellular stress, autophagy promotes homeostasis and bioenergy efficiency through the recycling of cellular components14. Necroptosis is a programmer form of necrosis that relies on the phosphorylation of mixed -lineage kinase-like (MLKL) by receptor interacting kinase-1(RIPK1) and receptor interacting kinase-3 (RIPK3) which characterized by plasma membrane infiltration, cell swelling, and loss of cellular and organelle integrity15. Parthanatos is a poly [ADP-Ribose] polymerase 1 (PARP1)-dependent PCD that is triggered by oxidative stress-induced DNA damage and chromatinolysis16. Entotic cell death is a form of cannibalism in which one cell engulfs and kills another cell. The presumable fates of engulfed cells include division, release, or death16. Netotic cell death is driven by NET release. NETs are extracellular net-like DNA–protein structures released by cells in response to infection or injury17. LCD, also known as lysosomal cell death, is mediated by hydrolytic enzymes that are released into the cytoplasm after lysosomal membrane permeabilization. LCD is associated with inflammation, tissue remodeling, aging, cardiovascular disorders and so on18. Ferroptosis is an iron-dependent form of cell death caused by the toxic accumulation of lipid peroxides on the cell membrane. The features of ferroptosis include reduced mitochondrial volume, mitochondrial outer membrane, reduced or absent mitochondrial crest and normal nuclear size without nuclear concentration19,20. Cuproptosis is a Cu-induced cell death pathway closely related to mitochondrial metabolism21. Compared with necroptosis and ferroptosis, pyroptosis is the major mechanism of host defense and is crucial for bridging innate and adaptive immunity. Pyroptosis is activated by the activation of inflammasome sensors, ultimately leading to the loss of plasma membrane integrity22,23. Oxeiptosis is a novel oxygen radical-induced caspase-independent PCD driven by the activation of the KEAP1-PGAM5-AIFM1 pathway16. Alkaliptosis is a novel PCD driven by intracellular alkalization16. Furthermore, when the process of PCD proceeds normally, the life span of cells does not increase abnormally, which results in insufficient time for division and proliferation of tumor cells. Similarly, failures in PCD promote the occurrence and development of cancer by means of creating a permissive environment for genetic instability and accumulation of gene mutations, supporting anchorage-independent survival during metastasis and strengthening the resistance to damage from the immune system24.

It is generally known that many cell surface proteins relay important signals that ultimately determine cell fate including cell differentiation, division, migration and death25. To strictly ensure the quality of these proteins, there is a protein folding area in endoplasmic reticulum (ER). When the protein folding demand overpowers the folding capacity of the ER, ER stress occurs26. The unfolded protein response (UPR) is an intracellular ER-to-nucleus signal transduction pathway that is triggered by the accumulation of unfolded proteins in the ER to cope with and adapt to ER stress27. However, with the high levels of ER stress persisting, a terminal UPR program induces PCD of cells28. The abnormal accumulation unfold protein in ER was regarded as a possible driver of various cancers27. Meanwhile, the hazardous microenvironment in malignant tumors may destroy the homeostasis of ER and induce the occurrence of UPR29. Herein, exploring the potential relationship between PCD and UPR in tumors is also significant.

Building upon prior investigations, it is evident that PCD holds promise as a potent tool in the treatment or control of malignant tumor cells. However, a comprehensive and in-depth understanding of the role that PCD plays in LUAD remains elusive based on current knowledge30. However, based on current research, we have no way to gain a profound and comprehensive understanding of the role of PCD in LUAD. Given the prominent characteristic of PCD, this research aimed to exclaim a novel signature, CDI, which was utilized in evaluating the prognosis, heterogeneity, and response to diversified therapy strategies in LUAD31.

Materials and methods

The step-by-step procedures of our study was exhibited in Fig. 1.

Fig. 1
figure 1

The general framework for the holistic analysis of diversified PCD patterns in LUAD.

Full size image

Data source and download

The genes related to the above 14 kinds of PCD were rigorously collected and arranged from several public websites and literatures, such as MSigDB, GeneCards, KEGG, and review articles16,31,32,33. Altogether, a total of 1515 genes that regulate the 14 kinds of PCD mutually were concentrated, including 580 genes of apoptosis, 52 of pyroptosis, 88 of ferroptosis, 368 of autophagy, 101 of necroptosis, 14 of cuproptosis, 24 of disulfidptosis, 9 of parthanatos, 15 of entotic cell death, 8 of netotic cell death, 220 of LCD, 7 of alkaliptosis, 6 of oxeiptosis, 25 of zinc-dependent cell death. And the detailed information of these 1515 genes were exhibited in Table S1.

The inclusion criteria of patients in our research were as follows: (1) Powerful evidence for the diagnosis of LC, (2) Confirming LUAD at the histopathological level, (3) Available survival time, status and clinical information. Herein, a total of 1602 LUAD patients from six various and independent cohorts were accessed from two public databases: The Cancer Genome Atlas (TCGA) and GEO (Table S2)34,35. The normalized and log2(x + 1) converted RNA-sequencing (RNA-seq) count profile of the TCGA-LUAD cohort were utilized for subsequent analysis. Besides, single-cell RNA (sc-RNA) transcriptome data of LUAD patients were collected from the eleven LUAD samples in the GSE131907.

Calculation of normalized weight of each PCD in LUAD

According to the PCD gene sets in Table S1 collected in advance, we utilized the method “single sample gene set enrichment analysis (ssGSEA)” to calculated the weight of each PCD in each sample of the TCGA-LUAD cohort via the “GSVA” package (version 1.40.1) in R software, in which we presumed the minimum gene set to 5 and the maximum gene set to 500036. The “survival” package and log-rank test was performed to evaluate the prognostic role of each PCD in LUAD37. And the criteria to select significant PCD types was p < 0.05.

Exploration of the expression pattern and mutation level of genes related to PCD

Initially, the differential expression analysis between the tumor and normal samples of the TCGA-LUAD cohort was performed via the “limma” package in the R software to extract the differentially expressed genes (DEGs) with the following criteria: p < 0.05 and |log2 FC|> 138. Additionally, the “maftools” package in the R software was used to explore the mutation landscape among the LUAD patients39. The intersection between the DEGs and genes related to significant PCD types were considered as the DEGs related to PCD (PCD-DEGs).

Establishment and validation of the CDI signature

To develop a consensus prognostic model in LUAD with accuracy and stability, firstly, we utilized the univariate-Cox regression to detect the prognostic role of the above candidate PCD-DEGs with the cut-off p < 0.05. Subsequently, totally 101 algorithm patterns combined via 10 machine learning methods were utilized in the above selected prognostic PCD-DEGs within the TCGA-LUAD cohort as training set, GSE72094 as test set and other four datasets as validation sets, respectively. The 10 independent algorithms for forming the 101 patterns were as follows: CoxBoost, elastic network (Enet), generalised boosted regression modelling (GBM), least absolute shrinkage and selection operator (Lasso), partial least squares regression for Cox (plsRCox), random survival forest (RSF), Ridge, supervised principal components (SuperPC), stepwise Cox, survival support vector machine (survival-SVM). The C-index of each pattern was calculated in each dataset and regarded as the criteria to ensure the most appropriate algorithm pattern40,41. Finally, we imported the significant prognostic PCD-DEGs into the most appropriate algorithm pattern to establish the prognostic signature with the following formula: CDI = Σ (Ci * Ei). The Ci indicates the coefficients calculated via the plsRcox of each PCD-DEG and the Ei represents the expression value of each PCD-DEG. Then, the Kaplan–Meier (KM) curve was drawn via the “survival” R package to evaluate the prognostic role of the CDI signature42. We used the “maxstat” R package (version 0.7–25) to calculate the most appropriate cut-off to divide LUAD patients into the high- and low-CDI group43. Via the “survivalROC” R package, we exhibited the time-dependent receiver operating characteristic (ROC) curves, in which the area under the curve (AUC) was regarded as the prognostic accuracy of our CDI signature44. Similar analysis was performed in the validation sets to verify our CDI signature.

Moreover, to exhibit the superior role of CDI signature in the field of PCD in LUAD patients, we comprehensively retrieved the published literatures with accessible prognostic models associated with PCD in PubMed up to August 1, 2023. Next, according to the model formula provided in these literatures, we systematically calculated the corresponding risk scores in all the six cohorts included in our research and assessed the prognostic role via univariate-Cox regression analysis and comparing the C-index of each model with CDI.

Enrichment analysis

For the functional enrichment analysis of PCD-DEGs, the “org.Hs.eg.db” R package (version 3.17) was used to obtain the annotation of Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) database45,46,47. Subsequently, the “clusterProfiler” R package (version 3.14.3) was used to finish the enrichment analysis48,49. Besides, within the gene set “h.all.v7.5.1.entrez.gmt” [HALLMARK], we performed the gene set enrichment analysis (GSEA) to investigate the biological background between the high- and low-CDI group via the R package “GSEABase” (version 3.17) and “GSVA” (version 1.40.1)36,50.

Description of the TME

Diversified algorithm of immune infiltration were adopted in our research, including CIBERSORT, ESTIMATE, and TIMER via the corresponding methods in the R package “IOBR”51. Meanwhile, the ssGSEA method via the R package “GSVA” was also performed to calculate the infiltration score of 28 different immune cells, the gene set utilized in which was exhibited in Table S336. Moreover, we used the “SeuratV5” package in the R software to detect the distribution of CDI and the model PCD-DEGs52,53,54.

Investigation of the novel therapy strategies

We firstly collected 9 common immune checkpoints (ICKs) and detected whether there was a significant difference of the expression of the 9 ICKs between the high- and low- CDI group in the TCGA-LUAD cohort55. The public website Tumor Immune Dysfunction and Exclusion (TIDE) was utilized to analysis the response for the immunotherapy via the inhibitors of ICKs in the TCGA-LUAD patients, such as anti-PD1, anti-CTLA4, and so on56. The LUAD patients with TIDE score > 0 were considered the non-responders while with TIDE score < 0 were regarded as the responders. Additionally, the Dysfunction and Exclusion score were also compared between the high- and low-CDI group. Besides, we verified the predictive role to immunotherapy of the CDI signature in other cancers via the immunotherapy datasets collected from the Tumor Immunotherapy Gene Expression Resource (TIGER) database, which were also exhibited in Table S257. Moreover, the prediction for the sensitivity of several drugs in the Genomics of Drug Sensitivity in Cancer (GDSC) database were performed via the “oncoPredict” package in the R software58,59.

Construction and evaluation of the nomogram combined with various clinical indexes

Integrating various clinical information (Age, Clinical T, N, M, and Pathological stage), we performed the univariate- and multivariate-Cox regression analysis to confirm whether CDI can act as an independent prognostic factor in LUAD patients, respectively60. Furthermore, a clinical nomogram was established among CDI and the above clinical indexes via the multivariate-Cox and stepwise regression analyses in the R packages “rms” and “regplot”61. The calibration curve was created via the “caret” package and the clinical decision curve was drawn via the “rmda” package in the R software.

RT-qPCR

To validate the expression level of our significant PCD-related features in our CDI model between Tumor and Normal group, we collected clinical tissue samples from the Second Affiliated Hospital of Nanchang University. Samples of LUAD (n = 10) were obtained exclusively from patients who underwent surgery. While the fresh tissues of normal group (n = 10) were obtained from the biopsy of lung nodule tissues with diagnosis of non-tumor. The method for qPCR was described in our previously published paper62. The protocol for collecting human tissue samples was approved by Ethics Committee of the Second Affiliated Hospital of Nanchang University (Jiangxi, China). Written informed consent was provided by all participants before enrollment.

IHC

LUAD and adjacent tissues were fixed, embedded in paraffin, sectioned, and deparaffinized. CDK5R1 and DAPK2 was detected using the anti-P35 antibody (MA5-14834; Thermo Scientific, Waltham, MA, USA) and anti-DAPK2 antibody [EP1633Y] (ab51061). A peroxidase/3,3’-diaminobenzidine-conjugated secondary antibody was used for visualization.

Western blotting

LUAD tissues were collected and cell extracts were prepared in lysis buffer. Protein expression was detected by incubation with above primary antibodies and secondary antibodies overnight at 4 °C. After washing the membranes with Tris-buffered saline with 0.1% Tween-20 (TBST), protein bands were detected by ECL.

Cell culture

The A549 LUAD cell lines were purchased from Shanghai Institute of Cell Research (Cell Resource Center, Shanghai Institutes for Biological Sciences, Chinese Academy of Sciences). These cells were cultured at 37 °C in a humidified atmosphere containing 5% CO2 in Dulbecco’s modified Eagle’s medium (cat.No.10–013-CVR, Corning Inc., USA) supplemented with 10% fetal bovine serum (FBS, Life Technologies). The DAPK2 (GenBank NM_014326) overexpressing lentiviral vector (Lv- DAPK2-1, Lv- DAPK2-2, Lv- DAPK2-3) and lentiviral negative control (NC-OE) were constructed by Shanghai Genechem Co., Ltd. Lentivirus production and transduction were done as described30,41. Cells were transfected using Lipofectamine RNAiMAX (Life Technologies) with siRNA duplexes against rat Cdk5, 5’-GUUCAGCCCUCCGGGAGAUTT-3’, or a previously described scrambled control, 5’-GAGACCCTATCCGTGATTA-3’. Following an overnight incubation, cells were transduced with AdCMV-GFP or AdCMV-Cdk5r1 or left untreated.

Transwell assay

For the transwell assays, logarithmic growth-phase melanoma cells were digested with trypsin and counted. Cells in 600μL of serum-free medium were inculated into the upper chamber and placed in a 24-well plate containing medium with 10% FBS. After 48 h of culture at 5% CO2 and 37 °C, melanoma cells were fixed in ethanol for 15 min and stained with 0.1% crystal violet for 25 min. Non-migrating cells in the upper layer were gently wiped away with a cotton swab. Finally, the chambers were washed three times with phosphate-buffered saline and photographed under a microscope.

Wound healing

For the wound healing assay, 5 × 105 cells were seeded in 6-well plates, which were scratched with a pipette tip and washed with 1640/DMEM. Next, these cells were cultured at 37 °C for 24 h. The photographs were taken at 0 and 24 h.

Xenograft assay

Female athymic BALB/c nude mice (4 weeks old) were purchased from the Model Animal Research Center of Nanjing University (Nanjing, China) and subcutaneously injected with 1 × 106 control and cells. The tumor volume (cm3) was measured every 7 days. The standard formula was width2 × length × 0.5. These nude mice were sacrificed at 48 days. Next, the mice were anesthetized with isoflurane and subsequently euthanized by cervical dislocation. And the xenografts were removed and weighed. Finally, these xenograft samples were dehydrated, embedded in paraffin, and sectioned for staining.

Statistical analyses

All statistical analyses in our research were finished via the R software (version 4.3.1) and the GraphPad Prism (version 8.0.2). Statistical differences between two groups were compared via the unpaired t test. And the one-way ANOVA was used to detect the statistical differences between multiple groups. The log-rank test was utilized to compare the survival difference. All the correlation analyses in this research were performed via the method “Pearson”. Throughout the text, p < 0.05 was considered with statistical significance.

Results

General investigation of PCD landscape in LUAD

To comprehensively explore the PCD landscape in LUAD, a total of 1515 genes from 14 kinds of PCD were enrolled in our research. At first, we performed the ssGSEA algorithm to generally investigate the normalized weight of each PCD. In the TCGA-LUAD cohort, except of 2 PCD without weight via ssGSEA, all the 12 PCD exhibited statistical significance between normal and tumor group (p < 0.05, Fig. 2A). Then, the KM curves revealed the prognostic role of each PCD (Fig. 2B). According to the criteria (p < 0.05), 7 kinds of PCD were selected as the crucial PCD form in LUAD, including autophagy, cuproptosis, disulfidptosis, entotic cell death, LCD, netotic cell death, zinc-dependent cell death. Therefore, the 635 genes of the above 7 PCD were utilized in the subsequent analysis.

Fig. 2
figure 2

General Exploration of the PCD landscape in LUAD patients. (A) Box plots exhibited the weights of each PCD types between tumor and normal samples. (B) KM curves revealed the prognostic role of each PCD types in LUAD patients of the TCGA-LUAD cohort. (C) Volcano plot of the DEGs between tumor and normal samples. (D) Venn-gram of the interaction between the significant DEGs and genes related to prognostic PCD types. (E, F) Enrichment analysis of the 80 PCD-DEGs via (E) GO and (F) KEGG database. (G) The oncoplot exhibited the SNV mutation landscape of the PCD-DEGs in LUAD. (H) A forest plot of the PCD-DEGs with p < 0.05 via univariate-Cox regression analysis.

Full size image

Furthermore, the differential expression analysis in the TCGA-LUAD cohort totally identified 4430 DEGs, including 1820 up-regulation and 2610 down-regulation (Fig. 2C). Via intersecting the above 635 genes and the 4430 DEGs, 80 DEGs related to PCD were obtained as PCD-DEGs (Fig. 2D). All the PCD-DEGs were listed in the Table S2. Additionally, the GO and KEGG enrichment analysis indicated that diversified biological pathways mediated the PCD landscape in LUAD, such as macroautophagy, mast cell activation, adrenergic receptor activity, NOD-like receptor signaling pathway, AMPK signaling pathway, etc. (Figs. 2E, F). Besides, we also evaluated the mutation role of 80 PCD-DEGs in TCGA-LUAD cohort. Overall, approximately 32.63% (185/567) LUAD patients occurred genetic mutations, in which the most prevalent mutation type was missense-mutation. And the top 15 mutant PCD-DEGs are exhibited, with KDR possessing the highest mutation frequency (8%) (Fig. 2G). Finally, to explore the prognostic role of 80 PCD-DEGs, the univariate-cox regression analysis was performed and 31 significantly prognostic PCD-DEGs were identified with p < 0.05 (Fig. 2H). The complete results of univariate-cox regression analysis were exhibited in Table S3.

Systematic construction of a consensus prognostic PCD-signature in LUAD

To systematically and rigorously construct our consensus prognostic PCD-signature in LUAD, a total of 101 artificial combinations consisted of 10 machine learning algorithms with a tenfold cross-validation framework was computed. In the execution process of the algorithms, we selected TCGA-LUAD cohort as our training set, the external dataset with the largest sample size GSE77294 as our testing set, the other four external datasets as our validation sets. The 31 significantly prognostic PCD-DEGs were imported as the source features and the average C-index was set as the selection criteria. Among the four combinations with the highest average C-index (0.66), the Lasso + plsRcox the highest average C-index in the four validation sets. As a result of comprehensive screening, the model Lasso + plsRcox was confirmed with high accuracy and translational relevance (Fig. 3A and Table S4).

Fig. 3
figure 3

Establishment of the consensus prognostic PCD-signature, CDI, in LUAD patients. (A) The 101 algorithm patterns consisted of 10 machine learning methods to construct the CDI signature and further calculated the C-index across all datasets. (B, C) Performing the Lasso algorithm determined the most suitable λ to select the most useful prognostic PCD-DEGs. (D) Coefficients of 14 PCD-DEGs calculated via the plsRcox method to form final CDI signature. (E–I) Scatter plots exhibited the distribution of CDI between different (E) survival status, (F) T, (G) N, (H) M, and (I) pathological stages. (J) The heatmap revealed the expression level of 14 signature PCD-DEGs in the low- and high-CDI group and various clinical characteristics.

Full size image

Firstly, via Lasso Cox regression, 14 PCD-DEGs with non-zero coefficients were further subjected into plsRcox analysis, in which all the 14 PCD-DEGs exhibited a statistical significance (p < 0.05) (Figs. 3B–D). Among the 14 PCD-DEGs, six of these genes were derived from autophagy, six of these genes were obtained from LCD, one belonged to disulfidptosis and one was derived from netotic cell death. The co-expression network of the 14 PCD-DEGs was exhibited in Fig. S1A with the co-expression of 50.03%, the physical interactions of 35.19%, the pathway of 9.81% and the genetic interactions of 4.97%. Besides, we explore the correlation of each model gene (Fig. S1B) and the prognostic role of each PCD-DEG was investigated via the KM curves (all p < 0.05) (Fig. S2A–N). The expression level of these 14 genes between normal and tumor group in the TCGA-LUAD cohort were exhibited in Fig. S3 (all p < 0.0001). Consequently, the consensus prognostic PCD-signature indicated the CDI of each patient with LUAD according to the following formula: (ExpGAPDH * 0.11545648)—(ExpDAPK2 * 0.11659387)—(ExpNAPSA * 0.08873154)—(ExpPLA2G3 * 0.08760877)—(ExpNUPR1 * 0.07496043) + (ExpCDK5R1 * 0.07815069) + (ExpSLC7A11 * 0.06791662)—(ExpPIK3CG * 0.06415352)—(ExpGATA2 * 0.06824659)—(ExpCAMP * 0.0717327)—(ExpCTSW * 0.06837049)—(ExpHGF * 0.06330508)—(ExpGAB2 * 0.05121809)—(ExpMAPT * 0.05882832). Interestingly, CDI exhibited a prominent clinical significance combined with survival status, clinical T, N, M and pathological stage. The CDI was higher in the dead patients and indicated an upward trend with the increase of clinical T, N, M and pathological stage. And all clinical features exhibited a statistical significance (p < 0.05) except clinical M (Fig. 3E–H).

Comprehensive evaluation for the biological and prognostic role of the CDI model

Via the prognostic heat map of our CDI model, we revealed the relationship between CDI and the survival status in the TCGA-LUAD cohort, obviously in which the survival rate of patients with high CDI is lower than that of patients with low CDI (Fig. S4A). And the KM curve for CDI also confirmed the above state (p < 0.05, Fig. S4A). Besides, the consistent phenomenon also occurred in all the validation sets (GSE77294, GSE50081, GSE30219, GSE31210, GSE37745), which further emphasizes the accurate prognostic role of the CDI model (Fig. S4B–F). Meanwhile, the LUAD patients with higher CDI exhibited remarkably poorer disease-specific survival (DSS) and disease-free interval (DFI) (all p < 0.05, Fig. S2O, P). Furthermore, to better demonstrate the discrimination of the CDI model, the time-dependent ROC curves were exhibited, with 365-, 1095-, and 1825-day AUCs of 0.72, 0.67, and 0.61 in the TCGA-LUAD cohort; 0.68, 0.70, and 0.71 in the GSE77294 cohort; 0.58, 0.59, and 0.59 in the GSE50081 cohort; 0.57, 0.59, and 0.61 in the GSE37745 cohort; 0.71, 0.61, and 0.61 in the GSE31210 cohort; and 0.58, 0.61, and 0.65 in the GSE30219 cohort, respectively (Fig. S4A–F).

Moreover, with the crucial role of PCD in diversified diseases and the remarkable development of artificial intelligence and machine learning, massive prognostic and predictive models related to PCD have been exploited. To more comprehensively delineate the advantages of our CDI model within the field of PCD in LUAD, we collected all 35 publicly published literature on constructing prognostic and predictive mRNA-models via machine learning in the past and compare the clinical performance with our CDI model. We performed KM prognostic analysis for CDI and all the previous models in all the six datasets included in our research and indicated that only a small portion of the models play a consistent and significant role in all the six datasets (Fig. 4A). Besides, via comparing the C-index between CDI and other models, notably, our CDI model stamped better property than the majority of published models in almost all the datasets except GSE31210, which further revealed the stability and superiority of the CDI model (Figs. 4B–G).

Fig. 4
figure 4

Comparison of the prognostic accuracy between the CDI signature and other published models associated with PCD in LUAD. (A)The heat map exhibited the prognostic characteristic of CDI and other published models in TCGA-LUAD, GSE30219, GSE31210, GSE72094, GSE50081, and GSE37745 via univariate-Cox regression analysis. (B–G) Lollipop diagrams indicated the C-index across all PCD-signatures in (B) TCGA-LUAD, (C) GSE30219, (D) GSE31210, (E) GSE72094, (F) GSE50081, and (G) GSE37745.

Full size image

Meanwhile, the biological background of our CDI model in LUAD was also evaluated. We performed GSEA to explore the functional landscape between high- and low-CDI group, which was divided according to the most suitable cut-off calculated via log-rank test. Noteworthily, the pathway “UPR” was found with the highest normalized enrichment score (NES) with p < 0.05, which exhibited a higher trend in the high-CDI group (Fig. S5A, B). Then, we demonstrated the correlated relationship between the expression levels of 14 PCD-DEGs in the CDI model and the weight of UPR. Surprisingly, all the 14 PCD-DEGs exhibited a significant correlation with the UPR, which further revealed the crucial role of UPR in LUAD (Fig. S5C–P).

Exploration of the TME mediated by CDI in LUAD

According to the high- and low-CDI group divided by the log-rank test, we explored the TME via various immune infiltration algorithms, including CIBERSORT, ssGSEA, ESTIMATE, and TIMER. For ESTIMATE, all the three indexes (Stromal Score, Immune Score, and ESTIMATE Score) exhibited a higher distribution in the low-CDI group (Fig. 5A). Moreover, the infiltration results calculated via other algorithms were consistent, which revealed that the majority of immune cells exhibited a higher infiltration score in the low-CDI group (Figs. 5B–D). And abundant studies have revealed that macrophages played a crucial role in PCD, inflammation, and cancers. Two various types of macrophages, M1-like and M2-like, maintain a balance between pro-inflammatory and anti-inflammatory responses63. Herein, we detected the ratio of M1-like macrophages/M2-like macrophages (M1/M2 Ratio) to evaluate the degree of immune and inflammatory response between the high- and low-CDI group in the TCGA-LUAD cohort. As a result, the M1/M2 ratio was higher in the high-CDI group and was positively correlated with the CDI (all p < 0.01, Figs. 5E, F). Subsequently, we investigated the specific distribution of CDI in the 11 tumor tissues of LUAD patients of sc-RNA dataset GSE131907. The major cell types in GSE1319071 were annotated. After filtering out cells with poor quality, a total of 9 cell types were identified (Fig. 5G). We found that the CDI was significantly higher in endotheliar cells, monocytes and macrophages (Figs. 5H, I). Besides, the bubble plot exhibited the average expression of the 14 PCD-DEGs in each cell type in detail (Fig. 5J).

Fig. 5
figure 5

Comprehensive Exploration of the TME in LUAD. (A) The distribution of stromal score, immune score and ESTIMATE score between the high- and low-CDI group calculated via the ESTIMATE algorithm. (B) The distribution of the infiltration score of B cells, Dentritic cells, Macrophages, Neutrophils, CD4 T cells, and CD8 T cells calculated via the TIMER algorithm. (C) The distribution of the infiltration score of 22 different immune cells calculated via the CIBERSORT algorithm. (D) The distribution of the infiltration score of 28 different immune cells calculated via the ssGSEA algorithm. (E) The difference of M1/M2 ratio between the high- and low-CDI group. (F) The correlation between the CDI and M1/M2 ratio. (G) UMAP plot visualized all cell types from the 11 LUAD patients in GSE131907. (H) UMAP plot exhibited the distribution of CDI in all the cell types. (I) Scatter plot exhibited the distribution of CDI across each cell type in detail. (J) Bubble plot revealed the expression levels of all 14 signature PCD-DEGs in each cell type.

Full size image

Accurate prediction of immunotherapy and potential therapeutic drugs

We firstly collected the expression level of several common ICKs in the TCGA-LUAD cohort and compared the expression difference between the high- and low-CDI group, including CD28, CD80, CD86, CD274, CTLA4, ICOS, ICOSG, PDCD1, and PDCD1LG2. Interestingly, all the 9 ICKs exhibited a higher expression in the low-CDI group (Fig. 6A). Subsequently, we performed TIDE analysis to predict the effect of immunotherapy in LUAD patients. As a result, the TIDE score was higher in the low-CDI group and the CDI was higher in the non-response patients (Figs. 6B, C). Besides, two classical indicators to predict the effect of immunotherapy (TMB and MSI), were also calculated and compared between the high- and low-CDI group (Figs. 6D, E). Moreover, the dysfunction score was higher while the exclusion score was lower in the low-CDI group, which were consistent with the above results (Figs. 6F, G). Meanwhile, we detected the prognostic role of TIDE score via the KM curve, which revealed that the higher TIDE score may suffer a worsen survival outcome (p < 0.05, Fig. 6H). Additionally, the response rate was higher in the low-CDI group (Fig. 6I). Overall, the sensitivity and effects of immunotherapy in the low-CDI group were significantly more superior than in the high-CDI group. To validate the predictive accuracy of the CDI, we evaluated our CDI model in the immunotherapy datasets of another three cancers, including IMvigor210 of bladder urothelial carcinoma (BLCA), Braun_2020 of renal cell carcinoma (RCC), and GSE91061 of melanoma (MM). Surprisingly, in all the three datasets, the prognostic role and prediction of immunotherapy were consistent with the phenomenon in LUAD, which further emphasize the stability and effectiveness of CDI in the field of immunotherapy (Fig. 6J–O).

Fig. 6
figure 6

Exhibition of the predictive accuracy in immunotherapy of CDI signature. (A) The expression difference of several common ICKs between the high- and low-CDI group. (B) The distribution of TIDE score between the high- and low-CDI group. (C) The trend of CDI between the non-responders and responders. (D–G) The distribution of (D) TMB, (E) MSI, (F) Dysfunction score, and (G) Exclusion score between the high- and low-CDI group. (H) KM curve indicated the prognostic role of TIDE score in TCGA-LUAD. (I) The percent of non-responders and responders in the high- and low-CDI group. (J, L, N) KM curves validated the prognostic role of CDI in (J) IMvigor210 of BLCA, (L) Braun_2020 of RCC, and (N) GSE91061 of MM. (K, M, O) The percent of PD/CD and CR/PR between the high- and low-CDI group in (K) IMvigor210 of BLCA, (M) Braun_2020 of RCC, and (O) GSE91061 of MM.

Full size image

Then, to deeply investigate and accurately predict the sensitivity of chemotherapeutics within our CDI model, we calculated the half maximal inhibitory concentration (IC50) of various drugs in each sample of the TCGA-LUAD patients. And the correlated coefficient (r) between the IC50 and CDI was calculated. Based on the criteria of p < 0.05 and the absolute value of r > 0.7, a total of 14 potential drugs were identified (Fig. 7A). In the high-CDI group, only trametinib exhibited a lower IC50 than in the low-CDI group (Fig. 7B–P). Therefore, for LUAD patients with high CDI and chemotherapy-resistant, trametinib may play an unexpected and surprising result.

Fig. 7
figure 7

General landscape of the effectiveness of the potential predictive therapeutic drugs. (A) Lollipop diagram exhibited the correlated relationship between the CDI and the IC50 of several predictive drugs. (B–O) The IC50 of (B) Acetalax, (C) Afatinib, (D) AZD8186, (E) Ipatasertib, (F) Alpelisib, (G) AZD1332, (H) AZD5582, (I) JQ1, (J) Sabutoclax, (K) AZD8055, (L) AZD8055 (M) Sinularin, (N) Tasclisib, (O) Tramctinib and (P) Uprosertib between the high- and low-CDI group.

Full size image

Emphasizing the prognostic role of CDI and integration of a nomorgam risk model with clinical information

The prognostic role of CDI was firstly evaluated via performing the univariate- and multivariate-Cox regression analysis including 337 patients in the TCGA-LUAD cohort with complete clinical information. Via univariate-Cox regression analysis, each clinical information and CDI were detected whether can be regarded as an independent prognostic factor. As a result, we considered CDI as an independent risk prognostic factor (p < 0.0001, HR = 2.04, 95% CI 1.64–2.55; Fig. 8A). Meanwhile, via multivariate-Cox regression analysis combined with all the clinical information to normalize the p, the result also demonstrated the remarkable prognostic role of CDI in LUAD (p < 0.0001, HR = 1.88, 95% CI 1.47–2.40; Fig. 8B). Then, to avoid the one-sidedness of using CDI as an independent prognostic indicator, we integrated various clinical information and CDI to perform multivariable-Cox and step-wise regression analysis to construct a comprehensive nomogram scoring model in the TCGA-LUAD cohort, in which CDI, Age, and Stage were included (p < 0.0001, C-index: 0.721, 95% CI 0.674–0.768; Fig. 8C). And the calibration curve described the predictive reliability of the nomogram model in the 1-, 3-, and 5-year (Fig. 8D). Moreover, the clinical decision curves in 1-, 3- and 5-year all indicated the superior performance of nomogram score than these independent prognostic indexes included in the research (Fig. 8E). To further validate the predictive prognostic characteristic of the nomogram model, the time-dependent ROC curve was exhibited and the AUCs of 0.78 in 1-year, 0.72 in 3-year, and 0.65 in 5-year (Fig. 8F). Meanwhile, the KM cure indicated the nomogram score as a risk prognostic factor in LUAD (p < 0.0001, Fig. 8G).

Fig. 8
figure 8

Indication of the clinical characteristic of CDI via a nomogram. (A, B) Description of the independent prognostic role of CDI combining with various clinical characteristic via (A) univariate-Cox regression analysis and (B) multivariate-Cox regression analysis. (C) The clinical nomogram was constructed by CDI, age and pathological stage. (D–G) The (D) calibration curves, (E) clinical decision curves, (F) time-dependent ROC curves, and (G) KM curve of the clinical nomogram.

Full size image

Validation and exploration for the functional background of CDI

Firstly, we collected an in-house cohort including 10 LUAD samples and 10 normal samples to validate the 14 PCD-related genes in the CDI model via the RT-qPCR in the mRNA level. As a result, all the model genes exhibited a consistent trend in the LUAD group except NAPSA and PLA2G3 without significance (Fig. 9A). Moreover, we selected the most crucial risk factor CDK5R1 and protective factor DAPK2 to perform further validation, respectively. Both the IHC and WB exhibited the higher expression of CDK5R1 and lower expression of DAPK2 in the tumor group in the protein level (Fig. 9B, C). And the Fig. 9D and E confirmed the statistical significance of the expression trend of CDK5R1 and DAPK2 protein.

Fig. 9
figure 9

Verification of our PCD-related features in the CDI model. (A) The verification of our PCD-related features in the mRNA level via RT-qPCR. (B–E) The verification of the most significant risk feature (CDK5R1) and protective feature (DAPK2) in the protein level via IHC and WB.

Full size image

Furthermore, we performed additional experiments in vitro to explore the specific role of CDK5R1 and DAPK2 in LUAD. The transwell experiment exhibited that both the inhibition of CDK5R1 and over-expression of DAPK2 obviously decreased the invasion of LUAD in the A549 cell line (Fig. 10A). Meanwhile, we also utilized the wound healing experiment to verify the migration ability of si-CDK5R1 (Fig. 10B, D) and oe-DAPK2 (Fig. 10C, E). The proliferation of A549 LUAD cell line was also significantly inhibited in the si-CDK5R1 and oe-DAPK2 group (Fig. 10F).

Fig. 10
figure 10

Indication of the crucial role of CDK5R1, DAPK2, and UPR in LUAD within CDI. (A) Representative images of LUAD A549 cells after CDK5R1 knockdown by siRNA and the over expression of DAPK2 in the clone formation assay. (B, D) The migratory capacity of LUAD A549 cells transfected with si-CDK5R1 in the 0 h and 24 h were detected using wound-healing assays. (C, E) The migratory capacity of LUAD A549 cells transfected with oe-DAPK2 in the 0 h and 24 h were detected using wound-healing assays. (F, G) EdU cell proliferation detection method for detecting the proliferation of LUAD A549 cells with (F) si-CDK5R1 and (G) oe-DAPK2. (H–J) Subcutaneous tumor development in nude mice and observation of the roles of CDK5R1 and DAPK2 in the growth of BC tumors. (K) The expression of markers of UPR, including CHOP, phosphorylated ERK (p-ERK) and total ERK (t-ERK) in the protein level. (L) The expression of CHOP and the ratio of p-ERK1/t-ERK1 and p-ERK2/t-ERK2 between the high-CDI and low-CDI group.

Full size image

Subsequently, to further verify the effect of CDK5R1 and DAPK2 in vivo, A549 cells transfected with si-CDK5R1 or oe-DAPK2 were subcutaneously injected into nude mice. We found that tumor volumes and weights were obviously smaller than in the NC group (Fig. 10H–J), which indicated that CDK5R1 could promote while DAPK2 could inhibit the tumorigenicity of LUAD cells in vivo.

Moreover, based on the crucial role of UPR in LUAD within the framework of CDI, we calculated the CDI of each sample. Then, we selected 6 LUAD samples with the three top-scored and the top three lowest of CDI to further verify several markers of UPR between the low- and high-CDI group. Surprisingly, we found that in the high-CDI group the expression of CHOP and p-ERK/t-ERK was significantly higher, which confirmed the potential positive correlation between CDI and UPR (Fig. 10K, L).

Discussion

Reflecting on years of sustained scientific inquiry into PCD in tumors, this study embarked on an exploration and integration of fourteen distinct PCD types in LUAD. The investigation introduced a novel predictive prognostic model, termed CDI, which underwent meticulous validation across five external cohorts (GSE72094, GSE50081, GSE37745, GSE30219, and GSE31210) in a systematic and comprehensive manner. Moreover, leveraging diverse clinical information, a prominent nomogram was meticulously constructed. Delving into the biological context, correlations with the TME, immunotherapy, and potential therapeutic drugs were elucidated. Ultimately, through validation and functional experiments, the study underscored the robustness of the CDI signature and explored potential connections between PCD and the UPR in LUAD.

To the best of our knowledge, as a crucial mechanism to maintain the normal cellular activity, more and more research confirmed that PCD plays an indispensable characteristic in various diseases, including the occurrence, development, and metastasis of malignant tumors12,64,65. To put it in a nutshell, the CDI signature constructed via rigorous machine learning algorithms was consisted of fourteen elements (CAMP, CDK5R1, CTSW, DAPK2, GAB2, GAPDH, GATA2, HGF, MAPT, NAPSA, NUPR1, PIK3CG, PLA2G3, SLC7A11), which can accurately predict the OS, DSS, and DFI of LUAD patients. CAMP encodes the antimicrobial peptide, which has been regarded as a novel inhibitor to suppress the proliferation, migration, and invasion of cancers via the extensive antibiotic activity and toxic effects on cancer cells66,67,68. In our research, CAMP was considered as a protective factor of LUAD. CDK5R1 participated in cell cycle and mitosis, whose up-regulation promotes the proliferation of tumors like hepatocellular carcinoma, head and neck squamous cell carcinoma, and diffuse large B-cell lymphoma69,70,71. Meanwhile, CDK5R1 was regarded as a therapeutic target of quercetin in LUAD72. In our research, there was a higher expression of CDK5R1 in LUAD, which was a risky prognostic factor of LUAD. CTSW is a member of the peptidase C1 family and several research revealed that CTSW was associated with the cytolytic activity of T cells, which has been regarded as the candidate gene in inhibiting cancers73. Our study first considered CTSW as the prognostic indicator of LUAD. GAB2, as an adapter protein, regulated diversified oncogenic signaling pathways via cytokine, antigen, hormone, and so on74. However, in our analysis, further clarification is needed to verify whether GAB2 is a friend or an enemy in LUAD. GAPDH is a key enzyme in the process of glycolysis, which exhibited a high expression in almost all tissues75. Interestingly, nowadays several studies provided a novel insight that GAPDH is a regulator in PCD and potential biomarker in LUAD and other tumors, which was consistent with our research and remains abundant experimental confirmation76,77,78. The lack of GATA2 has been widely regarded as the significant inhibitor of malignant hematopoietic system tumors79,80. In our research, we filled the blank in the characteristic of GATA2 in LUAD and PCD. SLC7A11 was regarded as the suppressor of ferroptosis and significantly high-expressed in most of malignant tumors, including prostate cancer, LUAD, gastric cancer, and so on, which was consistent with our research81,82,83. As a crucial factor in diversified PCD types, especially apoptosis and autophagy, DAPK2 has been widely accepted that played a critical characteristic in suppressing tumors84. And latest research demonstrated that novel therapeutic strategies targeting on DAPK family may become a rising star in cancer treatment85. Moreover, the loss expression of DAPK family due to the aberrant promoter methylation was regarded as the potential mechanism of the occurrence and progress in LUAD86,87. However, detailed mechanisms of DAPK2 in LUAD remains unknown, which were worthy to deeper explore in the future. Other PCD-DEGs in the CDI signature played a various role in regulating the biological functions of malignant tumors, which should discuss the mechanisms in the specific cancer types.

TME, as the crucial characteristic which contributes prominently in providing nutrients necessary for the survival of tumor cells and inducing the growth and metastasis of malignant tumors88. In our research, both the expression level of ICKs and the infiltration status of various immune cells exhibited a lower trend in the high-CDI group, which indicated that the immune functions and responses were significantly suppressed, respectively. Moreover, cancer cells may possess resistance to multiple therapy strategies due to mutations that disrupt PCD, and avoiding PCD is one of the representative indicators of cancer, which mean the lower activity of PCD occurred in the high-CDI group12. Nowadays, more and more research indicated that PCD was a potential agonist of anti-inflammatory response, respectively89. Consistently, we found that in the high-CDI group, the M1/M2 ratio was higher and positively correlated with the CDI. As we all know, the two types of macrophages (M1-like and M2-like) were both closely associated with inflammation, in which M1-like macrophages mainly played a pro-inflammatory role while M2-like macrophages played an anti-inflammatory characteristic63. Besides, Tang et al. revealed that in the TME of LUAD, the polarization of M2-like macrophages was remarkably inhibited, which powerfully supported our research findings90.

For most potential anti-cancer drugs predicted via “oncopredict”, the IC50 of them was higher in the high-CDI group, like JQ1, uprosertib, and taselisib, which revealed that LUAD patients with high CDI may be resistant for several chemotherapy. Herein, the poor prognosis of patients in the high-CDI group may exhibited a potential reasonable explanation. Interestingly, trametinib seemed to exhibited a more sensitive therapeutic effect in the high-CDI group, which may be considered as a candidate for chemo-resistant LUAD treatment. Notably, our CDI signature exhibited a stable predictive ability for immunotherapy in LUAD and other solid tumors, including BLCA, MM, and RCC. Given the above, the specific CDI model may provide personalized therapeutic strategies for LUAD patients.

Conclusions

In conclusion, the CDI signature related to PCD constructed in this research is a particular prognostic indicator, according to which enable a potential personalized therapeutic strategies and accurate assessment of survival status for LUAD patients. Besides, the clinical tissues and cell lines demonstrated the remarkably positive correlation between the CDI and UPR. Overall, the comprehensive and systematical exploration of PCD was endowed strong clinical and biological significance for the first time.