The messenger RNA (mRNA) expression data (including 535 LUAD samples and 59 adjacent nontumor samples) and clinical information were downloaded from TCGA database (https://cancergenome.nih.gov). We also downloaded the following gene expression profiles from the GEO: GSE10072 (including 33 LUAD samples and 33 paired adjacent nontumor samples), GSE32863 (including 58 LUAD samples and 58 paired adjacent nontumor samples), GSE7503 (including 83 LUAD samples and 83 paired adjacent nontumor samples), GSE30219 (including 84 LUAD samples and 14 adjacent nontumor samples), GSE31210 (including 226 LUAD samples), GSE37745 (including 106 LUAD samples), and GSE3141 (including 58 LUAD samples). These were extracted from the GEO database (https://www.ncbi.nlm.nih.gov/geo/) to further validate our results. The Human Protein Atlas (HPA) database (http://www.proteinatlas.org/) was used to verify the expression of GPI in LUAD at the protein level. We obtained the gene amplification and mutation status of GPI through the cBioPortal for Cancer Genomics (http://www.cbioportal.org/).

The expression level of the GPI gene in various types of cancers was identified in the Oncomine database (https://www.oncomine.org/resource/login.html) (16). Student’s t-test was used to compare the transcription levels of the GPI gene in cancer samples with those in normal controls. The cutoff p-value and fold change were defined as 0.05 and 1.5, respectively.

The Kaplan–Meier plotter (http://kmplot.com/analysis/) (17) and GEPIA were used to estimate the correlation between GPI expression and the survival rate of different clinical features in LUAD patients, and the hazard ratio (HR) and log-rank p-value of the 95% confidence interval were calculated.

Based on the Kyoto Encyclopedia of Genes and Genomes (KEGG) interaction (18), we constructed a pathway and gene regulatory network. Protein–protein interaction (PPI) data were extracted from the STRING database based on protein interactions and signaling pathways (https://string-db.org), and the network was constructed using Cytoscape 3.7.2 applications.

In this study, we selected all independent clinicopathological prognostic factors from Cox regression analysis and constructed a contingency table to evaluate the 1-, 3-, and 5-year overall survival (OS) probability of LUAD patients. By comparing the predicted probability of the line chart with the observed actual probability through a calibration curve, the accuracy of the line chart was verified. The overlapping reference lines show that the model is accurate. Receiver operating characteristic (ROC) analysis was used to compare the prediction accuracy of the line chart of the combined model and the line chart of various clinicopathological prognostic factors.

To understand the biological processes and pathways that GPI may participate in, we conducted the following analysis. The first 600 genes related to GPI in LUAD and 56 genes related to LUAD survival were obtained from GEPIA (http://gepia.cancer-pku.cn/) (19). These genes were enriched by Gene Ontology (GO) [including biological processes (BP), cellular components (CC), and molecular function (MF)] and KEGG pathway analyses using the Database for Annotation, Visualization and Integrated Discover (DAVID) and visualized by the R package ggplot2. A corrected p < 0.05 was determined to be statistically significant.

The LinkedOmics database (http://www.linkedomics.org) (20) contains multiomics data and clinical data for 32 cancer types and a total of 11,158 patients from TCGA project (20). In this study, the LinkedOmics database was used to explore the expression profile of GPI. We explored the GO biological process (GO_BP) and KEGG pathways of GPI and its co-expressed genes by using gene set enrichment analysis (GSEA) in the Link Interpreter module.

TIMER (http://www.cistrome.shinyapps.io) (21) is a convenient and accurate online analysis tool that can explore the expression levels of genes in normal and tumor tissues in multiple tumors in TCGA datasets, infer the abundance of tumor-infiltrating immune cells from the gene expression profiles, and evaluate their clinical impact. In this study, the TIMER database was used to analyze the expression levels of genes in normal and tumor tissues in multiple tumors. The deconvolution statistical method in TIMER was used to infer the abundance of tumor-infiltrating immune cells (TIICs) from the gene expression profiles of LUAD samples in TCGA datasets.

TISIDB is an integrated knowledge base portal that plays a vital role in detecting the interaction between tumors and the immune system (http://cis.hku.hk/TISIDB/) (22). To further clarify the immune correlation of GPI in cancer, we used the “Immunomodulator” module of the TISIDB database to analyze and evaluate the correlation between GPI expression and the levels of immune checkpoint genes. To further study the association between GPI and chemokine/chemokine receptor expression, we evaluated the expression levels of chemokine/chemokine receptors of TIICs through the “chemokine” module.

The lung adenocarcinoma cell lines H1975 and A549 were purchased from ATCC and cultured in Dulbecco’s modified Eagle’s medium (DMEM) (Gibco, Waltham, MA, USA) containing 10% fetal bovine serum (FBS) (Biological Industries, Israel) in a CO₂ incubator at 37°C. The cells were grown to ~60% confluency in six-well plates (NEST Biotechnology Co., Ltd., Wuxi, China) 1 day before transfection. We used Lipofectamine 3000 (Invitrogen, San Diego, CA, USA) to transfect the mall interfering RNA (siRNA) sequences according to the manufacturer’s instructions (23). The cells were cultured with serum-free medium supplemented with 10% FBS 6 h post-transfection and then incubated for 24–72 h. The siRNA sequences of GPI in H1975 and A549 cells were used as follows: siRNA negative control (NC) (5′-TTCTCCGAACGTCACGT-3′), siRNA-GPI #1 (5′-CCAAGATGATACCCTGTGA-3′), siRNA-GPI #2 (5′-GTGCTCA AGTGACCTCTCA-3′), and siRNA-GPI #3 (5′-GGCATATCCTGGTGGATTA-3′).

H1975 and A549 cells transfected with GPI siRNA were placed in a 96-well plate at 3000/well and allowed to adhere overnight. Ten microliters of Cell Counting Kit-8 (CCK-8 solution) was added to each well and incubated for 2 h. A microplate reader (Thermo Fisher Scientific, Shanghai, China) was used to measure the absorbance at 450 nm every day, and the experiment was repeated three times independently (24).

Lung cancer cells were lysed with protease inhibitor, phosphatase inhibitor, and RIPA cleavage buffer (Thermo Fisher Scientific, Shanghai, China). The protein lysates were separated by sodium dodecyl sulfate polyacrylamide gel electrophoresis (SDS-PAGE) (Servicebio, Wuhan, China) and imprinted on polyvinylidene fluoride (PVDF) membranes (Millipore, Billerica, MA, USA) for analysis (25). Anti-GPI (1:1,000 dilution, #57893; Cell Signaling Technology, Danvers, MA, USA) and anti-GAPDH (1:5,000 dilution, 60004-1-Ig; Proteintech, Wuhan, China) were incubated overnight at 4°C. Horseradish peroxidase (HRP)-conjugated secondary antibodies (1:2,000 dilution; A0181, Beyotime, Shanghai, China) were added for 2 h at room temperature. Western blot analysis was performed with ImageJ software.

A549 and H1975 cells were cultured in six-well serum-free culture with 1 × 10⁵ cells for 12 h and were transfected with GPI siRNA. The cells were collected after 48 h, washed twice with phosphate-buffered saline (PBS), and then fixed with 70% pre-cooled ethanol at 4°C for 12 h. The cells were stained with RNase A-containing PI buffer (KGA511; Keygen, Jiangsu, China) for 30 min. Flow cytometry (Beckman Coulter Quanta SC System) was used to detect the cell cycle distribution.

This study was approved by the Institutional Research Ethics Committee of Taihe Hospital. Forty-six paraffin-embedded lung adenocarcinoma tissues and paracarcinoma tissues were used for immunohistochemistry (IHC) staining. Each tissue was incubated overnight with a primary antibody for GPI (1:1,000 dilution, #94068; Cell Signaling Technology, Danvers, MA, USA) at 4°C. After washing with PBS, each section was incubated with an HRP-labeled goat anti-rabbit secondary antibody (1:2,000 dilution, A0181; Beyotime, Shanghai, China) for 2 h at room temperature. IHC staining was achieved using 3,3′-diaminobenzidine and counterstained with hematoxylin (Beyotime, Shanghai, China). The IHC staining results were analyzed and scored by two pathologists who were blinded to the sources of the clinical samples. ImageJ software was used to analyze the intensity of staining using a semiquantitative integration method.

Box plot and scatter plot methods were used to detect the expression level of the GPI gene in LUAD patients. The cutoff value of GPI expression was selected as the median method of gene expression. The Wilcoxon signed-rank test and logistic regression were used to analyze the relationship between the clinical features of LUAD and the expression of GPI. Log-rank test was used to test the p-value. Single-factor and multivariate Cox analyses were used to screen potential prognostic factors. In all analyses, *, **, and *** indicate p < 0.05, p < 0.01 and p < 0.001, respectively.

We developed a flow diagram to show our process ( Figure 1 ). To explore the expression level of GPI in normal and tumor tissues, we used the Oncomine database to analyze the expression levels of GPI mRNA in different tumors and normal tissues. The results showed that, compared with that in normal tissues, GPI expression was significantly higher in tumor tissues such as bladder cancer, colorectal cancer, gastric cancer, kidney cancer, lung cancer, ovarian cancer, and lymphoma. In addition, a lower expression was observed in brain tumors, central nervous system (CNS) tumors, leukemia, and sarcoma ( Figure 2A ). Similar results were also obtained from the TIMER database of TCGA clinical patients ( Figure 2B ). In addition, to understand the mutation level of GPI in LUAD, we analyzed its genome and copy number. The OncoPrint map of the GPI gene of LUAD patients in TCGA dataset was analyzed using a cBioPortal map ( Figure 2C ), and the results showed that GPI had less than 6% gene amplification and missense mutations.

We used TCGA database to evaluate the mRNA expression levels of GPI in LUAD patients and compared these with the expression levels in adjacent tissues. The results showed that the expression level of GPI in LUAD was significantly higher than that in paracarcinoma tissues (p < 0.01) ( Figure 2D ). The results were verified in LUAD and paired paracarcinoma tissues ( Figure 2E ).

To analyze the correlation between GPI expression and clinical characteristics in LUAD patients, we analyzed the mRNA expression levels of GPI in different clinical categories in TCGA database. The association identified between GPI expression and clinical features in patients with LUAD is summarized in Table 1 . The results showed that the high expression of GPI was significantly related to the clinical stage (p < 0.05), N stage (p < 0.05), initial treatment result (p < 0.05), and OS event (p < 0.01) of these patients ( Figures 2F–J and Table 1 ). To further elucidate the abnormal upregulated mechanisms of GPI in LUAD tissues, we also explored the correlation of the expression levels of GPI and their methylation status using public databases (DiseaseMeth, version 2.0). The data showed that the difference in the methylation levels between tumor tissues and paracarcinoma tissues was not significant ( Supplementary Table S1 ). These data suggest that GPI is significantly upregulated in LUAD.

To further verify the expression level of GPI in LUAD, we selected another four independent external GEO datasets (validation cohort) to analyze the GPI transcription levels of cancer tissues and adjacent tissues in LUAD, including GSE10072, GSE32863, GSE7503, and GSE30219. The results showed that the transcription level of GPI in LUAD was significantly higher than that in noncancerous adjacent tissues from four LUAD datasets (p < 0.001) ( Figures 3A–D ). The results of this comparison were also verified in LUAD tissues and accurately matched noncancerous tissues (GSE30219 does not contain accurately matched noncancerous tissues) and lung cancer cell lines ( Figures 3E–H ).

At the same time, we analyzed the protein expression level of GPI in LUAD based on the HPA database. The data showed that the expression level of GPI in tumors was higher than that in adjacent tissues ( Figure 3I ). In addition, the expression level of GPI in NSCLC cell lines was detected by Western blotting. The results showed that the expression level of GPI in A549, H1975, H1299, SK-MES-1, and H460 cells was significantly higher than that in the normal lung epithelial cell line BEAS-2B ( Figure 3J ). Finally, clinical specimens of LUAD were collected to investigate GPI expression, and the results indicated that the expression level of GPI in tumors was higher than that in paracarcinoma ( Figures 4A, B ). These data verify that GPI is highly expressed in LUAD.

To identify whether GPI expression affects patient survival, we classified LUAD patients in TCGA database into a high GPI expression group (the top 50% of samples with the highest expression) and a low GPI expression group (the remaining 50% of the samples) in order to perform survival analysis according to the mean expression value of GPI. The Kaplan–Meier survival analysis showed that the high expression of GPI was related to the poor prognosis of LUAD patients (HR = 1.44, p = 0.014) ( Figure 5A ). Subgroup analysis showed that a high GPI expression was significantly correlated with poor prognosis in LUAD in the following cases: patients over 65 years old (HR = 1.94, p = 0.002), female patients (HR = 1.60, p = 0.021), T1 (HR = 2.05, p = 0.25), T3 (HR = 4.22, p = 0.005), pathological stage II (HR = 1.90, p = 0.024), CR/PR (complete/partial response; HR = 1.59, p = 0.035), smokers (HR = 1.48, p = 0.018), and smokers over 40 years old (HR = 2.10, p = 0.006). These data are shown in Figures 5B–I .

In addition, we used three other independent external GEO datasets—GSE31210, GSE37745, and GSE3141—as the validation cohort to verify the reproducibility and transferability of GPI expression data in the prognosis of LUAD patients. The Kaplan–Meier survival analysis showed that patients with increased GPI expression in three independent external datasets had shorter OS ( Figures 5J–L ). Univariate Cox analysis demonstrated that a high GPI expression was significantly correlated with poor OS (HR = 1.435, 95% CI = 1.075–1.916, p = 0.014) ( Table 2 ). These data suggest that a high expression of GPI is an independent prognostic factor for the OS of LUAD patients.

To analyze the diagnostic value of GPI expression in LUAD, we performed ROC curve and nomogram analysis on the GPI gene expression data of TCGA database to evaluate the diagnostic value of the gene. The area under the ROC curve (AUC) was 0.940, suggesting a higher diagnostic value, as shown in Figure 6A . We combined the expression level of GPI with the clinical variables to construct a nomogram in order to predict the survival probability of patients at 1, 3, and 5 years. The nomogram indicated that the prognostic prediction of the expression level of GPI was better than that of the traditional clinical features of age and sex ( Figure 6B ).

To understand the biological function of GPI in LUAD, we used the LinkFinder module of the LnkedOmics website to detect the co-expression pattern of GPI in LUAD of TCGA database. The red dot indicates the top 25 genes that were positively correlated with GPI, and the green dot represents the bottom 25 genes that were negatively correlated with GPI ( Figure 7A ).

We used DAVID Functional Annotation Bioinformatics Microarray Analysis to identify the enriched GO functional enrichment and KEGG pathways among the GPI-related genes (top 600) and found that cell cycle signaling pathways, RNA replication, and mismatch repair were enriched among these genes ( Figure 7B ).

To identify genes with the same regulatory direction in high-GPI and non-survival patients, we crossed the 600 genes with the highest GPI correlation with the 148 survival-related upregulated genes in LUAD and detected 56 genes at the intersection that were related to GPI and LUAD survival ( Figure 7C ). These 56 protein-coding genes may be potential genetic biomarkers for LUAD patients. GO functional enrichment and KEGG pathway analysis were performed for these 56 genes, and the results showed that differentially expressed genes (DEGs) were significantly enriched in the cell cycle from the GO terms and KEGG pathways ( Figure 7D ).

After discovering significantly different pathways, PPIs and correlated analysis were used to identify the interactions between these 56 proteins. We found that there was a stronger enrichment network among these proteins than in random proteins, and these genes were particularly related to cell cycle pathways ( Figure 8A ). Gene co-expression correlation analysis showed that most of the proteins in the network had a strong positive correlation with each other ( Figure 8B ). Therefore, these GPI-associated established genes, especially cell cycle-related genes, have a strong intertwined interaction and can be used as multigene biomarkers to predict the survival of LUAD patients.

To explore the regulatory effect of GPI on the cell cycle, GPI-targeted siRNA sequences were used to interfere with GPI expression, and the knockout efficiency was detected by Western blotting. The results showed that the expression of GPI was significantly decreased after transfection of GPI siRNA ( Figure 9A ). CCK-8 assays were applied to investigate the effect of GPI on cell proliferation. The results showed that silencing GPI significantly inhibited the proliferation of A549 and H1975 cells ( Figures 9B, C ). The effect of silencing GPI on the cell cycle of LUAD cells was further determined by flow cytometry. The results showed that A549 and H1975 cells revealed G2/M cell cycle arrest after GPI knockdown ( Figure 9D ). These data indicated that GPI affects the proliferation of LUAD cells by regulating cell cycle progression.

The above bioinformatics analysis indicated that GPI is significantly enriched in the cell cycle pathway. Therefore, we further analyzed the correlation between GPI and the cell cycle regulation genes in TCGA-LUAD, and we found that the expressions of the cell cycle regulatory genes CCNA2, CCNB1, CCNB2, CCNE1, CHEK1, BUB1B, ESPL1, PTTG1, PCNA, PKMYT1, CDC45, PLK1, MCM2, MCM4, MCM6, E2F1, CDC6, CDC20, CDC25A, and CDC25C were positively correlated with GPI (r > 0.3, p < 0.001) ( Figures 10A–T ). These data verify that GPI is closely related to LUAD cell cycle regulation genes.

Research has shown that the levels of GPI are elevated in autoimmune diseases (26). In this study, we found that GPI is abnormally highly expressed in LUAD, so we speculate that GPI may be involved in regulating the tumor immune response. To explore the correlation between the expression level of GPI and tumor immune response, we used the TIMER database to investigate immune infiltration in LUAD with different GPI expression levels. The results showed that the infiltration levels of CD8⁺ T cells, dendritic cells (DCs), eosinophils, interdigitating cells, macrophages, mast cells, NK CD56 dim cells, central memory T cells, effective memory T cells, and follicular helper T cells in LUAD patients with high GPI expression were significantly lower than those in patients with low GPI expression. In contrast, the infiltration of Th2 cells, regulatory T cells (Tregs), and gamma delta T cells in LUAD patients with high GPI expression was significantly higher than that in patients with low GPI expression. In addition, there was no significant difference in the infiltration levels of Th1 cells, Th17 cells, B cells, natural killer (NK) cells, and neutrophils between patients with high and low GPI expression ( Figures 11A, B ).

We further analyzed the correlation between the expression level of GPI and immune infiltration in LUAD, and the results showed that the expression level of GPI was positively correlated with the infiltrating levels of Th2 cells (r = 0.47, p < 0.001) ( Figure 11C ), Tregs (r = 0.13, p < 0.002) ( Figure 11D ), gamma delta T cells (r = 0.110, p =0.008) ( Figure 11E ) and NK CD56bright cells (r = 0.230, p <0.001) ( Figure 11F ). In contrast, the expression level of GPI was negatively correlated with the infiltrating levels of DC (r = -0.200, p < 0.001) ( Figure 11G ), immature DC (r = -0.200, p < 0.001) ( Figure 11H ), Mast cells (r = -0.260, p < 0.001) ( Figure 11I ), central memory T cells (r = -0.240, p < 0.001) ( Figure 11J ), (K) Macrophages (r = -0.140, p = 0.001) ( Figure 11K ), Eosinophils (r = -0.190, p < 0.001) ( Figure 11L ), effective memory T cell (r = -0.150, p < 0.001) ( Figure 11M ), CD8+ T cells (r = -0.0960, p = 0.026) ( Figure 11N ). A Th1/Th2 balance drift exists in many tumors, such as lung cancer, cervical cancer, breast cancer, gastric cancer, and other cancers (27). Th2 cells are often dominant, which may be related to tumor immune escape (28). Tregs also play an important role in tumor immune escape (29). Therefore, these data suggest that GPI may inhibit the tumor immune response of LUAD by positively regulating Th2 and Treg infiltration into tumors and negatively regulating CD8⁺ T-cell, central memory T-cell, DC, macrophage, mast cell, and eosinophil infiltration into tumors.

Chemokines and chemokine receptors are essential for the infiltration of immune cells into tumors (30). Therefore, we used TISIDB to analyze the correlation between the expression level of GPI and immune cell chemokines and chemokine receptors in LUAD. The heatmap results showed that several chemokines and chemokine receptors were significantly correlated with the expression of GPI in LUAD ( Figures 12A, B ). These results demonstrate that the GPI gene may play an important role in tumor immunity. To further clarify the relationship between GPI expression and immune cell migration, we comprehensively analyzed the correlation between GPI expression and chemokines/chemokine receptors. The results showed that GPI expression was negatively correlated with CCL14 (r = −0.373, p < 2.2e−16), CCR4 (r = −0.321, p < 9.17e−14), CCR6 (r = −0.369, p < 2.2e−16), and CX3CR1 (r = −0.376, p < 2.2e−16) ( Figures 12C–F ), whereas GPI was not highly correlated with other chemokines/chemokine receptors (−0.3 < r < 0.3). These data indicated that GPI is negatively correlated with the expressions of chemokines/chemokine receptors in LUAD.

Immune checkpoint inhibitors (ICIs) are a significantly novel strategy for tumor immunotherapy that has gradually improved the prognosis of patients with many types of cancers (31). Subsequently, we analyzed the correlation between GPI and the expressions of immunoinhibitors and immunostimulators in different types of human cancers using the TISIDB database ( Figures 13A, B ). Interestingly, there was no significant correlation between GPI and immunoinhibitor expression. However, GPI was negatively correlated with the expression of several immunostimulators, such as CD40L (r = −0.366, p < 6.85e−19), IL6R (r = −0.365, p < 2.02e−18), and TMEM173 (r = −0.303, p < 2.37e−12), in LUAD ( Figures 13C–E ). Therefore, these results suggest that GPI may play a role in regulating tumor immunity.