Identification of a novel glycolysis-related gene signature that can predict the survival of patients with lung adenocarcinoma

ABSTRACT

Lung cancer is one of the most malignant cancers worldwide, and lung adenocarcinoma (LUAD) is the most common histologic subtype. Thousands of biomarkers related to the survival and prognosis of patients with this cancer type have been investigated through database mining; however, the prediction effect of a single gene biomarker is not satisfactorily specific or sensitive. Thus, the present study aimed to develop a novel gene signature of prognostic values for patients with LUAD. Using a data-mining method, we performed expression profiling of 1145 mRNAs in large cohorts with LUAD (n = 511) from The Cancer Genome Atlas database. Using the Gene Set Enrichment Analysis, we selected 198 genes related to GLYCOLYSIS, which is the most important enrichment gene set. Moreover, these genes were identified using Cox proportional regression modeling. We established a risk score staging system to predict the outcome of patients with LUAD and subsequently identified four genes (AGRN, AKR1A1, DDIT4, and HMMR) that were closely related to the prognosis of patients with LUAD. The identified genes allowed us to classify patients into the high-risk group (with poor outcome) and low-risk group (with better outcome). Compared with other clinical factors, the risk score has a better performance in predicting the outcome of patients with LUAD, particularly in the early stage of LUAD. In conclusion, we developed a four-gene signature related to glycolysis by utilizing the Cox regression model and a risk staging model for LUAD, which might prove valuable for the clinical management of patients with LUAD.

Introduction

Non-small cell lung cancer (NSCLC) accounts for 80%–85% of lung cancer cases and is the primary cause of cancer-related deaths [1,2]. Owing to the lack of specific clinical symptoms, lung adenocarcinoma (LUAD), which is the major histological subtype of NSCLC, is usually diagnosed at an advanced stage, resulting in the death of most patients from local recurrence or distant metastasis [3,4]. Despite the clinical applications of combination of surgical resection, chemotherapy, radiation therapy, and targeted therapy, the 5-year overall survival rate of LUAD remains at a low level of 15.9% [2,3]. In clinical practice, histopathology can often successfully predict the prognosis of patients with lung cancer. But it has limitations because patients with the same pathology will have distinct prognoses due to individual differences. In combination with the existing prognostic methods, novel molecular biomarkers might improve the prognosis and the appropriate treatment of LUAD. Therefore, additional molecular biomarkers must be urgently identified.

Currently, several biomarkers are used to predict the survival of patients with NSCLC. The downregulation of gasdermin D (GSDMD) attenuates tumor proliferation and predicts good prognosis [5], and overexpression of Beclin-1 in patients with lung cancer indicates a low survival rate. Moreover, the level of circulating caspase-4 is related to overall survival in patients with NSCLC [6,7]. With the development of high-throughput sequencing, several databases of the genomes of patients have been established, allowing us to better understand genomic changes. More and more biomarkers related to survival and prognosis have been investigated via database mining [8,9]. However, a single gene cannot accurately predict LUAD outcomes. By contrast, the combination of biomarkers might enhance the sensitivity and specificity of LUAD prognosis. Multigene prognostic signatures derived from primary tumor biopsies can guide clinicians in selecting appropriate treatment, provide insights about cancer progression, and uncover potential new therapeutic targets. Thus, an expression-based gene signature for predicting the survival of LUAD patients must be established.

In the present study, the Gene Set Enrichment Analysis (GSEA) [10] was used to further analyze certain genes. In previous studies with high-throughput data, these technologies identify genes differentially expressed across the phenotypes of interest. GSEA focuses on coordinated differential expression of annotated groups of genes, or gene sets, and produces results that can be easily interpreted in terms of the relevant biological processes. However, the follow-up analysis and interpretation of the results is challenging. By combined analysis of the localization, nature, and function of existing genes, the Molecular Signatures Database (MSigDB) was established according to GSEA on the basis of information, such as biological significance.

The current study aimed at the discovery of new prognostic markers in patients with LUAD using GSEA and Cox multivariate regression models to analyze whole genome expression profile data sets. We profiled the hallmark gene sets in 511 cases of LUAD with the whole mRNA expression dataset from The Cancer Genome Atlas (TCGA) database. We identified 198 mRNAs significantly associated with glycolysis and established a four-gene risk signature that can effectively predict patient outcome. Surprisingly, the glycolysis-related risk signature could independently identify patients in the high-risk group with poor prognoses. In addition, according to Cox multivariate hazard analyses, the risk score had a better performance than other clinical information in predicting the prognosis of patients with LAUD.

Materials and methods

Clinical information of patients and mrna expression dataset

The mRNA expression profile and clinical dataset of patients with LUAD were downloaded from the TCGA dataset (https://cancergenome.nih.gov/). Clinical information, including the total number of patients (n = 511), sex, age, TNM stage, T stage, N stage, M stage, personal neoplasm cancer status, and presence of residual tumor, was included in the study. The general clinical features are listed in Table 1.

Table 1.

clinical features of LUAD patients (n = 511) from TCGA database.

Variables	Patients, n (%)
Sex
Male	236 (46%)
Female	275 (54%)
Age, years
≤65	238 (48%)
>65	254 (52%)
TNM stage
I	273 (54%)
II	120 (24%)
III	84 (16%)
IV	26 (6%)
T stage
T1	168 (33%)
T2	274 (54%)
T3	47 (9%)
T4	19 (4%)
N stage
N0	329 (64%)
N1	94 (18%)
N2	74 (14%)
N3	2 (0.4%)
NX	12 (3.6%)
M stage
M0	342 (67%)
M1	165 (33%)
Person neoplasm cancer status
Tumor-free	304 (73%)
With tumor	110 (27%)
residual_tumor
No	342 (89%)
Yes	42 (11%)

Gene set enrichment analysis

GSEA (http://www.broadinstitute.org/gsea/index.jsp) was used to explore whether the identified sets of genes showed statistically significant differences between the two groups [10]. The expression of 1145 mRNAs, including 108 mRNAs in adjacent noncancerous tissue and 1037 mRNAs in LUAD samples that were obtained from the TCGA dataset, was analyzed. Normalized P value (P < 0.05) and normalized enrichment score (NES) were used to determine which function had to be investigated for further analysis.

Statistical analysis

The expression profiles of 1145 mRNAs were presented as raw data, and each mRNA was normalized by log2 transformation for further analysis. Univariate Cox regression analysis was carried out to identify the genes evidently related to overall survival (OS) with P values < 0.05.

The candidate genes were fitted in a stepwise multivariate Cox proportional regression model to identify the predictive model with the best explanatory and informative efficacy. A risk score staging model was established using the R package survival function coxph (). The risk score formula is described as follows:

$$\begin{array}{rl}& Riskscore=expressionofgene1\times \beta 1\\ & +expressionofgene2\times \beta 2\\ & +\dots +expressionofgenen\times \beta n.\end{array}$$

We classified 511 patients with LUAD into two subgroups (high-risk and low-risk groups) using the median risk score as the cutoff. Kaplan–Meier curves and the log-rank method were used to validate the prognostic significance of the risk score. Student’s t-test was utilized to show the different expression of optimal genes in adjacent normal and LUAD tissues. The alteration in selected genes is shown online (http://www.cbioportal.org/). All statistical analyses were performed with the Statistical Package for the Social Sciences software version 16.0 (SPSS Inc., Chicago, IL, the USA) and GraphPad Prism 7 (GraphPad Software, La Jolla, CA, the USA; www.graphpad.com).

Results

Initial screening of the genes using GSEA

The clinical information of 511 patients and 1145 mRNA expression dataset for LUAD were obtained from TCGA. The hallmark gene sets containing 50 specific gene sets are coherently expressed signatures derived by aggregating MSigDB version 6.2 gene sets to represent well-defined biological states or processes. GSEA was used for the abovementioned data to explore whether the identified sets of genes showed statistical differences between the lung adenocarcinoma and adjacent normal tissues. We found that five gene sets, including GLYCOLYSIS, E2F_TARGETS, G2M_CHECKPOINT,MYC_TARGETS_V2, and MYC_TARGETS_V1, were significantly enriched at normalized P value < 1% (Table 2, Figure 1). We then selected GLYCOLYSIS (P = 0.000, NES = −2.0285), which has the top-ranking function and contained 198 genes, for further analysis.

Table 2.

Gene sets enriched in LUAD (511 samples).

GS follow link to MSigDB	SIZE	NES	NOM p-val	FDR q-val
HALLMARK_GLYCOLYSIS	198	−2.0285	0.0000	0.0024
HALLMARK_E2F_TARGETS	197	−2.0658	0.0021	0.0064
HALLMARK_G2M_CHECKPOINT	196	−2.0549	0.0083	0.0037
HALLMARK_MYC_TARGETS_V2	58	−2.0509	0.0040	0.0028
HALLMARK_MYC_TARGETS_V1	197	−1.9677	0.0042	0.0043

Figure 1.

GSEA results for the enrichment plots of five gene sets (GLYCOLYSIS, E2F_TARGETS, G2M_CHECKPOINT, MYC_TARGETS_V2, and MYC_TARGETS_V1) that were significantly differentiated in normal and LUAD tissues based on TCGA. GSEA, Gene Set Enrichment Analysis; LUAD, lung adenocarcinoma; TCGA, The Cancer Genome Atlas.

Identification of glycolysis-related genes associated with survival in patients with progressive LUAD

To identify novel genetic biomarkers associated with the outcome of patients with LUAD, univariate Cox proportional hazard regression was applied to 198 genes that were enriched via glycolysis. A total of 12 genes were significantly correlated to OS (P < 0.05) and were entered into a stepwise multivariate Cox regression analysis. Finally, four independent genes (AGRN, AKR1A1, DDIT4, and HMMR) (Table 3) were selected via multivariable Cox regression analysis, and a gene-based prognostic model was established to evaluate the survival risk of each patient as follows:

Table 3.

The detailed information of four prognostic mRNAs significantly associated with overall survival in patients with LUAD.

mRNA	Ensemble ID	Location	β (Cox)	HR (95%CIs)	P
AGRN	ENSG00000188157	Chr1: 1,020,123–1,056,118	0.2486	1.2823 (0.0991,2.51)	0.01210
AKR1A1	ENSG00000117448	Chr1: 45,550,543–45,570,049	−0.6790	0.5071 (0.1394, −4.87)	1.1e-06
DDIT4	ENSG00000168209	Chr10: 72,273,920–72,276,036	0.1966	1.2173 (0.0800,2.46)	0.01397
HMMR	ENSG00000072571	Chr5: 163,460,203–163,491,945	0.3151	1.3704 (0.0892,3.53)	0.00041

$$\begin{array}{rl}& Riskscore=0.2486\times expressionofAGRN\\ & +(-0.6790)\times expressionofAKR1A1\\ & +0.1966\times expressionofDDIT4\\ & +0.3151\times expressionofHMMR\end{array}$$

Then, we analyzed the alteration of the four selected genes in 511 clinical LUAD samples in the cBioPortal database. Results showed that the genes were altered in 48 (9.5%) sequenced cases. The alterations in four genes are shown in Figure 2(a). The different expressions of the four genes in adjacent normal and LUAD tissues were also investigated. The four genes were all significantly upregulated in the tumor tissues (P < 0.0001, Figure 2(b)).

Figure 2.

Identification of mRNAs associated with patient survival. (a) The alteration proportion for the four selected genes in 511 clinical samples of lung adenocarcinoma in the cBioPortal database. (b) Different expression of four genes in the normal tissues (n = 108) and tumor tissues (n = 1037) based on The Cancer Genome Atlas. (***represent P < 0.0001).

Association between risk scores and patient outcome

According to the risk score formula, the patients with LUAD were classified under the high-risk group (n = 255) and low-risk group (n = 256) according to the median risk score used as a cutoff value (Figure 3(a)). The survival (days) of each patient is shown in Figure 3(b). The patients with a high-risk score had a higher mortality rate than those with a low-risk score. A heatmap (Figure 3(c)) is shown to present the expression profile of the four mRNAs. As the risk score of patients with LUAD increased, the expression of the risky-type mRNAs (AGRN, DDIT4, and HMMR) was distinctly upregulated; in contrast, the expression of the protective-type mRNA (AKR1A1) was downregulated.

Figure 3.

The four-mRNA signature related to risk score predicts the overall survival of patients with lung adenocarcinoma. (a) mRNA risk score distribution. (b) Survival status. (c) Heatmap of four genes expression profile in The Cancer Genome Atlas.

Risk score generated from the signature of four mRNAs as an independent prognostic indicator

To compare the risk score and conventional clinical factors, we performed univariate and multivariate Cox hazard analyses to evaluate the importance of these indicators in the patient cohort, which included risk score, age, sex, TNM stage, T stage, M stage, N stage, personal neoplasm cancer status, and presence of residual tumor as covariables. We found that risk score (hazard ratio [HR]: 2.826; 95% confidence interval [CI]: 1.946–4.102; P < 0.001), TNM stage (HR: 1.642; 95% CI: 1.394–1.943; P < 0.001), T stage (HR: 1.517; 95% CI: 1.225–1.879; P < 0.001), N stage (HR: 2.537; 95% CI: 1.777–3.621; P < 0.001), personal neoplasm cancer status (HR: 6.456; 95% CI: 3.906–10.670; P < 0.001), and presence of residual tumor (HR: 2.755; 95% CI: 1.555–4.882; P = 0.001) were associated with patient survival in the univariate analysis. In addition, risk score, personal neoplasm cancer status, and N stage had a remarkable independent prognostic value not only in the univariate analysis but also in the multivariate analysis with P value < 0.0001 or < 0.05, which indicates that the prognostic value of the four-gene signature is competitive for survival prediction. These results indicated that the risk score was robust in predicting the prognosis of patients with LUAD (Table 4).

Table 4.

Univariable and multivariable analyses for each clinical feature.

		Univariate analysis			Multivariate analysis
Clinical feature	Number	HR	95%CI of HR	P value	HR	95%CI of HR	P value
Risk score (Low-risk/High-risk)	256/255	2.826	1.946–4.102	< 0.001	2.648	1.469–4.772	< 0.0001
Sex (Male/Female)	236/275	1.059	0.744–1.507	0.751	1.057	0.610–1.834	0.843
Age (≤65/>65)	238/254	2.001	0.990–4.044	0.053	1.133	0.639–2.008	0.670
TNM stage (I/II/III/IV)	273/120/84/26	1.642	1.394–1.943	< 0.001	0.836	0.528–1.324	0.446
T stage (T1/T2/T3/T4)	168/274/47/19	1.517	1.225–1.879	< 0.001	1.436	0.929–2.220	0.104
N stage (N0/N)	329/181	2.537	1.777–3.621	< 0.001	2.358	1.079–5.154	< 0.05
M stage (M0/M1)	342/165	1.169	0.781–1.749	0.448	0.922	0.415–2.049	0.842
Person neoplasm cancer status (Tumor-free/With tumor)	304/110	6.456	3.906–10.670	< 0.001	5.351	2.841–10.078	< 0.0001
Residual tumor (No/Yes)	342/42	2.755	1.555–4.882	0.001	1.223	0.532–2.814	0.635

Validation of the four-mRNA signature in predicting survival using Kaplan–Meier curves

The Kaplan–Meier curves revealed that clinical features, TNM stage (III and IV) (P < 0.0001), T stage (T3 and T4) (P < 0.001), N stage (N) (P < 0.001), personal neoplasm cancer status (with tumor) (P < 0.001), presence of residual tumor (P < 0.0001), and high risk score (P < 0.0001) were significantly associated with poor OS (Figure 4(a, b)).

Figure 4.

Kaplan–Meier survival analysis for patients with lung adenocarcinoma in The Cancer Genome Atlas dataset. (a) Kaplan–Meier curve for patients divided into the high-risk and low-risk groups. (b) Different clinical features that can predict patient survival (TNM stage, T stage, N stage, personal neoplasm cancer status, and presence of residual tumor).

The Kaplan–Meier curves showed that the risk score was a stable prognostic marker for patients with LUAD stratified by sex (male or female), N stage (N0 or N), M stage (M0 or M), and personal neoplasm cancer status (tumor free or with tumor) (Figure 5(a), (d), (e), and (f)). However, when the patients were stratified into different subgroups based on TNM stage, T stage, and residual tumor, the role of risk score differed. The patients in the high-risk group had significantly shorter OS than those in the low-risk group in the stage I and II subgroups (P < 0.01), and no statistical difference was observed between the stage III and IV subgroups (P = 0.1492; P = 0.1617). The prognostic power of the risk score in individuals with T1 stage disease (P < 0.001) and T2 stage disease (P < 0.0001) was better than that in individuals with T3 stage disease (P = 0.1744) and T4 stage group (P = 0.1333). Similar results were observed for residual tumor (residual tumor: no, P < 0.0001; residual tumor: yes, P = 0.5072). These findings indicate that risk score may be a more effective prognostic marker in the early stages of the disease (Figure 5(b), (c), and (g)).

Figure 5.

Kaplan–Meier curves for the prognostic value of risk score signature for the patients grouped according to each clinical feature.

(a)Kaplan–Meier survival curves of the male patient group (n = 236) and female patient group (n = 275). (b)Kaplan–Meier survival curves of the stage Ⅰ patient group (n = 273) and stage Ⅱ patient group (n = 120). (c)Kaplan–Meier survival curves of the T1 stage patient group (n = 168), T2 stage patient group (n = 274), T3 stage patient group (n = 47), and T4 stage patient group (n = 19). (d)Kaplan–Meier survival curves of the N0 stage patient group (n = 329) and N stage patient group (n = 181). (e)Kaplan–Meier survival curves of the M0 stage patient group (n = 342) and M stage patient group (n = 165). (f)Kaplan–Meier survival curves of the without tumor patient group (n = 304) and with tumor patient group (n = 110). (g)Kaplan–Meier survival curves of the without residual tumor patient group (n = 342) and with residual tumor patient group (n = 42).

Figure 5.

(Continued).

Discussion

Among the four biomarker genes (AGRN, AKR1A1, DDIT4, and HMMR) discovered by the present study, DDIT4 was highly upregulated in response to hypoxia-inducible factor 1 (HIF-1), and it regulated the generation of cellular reactive oxygen species. These findings indicate that the gene may play a role in glycolysis process [11]. DDIT4, as an oncogene [12], is expressed in various cancer tissues, and a high DDIT4 tumor expression was associated with worse outcomes compared with a low DDIT4 tumor expression [13]. AGRN promotes acetylcholine receptor synthesis in synapses, maintains functional neuromuscular junctions, and conducts neural signals. The relationship between AGRN and cancer has rarely been reported. In liver cancer, the agrin protein encoded by AGRN can be used as an extracellular matrix sensor, and it then promotes the growth and reproduction of liver cancer cells through the agrin-LRP4-MuSK signal axis [14]. HMMR is a protein-coding gene. Diseases associated with HMMR include breast cancer and fibrosarcoma. Cell cycle, mitotic, and apoptotic pathways in synovial fibroblasts are some of the related pathways. Gene Ontology annotations related to HMMR include hyaluronic acid binding (Genecards, https://www.genecards.org). Laura E et al. have pointed out that LUAD can be categorized into three distinctive subgroups based on their bulk gene expression profiles and morphology [15]. These subtypes, which are referred to as terminal respiratory unit, proximal proliferative, and proximal inflammatory (PI) subtypes, partially overlap with other transcriptomic classifications. The expression of HMMR is enriched in the PI subtype and correlates with survival, clinical progression, and distant metastasis via extracellular matrix-mediated signaling; the members of the human aldo-keto reductase (AKR) superfamily have been involved in cancer progression. However, the final conclusion is not generally accepted. AKR1A1 contributes to the acquisition of radio resistance by the laryngeal cancer cells via the suppression of p53 activation through inhibitory interaction [16], but the relationship between ALR1A1 and lung cancer and molecular mechanism must be explored in depth at a later time.

GSEA is a type of gene set enrichment analysis, which can integrate data from different levels and sources. In the present study, we attempted to run the GSEA using 1145 mRNA expression data of 511 patients with lung adenocarcinoma and found that five functions had significant differences. According to the NES and N and P value, GLYCOLYSIS, which had the minimum P value, was selected for further analysis. We focused on a specific function to select genes with GSEA for predicting patient survival and widely explored these genes. Univariate and multivariate Cox regression analyses were performed to identify the combination of four genes with prognostic value for patients with LUAD instead of a single gene. Compared with some known prognostic biomarkers, this selected risk signature may have a targeted and more powerful prognostic role in supporting clinical outcome and acting as an effective classification tool for patients with LUAD. We collected GLYCOLYSIS-related genes using the dataset of LUAD in TCGA and compared the tumor tissue with adjacent noncancerous tissues. Kaplan–Meier curve analysis showed that a high-risk score was associated with a poor survival rate. This result suggested that detecting and calculating the risk score of patients with LUAD had an important prognostic significance,as it may enrich the existing methods of predicting patient survival and prognosis. Moreover, the level of risk score may assist clinicians in selecting optimal treatment methods.

Warburg et al. have found that tumor cells require more glucose than normal cells during energy metabolism of cells, finally converting this to lactic acid through metabolic pathways. Thus, cancer is considered a metabolic disease [17]. Even under oxygen-sufficient conditions, tumor cells prefer glycolysis than mitochondrial oxidative phosphorylation for glucose metabolism. The special kind of metabolic pathway is called aerobic glycolysis, also known as the Warburg effect, which is a major feature that differentiates tumor cells from normal cells. Researchers believe that tumor cells prefer to use the aerobic glycolysis metabolic pathway mainly for the following reasons: first, tumor cells use a non-economical glycolysis pathway as the main source of energy probably because it can produce ATP faster through the glycolytic pathway than the oxidative phosphorylation pathway [18]. Second, the final product of aerobic glycolysis is a large amount of lactic acid, which can provide a material basis for the synthesis of biological macromolecules, such as nucleic, amino, and fatty acids. In addition, a local extracellular acidic environment is beneficial for tumor growth and invasion [19]. Moreover, aerobic fermentation can activate the pentose phosphate pathway, leading to the increased production of reduced nicotinamide adenine dinucleotide phosphate and glutathione, both of which help tumor cells to overcome oxidative damage. The three-way combination is a powerful explanation for the function of the glycolysis pathway as the primary mediator of tumor cell growth [20].

To date, most studies have focused on the relationship between glycolysis and tumor oncogenesis, development, proliferation, and invasion. Several studies have examined the prediction of patient survival in relation to glycolysis. For example, TCF7L2 positively regulates aerobic glycolysis in pancreatic cancer via the EGLN2/HIF-1α axis, and patients with higher TCF7L2 expression had worse prognosis [21]. Chakraborty et al. have shown that silencing MICU1 in vitro increases oxygen consumption, decreases lactate production, and inhibits clonal growth, migration, and invasion of ovarian cancer cells, whereas silencing the protein in vivo inhibits tumor growth as well as increases cisplatin efficacy and OS [22]. However, no set of glycolysis-related genes can predict patient survival.

The present study first used the public TCGA database to identify and comprehensively analyze glycolysis-related mRNAs that are significantly associated with the prognosis of patients with LUAD. Although the four-gene signature can provide an effective model for predicting the prognosis of LUAD, the present study also had certain limitations. First, the risk score model was established based on the TCGA database and should be validated in other cohorts in future studies. Second, the biological functions of the predictive genes were annotated using computational methods, and additional studies should be performed to further reveal the mechanisms of the genes involved in the tumorigenesis of LUAD. Moreover, although the gene signature may be most effective at early stages, its prognostic role in early LUAD must be further evaluated. At present, we are actively collecting clinical specimens and data for our subsequent studies.

Conclusion

To the best of our knowledge, we first identified four glycolysis-related genes associated with survival in patients with LUAD using GSEA and Cox regression. Further analysis revealed that the four-gene signature might be an independent factor that can predict the prognosis of patients with LUAD. In combination with the existing methods used in predicting survival rate, the use of the four-gene signature might improve the success rate of customized cancer therapies. These findings may indicate a potential glycolysis-related biomarker for the prognosis of LUAD and provide insights about theoretical guidance toward further advances.

Funding Statement

This study was supported by theNational Natural Science Foundation of China [No. U1608281]; Natural Science Foundation of Liaoning Province [No.2013225021]; Program for New Century Excellent Talents in University [No. LJQ2015118].

Disclosure statement

No potential conflict of interest was reported by the authors.