key: cord-0775586-5x2vniq2
authors: Zhang, Yankai; Yan, Yichao; Ning, Ning; Shen, Zhanlong; Ye, Yingjiang
title: A signature of 24 aging‑related gene pairs predict overall survival in gastric cancer
date: 2021-04-06
journal: Biomed Eng Online
DOI: 10.1186/s12938-021-00871-x
sha: a88030a1305ba32addc71d56476fd41ef8121d42
doc_id: 775586
cord_uid: 5x2vniq2

BACKGROUND: Aging is the major risk factor for most human cancers. We aim to develop and validate a reliable aging-related gene pair signature (ARGPs) to predict the prognosis of gastric cancer (GC) patients. METHODS: The mRNA expression data and clinical information were obtained from two public databases, The Cancer Genome Atlas (TCGA) dataset, and Gene Expression Omnibus (GEO) dataset, respectively. The best prognostic signature was established using Cox regression analysis (univariate and least absolute shrinkage and selection operator). The optimal cut-off value to distinguish between high- and low-risk patients was found by time-dependent receiver operating characteristic (ROC). The prognostic ability of the ARGPS was evaluated by a log‐rank test and a Cox proportional hazards regression model. RESULTS: The 24 ARGPs were constructed for GC prognosis. Using the optimal cut-off value − 0.270, all patients were stratified into high risk and low risk. In both TCGA and GEO cohorts, the results of Kaplan–Meier analysis showed that the high-risk group has a poor prognosis (P < 0.001, P = 0.002, respectively). Then, we conducted a subgroup analysis of age, gender, grade and stage, and reached the same conclusion. After adjusting for a variety of clinical and pathological factors, the results of multivariate COX regression analysis showed that the ARGPs is still an independent prognostic factor of OS (HR, 4.919; 95% CI 3.345–7.235; P < 0.001). In comparing with previous signature, the novel signature was superior, with an area under the receiver operating characteristic curve (AUC) value of 0.845 vs. 0.684 vs. 0.695. The results of immune infiltration analysis showed that the abundance of T cells follicular helper was significantly higher in the low-risk group, while the abundance of monocytes was the opposite. Finally, we identified and incorporated independent prognostic factors and developed a superior nomogram to predict the prognosis of GC patients. CONCLUSION: Our study has developed a robust prognostic signature that can accurately predict the prognostic outcome of GC patients. SUPPLEMENTARY INFORMATION: The online version contains supplementary material available at 10.1186/s12938-021-00871-x.

Globally, gastric cancer (GC) is an important public health problem, and its morbidity and mortality are at the forefront of cancer [1, 2] . As research continues, treatment approaches for GC have been refined and optimized. However, due to the occult course of GC, most patients who are first diagnosed are already at an advanced stage, with a 5-year survival rate of only about 30% [3] . In addition, GC is a highly heterogeneous cancer, has multiple histological types and each type has a unique biological behavior [4, 5] . It is well known that the most common prognostic tool used by clinicians is the American Joint Committee on Cancer (AJCC) tumor node metastasis (TNM) stage system [6] . However, some patients with GC have the same TNM stage and finally get different clinical outcomes, which show that the AJCC stage system ignores the biological heterogeneity of the tumor. Therefore, novel and powerful prognostic signature is needed in clinical work to accurately identify high-risk patients in order to improve survival outcomes in GC patients.

Aging is the process of loss and degeneration of the body from constitutive substances and tissue structures to physiological functions [7, 8] . Previous studies have shown that aging is the risk factor for many diseases, such as heart disease [9] , neurodegenerative diseases [10] , type 1 diabetes [11] and cancer [12] . Interestingly, aging-related genes (AGs) have two different biological roles in regulating cell senescence. It can inhibit tumor growth by regulating the senescence of tumor cells, and can also promote tumor invasion and metastasis [13] [14] [15] [16] . Selecting those gene sets with potential biological significance for the development of predictive models can provide new insights for individualized treatment. Given the key role of aging in the development of GC, we selected a set of AGs to build a gene signature to distinguish different risk groups, which are of great benefit in selecting personalized treatment plans for patients.

The rapid development of high-throughput sequencing technologies and bioinformatics tools has allowed us to study cancer in greater depth. For GC, many previous studies have established gene signature to predict the prognosis of patients [17] [18] [19] . However, for various reasons (technical limitations and diversity of data types) these gene signature have not been widely used in daily clinical work. Fortunately, in order to overcome above problems and make the developed gene signatures available to clinicians as early as possible, researchers have developed a new algorithm based on the relative ranking of gene expression levels, and many studies using this algorithm have yielded reliable results [20] [21] [22] . Therefore, the main focus of this study is to construct a robust and stable individualized prognostic signature based on AGs using a new algorithm. In addition, the robustness of our signature was validated using another independent cohort. Subsequently, we also compared the performance of the new signature with previously published prognostic models for GC.

The analysis process of present study is shown in Fig. 1 . As described in the method, a total of 155 ARGPs related to prognosis was identified. Next, we applied a Lasso-penalized Cox regression to further reduce the number of ARGPs in the prognostic model. Finally, we filtered to obtain 24 ARGPs and their corresponding coefficients (Table 1) . Based on the score of each gene pair and the corresponding coefficient, the following formula was constructed to calculate the risk score:

Risk 

According to the above formula, the risk score for all patients was calculated in the TCGA dataset. The optimal cut-off points for the risk score using ROC curve analysis were − 0.270, which can categorize the patients into two groups (high risk vs. low risk) (Additional file 1: Figure S1 ). The K-M curve and log-rank test showed that patients in the high-risk group have a shorter survival time than in the low-risk group (P < 0.001, Fig. 2a ). Next, we used the same cut-off value to divide patients in the GEO dataset into two groups, and got the same results (P = 0.002, Fig. 2c ). In addition, to further evaluate the prediction accuracy of ARGPs signature, a timedependent ROC curve analysis was performed. In the TCGA cohort, the AUC values of the prognostic models predicted by 1-year, 3-year and 5-year survival rates were Aging-related genes TCGA GEO TCGA Aging-related genes GEO Aging-related genes 2b ). This demonstrated that the predictive performance of our established prognostic signature was reliable. The AUC values for OS in the GEO cohort at 1 year, 3 years and 5 years were 0.637, 0.685 and 0.690, respectively (Fig. 2d) . These results were sufficient to confirm that the prognosis model can accurately predict the prognosis of GC patients. Subsequently, we performed a stratified survival analysis to determine whether our model still had predictive value in different subgroups of clinicopathological parameters. As expected, the K-M curves illustrated that the signature was a stable prognostic marker for patients with GC stratified by age (≤ 65 or > 65), gender (female or male), grade (G1-2 or G3), and stage (I-II or III-IV) (Fig. 3 ).

In addition, univariate and multivariate Cox regression analyses were performed in the TCGA and GEO datasets to assess whether the prognostic power of ARGPs signature was independent of other clinicopathological parameters (age, sex, stage, grade). After two COX regression analyses, the risk score remained an independent prognostic indicator in the TCGA cohort (HR = 4.919, 95% CI [3.345-7.235], P < 0.001; Fig. 4a, b) . As shown in Fig. 4c , d, the independent value of the ARGPs signature risk score had been verified in the GEO dataset (HR = 1.529, 95% CI [1.083-2.158], P = 0.016). + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + ++ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + ++ ++ + Time(years) Risk + + + + + + + + + ++ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + ++++ + + + + + ++ + + + + +++ + ++ + ++ ++ + ++ ++ ++ ++ ++ + ++ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + ++ + + + + + ++ + + + + ++ + + at 5 years of OS for the ARGPs was 0.845, which was significantly higher than that of RNA-binding proteins signature (AUC = 0.684) and metabolic-related gene signature (AUC = 0.695). Collectively, the above result indicated that our model was a promising tool for predicting the prognosis of GC patients.

Additional file 2: Figure S2a shows the results of GO enrichment analysis. These genes were significantly enriched in response to oxidative stress, aging, apoptotic signaling pathway, regulation of reactive oxygen species metabolic process, negative regulation of cell proliferation, reproductive structure development. The KEGG pathway analysis revealed that these genes were highly associated with the pathways in cancer, HTLV-I infection, Kaposi sarcoma-associated herpesvirus infection, AGE-RAGE signaling pathway in diabetic complications, cellular senescence (Additional file 2: Figure S2b ).

Tumor-infiltrating immune cells have been shown to correlate with the prognosis of many cancer patients [25, 26] . We applied the CIBERSORT algorithm to estimate the relative abundance of 22 immune cells within different risk groups. The radar charts depict a comparative summary of various immune cells in these two risk groups (Fig. 6a) . We found that immune cells, such as T cells follicular helper and monocytes were enriched in different risk groups. Among them, T cells follicular helper was significantly and highly expressed in the low-risk group, and the monocytes were highly expressed in the high-risk group (Fig. 6b) .

Multivariate COX regression analysis showed that in addition to the risk score, age and stage are also independent prognostic factors for OS. To further improve the accuracy of clinicians in predicting patient prognosis, we constructed a practical and novel nomogram tool combining all parameters with independent prognostic significance (Fig. 7a) . The total score was obtained by calculating the patient's score for each parameter separately, and by making a vertical line downward from the total score on the nomogram, the clinician can easily derive the probability of OS at 1, 3 and 5 years. In addition, we also evaluated the performance of the nomogram. The calibration curves consisting of calibration points were very close to the standard curves (Fig. 7b) . This was sufficient to show that our nomogram prediction performance was very strong, especially when predicting the 3-year OS.

In this study, the 24 ARGPs obtained from the screening were constructed into a new prognostic signature and its performance was further validated in an independent GEO cohort. The results showed the robust performance of our gene signature in predicting the prognosis of GC patients and accurately distinguishing high-risk populations. Then, subgroup survival analysis of age, sex, grade and stage also verified the stability of the signature. In addition, multivariate Cox analysis showed that the risk score was an independent prognostic factor for GC patients. In addition, the new prognostic signature identified performed well compared to the two published prognostic signatures, which is an important strength of this study. Finally, our nomogram constructed by combining risk score and clinicopathological parameters with independent prognostic significance can improve the predictive accuracy of OS in patients with GC. Due to the heterogeneity of GC and the technical noise caused by cross-platform sequencing, the previously established prognostic signature needs to standardize the mRNA expression profile, which is a difficult but necessary step in data analysis. However, our study differs from previous literature in the building of prognostic signatures. The construction of the prognostic signature is based on gene pairs rather than single Year Survival

Year Survival

Year survival genes in this study. This method of data analysis using gene expression values for pairwise comparison is a novel algorithm that can overcome the above-mentioned problems. The use of novel algorithms to construct signatures is an advantage in this study, and has been proven to be reliable in previous studies [27] . ARGPs signature not only have a predictive prognostic function, but also reduce the detection process to some extent. For instance, both genes, RECQL4 and NUDT1, were overexpressed in GC tissues. Increased RECQL4 expression was associated with poor prognosis in GC patients [28] . The overexpression of NUDT1 was closely related to the increased depth of invasion and lower survival rate of GC patients [29] . These two mRNAs formed an ARGP in this study, and when the expression value of RECQL4 was lower than the expression level of NUDT1, a high-risk score was acquired, leading to poorer prognostic outcomes.

The results require only a brief comparison of the magnitude of expression values for the constituent genes in the prognostic gene pairs, without the need for specific expression data. Thus, its feasibility is better than traditional gene signature detection. This novel signature is composed of 36 unique AGs, some of which have been intensively studied in GC and other cancers. For example, previous studies have shown that the expression of PDFGRB was related to lymphatic metastasis and poor prognosis of GC [30] . CDKN2B is highly expressed in GC tissues compared with adjacent tissues, and can be used as a prognostic marker for GC [31] . Ai et al. 's study showed that PIK3R1 overexpression can promote the progression of hepatocellular carcinoma [32] . The results of Irene Arroyo-Solera et al. [33] showed that high SERPINE1 expression promotes tumor invasion and metastasis, and is also associated with poor prognosis in head and neck squamous cell carcinoma patients. The above evidence is sufficient to suggest that ARGPs are not surprising as a prognostic signature for GC.

Recently, a lot of evidence supports the fact that the tumor microenvironment plays a huge role in determining the occurrence and development of tumors [34] [35] [36] . The results of Zeng et al. on the role of tumor immune infiltrating cells in GC suggest that it can be used as independent prognostic markers [37, 38] . Therefore, in this study we compared the differences in the abundance of tumor immune infiltrating cells between the high-risk and low-risk groups, and the results showed that two types of immune cells, T cells follicular helper, monocytes, differed between the two groups. There into, T cells follicular helper was more enriched in the low-risk group. Pan et al. found that T cells follicular helper was associated with poorer results in clear cell renal cell carcinoma [39] . The biological significance of T cells follicular helper in GC needs further study. In addition, compared with the low-risk group, the high-risk group had higher levels of monocytes. Increased monocytes have a negative impact on tumor immunity and are related to the poor prognosis of GC [40] .

Nomogram is a powerful and easy-to-use tool that has been widely used in previous biomedical studies [41] [42] [43] . The clinicopathological parameters we included in the nomogram were all factors with independent prognostic significance obtained by multivariate Cox regression analysis; in this way, the nomogram plot is more convincing for predicting the prognosis of GC patients, and we also tested the accuracy of the tool using calibration plot curves. Although studies have been developed to predict prognosis nomogram plots for GC patients, this is the first one to be constructed in conjunction with ARGPs signature, providing new ideas for the development and use of nomograms. Nevertheless, some limitations to our study should be acknowledged. First of all, we cannot avoid that our research results are based on retrospective analysis, and the results should be verified in prospective studies. Second, functional experimental studies should be conducted on these AGs to progressively elucidate the function of these ARGPs in GC.

In summary, we have developed a robust ARGPs signature using a new algorithm. It can still accurately predict the prognosis of individual patients in independent GEO data set. In addition, its performance was better than previously published prognostic models.

The preprocessed level 3 RNA-seq data and associated clinical information of GC were mined from TCGA data portal. Similarly, independent validation dataset GSE15459 was downloaded from GEO, including gene expression files and clinical follow-up information of 192 GC sample. 307 human AGs were obtained from the human aging genome resource (HAGR, http:// genom ics. senes cence. info/ genes/). Probe IDs correctly changed to the corresponding gene symbols based on the annotated information on the platform without further standardization. If multiple probe IDs are mapped to a single gene symbol, the average value is calculated as the expression value of the gene for further analysis. Expression data for 307 AGs were extracted from the TCGA cohort and the GSE15459 dataset, respectively, and the AGs common to both datasets were thus obtained.

First, the median absolute deviation (MAD) < 0.5 was used as a filtering condition to discard those AGs that were imbalanced in distribution or had fairly little mutations [21] . The values of gene expression in a given sample or profile were compared pairwise to obtain a score for each ARGP. Briefly, the value was assigned as 1 when the first AG exhibit higher expression relative to the second AG in each ARGP; otherwise, the output was 0. In addition, we also filter out those ARGPs with a ratio of "0" or "1" < 20%. This novel algorithm was one of the strengths of this study because ARGP was derived from pairwise comparisons and was based entirely on gene expression in the same sample, which can overcome the bias associated with sequencing across different platforms and does not require additional processing of the data. A total of 1278 ARGPs were used as initial candidate factors for prognostic prediction in two datasets. Through univariate Cox regression and the Kaplan-Meier (K-M) method, a summary of 155 ARGPs related to the prognosis of survival were screened out (P < 0.05). To get a more practical model, we used least absolute shrinkage and selection operator (LASSO) regression to eliminate highly correlated ARGPs. Finally, 24 ARGPs were reserved for constructing prognostic signature. For every patient, the risk score was obtained according to the following model formula we constructed:

where "coef " was the coefficient estimated from the LASSO regression analysis and "Exp" represents the score of the ARGP. We used the R package "survival" and "survival-ROC" to access the best cut-off value of a risk score by the time-dependent ROC curve analysis. We then categorized the GC patients into high-and low-risk groups, according to the best cut-off value.

Survival curves for OS were generated according to the K-M method and the log-rank test was applied to compare the differences between low-and high-risk groups. The predictive ability of the model was then determined by the area under the ROC curve (AUC). Univariate and multivariate Cox analysis were performed to estimate the influence of risk score model on prognosis and other clinicopathological factors in the TCGA dataset. Hazard ratios (HRs) with their corresponding 95% confidence intervals (CIs) were obtained.

To demonstrate the powerful performance of the prognostic model we have established, we use the R packages "limm'' , "survival", "survminer" and "timeROC" to compare it with previously published models. The AUC was used to measure the predictive performance of each model.

To further explore the underlying biological significance of AGs in GC, 39 AGs of the constructed prognostic model were used for functional enrichment analysis. Metascape is an online web tool with comprehensive functions and fast update speed. Therefore, we used this tool to investigate the biological functions and pathways involved in AGs.

The CIBERSORT algorithm is a linear support vector regression-based machine learning method that accurately identifies the relative abundance of 22 infiltrating immune cells in samples with unknown content and noise (e.g., solid tumors) compared to other deconvolution methods [44] . Therefore, we chose to use the CIBERSORT algorithm to quantify the relative abundance of 22 infiltrating immune cells, including T cells, helper follicular cells and monocytes, in all samples. In addition, to assess reliability of the results, CIBERSORT derives a P-value for the deconvolution of each sample using Monte Carlo sampling.

To optimize prognosis prediction, we integrated parameters with independent prognostic significance from multivariate Cox regression analysis to construct nomogram Risk score = coef gene1 * Exp gene1 + Coef gene2 * Exp gene2 + · · · + coef genen * Exp genen ,

Cancer statistics

Global cancer statistics

Gastric cancer prevention strategies: a global perspective

Gastric cancer: epidemiology, risk factors, classification, genomic characteristics and treatment strategies

Molecular analysis of gastric cancer identifies subtypes associated with distinct clinical outcomes

Proposal and validation of a modified staging system to improve the prognosis predictive performance of the 8th AJCC/UICC pTNM staging system for gastric adenocarcinoma: a multicenter study with external validation

The hallmarks of aging

Senescence in health and disease

Heart disease and stroke statistics-2016 update: a report from the

Misfolded protein aggregates: mechanisms, structures and potential for disease transmission

Stem cell therapy to cure type 1 diabetes: from hype to hope

Aging and epigenetic drift: a vicious cycle

The dynamic nature of senescence in cancer

mTOR is a key modulator of ageing and age-related disease

Proteins of the Myc network: essential regulators of cell growth and differentiation

p53: puzzle and paradigm

Prognostic prediction using a stemness index-related signature in a cohort of gastric cancer

Development and validation of a scoring system based on 9 glycolysis-related genes for prognosis prediction in gastric cancer

Angiogenesisrelated gene expression signatures predicting prognosis in gastric cancer patients

Development and validation of a RNA binding protein gene pair-associated prognostic signature for prediction of overall survival in hepatocellular carcinoma

Gene-pair expression signatures reveal lineage control

Development and validation of an individualized immune prognostic signature in early-stage nonsquamous non-small cell lung cancer

Development of RNA binding proteins expression signature for prognosis prediction in gastric cancer patients

Identification and prognostic value of metabolism-related genes in gastric cancer

Tumor infiltrating immune cells in gliomas and meningiomas

The prognostic landscape of tumor-infiltrating immune cell and immunomodulators in lung cancer

Identification of a poor-prognosis BRAF-mutant-like population of patients with colon cancer

Overexpression of RECQL4 is associated with poor prognosis in patients with gastric cancer

The role of miR-485-5p/NUDT1 axis in gastric cancer

Expression of platelet-derived growth factor (PDGF)-B and PDGF-receptor β is associated with lymphatic metastasis in human gastric carcinoma

Overexpression of CDKN2B is involved in poor gastric cancer prognosis

Overexpression of PIK3R1 promotes hepatocellular carcinoma progression

Effect of serpinE1 overexpression on the primary tumor and lymph node, and lung metastases in head and neck squamous cell carcinoma

A mimic of the tumor microenvironment: a simple method for generating enriched cell populations and investigating intercellular communication

The microenvironment of the tumour-host interface

The tumor microenvironment and immunoscore are critical determinants of dissemination to distant metastasis

Tumor microenvironment characterization in gastric cancer identifies prognostic and immunotherapeutically relevant gene signatures

Gene expression profiles for a prognostic immunoscore in gastric cancer

The immune infiltration in clear cell renal cell carcinoma and their clinical implications: a study based on TCGA and GEO databases

Peri-operative monocyte count is a marker of poor prognosis in gastric cancer: increased monocytes are a characteristic of myeloid-derived suppressor cells

20:35 • fast, convenient online submission • thorough peer review by experienced researchers in your field • rapid publication on acceptance • support for research data, including large and complex data types • gold Open Access which fosters wider collaboration and increased citations maximum visibility for your research: over 100M website views per year research ? Choose BMC

Comment on "factors affecting sentinel node metastasis in thin (t1) cutaneous melanomas: development and external validation of a predictive nomogram

Development and validation of a model to predict survival in colorectal cancer using a gradient-boosted machine

Individualizing risk prediction for positive coronavirus disease 2019 testing: results from 11,672 patients

Robust enumeration of cell subsets from tissue expression profiles

The authors gratefully acknowledge the Gene Expression Omnibus (GEO) database, and the Cancer Genome Atlas (TCGA) database, which made the data available.

Page 12 of 14 Zhang et al. BioMed Eng OnLine (2021) 20:35 to predict 1-year, 3-year, and 5-year OS survival in GC patients. The calibration curve was utilized to assess the predictive accuracy and discriminatory power of the developed nomogram.

All statistical analyses were performed with R statistical software. Wilcoxon test was used to compare the differences high-and low-risk groups. For all assays, two-sided P value of less than 0.05 indicated a statistically significant difference.

The online version contains supplementary material available at https:// doi. org/ 10. 1186/ s12938-021-00871-x.Additional file 1: Figure S1 . The optimal cut-off value of the ARGPs risk-score obtained by the time-dependent ROC curve analysis. 

This research was conducted in collaboration with all authors. YZ and YY performed the data curation and analysis. YZ and NN analyzed and interpreted the results. YZ, ZS and YY drafted and reviewed the manuscript. All authors read and approved the final manuscript.

We are very grateful for the support of the National Natural Science Foundation of China (Grant number: 81702336) and the Peking University International Hospital Research Grant (No.YN2019XQ01).

The raw data of this study are derived from the TCGA database (https:// portal. gdc. cancer. gov/) and GEO data portal (https:// www. ncbi. nlm. nih. gov/ geo/), which are publicly available databases.

Ethics approval and consent to participate Not necessary.

Not applicable.

The authors declare that they have no competing interests.