Integrated analysis reveals potential long non-coding RNA biomarkers and their potential biological functions for disease free survival in gastric cancer patients

The study datasets in the current study were obtained from The Cancer Genome Atlas (TCGA) database (https://​cancergenome.​nih.​gov/​) and the Gene Expression Omnibus (GEO) database (https://​www.​ncbi.​nlm.​nih.​gov/​gds/​). The current study downloaded and analyzed the study datasets according to the data policies of TCGA database and GEO database. The study datasets obtained from TCGA database and GEO database were fully anonymous and therefore the ethics approval was not required.

The model dataset was downloaded from TCGA database (https://​tcga-data.​nci.​nih.​gov/​docs/​publications/​tcga/​). The gene expression values were generated by using the Illumina HiSeq 2000 RNA Sequencing platform. In the current study, the selected lncRNA IDs were defined according to GENCODE Version 29 (https://​www.​gencodegenes.​org/​human/​). Finally, there were 14,449 lncRNAs obtained from 375 tumor specimens and 32 normal specimens in model dataset. The original gene expression counts were converted to 0 for low expression and 1 for high expression according to the median values of original gene expression values. The clinical information of 443 GC patients in model dataset were downloaded from cBioPortal database (http://​www.​cbioportal.​org/​datasets). The GC patients without disease free survival information were excluded from the current study (n = 120). In order to avoid or reduce the interferences of confounding factors, there were 10 GC patients excluded from the current study because disease free survival time less than 1 month (n = 10). After interaction between gene dataset and clinical dataset, there were 265 GC patients with gene expression information and survival information included in the model cohort (Fig. 1). The maximum DFS time was 122.21 months and the minimum DFS time was 1.02 month in model cohort. The study time was from January 13, 2002 to November 24, 2014. The missing data were coded as “NA” in the model dataset. The mean ± standard deviation (SD) age of GC patients in the model cohort was 64.4 ± 10.6 years. Ninety-eight (37.0%) patients out of 265 GC patients died within the follow-up period (mean ± SD: 618 ± 564 days).

We explored and identified potential study datasets in GEO database according to the following criteria: (1) gene expression values detected by using the Affymetrix Human Genome U133 Plus 2.0 array; (2) DFS time and DFS status were provided in the corresponding studies; (3) patient number more than 100. Finally, we identified GSE62254 as an independent external validation dataset [18]. The gene expression information and corresponding clinical information were downloaded from the GEO database (http://​www.​ncbi.​nlm.​nih.​gov/​geo/​). There were 300 GC patients with gene expression information and clinical information from GSE62254 (https://​www.​ncbi.​nlm.​nih.​gov/​geo/​query/​acc.​cgi?​acc=​GSE62254). The gene expression information was detected on Affymetrix Human Genome U133 Plus 2.0 chip platform. The patient selection flow chart was described in Fig. 1. The Affymetrix HG-U133 Plus 2.0 probe set IDs were mapped to the Ensembl gene IDs by using the platform background file of GPL570 (https://​www.​ncbi.​nlm.​nih.​gov/​geo/​query/​acc.​cgi?​acc=​GPL570).

Based on the lncRNA IDs defined in GENCODE Version 29 (https://​www.​gencodegenes.​org/​human/​), 1978 lncRNAs with corresponding gene symbols from 300 GC patients were extracted and analyzed for further survival study.

The prognostic nomogram and the corresponding calibration plots were generated by using “rms” package of R software. Calibration plots were performed to evaluate the predictive performance of the prognostic nomogram. The predicted survival and observed survival were plotted on the x-axis and y-axis respectively. The 45-degree line represented the best predictive curve. Time-dependent ROC curves were conducted to assess the predictive performance of the prognostic nomogram by using the “pROC” package.

The decision curve analysis (DCA) was performed to evaluate the clinical utility of the prognostic nomogram for disease free survival. The DCA is a method for evaluation and comparison of the predictive value between different prediction models [19‐21]. The x-axis of DCA represented the percentage of threshold probability, and the y-axis represented the net benefit of the predictive model. The net benefit was calculated according to the following formula: Net benefit = (True positives/n) − (False positives/n) * (p_t/(1 − p_t).

To explore the potential biological functions of lncRNAs included in the prognosis signature, the co-expressed mRNA were obtained through the model dataset according to the thresholds of P value < 0.05 and |spearman correlation coefficient| > 0.5. Functional enrichment analyses of these co-expressed mRNAs were performed by using the Database for Annotation, Visualization and Integrated Discovery (DAVID, https://​david.​ncifcrf.​gov/​) [22].

Continuous variables were displayed as mean ± standard deviation (SD). Continuous variables were compared by t test or Mann–Whitney U test as appropriate. Categorical variables were compared by using Chi squared test or Fisher’s exact test as appropriate. To explore the potential associations between lncRNAs and DFS, univariate Cox regression analyses and multivariate Cox regression analyses were performed to identify the prognostic biomarkers for development of prognosis signature. The GC patients were divided into two subgroups according to the scores generated by the prognosis signature. Kaplan–Meier analysis was used to compare the difference of DFS between high risk group and low risk group. The predictive performance and clinical utility of prognostic signatures were evaluated by using the Harrell’s concordance index and time-dependent receiver operating characteristic (ROC) curve. The statistical analyses in the present study were conducted by SPSS Statistics 19.0 (SPSS Inc., an IBM Company) and R software (version 3.4.5). The R packages, including “pROC”, “plyr”, “rms”, “survival”, “timeROC “ and “glmnet “, were used as needed in the current study. P value < 0.05 was defined as statistically significance in the current study.

The flow chart of patient selection in the current study was showed in Fig. 1. There were 265 GC patients in the model group (Additional file 1) and 300 GC patients in the validation group (Additional file 2). There were 98 (37.0%) patients out of 265 patients died within the follow-up period in the model group, whereas there were 161 (53.7%) patients out of 300 patients died within the follow-up period in the validation group. The basic clinical characteristics of GC patients in the model group and validation group were presented in Table 1.

The univariate Cox proportional regression analyses were carried out to explore the potential lncRNA predictors for disease free survival in GC patients. There were 249 lncRNAs that were significantly related with DFS in the model group. According to the multivariate Cox regression analyses, there were 13 lncRNAs identified as independent biomarkers for disease free survival in GC patients. The relevant estimated regression coefficients of these 13 prognostic lncRNAs were showed in Table 2. Therefore, a prognostic signature was conducted according to the following formula: Prognostic signature score = (− 0.712*GAS5-AS1) + (− 0.725*AL109615.3) + (− 0.830*KDM7A-DT) + (− 0.837*AP000866.2) + (− 0.984*KCNJ2-AS1) + (− 1.150*LINC00656) + (0.764*LINC01777) + (0.769*AC046185.3) + (0.775*TTTY14) + (0.783*LINC01526) + (0.803*LINC02523) + (0.941*LINC00592) + (1.162*C5orf66). According to the multivariate Cox regression analyses, a prognostic nomogram for prediction of disease free survival in gastric cancer patients was presented in Fig. 2.

This nomogram could be used to predict the individual mortality risk. The value of each lncRNA was calculated according to the corresponding scale. The total score was obtained by adding the values of these thirteen lncRNA. The total score was projected to the probability of DFS in 1 year, 3 year, and 5 year respectively.

For displaying the distribution characteristics of prognostic signature score, the violin plot (Fig. 3a), density plot (Fig. 3b), scatter plot (Fig. 4a), the interaction distribution scatter plot among DFS time, DFS status, and predictive value (Fig. 4b) were presented in Figs. 3 and 4.

According to the median value of prognostic signature score, GC patients in model cohort (n = 265) were stratified into high risk group (n = 133) and low risk group (n = 132). The disease free survival rate (Fig. 5a) in high risk group was significantly poorer than that in low risk group (P < 0.001). The cumulative proportion surviving at 1-year, 3-year, and 5-year were 94.3%, 81.7%, and 77.1% in low risk group, whereas it were 59.2%, 27.8% and 11.3% in high risk group respectively (all P < 0.001). The Harrell’s concordance indexes (C-indexes) of prognostic signature for disease free survival in the model group were 0.849 (95% CI 0.803–0.895), 0.859 (95% CI 0.813–0.905) and 0.888 (95% CI 0.842–0.934) for 1-year disease free survival, 3-year disease free survival and 5-year disease free survival respectively (Fig. 5b). The calibration curves for 1-year, 3-year, and 5-year disease free survival were presented in Fig. 6.

We validated the clinical utility of prognostic signature by using GSE62254 dataset. The prognostic signature scores for GC patients in validation dataset were calculated according to the previous prognosis signature formula in model dataset. Then the GC patients in validation set (n = 300) were divided into low risk group (n = 150) and high risk group (n = 150). The cumulative proportion surviving at 1-year, 3-year, and 5-year were 89.3%, 70.1%, and 62.3% in low risk group, whereas it were 42.4%, 24.1% and 21.4% in high risk group respectively (all P < 0.001). Kaplan–Meier analysis (Fig. 7a) indicated that there was significant difference in term of DFS between low risk group and high risk group in validation set (P < 0.001). The C-index of prognostic signature for disease free survival in the validation group were 0.870 (95% CI 0.834–0.906), 0.822 (95% CI 0.786–0.858) and 0.816 (95% CI 0.780–0.894) for 1-year, 3-year, and 5-year disease free survival respectively (Fig. 7b). The calibration curves for 1-year, 3-year, and 5-year disease free survival were presented in Fig. 8 for validation cohort.

We carried out multivariate Cox regression analyses to explore whether prognostic signature was independent to other clinical parameters for DFS in GC patients. The pathological diagnosis was performed according to the suggestions of American Joint Committee on Cancer (AJCC). Table 3 indicated that prognostic signature was an independent risk factor for DFS after adjustment for confounding effects of gender, age, and pathological stage in the model group. In the validation group, multivariate Cox regression analyses demonstrated that prognostic signature, age, and AJCC stage were independent risk factors for DFS.

According to a threshold of P value < 0.05 and |spearman correlation coefficient| > 0.5, there were 1280 mRNA genes that significantly co-expressed with the prognostic lncRNAs in the prognostic signature. Functional enrichment analyses were carried out through the Database for Annotation, Visualization, and Integrated Discovery (DAVID) Bioinformatics Resources (https://​david.​ncifcrf.​gov/​). The gene ontology (GO) biological process enrichment analyses and the Kyoto Encyclopedia of Genes and Genomes (KEGG) signaling pathways were presented in Fig. 9. Functional enrichment analyses of the 1280 co-expressed mRNA genes demonstrated that the co-expressed mRNA genes were mainly enriched in regulation of transcription, RNA splicing, protein ubiquitination, cellular response to DNA damage stimulus, cilium assembly, cilium morphogenesis, centrosome organization, G2/M transition of mitotic cell cycle, regulation of transcription from RNA polymerase II promoter.

As shown in Fig. 10, the prognostic signature (red line) had the higher net benefit than the pathological stage (green line). The decision curve analyses indicated that prognostic signature could gain more benefit than either the dead-all-patients scheme or the dead-none-patients scheme for prediction of 1-year DFS (Fig. 10a), 3-year DFS (Fig. 10b), and 5-year DFS (Fig. 10c). Clinical impact curve (Fig. 10d) depicted the prediction of risk stratification of 1000 patients by using resample bootstrap method. “Number high risk” indicated the number of patients classified as positive (high risk) by prognostic signature according to various threshold probabilities. “Number high risk with event” was the true positive patient number according to various threshold probabilities.

To explore the predictive performance of prognostic signature for DFS, a 10-group risk stratification chart was presented in Fig. 11. For model cohort, the discriminative ability of prognostic signature for 1 year, 2 year, and 3 year DFS were showed in Fig. 11a–c. For validation cohort, the discriminative ability of prognostic signature for 1 year, 2 year, and 3 year DFS were showed in Fig. 11d–f. Figure 11 demonstrated that patients in lower risk groups had higher survival probability and patients in higher risk groups had lower survival probability.

Firstly, this prognostic nomogram can directly provide the individual mortality percentage forGC patients, which is important for improvement of individualized treatment decision-making. Secondly, the thirteen-lncRNA prognostic signature is easy to calculate and understand by users without medical knowledge and professional calculation tool. Thirdly, for patients without pathological diagnosis or unwilling to undergo surgery, the thirteen-lncRNA prognostic signature provides a simple non-invasive preoperative predictive method for prognosis of GC patients, which is of clinical significance for individualized clinical decision-making before surgery.

As a clinical study by using study datasets downloaded from public databases, the model dataset and validation dataset did not contain detailed study information of drug regimen and other postoperative treatments, which might influence the therapeutic effect and prognosis. Additionally, the results in the current study depended on gene mining approach and lacked evidences from clinical experimental researches. Therefore, it is necessary to carry out large prospective studies to elucidate the relationship between the prognostic lncRNA biomarkers and DFS in GC patients.