A seven-lncRNA signature predicts overall survival in esophageal squamous cell carcinoma

Esophageal cancer ranks the 8th most common type of cancer worldwide and the 6th leading cause of cancer mortality¹. There are two main histological types of esophageal cancer: esophageal adenocarcinoma (EAC) and esophageal squamous cell carcinoma (ESCC). These two cancer types differ from each other in terms of causes, incidence patterns and biology features. Although the incidence of EAC is increasing rapidly in Western countries, ESCC still remains dominant in East Asian². Besides, the overall 5-year survival rate of ESCC remains extremely poor with a high probability of recurrence and metastasis³. Despite the tumor-node-metastasis (TNM) system has been widely used as prognostic factors, substantial differences exist in survival among patients within the same clinical stage, as a result of the heterogeneous of ESCC. Hence, there is an urgent need for fully comprehensive research into the crucial molecular mechanisms associated with the prognosis of ESCC.

Long non-coding RNAs (lncRNAs) are defined as RNA transcripts longer than 200 nucleotides that lack protein-coding abilities⁴. Nowadays, lncRNAs have attracted increasing scientific interest and recent evidence revealed their role as an important molecular players in modulating diverse biological processes. They have been reported to regulate gene expression through chromatin modification, transcriptional and post-transcriptional processing⁵. For instance, the well-known lncRNA HOTAIR induce the transcriptional repression of HOX loci and genome-wide retargeting of PRC2 (polycomb repressive complex 2) which results in altered histone H3K27 methylation and metastasis-related gene expression⁴.

In addition to the regulation of biological process, recent studies have revealed that lncRNAs can serve as potential prognostic biomarkers and several prognostic lncRNA signatures have been identified and validated in many cancer types, such as gastric cancer, colorectal cancer and clear cell renal cell carcinoma^6–8. However, the prognostic role of lncRNA in ESCC remain largely unknown, mainly due to the lack of the comprehensive and systemic analysis of lncRNA profiling analysis in ESCC⁹. Presently, since the recent release of gene expression data and related prognosis information in Gene Expression Omnibus (GEO) and The Cancer Genome Atlas (TCGA), we mined the LncRNA data from the GEO and conducted lncRNA profiling on ESCC patients. We identified a prognostic, seven-lncRNA signature for ESCC from the training set of GEO and validated its prognostic value in two independent test sets including the GEO validation set and another independent TCGA test set.

For most type of cancers, including ESCC, TNM stage still act as the main reference to direct the treatment strategies and is used as a prognostic predictor. However, as a result of the heterogeneity of cancer at the molecular and genetic levels, the clinical outcome and prognosis of patients diverse even if they are in the same stage and received similar treatment^12,13. Currently, with the advancements of high-throughput technologies including microarray and RNA sequencing, gene expression profiling has become a powerful technique to identify the molecular biomarkers of esophageal cancer phenotypes or prognosis¹⁴. Multigene signatures which is designed to analyze the activity of a group of genes that strongly correlated with the behavior of the cancer have been marketed already, such as Oncotype DX Test for breast cancer or ColoPrint for colon cancer. These signatures can be applied to help cancer treatment and prognosis management¹⁵.

Growing evidence suggests that the aberrant expression of specific lncRNAs may acts as major contributor to tumorigenesis and intimately correlated with tumor progression. Recent studies have focused on the role of specific lncRNAs which serves as independent markers for predicting prognosis in disease such as colorectal cancer, glioma and pancreatic cancer^8,15–17. Although a series previous articles have revealed the potential value of lncRNAs in ESCC prognosis predicting, such as HOTAIR¹⁸, CCAT2¹⁹ and MALAT1^20,21. However, the use of the combination of lncRNAs in predicting ESCC prognosis have not been elucidated clearly.

Here, we analyzed the lncRNAs expression profiles of patients with ESCC downloaded from GEO and identified a robust seven-lncRNAs signature associated with overall survival which was independent of classical prognostic factors and molecular subtypes. The prognostic value of the lncRNAs signature was further validated in the GEO validation set and an external independent test set from TCGA. When we tried to identify the prognosis related lncRNAs from GEO training set which is the high-throughput biological data, the common problem, ‘curse-of-dimensionality’ (small sample size combined with a very large number of genes) was taken into consideration. In view of this, we applied the RSF algorithm to pick out lncRNAs and narrow down the high dimension. The random forests method bases predictions on majority voting of a collection of decision trees which exploits maximal sub-trees for effective variable selection. The criteria of gene importance is used to filter the original gene set iteratively which results in good performance in feature selection²². The random sampling and ensemble strategies used in the RSF method achieves greater stability and accurate predictions while running efficiently on ‘curse-of-dimensionality’ data²³.

Next, the prognostic related lncRNAs were further selected to construct a risk score formula by Cox regression model. Cox Regression model builds a predictive model for time-to-event data. The model produces a survival function that predicts the probability that the event of interest has occurred at a given time for given values of the predictor variables. The shape of the survival function and the regression coefficients for the predictors are estimated from observed subjects; the model can then be applied to new cases that have measurements for the predictor variables²⁴. After subjecting the selected genes to Cox regression analysis, a risk score formula was constructed based on their estimated regression coefficients. By applying the seven-lncRNA signature to the GEO training set, GEO validation set and TCGA test set, obvious separation was observed in the survival curves of the high-risk group and low-risk group classified by the same criteria in all three sets which indicated the high reproducibility of this lncRNA signature in ESCC. Further analysis showed that the seven-lncRNA signature was of prognostic significance no matter it was considered as a continuous variable (in multivariable Cox regression analysis) or category variable (in log-rank p test). Moreover, multivariable Cox regression and stratification analysis demonstrated that the prognostic value of the seven-lncRNA signature was independent of the TNM stage and lncRNA signature had prognostic predict ability within clinical stages.

In order to evaluate and compare the predictive efficiency, we introduced the time-dependent ROC curves and dynamic AUC which are more useful when the data is a censored survival time. A number of previous research have applied familiar binary outcome methods such as ROC curves to evaluate the specificity and sensitivity of a marker in survival prediction. Routine ROC analysis can only characterize the accuracy of a marker by focusing on the correct classification rates of the final status. However, the survival data is usually a combination of the status at the end of follow-up (binary) and the length of follow-up (continuous). The methods which estimates only the classification of binary outcome may not be extended for survival outcomes. Hence we constructed time-dependent ROC curves to assess the sensitivity and specificity of variables in classifying death and survival on the 12th month of follow up and calculate the corresponding AUC. Then the AUC for time specific ROC curves at continuous time point was calculated and further plotted as a function of time to characterize temporal changes in accuracy. In this way, we showed that the predictive accuracy of LncRNA signature were much better than that of TNM stage. Moreover, a new variable, which combined both LncRNA and TNM, displayed better prognostic predict ability of overall survival than lncRNA signature or TNM stage alone.

Despite growing studies began focus on the molecular mechanisms of lncRNA functions in malignancy, most lncRNAs are not yet functionally annotated. LncRNAs generally function as the ceRNAs which regulate gene expression through epigenetic mechanisms or posttranscriptional events such as mRNA processing and degradation¹⁰. Hence, we can infer the possible effect of the lncRNAs on ESCC through performing functional enrichment of their related genes²⁵. The results suggested that these genes were enriched in cancer related biological process such as cell cycle regulation and histone methylation and signaling pathways such as PI3K-Akt and HIF-1 pathway. Here, we also observed the role of LINC00173 in modulating the proliferation and cell cycle of ESCC cells. Besides, the correlations between some of the LncRNAs and other tumors have already been elucidated. For instance, a recent research revealed the role of LINC00173 in the formation and maintenance of the human blood hierarchy, highlighting the function of LINC00173 in leukemia¹¹. Hence, the functional enrichment analysis uncovered the underlying molecular mechanisms of these lncRNAs in playing roles in survival prediction.

The whole process of our analysis have been plotted as a flowchart in the supplementary material (Supplementary figure).

In conclusion, this study reported a seven lncRNA signature to predict prognosis in ESCC patients by integrating and mining currently available microarray data. Moreover, the time-dependent ROC curves and dynamic AUC were introduced to evaluated its predictive accuracy which showed that the new variable combined both lncRNA and TNM displayed better prognostic predict ability than either alone. The functional enrichment analysis and experiments suggested that the lncRNAs in signature might be correlated with several cancer related processes and pathways, which supported the prognosis predictive ability of the lncRNAs. Future studies will focused on the validation of the signature in prospective clinical trials and the molecular mechanisms exploration and explanation of these lncRNAs.