A novel lncRNA-focus expression signature for survival prediction in endometrial carcinoma

Patient datasets

Acquisition and processing of mRNA and lncRNA expression profiles in UCEC patients

Genome-wide mRNA and lncRNA expression profiles (RPKM expression levels) were downloaded from TCGA long non-coding RNAs database (http://larssonlab.org/tcga-lncrnas/index.php) according to Akrami’s study [17]. Briefly, the acquisition and processing of mRNA and lncRNA expression profiles were performed by Akrami et al. as follows [17]: TCGA RNA-seq data in FASTQ format was realigned to the Hg19 assembly using TopHat software and read counts for each lncRNA and mRNA were obtained using HTSeq-count. Then, RPKM values were used to quantify expression levels of lncRNAs and mRNAs by normalizing for lncRNA or mRNA length and library size and were log transformed using log2 (RPKM + 0.01) [17]. A total of 20,462 mRNAs and 10,419 lncRNAs were finally retained in the further analysis.

Statistical analysis

Univariate Cox regression analysis was used to select candidate prognostic lncRNAs that were significantly correlated with overall survival at the significance level of 1%. All candidate prognostic lncRNAs were subjected to the multivariate analysis with Cox proportional hazard model for identifying lncRNA biomarkers with independent prognostic value. The survival rate and median survival for each prognostic risk group were calculated using the Kaplan-Meier method. The survival difference between the high-risk group and the low-risk group was assessed by log-rank test with 5% significant level. Univariate Cox analysis was performed to evaluate the prognostic value of lncRNA signature. To assess the independence between lncRNA signature and the key clinical factors, multivariate Cox regression and stratification analyses were conducted. Hazard ratios (HRs) and 95% confidence intervals (CIs) were computed by the Cox analysis. The comparison of survival prediction based on lncRNA signature and key clinical characteristics were performed by the time-dependent receiver operating characteristic (ROC) analysis. Kruskal-Wallis test was used to compare expression levels for each lncRNAs across four UCEC subtypes. All statistical analyses were performed using R/Bioconductor.

Formulation of lncRNA-focus expression signature

Here, Risk Score(patient) is a LFES-based risk score for UCEC patient. lncRNA_i represents the ith prognostic lncRNA and expression(lncRNA_i) is the expression level of lncRNA_i for the patient. Regression coefficient of multivariate Cox analysis was denoted as coefficient(lncRNA_i) which represents the contribution of lncRNA_i for prognostic risk scores. Patients with higher risk score tend to have a poor survival outcome. The median risk score for discovery cohort was selected as the cutoff point. Based on this cutoff, patients in the discovery cohort, validation cohort and entire TCGA cohort can be assigned to a high-risk group or a low-risk group.

In silico analysis of lncRNA function

Co-expression relationship was evaluated between lncRNAs and mRNAs using paired expression profiles of lncRNAs and mRNAs in entire TCGA UCEC patients, and lncRNA-mRNA co-expression network was constructed. Functional enrichment analysis of mRNAs in the lncRNA-mRNA co-expression network was used to infer potential biological processes and pathways of prognostic lncRNAs according to Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) through DAVID Bioinformatics Resources (https://david.ncifcrf.gov/, version 6.8) [18]. Finally, the top one of significantly enriched GO terms or KEGG pathways was considered as a potential function of prognostic lncRNAs.

Patient’s characteristics

A total of 150 UCEC samples were randomly selected from 301 UCEC samples as discovery cohort, and other 151 UCEC samples composed the validation cohort. The details of clinical characteristics for both cohorts were listed in Table 1. The clinical variables, including stage, grade, histology and vital status, were similar in the training and validation cohorts. Results of the statistical analysis exhibited that the random assignment with the discovery and validation cohorts was in equilibrium with these clinical characteristics.

Development of lncRNA-focus expression signature for survival prediction in UCEC

To build a lncRNA-focus expression signature for survival prediction, lncRNA expression profiles of the selected 11 independent prognostic lncRNAs were used to build the multivariable Cox regression model for evaluating their relatively predictive power. We constructed lncRNA-focus expression signature (LFES) for survival prediction by weighted scoring method using expression level of independent prognostic lncRNAs weighted by their regression coefficients in above multivariate Cox analysis as follows: Risk Score (patient) = (5.0432 * expression value of RP11-1072A3.3.1) + (0.8462 * expression value of ACVR2B-AS1) + (6.3725 * expression value of RP4-781 K5.7.1) + (1.9110 * expression value of AC073046.25) + (1.9166 * expression value of AP001347.6) + (0.3553 * expression value of DOCK9-AS2) + (−0.2987 * expression value of NRAV) + (−6.896 * expression value of GTF3C2-AS1) + (−0.8517*expression value of LINC01006) + (0.5747 * expression value of RP11-531A24.5) + (0.2325 * expression value of AC004947.2).

Prognostic validation of LFES in the discovery cohort

To assess the prognostic value of the predictive model, a LFES-based risk score was generated for each patient in the discovery cohort by the expression level of 11 lncRNAs. The median risk score was obtained from the discovery cohort and was selected as the threshold point (1.703). According to the risk score and the threshold point, patients of discovery cohort were classified into high-risk group (n = 75) and low-risk group (n = 75). Survival analysis showed that there was a significant difference in overall survival (p < 0.001, log-rank test) (Fig. 1a) and progression-free survival (p = 0.006, log-rank test) (Fig. 1b) between patients in the high-risk group and low-risk group. As shown in Fig. 1a, patients in the high-risk group only have 3- and 5-year survival rates of 71.2% and 65.2%, respectively, compared to the patients in the low-risk group with 3- and 5-year survival rates of 100%. In a univariate Cox regression analysis, the hazard ratios of high-risk group versus low-risk for overall survival was 2.718 (p < 0.001, 95% confidence interval (CI) = 1.923–3.842) (Table 3).

Variables	Unfavorable/Favorable	Univariate			Multivariate
		HR	95% CI	P vaule	HR	95% CI	P vaule
Discovery cohort (n = 150)
11-lncRNA risk score	High/Low	2.718	1.923–3.842	<0.001	2.649	1.788–3.923	<0.001
Age		1.055	1.004–1.108	0.035	0.992	0.931–1.057	0.797
Stage	(III + IV)/(I + II)	4.122	1.588–10.7	0.004	1.219	0.392–3.795	0.732
Grade	3/(1 + 2)	4.397	1.258–15.37	0.020	1.609	0.397–6.52	0.506
Histology	Serous/Endometrioid	1.731	0.604–4.962	0.307	0.728	0.217–2.445	0.607
Validation cohort (n = 151)
11-lncRNA risk score	High/Low	6.903	1.521–31.340	0.012	6.158	1.205–31.465	0.029
Age		1.042	0.986–1.101	0.141	1.046	0.975–1.122	0.208
Stage	(III + IV)/(I + II)	7.160	2.196–23.340	0.001	7.153	1.601–31.955	0.010
Grade	3/(1 + 2)	2.632	0.879–7.885	0.084	0.681	0.150–3.083	0.618
Histology	Serous/Endometrioid	4.873	1.627–14.600	0.005	0.691	0.099–4.830	0.709
Entire TCGA cohort (n = 301)
11-lncRNA risk score	High/Low	11.767	3.568–38.810	<0.001	10.793	3.084–37.777	<0.001
Age		1.050	1.012–1.09	0.009	1.064	1.018–1.112	0.006
Stage	(III + IV)/(I + II)	4.835	2.359–9.906	<0.001	3.948	1.759–8.859	0.001
Grade	3/(1 + 2)	3.206	1.433–7.177	0.005	1.263	0.490–3.257	0.628
Histology	Serous/Endometrioid	2.584	1.236–5.402	0.012	0.509	0.209–1.240	0.137

The expression pattern of 11 prognostic lncRNAs, the distribution of the risk score and the survival status of UCEC patients for the discovery cohort was shown in Fig. 1c. Ten risky lncRNAs are over-expressed among patients with the high-risk score, but the protective lncRNA, NRAV, often would express in the low-risk cases.

Further confirmation of LFES for survival prediction in the validation cohort and entire TCGA cohort

To validate the universality of LFES for identification of UCEC patients with poor outcome, we examined the ability of LFES in the independent validation cohort. By using the same LFES-based risk score model, the patients of the validation cohort were divided into high-risk group (n = 78) and low-risk group (n = 73) according to the same threshold point as for the discovery cohort. Patients with high-risk LFES had significantly shorter overall survival and progression-free survival than those with the low-risk signature (p = 0.004, log-rank test) (Fig. 2a and b). The 3- and 5-year survival rates of the high-risk group were 82.5% and 57.9%, respectively, whereas the corresponding rates in the low-risk group both were 95.6%. Notably, there were 11 cancer-related deaths in the high-risk group and only three death events in patients with low-risk scores. The hazard ratios of high-risk group versus low-risk group for overall survival was 6.903 (p = 0.012, 95% CI = 1.521–31.340) (Table 3).

We also elevated the prognostic value of LFES in the entire TCGA cohort. The LFES could also distinguish between patients with the good and poor outcome, which is consistent with the findings from the discovery and validation cohorts. Kaplan-Meier survival curves based on the LFES were significantly different (p < 0.001, log-rank test) (Fig. 2c and d). The median survival time for patients with high-risk scores was 108 months. In sharp contrast, the patients with low-risk scores had not reached the threshold to calculate their median survival time. The survival rates at 3- and 5-year were 77.5% and 63.5% for patients in the high-risk group compared with both 97.8% for patients in the low-risk group. By subjecting the risk scores to univariate Cox regression analysis, patients with high-risk scores exhibited an 11.767-fold increased risk than patients with low-risk scores (Table 3). The expression pattern of 11 prognostic lncRNAs, the distribution of the risk score and the survival status of UCEC patients for the validation and entire TCGA cohorts was shown in Fig. 2e and f, which is consistent with findings in the discovery cohort.

Correlation between LFES and other clinicopathologic characteristics or subtype

To evaluate independent prognostic values of the LFES in survival prediction, we performed multivariate Cox regression analysis to test the performance of the LFES, including LFES-based risk scores, age, stage, grade and histology as covariates and overall survival as the dependent variable. In the discovery cohort, only the LFES was significant in multivariate analysis (p < 0.001, Table 3) compared to these clinical characteristics of age, stage and grade. Furthermore, the hazard ratios of high-risk group versus low-risk group for overall survival were 6.158 (p = 0.029, 95% CI = 1.205–31.465) in the validation cohort and 10.793 (p < 0.001, 95% CI = 3.084–37.777) in the entire TCGA cohort after adjustment by these clinical characteristics (Table 3), respectively, indicating that the LFES maintained an independent correlation with overall survival.

Additionally, we found that age (HR = 1.064, 95% CI = 1.02–1.11, p = 0.006) and stage (HR = 3.948, 95% CI = 1.76–8.86, p = 0.001) were both significantly prognostic factors associated with survival for all UCEC patients (Table 3). The stratification analysis was performed to ascertain that lncRNA signature was independent of age and stage. The 301 UCEC patients were assigned into a young set (age < =63, n = 152) and an old set (age > 63, n = 149). For the young set, the lncRNA risk score could further divide patients into a better survival subgroup (n = 68) or poorer survival subgroup (n = 84) (p = 0.001, log-rank test) (Fig. 3a). Patients in the old set exhibit the same trend (Fig. 3b). For elder patients, the LFES also assigned the patients into two subgroups with significantly different survival (p < 0.001, log-rank test) (Fig. 3b). The analysis demonstrated that the LFES was free from age. To evaluate whether the LFES may predict the survival of patients within each stage stratum, stratified analysis based on stage was carried out. All UCEC patients were divided into an earlier stage stratum (stage I and II patients) or a later stage stratum (stage III and IV patients). The LFES was performed to distinguish high-risk and low-risk patients in each stage stratum. By the KM curves shown in Fig. 3c and d, patients with high-risk scores have significantly shorter survival than those with low-risk scores for earlier stage stratum (p = 0.012, log-rank test) and later stage stratum (p < 0.001, log-rank test) (Fig. 3c and d). Multivariate and stratification analysis shows that prognostic power of the LFES was independent of other clinicopathological factors for survival prediction in UCEC patients.

We compared the prognostic performance of the LFES with other clinical characteristics used for risk stratification of UCEC patients, including age, stage and BMI. Time-dependent ROC analysis was conducted to compare the sensitivity and specificity of survival prediction. The AUC for each of the prognostic factors was calculated and compared. As shown in Fig. 4, the AUC of LFES was 0.887 that is significantly higher than age (AUC = 0.63), stage (AUC = 0.763) and BMI (AUC = 0.551). These results showed that the LFES had a better prognostic performance than other prognostic factors.

Finally, we compared expression level of 11 lncRNAs in the LFES across four UCEC subtypes (Ultramutated (POLE), Hypermutated (MSI), Low CN (MSS) and High CN (Serous-like)) identified by The Cancer Genome Atlas Research Network based on a combination of somatic nucleotide substitutions, MSI and SCNAs [16]. The results indicated no significant difference in the distribution of expression levels for all 11 prognostic lncRNAs across four UCEC subtypes (Additional file 2), implying that the LFES is not a subtype-specific marker.

Functional roles of prognostic lncRNAs in the signature in UCEC biology

In order to understand functional roles behind the LFES in UCEC biology, we performed in silico analysis for lncRNA function through functional enrichment analysis. An integrated lncRNA-mRNA co-expression network was generated by calculating the Pearson correlation coefficient between expression values of prognostic lncRNAs and those of mRNAs in the entire TCGA patients. Functional enrichment analysis of GO and KEGG was performed for co-expressed mRNAs to infer potential biological processes and pathways of prognostic lncRNAs. We found that these prognostic lncRNAs may be involved in Wnt signaling pathway, Rho protein signal transduction, cell cycle, protein ubiquitination, phosphatase signaling pathway, epidermal growth factor receptor (EGFR) signaling pathway, Notch signaling pathway, immune response, PPAR signaling pathway, ion transmembrane transport and cell proliferation (Fig. 5). It suggested that lncRNAs in the LFES played important roles in UCEC biology.