Development and internal validation of a nine-lncRNA prognostic signature for prediction of overall survival in colorectal cancer patients

Study protocol approval

The data download and analyses were performed according to the policies of The Cancer Genome Atlas (TCGA) database. Since the study datasets in the present study were all downloaded from TCGA database, additional ethics approval was not needed. All data collections and analyses were in accordance with the principles of Declaration of Helsinki.

The gene expression dataset

The original RNA sequencing dataset was obtained from GDC Data Portal (https://portal.gdc.cancer.gov/exploration). The dataset was downloaded in May 31, 2018. The RNA sequencing data were generated on the Illumina HiSeq 2000 RNA Sequencing platform. The downloaded RNA sequencing data contained the original gene read counts of 60,483 genes from 458 CRC tumor tissues and 41 non-tumor normal tissues. The duplicated samples were removed from the present study (n = 22). In the present study, the lncRNAs ID were downloaded from GENCODE Version 7 (Derrien et al., 2012). Finally the extracted gene expression data of 14,449 lncRNAs from 458 CRC tissues and 41 non-tumor normal tissues were selected for differentially expressed lncRNAs.

Differential expression analyses

The “edgeR” package was applied for the differential expression analyses and the original gene expression read counts was normalized by the Trimmed Mean of M (TMM) method (Robinson & Oshlack, 2010). The genes whose average expression read counts lower than 1 were filtered out from the present study. The F-tests was used for assessments of quasi-likelihood. Thresholds of differentially expressed lncRNAs were P adj <0.05 and —log2 fold change— >2.

Heat map and volcano map

The heat map and volcano map were drawn according to the normalized gene expression values of differentially expressed lncRNAs. The darker color represented higher expression level of differentially expressed lncRNA.

The model cohort

The clinical dataset of 629 CRC patients were downloaded from cBioPortal database (May 31, 2018, http://www.cbioportal.org/datasets). The patients with overall survival time less than one month were excluded from the present study to avoid the impact of unrelated causes of death (n = 49). The clinical dataset was matched to the gene expression dataset. The patients without gene expression information were excluded from the present study (n = 156). Finally 424 CRC patients were enrolled for further survival analysis. Figure 1 was the study flowchart for patient selection in the present study. The end-point in the present study was overall survival (OS). The average follow-up time was 30.0 ± 25.5 months. Overall survival time was calculated from initial diagnosis time to death time or the last follow-up time. The maximum value and the minimum value of overall survival time were 140.9 months and 1.0 month. The study time was from May 18, 2010 to January 27, 2015. The missing data in the present study were coded as “NA” in all tables and analyses.

Assessment of predictive performance

The Harrell’s concordance index (C-index), calibration plot and time-dependent receiver operating characteristic (ROC) curves were used to assess the predictive performance of prognostic signatures. The predictive performance of prognostic signature in the current study was compared with two previous prognostic signatures (named RS lncRNA score and Risk score). The formulas of the RS lncRNA score and the Risk score were as follows:

RS lncRNA score = (0. 1337 × expression value of RP1-170O19.17) + (0. 0633 × expression value of RP11-785D18.3) + (0. 8131 × expression value of RP11-798K3.2) + (−1.5194 × expression value of XXbac-B476C20.9) + (4. 3132 × expression value of RP11-481J13.1) + (0. 4497 × expression value of RP11-167H9.4) (Fan & Liu, 2018). In the formula of RS lncRNA score, the lncRNA expression values were normalized by “DEseq” packages.

Risk score =exp_LINC01555∗(−.191) + exp_{RP11−108K3.1}∗(0.318) + exp_LINC01207∗(−0.191) + exp_{RP11−610P16.1}∗(−0.163) (Zeng et al., 2017). In the formula of Risk score, the lncRNA expression values were log₂-transformed.

Internal validation by using bootstrap resampling method

The bootstrap resampling method has been recommended for internal validation of the predictive model (Blackstone, 2001; Grunkemeier & Wu, 2004). The validation cohort in the present study was constructed by drawing 424 CRC patients with replacements from the original model cohort.

Statistical analysis

The statistical analyses were carried out by using the R software (version 3.4.1) and SPSS Statistics 19.0 (SPSS Inc., Armonk, NY, USA). The following R packages were carried out as needed in the current study: “survival”, “rms” , “pROC”, “plyr”, “glmnet”, and “timeROC”. Continuous variables in the present study were presented as mean ± standard deviation. Continuous variables were compared by t-test or Mann–Whitney U test. Categorical variables were compared by chi-squared test or Fisher’s exact test. A 2-tail P < 0.05 was defined as statistically significant in the present study.

Study cohorts

The present study finally included 424 CRC patients with total lncRNA expression information and overall survival information. The average age was 66.7 ± 13.0 years and the average overall survival time was 30.0 ± 25.5 months in model cohort. There were 102 (24.0%) patients died during the follow-up period in model cohort, whereas there were 108 (25.5%) patients died during the follow-up period in validation cohort. The clinical characteristics of CRC patients in model cohort and validation cohort were summarized in Table 1. There were no significant differences in terms of clinical characteristics and lncRNA expression between model cohort and validation cohort. There were no missing data in terms of survival status, survival time and lncRNA expression value.

Differentially expressed analyses

We performed differential expression analyses by comparing all gene expression values between 458 tumor tissues and 41 normal colon tissues. The “edgeR” package identified 574 differentially expressed genes for overall survival. The heat map and volcano map of differential expression genes were presented in Figs. S1 and S2, respectively.

Construction of prognostic signature for overall survival

The univariate Cox regression analysis was performed to explore the potential prognostic lncRNAs for overall survival. The univariate Cox proportional regression analysis identified 33 potential lncRNA predictors for overall survival. Using multivariate Cox proportional regression analysis, a nine-lncRNA prognostic signature (Fig. 2) was constructed based on the potential prognostic lncRNA predictors determined by Cox regression analysis. The overall information of nine prognostic lncRNA predictors were summarized in Table 2. The formula of nine-lncRNA prognostic signature was as follows: nine-lncRNA prognostic signature score = (0.837* RP11_815M8.1) + (0.822* RP11_342A23.2) + (0.905* RP11_264B14.1) + (−0.529* AC064834.1) + (0.907* RP11_108K3.2) + (−0.745* LINC01571) + (1.241* RP11_383I23.2) + (−0.737* AC079612.1) + (0.725* AC005256.1).

Performance of nine-lncRNA prognostic signature in model cohort

The nine-lncRNA prognostic signature score were generated according to the previous formula. The distributions of nine-lncRNA prognostic signature score (Fig. 3A), overall survival status and overall survival time (Fig. 3B) in model cohort were shown in Fig. 3. In model cohort, the Harrell’s concordance-index (C-index) of nine-lncRNA prognostic signature was 0.757 (95% CI [0.706–0.808]).

Kaplan–Meier survival curves and log-rank test in model cohort

The median of nine-lncRNA prognostic signature score was used as cutoff value to stratify CRC patients into high risk group and low risk group. Kaplan–Meier survival curves (Fig. 4) and log-rank test were used to compare the difference of overall survival rate between high risk group and low risk group. As shown in Fig. 4, patients with high nine-lncRNA prognostic signature score had poorer overall survival rate than patients with low nine-lncRNA prognostic signature score (P < 0.001).

Time-dependent receiver operating characteristic curves in model cohort

We further explored the predictive accuracy of nine-lncRNA prognostic signature score compared with two previous prognostic signatures by using time-dependent receiver operating characteristic curves (Fig. 5). The C-indexes of nine-lncRNA prognostic signature, RS lncRNA score and Risk score were 0.768, 0.654 and 0.658 for 1-year overall survival (Fig. 5A) respectively, whereas it were 0.778, 0.666 and 0.582 for 3-year overall survival (Fig. 5B). For 5-year overall survival (Fig. 5C), the C-indexes of nine-lncRNA prognostic signature, RS lncRNA score and Risk score were 0.870,0.681 and 0.633, respectively.

Calibration curves in model cohort

The calibration curves were used to assess the predictive performance of nine-lncRNA prognostic signature. The calibration curves for 1-year (Fig. 5D), 3-year (Fig. 5E) and 5-year (Fig. 5F) overall survival demonstrated that there were a good agreement between the predictive probability and the actual overall survival in model cohort.

Internal validation of nine-lncRNA prognostic signature

The validation cohort (n = 424) was generated by random drawing from the model cohort with replacement method. The nine-lncRNA prognostic signature score were generated according to the previous formula for patients in validation cohort. The distributions of nine-lncRNA prognostic signature score (Fig. 6A), overall survival status and overall survival time (Fig. 6B) in validation cohort were shown in Fig. 6. The C-index of nine-lncRNA prognostic signature was 0.751 95% CI [0.700–0.802]) in validation cohort.

Kaplan–Meier survival curves and log-rank test in validation cohort

The previous cutoff value of nine-lncRNA prognostic signature score in model cohort was used as the cutoff value to stratify CRC patients into high risk group and low risk group in validation cohort. As shown in Fig. 7, the log-rank test indicated that the overall survival rate in high risk group was significantly lower than that in low risk group (P < 0.001).

Time-dependent receiver operating characteristic curves in validation cohort

In validation cohort, the C-indexes of nine-lncRNA prognostic signature, RS lncRNA score and Risk score were 0.761, 0.695 and 0.664 for 1-year overall survival (Fig. 8A) respectively, whereas it were 0.801, 0.660 and 0.582 for 3-year overall survival (Fig. 8B). For 5-year overall survival (Fig. 8C), the C-indexes of nine-lncRNA prognostic signature, RS lncRNA score and Risk score were 0.883,0.694 and 0.616 respectively.

Calibration curves in validation cohort

The calibration curves for 1-year (Fig. 8D), 3-year (Fig. 8E) and 5-year (Fig. 8F) overall survival indicated a good agreement between the predictive probability of overall survival and the actual overall survival in validation cohort.

Independence of nine-lncRNA prognostic signature for overall survival

We further carried out multivariate Cox regression analyses to explore whether nine-lncRNA prognostic signature was an independent influence factor for overall survival in model cohort and validation cohort. After adjustment of other clinical variables, including age, gender, the American Joint Committee on Cancer (AJCC) PT, AJCC PN, AJCC PM and AJCC stage (Table 3), the results of multivariate Cox regression analysis indicated that the nine-lncRNA prognostic signature was an independent influence factor for overall survival of CRC patients.

Functional enrichment analysis of prognostic signature

We performed functional enrichment analysis to explore the biological pathway and process correlated with this nine-lncRNA prognostic signature. At first, the pearson correlation coefficients between these prognostic lncRNA expression values and the mRNA expression values were calculated in the TCGA dataset. Then the genes correlated with at least one of these prognostic lncRNAs (defined as —Pearson correlation coefficient— >0.5) were included into the following functional enrichment analysis. The gene ontology (GO) biological process enrichment analysis and the Kyoto Encyclopedia of Genes and Genomes (KEGG) signaling pathway analysis were presented in Fig. 9 by using the above identified genes in the Database for Annotation, Visualization, and Integrated Discovery (DAVID) (https://david.ncifcrf.gov/) Bioinformatics Resources. The results of functional enrichment analysis demonstrated that the co-expressed genes were mainly enriched in G-protein coupled receptor signaling pathway, potassium ion transmembrane transport, carboxylic acid metabolic process response to hormone, regulation of ion transmembrane transport (Fig. 9).