Identification of a 6-gene signature predicting prognosis for colorectal cancer

RNA sequencing data from COAD and READ cohort consisted of 647 CRC and 51 normal samples obtained from TCGA data portal (https://​tcga-data.​nci.​nih.​gov/​docs/​publications/​tcga/​). TCGA-COAD cohort consisted of 480 COAD tissue samples and 41 adjacent normal colon tissue samples. TCGA-READ consisted of 167 READ tissue samples and 10 adjacent normal rectal tissue samples.

In addition, publicly available CRC information derived from 562 individuals with clinical follow up data were collected from the TCGA database. Among them, 419 were associated to COAD patients and 143 to READ ones. Data were downloaded from the TCGA database, thus, additional approval by an ethics committee was not required.

Raw data were normalized using the trimmed mean of M-values method [19], and DEMs in adjacent normal vs. COAD, and adjacent normal vs. READ samples were identified using the individual R package EdgeR (version 3.20.9) [20, 21]. DEMs were detected if ∣log2-fold change (FC)∣ > 1 and P value < 0.05. Volcano plots were created using the R package ggplot2 [22], and the hierarchical cluster analysis was conducted on the basis of the expression value of these DEMs using the pheatmap package (Version: 1.0.8, https://​cran.​r-project.​org/​web/​packages/​pheatmap/​index.​html) [23].

The scheme of our study is illustrated in Fig. 1. The intersection of up-regulated and down-regulated mRNAs was selected for further analysis. After that, a univariate Cox proportional hazards regression analysis was used to investigate the association between DEMs expression and OS in COAD/READ patients with the purpose of evaluating which mRNAs could be potentially used as prognostic indicators for COAD/READ. Subsequently, only the common DEMs in COAD and READ with a P-value < 0.05 and hazard ratio (HR) > 1 were considered as candidates and subjected to a step-wise multivariate Cox regression model to extract the predictive mRNA-based model with the best explanatory and informative efficacy. Next, an mRNA-based prognostic model was used to predict the risk score for each patient as follows:where “exp” represents the mRNA expression, and “β” is referred to the mRNA coefficient derived from the multivariate Cox regression analysis.

Based on the mRNA-based risk score equation, a risk score was obtained for each patient, and CRC patients in each cohort could be divided into high- or low-risk group using the median risk score as the threshold [24]. The receiver operating characteristic (ROC) curve was used to evaluate the sensitivity and specificity of the survival prediction according to the mRNA expression-based biomarker through analyzing the area under the curve (AUC) using the R package “survivalROC” [25]. The defining point set up by 3-year and 5-year time-dependent ROC curve analysis was employed to assess the predictive value of the risk score for time-dependent outcomes [25]. The Kaplan–Meier survival curve combined with a log-rank test was used to evaluate the differences in the patients’ survival time in the high- and low-risk group by the univariate analysis using the R package “survival”.

Univariate Cox regression model was used to evaluate the prognostic value of the gene signature and clinical variables (including age, gender, new tumor after initial treatment, history of colon polyps, residual tumor, pathologic stage metastasis (M)/node (N)/tumor (T), tumor site, and risk score) in their relationships with patients’ OS in the CRC entire cohort, COAD cohort, and READ cohort. Then multivariate Cox regression analysis was performed to investigate whether the predictive ability of the gene signature was independent of other clinical parameters, using OS as the dependent variable, mRNA risk score and other clinical characteristics as the explanatory variables.

To evaluate the prognostic performance of the gene signature in the cancer subgroups considered (CRC entire cohort, COAD cohort, and READ cohort), a stratified analysis was implemented according to clinical factors. The patients in each cohort were stratified into the two subgroups (for example, according to age, the patients were divided into ≤ 67 subgroup and > 67 subgroup), and then each subgroup was further classified into high- and low-risk group using the gene signature-based risk score. Stratification analysis was carried out using with univariate Cox regression model and the log rank test.

To evaluate the gene expression pattern in CRC, differential expression analysis was performed in COAD vs. adjacent normal samples, and READ vs. adjacent normal samples. When the criteria was set at P < 0.05 and |log FC| > 1, 2861 up-regulated mRNAs in COAD samples, 2944 up-regulated mRNAs in READ samples, and 2456 commonly up-regulated DEMs were found in these two cohorts. In addition, a total of 2480 down-regulated mRNAs in COAD samples and 2650 down-regulated mRNAs in READ samples were found and among them, 2271 DEMs were commonly down-regulated in the two cohorts. The volcano plot is referred to the DEMs of COAD and READ (Additional file 1: Figure S1). Hierarchical clustering results showed that COAD (Additional file 2: Figure S2) and READ (Additional file 3: Figure S3) were clearly distinguished from the adjacent normal tissue according to DEMs.

Subsequently, with the goal of extracting the predictive signature having the best explanatory and informative efficacy, the 14 candidate mRNA were subjected to the step-wise multivariate Cox’s model, resulting in a total of 6 mRNAs identified as survival predictors, such as EPH Receptor A6 (EPHA6), Tissue Inhibitor Of Metallopeptidase Inhibitor 1 (TIMP1), Iroquois Homeobox 6 (IRX6), ADP-Ribosyltransferase 5 (ART5), Histone Cluster 3 H2B Family Member B (HIST3H2BB), and Forkhead Box D1 (FOXD1). The information related to these 6 genes is listed in Table 1.

For each patient belonging to CRC, COAD, and READ, we computed a 6-gene expression-based survival score and we assigned these scores to the high- or low-risk group based on the median risk score that was used as the cutoff point. In the CRC cohort, 562 cases were classified into high- and low-risk group using the median risk score as the threshold (Fig. 2a). Figure 2b shows the expression pattern of the 6 selected mRNAs (EPHA6, TIMP1, IRX6, ART5, HIST3H2BB, and FOXD1) in the high- and low-risk group, with the blue color representing the low expression and the red representing the high expression. The mortality rate in the high-risk group was higher than that of the low-risk group, as shown in Fig. 2c. The risk score distribution of the 6-gene, their expression and the survival status of CRC patients are shown in Fig. 2a–c respectively. The Kaplan–Meier curve with the log-rank analysis showed that the survival rate of patients in the high-risk group was lower compared to that in the low-risk group (Fig. 2d, log-rank P = 2.58e−08). Moreover, univariate Cox’s regression model showed that patients in the high-risk group had a significantly lower survival rate compared to that in the low-risk group (Fig. 2d, Cox P = 1.36e−07). Thus, high risk score was a poor prognostic factor for CRC patients (HR = 3.08, 95% CI = 2.03–4.69). The 3-year and 5-year survival as predicted by the risk scores are shown in Fig. 2e, f, with an AUC of 0.711 and 0.683 respectively, implying that this 6-gene signature possessed a high specificity and sensitivity in the prediction of OS.

Based on the median risk score of the COAD cohort, 419 COAD patients were divided into high- and low-risk group (Fig. 3a). Figure 3b, c show the expression of the 6 mRNAs and the survival status of COAD patients, respectively. The Kaplan–Meier OS curve of the two groups showed that the patients in the high-risk group had worse prognosis than that in the low-risk group (Fig. 3d, log-rank P = 2.69e−06). The prognostic ability of the 6-gene signature was assessed by computing the AUC value of the ROC curve. Higher AUC corresponds to a better performance and the AUC for the 6-gene signature achieved 0.679 and 0.653 for the 3-year and 5-year survival, respectively (Fig. 3e, f), implicating the better performance of the 6-gene signature model in predicting COAD patient survival.

In the READ cohort, 143 patients were also classified into the high- and low-risk group according to the median risk score of the READ cohort. Figure 4a–c displays the risk score distribution, the expression of the 6 genes and survival status in the READ cohort. In line with the results in the CRC and COAD cohort, the patients in the high-risk group had a worse prognosis than that of the low-risk group (Fig. 4d, log-rank P = 1.49e−03). The ROC curve analysis achieved AUC values for the 3-year and 5-year survival of 0.845 and 0.74, respectively (Fig. 4e, f). These results confirmed that the 6-gene biomarker was able of predicting the prognosis of READ patients.

Figure 5 shows the expression patterns of all the 6 mRNAs in the three cohorts and two groups. From this figure, we found that ART5 FOXD1, HIST3H2BB, and TIMP1 expression in CRC, COAD, and READ was significantly higher than that in normal tissues (P < 0.05 or P < 0.001), while EPHA6 and IRX6 expression in CRC, COAD, and READ was lower than that in normal samples (P < 0.0001). The expression of all the 6 genes was significantly higher in the high-risk group compared to the low-risk groups in the three cohorts (P < 0.05 or P < 0001, Fig. 6).

As shown in Table 2, univariate Cox regression model demonstrated that the 6-gene signature risk score (HR = 3.08, 95% CI = 2.03–4.69, P = 1.36E−07 for CRC, and HR = 2.9, 95% CI = 1.82–4.62, P = 7.31E−06 for COAD), age (HR = 1.03, 95% CI = 1.01–1.04, P = 7.99E−04 for CRC, and HR = 1.02, 95% CI = 1.00–1.04, P = 4.20E−02 for COAD), new tumor after initial treatment (HR = 2.52, 95% CI = 1.70–3.74, P = 4.26E−06 for CRC, and HR = 2.53, 95% CI = 1.63–3.93, P = 3.86–05 for COAD), residual tumor (HR = 3.96, 95% CI = 2.26–6.96, P = 1.65E−06 for CRC, and HR = 3.81, 95% CI = 1.89–7.68, P = 1.81E−04 for COAD), pathologic stage (HR = 3.17, 95% CI = 2.09–4.79, P = 4.83E−08 for CRC, and HR = 3.07, 95% CI = 1.95–4.85, P = 1.41E−06 for COAD), stage M (HR = 4.51, 95% CI = 2.95–6.89, P = 3.92E−12 for CRC, and HR = 4.69, 95% CI = 2.87–7.66, P = 6.55E−10 for COAD), stage N (HR = 2.90, 95% CI = 1.96–4.30, P = 1.04E−07 for CRC, and HR = 2.86, 95% CI = 1.85–4.42, P = 2.21E−06 for COAD), and stage T (HR = 2.18, 95% CI = 1.13–4.18, P = 1.94E−02 for CRC, and HR = 2.87, 95% CI = 1.25–6.59, P = 1.30E−02 for COAD) were significantly related to the patients’ OS in the CRC entire cohort and COAD cohort, but other factors did not exhibit any significant correlation with OS. To further investigate whether the prognostic performance of the 6-gene signature was independent of clinical factors of CRC and COAD cases, the multivariate Cox regression analysis was performed based on the 6-gene biomarker and other clinical parameters as explanatory variables and OS as the dependent variable. As shown in Table 3, the results of multivariate Cox regression model suggested that the 6-gene signature still remained an independent factor of OS after adjustment for clinical factors, including age (HR = 1.03, 95% CI = 1.02–1.05, P-value = 2.19E−04 for CRC, and HR = 1.02, 95% CI = 1.01–1.04, P = 8.93E−03 for COAD), pathologic stage (HR = 3.26, 95% CI = 2.14–4.98, P-value = 4.45E−08 for CRC, and HR = 3.15, 95% CI = 1.96–5.05, P = 1.93E−06 for COAD), and risk score (HR = 2.37, 95% CI = 1.53–3.68, P-value = 1.07E−04 for CRC, and HR = 2.28, 95% CI = 1.4–3.71, P = 9.44E−04 for COAD). Similar results were obtained from the READ cohort. Univariate Cox regression model suggested that the 6-gene signature risk score (HR = 4.34, 95% CI = 1.62–11.64, P = 3.57E−03), age (HR = 1.09, 95% CI = 1.04–1.14, P = 2.68E−04), new tumor after initial treatment (HR = 2.53, 95% CI = 1.04–6.15, P = 4.03E−02), residual tumor (HR = 3.38, 95% CI = 1.23–9.28, P = 1.79E−02), pathologic stage (HR = 3.72, 95% CI = 1.35–10.26, P = 1.10E−02), stage M (HR = 4.05, 95% CI = 1.70–9.68, P = 1.62E−03), and stage N (HR = 3.24, 95% CI = 1.27–8.29, P = 1.39E−02) were significantly associated with the patients’ OS in READ cohort, but other factors did not exhibit any significant correlation with OS (Table 2). The multivariate Cox regression model implicated that the 6-gene signature was an independent factor of prognosis after adjusting for other clinical factors, including age (HR = 1.08, 95% CI = 1.03–1.14, P = 3.03E−03) and risk score (HR = 2.97, 95% CI = 1.07–8.22, P-value = 3.67E−02).

In summary, the 6-gene risk score was an independent adverse prognostic factor for the three cohorts.

With the goal of evaluating the prognostic performance of the 6-gene signature, the patients in each cohort were firstly stratified into two subgroups based on clinical parameters (such as age (≤ 67/> 67), gender (Female/Male), and stage (I–II/III–IV)), and then each subgroup was further classified into high- and low-risk group using the 6-gene signature. In all the subgroups of the CRC entire cohort, patients in the high-risk group had a significantly shorter survival time than that in the low-risk group (Fig. 7, P < 0.05), suggesting that the 6-gene risk score was an adverse prognostic factor in CRC. In the subgroups of the COAD cohort (except the male subgroup), patients in the high-risk group had also a significantly poorer prognosis compared to that of the patients in the low-risk group (Fig. 7, P < 0.05), demonstrating that the 6-gene risk score can predict the survival status in COAD patients. In the READ cohort, except the subgroup of stage I–II, patients of the other subgroups in the high-risk group had also a significantly poorer prognosis compared to that of the patients in the low-risk group (Fig. 7, P < 0.05), implying that the 6-gene risk score was an adverse prognostic indicator able to predict the survival status in READ patients. Combining all these results, the 6-gene signature was an independent predictor of other clinical factors for predicting survival in CRC patients.