DNA methylation profile in CpG-depleted regions uncovers a high-risk subtype of early-stage colorectal cancer

Abstract

Background: The current risk stratification system defined by clinicopathological features does not identify the risk of recurrence in early-stage (stage I-II) colorectal cancer (CRC) with sufficient accuracy. We aimed to investigate whether DNA methylation could serve as a novel biomarker for predicting prognosis in early-stage CRC patients.

Methods: We analyzed the genome-wide methylation status of CpG loci using Infinium MethylationEPIC array run on primary tumor tissues and normal mucosa of early-stage CRC patients to identify potential methylation markers for prognosis. The machine-learning approach was applied to construct a DNA methylation-based prognostic classifier for early-stage CRC (MePEC) using the 4 gene methylation markers FAT3, KAZN, TLE4, and DUSP3. The prognostic value of the classifier was evaluated in 2 independent cohorts (n = 438 and 359, respectively).

Results: The comprehensive analysis identified an epigenetic subtype with high risk of recurrence based on a group of CpG loci in the CpG-depleted region. In multivariable analysis, the MePEC classifier was independently and statistically significantly associated with time to recurrence in validation cohort 1 (hazard ratio = 2.35, 95% confidence interval = 1.47 to 3.76, P < .001) and cohort 2 (hazard ratio = 3.20, 95% confidence interval = 1.92 to 5.33, P < .001). All results were further confirmed after each cohort was stratified by clinicopathological variables and molecular subtypes.

Conclusions: We demonstrated the prognostic statistical significance of a DNA methylation profile in the CpG-depleted region, which may serve as a valuable source for tumor biomarkers. MePEC could identify an epigenetic subtype with high risk of recurrence and improve the prognostic accuracy of current clinical variables in early-stage CRC.

Figure 1.

DNA methylation signature in CpG-depleted regions related to colorectal cancer (CRC) recurrence. A) DNA methylation profiles of recurrence-specific differentially methylated probes (DMPs), with most variable DNA methylation values among 21 recurrent and 24 recurrence-free CRC patients in the discovery cohort. The DNA methylation β values are represented by using a color scale from dark blue (low DNA methylation) to yellow (high DNA methylation). Two subgroups were derived by clustering analysis and are indicated above the heatmap: low-methylation (blue, n = 13) and high-methylation (yellow, n = 32) groups. The presence of KRAS mutations (blank = wild type), TP53 mutations (blank = wild type), high-frequency microsatellite instability (MSI-H; blank = microsatellite stability or low-frequency microsatellite instability [MSI-L]), CpG island methylator phenotype (CIMP)-positive (blank = CIMP-negative), and clinical variables are indicated by a colored block. All the tumors were BRAF wild type. Probes targeted to CpG islands, CpG open seas, gene promoters, and gene bodies are indicated by the vertical color bars to the left of the heatmap. Each column and row represents 1 patient and probe, respectively. The probes were arranged based on the order of unsupervised hierarchal cluster analysis using a correlation distance metric and average linkage method. See also Supplementary Figure 4 (available online) for clustering analysis using cancer-specific DMPs among 21 recurrent and 24 recurrence-free CRC patients. B) Time to recurrence in the high- and low-methylation groups based on recurrence-specific DMPs. See also Supplementary Figure 4 (available online) for survival analysis using cancer-specific DMPs. C) Clustering analysis of DNA methylation using the recurrence-specific DMPs located in CpG-depleted regions (left) and CpG-rich regions (right). High met = high methylation; HR = hazard ratio; Low met = low methylation.

Figure 2.

Development and validation of methylation-based prognostic classifier for early-stage colorectal cancer (CRC). A) The 4 candidate probes in the machine-learning model were arranged based on the unsupervised hierarchal cluster analysis. Each column represents 1 patient, red indicates high methylation levels, and blue indicates low levels. B) Time-dependent receiver operating characteristic (ROC) at 5 years are shown for the 4 methylation-based markers in the discovery cohort. The 4 probes targeted to CpGs in the bodies of KAZN, FAT3, DUSP3, and TLE4 showed high prognostic accuracy in the discovery cohort with areas under the curve (AUCs) ranging from 0.769 to 0.785. The methylation-based prognostic classifier for early-stage CRC (MePEC) classifier that combined these 4 markers demonstrated the best prognostic accuracy with an AUC of 0.876. C-D) Long-term outcome of MePEC classifier in the National Basic Research Program of Evolution from Precancerous Disease to Cancer in China (NEPDC) cohort (C) and Colon Cancer Family Registry (CCFR) cohort (D). The patients were divided into MePEC-high and MePEC-low groups based on the prespecified cutoff value of 3. Data are presented as hazard ratio (95% confidence interval). Log-rank test P value is given in each plot.

Figure 3.

Hazard ratios (HRs) of methylation-based prognostic classifier for early-stage colorectal cancer (MePEC) classifier for recurrence in 2 validation cohorts. Multivariable Cox regression analysis was performed for MePEC classifier and clinical variables in time to recurrence prediction. The prognostic value of each variable was graphed as a forest plot. Data are presented as hazard ratio. The error bars indicate the 95% confidence intervals. CCFR = Colon Cancer Family Registry; CIMP = CpG island methylator phenotype; MSI-H = high-frequency microsatellite instability; MSI-L = low-frequency microsatellite instability; MSS = microsatellite stability; NEPDC = National Basic Research Program of Evolution from Precancerous Disease to Cancer in China.

Figure 4.

Prognostic value of methylation-based prognostic classifier for early-stage colorectal cancer (MePEC) classifier in the subgroups stratified by clinicopathological variables. The multivariable Cox model, including MePEC classifier, age, sex, tumor stage, tumor differentiation, tumor location, lymph nodes removed, mutation status of KRAS and BRAF, microsatellite status, and CpG island methylator phenotype (CIMP) status, was generated in each subgroup. The hazard ratio (HR) with 95% confidence interval (CI) of MePEC classifier in each subgroup was graphed as a forest plot including interaction P values. The error bars indicate the 95% confidence intervals. TTR = time to recurrence.