Colorectal cancer (CRC) is a frequently lethal disease with high incidence. In most cases, CRC usually starts as a polyp (a noncancerous growth that develops in the lining of the colon and rectum), and does not have obvious symptoms until it becomes difficult to cure [1]. Therefore, CRC can largely be prevented by the early detection and removal of precursor lesions [2]. There are two main types of CRC screening strategies: stool tests (such as fecal occult blood testing) and structural exams, including colonoscopy, double-contrast barium enema, and computed tomographic colonography, etc. [3]. However, the effects of these tests are not satisfactory in clinical applications because of their complex protocols and limited sensitivity and specificity [4]. Since the occurrence of CRC is driven by the accumulation of genetic abnormalities, there has been an increasing number of gene abnormality-based technologies available for CRC screening in the last decade [5]. Even though several tests have already been used in clinical practice, current options still do not have enough sensitivity and specificity to serve as general screening [6].

The results of CRC epigenome assessment reveal that almost all CRCs have aberrantly methylated genes, which play pathological roles in CRC development [7]. Meanwhile, the heterogeneity of methylation characteristics among individuals is closely associated with the prognosis [8]. There are two types of methylation associated with CRC progression: age-related methylation (type A), and cancer-specific methylation (type C) [9]. Among the type C methylation, DNA hypermethylation of CpG-rich promoters, which result in switching off tumor suppressor genes, has been recognized as a subgroup of CRCs. CpG island methylator phenotype (CIMP) defines the overall methylation-mediated gene expression pattern in a sample by the methylation status of specific gene promoters [10]. Some studies proposed CIMP status as the most promising predictor of all CRC biomarker candidates [11], however, this view is still controversial and needs more research to provide rigorous evidence.

In this study, we performed an integrated analysis of DNA methylation and gene expression profiles of CRC. The data downloaded from The Cancer Genome Atlas (TCGA) were used as a training set, and that from Gene Expression Omnibus (GEO) datasets were used as a testing set. The differential expression genes (DEGs) and methylation CpG sites (dmCpGs) in CRC samples were identified, and a logistic regression model was constructed based on the hypermethylated CpG sites which were located in downregulated genes for CRC diagnosis. Finally, the prediction accuracy of the constructed model was evaluated. We believe that these results can contribute to research on the screening of early diagnostic markers for CRC.

DNA methylation and gene expression profiles were first processed, including the removal of CpG sites and genes with missing values in more than 10% of samples, and then the remaining missing values were added through the R Bioconductor impute package (https://bioconductor.org/packages/release/bioc/html/impute.html). We used paired t-test to screen differential methylation CpG sites (DMCs) between the 44 pairs of CRC and adjacent samples with the thresholds of absolute β (methylation level) difference >0.2 and FDR adjusted p-value < 0.05. Differential expression genes (DEGs) between paired CRC and adjacent samples were measured through the edgeR Bioconductor package [14] based on the raw count data. Genes with absolute log2 (fold change) > 1 and FDR adjusted p-value < 0.05 were determined as differentially expressed.

To screen reliable CRC diagnostic biomarkers, we selected hyper-methylated DMCs in CRC samples and filtered out those not in promoters and with β > 0.05 in adjacent samples; the remaining DMCs are hereafter referred to as ProHyperDMCs. Hyper-methylation in promoters were usually associated with repressed gene expression, so we further selected CpGs from ProHyperDMCs that were located in downregulated genes in CRC samples as promising CRC diagnostic biomarkers. A logistic regression model was finally constructed using sample type, i.e., CRC or normal, as response variables and CpGs’ β values were used as predict variables in the training set.

The vast majority of human cancer cells harbor both genetic and epigenetic abnormalities, which allow them to escape from chemotherapy and host immune surveillance [17]. Hence, a growing number of efforts on the analysis of high-throughput sequencing-based epigenome data, including DNA methylation and histone modifications, has been advanced for the need of individualized therapies [18]. In addition, methylation characteristics were also closely related to the prognosis of CRC patients [19]. For example, UHRF1, FOXE1, AXIN2, and DKK1 have recently been defined as biomarkers that support oncogenic properties, and high expressions of these genes predict reduced CRC patient survival [20–22].

CIMP status was first found in CRC, and this subtype had distinct histological and molecular features [23]. In this study, we first clustered the CRC samples based on methylation of CpG sites, and identified the patients with CIMP. The OS analysis revealed that the CIMP status was significantly associated with the prognosis of CRC patients (Figure 1), which was consistent with the literature report. Furthermore, we performed a cross-analysis between differential methylation sites and differential genes, and identified cg07945335 and cg00321288 as the key genes for CRC diagnostic model construction, which were located in the promoter of CD300LG and MGAT4C, respectively.

CD300LG protein, a member of the CD300 family, is a type I cell surface glycoprotein is that exclusively expressed in the capillary endothelium [24]. CD300LG mediates molecular traffic across the capillary endothelium, responds to the immunological environment, and is implicated in lymphocyte binding and transmigration [24, 25]. Herein, we reported on the important role of CD300LG in the CRC process for the first time, since leukocyte diapedesis through the vasculature involves critical adhesive interactions with endothelial cells, and both leukocytes and cancer cells express similar surface receptors capable of binding endothelial adhesion molecules [26]. Therefore, we speculated that CD300LG probably affected transendothelial migration of CRC cancer cells by regulating the response of cancer cells to the immune microenvironment, which will be confirmed in our future studies. Mannosyl (alpha-1,3-)-glycoprotein beta-1,4-N-acetylglucosaminyltransferase (MGAT4C), is a member of the MGAT4 family [27]. Demichelis F et al. investigated the possible function of MGAT4C in prostate cancer through gene overexpression and knockdown experiments [28]. The results revealed that MGAT4C expression was related to the proliferation and migration of prostate cancer cells. However, the function of this MGAT4C in CRC still needs more exploration.

We constructed a CRC diagnostic model based on cg07945335 and cg00321288, and used GEO data as a validation set to determine the specificity and sensitivity of these two key genes as diagnostic biomarkers, and the results indicated the good performance of the diagnostic model in CRC.

In conclusion, this study identified promising DNA methylation biomarkers for CRC diagnosis through an integrative analysis of DNA methylation and gene expression data. Nevertheless, there are also some limitations in this study. First, the expressions of these biomarkers have not been verified by clinical samples, and the biological function of them is not clear. Second, since the occurrence and development of CRC are related to some high risk factors such as age, the inclusion of other clinical factors and the expansion of the sample size will help to improve the accuracy of the model.