Glycosyltransferase Gene Expression Identifies a Poor Prognostic Colorectal Cancer Subtype Associated with Mismatch Repair Deficiency and Incomplete Glycan Synthesis

Despite major advances in diagnosis and treatment, colorectal cancer remains one of the leading causes of cancer-related death worldwide (1, 2). Colorectal cancer is commonly grouped into two categories: tumors with microsatellite instability (MSI), caused by defective function of the DNA mismatch repair (MMR) system, and tumors that are microsatellite stable but exhibiting chromosomal instability (CIN; refs. 3–5). The majority of colorectal cancer (∼85%) follows the CIN pathway, often accompanied by KRAS mutations and TP53 inactivation. Approximately 15% of colorectal cancers that exhibit deficient MMR (dMMR) frequently carry BRAF mutations (3, 5). Clinical trials implicated MMR status as a potential therapeutic classifier for stage II patients in the adjuvant setting (6–8). In the metastatic setting, KRAS and BRAF mutations are used for predicting unresponsiveness to EGFR-targeted therapies (4). Despite increasing knowledge, clinicopathologic staging system remains the only prognostic classification currently used in clinical practice. However, clinicopathologically similar tumors can strikingly differ in clinical behaviors that likely reflect the molecular heterogeneity. Although it is recommended that stage III patients receive postoperative chemotherapy, approximately 30%–40% of patients develop recurrence even after standard treatment (9–12).

Glycosylation is a common posttranslational modification that involves sequential addition of single sugar residues to target structures, resulting in glycan elongation. Further chemical modifications and branching can finally form a vast array of glycan structures (13). Those procedures are regulated by the multienzymatic reaction of glycosyltransferases, whose encoding genes, namely “glycogenes”, are equivalent to 1% of human genome. Cell surface glycans undergo changes during malignant transformation and tumor progression accompanied by distinct biological functions and unique tumor phenotypes, thereby making glycans as potential cancer biomarkers (13, 14). For instance, a cancer-associated glycan epitope, CA19-9, called sialyl Lewis A (sLe^a), is routinely utilized as a serum tumor marker (15). CA19-9 and several other cancer-associated glycans, including sialyl Lewis X (sLe^x), sialyl Tn, Tn, and T antigens, are associated with tumorigenesis and poor prognosis of colorectal cancer (13, 16). Such glycans can be attributed to transcriptional dysregulation of glycosyltransferases that has been postulated as two principal mechanisms, “incomplete synthesis” and “neosynthesis” (13, 16–18). Some glycogenes are repressed by epigenetic silencing during early stages of tumorigenesis, which lead to the biosynthesis of truncated structures, such as Tn and STn expression, called incomplete synthesis. Conversely, in the neosynthesis process, transcriptionally induced glycogenes can result in the de novo expression of cancer antigens, such as sLe^a and sLe^x.

In this study, with the aim to discover distinct classes of colorectal cancer on the basis of the expression of glycosyltransferases, we compiled an extensive number of transcriptomic profiles obtained from multiple cohorts by integrating other available data sources, including mutations, MMR status, methylation, protein expression as well as nontumor, precancerous, preinvasive, and cancerous samples. We initially described a novel subtype based upon clustering analysis of genome-wide “glycogene” expression patterns, and this led us to identify a glycogene, GALNT6 as a promising biomarker for disease prognosis. Moreover, we found the functional characteristics of GALNT6 involved in tumor progression and glycosylation, suggesting the contribution of epigenetic silencing of GALNT6 to colorectal carcinogenesis, through the incomplete synthesis of cell surface glycans.

All microarray and methylation array data are publicly available in the Gene Expression Omnibus (GEO) database (http://www.ncbi.nlm.nih.gov/geo) as shown in Supplementary Table S1. We utilized the normalized expression values obtained from each dataset. If a gene is represented by multiple probes, they were averaged. To generate a list of glycogenes, official gene symbols and Entrez Gene IDs for 190 glycogenes were obtained from GGDB (GlycoGene DataBase; http://acgg.asia/ggdb2/). Among 190 glycogenes, 185 unique genes were converted to Affymetrix_3PRIME_IVT_ID using DAVID Bioinfomatics Resources6.7 (http://david.abcc.ncifcrf.gov/home.jsp) as shown in Supplementary Table S2.

Hierarchical clustering was initially performed using an Affymetrix dataset, GSE17536, consisted of 177 colorectal cancer patients with survival information. Expression levels of 185 glycogenes were median-centered, and then genes and samples were subjected to an unsupervised clustering by the centroid linkage method using the Cluster3.0 and the Java Treeview program (19). Among 39 differentially expressed genes between two major clusters (Cluster A vs. B, P < 0.001 by t test), 15 genes exhibited significant differential expression between the subcluster (Cluster A1) and the remaining subclusters (Clusters A2, B1, and B2) with stringent P values at <0.0001 (Supplementary Tables S3 and S4). We then obtained two Affymetrix datasets, and 121 stage I–III patients in GSE41258 and 89 stage II patients in GSE33113 with available survival information were used for hierarchical clustering. On the basis of three independent clustering analyses, 3 glycogenes that were consistently upregulated or downregulated between clusters with log₂ fold-change > 0.4 were identified.

Level 3 Illumina RNA-Seq data for colon and rectal adenocarcinoma (COADREAD) were downloaded through cBioPortal (http://www.cbioportal.org/; ref. 20). Clinicopathologic and molecular features were obtained from the TCGA data portal (http://tcga-data.nci.nih.gov/) in June 2015 (3). We utilized two different versions of RNA-Seq data normalized either by RPKM or RSEM methods. These two TCGA datasets, namely, RNA-Seq RPKM and RNA-Seq V2 RSEM, contained 193 and 361 colorectal cancer samples, respectively, after removing three redundant samples from the latter dataset. Hierarchical clustering based on the mRNA expression Z-scores for the 15 glycogenes was applied to each TCGA datasets as described above. For the analysis of GALNT6, both mRNA expression Z-scores by RNA-Seq V2 RSEM and DNA methylation β-values by Illumina Infinium HumanMethylation450 for 357 samples with available MMR status were also downloaded from cBioPortal.

To analyze the relationship between glycogenes and molecular features, 9 additional datasets were downloaded from GEO, including GSE39582, GSE39084, GSE42284, GSE75315, GSE26682, GSE13294, GSE4554, GSE13067, and GSE18088 (Supplementary Table S1). They were discovered by carefully searching the GEO database according to the availability of more than 10 dMMR samples in each dataset. We also used an Illumina microarray dataset GSE59857, in which mutational and transcriptional profiles of 151 colorectal cancer cell lines were available (Supplementary Table S1).

We obtained formalin-fixed paraffin-embedded (FFPE) specimens of endoscopically resected colorectal adenomas from 40 patients and surgically resected colorectal adenomas from 20 patients treated at Fukushima Medical University Hospital (Fukushima City, Japan). We also obtained 8 endoscopically resected specimens that were pathologically diagnosed as carcinoma in adenoma. In addition, transcriptomic and epigenomic data from a total of 345 colon adenoma samples with 213 normal colon and 570 carcinoma samples were analyzed. Briefly, we obtained datasets of colon biopsy specimens from normal colon, adenoma, and carcinoma (GSE4183, GSE77953, GSE37364, GSE20916, GSE41657, and GSE71187) and four additional datasets (GSE45270, GSE79460, GSE4045, and GSE36758) of conventional tubular adenomas/adenocarcinomas and serrated adenomas/adenocarcinomas (Supplementary Table S1). Also, we utilized epigenome-wide data based on Illumina Infinium HumanMethylation450 BeadChip platform for normal colon, adenoma and cancer tissues (GSE48684 and GSE77954), and 9 dMMR and 34 pMMR colorectal cancer samples (GSE68060; Supplementary Table S1). In those analyses, methylation levels were reported as β-values or M-values, and we examined probe cg19265103 located in the GALNT6 promotor region, as it was utilized in the cBioPortal as described earlier.

We enrolled 368 consecutive patients with primary colorectal cancer, who underwent surgery between 1990 and 2010 in Fukushima Medical University Hospital (Fukushima City, Japan). Tumors were classified according to the TNM classification of malignant tumors (21). After exclusion of patients who received preoperative chemotherapy or radiotherapy, 335 stage 0 to IV patients with available FFPE tumor sections were used. Adjacent normal mucosae from 304 sections were also available for evaluation. Clinical information was retrospectively obtained by reviewing medical records, with the last follow-up in February 2016. For survival analysis, 17 patients with stage 0 tumors (carcinoma in situ) were omitted, and 267 stage I to IV patients who underwent curative resection (R0), with survival information, were utilized. We analyzed disease-specific survival (DSS), disease-free survival (DFS), and overall survival (OS), which were defined as time from the date of surgery to the date of disease recurrence, cancer-related death, and death from any cause, respectively. The study was conducted in accordance with the Declaration of Helsinki and was approved by the Institutional Review Board of Fukushima Medical University.

IHC was performed as described previously (22), with primary rabbit polyclonal anti-GALNT6 antibody (HPA011762, Prestige Antibodies Powered by Atlas Antibodies, Sigma-Aldrich, Co. LLC), identified using the Human Protein Atlas database (www.proteinatlas.org; ref. 23). Briefly, antigens were retrieved by autoclave, and anti-GALNT6 antibody was incubated in a 1:500 dilution at 4°C overnight, and subsequently detected by a horseradish peroxidase (HRP)-coupled anti-rabbit polymer followed by incubation with diaminobenzidine (EnVision+ System, Dako). IHC slides were evaluated by two independent observers without knowledge of patients' clinical information. Several adenocarcinoma specimens from lung (24), pancreas (25), breast (26), and stomach (27) were used as positive controls. Each sections were considered positive for GALNT6 staining when more than 10% of tumor cells were stained in the cytoplasm according to the procedure as described previously (24, 25).

IHC for MMR protein was performed as described elsewhere (28), with primary antibodies against MLH1 (ES05, 1:50, Dako), MSH2 (FE11, 1:50, Dako), MSH6 (EP49, 1:200, Dako), and PMS2 (EP51, 1:50, Dako). Loss of a MMR protein was defined as the absence of nuclear staining of tumor cells in the presence of positive nuclear staining in normal colonic epithelium and lymphocytes (6).

In the expression datasets, MSI testing data (MSI-H, MSI-L and MSS) were obtained through the GEO or the TCGA data portal. Tumors demonstrating MSI-H or loss of at least one MMR protein were collectively designated as dMMR, and tumors with MSS/MSI-L or intact MMR protein expression as proficient MMR (pMMR).

Stage II and III colorectal cancer samples from three independent datasets were aggregated as an independent validation meta-cohort, herein termed microarray validation cohort (n = 364). We utilized GSE37892, GSE24551 and GSE38832, because they were not used in the previous GALNT6 analyses of this study and had enough number of stage II and III colorectal cancer samples with available DFS information. On the basis of the fact that 14.6% of colorectal cancer showed loss of GALNT6 protein by IHC, we simply used the same percentile as cutoff, namely patients in the lowest 14.6th percentile of GALNT6 mRNA expression were defined as GALNT6-low and the remaining patients were considered GALNT6-high within each dataset.

Short tandem repeat (STR)-authenticated colorectal cancer cell lines, including SW480, SW620, and RKO, were purchased from ATCC. SW837 and HCT116 were obtained from JCRB Cell Bank and RIKEN Cell Bank, respectively. HCT15, SW48, LS180, and Colo205 were previously obtained and authenticated by STR analysis (Promega). RKO and LS180 cells were maintained with DMEM; others with RPMI1640 containing 10% FBS and penicillin/streptomycin (Thermo Fisher Scientific) at 37°C in a humidified atmosphere of 5% CO₂. A demethylation reagent, 5-aza-2′-deoxycytidine (5-aza-dC; Sigma-Aldrich) was dissolved in DMSO at 10 mmol/L and stored in aliquots at −80°C until use.

Knockdown experiments were conducted using siRNA oligonucleotides of GALNT6 or scramble control with Lipofectamine RNAiMAX Reagent, according to manufacturer's instructions (Ambion Silencer Select; s22154, s22155, and negative control #1, Thermo Fisher Scientific).

Total RNA was extracted using TRIzol Reagent, and 1 μg of total RNA was reverse transcribed to cDNA using the SuperScript III First-Strand Synthesis System (Thermo Fisher Scientific) according to the manufacturer's instructions. qRT-PCR was carried out using TaqMan Gene Expression Master Mix on the 7500 Real-Time PCR system in triplicate with TaqMan assays, including GALNT6 (Hs00926629_m1), MLH1 (Hs00179866_m1), and ACTB (Hs99999903_m1; Thermo Fisher Scientific). Relative expression levels were determined with SDS software by the 2^−ΔΔC_t method as described by the manufacturer.

Total protein was extracted using RIPA lysis buffer supplemented with Halt Protease Inhibitor Cocktail, and were boiled in Tris-Glycine SDS Sample Buffer (Thermo Fisher Scientific). Equal amount of protein was loaded and separated by 10% SDS-PAGE gel, and then transferred onto polyvinylidene difluoride membranes (Thermo Fisher Scientific). The membrane was blocked with 5% nonfat dried skimmed milk powder (Cell Signaling Technology), and incubated with primary rabbit anti-GALNT6 (#HPA011762, 1:250, Atlas Antibodies) or mouse anti-β-actin (#SC-69879, 1:2,000, Santa Cruz Biotechnology). The membrane was incubated with goat anti-rabbit or anti-mouse HRP secondary antibody (Santa Cruz Biotechnology), and developed with the SuperSignal West Pico Chemiluminescent Substrate (Thermo Fisher Scientific) using LAS4000 imager (GE Healthcare).

Cell suspensions were incubated with mouse monoclonal anti-Tn antibody (MLS128, 1:100, Wako), followed by staining with goat anti-mouse IgG H&L (Alexa Fluor 488; ab150113, 1:2,000; Abcam). The data were acquired on a FACSCanto II (Becton Dickinson) and analyzed with FlowJo software (TOMY Digital Biology).

Cell proliferation was measured using the Cell Counting Kit-8 (CCK-8, DOJINDO) according to the manufacturer's instructions. Cytotoxicity was assessed by CCK-8 assay using a series of 5-FU (Sigma-Aldrich) concentrations. We preliminarily applied a series of 5-FU concentrations ranging from 0.1 to 1,000 μg/mL or vehicle alone for generating dose–response curves. We then used 1, 5, 10, 50, and 100 μg/mL of 5-FU for experiments. Apoptotic cells were detected using the Annexin V-PE/7-AAD Apoptosis Detection Kit (BD Biosciences) according to the manufacturer's protocol. Annexin V–positive cells were regarded as apoptotic cells.

For wound-healing assay, cells were seeded on a 6-well plate and allowed to reach confluency. After scratching the bottom of the well with a pipette tip, the monolayer of cells was washed, and the wound closure photographs were captured at 0, 6, 12, 18, and 24 hours using a phase-contrast microscope. The percent of wound closure was calculated as the cell migration distance to the initial wound distance. Invasion assay was performed using Corning BioCoat 24-Multiwell Tumor Cell Invasion Systems (Corning) according to the manufacturer's protocol. Fluorescence of invaded cells labeled with Calcein-AM (Corning) was measured using SkanIt RE for Varioskan Flash 2.4 (Thermo Fisher Scientific).

The lectin microarray was performed essentially as described elsewhere (29). Briefly, the membrane fractions of cultured cells were obtained using the ProteoExtract Subcellular Proteome Extraction kit (Merck Millipore) and the total protein content was determined using the Micro BCA Protein Assay kit (Thermo Fisher Scientific), and then Cy3-labeled proteins with Cy3 monoreactive dye pack (GE Healthcare Life Science) were analyzed on a lectin microarray glass slide (LecChip ver 1.0; GlycoTechnica). Fluorescent images were acquired using an evanescent-field fluorescence scanner (GlycoStation Reader 1200; GlycoTechnica). The raw fluorescence intensity was first processed with the gain-merging procedure, followed by average normalization (29). Data were analyzed with GlycoStation Tools Pro Suite1.5 (GlycoTechnica).

Fisher exact test, χ² test, unpaired t test, and Mann–Whitney U test were used to determine differences between two variables. Spearman correlation was used to evaluate the correlations between levels of expression and methylation. Cumulative survival was estimated by the Kaplan–Meier method, and differences between the two groups were analyzed by log-rank test. Univariate and multivariate models were computed using Cox proportional hazards regression. All statistical analyses were two-sided and were conducted using GraphPad Prism v6.0 (Graphpad Software Inc.) and SPSS Statistics version 24 (IBM Corporation). All P values were two-sided, and P values less than 0.05 were considered statistically significant.

The overall study design is demonstrated in Supplementary Fig. S1. We initially conducted an unsupervised hierarchical clustering analysis in 177 patients from GSE17536 using the 185 glycogenes (Supplementary Table S2; Fig. 1A; Supplementary Fig. S2A), resulting in two major clusters (Cluster A and B) and four subclusters (Cluster A1, A2, B1, and B2; Supplementary Table S5). Cluster A showed a clear tendency to be associated with worse clinical outcomes (Supplementary Fig. S3A and S3B). Moreover, patients segregating to Cluster A1 had significant poor DSS and DFS compared with the remaining subclusters (Fig. 1B–E; Supplementary Fig. S3C and S3D). Multivariate Cox analysis demonstrated that Cluster A1 was significantly associated with DSS [HR, 5.71; 95% confidence interval (CI), 2.69–12.13; P = 6.0E−06], DFS (HR, 2.83; 95% CI, 1.38–5.80; P = 0.005), and OS (HR, 3.71; 95% CI, 2.02–6.82; P = 2.5E−05; Supplementary Table S6). Cluster A1 was also associated with poorly differentiated histology (Supplementary Table S5).

As Cluster A and its subcluster Cluster A1 exhibited worse survival outcome, we next sought to identify a minimum set of genes whose expression was closely related to these poor prognosis subgroups. Thirty-nine differentially expressed genes between Cluster A and B were further narrowed down to 15 genes (GCNT3, FUT8, B3GAT2, GALNT6, POFUT1, GALNT1, B3GNT8, DPM1, HS3ST3B1, SLC35A1, MGAT2, GALNT5, GYLTL1B, MGAT5, and HS3ST1, designated the 15-glycogene signature) that were significantly altered between Cluster A1 and the remaining subclusters (Supplementary Tables S3 and S4).

To test the hypothesis that the 15-glycogene signature can discriminate prognostic subgroups, independent datasets were utilized. Clustering analysis showed that 121 patients with stage I to III diseases from GSE41258 were clearly separated into two clusters, designated as 15-Glycogene Cluster A and 15-Glycogene Cluster B, with significant DFS difference (Fig. 1F–H; Supplementary Fig. S2B). We next studied a homogeneous group of 89 stage II patients from GSE33113. This analysis verified the prognostic subgroups, demonstrating that 15-Glycogene Cluster A patients had significant shorter DFS than that of 15-Glycogene Cluster B (Fig. 1I and J; Supplementary Fig. S2C, P = 0.0080). Multivariate analysis revealed that prognostic significance of 15-Glycogene Cluster A was independent of clinical features in GSE41258 (HR, 4.22; 95% CI, 1.50–11.84; P = 0.006) and in GSE33113 (HR, 4.03; 95% CI, 1.31–12.37; P = 0.015; Supplementary Table S7). In all three independent clustering analyses, the expressions of the 15 genes were each consistently altered between clusters (Supplementary Fig. S3E). Specifically, we identified upregulation of GCNT3 and FUT8, and downregulation of GALNT6 as common features of Cluster A1 and 15-Glycogene Cluster A. Of note, the expression of GCNT3 and FUT8, but not GALNT6, has been reported to be associated with prognosis in colorectal cancer (30, 31).

We further analyzed the association between the 15-Glycogene clusters and known molecular markers, such as MMR, RAS, BRAF, and TP53 status. In GSE41258, the 15-Glycogene Cluster A was significantly associated with dMMR (P = 0.003) and wild-type TP53 (P = 0.050; Fig. 1K; Supplementary Table S8). The same clustering procedure was applied to two independent RNA-Seq datasets obtained from TCGA, consisting of RNA-Seq RPKM (n = 193) and RNA-Seq V2 RSEM (n = 361). This validated the association of the 15-Glycogene Cluster A with dMMR and wild-type TP53 (Fig. 1F, K, and L; Supplementary Fig. S2B, SD and S2E; Supplementary Table S8). Intriguingly, we found that proximal location, mucinous histology, mutant RAS, and mutant BRAF were statistically significantly enriched in the 15-Glycogene Cluster A.

We attempted to focus on single glycogenes, including GCNT3, FUT8, and GALNT6, altered expression of which might be characteristics of tumors with dMMR or genetic alterations in BRAF, RAS and TP53. Also, CpG island methylator phenotype (CIMP) was included in this analysis, as CIMP-positive tumors are known to be closely related to dMMR and BRAF mutation (32). Nine additional datasets were assembled and we were thus able to analyze 12 independent cohorts containing a total of 2,472 patients with colorectal cancer (Fig. 2). This revealed the association between GALNT6 expression and MMR status with high reproducibility, where GALNT6 was statistically significantly downregulated in dMMR tumors in all cohorts, comprised of 417 dMMR and 1,800 pMMR tumors. It is worth noting that the association of decreased GALNT6 with dMMR was clearly reproduced by the analysis of 151 colorectal cancer cell lines (Fig. 2; Supplementary Fig. S4A). This tight correlation between GALNT6 downregulation and dMMR prompted us to focus specifically on the significance of GALNT6, which encodes one of the polypeptide GalNAc transferase (ppGalNAc-T) family enzymes involved in the initiation of O-glycosylation.

Downregulation of glycogenes is an important step in colorectal cancer development and progression (14, 18). Thus, we hypothesized that downregulated GALNT6 is involved in carcinogenesis. To this end, we analyzed multiple datasets containing normal colon (n = 161, in total), colon adenoma (n = 264, in total), and carcinoma (n = 387, in total) samples. In all 7 analyses, GALNT6 mRNA expression was significantly higher in adenomas than that of normal colon, while it was significantly decreased in carcinomas, compared with adenomas (Fig. 3A–G). To further validate this finding, IHC for GALNT6 protein was conducted using our large series of colorectal adenoma (n = 60) and carcinoma specimens (n = 335). IHC demonstrated that GALNT6 protein expression was not detected in the vast majority of normal colon mucosal cells (92.8% of 304 normal tissues were GALNT6-negative). Whereas, virtually all samples of adenoma and carcinoma in situ (Tis) showed strong granular cytoplasmic staining of GALNT6 in tumor cells essentially throughout the tumor area (98.3% of adenoma and 100.0% of Tis; Fig. 4A–C). Likewise, intense GALNT6 staining was diffusely found in carcinoma cells (Fig. 4D and E). However, approximately 15% of carcinomas lacked GALNT6 protein expression (Fig. 4F and G; Table 1). We also examined the expression patterns of GALNT6 in adenoma-to-carcinoma transition within the same lesion using 8 specimens of carcinoma-in-adenoma, showing that in one of 8 lesions (12.5%) GALNT6 staining was lacking in the carcinoma component, but all the adenoma components exhibited positive-GALNT6 (Supplementary Fig. S5A–S5H). GALNT6 staining was frequently lost in dMMR tumors (52.0% were negative), although the majority of pMMR tumors showed positive GALNT6 (11.6% were negative; Fig. 3H; Table 1). Collectively, in both mRNA and protein levels, GALNT6 expression was the highest in precursor and preinvasive tumors, and was subsequently downregulated or lost in a subset of carcinomas, which was associated with dMMR tumors.

It has become apparent that more than 15% of colorectal cancer is known to originate from serrated precursor lesions and is often characterized by activating BRAF mutations and CIMP that greatly overlaps with dMMR tumors (4, 5, 33). Because decreased GALNT6 mRNA expression was associated not only with MMR status, but also with CIMP and BRAF mutations (Fig. 2), it was speculated that GALNT6 downregulation could be associated with the serrated neoplasia pathway. Indeed, dMMR, CIMP, and BRAF mutations were each highly enriched in tumors with decreased levels of GALNT6 expression in three cohorts (Supplementary Fig. S6A–S6C). However, we observed no difference in GALNT6 expression between serrated adenomas and conventional adenomas, or between serrated adenocarcinomas and conventional adenocarcinomas in 4 datasets of histologically confirmed adenoma and adenocarcinoma samples (Supplementary Fig. S7A–S7D).

Because downregulation of some glycogenes results from epigenetic silencing mainly by DNA hypermethylation upon malignant transformation (14), we addressed the possibility that DNA methylation contributes to decreased GALNT6 expression. We observed significant inverse correlation between mRNA expression and methylation of GALNT6 (Fig. 3I; P < 0.0001). Higher levels of GALNT6 promotor methylation were observed in colorectal cancer tissues than adenomas in two additional cohorts (Fig. 3J and K), which was in clear contrast to the downregulated GALNT6 in colorectal cancer tissues compared with adenomas (Fig. 3A–F). Moreover, in colorectal cancer tissues, GALNT6 methylation levels were significantly higher in dMMR tumors than that of pMMR (Fig. 3L). To further confirm the methylation of GALNT6 in vitro, colorectal cancer cell lines, including HCT116, SW48, and RKO, which displayed relatively lower GALNT6 expression levels (Supplementary Fig. S4B–S4D), were treated with a DNA methyltransferase inhibitor, 5-aza-dC. Demethylation treatment restored MLH1 expression in MLH1-methylated cell lines, including RKO and SW48, while GALNT6 expression was induced only in SW48 cells (Fig. 3M and N).

We next examined the clinicopathologic and prognostic significance of GALNT6 protein expression in the FFPE cohort. Tumors lacking GALNT6 protein were associated with poorer histologic differentiation (P < 0.0001), but exhibited no association with other clinical features (Table 1). Intriguingly, patients with negative GALNT6 had significantly poorer DSS and OS, compared with those with positive GALNT6 (Supplementary Fig. S8A and S8B; P = 0.0038 and P = 0.022, respectively). This remained statistically significant when the analysis was conducted in 195 stage II and III patients (Fig. 4H; Supplementary Fig. S8C; P = 0.0008 and P = 0.014, respectively). Multivariate Cox analysis demonstrated that the lack of GALNT6 protein was significantly associated with poor DSS (HR, 3.39; 95% CI, 1.28–9.02; P = 0.014) and OS (HR, 2.34; 95% CI, 1.08–5.05; P = 0.031), independent of stage and other conventional factors (Supplementary Tables S9 and S10). Stratified analyses also showed that negative GALNT6 had significant prognostic impact on DSS and OS in stage III patients (P < 0.0001 and P = 0.0016, respectively), but not evident in stage II patients (Fig. 4I and J; Supplementary Fig. S8D and S8E). Concerning DFS, the prognostic values of GALNT6 expression showed only a trend, which did not reach statistical significance (Supplementary Fig. S8F–S8I).

Because GALNT6 was chosen from the 15 genes for detailed evaluation primarily because of its tight relationship with MMR status, patients were further divided into four subgroups based on GALNT6 and MMR status to explore the clinicopathologic and prognostic significance in those groups. Interestingly, although dMMR tumors shared similar clinicopathologic features irrespective of GALNT6 expression (Supplementary Table S11), striking survival differences were found between GALNT6-negative/dMMR and GALNT6-positive/dMMR subgroups. In stage II–III analysis, GALNT6-negative/dMMR patients had significant poorer DSS, OS, and DFS compared with those of GALNT6-positive/dMMR (Supplementary Fig. S9A–S9C). In addition, GALNT6-negative/pMMR patients demonstrated significantly poorer prognosis than those of GALNT6-positive/pMMR, particularly in stage III analyses (Supplementary Fig. S9D–S9F). It appears that there was no clinicopathologic similarity between GALNT6-negative/dMMR and GALNT6-negative/pMMR, although those two subgroups were sharing poor survival outcomes (Supplementary Table S11).

Because loss of GALNT6 protein was associated with worse survival in stage II and III patients, we hypothesized that decreased GALNT6 mRNA levels may also be prognostic. We assembled three additional datasets, combining them into a microarray validation meta-cohort containing 364 patients with stage II and III colorectal cancer (Fig. 4K; Supplementary Fig. S10). Low GALNT6 mRNA was significantly associated with worse DFS in patients with stage II and III colorectal cancer (Fig. 4K, P = 0.0241), and it was independent of clinical factors by multivariate analysis (Supplementary Table S12; HR, 1.88; 95% CI, 1.16–3.06; P = 0.011). Consistent with IHC analysis, the prognostic value of GALNT6 mRNA expression was clearly demonstrated in stage III patients (P = 0.0036), but not in stage II (Fig. 4L and M; Supplementary Fig. S10).

It is well recognized that stage II and III patients with dMMR colorectal cancer may not benefit from 5-FU–based adjuvant chemotherapy (7, 34). We sought to determine whether the expression of GALNT6 was associated with response to adjuvant chemotherapy. Among 190 stage II and III patients in the IHC cohort for which information on the administration of adjuvant chemotherapy was available, 114 patients received intravenous or oral 5-FU–based adjuvant chemotherapy after surgery, while 76 patients were treated by surgery alone. We conducted DFS analyses for GALNT6 expression by stratifying stage II and III patients on the basis of adjuvant treatment history (Supplementary Fig. S11A–S11F). Among patients who received chemotherapy, negative-GALNT6 showed a nonsignificant trend toward worse DFS (Supplementary Fig. S11A). Notably, in stage III patients receiving adjuvant chemotherapy, negative-GALNT6 was associated with poor therapeutic outcome (HR, 5.56; 95% CI, 1.57–19.69; P = 0.0079, Supplementary Fig. S11C). This effect was not observed in stage III patients treated by surgery alone (HR, 0.31; 95% CI, 0.04–2.68; P = 0.290), although the number of patients in each group was limited (Supplementary Fig. S11D). There was no clear trend when stage II patients were analyzed (Supplementary Fig. S11E and S11F).

To understand the biologic function of GALNT6, a pMMR cell line, SW480, with relatively higher GALNT6 mRNA and protein expression were selected for further analyses (Supplementary Fig. S4B–S4D). We used two different siRNAs targeting GALNT6, demonstrating that GALNT6 was effectively silenced, confirmed by qRT-PCR and Western blotting (Fig. 5A and B). Although silencing of GALNT6 had no significant impact on cell proliferation (Fig. 5C), it enhanced both cell migration and invasion determined by wound-healing assay and transwell invasion assay, respectively (Fig. 5D and E; Supplementary Fig. S12). As we found the association between negative GALNT6 and poor response to 5-FU–based chemotherapy (Supplementary Fig. S11), we tested the in vitro contribution of GALNT6 expression to the sensitivity to 5-FU treatment. We found a moderate, but significant increase of 5-FU resistance in GALNT6-knockdown cells as compared with cells treated with control siRNA (Fig. 5F). Correspondingly, apoptosis was significantly suppressed in GALNT6-knockdown cells treated with 5-FU (Fig. 5G).

Dysregulated glycogenes can result in alteration of cell surface glycosylation. Thus, we tested to determine whether the depletion of GALNT6 could affect the cell surface glycan profiles. Lectin microarray analysis was conducted to examine the glycomic profiles of surface membranous fractions in SW480 cells. Compared with siRNA control, GALNT6-silenced cells demonstrated decreased lectin Jacalin and ACA, each of which can bind to core 1 (Galβ1-3GalNAcα-Ser/Thr) and core 3 (GlcNAcβ1-3GalNAcα-Ser/Thr) extension of O-glycan, respectively (Supplementary Fig. S13). GALNT6 silencing also led to the increased signal intensity of lectin HPA that is highly specific to GalNAcα-Ser/Thr, a truncated O-glycan structure, also known as Tn-antigen (Fig. 5H; Supplementary Fig. S13; ref. 35). We confirmed that GALNT6 knockdown increased the cell surface expression of Tn-antigen by flow cytometry using a mAb MLS128 (Fig. 5I and J; ref. 36). Conversely, 5-aza-dc treatment in SW48 cells resulted in decreased expression of Tn antigen along with the concomitant induction of GALNT6 (Figs. 3N and 5K and L).

This study provides several lines of evidence that the expression of GALNT6 is a potential biomarker for identifying a prognostic subgroup and is implicated in colorectal carcinogenesis. First, a glycogene-derived transcriptional subtype, namely, the 15-Glycogene Cluster A, was identified and validated using a total of 941 samples from multiple transcriptomic datasets. This novel subgroup, in which GALNT6 was downregulated, was characterized by poor prognosis, poorly differentiated histology, proximal location, and dMMR. Moreover, strong association between decreased GALNT6 mRNA expression and dMMR was robustly confirmed in 12 patient cohorts and a dataset of cell lines, followed by the analysis of a FFPE cohort at GALNT6 protein levels. Second, downregulation of GALNT6 mRNA and protein seemed to occur during transition from adenoma to carcinoma, possibly through epigenetic silencing, where GALNT6 was expressed in most of premalignant/preinvasive lesions but was subsequently decreased in a subset of carcinomas. This suggests a crucial role of GALNT6 in colorectal cancer, especially contributing to the mechanism referred to as incomplete synthesis of glycans. Indeed, GALNT6 depletion not only increased the invasive and migratory potentials but also upregulated the cancer-associated truncated glycan, Tn-antigen. In contrast, demethylation resulted in a decrease of Tn-antigen along with GALNT6 reactivation. Third, decreased GALNT6 expression in both mRNA and protein levels discriminated a poor prognostic subgroup that was largely consistent with that of the 15-Glycogene Cluster A, reinforcing the notion that the glycogene-derived transcriptional subtype is recapitulated by tumors lacking GALNT6 protein.

Our strategy integrated various gene expression platforms, including Affymetrix, Agilent, and Illumina microarrays and RNA-Seq, obtained from different laboratories, and even technologically independent approach by IHC, consisting of a total of more than 4,500 samples. Likewise, downregulation and methylation of GALNT6 in carcinoma tissues, compared with adenoma tissues, was clearly reproduced in multiple series of nonmalignant, premalignant, and malignant lesions using epigenomic and transcriptomic datasets and IHC analysis. This finally led us to identify a distinct subgroup lacking GALNT6 protein expression in approximately 15% of colorectal cancer. Those integrated, multistep analyses could minimize false-positive results. It is therefore unlikely that the presence of this subgroup is related to false discoveries or batch effects from high-throughput data analyses. Notably, this GALNT6-negative subgroup could be identified in both mRNA and protein levels, and its prognostic values were statistically independent of clinical factors. Therefore, it is suggested that GALNT6 expression can be a robust prognostic biomarker for colorectal cancer. Because its prognostic performance was particularly remarkable in stage III patients, GALNT6 expression may help guide clinical decisions, including adjuvant chemotherapy and surveillance plans after curative surgery for patients with stage III colorectal cancer. It is also important that IHC for GALNT6 protein is a practical assay that can be routinely analyzed on readily available FFPE specimens in clinical practice.

GALNT6 downregulation was originally identified to be tightly correlated with dMMR, and finally we noticed that it had significant impact on prognosis. It is worth noting that GALNT6-negative tumors shared poor survival outcomes even when the dMMR and pMMR tumors were analyzed separately, but the prognostic impact of negative GALNT6 seemed to be more remarkable in the analysis of dMMR tumors. Therefore, we suggest that GALNT6 can be a promising prognostic biomarker for both dMMR and pMMR tumors, and GALNT6 IHC combined with MMR may provide more useful prognostic stratification that can discriminate an extremely poor prognostic subset displaying negative GALNT6/dMMR, from GALNT6-positive/dMMR tumors with excellent prognosis. These results warrant confirmation in large-scale prospective studies.

It is likely that GALNT6-negative patients receiving 5-FU–based adjuvant chemotherapy were associated with poor therapeutic outcome. Despite the exploratory nature with small number of patients and low number of events in each subgroup, the negative prognostic effect on DFS was evident in stage III patients who received 5-FU–based adjuvant therapy, demonstrating a striking contrast to those who were treated by surgery alone. This was further supported by the finding that GALNT6-depleted colorectal cancer cells demonstrated an increase of chemoresistance to 5-FU treatment. This implicated that negative GALNT6 may also have a predictive value for poor response to 5-FU–based adjuvant chemotherapy in stage III colorectal cancer. Therefore, alternative therapeutic strategies, including combination regimens or targeted drugs, may be more effective and appropriate for stage III patients with negative-GALNT6 tumor. Recent clinical trials revealed that stage III dMMR patients may benefit from adjuvant 5-FU treatment combined with oxaliplatin (37, 38). It would be interesting to address the effect of adding oxaliplatin compared with the conventional 5-FU–based therapy alone, in relation to GALNT6 status, although no patients in this study were treated with oxaliplatin in the adjuvant setting. Because dMMR colorectal cancer has recently been reported to be effectively treated with anti-PD-1 immune checkpoint inhibitors (39, 40), detailed analysis of downregulated GALNT6 in relation to ant-tumor immunity may help to understand the dMMR–colorectal cancer biology.

In normal tissues, GalNAc type O-glycans are modified by glycosyltransferases to generate core structures, and core O-glycans are further extended and capped by the addition of sialylated and fucosylated terminal structures (13, 41). The ppGalNAc-Ts, which catalyze the transfer of GalNAc to Ser/Thr residues on substrate proteins, control the initiation step of GalNAc-type O-glycosylation. The ppGalNAc-Ts form a family of 20 distinct isoenzymes expressed in a cell-type–specific manner, with different but overlapping substrate specificity, thus O-glycans are synthesized through concerted and occasionally competitive action of ppGalNAc-Ts (41, 42). GALNT6 was reported to be expressed in high percentages of adenocarcinoma cells from breast, lung, pancreas, and renal cancer, whereas it was undetectable or very weakly found in their normal counterpart (24–26, 43). We showed that GALNT6 staining was undetectable in normal colonic tissue, but was invariably overexpressed in virtually all tumor cells of premalignant/preinvasive lesions. Transcriptomic and epigenomic data confirmed the upregulation of GALNT6 mRNA along with demethylation of GALNT6 promotor in adenoma samples, compared with normal colon. This suggests a role of GALNT6 in the early stage of tumorigenesis, where GALNT6 may even support adenoma formation, irrespective of conventional or serrated carcinogenesis pathways. GALNT6 expression in premalignant tumors seemed to be maintained in the majority of colorectal cancer as well, with the exception of lower percentages (∼15%) of colorectal cancer showing GALNT6 loss. Because GALNT6 silencing could promote the capacity of invasion and migration in vitro, it is likely that in a subset of colorectal cancer, decreased GALNT6 mRNA and loss of GALNT6 protein contribute to transition from premalignant/preinvasive lesions to invasive carcinomas.

Altered expression of GALNT6 has been investigated in other tumor types, suggesting their potential as cancer biomarkers (13). In pancreatic cancer, Li and colleagues reported that loss of GALNT6 expression was associated with poor differentiation and poor OS (25). Conversely, the same group from Li and colleagues has recently reported that GALNT6 expression predicted poor OS in lung adenocarcinoma (24). In breast cancer, GALNT6 overexpression may contribute to mammary carcinogenesis through aberrant glycosylation (26). Such conflict between different cancers has also been observed in several studies investigating other ppGalNAc-Ts. For instance, GALNT3 expression correlated with poor survival in ovarian cancer (44) and renal cell carcinoma (43), whereas it was associated with better survival in lung adenocarcinoma (45), gastric cancer (46), and colorectal cancer (47). GALNT7 was shown to be targeted by miRNA-214 in cervical (48) and esophageal cancer (49); its overexpression enhanced proliferation, invasion, and migration. In contrast, in melanoma cells, microRNA-30b/30d promoted invasion and metastasis by direct suppression of GALNT7 (50). Although there is no direct explanation for the contradictory influence of ppGalNAc-Ts in different cancers, these conflicting data among different cancer types may indicate the complexity of O-glycosylation along with the diversity and distinct substrate specificities of ppGalNAc-Ts that can confer specific roles in specific cellular contexts. Future studies would be required to address this complexity of O-glycosylation associated with deregulated ppGalNAc-Ts during tumorigenesis of various malignancies.

Upon malignant transformation, epigenetic alterations are recognized as key characteristics that can cause dysregulation of glycogenes, resulting in aberrant expression of cell surface glycans (16). We found an inverse correlation between the expression and methylation of GALNT6, and demethylation treatment reactivated GALNT6 expression in colorectal cancer cell lines. Also, cell surface lectin microarray analysis revealed that the levels of lectin HPA-recognized GalNAcα-Ser/Thr, known as cancer-associated Tn antigen, were specifically increased in GALNT6-depleted cells, confirmed by using a mAb MLS128 (35, 36). This truncated O-glycan structure, Tn, is involved in tumor progression in many types of cancer, including colorectal cancer (35, 51–53). Indeed, Tn antigen is known to be a marker of poorly differentiated and mucinous adenocarcinoma, and poor patient prognosis in colorectal cancer (13, 42). GALNT6 knockdown promoted invasion and migration, which was in agreement with the tumor phenotype with Tn antigen overexpression. Conversely, cell surface Tn antigen was diminished by DNA demethylating agent along with GALNT6 induction. Those findings are highly consistent with the concept of incomplete synthesis that glycan elongation in nonmalignant cells are impaired upon malignant transformation by silencing of glycogenes, resulting in the expression of cancer-associated truncated glycans (13, 16, 18). Taken together, our results suggest that GALNT6 expression is epigenetically regulated during preinvasive neoplasm-invasive carcinoma transition in a subgroup of colorectal tumors that contribute to cancer progression possibly through the incomplete synthesis mechanism.

We found that not only dMMR but also CIMP and BRAF mutation were each enriched in tumors showing decreased levels of GALNT6 expression, and thus they were overlapping considerably each other. Because CIMP-positive tumors are known to have poor prognosis and are frequently accompanied by BRAF mutation, our study might raise the possibility that the prognostic impact of GALNT6 can be at least in part attributed to CIMP phenotype. Although the underlying driver biology of GALNT6 loss remains inconclusive, it appears that CIMP might be one of the putative mechanisms that can explain the epigenetic silencing of GALNT6. Further investigation would be required for understanding the pathogenesis of GALNT6-negative colorectal cancer associated with promotor methylation of GALNT6 and methylator phenotype.

This study had several limitations, including its retrospective nature and lack of IHC validation in independent cohorts. In addition, CIMP status and mutational profiles, such as RAS and BRAF, were unavailable in the FFPE cohort. Therefore, those results presented here would need to be validated in the future investigations, by combining GALNT6 IHC with other genetic and epigenetic biomarkers, for instance, BRAF, RAS, CIMP, and PIK3CA (4, 5). Concerning the in vitro loss-of-function experiments, this study might not provide conclusive evidence that epigenetically silenced GALNT6 could directly contribute to tumor progression. Thus, we suggest that functional assays using panels of cell lines harboring several genetic profiles and in vivo tumorigenicity assays would be interesting future directions.

In conclusion, we developed and validated the 15-glycogene signature that can identify a genomically distinct subgroup exhibiting dMMR, decreased GALNT6 expression and poor outcomes. Also, GALNT6 expression can be a novel prognostic biomarker that can be applied to FFPE specimens. GALNT6 downregulation, in part, due to epigenetic silencing, may contribute to the incomplete O-glycan synthesis and increased expression of the cancer-associated Tn antigen, highlighting the possible role of GALNT6 in colorectal carcinogenesis and poor prognosis.

No potential conflicts of interest were disclosed.

Conception and design: M. Noda, H. Okayama, K. Kono

Development of methodology: M. Noda, H. Okayama, K. Tachibana, K. Kono

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): M. Noda, H. Okayama, K. Saito, T. Nakajima, T. Momma, K. Katakura, S. Ohki, K. Kono

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): M. Noda, H. Okayama, K. Saito, T. Nakajima, K. Aoto, K. Kono

Writing, review, and/or revision of the manuscript: M. Noda, H. Okayama, K. Kono

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): H. Okayama, K. Tachibana, W. Sakamoto, A.K. Thar Min, M. Ashizawa, T. Momma

Study supervision: H. Okayama, K. Saito, K. Kono

This work was supported by JSPS KAKENHI grant numbers 15K10143 and 25870582. H. Okayama and M. Noda were supported by Takeda Science Foundation. The authors thank Dr. Yuuichirou Kiko for providing pathologic advice.

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.