A flowchart for this study is shown in Figure 1 . Briefly, this study was carried out in three cohorts. First, a genome-wide methylation scan by 850K array on cancerous and paired normal tissues from 12 CRC patients in cohort I was performed, followed by cross-validation using DNA methylation data from the TCGA database (https://cancergenome.nih.gov) and the GEO database (https://www.ncbi.nlm.nih.gov/geo/). The DNA methylation data from the TCGA and the GEO were generated using an Illumina HumanMethylation450 BeadChip (450K array) in 438 CRC tissue samples (393 tumor, 45 normal) and 208 CRC tissue samples (104 tumor, 104 normal), respectively. An overview of the external datasets used in this study is shown in Supplementary Table S1 . Then, 48 pairs of CRC tissue samples from cohort II were tested. Additionally, the methylation levels of DLX6-AS1 were further validated in cohort III, which consisted of 286 CRC patients, 81 AA patients and 81 NAA patients. The characteristics of the participants in each cohort subjected to the tissue-based methylation analysis are shown in Table 1 .

To evaluate the DNA methylation levels in peripheral blood, we randomly sampled 60 CRC patients and 60 adenoma patients with complete tissue-based DNA methylation data from cohort II and cohort III, and 60 healthy controls from a population-based cohort. The DNA methylation status of the same region as measured in tissue samples was tested in each sample of peripheral blood leucocyte DNA. The characteristics of the participants subjected to the peripheral blood-based methylation analysis are shown in Supplementary Table S2 .

To evaluate the DNA methylation levels in cfDNA, the DNA methylation data generated by 850K array in 7 cfDNA samples (3 CRCs, 4 healthy controls) were obtained from GEO database ( Supplementary Table S1 ).

To evaluate the influence of DLX6-AS1 methylation on survival, CRC patients with successfully measured DNA methylation data in our cohort II and cohort III were pooled together, and CRC patients with both available methylation data and survival information from the TCGA database were used as an external validation.

CRC patients from Shaoxing People’s Hospital were enrolled between January 2015 and July 2018. Participants with AA or NAA and healthy controls were selected from an ongoing population-based cohort since 1989 in Jiashan County, which has been described previously (11). All participants were ethnic Han Chinese from Zhejiang Province and were pathologically confirmed, with no familial adenomatous polyposis (FAP), no previous history of CRC and no preoperative anticancer treatment. For each participant, histologically confirmed tissue samples, including a colorectal lesion (carcinoma or adenoma) and an adjacent normal mucosa sample, and peripheral blood samples were obtained. The adjacent normal mucosa was collected from the colonic mucosa 5 cm distal from the main neoplasm. Adenomas were classified as AA (any adenoma ≥ 1 cm, high-grade dysplasia, or with tubulovillous or villous histology) and NAA (adenomas < 1 cm without advanced histology) according to current guidelines (20). The TNM staging classification for CRC was determined according to the 7th edition of the American Joint Committee on Cancer (AJCC) cancer staging manual (21).

The study protocol was approved by the Medical Ethics Committee of Zhejiang University School of Medicine. Before basic information and sample collection, written informed consent was obtained from all recruited participants.

Genomic DNA from fresh-frozen samples and peripheral blood leukocytes was isolated using a DNA Tissue Kit (TianLong Biotech, Xi’an, China) and a RelaxGene Blood DNA System (TianGen Biotech, Beijing, China), respectively. Bisulfite treatment was conducted on genomic DNA (500 ng) using the EZ Methylation Gold Kit (Zymo Research, Irvine, CA, USA). All procedures were conducted in accordance with the manufacturer’s instructions.

Genome-wide DNA methylation profiling was analyzed using the 850K array in 12 pairs of cancerous and adjacent normal tissues according to the manufacturer’s instructions as described in a previous study (11). In this study, the raw array data were processed using the ChAMP package in R software for deriving the methylation level, which was generated as beta values (fraction methylation values between 0 and 1). We focused mainly on probes located in the promoter region of lncRNAs, which was defined as 1500 bp upstream and downstream from the transcription start site (TSS). The lncRNA annotation file was obtained from LNCipedia (https://hg19.lncipedia.org/) and the mapping procedure was conducted using the bedtools (22). Probes were selected on the basis of showing a difference in methylation of ≥ 0.20 and an adjusted P value (Benjamini-Hochberg method) < 0.05. To cross-validate the results based on our samples, the eligible methylation data in TCGA and GEO were obtained and analyzed. The detailed procedures of data processing have been supplemented in the Supplementary Methods . Due to the larger coverage of the 850K array as compared to 450K array, the new probes in 850K array were cross-validated by the average beta value of the promoter regions of the target genes in 450K array.

The methylation levels of particular CpG sites located in the promoter region of candidate genes were verified using MassARRAY EpiTYPER (Sequenom, San Diego, CA). The schematic representation of each candidate gene is provided in the UCSC browser (http://genome.ucsc.edu). The primers were designed using EpiDesigner (http://epidesigner.com, Supplementary Table S3 ). The analyzed sequences are shown in Supplementary Figures S1 – 6 . In some cases, fragments resulting from the T-cleavage reaction may contain small groups of adjacent CpG sites and are therefore referred to as “CpG units”. CpG sites that were outside of the mass spectrometry analytical window (low or high mass) were filtered out. The mass spectra were collected on a MassARRAY Compact MALDI-TOF system (Sequenom, BioMiao Biological Technology, Beijing, China), and the methylation proportions of individual units on the spectra were generated by EpiTYPER software (Sequenom, San Diego, CA). Methylation levels ranging from 0 (completely nonmethylated) to 1 (fully methylated) are presented. For each gene, CpG unites with missing values in more than 20% of the samples were removed, as well as samples with missing values in more than 20% of CpG unites. The average methylation value of all CpG units was calculated as a representation of the region-specific gene methylation level.

Statistical analyses were performed in R software (version 3.6.2, R Foundation for Statistical Computing, Vienna, Austria). Continuous variables are presented as the mean and standard deviation (SD), and categorical variables are presented as the frequency.

A paired Student’s t test was used to assess the differences in DNA methylation levels between colorectal lesion tissues and paired normal tissues. Analysis of variance (ANOVA) followed by Bonferroni’s posttest was used to examine significant differences between different groups. Pearson correlation analyses were used to evaluate the consistency of DLX6-AS1 methylation levels between peripheral blood and local lesions of the same patients with CRC or adenoma. The performance of the mean methylation level of candidate genes in distinguishing colorectal lesion tissues from their adjacent normal tissues was tested by receiver operating characteristic (ROC) curve analysis, and the area under the curve (AUC), sensitivity, and specificity were calculated. In the survival analysis, we adopted the best Youden index based on the time-dependent ROC curve as an optimal cutoff to dichotomize the study patients into high-risk and low-risk groups. Survival differences between groups were assessed using the Kaplan-Meier test and compared by the log-rank test. Hazard ratios (HRs) and 95% confidence intervals (95% CIs) were calculated by univariate and multivariate Cox proportional hazards regression analyses. The multivariate analysis was adjusted for age, sex and TNM stage. A nomogram was established to predict the 1-, 2-, 3- and 4-year survival for CRC patients. Harrell’s concordance index (C-index) was measured to quantify the discrimination ability of the nomogram, while the calibration curves were used to evaluate whether the predicted survival probabilities were consistent with those observed. All analyses were carried out in a two-sided manner, with a P value < 0.05 regarded as statistically significant.

By DNA methylation profiling, a total of 185 differentially methylated CpG sites mapping to the promoter of lncRNAs (all with P _adj < 1*10^-5 and β difference > 0.20) were identified by the 850K array generated from 12 pairs of colorectal cancerous and adjacent normal tissues, followed by cross-validation using DNA methylation data generated by the 450K array in CRCs from the TCGA database (tumor=393, normal=45) and GEO database (tumor=104, normal=104), respectively ( Supplementary Table S4 ). Among them, 95.14% (176/185) of the identified CpG sites were significantly hypermethylated and 4.86% (9/185) were significantly hypomethylated. The methylation levels for each differentially methylated CpG sites are shown by heat maps ( Figure 2 ). Among the list of CpG sites, we focused on six sites ranking on the top (cg24014202 in DLX6-AS1, cg18323466 in lnc-DPH5-1, cg08430489 in lnc-PRSS2-6, cg17722675 in lnc-RPS12-6, cg00159100 in lnc-SFRP4-2, cg27442308 in SOX21-AS1) for following technical confirmation analysis ( Figure 3 ), which were considered as candidate biomarkers.

To confirm the above findings, target regions covering the identified CpG units in the promoter of the above 6 candidate genes were amplified in 48 CRCs by MassARRAY EpiTYPER in cohort II. The methylation levels of the mean and individual CpG units for each gene, which were significantly higher in colorectal cancerous tissues than in adjacent normal tissues, are shown in bar plots ( Figures 4A–F ). Specifically, the increases in methylation status between cancerous and paired normal mucosa were found in 95.83% (46/48), 79.17% (38/48), 100% (42/42), 93.75% (45/48), 80.43% (37/46) and 91.67% (44/48) of CRCs, respectively, for DLX6-AS1, lnc-DPH5-1, lnc-PRSS2-6, lnc-RPS12-6, lnc-SFRP4-2 and SOX21-AS1 ( Supplementary Figure S7 ).

ROC curve analyses revealed that methylation status of each individual genes could significantly distinguish primary carcinoma from adjacent normal mucosa, as measured by AUC value (DLX6-AS1: 0.941; lnc-DPH5-1: 0.833; lnc-PRSS2-6: 0.913; lnc-RPS12-6: 0.916; lnc-SFRP4-2: 0.830; SOX21-AS1: 0.921) ( Figure 4G ). Among them, DLX6-AS1 showed a high discriminative performance and was therefore chosen for further validation.

To elucidate the DLX6-AS1 methylation pattern during colorectal neoplastic progression, the methylation status was assessed in colorectal lesion tissues and adjacent normal tissues from 286 CRCs, 81 AAs and 81 NAAs in cohort III with MassARRAY EpiTYPER. Among them, 433 histologically confirmed colorectal lesion tissues (283 CRCs, 76 AAs and 74 NAAs) and 441 adjacent normal tissues (284 CRCs, 80 AAs and 77 NAAs) were successfully measured. DLX6-AS1 hypermethylation was detected at all stages of colorectal neoplasms, even as early as the NAA stage. Compared to their adjacent normal tissues, 94.31% (265/281) of CRCs, 85.33% (64/75) of AAs and 80.00% (56/70) of NAAs presented higher DLX6-AS1 methylation levels, with statistically significant differences (all P < 0.001) ( Figure 5 and Supplementary Figure S8 ). The mean DLX6-AS1 methylation levels could distinguish primary lesions from their adjacent normal mucosa, with AUC values of 0.944 (95% CI: 0.922-0.962), 0.811 (95% CI: 0.739-0.870) and 0.767 (95% CI: 0.688-0.834) for CRC, AA and NAA, respectively ( Figure 5 ). When comparing the DLX6-AS1 methylation levels between different stages of colorectal lesions ( Table 2 ), the DLX6-AS1 promoter was revealed to be significantly hypermethylated between CRC vs. NAA (P < 0.001) and AA vs. NAA (P = 0.004) but not between CRC vs. AA (P = 1.000).

To evaluate the potential of DLX6-AS1 methylation as a noninvasive biomarker for the diagnosis of colorectal neoplasms, DLX6-AS1 methylation levels were measured in the peripheral leucocyte DNA of 60 CRC patients, 60 adenoma patients and 60 healthy controls. However, there were no significant differences in peripheral blood-based DLX6-AS1 methylation levels in multiple comparisons between CRC patients, adenoma patients and healthy controls ( Supplementary Table S5 ). Even though some CpG units, such as CpG_2.3, reached a statistically significant level (P = 0.017), the methylation levels did not differ much across the different groups. When evaluating the consistency between peripheral blood and local lesions from the same patients ( Supplementary Table S6 ), the Pearson correlation analysis showed poor correlations between matched peripheral blood and local lesions (P = 0.362 for CRCs and 0.893 for adenomas, respectively, in average methylation levels).

To identify the methylation status of the DLX6-AS1 promoter in cfDNA of CRC patients, we analyzed the methylation data generated by the 850K array in cfDNA from 3 CRC patients and 4 healthy controls in GEO dataset. Among the available 35 CpG sites in the promoter of DLX6-AS1, 18 significantly hypermethylated CpG sites were identified (ß difference > 0.20 and P _adj < 0.05). The mean methylation values were 0.40 and 0.16 in CRC patients and healthy controls, respectively (P _adj = 0.003). Details of the results are presented in Supplementary Figure S9 .

331 CRC patients with DNA methylation successfully measured by MassARRAY EpiTYPER were included. During follow-up, there were 53 CRC-specific deaths, and the median follow-up time was 3.60 years. Kaplan-Meier analysis revealed that patients with a high DLX6-AS1 methylation status had poorer disease-specific survival (DSS) rates than those with a low methylation status (P = 0.017, Figure 6A ). A multivariate Cox regression model considering the most relevant risk factors, including age, sex and TNM stage, confirmed that DLX6-AS1 methylation was an independent prognostic biomarker for poorer DSS (HR = 2.52, 95% CI: 1.35-4.69, P = 0.004, Figure 6B ). This unfavorable effect was also identified in most individual CpG units, among which CpG_22 was the strongest indicator both in univariate regression model (HR = 2.67, 95% CI: 1.45-4.92, P = 0.002) ( Figure 6C ) and multivariate regression model (HR = 3.97, 95% CI: 2.08-7.57, P < 0.001) ( Figure 6D ). External validation based on additional 379 CRC patients (1.86 years of median follow-up time, 86 overall deaths) from the TCGA database confirmed the poorer prognosis with hypermethylated DLX6-AS1 (P = 0.007), and the HR (95%CI) for overall survival (OS) by multivariate Cox proportional hazards regression model was 1.64 (1.02-2.64, P = 0.042) ( Supplementary Figure S10 ).

We further built a nomogram, including the methylation status of DLX6-AS1 and clinical factors (age, gender, and TNM stage). The nomogram served as an individual’s prognostic predictor to predict the probability of disease-specific survival with 1-, 2-, 3-, and 4-year for CRC patients ( Figure 7A ). The C-index of the nomogram for predicting the DSS of CRC patients was 0.789 (95%CI: 0.681-0.897), and calibration curves for the 1-, 2-, 3-, and 4-year survival probability demonstrated optimal agreement between the prediction and actual observation ( Figures 7B–E ). Similar results were observed in the TCGA dataset ( Supplementary Figures S11 ).