Genetic and epigenetic alterations in normal tissues have differential impacts on cancer risk among tissues

Significance

The relative importance of genetic and epigenetic alterations in normal tissues on cancer risk was clearly different between esophageal squamous cell and gastric cancers, implying a variety of differences in various types of cancers. The difference observed was well explained by known etiologies: tobacco mutagens for esophageal cancer and chronic inflammation for epigenetic alterations in gastric cancer. The study showed that, if epigenetic and genetic alterations in normal tissues are combined, reflecting their relative contributions, patients with cancer can be precisely discriminated, opening up an avenue to precision cancer risk diagnosis. The study also indicated that for effective cancer prevention, allocation of resources and efforts against genetic and epigenetic alterations should consider their relative contributions.

Abstract

Genetic and epigenetic alterations are both involved in carcinogenesis, and their low-level accumulation in normal tissues constitutes cancer risk. However, their relative importance has never been examined, as measurement of low-level mutations has been difficult. Here, we measured low-level accumulations of genetic and epigenetic alterations in normal tissues with low, intermediate, and high cancer risk and analyzed their relative effects on cancer risk in the esophagus and stomach. Accumulation of genetic alterations, estimated as a frequency of rare base substitution mutations, significantly increased according to cancer risk in esophageal mucosae, but not in gastric mucosae. The mutation patterns reflected the exposure to lifestyle risk factors. In contrast, the accumulation of epigenetic alterations, measured as DNA methylation levels of marker genes, significantly increased according to cancer risk in both tissues. Patients with cancer (high-risk individuals) were precisely discriminated from healthy individuals with exposure to risk factors (intermediate-risk individuals) by a combination of alterations in the esophagus (odds ratio, 18.2; 95% confidence interval, 3.69–89.9) and by only epigenetic alterations in the stomach (odds ratio, 7.67; 95% confidence interval, 2.52–23.3). The relative importance of epigenetic alterations upon genetic alterations was 1.04 in the esophagus and 2.31 in the stomach. The differential impacts among tissues will be critically important for effective cancer prevention and precision cancer risk diagnosis.

The vast majority of human cancers develop via the accumulation of genetic and epigenetic alterations (1, 2). Both alterations can be accumulated, affecting both driver and passenger genes, in normal tissues at low levels long before cancer develops (3, 4), and this accumulation constitutes cancer risk (5). In particular, accumulation of epigenetic alterations, namely, aberrant DNA methylation, in normal tissues of patients with cancer has been well-documented for multiple types of cancer (6–10) and is associated with cancer risk for some cancers, including gastric and esophageal squamous cell cancers (ESCCs) (11, 12). Especially for gastric cancer, a multicenter prospective cohort study established a strong association between accumulation of aberrant DNA methylation and cancer risk, even in a clinical setting (13, 14).

Accumulation of genetic alterations, represented by point mutations (15), also constitutes cancer risk, as demonstrated by many studies in transgenic animals with a marker gene (16). By using a marker gene, whose point mutations at the 10⁻⁵ to 10⁻⁴ per gene level can be accurately measured, associations between carcinogen exposure and accumulation of point mutations and between the accumulation and cancer risk have been shown (17). In contrast to these animal studies, in humans, it has been difficult to measure accumulation levels of point mutations, mainly because of their very low frequencies. Taking advantage of clonal patches of cells or clonal expansion of cells, recent studies demonstrated that mutations are indeed accumulated in normal human tissues (18, 19).

We also recently developed a method to quantify rare point mutations in any human DNA sample by analyzing only a small number of DNA molecules (20). Because of its technical simplicity, quantitative assessment of accumulation of point mutations in human normal tissues can be now performed. It is expected that, by analyzing both genetic and epigenetic alterations in normal tissues of patients with cancer and healthy individuals (with and without exposure to cancer risk factors), we can estimate relative impacts of genetic and epigenetic alterations on cancer risk. Clarification of their relative importance is expected to have great value for precise cancer risk estimation based on their accumulation and for allocation of resources for effective cancer prevention, as etiologies for genetic and epigenetic alterations are different (21).

In the current study, we evaluate the impacts of accumulation of point mutations and DNA methylation on ESCC and gastric cancer risk and reveal their differential impacts.

Results

Genetic Alterations in Normal Esophageal and Gastric Mucosae and Cancer Risk.

First, we measured mutation frequency in normal esophageal and gastric mucosae with three cancer risk levels (samples in the test set, SI Appendix, Table S1; flowchart, SI Appendix, Fig. S1) (20). In esophageal mucosae, mutation frequencies in the low-risk (n = 30), intermediate-risk (n = 32), and high-risk (n = 31) groups were 1.5 ± 0.2 (mean ± SD), 1.8 ± 0.4, and 2.4 ± 0.6 × 10⁻⁵ per base, respectively (Fig. 1A). The frequency of genetic alterations clearly and significantly increased in a stepwise manner according to the risk level. In particular, the odds ratio to distinguish patients with ESCC (high-risk group) from healthy individuals exposed to risk factors, namely, alcohol drinking, betel quid chewing, and cigarette smoking (ABC) (12, 22) (intermediate-risk group), was as high as 18.2 (95% confidence interval, 3.7–89.9). The mutation frequency was weakly correlated with alcohol drinking (r = 0.362; P = 0.04, SI Appendix, Table S2), but not with smoking or chewing betel quid.

Fig. 1.

Accumulation of somatic mutations in esophageal and gastric mucosae. (A) Mutation frequencies in esophageal (n = 93) and gastric (n = 96) mucosae measured by the 100-copy method (20). In esophageal mucosae, the mutation frequency showed a stepwise increase from healthy individuals without lifestyle risk factors (low risk, n = 30) to healthy individuals with lifestyle risk factor exposures (intermediate risk, n = 32) and patients with cancer (high risk, n = 31). In contrast, in gastric mucosae, the mutation frequency did not show a stepwise increase with respect to risk level [low risk (n = 32), intermediate risk (n = 32), and high risk (n = 32)], but the difference between the low- and intermediate-risk groups was significant. (B) Mutation patterns in esophageal and gastric mucosae depicted using a 96-substitution classification. N shows the number of mutations analyzed in each panel. The mutation patterns were distinct between the esophageal and gastric mucosae.

In gastric mucosae, surprisingly, a stepwise increase with respect to risk level was not observed (Fig. 1A). Namely, the mutation frequencies in the low-risk (n = 32), intermediate-risk (n = 32), and high-risk (n = 32) groups were 0.6 ± 0.3, 0.9 ± 0.4, and 0.7 ± 0.4 × 10⁻⁵ per base, respectively. Although the difference between healthy individuals exposed to the major risk factor, Helicobacter pylori infection, and those without was significant, the difference between healthy individuals with H. pylori infection and patients with cancer who have had H. pylori infection was not. This suggested a larger contribution of epigenetic alterations than that of genetic alterations in gastric carcinogenesis.

Characteristics of Mutation Patterns in Normal Tissues.

We also analyzed mutation patterns in esophageal and gastric mucosae for their characteristics and their association with mutation signatures of specific lifestyle risk factors (Fig. 1B and SI Appendix, Fig. S2). In esophageal mucosae, no significant difference in mutation patterns was detected among the three risk groups. However, when the analysis was limited to individuals most heavily influenced by smoking (individuals in the high-risk group, smoking score, 4; age, <60 y), the mutation type associated with tobacco smoking (C:G to A:T transversions) (23, 24) had a significantly higher frequency than the low-risk group (P = 0.032; SI Appendix, Fig. S3A and Table S3). In gastric mucosae, importantly, the mutation signature of activation-induced cytidine deaminase (RC to RT) was more frequent in individuals with H. pylori infection than in those without H. pylori infection (P = 0.006; SI Appendix, Fig. S3B and Table S4), in accordance with a previous report (25).

To assess selection bias for protein-altering mutations in normal tissues, we compared the ratio of nonsynonymous to synonymous mutations (dN/dS). In both esophageal and gastric mucosae, regardless of risk groups, nonsynonymous mutations were significantly less frequent than synonymous mutations (dN/dS = 0.50–0.58; SI Appendix, Fig. S4), suggesting that nonsynonymous mutations are strongly selected against in normal human tissues. We further analyzed similarity of mutation patterns between normal and cancer tissues (SI Appendix, Fig. S5), using mutations reported in esophageal cancers (n = 88) (26) and gastric cancers (n = 100) (27). Cosine similarity was very high in the stomach (cosine similarity = 0.97) and high in the esophagus (cosine similarity = 0.88), suggesting that mutation signatures in normal tissues are reflected in cancer tissues.

Epigenetic Alterations in Normal Esophageal and Gastric Mucosae and Cancer Risk.

We then focused on the impact of epigenetic alterations on cancer risk (flowchart, SI Appendix, Fig. S1). To assess overall epigenome damage accumulated in a tissue, it is critical to use appropriate marker genes for individual tissues. For esophageal mucosae, such methylation markers have not been extensively screened. Therefore, we performed genome-wide methylation analyses of samples in a screening set (n = 27), using two different algorithms (SI Appendix, Fig. S6). One method involved screening for probes with large differences between the two groups. The other method involved screening for probes with high variance within the high-risk group [i.e., the epigenetic variable outliers for risk prediction algorithm (iEVORA) method] (10, 28). We obtained a total of eight candidate regions with high-quality quantitative methylation-specific PCR (qMSP) primers (four by each method; SI Appendix, Tables S5 and S6). Methylation levels of the eight regions and one ESCC risk marker previously isolated (HOXA9) (12, 29) were measured by qMSP in the validation set (n = 154; SI Appendix, Table S1). Seven markers (TFAP2E, OTX1, OPLAH, CHAD, MARCH11, GALR1, and HOXA9) showed significant differences between the high-risk and intermediate-risk groups and were considered definite risk markers for ESCC (SI Appendix, Fig. S7).

Then we analyzed the methylation levels of the seven markers in the test set used for the mutation analysis (Fig. 2A and SI Appendix, Fig. S8). The methylation levels of all the seven markers were significantly different between the low- and intermediate-risk groups (due to exposure to lifestyle risk factors) and between the intermediate- and high-risk groups (due to cancer risk). The areas under the curve of the seven markers to distinguish patients with cancer among individuals with exposure to risk factors were 0.61–0.79 (SI Appendix, Fig. S9A). Thus, epigenetic alterations in esophageal mucosae were strongly associated with cancer risk. Age-related methylation was analyzed in the three risk groups. Especially in the low-risk group, the influence of exposure to lifestyle risk factors was minimum, but no significant correlations were observed in this group (SI Appendix, Fig. S10).

Fig. 2.

Accumulation of aberrant DNA methylation in esophageal and gastric mucosae. (A) DNA methylation levels of definite marker genes were measured by qMSP in the same samples as in Fig. 1. In esophageal mucosae, the methylation level of a representative marker (TFAP2E) showed a stepwise increase according to the risk level. In gastric mucosae, the methylation level of a representative marker (miR-124a-3) showed a prominent stepwise increase. (B) DNA methylation of multiple CpG sites analyzed by deep bisulfite sequencing. For the low-, intermediate-, and high-risk groups, samples at the 75th, 50th, and 25th percentiles of the methylation level were analyzed. For individual samples, more than 300 molecules were sequenced for 11 CpG sites (TFAP2E) and 23 sites (miR-124a-3), and the sequences of 300 randomly selected molecules are shown. The fraction of densely methylated molecules was in good accordance with the methylation level, and the presence of dense methylation was observed. This analysis showed a prominent effect of epigenetic alterations on gastric cancer risk.

In gastric mucosae, we previously identified methylation markers in a genome-wide methylation analysis and demonstrated their usefulness as cancer risk markers in a prospective clinical study (13, 14). Therefore, we used three markers established in our previous studies (miR-124a-3, EMX1, and NKX6.1) to evaluate the impact of epigenetic alterations in gastric mucosae. The methylation levels of the three markers increased prominently in a stepwise manner according to the risk level (Fig. 2A and SI Appendix, Fig. S11). The areas under the curve of the three markers were 0.65–0.79 (SI Appendix, Fig. S9B). Epigenetic alteration was thus also strongly associated with cancer risk in gastric mucosae.

Significant correlations among methylation levels of all of the methylation markers were observed, especially in gastric mucosae (SI Appendix, Tables S7 and S8). This finding confirmed that these markers measured a single component (i.e., the overall level of epigenetic alterations in the genome) and indicated that a marker with the highest area under the curve (TFAP2E for esophageal mucosae and miR-124a-3 for gastric mucosae) can be used as a representative for each tissue. In contrast, regardless of tissue type, a correlation between mutation frequency and methylation levels for the risk marker genes was not observed in any risk groups, showing that genetic and epigenetic alterations are independent indicators of cancer risk. Age-related methylation was analyzed, and no significant correlations were observed in the low-risk group (SI Appendix, Fig. S12). The correlation in the high-risk group was considered to reflect the duration of H. pylori infection (30).

We analyzed the presence of dense methylation, which can silence the transcription of a downstream gene when it is present in a promoter CpG island (CGI) (31, 32), by deep bisulfite sequencing. Both in esophageal and gastric mucosae, not only sparse methylation but also dense methylation increased significantly in a stepwise manner according to the risk level (TFAP2E and miR-124a-3; Fig. 2B), and the results by qMSP were confirmed.

Differential Effect of Genetic and Epigenetic Alterations on Cancer Risk.

Finally, we assessed the relative importance of genetic and epigenetic alterations on cancer risk by logistic regression analyses, using the intermediate-risk group (controls) and the high-risk group (cases; flowchart, SI Appendix, Fig. S1). Based on a univariate analysis, the mutation frequency in esophageal mucosae was associated with ESCC risk with a high odds ratio (18.2; 95% CI, 3.7–89.9), and the methylation level of a representative marker (TFAP2E) was also strongly associated (odds ratio, 10.9; 95% CI, 3.2–36.9), showing the importance of both genetic and epigenetic alterations (Fig. 3A). In contrast, in gastric mucosae, the mutation frequency was not associated with gastric cancer risk, whereas the methylation level was closely associated with cancer risk (odds ratio, 6.6; 95% CI, 2.2–19.7). This clearly showed the greater importance of epigenetic alterations on gastric cancer risk compared with genetic alterations.

Fig. 3.

Evaluation of impacts of genetic and epigenetic alterations. (A) A univariate analysis and (B) a multivariate logistic regression analysis, respectively, in the model of discrimination of patients with cancer (high-risk group) among individuals with exposure to lifestyle risk factors (high- and intermediate-risk groups). In the univariate analysis (A), a higher area under the curve was obtained for genetic alterations in esophageal mucosae and for epigenetic alterations in gastric mucosae. In the multivariate analysis (B), in esophageal mucosae, the addition of mutation frequency to traditional risk factors (age + lifestyle factors, model 1) significantly improved the discrimination ability (likelihood ratio test P < 0.05). Further, a significantly larger c-index was obtained by adding the TFAP2E methylation level to the model. In gastric mucosae, the addition of a methylation risk marker to the traditional risk factors (model 1) significantly improved the discrimination ability, but further addition of mutation frequency did not. An appropriate combination of genetic and epigenetic alterations was shown to be critical for precision cancer risk diagnosis, taking account of life history. (C) The relative importance of the methylation level and mutation frequency quantified by computing the ratio of the standardized coefficients and the generalization of this ratio. (D) A schema of the relative importance. In esophageal mucosae, genetic and epigenetic alterations had equal impacts on cancer risk, and, in gastric mucosae, epigenetic alterations had a greater impact on cancer risk than genetic alterations.

We further incorporated the influence of traditional risk factors (i.e., age and lifestyle risk factors) by a multivariate logistic regression analysis (Fig. 3B). In esophageal mucosae, the addition of mutation frequency to traditional risk factors significantly improved the risk prediction. The addition of the TFAP2E methylation level also significantly improved the risk prediction, showing that combined measurement of genetic and epigenetic alterations was effective for precise cancer risk estimation, with the relative importance of methylation being 1.0 (Fig. 3C). In contrast, in gastric mucosae, the addition of the miR-124a-3 methylation level to a traditional risk factor significantly improved risk prediction, but further addition of mutation frequency did not (Fig. 3B), with the relative importance of methylation being 2.3 (Fig. 3C). In both esophageal and gastric mucosae, the same relative importance was observed using other methylation markers (SI Appendix, Fig. S13). These results demonstrated that the impacts of genetic and epigenetic alterations on cancer risk are strikingly different between ESCC and gastric cancer, and the impact of epigenetic alterations exceeded that of genetic alterations in gastric cancer risk (Fig. 3D).

Discussion

Our analyses demonstrated that the impacts of genetic and epigenetic alterations on cancer risk are strikingly different between ESCC and gastric cancer, and epigenetic alterations can have a greater impact on cancer risk than genetic alterations for some types of cancers. The relative importance of genetic and epigenetic alterations could have great value in cancer prevention, including the search for carcinogenic factors and the mobilization of social resources. In addition, combining genetic and epigenetic alterations is expected to be useful for precision cancer risk diagnosis, and this approach is applicable to a broad range of cancer types. The mutation analysis also supported a long-postulated hypothesis that genetic alterations accumulate in normal-appearing tissues after exposure to risk factors and constitute cancer risk, which has been termed field cancerization (33, 34).

The difference in impacts of genetic and epigenetic alterations on cancer risk was in accordance with the effects of epidemiologically established lifestyle risk factors on ESCCs and gastric cancers (35, 36). As for ESCCs, alcohol drinking, betel chewing, and tobacco smoking have been established in the Taiwanese population (22, 37). Among these, tobacco contains multiple mutagens (23), and a high concentration of alcohol can produce acetaldehyde, which is mutagenic (38). At the same time, tobacco has been reported to be also associated with methylation changes in buccal tissue (39). In contrast, for gastric cancer, H. pylori infection is the major cause by far (36). H. pylori infection induces chronic inflammation, which is known to induce aberrant DNA methylation in gastric mucosae (8, 40, 41) and other tissues (6, 42). These differences of cancer-causing agents and their carcinogenic mechanisms, induction of mutation or methylation, can explain the difference in the relative importance of genetic and epigenetic alterations in the two tissues. The difference in the impacts was further in accordance with the fact that, even after extensive sequencing of gastric cancers, only limited numbers of genetic alterations of driver genes with low incidences were identified (43, 44).

Mutation signatures that reflect exposure to cancer lifestyle risk factors were observed in the two normal tissues. In esophageal mucosae, the mutation type associated with tobacco smoking and a hallmark signature of ESCC (C to A) (26, 45, 46) was slightly, but significantly, higher in the high-risk group with a severe smoking history than in the low-risk group. In contrast, the mutation type possibly associated with gastric acid reflux and a hallmark signature of esophageal adenocarcinoma (T to G) (47, 48) was barely observed. This was thought to be because esophageal mucosae analyzed in this study were from individuals in Taiwan, where ESCC is the major histological type by far, and because all the noncancerous esophageal mucosa samples were obtained from patients with ESCC. In gastric mucosae, the mutation signature of activation-induced cytidine deaminase was clearly observed, indicating the prominent effect of H. pylori infection on gastric mucosae. In addition, the mutation patterns of normal mucosae and cancers were similar both in the esophagus and in the stomach. Nonsynonymous mutations were significantly less frequent than synonymous mutations, regardless of the risk groups, in both esophageal and gastric mucosae. These low dN/dS ratios showed that nonsynonymous mutations were selected against. This experimentally supported the neutral theory of molecular evolution (49) for somatic mutations, which has been another long-postulated hypothesis.

The methylation markers used here were considered to reflect overall epigenome damage in individual tissues. To support this idea, methylation levels of multiple markers in a tissue were highly correlated (SI Appendix, Tables S7 and S8), suggesting that the markers measured a single entity. Also, their levels were clearly increased by exposure to lifestyle risk factors, but were not correlated with age. In addition, methylation of the marker CGIs did not appear to lead to any biological consequences. Some of the marker CGIs were located outside promoter regions, methylation of which does not necessarily lead to gene silencing (50). Even when a marker CGI was located in a promoter region, its downstream gene tended to have very low or no expression in the normal tissue, including HOXA9 (12, 29). These suggested that the marker CGIs were selected not because of their gene functions but because of their correlations with overall epigenome damage.

In summary, we demonstrated that epigenetic alterations can have greater impact on cancer risk than genetic alterations in some tissues. The different impacts of genetic and epigenetic alterations have the potential to be important for effective cancer prevention and precision cancer risk diagnosis.

Materials and Methods

Esophageal Mucosal Samples.

A total of 274 esophageal mucosa samples were collected endoscopically from adults who underwent cancer screening at the National Taiwan University Hospital from September 2008 to April 2013. The study was approved by the Research Ethics Committee C National Taiwan University Hospital (approval no. IRB200806039R), and written informed consent was obtained from all participants. The lifestyle ABC risk factors, whose very strong influence on the risk of ESCCs has been established (12, 22), were evaluated for patients based on interviews, as previously described (SI Appendix, Table S9) (12). The 274 samples were classified into three risk groups based on ABC risk factors and healthy/cancer statuses (SI Appendix, Table S1). The low-risk group was composed of 67 normal esophageal mucosae from healthy individuals without a history of ABC. The intermediate-risk group was composed of 96 normal-appearing esophageal mucosae from healthy individuals with a history of ABC. The high-risk group was composed of 111 noncancerous esophageal mucosae from ESCC patients with a history of ABC. The 274 samples were divided into three sample sets (screening set, 27 samples for the genome-wide DNA methylation analysis; validation set, 154 samples for the gene-specific DNA methylation analysis; and test set, 93 samples for the analysis of mutation frequency and methylation levels of definite risk marker genes). Further information on risk exposure levels is given in SI Appendix, Tables S10–S12.

Gastric Mucosal Samples.

A total of 96 gastric mucosa samples were collected endoscopically from the antral region of adults who underwent cancer screening at the Research Center for Cancer Prevention and Screening (51), National Cancer Center, Japan, and who underwent endoscopic submucosal dissection at the National Cancer Center Hospital (13), and were used for analysis of DNA methylation levels and cancer risk in our previous studies (13, 51). The previous studies and the current study were approved by the National Cancer Center Ethics Committee (approval nos. 2008-104, 2015-139), and written informed consent was obtained from all participants. Point mutations were newly analyzed in this study, and DNA methylation results and clinical information were obtained from the previous studies (13, 51).

The 96 samples were classified into three groups based on past H. pylori infection and healthy/cancer statuses (SI Appendix, Table S1). The low-risk group was composed of 32 normal gastric mucosae from healthy individuals without H. pylori infection. The intermediate-risk group was composed of 32 normal-appearing gastric mucosae from healthy individuals with past H. pylori infection (six or more months after the eradication of H. pylori). The high-risk group was composed of 32 noncancerous gastric mucosae from patients with gastric cancer with past H. pylori infection.

Both for the esophageal and gastric samples, the distributions of sexes and ages were comparable between the intermediate- and high-risk groups. In contrast, the low-risk group included more females for the esophageal samples (P < 0.001), and the low-risk group had younger ages compared with the other groups for gastric samples (P < 0.001).

Measurement of Point Mutations.

Rare base substitution mutations were measured by a method recently established: the 100-copy method (20). Briefly, a sequence library (291 regions of 55 cancer-related genes, covering 48,005 base positions) was prepared by multiplex PCR, using Ion AmpliSeq Library Kits 2.0 (Thermo Fisher Scientific) and 100 copies of genomic DNA as a template. Libraries from different samples were uniquely barcoded and sequenced using an Ion PI chip, Ion PI Hi-Q Sequencing 200 Kit, and Ion Proton sequencing system (Thermo Fisher Scientific) with an average sequencing depth of at least 5,000 reads. A base substitution variant with an allele frequency above a cutoff value (0.8%) was counted as a somatic mutation, and insertions and deletions were disregarded. Among the 48,005 bases amplified in the library, 15,313 base positions showed insufficient amplification (<2,500×), and 17,140 positions had a variant allele frequency ≥0.2% in any of the three analyses of the same sample (20). The remaining 15,552 base positions from 201 genomic regions (53 genes) were selected as error-resistant bases and were used to measure rare point mutations. The mutation frequency was estimated as the number of somatic mutations at 15,552 bases divided by 1,555,200 (=15,552 × 100). High reproducibility of the estimation of the frequency of rare mutations was shown in our previous study (20), and sequencing was conducted only once in this study.

Analysis of Mutation Patterns.

A mutation pattern was depicted using a six-substitution classification or a 96-substitution classification defined by the substitution type and sequence context immediately 5′ and 3′ to a mutated base. The similarity of mutation patterns was evaluated using cosine similarity (52). The proportions of specific mutation types were compared between groups using Pearson’s χ² test. Effects of isolated variants were estimated using Variant Effect Predictor (grch37.ensembl.org/Homo_sapiens/Tools/VEP). A dN/dS value was calculated as the actual ratio of nonsynonymous and synonymous mutations divided by the expected ratio of nonsynonymous and synonymous mutations with the assumption that all the base substitutions are expected to occur with the same probability. Expected consequences of base substitutions in the regions analyzed were determined as in our previous study (20).

Genome-Wide DNA Methylation Analysis.

A genome-wide DNA methylation analysis was performed using an Infinium HumanMethylation450 BeadChip Array (Illumina) covering 485,577 CpG sites. To adjust for probe design biases, intra-array normalization was conducted using a peak-based correction method, Beta Mixture Quantile dilation (53). The methylation level of a CpG site was represented by a beta value, which ranged from 0 (completely unmethylated) to 1 (completely methylated). Data obtained from the microarray have been deposited in the Gene Expression Omnibus (https://www.ncbi.nlm.nih.gov/geo, accession number GSE77991).

Gene-Specific DNA Methylation Analysis.

Genomic DNA was treated with sodium bisulfite and purified as previously described (29). qMSP was performed by real-time PCR, using bisulfite-modified DNA with primers specific to methylated DNA and primers universal to both methylated and unmethylated DNA (SI Appendix, Table S13). The methylation level was obtained as the number of methylated DNA molecules at a target locus divided by the number of all DNA molecules at a control (the RPPH1 gene) locus and normalized to the same measurement obtained using SssI-treated genomic DNA (percentage of the value of methylated DNA reference). High reproducibility of the methylation measurement was established previously (8), and methylation levels were measured once in this analysis. Deep bisulfite sequencing was performed using Ion PGM (Thermo Fisher Scientific) with bisulfite-modified DNA and universal primers (SI Appendix, Table S13), as previously described (54).

Isolation of Methylation Risk Markers for ESCC.

To isolate methylation risk markers for ESCC, two algorithms were used to compare methylation levels of normal esophageal mucosae between healthy individuals (low- and intermediate-risk groups) and patients with cancer (high-risk group). First, probes were screened to identify those with large differences between the low- and intermediate-risk groups and the high-risk group. Specifically, probes that had almost no methylation in normal mucosae (low- and intermediate-risk groups) and peripheral blood (beta value < 0.2) and high methylation in noncancerous mucosae (difference of median beta value > 0.2) were obtained.

Second, probes were screened to identify those with high variance as markers for a high-risk group by the iEVORA method, using the R script iEVORA.R (10, 28). Probes with a Bartlett’s test false discovery rate of less than 0.001 and an unadjusted P value of less than 0.05 based on a t test were selected.

From the CpG sites selected by the two screening methods, those within promoter CGIs or intragenic CGIs were further selected. Cross-reactive/polymorphic probes, as identified by Chen et al. (55), were removed. When three or more consecutive probes at a locus showed such differences, the locus was considered as a candidate cancer risk marker.

Statistical Analysis.

The mutation frequency and methylation levels were compared between groups, taking their distributions into account using Welch’s t tests and Mann–Whitney U tests, respectively. The correlations between mutation frequency and methylation levels for risk marker genes, among methylation levels for risk marker genes, and between age and methylation levels, were assessed by calculating the Pearson’s product-moment correlation coefficient. A receiver operating characteristics analysis was performed to determine an optimal cutoff value for a mutation frequency and a methylation level. The ranges of the ratio, such as dN/dS, and fraction of mutation signature, were assessed by interval estimation by calculating 95% confidence intervals. These statistical analyses were conducted using R version 3.2.1 (https://cran.r-project.org/). All statistical analyses were two-sided, and a P-value of less than 0.05 was considered to indicate statistical significance.

To study the effect of a marker on cancer risk, logistic regression analysis was performed based on individuals in the intermediate-risk group (controls) and the high-risk group (cases) in the test set. First, the effect of each methylation marker (or mutation frequency) was assessed in a univariate model, using previously determined cutoff values to define dichotomous variables corresponding to low/high levels of methylation (or mutation frequency). Second, a multivariate analysis including known (or traditional) risk factors (age and lifestyle risk factors), methylation levels, and mutation frequency was conducted. Improvement by an additional factor was assessed by the likelihood ratio test, and the performance of a model was assessed by the c-index (equivalent to the area under the curve). The relative importance of genetic and epigenetic alterations was estimated by calculation of the relative importance of the methylation levels compared with the mutation frequency based on the ratio of the corresponding standardized coefficients (56).