Across-Site Differences in the Mechanism of Alcohol-Induced Digestive Tract Carcinogenesis: An Evaluation by Mediation Analysis

Digestive tract cancers, including head and neck, esophageal, stomach, small intestine, and colorectal cancers, account for 27% of all cancer-related deaths worldwide (1, 2), and represent a major public health burden. Alcohol is a major risk factor for digestive tract cancers, and known to exhibit a dose-dependent effect on head and neck cancer, esophageal, and colorectal cancer risk (3). Although DNA damage by acetaldehyde, the first metabolite of alcohol, is one important hypothesized mechanism, several other possible mechanistic hypotheses for alcohol-mediated carcinogenesis have been proposed (4), and different mechanisms may contribute to alcohol-mediated carcinogenesis, possibly depending on cancer site.

Among aldehyde dehydrogenase (ALDH) isoforms, ALDH2 has primary responsibility for the oxidation of acetaldehyde (5). ALDH2 rs671 [c.1510G>A (p.Glu504Lys)] is a functional SNP, which is highly prevalent in East Asians (6). Its minor allele, the rs671 Lys allele, inactivates ALDH2 enzymatic activity, elevating acetaldehyde exposure and thereby likely increasing susceptibility to carcinogenesis in humans (6). Previous epidemiologic studies have shown that the ALDH2 rs671 Lys allele (ALDH2 Lys allele) increases susceptibility to head and neck cancer and esophageal cancer among drinkers due to strong gene–environment interaction between the Lys allele and alcohol drinking (7). The International Agency for Research on Cancer has therefore classified acetaldehyde associated with the consumption of alcoholic beverages as carcinogenic for head and neck cancer and esophageal cancer (8). Similar interaction between ALDH2 Lys allele and alcohol drinking is also reported in stomach cancer risk (9, 10). Accordingly, acetaldehyde appears important in the carcinogenesis of head and neck cancer, esophageal cancer, and stomach cancer. In contrast, acetaldehyde's role in colorectal carcinogenesis has not been fully elucidated (11), and no clear evidence for a gene–environment interaction between ALDH2 Lys allele and alcohol drinking in colorectal cancer risk is yet available.

Against this evidence, a paradoxically opposite effect for this polymorphism has also been observed. Namely, because individuals with the ALDH2 Lys allele experience a rapid accumulation of blood acetaldehyde after alcohol ingestion, this allele is known to be associated with decreased alcohol consumption (12, 13) due to increased exposure to the unpleasant effects of acetaldehyde (e.g., flushing, headache, palpitation, and nausea) meaning that this allele confers a protective effect against alcohol-induced carcinogenesis. The impact of drinking amount as a surrogate for acetaldehyde exposure (direct effect of ALDH2 Lys allele) therefore appears attenuated by the indirect effect mediated by reduced drinking intensity. However, because previous case–control studies on digestive tract cancer risk used conventional multivariate analysis and did not fully consider possible mediation effects from ALDH2 Lys allele on digestive tract cancer risk through drinking intensity, it remains unclear whether and to what extent the ALDH2 Lys allele operates through the direct and/or indirect pathways in the development of these cancers.

To separate these two opposing effects of ALDH2 Lys allele on the development of digestive tract cancers, we conducted case–control studies for five digestive tract cancers and applied the mediation analysis introduced by VanderWeele and colleagues (14). These authors showed that the genetic variants on 15q25.1, which are related to both lung cancer risk and smoking behavior, were associated with lung cancer risk primarily through direct pathways, independent of smoking behavior. Here, by quantifying the respective direct and indirect effects of the ALDH2 Lys allele on digestive tract cancer risk according to site, we aimed to elucidate potential site-dependent heterogeneity in the attribution of alcohol- or acetaldehyde-induced carcinogenesis.

The study sample was selected from participants in the Hospital-based Epidemiologic Research Program at Aichi Cancer Center (HERPACC)-2 (2001–2005) and HERPACC-3 (2005–2013). These studies have been detailed elsewhere (15, 16). We conducted a case–control study for each cancer using cases and age- and sex-matched controls: head and neck cancer with 748 and 1,113, esophageal cancer with 617 and 882, stomach cancer with 1,491 and 2,187, small intestine cancer with 29 and 111, and colorectal cancer with 1,214 and 1,772, respectively. Cases were first-visit outpatients at Aichi Cancer Center Hospital, who were diagnosed with each cancer, and controls were first-visit outpatients who were confirmed to have no cancer or history of neoplasm. We defined noncancer first-visit outpatients as the population among whom cases may arise under the assumption that they will visit the Aichi Cancer Center Hospital, if they develop cancer in the future (15), indicating that controls within HERPACC were appropriately selected from the source population (17). In addition, as shown in Supplementary Table S1 and Supplementary Fig. S1, the randomly sampled controls did not substantially differ between studies in the distribution of hospital unit they first visited, indicating that the appropriateness of the controls was likely to hold in each study. All participants gave written informed consent, completed a self-administered questionnaire, and provided a peripheral blood sample. The study was conducted in accordance with the Declaration of Helsinki and approved by the Institutional Ethics Committee of Aichi Cancer Center. As we have already reported on the mediation effect between the ALDH2 Lys allele and drinking intensity on stomach cancer risk (18), we reanalyzed the results for this cancer using the updated dataset.

Head and neck cancer included the following the International Classification of Diseases for Oncology, Third Edition (ICD-O-3) codes (19): C00–C14 and C30–C32. Similarly, cancers of the oral cavity and oropharynx included ICD-O-3 C00–C06, C09, and C10 (except C00.0, C00.1, and C00.2). Cancers of the hypopharynx and larynx included ICD-O-3 C12, C13, and C32. Esophageal cancer included ICD-O-3 C15. Cancers of the upper/middle esophagus included ICD-O-3 C15.0, C15.1, C15.3, and C15.4. Cancers of the lower esophagus included ICD-O-3 C15.2, and C15.5. Stomach cancer included ICD-O-3 C16. Cancers of the cardia stomach included ICD-O-3 C16.0. Cancers of the noncardia stomach included ICD-O-3 C16.1–C16.6. Small intestine cancer included ICD-O-3 C17. Colorectal cancer included ICD-O-3 C18–C20. Cancers of the proximal (right-sided) colon included ICD-O-3 C18.0–C18.5. Cancers of the distal (left-sided) colon included ICD-O-3 C18.6 and C18.7. There is an inconsistency between studies regarding inclusion of the right or left subsite in the case of C18.5 (splenic flexure of the colon; ref. 20). In this study, we included C18.5 (n = 1) in the proximal (right-sided) colon, in accordance with our previous study (21). Cancers of the rectum included ICD-O-3 C19 and C20.

Each participant's DNA was extracted from the buffy coat fraction with a DNA Blood Mini Kit (Qiagen). ALDH2 rs671 was genotyped using TaqMan Assays on a 7500 Real-Time PCR System (Applied Biosystems). Genotype distribution in the control was assessed for Hardy–Weinberg equilibrium using the χ² test.

Information on environmental risk factors was collected via a self-administered questionnaire, which asked the participants to answer questions on their exposure status before the development of the symptoms for which they first visited the hospital. Trained interviewers checked for the presence of incomplete responses and inconsistencies between responses.

Daily alcohol intake (g/day) was used as a measure of drinking intensity and calculated using the information on the frequency of alcohol drinking and the total amount of pure alcohol consumption during each drinking session. In the questionnaire related to alcohol consumption in HERPACC-2 and -3, the participants were first asked about drinking status, as never (including “almost never”), former, or current drinkers. Former and current drinkers were then asked about the frequency of alcohol drinking (HERPACC-2: <1 day/week, 1–2 days/week, 3–4 days/week, or ≥5 days/week; HERPACC-3: 1–3 days/month, 1–2 days/week, 3–4 days/week, 5–6 days/week, or everyday), along with type of beverage (Japanese sake, beer, shochu, whiskey, and wine) and average consumption during each drinking session. In particular, we required the participants to answer the latter three questions (frequency of alcohol drinking, type of beverage, and average consumption during each drinking session) based on their average drinking behavior during the year before the development of the symptoms or diagnosis. To calculate daily alcohol intake (g/day), each category of drinking frequency was assigned a score as follows: 0 for never drinkers, 0.5 for <1 day/week, 1.5 for 1 to 2 days/week, 3.5 for 3 to 4 days/week, and 6 for ≥5 days/week in HERPACC-2; and 0 for never drinkers, 0.5 for 1 to 3 days/month, 1.5 for 1 to 2 days/week, 3.5 for 3 to 4 days/week, 5.5 for 5 to 6 days/week, and 7 for every day in HERPACC-3. Alcohol intake per drinking session was estimated on the basis of the concentration of ethanol in each beverage. In the HERPACC study, one unit of drink is assumed to contain 23 g of ethanol for 180 mL (one “go”) of Japanese sake, 633 mL (one large bottle) of beer, 90 mL (a half “go”) of shochu (distilled spirit), 60 mL (double shot) of whiskey, and 200 mL (two and a half glasses) of wine. We therefore estimated alcohol intake per drinking session by multiplying the summed units of each type of beverage by 23. Finally, daily alcohol intake was estimated by multiplying the alcohol intake per drinking session by the frequency score/7.

Cumulative smoking exposure was evaluated as pack-years, calculated by multiplying the number of packs consumed per day by the number of years of smoking. Physical activity was evaluated as metabolic equivalent (MET) hours per week (22), calculated by frequency, intensity, and the amount of time per session. Participants were divided into three groups according to the distribution of MET hours per week among controls (tertiles). Body mass index (BMI) was calculated as the self-reported weight in kilograms divided by the square of the height in meters. Energy-adjusted fruit/vegetable intake was estimated by the residual method (23, 24) using information from a validated food frequency questionnaire (25). For fruit/vegetable intake and total energy intake, participants were classified into three groups according to the distributions of the respective factors among controls (tertiles). The frequencies of meat (beef or pork) intake and processed meat intake were classified into three categories: <1, 1 to 4, and ≥5 times/week. A family history of each cancer in parents and siblings and history of diabetes were coded as “yes” or “no,” based on self-reporting. For Helicobacter pylori (H. pylori) infection status, plasma IgG antibody levels for H. pylori were measured by a direct ELISA Kit (‘E Plate “Eiken” H. pylori Antibody'; Eiken Kagaku), with anti-H. pylori IgG > 10 U/mL regarded as a positive infection. For atrophic gastritis status, serum pepsinogens (PG) levels were measured by chemiluminescence enzyme immunoassay, with PG I ≤70 ng/mL and PG I/PG II ≤3 regarded as a positive gastric mucosal atrophy.

Differences in the distribution of risk factors between cases and controls were evaluated using the t test or χ² test. Frequencies of never/ever drinkers across ALDH2 genotypes were examined using Fisher exact test. Among ever drinkers, alcohol intake across ALDH2 genotypes was examined using the nonparametric Kruskal–Wallis test. Before conducting mediation analyses, we estimated ORs and their 95% confidence intervals (CI) for a 10-g change in daily alcohol intake in a conditional logistic regression model, including age (years), pack-years, physical activity (tertiles), BMI (kg/m²), fruit/vegetable intake (tertiles), total energy intake (tertiles), frequency of meat intake (<1, 1–4, ≥5 times/week), frequency of processed meat intake (<1, 1–4, ≥5 times/week), family history of each cancer (yes/no), history of diabetes (yes/no), and ALDH2 rs671 genotype (Glu/Glu, Glu/Lys, Lys/Lys) as covariates. The variant-drinking interaction was evaluated on both the additive and multiplicative scales by including the interaction term corresponding to a 1-allele change in the genetic variant multiplied by a 10-g change in daily alcohol intake into the model. We also evaluated the smoking-drinking interaction on both the additive and multiplicative scales by including the interaction term of a 10-pack-year change in smoking exposure multiplied by a 10-g change in daily alcohol intake into the conditional logistic regression model, to address the impact of the synergistic effect of smoking and alcohol drinking on each cancer risk. Test for additive interaction was performed using the relative excess risk due to interaction (26), whereas multiplicative interaction was assessed using a Wald test of the coefficient of the interaction term.

Then, we performed mediation analyses to decompose the total effect of ALDH2 Lys allele on each cancer risk into direct and indirect effects using the Stata command paramed (27). In Fig. 1A, we show a directed acyclic graph (DAG) depicting a hypothetical causal structure between exposure, mediator, and outcome. Total-effect OR (tOR) can be written as the product of direct-effect OR (dOR) and indirect-effect OR (iOR; ref. 28). As shown in Fig. 1B, the total effect of ALDH2 Lys allele on cancer risk is the contrast between the outcome in those with the ALDH2 Lys allele and the outcome in those without the ALDH2 Lys allele (Glu/Glu; reference group). As shown in Fig. 1C, the direct effect is the contrast between the outcome in those with the ALDH2 Lys allele under the counterfactual situation of drinking intensity being the same as what it would be without the ALDH2 Lys allele (Glu/Glu) and the outcome in those without the ALDH2 Lys allele (Glu/Glu; reference group). Finally, as shown in Fig. 1D, the indirect effect is the contrast between the outcome in those with the ALDH2 Lys allele and the outcome in those with the ALDH2 Lys allele under the counterfactual situation of drinking intensity being the same as what it would be without the ALDH2 Lys allele (Glu/Glu) (reference group). The direct effect can be interpreted as the effect of ALDH2 Lys allele on each cancer risk through pathways independent of drinking intensity (Fig. 1C), and the indirect effect as the effect mediated through drinking intensity (Fig. 1D).

In the mediation analyses, ALDH2 genotypes of Glu/Lys and Lys/Lys were collectively treated as ALDH2 Lys allele and defined as the presence of exposure. Details of mediation analysis technique are described elsewhere (14, 28, 29). In brief, we estimated the tOR, dOR, and iOR by combining two models: the linear regression model for the mediator (i.e., drinking intensity) conditional on the exposure and covariates (age, sex, pack-years, physical activity, BMI, fruit/vegetable intake, total energy intake, frequency of meat intake, frequency of processed meat intake, family history of each cancer, history of diabetes, and HERPACC version); and the logistic regression model for the outcome conditional on the exposure, mediator, and covariates. Given that this mediation analysis technique relies on parametric modeling assumptions, drinking intensity [daily alcohol intake (g/day)] was entered in the analysis as its square root to better approximate a linear fit. To obviate concerns about residual confounding by smoking, we also performed this analysis in never smokers only. Moreover, to investigate whether the direct and indirect effect of the ALDH2 Lys allele was homogeneous within strata of smoking intensity, we conducted analyses stratified by smoking intensity (0, <20, 20 to <40, and ≥40 pack-years). Between-strata heterogeneity was evaluated using the Q and I² statistic (30). The Q statistic was considered statistically significant when P < 0.10, whereas 0% of the I² statistic represented no heterogeneity.

Some data for daily alcohol intake, pack-years, physical activity, BMI, fruit/vegetable intake, total energy intake, frequency of meat intake, frequency of processed meat intake, and history of diabetes, were missing. We therefore performed multiple imputation using chained equations (31). Ten imputed data sets were generated, and the results were combined using Rubin's rules (32). All analyses were carried out using Stata version 15 (Stata Corporation). We interpreted two-sided P values of <0.05 as statistically significant.

Supplementary Table S2 shows the baseline characteristics of participants. Genotype frequencies among controls did not deviate from values predicted from the Hardy–Weinberg equilibrium. To demonstrate the relationship between exposure (ALDH2 genotype) and mediator (drinking intensity), Supplementary Table S3 shows the frequencies of never/ever drinkers and the amount of alcohol intake among ever drinkers across ALDH2 genotypes in control subjects stratified by sex. Supplementary Table S4 shows the sample size cross-tabulated by the exposure and mediator as well as the mediator and outcome. Figure 2 shows adjusted ORs per 10-g increase in daily alcohol intake, along with tests for variant-drinking interaction. Significant dose-dependent association was observed between alcohol intake and risk of head and neck cancer (OR = 1.13; 95% CI, 1.09–1.18), esophageal cancer (OR = 1.46; 95% CI, 1.36–1.57), stomach cancer (OR = 1.03; 95% CI, 1.01–1.06), and colorectal cancer (OR = 1.06; 95% CI, 1.02–1.10), but not risk of small intestine cancer (OR = 0.99; 95% CI, 0.76–1.28). On stratification by subsite, we observed significant dose-dependent associations of alcohol intake with risk of cancers of the oral/oropharynx, hypopharynx/larynx, upper/middle esophagus, lower esophagus, noncardia stomach, and distal colon, and a marginally significant dose-dependent association for rectal cancer risk (P = 0.081). However, we also observed nonsignificant dose-dependent associations between alcohol intake and risk of cardia stomach and proximal colon cancers. Tests for interaction were nonsignificant on either the additive or multiplicative risk scales in cancers of the cardia stomach, small intestine, colorectum, proximal colon, distal colon, and rectum (P values for both the additive and multiplicative interactions >0.1; Fig. 2). We therefore used mediation analysis without considering exposure–mediator interaction when evaluating cancers of the cardia stomach, small intestine, colorectum, proximal colon, distal colon, and rectum; when evaluating cancers of the head and neck, oral/oropharynx, hypopharynx/larynx, esophagus, upper/middle esophagus, lower esophagus, stomach, and noncardia stomach, we considered exposure–mediator interaction in the mediation analyses.

The estimated tOR, dOR, and iOR of the ALDH2 Lys allele on each cancer risk are shown in Supplementary Table S5 and Fig. 3. For head and neck cancer, esophageal cancer, and stomach cancer, significant positive dORs were observed even when analysis was restricted to never smokers. In contrast, nonsignificant dORs were observed for small intestine cancer and colorectal cancer. Notably, ALDH2 Lys allele was directly associated with a large increase in esophageal cancer risk (dOR = 21.15; 95% CI, 9.11–49.12). Analysis by subsite suggested the heterogeneity of the magnitude of the direct effect among subsites in stomach cancer and head and neck cancer. There was a significant direct effect for the risk of noncardia stomach cancer (dOR = 1.92; 95% CI, 1.49–2.46), but no such effect for the risk of cardia stomach cancer (dOR = 0.90; 95% CI, 0.44–1.83). For head and neck cancer, the dOR for cancer of the oral/oropharynx was 1.37 (95% CI, 1.03–1.82), whereas that for the hypopharynx/larynx cancer was 3.49 (95% CI, 2.01–6.05). With regard to indirect effects, we observed a significant protective effect of ALDH2 Lys allele on risk of all cancers except small intestine cancer, namely head and neck cancer (iOR = 0.67; 95% CI, 0.59–0.76), esophageal cancer (iOR = 0.16; 95% CI, 0.12–0.22), stomach cancer (iOR = 0.81; 95% CI, 0.74–0.89), and colorectal cancer (iOR = 0.87; 95% CI, 0.81–0.94). For never smokers, a significantly protective indirect effect was only observed in risk of head and neck cancer and esophageal cancer, albeit that the point estimates showed a protective indirect effect for all cancer risk. Analysis by subsite suggested heterogeneity in the magnitude of the indirect effect, especially in head and neck cancer and colorectal cancer. In head and neck cancer, the iORs for cancer of the oral/oropharynx and hypopharynx/larynx were 0.79 (95% CI, 0.69–0.90) and 0.46 (95% CI, 0.34–0.62), respectively. In colorectal cancer, significant protective iORs were observed for distal colon (iOR = 0.79; 95% CI, 0.69–0.92) and rectal (iOR = 0.88; 95% CI, 0.79–0.99) cancers, versus no effect for proximal colon cancer risk (iOR = 0.96; 95% CI, 0.83–1.12). In the analysis of overall stomach cancer risk, the dOR and iORs in this study were similar to those in the previous study (Fig. 3; ref. 18). Given that H. pylori infection and atrophic gastritis are well-known risk factors for stomach cancer (33, 34), we further included these two factors as covariates, but the results did not change substantially (Supplementary Table S6). Further analysis by treating ALDH2 Glu/Lys and Lys/Lys separately suggested a greater increase in dORs for head and neck cancer and stomach cancer and a greater decrease in iORs for head and neck cancer, stomach cancer, and colorectal cancer in ALDH2 Lys/Lys compared with ALDH2 Glu/Lys, albeit that most estimates were nonsignificant (Supplementary Table S7).

To address the impact of the synergistic effect of smoking and alcohol drinking on each cancer risk, we performed two analyses of smoking–drinking interaction (Supplementary Table S8) and stratification by smoking intensity (Supplementary Table S9). Regarding head and neck cancer and esophageal cancer, the multiplicative interaction was significantly negative, whereas the additive interaction was significantly positive. However, these point estimates in both multiplicative and additive scales were close to the respective null values of 1.0 and 0 (head and neck cancer, 0.99 and 0.012, respectively; esophageal cancer, 0.98 and 0.078, respectively). Regarding stomach cancer, small intestine cancer, and colorectal cancer, there was no significant interaction in either multiplicative or additive scale. In addition, we observed consistent results for dOR and iOR across each stratum for head and neck cancer, esophageal cancer, stomach cancer, and colorectal cancer (Supplementary Table S9). The test for between-strata heterogeneity suggested that there was no significant heterogeneity in either dOR or iOR for all cancers, except for iORs for esophageal cancer.

We performed mediation analyses to estimate the direct and indirect effects of the ALDH2 Lys allele on digestive tract cancer risk in a case–control study with 4,099 cases and 6,065 age- and sex-matched controls. Significant dORs were observed for head and neck cancer, esophageal cancer, and stomach cancer, but the effects varied substantially by site. In contrast, nonsignificant dORs were observed for small intestine cancer and colorectal cancer. Regarding indirect effects, we observed a significant protective effect on all cancer sites except for small intestine.

Alcohol absorption starts in the upper digestive mucosa, and happens primarily in the stomach and small intestine (35, 36). We showed a significantly increased direct effect of ALDH2 Lys allele for head and neck cancer, esophageal, and stomach cancer risk, even when restricted to never smokers, supporting the idea that the putative carcinogenic effect of alcohol-derived acetaldehyde may exist only in those organs that are locally exposed to alcohol. Although acetaldehyde production occurs mainly in the liver, acetaldehyde formation starts in the mouth and continues along the digestive tract (35). In fact, drinkers with the ALDH2 Lys allele are considered to be exposed to two to three times (via saliva) and five to six times (via gastric juice) higher local acetaldehyde levels than those without this allele (37). Moreover, one experimental study showed higher levels of aldehyde-derived DNA adducts in drinkers with the ALDH2 Lys allele than in those without ALDH2 Lys allele, indicating the link between the ALDH2 Lys allele and aldehyde-derived DNA damage in humans (38). In addition, acetaldehyde-derived DNA adducts are more frequent in Aldh2-knockout than wild-type mice treated with ethanol in the esophagus (39) and stomach (40). The Cancer Genome Atlas data showed that mutational signature 16, which is related to alcohol consumption (41), was more frequently observed among those with than without the ALDH2 Lys allele in esophageal cancer (P < 0.0001) and stomach cancer (P = 0.026; ref. 42), indicating that acetaldehyde may be the actual cause of somatic mutations in upper gastrointestinal tract cancers.

To date, the association of alcohol and acetaldehyde with stomach cancer risk has been inconclusive (35). By showing significant dOR of the ALDH2 Lys allele, however, we consider that our study provides additional evidence for the involvement of acetaldehyde in stomach cancer carcinogenesis, which is in line with previous epidemiologic evidence on ALDH2 polymorphisms (9, 10). Nevertheless, the sample size may not be large enough to make a significant contribution to this issue, and a conclusive investigation will require additional data.

Of note, the direct effect of ALDH2 Lys allele was heterogenous between head and neck cancer and esophageal cancer, although acetaldehyde is classified as a group 1 carcinogen for both (8). Our results also demonstrated that both the significant direct and indirect effects disappeared in oral/oropharyngeal cancer risk when restricted to never smokers (Fig. 3), suggesting heterogeneity in the magnitude of the association of alcohol consumption with carcinogenesis between these cancers. A meta-analysis of 28 case–control studies reported that smoking was a stronger risk factor than drinking for head and neck cancer (43). A study evaluating somatic mutations of head and neck cancer and esophageal cancer reported that mutational signature 16 accounted for 7.1% and 16.5% of total mutations, respectively (41). Further, Aldh2-knockout mice treated with intraperitoneal ethanol injection for 8 weeks showed an elevated esophageal acetaldehyde-derived DNA adduct level over control mice, suggesting that circulating ethanol-derived acetaldehyde causes esophageal DNA damage (44). Therefore, esophageal carcinogenesis might involve both local and circulating acetaldehyde. This may explain the larger direct effect of ALDH2 Lys allele on esophageal cancer risk, at least in part.

An East Asian-specific SNP of ALDH2, rs671, may explain geographic differences in the incidence and etiology of some cancers. Geographic differences in esophageal cancer incidence are well known: over half of cases occur in East Asians (1, 2, 45). Moreover, the predominant histologic type of esophageal cancer varies by ethnicity; squamous cell carcinoma is predominant in East Asians, whereas adenocarcinoma, with obesity and gastroesophageal reflux disease as major risk factors, is more common in Caucasians (45). Furthermore, it was experimentally demonstrated that acetaldehyde-induced overexpression of wild-type ALDH2 reduces acetaldehyde-derived DNA damage in human esophageal keratinocytes, suggesting that esophageal wild-type ALDH2 is protective against the carcinogenic effect of alcohol on esophageal epithelium (39). For stomach cancer, the World Cancer Research Fund (WCRF) and American Institute of Cancer Research (AICR) concluded that a positive association between alcohol drinking and stomach cancer risk was probable, but noted that this was largely on Asian population studies (46). A recent large prospective cohort study involving mostly non-Hispanic whites showed no association of alcohol with either cardia or noncardia stomach cancer risk (47). The geographic differences in the incidence and etiology of esophageal cancer and stomach cancer might therefore be at least partly due to the different allele frequencies of the ALDH2 Lys allele.

We observed a protective indirect effect on all cancer sites except for small intestine. Interestingly, this notable behavior-related effect of the ALDH2 Lys allele reduced the risk of most digestive tract cancers in this population. In particular, on esophageal cancer, this mediation analysis showed that the tOR of ALDH2 Lys allele (tOR = 3.45; 95% CI, 1.58–7.54; Supplementary Table S5) can be decomposed into dOR of 21.15 (95% CI, 9.11–49.12) and iOR of 0.16 (95% CI, 0.12–0.22; Fig. 3), revealing that the direct effect was largely masked by the protective indirect effect. Moreover, although a recent meta-analysis showed a protective effect of ALDH2 Lys allele on colorectal cancer risk (48), our study revealed that this effect is mainly due to decreased drinking intensity (indirect effect), without any direct effect. This may be partially explained by a differential level of ALDH2 expression between the upper and lower gastrointestinal tracts, which was suggested by one experimental study showing a much higher rate of ALDH2 expression in stomach than in colon when evaluating the expression of ALDHs in mouse tissues (49). Nevertheless, considering that ALDH2 has by far the highest affinity for acetaldehyde (K_m < 1 μmol/L) among ALDH isoforms and has been considered to be probably the only isoform that significantly contributes to acetaldehyde metabolism in humans (5, 6), we speculate that acetaldehyde may not play a large role in colorectal carcinogenesis.

Several previous epidemiological studies suggested a synergistic effect of smoking and alcohol drinking on head and neck cancer and esophageal cancer risk (50). However, the mediation analysis in this study assumes no interaction between the covariates and the exposure or mediator (28), meaning that we cannot include the product term of smoking (the covariate) and alcohol drinking (the mediator). To address this concern regarding a synergistic effect, we performed two analyses of smoking–drinking interaction (Supplementary Table S8) and stratification by smoking intensity (Supplementary Table S9). In the smoking–drinking interaction analysis, the point estimates were close to the null value in both multiplicative and additive scales for head and neck cancer and esophageal cancer (Supplementary Table S8), indicating that the impact of their synergistic effect might be small on head and neck cancer and esophageal cancer risk. In addition, the most recent comprehensive review of evidence by the WCRF/AICR noted that the results for interaction between the two factors in head and neck cancer were inconsistent and that the number of cases was limited (51). With regard to esophageal cancer, although several case–control studies showed a significant interaction between the two factors, no cohort studies have shown a synergistic effect, at least in the multiplicative scale (52). Therefore, even though their synergistic effect had some impact on our estimates, these recent findings and the small values of the point estimates observed in our interaction analysis indicate that any impact of the synergistic effect would likely be relatively small or negligible. We therefore consider that the synergistic effect between smoking and alcohol drinking might have had little or no substantial effect on our results. Moreover, the finding that the stratified analyses by smoking intensity had consistent results for dOR and iOR across each stratum for most digestive tract cancers further supports the robustness of our results (Supplementary Table S9).

As suggested in Supplementary Table S3, the effects of alcohol consumption substantially varied depending on whether the subject carried the Glu/Lys or Lys/Lys genotype. Although the estimates on cancers of the esophagus and small intestine included only a small number of cases with ALDH2 Lys/Lys and were accordingly unstable, and most estimates were nonsignificant when the analysis treated ALDH2 Glu/Lys and Lys/Lys separately (Supplementary Table S7), a greater increase in dORs for head and neck cancer and stomach cancer and a greater decrease in iORs for head and neck cancer, stomach cancer, and colorectal cancer with ALDH2 Lys/Lys than with ALDH2 Glu/Lys might reflect higher acetaldehyde accumulation caused by the inability of acetaldehyde detoxification with ALDH2 Lys homozygosity. However, the sample size of those homozygous for ALDH2 Lys was small, and conclusive investigation of the effects on alcohol consumption with this genotype requires additional data.

Among methodologic strengths of our study, we conducted case–control studies of multiple digestive tract cancers in one dataset with many subjects, with detailed information on potential confounders and alcohol consumption. Further, the following characteristics of the study are noteworthy: (i) quantification of two opposing effects of ALDH2 Lys allele on multiple digestive tract cancer risks; (ii) detection of across-site differences in the mechanism of alcohol-induced digestive tract carcinogenesis through epidemiologic study; (iii) demonstration of the protective effect of ALDH2 Lys allele on most digestive tract cancers; and (iv) exposure of the direct effect, which was masked by the protective indirect effect on risk of head and neck cancer, esophageal cancer, and stomach cancer.

Several potential limitations also warrant mention. First, the use of self-administered questionnaires raises the possibility of recall bias, albeit that we collected lifestyle information before the first examination, limiting this bias. Second, unmeasured covariates may have also introduced bias. Mediation analyses assume that conditional on the covariates, there are (i) no exposure–outcome confounders; (ii) no mediator–outcome confounders; (iii) no exposure–mediator confounders; and (iv) no mediator–outcome confounding affected by the exposure (14). In this study, assumptions (i) and (iii) hold because the exposure was a genetic variant in a single ethnic group. Assumption (iv) also likely holds given the specific function of this variant in alcohol drinking. However, whether assumption (ii) is upheld remains unclear. Nevertheless, we controlled for the most important potential confounders, and therefore consider that assumption (ii) does hold, partially at least. Third, our findings were based on a genetic epidemiologic study and were unable to directly demonstrate that the accumulation of acetaldehyde with the ALDH2 Lys allele was the cause of carcinogenicity. Although the ALDH2 Lys allele is considered to largely contribute to the accumulation of toxic alcohol-derived acetaldehyde, other toxic effects through other possible pathways may also contribute to the estimated direct effect of the ALDH2 Lys allele, given that ALDH2 is also capable of oxidizing both endogenous aldehyde and exogenous aldehyde (contained in tobacco smoke, car exhaust, etc.; refs. 6, 53). Nevertheless, on the basis of both biological plausibility (4, 38) and the variant-drinking interaction consistently observed in previous genetic epidemiologic studies (7), we consider that the accumulation of acetaldehyde with the ALDH2 Lys allele is a reasonable and most likely mechanism for subsequent carcinogenesis. Whether other toxic effects, including endo- and exogenous aldehydes, are involved in the carcinogenesis of upper gastrointestinal tract cancer warrants further exploration. Fourth, among our findings, we did not observe any significant ORs for drinking intensity or direct-/indirect-effect estimates on the three cancer subsites of cardia stomach, small intestine, and proximal colon. The respective sample sizes (case/control) for these cancers were 87 of 108, 29 of 111, and 255 of 369, and the nonsignificant results for them were accordingly considered partially due to the small sample size. In particular, regarding small intestine cancer, some studies have suggested an association of alcohol drinking with small intestine cancer risk (54, 55), although this remains inconclusive (3). Because of small intestine cancer is rare, the results obtained for this tumor should be interpreted with caution and conclusive investigation awaits additional data. On the other hand, our results for cardia stomach and proximal colon cancers are consistent with those of previous studies in large populations. One meta-analysis (56) and three large cohort studies (47, 57, 58) showed no association between alcohol drinking and cardia stomach cancer risk, and two (56, 57) showed positive association for noncardia stomach cancer risk. With regard to proximal colon cancer, which has been considered to have a different etiology to distal colon cancer (59), the evidence of an association between alcohol drinking and colon cancer by anatomical subsite suggested a stronger association for distal colon cancer than for proximal colon cancer (60, 61). We consider that such consistency of results between these previous and our present studies supports the plausibility of our findings on these cancers. Nevertheless, studies evaluating the association between alcohol drinking and cancers of the cardia stomach and proximal colon are limited. In addition, studies evaluating the association of ALDH2 genotypes with these cancers are scarce. Therefore, further investigation of the association of ALDH2 genotype and alcohol drinking with the risk of cancers of cardia stomach, small intestine, and proximal colon are warranted. Fifth, our questionnaire related to alcohol consumption did not include binge drinking. Several studies have suggested that binge drinking is associated with an increased risk of some cancers (62–64). One case–control study reported that binge drinking was associated with an increased risk of pancreatic cancer even after adjustment for lifetime alcohol consumption (62). Assessment of the impact of binge drinking is therefore important, and the lack of information on this is a limitation of our study. Moreover, considering the impact of the ALDH2 Lys allele on drinking behavior, individuals with the ALDH2 Lys allele are likely to feel even sicker than those without the ALDH2 Lys allele (Glu/Glu) following binge drinking episodes, which has in fact been suggested by a few studies, specifically studies on college students (65). To our knowledge, however, this relationship between ALDH2 genotype and binge drinking has not been fully evaluated. More detailed investigation of the impact of this allele on drinking behavior is necessary.

In conclusion, alcohol is associated with an increased risk of most digestive tract cancers, but its impact as a surrogate for acetaldehyde exposure appears heterogeneous, supporting the idea that the mechanisms of cancer induction by alcohol differ by site. Moreover, our mediation analyses highlight the protective effect of ALDH2 Lys allele through drinking behavior, suggesting the importance of encouraging individuals without the ALDH2 Lys allele (Glu/Glu) to reduce alcohol consumption from a preventive viewpoint.

No potential conflicts of interest were disclosed.

Conception and design: Y.N. Koyanagi, K. Matsuo

Development of methodology: Y.N. Koyanagi, E. Suzuki, K. Matsuo

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): Y.N. Koyanagi, Y. Kasugai, I. Oze, M. Iwase, Y. Usui, Y. Kawakatsu, Y. Hirayama, T. Tanaka, T. Abe, S. Ito, K. Komori, N. Hanai, M. Tajika, Y. Shimizu, Y. Niwa, K. Matsuo

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): Y.N. Koyanagi, Y. Kasugai, I. Oze, T. Ugai, M. Iwase, Y. Usui, Y. Kawakatsu, H. Ito, K. Matsuo

Writing, review, and/or revision of the manuscript: Y.N. Koyanagi, E. Suzuki, I. Imoto, I. Oze, T. Ugai, M. Iwase, Y. Usui, Y. Kawakatsu, M. Sawabe, S. Ito, K. Komori, N. Hanai, M. Tajika, H. Ito, K. Matsuo

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): Y.N. Koyanagi, S. Ito, K. Komori, K. Matsuo

Study supervision: K. Matsuo

This study was supported by grants-in-aid for Scientific Research from the Ministry of Education, Science, Sports, Culture and Technology of Japan, consisting of Priority Areas of Cancer (No. 17015018 to K. Matsuo), Innovative Areas (No. 221S0001 to K. Matsuo and H. Ito), and JSPS KAKENHI Grants (No. JP26253041 to K. Matsuo and H. Ito, No. JP15H02524 to K. Matsuo, No. JP16H06277 to K. Matsuo, No. JP18H03045 to K. Matsuo, H. Ito, and I. Oze, and No. JP19K19425 to Y.N. Koyanagi), and by a grant-in-aid for the third term comprehensive 10-year strategy for cancer control from the Ministry of Health, Labour and Welfare of Japan (to K. Matsuo).

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.