DNA methylation at AHRR as a master predictor of smoke exposure and a biomarker for sleep and exercise

Abstract

Background

DNA methylation profiling may provide a more accurate measure of the smoking status than self-report and may be useful in guiding clinical interventions and forensic investigations. In the current study, blood DNA methylation profiles of nearly 800 Polish individuals were assayed using Illuminia EPIC and the inference of smoking from epigenetic data was explored. In addition, we focused on the role of the AHRR gene as a top marker for smoking and investigated its responsiveness to other lifestyle behaviors.

Results

We found > 450 significant CpGs associated with cigarette consumption, and overrepresented in various biological functions including cell communication, response to stress, blood vessel development, cell death, and atherosclerosis. The model consisting of cg05575921 in AHRR (p = 4.5 × 10^–32) and three additional CpGs (cg09594361, cg21322436 in CNTNAP2 and cg09842685) was able to predict smoking status with a high accuracy of AUC = 0.8 in the test set. Importantly, a gradual increase in the probability of smoking was observed, starting from occasional smokers to regular heavy smokers. Furthermore, former smokers displayed the intermediate DNA methylation profiles compared to current and never smokers, and thus our results indicate the potential reversibility of DNA methylation after smoking cessation. The AHRR played a key role in a predictive analysis, explaining 21.5% of the variation in smoking. In addition, the AHRR methylation was analyzed for association with other modifiable lifestyle factors, and showed significance for sleep and physical activity. We also showed that the epigenetic score for smoking was significantly correlated with most of the epigenetic clocks tested, except for two first-generation clocks.

Conclusions

Our study suggests that a more rapid return to never-smoker methylation levels after smoking cessation may be achievable in people who change their lifestyle in terms of physical activity and sleep duration. As cigarette smoking has been implicated in the literature as a leading cause of epigenetic aging and AHRR appears to be modifiable by multiple exogenous factors, it emerges as a promising target for intervention and investment.

Background

Cigarette consumption is associated with numerous adverse health effects and accounts for more than 8 million deaths worldwide each year, according to the WHO. All forms of tobacco use are known to be harmful, including exposure to second-hand smoke and exposure in fetal life [1]. A growing body of research agrees that habitual smoking leaves a significant signature on DNA methylation [2, 3], and there is increasing evidence that some of the diseases associated with smoking may be mediated by smoke-induced changes in DNA methylation [4,5,6]. Smoke is also known to be a major cause of increased epigenetic aging [7,8,9]. In our recent study, we showed that those who smoke are, on average, four years older epigenetically than those who have never smoked [10]. The mechanism by which smoking affects DNA methylation is still under investigation, but current evidence suggests that it involves smoke-induced DNA damage and recruitment of DNA methyltransferases that methylate adjacent cytosines in CpG dinucleotides, or nicotine-induced downregulation of DNA methyltransferases [3, 11].

Monitoring the dynamics of changes in the DNA methylation profile with smoking exposure or after smoking cessation may be of great interest in diagnostics or clinical interventions to describe a patient’s smoking status better than self-report and to estimate the risk of smoking-related diseases such as myocardial infarction, lung cancer, diabetes, or chronic obstructive pulmonary disease (COPD) [12, 13]. DNA methylation is a promising biomarker of health and biological aging due to the potential to revert. However, cigarette-sensitive markers may differ in their ability to return to the pre-exposure methylation state [14,15,16,17]. An association between methylation profile and time since smoking cessation has also been reported [16,17,18]. Moreover, inferring an individual’s smoking status can also be a useful piece of information in a criminal investigation to complement an offender’s genetic profile [19,20,21].

Numerous CpG sites (CpGs) have been associated with tobacco smoking in blood, but the strongest smoking-induced epigenetic response has been reported for the AHRR gene [14, 16, 22,23,24], with cg05575921 included in most of predictive models available in the literature [21, 25,26,27,28,29]. A protein encoded by the AHRR gene participates in the aryl hydrocarbon receptor (AhR) signaling cascade, which mediates the degradation of environmental toxins, and regulates cell growth and differentiation. AHRR was found to be associated with smoking in Europeans and Asians, although population differences in methylation levels for AHRR and other markers were observed and an interaction between smoking status and ethnic group was identified at the AHRR locus [25]. Importantly, AHRR methylation was found to be informative for smoking inference in both blood and saliva tissues [27]. AHRR is also a predictor of epigenetic aging, included in the pace of aging model [30] and mortality risk score models [8], but it is not clear whether its role in epigenetic aging is limited to smoking exposure only. In the literature, AHRR has been associated with smoking-related diseases such as COPD incidence, lung function, lung cancer [31, 32] and atherosclerosis [33].

In the current study, blood DNA methylation profiles of nearly 800 individuals were assayed using Illuminia EPIC arrays. By conducting a novel epigenome-wide association study, we described epigenetic markers of smoking in the Polish population and focused on the role of the AHRR gene as a top marker of smoking and other lifestyle habits. We compared the methylation profiles of smokers, never smokers, and former smokers, which allowed us to infer the temporality of DNA methylation at smoke-sensitive markers. Furthermore, we developed a new compact 4-CpG prediction model for smoking, which may be of practical importance for forensic and diagnostic purposes.

Methods

Study population and smoking information

The study comprised of a total of 772 participants. First, whole blood samples were collected from 737 Polish unrelated individuals aged ≥ 20 years old (age mean ± SD = 46.4 ± 14.8) along with information on cigarette smoking history (Table S1). Participants were asked whether they were current or former cigarette smokers. Current smokers were asked to define their smoking frequency and were categorized as occasional smokers (≤ 1–2 times per week), regular light smokers (< 5 cigarettes per day), regular medium smokers (5–20 cigarettes per day) or regular heavy smokers (> 20 cigarettes per day). Former smokers were asked to define the number of years since quitting. For the purpose of statistical tests, current smokers were all included in one group of smokers or in two groups of light smokers or heavy smokers, with light smokers including individuals who smoke occasionally or regularly in small amounts, while the second group included individuals who smoke regularly in medium amounts or regularly in large amounts. Blood samples were also collected from a group of N = 35 children under the age of 13 who were treated as an additional validation group, although no smoking questionnaire was available at the time of sample collection.

Written informed consent was obtained from all participants, and the study was approved by the Bioethics Committee of the Jagiellonian University in Krakow (decision no. 1072.6120.132.2018).

Description of other lifestyle habits

Information on other lifestyle and socio-demographic characteristics was also available in this study (N = 737) and included education (university degree, primary school, high school, or vocational school); socioeconomic status (SES) (high vs. above average vs. average vs. low vs. very low); job type (physical/manual, mental work partially sedentary (up to 4 h per day; sedentary mental; SM), mental work only sedentary (more than 4 h per day; long sedentary mental; LSM), or retired/unemployed); workplace exposures (low/high workplace temperature, exposure to pesticides/chemicals, toxins/heavy metals/air pollution, ionizing radiation, sun, and stress); physical activity (e.g., exercise, jogging, cycling, yoga, etc.); frequency of alcohol consumption (non-drinkers, occasional drinkers (drinking once a week), and frequent drinkers (drinking at least three times a week)); number of meals per day; servings of fruits and vegetables; frequency of fish and meat consumption; cups of coffee per day; and hours of sleep. Frequencies of lifestyle-related characteristics in the study population are shown in Table S2.

DNA methylation data analysis

DNA was extracted from blood samples using an automated method and the Maxwell RSC Blood DNA Kit (Promega Corporation), and its quantity and quality were evaluated using NanoDrop (Thermo Scientific, MA, USA) and the Qubit dsDNA HS Assay Kit (Thermo Fisher Scientific). The amount of 500 ng DNA per sample was subjected to bisulfite conversion using the EZ-96 DNA Methylation Kit (Zymo Research Corp., CA, USA) according to the manufacturer’s instructions for Infinium assays. DNA methylation profiles were then generated using the Illumina Infinium Methylation EPIC microarray (Illumina, San Diego, CA, USA) [34]. To minimize batch effects, samples were analyzed at similar times and randomized on microarray plates prior to analysis using the web-based RANDOMIZE application [35]. In addition, at the data analysis stage, principal component analysis was used to detect potential batch effects, and the associated scatter plot of the top PCs did not indicate a high source of variation. Primary quality control of raw array data, filtering of low-quality samples and probes was performed as described in our previous study [10]. Functional normalization (FunNorm) on filtered data (750 821 probes) was performed using the minfi package in R version 4.3.1 [36].

Statistical data analysis strategy

Using specific sample sets and methylation data, a series of statistical analyses were performed, including an epigenome-wide association study (EWAS) for categorical outcomes and associated functional annotation analysis, followed by model building using logistic regression, again for categorical outcomes and model validation, and finally an association analysis of lifestyle factors with DNA methylation at the AHRR cg05575921 and since DNA methylation was considered the dependent variable in this type of analysis, linear regression was used. In the EWAS analysis, all current smokers (N = 171) were compared to all never smokers (N = 404). In turn, a subset of the EWAS sample was used for prediction model training (70%, N = 238) and testing (30%, N = 100), and samples within each group were matched for age and sex. Former smokers (N = 162) and children (N = 35) were used as external validation groups. The association of sociodemographic and lifestyle characteristics with DNA methylation at the AHRR was assessed in the dataset of blood samples from N = 737 individuals, including the EWAS sample set and former smokers. Summary statistics of the study samples are shown in Table S1.

Epigenome-wide association study

Differentially methylated positions (DMPs) for smoking, defined in a binary manner as smoking (N = 171) vs. nonsmoking (N = 404), were identified by a EWAS study using the limma package [37]. The results of the EWAS analysis were adjusted for age, sex, and blood cell counts (T-cells: CD8 + , CD4 + , CD8 naive, CD4 naïve; NK, B-cells; monocytes; granulocytes) predicted with a Houseman method [38]. Associations between differentially methylated CpGs and Gene Ontology (GO) and KEGG terms were investigated using the missMethyl R package [39]. The Benjamini–Hochberg false discovery rate (FDR) method was used to correct for multiple comparisons, and results with FDR p < 0.05 were considered significant.

Development and validation of predictive model

Predictive model training was performed on 238 samples, including 119 non-smokers and 119 current smokers, matched for age and sex. No significant differences in sex (Pearson’s chi-square p = 1.0) and age (Mann–Whitney U p = 0.999) between the groups were confirmed with appropriate statistical tests (Table S1). Variable selection was performed on the DMPs identified in the EWAS study by stepwise logistic regression, using the significance of the score statistic as the criterion for predictor entry and the probability of a likelihood ratio statistic based on maximum partial likelihood estimates as the test of exclusion. Nagelkerke’s R² statistic was calculated, which is an approximation of the R² statistic for linear regression. The R² statistic measures the goodness of fit of a model and summarizes the proportion of variance explained by each predictor. Bootstrap analysis with n = 10,000 permutation tests was applied to evaluate the robustness of the model. Two types of models were developed, the binomial model for comparing two categories (current smokers vs. never smokers) and the multinomial model, where three categories were considered (never smokers vs. light smokers vs. heavy smokers). The performance of the developed models was tested on an independent set of 100 blood samples from 50 smokers and 50 never smokers and the accuracy was described by AUC (the area under the ROC curve), sensitivity (true positive rate, for the binomial model it is the percentage of smokers correctly identified as smokers), specificity (true negative rate, for the binomial model it is the percentage of never smokers correctly identified as never smokers), and total number of correct predictions (the percentage of correct scores calculated for the combined group of smokers and never smokers). The newly developed biomarker for smoking was further validated using two additional sample sets, former smokers (N = 162) and children (N = 35). The epigenetic score for smoking, i.e. the probability of smoking and its distribution in different groups, was analyzed. Prediction modelling was done using IBM SPSS Statistics 29.

Analysis of the correlation between epigenetic score for smoking and epigenetic age

In our previous study, we showed that in our database of Polish samples, epigenetic clocks correlate strongly with lifestyle habits, including self-reported smoking [10]. Here, we tested whether the epigenetic score for smoking correlates with different measures of epigenetic age acceleration (EAA) as determined by different clocks, including Horvath 2013 [40], Hannum 2013 [41], Horvath Skin&Blood [42], PhenoAge [43], GrimAge [7], FitAge [44], Mortality Risk Score (MRS) [8] and DunedinPACE [45]. For all clocks except the latter two, epigenetic age acceleration (EAA) was calculated as described in our previous study [10]. The pairwise correlation between the epigenetic score for smoking measured by our new binomial model and different EAAs, MRS or PACE was analyzed by Pearson’s correlation test.

Association tests for AHRR and lifestyle

We tested for association between DNA methylation at cg05575921 in AHRR and lifestyle data. DNA methylation at cg05575921 was treated as the outcome, while each socio-demographic and lifestyle characteristic was independently entered into the linear regression model as an independent variable. The strength of association was interpreted using standardized beta coefficients, and the R² statistic was used to calculate the proportion of variation in DNA methylation at AHRR explained by lifestyle factors. Results were always adjusted for age and sex, smoking status, blood cell components, and DNA methylation pack-years calculated with the Horvath online tool (https://dnamage.clockfoundation.org/); and results p < 0.05 were considered statistically significant.

Results

Epigenome-wide association study for smokers vs. never smokers

The EWAS analysis for smoking yielded 459 age- and sex-adjusted CpG associations in a group of 575 individuals (Fig. 1). The top 12 CpGs (FDR p < 1 × 10^–7) are shown in Table 1, while the full list of FDR-significant CpGs (FDR p < 0.05; raw p < 5 × 10^–5) is shown in Table S3. When the results were additionally corrected for several lifestyle factors, including socioeconomic status, education, coffee consumption, and sleep duration, 99 CpGs remained significant (Table S4). The highest ranking hit was cg05575921 AHRR (p = 4.5 × 10^–32 adjusted for sex and age; p = 8.7 × 10^–27 adjusted for age, sex and lifestyle factors), and 6 other CpGs in AHRR also reached EWAS significance (Table S3). Downstream pathway analysis revealed enrichment in multiple biological functions (FDR p < 0.05), including among others organism development, cell communication, regulation of signaling, response to stress, blood vessel development, and cell death (Table S5). When KEGG pathways were tested in the overrepresentation analysis, although no significant groups of genes were revealed after FDR correction, the top significant (raw p < 0.05) terms included pathways in cancer, lipid and atherosclerosis, purine metabolism, and axon guidance (Table S6). When all 459 EWAS-significant CpG markers were analyzed in an independent group of former smokers, 117 of them were found to be statistically significantly (nominal p < 0.05) associated with time since quitting, with 21 reaching p < 0.0001 (Table S7).

Fig. 1

Manhattan plot of EWAS for smoking adjusted for age, sex and blood cell count. Red horizontal line corresponds to FDR p < 0.05, with the top annotated markers meeting this threshold

Table 1 Top 12 (FDR p < 10^–7) independent EWAS-based CpGs associated with smoking

Predictive analysis of smoking status

Prediction models were developed using logistic regression and stepwise marker selection. Two types of data categorization were used, including a simple binary definition of smoking and including 3 smoking categories defined as non-smokers, occasional smokers, and regular smokers. Four CpGs (cg05575921 in AHRR, cg09594361 at chr1: 54905423, cg21322436 in CNTNAP2 and cg09842685 at chr12: 4492769) were selected to explain a total of Nagelkerke R² = 38.9% of the observed variation in smoking. The analysis of multicollinearity showed low values of variance inflation factor (VIF) for the predictors in the model (between 1.1 and 1.4). The characteristics of a binomial regression model are shown in Table 2.

Table 2 Variables in the binomial logistic regression model and bootstrap analysis (n permutation tests = 10,000)

The binomial model predicted smoking with high accuracy described by the AUC parameter at the level of 0.80, in both training (N = 238) and test (N = 100) sets (Table 3). The sensitivity and specificity were 66% and 82%, respectively. The final number of correct classifications was 74% in the test set. The epigenetic score for smoking in the smoker and never smoker categories was compared and shown in Fig. 2. When the distribution of probabilities was analyzed in a larger number of categories, taking into account the frequency of smoking, a gradual increase in the probability of smoking was observed, starting from occasional smokers to regular smokers who smoke in large quantities. Importantly, the mean score (probability) of smoking was higher for former smokers than for never smokers, but lower than for current smokers (Fig. 3). In addition, the specificity of the model was tested in children aged < 13 years and, as expected, very low probabilities of smoking were achieved in this group. Notably, the mean score for smoking was higher in the never smokers’ group (0.37 ± 0.19) than in the children (0.20 ± 0.15).

Table 3 Predictive performance for binomial and multinomial models for smoking status

Fig. 2

Epigenetic score (probability) for smoking for never smokers and current smokers within training and test sets, as determined by a binomial regression model

Fig. 3

Epigenetic score (probability) for smoking for different sample groups as determined by a binomial regression model. Smokers were divided into four groups based on frequency of smoking: occasional smokers, regular light smokers, regular medium smokers and regular heavy smokers

In the next step, the categories of occasional smokers and regular light smokers were combined into one category of light smokers, while regular medium smokers and regular heavy smokers were combined into another category of heavy smokers. In the multinomial model, the AUC values were very high for never smokers and heavy smokers (0.8 and 0.88, respectively), and lower, but close to 0.7 level of prediction accuracy was achieved for the light smoker category, intermediate between never smokers and heavy smokers. High sensitivity was obtained for the never smoker category (84%), which was significantly lower for light smokers (44%), while high specificity of prediction was achieved for the light smoker and heavy smoker categories (80% and 97%, respectively).

All four CpGs selected in the models were significantly correlated with smoking status in both types of EWAS analyses (Table S3 and S4), and positively influenced the AUC value in both the training and test sets (Table S8). The single AHRR cg05575921 was found to explain 21.5% of the variation in smoking and, when used alone, predicted smoking at the AUC level of 0.76 in the test set (Table 3). The AHRR gene was also found to have the highest observed change in DNA methylation when comparing current smokers to never smokers with a mean difference in DNA methylation beta of 0.13 ± 0.01 (Fig. 4). Importantly, the AHRR cg05575921 and the other three DMPs included in the prediction models were all hypomethylated in smokers.

Fig. 4

Boxplots of methylation levels for 4 CpGs included in the predictive model for smoking and comparison between current smokers, former smokers and never smokers. Groups were matched for age and mean age of each group was 46.3

To analyze the dynamics of DNA methylation in smoke associated CpGs, we also compared non-smokers and current smokers with former smokers and found that there were no significant differences in DNA methylation beta levels between never smokers and former smokers for all CpGs in the model except cg05575921 in AHRR (Fig. 4). However, the methylation profile of former smokers at cg05575921 was more different from current smokers (p < 0.001) than from never smokers (p = 0.006). This finding may indicate the potential recovery of DNA methylation after smoking cessation and suggests that the rate of these changes may differ for different loci in the genome.

Epigenetic score for smoking correlates with epigenetic age acceleration

Analysis of the correlation between the epigenetic score for smoking (the probability of smoking) and various measures of epigenetic age acceleration (EAA) showed a significant effect for most clocks, except Horvath 2013 and Horvath Skin&Blood, which belong to the first generation of epigenetic clocks trained solely on chronological age. The strongest correlation and highest significance was observed for the GrimAge clock (Pearson’s R = 0.665, p = 1.25E-63) and the Mortality Risk Score (Pearson’s R = 0.525, p = 6.30E-36). Results are shown in Table S9.

Associations between AHHR and various lifestyle factors

Methylation of AHRR cg05575921 decreases slightly with age (p = 1.89 × 10^–5) and is lower in males (p = 1.67 × 10^–6). Of all habits examined, cigarette smoking was the most strongly correlated with methylation at cg05575921 (p = 4.71 × 10^–41). Higher methylation was observed in physically active individuals who exercised daily, but not less than once a week, and the results were significant after adjustment for age, sex, smoking status, blood cell count and DNA methylation pack-years (p = 0.028). People who slept at least 8 h per night also showed a more favorable methylation profile in cg05575921 compared to people with a sleep deficit (p = 0.027, Table 4). The full association tests for all lifestyle factors are shown in Table S10.

Table 4 List of lifestyle behaviors significantly associated with DNA methylation at AHRR cg05575921 in a group of N = 737 individuals

Discussion

To address the problem of the aging population and as a prevention strategy for age-related diseases, epigenetic reprogramming is a promising solution [46, 47]. As cigarette smoking has been implicated as a major cause of epigenetic aging [9, 10, 48], and literature data show that smoke-induced changes in DNA methylation patterns may mediate disease development [4,5,6], smoking-responsive markers appear to be promising targets for intervention and investment. DNA methylation-based smoking inference is also of interest in the field of forensic DNA phenotyping, which aims to describe the physical appearance, biogeographic ancestry, age, and lifestyle information of a perpetrator or human remains subject to identification [49, 50].

DNA methylation at a number of genomic loci has been associated with smoking exposure, and a clear trend towards hypomethylation of the genome has been observed in smokers [22]. This phenomenon was also observed for four markers included in our new smoking prediction model. The largest magnitude of effect was observed for the aryl hydrocarbon receptor repressor gene (AHRR), which has an established role in the response to smoking [23,24,25, 33, 51]. Reduced methylation at AHRR cg05575921 (chr5: 373378) results in increased expression of an AHRR-encoded protein [52] that represses the AhR receptor, thereby negatively impacting toxic clearance processes [53]. AhRR protein can also affect other signaling pathways and regulates inflammatory responses [54]. We noted the potential reversibility of DNA methylation following smoking cessation, although the DNA methylation pattern in AHRR in former smokers was still significantly different from that in never smokers, unlike the other three markers in the model for which we observed that methylation profiles did not differ significantly between never-smokers and ex-smokers. Different markers may require different times to return to levels seen in never smokers, and the literature suggests that methylation at AHRR returns to normal after 5 years of abstinence [22]. Our study suggests that this time may be longer or population dependent, as the mean time since quitting smoking in our group of former smokers was 14.6 ± 10.5 years. The obtained data also suggest that a faster return to methylation levels characteristic of never-smokers after quitting could potentially be achieved in people who change their lifestyle in terms of physical activity, sleep hours and diet. This shows the potential of interactions in shaping the final profile of DNA methylation.

AHRR is recognized as a strong marker of smoking in Europeans and Asians [55], but also as a biomarker of epigenetic aging [30], mortality [8], smoking-associated diseases and lung cancer [31, 32, 56]. In our study, we demonstrated that AHRR is also sensitive to lifestyle factors other than cigarette smoking. Favorable effect and increased DNA methylation was observed for enough sleep hours and physical activity (at least once a week). However, the significance of the association and the magnitude of the effect were smaller than for smoking. These results suggest the role of AHRR in epigenetic aging via mechanisms independent of smoking consumption and its power as a general biomarker of health and fitness. Importantly, members of the AhR-signaling pathway have been linked in the literature to the circadian rhythms [57, 58], and SNP polymorphisms in the AHRR gene associated with insomnia and early awakening [59]. These findings fit in well with our results showing that DNA methylation at AHRR is sensitive to sleep duration. The beneficial effects of physical activity on epigenetic aging have been widely described [60, 61] and confirmed in our recent study [10]. Daily exercise was associated with reduced epigenetic aging as measured by most of the known epigenetic clocks. Studies showed that exercises can lead to mobilization of natural killer cells involved in cell cytotoxicity via AhR/IDO pathway [62]. A recent study showed that acute exercise can affect AhR signaling, which in turn can influence the expression of the programmed cell death protein (PD-1), a promising target in cancer therapy [63]. Physical activity is therefore emerging as an important factor in immune regulation.

We showed that the AHRR cg05575921 was the main predictor of smoking in our study population, allowing high accuracy of smoking prediction described by the AUC parameter at the level of 0.76 when used alone. The value of AUC considered acceptable for implementation is 0.7 [64, 65]. To augment the performance and accuracy of the model, larger number of predictors is needed. In our study, three additional methylation markers were selected that independently contributed to the accuracy of the smoking inference. Two of them have a well-established role in smoking, CNTNAP2 cg21322436 (chr7:145812842) [16, 66,67,68] and cg09842685 (chr12: 4492769) [69], and their association with the risk of COPD and lung cancer has also been described [70]. CNTNAP2 is one of the largest genes in the human genome and encodes a contactin-associated protein-like 2a, which is a type of neurexin protein involved in cell adhesion and nervous system development. CNTNAP2 has been implicated in many neurodevelopmental diseases. Maternal smoking during pregnancy was found out to impact DNA methylation at CNTNAP2 and decreased level of methylation was suggested as a protective mechanism against adverse effects of smoking [71]. CpG cg09594361 (chr1:54905423) is a novel marker whose role in smoking has not been described so far. Importantly, this CpG was not covered by an older Infinium 450K array.

The developed 4-CpG model is characterized by high prediction accuracy with AUC = 0.8 and number of correct classifications at 74% in the test set. This result is more or less comparable to other published prediction models, using from 1 to 13 CpG markers for model training [21, 26, 29], although it should be noted that AUCs reaching 0.9 which is higher than in our study were also reported [21]. Our model allows to balance predictive accuracy and compactness of the marker panel, avoiding the overfitting often discussed in the literature [21, 29] and fulfilling the criteria often set in forensic genetics when dealing with difficult and degraded DNA material. The small number of markers also allows the use of alternative, less expensive and more available methods of DNA methylation analysis, such as pyrosequencing or high-throughput targeted DNA sequencing [72]. Furthermore, in our study we provided sufficiently accurate smoking inference at a higher prediction resolution by considering three possible categories of smoking status (never smoker vs. light smoker vs. heavy smoker), with former smokers showing similar methylation profiles to light smokers. Lower predictive accuracy for former smokers category and misclassification of former smokers as never smokers have been observed in previously published prediction models [21, 29, 74]. Furthermore, as we have shown in our study that AHRR is modifiable by other lifestyle factors, this may also have an impact on the accuracy of smoking prediction. Importantly, high predictive accuracy was observed in our study for a group of children used as an additional set for model validation. The probabilities of smoking were lower for children compared to never-smokers, suggesting that passive smoking and the cumulative effects of other exogenous agents may affect DNA methylation at smoke-associated CpGs over time. However, although second-hand smoke exposure is a known risk factor for diseases, passive smoking was shown to be much less pronounced in the DNA methylation pattern than active smoking [74].

Our study does not come without its limitations. The sample groups used to train and test the predictive models were part of the sample set used to perform the EWAS analysis. Therefore, the results should be validated in the future. On the other hand, the use of two additional independent groups, i.e. former smokers and children, strengthens the conclusions of the study. Another weakness of the study is the lack of questionnaire data on smoking among those under 13 years of age. Therefore, the assumption of non-smoking in this group is biased. In addition, given the large number of smoking-associated markers known in the literature and discovered in our project, alternative methods of variable selection could be applied in the future [75, 76] to improve the selection of the most relevant variants, facilitate the detection of epistatic effects [77], and ultimately further improve the accuracy of prediction.

Conclusions

We confirmed that DNA methylation is highly predictive of smoking status and revealed the role of > 400 CpGs, including novel loci involved in smoking. We confirmed AHRR as the top locus associated with smoking and provided novel data showing that changes in DNA methylation at the AHRR can be achieved by multiple lifestyle factors. Although there are reports in the literature suggesting the role of AHRR polymorphism in insomnia, this is the first study to link DNA methylation at AHRR with sleep duration. We developed a competitive model for smoking inference from blood consisting of only 4 CpGs. High predictive accuracy was obtained for the binomial model, but importantly, the multinomial model was able to predict three categories of smoking with reasonable accuracy. We also demonstrated a high correlation between the epigenetic score for smoking and epigenetic age acceleration. We provided novel evidence for DNA methylation reversion after smoking cessation, which is of utmost importance given that DNA methylation at the AHRR and other loci has been associated with the risk of developing lung cancer and other diseases in the literature.