Tobacco use causes up to 6 million deaths per year [1] – 10% of adult deaths [2] – primarily through cancer, chronic obstructive pulmonary disease and cardiovascular disease [3]. The link between smoking and adverse health outcomes is well established but the precise causal molecular and cellular mechanisms are still under investigation [4–6]. Epigenetic modifications, such as DNA methylation at cytosine-guanine dinucleotides (CpGs), are thought to be potential mediators in the course from exposure to disease.

The link between smoking and methylation is receiving growing attention [7], with numerous recent cross-sectional studies revealing epigenome-wide associations [8–12]. The most recent, comprehensive analyses include a systematic review of 17 studies [13], and a meta-analysis of 16 cohorts [12], the latter revealing 2623 CpGs sites significantly associated with smoking behavior at a strict multiple testing threshold. However, longitudinal studies – here defined as those examining methylation and smoking habits in individuals at two or more time-points – are scarce. These studies are an important next step in the understanding of the mechanisms of methylation [14]. To our knowledge, there have been only three reports investigating smoking and methylation with repeated measures: two candidate locus studies, one focusing on a small group of young women [15] and another on a small group of individuals attempting to quit smoking [16]; and a family-based study focused on maternal smoking [17]. Studies which cover larger sample sizes and different life stages are needed to better understand the effect of smoking on methylation.

A key question in the field of smoking-mediated methylation is whether quitting smoking allows differentially methylated CpG sites to return to levels found in individuals who have never smoked. Previous evaluations of the effect of time since smoking cessation have revealed CpG sites that could be classified as reversible and sites that could be classified as persistently differentially methylated [8, 18–20]. Reversible sites are those that are differentially methylated between smokers and non-smokers but with the difference disappearing some time after cessation of smoking. Persistently differentially methylated sites remain differentially methylated, perhaps indefinitely; this persistence has been observed even up to 35 years after cessation [20]. However, there is again a lack of longitudinal studies with regard to reversion of methylation levels; this lack impedes the identification of potential long-term biomarkers of smoking [11, 21], and hinders insights into the increased risk for disease faced by former smokers decades after cessation [22, 23].

The focus of our investigation is the linking of exposure – current and prior smoking behavior – to changes in DNA methylation at CpG sites over time, i.e. dynamism in methylation. Our previous study, a cross-sectional epigenome-wide association study (EWAS), compared the methylation of current, former and never smokers at approximately 450,000 CpG sites, and revealed an extensive effect of smoking across the methylome [8]. Here, we extend this investigation to include an earlier time-point for the same cohort, thus gaining information on changes in smoking habits, health characteristics and DNA methylation. Through a longitudinal site-by-site analysis, our goals were to (i) examine changes in methylation over time associated with quitting smoking or remaining a former smoker; and (ii) examine the effects of continued smoking on changes in DNA methylation, including the effect of intensity of smoking habit.

We have used longitudinal data with repeated measures of DNA methylation to examine changes in methylation over time associated with quitting or continuing smoking.

For the smoking cessation analysis, site-specific methylation changed the greatest (relative to changes observed for NS) for people who quit within 20 years prior to the follow-up examination. A total of 52 CpG sites were significant in our primary analysis of FS vs NS, with 32 significant for individuals quitting within (>0)-4 years and 30 significant for those who quit within 5–14 years prior to the follow-up examination (10 shared between them). Only one site showed a significant difference for individuals who had quit more than 14 years prior to the follow-up assessment.

Of the 52 significant sites, there were many from the same chromosome regions or genes. AHRR (8 sites), the aryl hydrocarbon repressor gene, located on chromosome 5, is involved in mediation of dioxin toxicity and functions also in cell regulation and growth. It is theorized that altered AHRR expression may have a deleterious effect on the body’s ability to eliminate environmental chemicals which may act as carcinogens [7]. MYO1G (3 sites), or unconventional myosin IG, is a plasma-membrane associated class 1 myosin and is found in abundance in lymphocytes [45]. It has been found to be related to cell death and a potential factor in cancer [46]. HIVEP3 (2 sites) was found to be the gene with the most strongly associated CpG site (by p-value) in the meta-analysis of Joehanes et al. [12], where they note its role in bone formation. GFI1 (Growth Factor Independent 1 Transcriptional Repressor) (4 sites) is a protein-coding gene, and its methylation was found to be the most robust mediator of the association between maternal smoking and birthweight in a recent study of children and newborns exposed to maternal and paternal smoking [47]. ZNF668 (2 sites) is a zinc finger protein whose role in DNA repair, cell proliferation and cancer has been investigated [48]. Chromosome region 1p36 (2 sites) has been investigated as a region containing a possible tumor suppressor [49]; with investigations including the effects of methylation [50]. Methylation of regions on chromosome 2, 2p25 (2 sites) and 2q37 (3 sites), has recently been found to be associated with all-cause mortality [51], as was methylation of many of the other genes significant in this analysis, including F2RL3, AHRR, and region 14q32. LRP5 (2 sites) plays a role in the Wnt signaling pathway, which influences bone formation and is a factor in skeletal disorders [52].

A site that is significant longitudinally is properly interpreted as one that is rapidly changing compared to within NS. Non-significant change does not necessarily imply that the methylation of FS has recovered to the level of NS. If indeed there is no significant change, but the methylation levels themselves remain different, it may indicate that the levels are only very slowly returning to those of NS, or not at all. Power considerations are also important: those sites that were significantly different between CS and NS for the EWAS may never show significant change in the TSQ analysis due to much smaller sample size for each category, a model that has fewer degrees of freedom and a smaller effect size. This does not imply that the methylation of these sites remains different to that of NS. This is most apparent for the difference between TSQ_L categories 1 and 2, where the difference between the significant cross-sectional and longitudinal coefficients for TSQ_L category 1 (264 and 32, respectively, a difference of 232) is much larger than the number of cross-sectionally significant sites for TSQ_L category 2 (104).

The sites that are not significant up to 4 years TSQ but are for between 5 and 14 years are difficult to interpret due to the implications of our categorization. It is possible the sites maintain the methylation levels of CS for years before rapidly changing later, a cascading effect in which certain methylation changes occur prior to others. Further possible explanations are attributable to the fact that the individuals with TSQ up to 4 years (TSQ_L category 1) are those who quit smoking between baseline and follow-up. Firstly, we do not have their methylation levels at cessation, only from some time prior, at baseline, at which point they were CS. Had their continued smoking to the time of cessation further driven their methylation levels and those of NS apart, the longitudinal effects seen would be diluted. Secondly, for other TSQ_L categories, we are examining the change in methylation for on average 7 years of non-smoking. Even considering a theoretically more rapid yearly change for individuals of TSQ_L category 1, the total effects measured (i.e., the regression coefficients) are based on at most 4 years of non-smoking. This change could thus be smaller in magnitude than a smaller per year change totaled over 7 years of non-smoking. Future studies could measure methylation directly before cessation to avoid these consequences.

The finer scale examination of reversion using medians of methylation change seems to indicate that methylation remains in general more dynamic for FS than NS even up to 40 years after cessation. Figure 2 shows in detail greater fluctuating methylation levels amongst FS than NS: after a certain length of time after quitting, methylation levels are not necessarily moving in opposite direction to the general effects of smoking, and in general changes remain stronger in magnitude than for NS. These indications, which warrant further investigation, seem to concur with Philibert et al. [16] that there can be overcorrection of methylation after cessation of smoking.

Regarding overall rates of reversion of disturbed methylation over time, it is of debate as to whether the relative persistence of differential methylation following cessation for certain CpG sites is due to slower reversion rates, or comparable reversion rates but a larger initial disturbance in methylation [8, 20]. Our results indicate the rate at which the methylation levels change in FS is related to the difference in methylation to NS, implying perhaps a type of exponential decay in the difference in methylation after cessation. Further investigation could investigate if this would be an appropriate model for some CpG sites, and if so, to determine a decay rate of the difference. Of particular interest are those sites with the largest absolute cross-sectional coefficient to longitudinal TSQ_L coefficient ratios. The implication of such a ratio being large is that these sites display an initially large difference between FS and NS, and this difference decays more slowly, relative to other sites. Of the 10 sites with the largest ratios in our study, 8 have been found to be “persistently differentially methylated” amongst FS in a previous study [20]. CpG sites which remain differentially methylated for long periods after cessation can aid us in understanding the course of smoking-related diseases – and thus the ongoing increased risk faced by FS – and act as biomarkers for past smoking exposure. Other markers, such as cotinine, a metabolite of nicotine, have short half-lives and their usefulness is thus limited [53]. Past candidates for CpG biomarkers of smoking include those at AHRR [11, 54], F2RL3 [21] and the position 2q37 [11]. Our study confirms their utility in this regard, as our longitudinal results indicate that the methylation difference is large enough that reversion can still be measured significantly many years after cessation. Further, if multiple methylation measures per individual are available, both the levels and change could be combined in a biomarker to provide more accurate estimates of smoking history and risk of disease.

In terms of future avenues of investigation with regard to disease, additional further analysis could focus on the biology pertaining to those sites persistently differentially methylated, or even those sites with greater fluctuation in methylation, and the implications in disease etiology. We have presented here an array of diseases associated with both the significant loci and with smoking, but did not investigate incidence of these conditions in our population.

The investigation into continued smoking reveals that there are complex mechanisms behind the methylation levels within the blood. The results that indicate the effects are not consistently positively correlated with the effects of smoking in general – as seen through the significant longitudinal effects not necessarily showing the same direction with the cross-sectional effects – may be surprising considering that methylation of some sites has been found to be associated with cumulative smoking exposure (often given as “pack-years”, the product of average intensity of smoking and duration of smoking) [9, 18, 21]. Although the decrease in smoking intensity may partially explain the opposing longitudinal and cross-sectional effects, we see no conclusive evidence. Another explanation may lie in the fact that since our cohorts are composed of mostly older individuals, most have been smoking for many years (mean duration of smoking at baseline: 31.4 years). They may have reached or are approaching a methylation “peak” and changes at this stage may be too small to notice. Zhang et al. [21] show that for selected CpG sites annotated to the gene F2RL3 the methylation response to dose of pack-years flattens after a certain level is achieved: the methylation effect is saturated. If such a model is accurate, and since many of the individuals in our sample are long-time smokers, longitudinal effects may be incorporating many individuals already at this stage, thus potentially “diluting” the contributions from earlier stage smokers. Indeed, since a CpG site does indeed have minimum and maximum methylation (0% - 100%) the effects cannot continue to compound across time after a certain point. The lack of significance results for duration (1 site), intensity (32 sites) and pack-years (we found 66 CpG sites of the 590 to be significant for pack-years, results not shown) may indicate a lack of power, a complex mechanism not addressed by our simple model, or a true lack of effect; this last possibility indicates that presence/absence of smoking habit (beyond a certain cut-off) may be the only relevant factor for certain CpG sites. A study of very early stage smokers may help to address this question. This lack of “new” smokers – either those that have just begun to smoke, or those with a very short history of smoking – is a weakness of this study, as we cannot examine the longitudinal effects within the early years of a smoking habit. These “new” smokers would likely be more informative than our long-term smokers on progression of longitudinal change in methylation. A further possibility for the seemingly conflicting results is interaction with other molecular factors. Recent studies [55, 56] have indicated that single-nucleotide polymorphisms influence the smoking-associated CpG sites. These studies further highlight the complexity of molecular networks, and underline the need for functional analysis.

The study has further limitations. Although longitudinal, only two time points were used, separated by approximately seven years. It is thus difficult to capture the shorter term, i.e. within the first few months, or longer term, i.e. decades, longitudinal effects of quitting smoking. Further, smoking studies often suffer from under-reporting, and a lack of data on consistency of smoking habits, particularly with regard to smokers attempting to quit.

Another weakness of the study is a type of selection bias. By the nature of the study – two time-points featuring identical individuals – we are likely excluding individuals most strongly affected by smoking: those that would have died, perhaps due to smoking-related illnesses, in the years between baseline and follow-up.

Finally, the lack of independent replication data weakens the reliability and generalizability of the results. It should be noted, however, that all CpG sites identified in this study as showing dynamism associated with continued smoking or with past smoking (TSQ analysis) were identified as cross-sectionally associated with smoking behavior in the extensive meta-analysis of Joehanes et al. [12], thus providing additional confidence in the results.

Strengths of the study include the longitudinal data – as mentioned, there is a scarcity of multiple time-point methylation data – and a relatively large sample size. The inclusion of well-documented covariates from the extensive KORA study lessens the possibility we are seeing confounded results.

Our results provide insights into the rates of reversion of smoking-disturbed methylation levels and their continued fluctuation upon cessation of smoking. The results indicate that the most rapid reversion of methylation occurs within the first two decades following cessation of smoking, but that levels continue to fluctuate more for former smokers than for never smokers even beyond 30 years after cessation. Rates of reversion are related to the initial disturbance of methylation, with greater disturbance showing greater change across time. Site-specific results, including those for the previously identified genes AHRR, F2RL3, GFI1, and MYO1G, and chromosome regions 1p36, 2p25 and 2q37, have potential implications for both biomarkers and the treatment of human disease. We note that before-and-after studies on the short-term effects of smoking cessation would be beneficial to fuller understanding. We also demonstrate that the effects of continued smoking on methylation are complex, where duration and intensity of smoking habit and range of possible methylation all play interconnected roles.