Using DNA methylation to validate an electronic medical record phenotype for smoking

Introduction

Twin and family-based studies show that approximately 50 percent of the risk of tobacco dependence is heritable (Goldman et al. 2005). However, the numerous genetic variants that have been linked to smoking behavior explain only a small proportion of the phenotypic variation. A major challenge for gene discovery is phenotypic ambiguity, especially that stemming from self-reported health behaviors such as smoking in which social desirability bias and/or lack of documentation in the health record may obscure actual behavior (Kinsinger et al. 2017). Additional inaccuracy can result from cross-sectional rather than longitudinal approaches to characterizing behavioral phenotypes. Both factors can contribute to inadequate statistical power to detect small individual genetic effects that characterize complex disorders like smoking.

Several biological markers could be employed to validate self-reported smoking data. Cotinine, a major metabolite of nicotine in tobacco smoke, is the most commonly used biomarker. However, cotinine is elevated from smoking for only days to weeks (Jarvis et al. 1998). Expired carbon monoxide, although sensitive to very recent smoking, lasts for only up to 8 hours (Sandberg, et al. 2011).

DNA methylation is an epigenetic process in which a methyl group is added to a DNA molecule. Environmental factors, such as exposure to cigarette smoke, can result in epigenetic changes (Bollati and Baccarelli 2010), and these changes can serve as biomarkers for long-term environmental exposure (Ladd-Acosta 2015). Epigenome-wide association studies have identified hundreds of DNA methylation sites, referred to as CpG or cg sites, associated with smoking (Breitling et al. 2011; Harlid et al. 2014; Gao et al. 2015; Ambatipudi et al. 2016, Joehanes et al. 2016). The two sites most frequently associated with smoking are methylation of cg05575921 in the AHRR gene and cg03636183 in the F2RL3 gene (Gao et al. 2015; Ambatipudi, et al. 2016; Lee et al. 2016; Bojesen et al. 2017), and these two CpG sites have been proposed as candidate biomarkers to determine smoking status (Philibert et al. 2013). These sites have been consistently linked to smoking status in multiple population groups (Elliott et al. 2014; Guida et al. 2015; Beach et al. 2017), among both sexes (Dogan et al. 2015), across a wide range of ages (Beach et al. 2015) and different tissues (Stueve et al. 2017) and across different methylation platforms (Breitling et al. 2011, Joehanes et al. 2016).

The performance of cg05575921 as a predictor of smoking had an area under the curve of 0.99 (Philibert et al. 2015) in a small study. A recent study in two independent samples showed that cg05575921 is a robust indicator for smoking, as confirmed by serum cotinine concentrations (Andersen et al. 2017). Methylation at cg05575921 and cg03636183 along with other CpG sites has also been shown to predict lung cancer risk (Zhang et al. 2015; Baglietto et al. 2017) and lung cancer incidence with an area under of curve of approximately 0.8 (Zhang et al. 2016).

Smoking-associated methylation persists after smoking cessation (Joehanes et al. 2016). Methylation at cg05575921 takes approximately 10 years (Fasanelli et al. 2015), and methylation at cg03636183 takes approximately 20 years (Zhang et al. 2014) after quitting smoking to return to a level similar to that of never smokers. Among Framingham Heart Study participants, cg03636183 was one of the top 36 most statistically significant methylation sites that did not return to never-smoker levels, even 30 years after smoking cessation (Joehanes et al. 2016). The recovery of methylation at cg05575921 has been suggested as a quantitative marker for smoking cessation (Philibert et al. 2016) and cigarette consumption (Philibert et al. 2018). These findings suggest that DNA methylation changes at cg05575291 and cg03636183 can be reliably detected and have long-term stability as measures of smoking exposure.

The Veterans Health Administration (VHA) benefits from one of the most highly developed health information systems in the world (Corrigan et al. 2002; McQueen et al. 2004). Smoking data from the VHA electronic medical record (EMR) date back to 2000 and have been previously validated against two sources of survey data (McGinnis et al. 2011). An emerging source of large genome-wide association studies data is the Million Veteran Program (MVP) of the Department of Veterans Affairs, which is in the process of obtaining phenotypic and genomic data. One goal is to identify genetic variation contributing to nicotine dependence risk. Thus, the phenotypic refinement effort described here has important implications for achieving this aim. Our goal was to use EMR data to develop and validate a longitudinal phenotype for smoking behavior that could be used to identify novel genetic variants in genome-wide association studies. More specifically, we aimed to determine which EMR summary smoking metric (most recent, modal or trajectory) was most strongly associated with DNA methylation markers (cg05575921 and cg03636183).

Materials and Methods

Results

Of the 140 152 patients in VACS with available smoking data, the mean age was 47 years at enrollment, 31 percent were HIV+, 97 percent were male, 48 percent were African American, 40 percent were white and 12 percent were Hispanic or other. Based on modal smoking, 53 percent were currently smoking, 18 percent smoked in the past and 29 percent never smoked. Of the 728 participants in VACS-BC with methylation data, the mean age was 53 years at the time of the blood draw, 84 percent were HIV+, 97 percent were male, 81 percent were African American, 15 percent were white and 4 percent were Hispanic or other. Based on modal smoking, 68 percent were currently smoking, 12 percent smoked in the past and 20 percent never smoked (Table 1).

There were differences in how VACS participants were classified using the three self-reported EMR smoking metrics and the agreement differed between each pair (Table 2). The correlations among the three EMR smoking phenotypes were high: 0.80 for most recent and modal, 0.85 for most recent and trajectory and 0.89 for modal and trajectory.

In VACS-BC, all EMR smoking metrics were associated with both smoking methylation markers. The percentage of patients identified as currently smoking based on both methylation markers increased monotonically from never to past to current smoking for each self-reported EMR smoking metric (Fig. 1). The gradient was steepest for the trajectory smoking metric for cg05575921. When we included only patients whose probability of trajectory membership was ≥0.80, we found similar patterns (Fig. 1), with only a modest improvement in the association between the EMR smoking measures and methylation markers at the two sites.

Agreement (reflected in higher C-statistics) was greater for cg05575921 than for cg03636183 for all EMR smoking metrics (Fig. 2). Limiting the sample to patients with ≥0.80 probability of trajectory assignment resulted in higher C-statistics for all comparisons, but did not alter the relative comparisons among metrics. In all comparisons, the smoking trajectory metric had the highest C-statistics (from 0.67 to 0.89), followed by modal smoking (from 0.67 to 0.86), most recent smoking (0.64 to 0.80) and closest smoking (0.63 to 0.81). For cg05575921, the C-statistics were significantly higher for the trajectories than most recent (p < .001), closest (p < .001) and modal (p = .03) metrics and for modal compared to most recent (p = .002) and closest (p = .01). When we limited the analyses to patients with ≥0.80 probability of trajectory group membership, the differences with cg05575921 were significantly greater for trajectories than for most recent and closest (both p < .001), and for modal than for most recent (p < .001) and closest (p = .02). For cg03636183 among participants of African ancestry limited to those with ≥0.80 probability of trajectory group membership, C-statistics were significantly higher for trajectory than for modal and closest (p = .026 and .049). For cg03636183 among participants of European ancestry, C-statistics were significantly higher for trajectory than for most recent (p = .012) and modal (p = .05). The sample size was too small for comparisons of C-statistics among participants of European ancestry and with ≥0.80 probability of trajectory group membership. Including only the 567 participants who reported using marijuana or hashish less than monthly or never, cg05575921 results were similar to when all 728 were included.

Discussion

Although methylation-defined smoking status was statistically significantly associated with the single, recent, self-reported smoking status in EMR, there was a greater association of methylation-defined smoking status with longitudinal measures (either modal or trajectory). Using methylation as a criterion standard, regardless of methylation cutoff, we found that longitudinal EMR data provided a valid phenotype for smoking status and the validity was enhanced by the integration of longitudinal data into statistical trajectories compared to modal or most proximal phenotypes. In all cases, associations were strong for the methylation site cg05575921 and for cg03636183 for participants of European ancestry. In addition, associations were stronger for cg03636183 for participants of European ancestry than for those of African ancestry, as expected based on previous studies. These results were robust whether we considered only patients for whom the probability of trajectory group assignment was ≥0.80 or irrespective of the certainty of assignment.

This study both confirms and extends prior work on smoking and DNA methylation levels. This work is in line with prior studies demonstrating that cg05575921 and cg03636183 are associated with smoking status (Gao et al. 2015; Joehanes, et al. 2016) and both have been applied as biomarkers to define smoking status (current, past and never) (Shenker et al. 2013), smoking intensity (quantity and duration) (Joehanes et al. 2016; Wilson et al. 2017) and smoking cessation (Philibert et al. 2016). To our knowledge, we are the first to use methylation data as a criterion standard for developing a longitudinal phenotype of smoking behavior for genetic discovery. Methylation data are amenable to this application for several reasons. First, they are not subject to the social desirability bias that can confound self-reported smoking status. Second, changes in methylation are only reversed following a long period of smoking abstinence (i.e. decades), so that associations with these biomarkers would be expected to demonstrate an increasing dose–response association between never, past and current smoking, which we observed in our analyses. Third, the fact that we found similar associations for two separate methylation sites supports the validity of the results.

We demonstrated substantially stronger associations between the criterion standard (cg05575921 and cg0363183) and summary metrics of longitudinal, repeated self-report measures of smoking than with a single report (i.e. most recent assessment of smoking status). As we found in prior analyses of repeated longitudinal self-reported measures of alcohol consumption (Justice et al. 2017; Justice et al. 2018), summary measures of repeated self-reported smoking also demonstrate stronger associations than single, cross-sectional reports that are often employed as phenotypes in large-scale genetic studies (Sanchez-Roige et al. 2017). Like drinking behaviors, smoking behaviors among middle-aged individuals are typically stable with some decrement with advancing age. In this context, multiple observations, summarized over time and adjusted for age, can reduce the degree of ‘noise’ in the measurement and provide improved phenotypic ‘signal’. In our prior analyses of self-reported alcohol consumption, an age-adjusted mean AUDIT-C score provided the best association with our criterion standard (the minor allele frequency for an ADH1B polymorphism that has been previously shown to be protective for alcohol use disorder) (Justice et al. 2018).

In the current analysis, statistical trajectories using age as the time scale were superior to the modal response. These findings suggest that it is both important to employ multiple measures of self-reported health behaviors over time and to adjust these for the age of the individual at the time of the report. Otherwise, an older individual's drinking or smoking behaviors might be misclassified based upon a lower level of use at an advanced age, obscuring higher levels of use at younger ages.

It is also important to note that our findings were highly consistent whether the analyses were restricted to patients for whom the statistical trajectory assignments were most certain (trajectory probability ≥0.80 versus <0.80). Patients for whom the trajectory assignments were most certain had more consistent self-reported smoking behavior (concordance = 52 percent vs 48 percent for all reports, p <.001) and/or had more reports (mean = 13.6 versus 7.0, p <.001). The findings were also similar when we limited the sample to participants who reported using marijuana or hashish less than monthly or never. These findings underscore that our measure is robust to missing data and to multisubstance use.

There are limitations to this study. First, there are relatively few women in the study, potentially reducing generalizability, although sex has not been identified as a moderator of DNA methylation in smokers. In addition, these data lack granular assessments of the dose of cigarette exposure, such as smoking pack years, which have been shown to predict methylation (Joehanes et al. 2016). Similarly, knowing whether a smoker was daily or non-daily would also be important in the assessment of total dose, as non-daily smokers tend to have lower overall exposure (although exposure can still be substantial) (Shiffman et al. 2012). HIV infection results in the depletion of a particular T-cell subset and a shift of overall cell frequencies. To assess whether the reduction of T-cells in HIV+ impacts DNA methylation of cg0755921 and cg03636183 in our sample, we compared DNA methylation of these two loci between HIV+ and uninfected samples and found that they did not differ significantly (cg05575921: t = −1.69, p = 0.09 and cg3636183: t = 0.14, p = 0.88), suggesting that DNA methylation of these loci are associated with smoking but not with HIV status. Finally, the VACS-BC does not contain biomarkers of nicotine metabolism, which is primarily driven by genetic variation in hepatic cytochrome P450 enzymes (Dempsey et al. 2004), and which has been found to influence DNA methylation (Loukola et al. 2015). Although biomarkers may be the most accurate way to identify smoking status, they are impractical on a large scale. We previously developed an algorithm to identify the smoking status of patients in VA care using text fields from EMR smoking data validated against self-reported confidentially collected research survey data (McGinnis et al. 2011). We are currently collecting cotinine concentrations in a separate study of patients and will be able to compare EMR smoking data to cotinine concentrations in future research.

As smoking status is often reported in text fields in non-VA systems as well, these methods can be used outside the VA, so that the findings are generalizable to other health systems. In addition, the VA is rapidly developing a cohort of a million veterans with genetic data in the MVP (Gaziano et al. 2016), and this study will be an important resource as a potential method to identify smoking status in the MVP.

In conclusion, we have demonstrated that longitudinal trajectories of smoking status based upon repeated assessments of smoking that are recorded in the EMR are strongly associated with two DNA methylation sites, which are sensitive biomarkers of tobacco use. We anticipate that these trajectories will prove to be effective, efficient phenotypes for large-scale genetic and other ‘omics’ discovery efforts.