A Droplet Digital PCR Assay for Smoking Predicts All-Cause Mortality

Abstract

Objectives

Determine whether an epigenetic assay for smoking predicts all-cause mortality in adults participating in a longitudinal study of Iowa adoptees.

Background

Improved biomarkers for smoking are needed given its large public health impact and significant limitations of both self-report and current biomarkers such as cotinine in detecting smoking. In the past 5 years, multiple epigenome-wide association studies of smoking have identified loci suitable for translation as epigenetic biomarkers for smoking, in particular the CpG cg05575921. Digital polymerase chain reaction methods hold promise for the development of this and other epigenetic biomarkers.

Methods

Participants in the Iowa Adoption Studies were interviewed regarding their smoking habits. DNA was prepared from whole blood and bisulfite-converted for methylation analysis and digital droplet polymerase chain reaction assay of methylation at cg05575921 was performed. National Death Index records were requested for 584 study participants, resulting in 24 complete matches, 210 partial matches and 350 non-matching records. Complete matches were coded as deceased while the remainder were coded as alive (i.e., censored). In total, methylation data and vital status information were available for a total of N=193 subjects, including 15 deceased and 178 non-deceased. Cox regression was used to examine the ability of cg05575921 methylation as a continuous value to predict the timing of mortality with and without the inclusion of age, sex, race, BMI, marital status, educational status, socioeconomic status, cardiovascular risk factors, and a history of cancer as covariates.

Results

Methylation at cg05575921 predicted the hazard of mortality as the sole predictor and after accounting for major demographic and clinical risk factors. The fitted model showed the hazard ratio increased by 3.5% for every 1% decrease in methylation.

Conclusions

Decreased methylation at cg05575921, an emerging epigenetic biomarker for smoking, was associated with early mortality in a longitudinal study of adults after accounting for the impact of major demographic and clinical risk factors for all-cause mortality. This approach may be useful in clinical research or actuarial assessments.

Introduction

Smoking is the number one cause of death in the United States, with nearly half a million attributable deaths per year. ¹ Despite the magnitude of smoking’s public health impact, the clinical assessment of smoking is typically limited to self-report, ² which may be inaccurate due to recall bias or intentional underreporting. ^3,4 The accuracy of self-reported smoking behavior may be even more limited in several important clinical populations, including pregnant women, ⁵ adolescents, ^6,7 African Americans, ⁸ and those in treatment for nicotine dependence. ^9,10 Biomarkers for smoking are an attractive alternative or supplement to self-report for clinicians and others interested in assessing patient risk and guiding treatment. Unfortunately, two of the most commonly used existing biomarkers for smoking have a major limitation: short half-lives for detection. Cotinine, the major metabolite of nicotine cotinine, ¹¹ has a half-life of approximately 20 hours, ¹² while exhaled carbon monoxide’s half-life is less than six hours. ¹³ Furthermore there is the potential for false-positives with cotinine in individuals receiving nicotine replacement therapy or using e-cigarettes. ^12,14

Recently, measurement of the methylation status of the cytosine-phospho-guanine dinucleotide (CpG) residues at multiple sites across the genome have emerged as potential biomarkers for smoking that can overcome the above limitations. Epigenetic marks such as CpG methylation are modifications to the genome that provide structural and regulatory functions without a change in the base pair sequence. ^15,16 In some cases, epigenetic marks are responsive to exposures such as tobacco smoke and other environmental pollutants. ^17–20 In particular, demethylation of the CpG cg05575921 has been found to be the most sensitive and specific indicator of smoking in the in the epigenome in a number of different studies. ^18,20,21

The performance characteristics of cg05575921 in detecting smoking are excellent, with an AUC of 0.99 reported in two studies that used serum cotinine to confirm smoking status. ^22,23 In addition to signaling both nascent smoking in young adults with as little as a 1/2 pack-year smoking history ²⁴, methylation at cg05575921 also has dose-response characteristics indicative an individual’s cumulative smoke exposure. ^23,25,26 Reversion of demethylation at cg05575921 with smoking cessation has also been demonstrated ^27,28 although the exact extent of this phenomenon is unclear and in need of further study. Finally, methylation of the locus has been directly linked to an increased risk of lung cancer, even after adjusting for reported smoking status. ^27,28

In contrast to the expensive chip-based assays used in most of the above studies, methylation sensitive, digital polymerase chain reaction methods may be more suitable for clinical translation as a means of ascertaining methylation status at specific CpG sites influenced by smoking. For methylation sensitive, digital PCR assays, an individual’s DNA is treated with bisulfite in a process that converts unmethylated cytosines to uracils (which are functionally interpreted as thymines by DNA polymerases, while methylated cytosines remain unchanged. ²⁹ The subsequent measurement of the proportion of converted and unconverted CpG residues can then serve as an indicator of the proportion of methylated CpG residues in the original sample. In contrast to older quantitative PCR techniques, methylation sensitive digital PCR methods allow measurement of these proportions without the need for an external reference ³⁰ while maintaining a high level of precision. ³¹

In this study, we employ a digital droplet PCR assay for methylation at cg05575921 in a cotinine-validated longitudinal sample with a high rate of smoking. We show that cg05575921 methylation is a significant predictor of the timing of all-cause mortality, in contrast to self-report of smoking history. This effect was found when methylation was the sole predictor and when it was adjusted for major risk factors of mortality.

Materials and Methods

Sample

The Iowa Adoption Studies (IAS) is a cohort of 475 adoptees from the state of Iowa with family histories of substance use disorders, antisocial personality disorder, and other psychopathology, along with 475 matched controls. Adoptees and controls have been followed for over 30 years with serial assessments using the Semi Structured Assessment for the Genetics of Alcoholism, Version 2 (SSAGA-II). ³² The SSAGA-II was administered by a trained research assistant during the most recent wave (2005–2009) and included self-report data on tobacco use, health, and demographic information along with other psychopathology and substance use patterns. Whole blood samples used for DNA preparation for use in the current study were obtained by phlebotomy at the same time. Demographic and self-reported clinical data including age, race/ethnicity, marital status, educational level, household income, history of hypertension, cerebrovascular accident, cardiovascular disease, diabetes mellitus, hyperlipidemia, and history of any cancer diagnosis were included in the assessment.

In the original IAS study, subjects were not followed to death or censoring, so that neither state could be verified directly. Proxies of death or censoring were determined in the following manner. Vital status and cause of death were obtained via a search of the National Death Index (NDI) database. The full name including middle initial if available, sex and date of birth of each of the 584 IAS subjects who participated in the last wave of the IAS study, at which time biomaterials were collected, were submitted to the NDI, yielding a total of 234 complete or partial matches and 350 non-matching records. Filtering of the 234 matches yielded a total of 24 full matches, defined as an exact match of first name, last name, middle initial if applicable, and full date of birth. Full matches were coded as deceased in the dataset, indicating the subject was deceased, while partial and non-matches were coded as censored for use in subsequent analyses. Time to death or censoring based on the NDI report was calculated for each subject as detailed in the statistical analysis section. For each death, the ICD-10 codes for primary cause of death and additional entity axis codes included in the NDI report were individually inspected.

All procedures and protocols used in the IAS sample were approved by the University of Iowa Institutional Review Board (IRB).

Biomaterials

DNA from whole blood was prepared according to our previously published methods. ^24,33,34 Measurement of cg05575921 methylation by ddPCR was performed as previously described. ^35,36 Bisulfite conversion of 1μg of DNA from each subject was done with an EpiTect Fast 96 DNA Bisulfite kit (Qiagen, Germany) following the manufacturer’s protocols. Measurement of percent methylation at cg05575921 was then done by comparing the proportion of C (methylated) and T (unmethylated) alleles following bisulfite conversion using the Smoke Signature™ Assay (IBI Scientific, Peosta, IA) and a QX200 Droplet Digital PCR System™ (Bio-Rad, Hercules, CA) following the manufacturer’s protocols. First, bisulfite converted DNA was pre-amplified with the Smoke Signature™ Pre-Amp Master Mix, then diluted between 1:1000 and 1:5000. Subsequently a 5 μl aliquot of the resulting solution was mixed with 1.1 ul of 20X Smoke Signature Assay, 4.9 ul of water, and 11 ul of BioRad 2X ddPCR Supermix (no dUTP), and vortexed. The resulting mixture was then processed with a Bio Rad Automated Droplet Generator, which generated approximately 20,000 micelles each containing approximately 1 nanoliter of PCR mixture, and quickly PCR amplified (95^oC x 10’, then 40 cycles of 95^oC x 15” and 55^oC x 60”, and finally 98^oC x 10’). After amplification was complete, the post-amplification allele content status (either C, T, C+T, or blank) of each micelle by the QX200 Droplet Reader and the percent methylation status of each sample calculated using BioRad’s QuantaSoft software (v1.7).

Statistical analysis

All subsequent analyses were conducted with the R statistical software, ³⁷ using the survival package. ³⁸ In total, methylation data on cg05575921 were available for 220 IAS subjects following ddPCR processing. Subjects failing quality control, defined as greater than a 5% confidence interval for the ddPCR assay or methylation value < 1%, were excluded from further analysis (N=27), leaving a total of N=193 subjects with both methylation data and vital status data for analysis. These data were then merged with the previously described NDI search data reflecting vital status. Of the N=193 subjects, N = 15 were classified as having died (8%), and N = 178 were classified as censored (92%)

Time to death or censoring was computed in the following manner. For a subject with a complete match with the database, time to death was computed as the difference between the date of death and the date of biomaterial sampling (which was the last wave of the original study). For all other subjects, time to censoring was computed as the difference between the date of the NDI search and the date of the biomaterial sampling.

In the first analysis, a Cox proportional hazards regression model was fit using the percent methylation at cg05575921 at the time of the last wave of the study (biomaterial sampling) as a single predictor of the hazard of mortality. In the second analysis, major demographic and clinical variables known to influence mortality risk were added to the model as follows: age at date of interview, sex, race/ethnicity (white vs. nonwhite), body mass index (BMI), marital status at the time of interview (married vs. non-married), educational status (high school diploma or less vs. some college or more), household income ($29,999 per year or less vs. $30,000 or more), and history of hypertension, cerebrovascular accident, cardiovascular disease, diabetes mellitus, hyperlipidemia, and a history of any cancer. For comparison, equivalent single-predictor or multiple-predictor models were fit using subjects’ self-report of having consumed a total of 100 or more lifetime cigarettes as a predictor in place of methylation.

Results

Demographic, clinical, and laboratory study characteristics of deceased and alive subjects are given in Table 1. In total, 15 of 193 subjects retained for analysis were confirmed deceased by death certificate matching. Deceased subjects were older, more likely to be male, white, and have less education. Those deceased also had a slightly lower average BMI (27.2) than living subjects (28.8) and were more likely to have a history of hypertension, cardiovascular disease, diabetes mellitus, and any cancer. Deceased subjects were more likely to have a smoking history of 100 or more cigarettes (73.3%) than those alive (52.8%). Finally, deceased subjects had lower mean percent methylation at cg05575921 (57.5%) than living subjects (72.1%), although the difference was not significant (t=2.00, df=15.339, p=0.06391).

Table 1.

Sample Demographics

Sample	Living	Deceased
N	178	15
Median Age*	55.3	63.7
Sex
Male (%)	89 (50)	12 (80)
Female (%)	89 (50)	3 (20)
Race/Ethnicity
White (%)	170 (95.5)	15 (100)
Non-White (%)	8 (4.5)	0 (0)
Education (N=190)
High School or Less	34 (19.4)	5 (35.7)
Some College or More	141 (80.6)	9 (64.3)
Household Income (yearly) (N=186)
< $30,000	24 (19.4)	2 (35.7)
$30,000 or more	147 (80.6)	12 (64.3)
Marital Status (N=193)
Currently Married	124 (69.7)	10 (66.7)
Not Currently Married	54 (30.3)	5 (33.3)
Medical Risk Factors
Mean BMI (SD)	28.8 (6.2)	27.2 (6.7)
Hypertension (%)	21(11.8)	4 (26.7)
Stroke (%)	4 (2.2)	0 (0)
Cardiovascular disease (%)	2 (1.1)	1 (6.7)
Diabetes mellitus (%)	14 (7.9)	2 (13.3)
Hyperlipidemia (%)	4 (2.2)	1 (6.7)
Cancer (%)	8 (4.5)	2 (13.3)
Smoking Indicators
Lifetime consumption of >=100 cigarettes (%)	94 (52.8)	11 (73.3)
Cg05575921 methylation (%) (SD)	72.1(20.6)	57.5 (27.7)

^*Indicates age at censoring for living subjects, or age at death for deceased subjects.

Examination of the primary causes of death and additional entity axis codes for deceased subjects revealed a strong pattern of smoking-associated morbidity, as shown in Table 2. Of the 15 primary causes of death, smoking is a known major risk factor for 10, including 6 neoplasms, chronic obstructive pulmonary disease, atrial fibrillation, diabetes, and stroke. In one other case atherosclerotic disease was noted as an additional entity axis code.

Table 2.

Causes of death among deceased subjects.

Sex	Age at death	Primary cause of death	Additional codes
M	65	Necrotizing fasciitis
M	65	Acute myeloblastic leukemia
M	62	Chronic obstructive pulmonary disease with acute lower respiratory infection	Peripheral vascular disease, unspecified
F	63	Malignant neoplasm of unspecified part of bronchus or lung
M	73	Malignant neoplasm of pancreas, unspecified
M	62	Pneumonitis due to inhalation of food and vomit
M	64	Atrial fibrillation and flutter	Other disorders of lung
M	52	HIV disease resulting in other bacterial infections
M	54	Accidental poisoning by and exposure to antiepileptic, sedative hypnotic, antiparkinsonism, and psychotropic drugs, not elsewhere classified	Atherosclerotic cardiovascular disease, so described
M	71	Malignant neoplasm of unspecified part of bronchus or lung
M	58	Unspecified diabetes mellitus with multiple complications
F	56	Malignant neoplasm of breast of unspecified site
M	68	Cerebrovascular disease, unspecified
F	64	Car driver injured in collision
M	65	Malignant neoplasm of tongue, unspecified

The Cox regression model showed that methylation at cg05575921 was a significant predictor of all-cause mortality (p<0.05), with a model R² of 2.8% and a concordance value of C = 0.626. In comparison, a self-reported history of lifetime consumption of 100 or more cigarettes was not a significant predictor of mortality.

In a model adjusted for major demographic variables and clinical risk factors, methylation at cg05575921 remained a significant predictor of mortality (p<0.005), the R² of the model improved to 13.7% and the C was 0.813, indicating strong predictive ability of the full model. Additional significant predictors in the full model included age (p<0.05) and male gender (p<0.05). The estimated hazard ratio for methylation was 0.9665 indicating that the hazard was increased by 3.5% for a 1% decrease in methylation, adjusting for all other predictors. As an illustration of the methylation effect, Figure 1 shows the model-based survival probability as a function of time for subjects at the 25^th and 75^th percentiles of methylation.

Figure 1.

Model-based survival probability based on cg05575921 methylation with clinical and demographic risk factors as a function of time for subjects at the 25^th (red line) and 75^th percentiles (blue line) of methylation. The y-axis indicates the probability of survival while the x-axis indicates time in years.

In an equivalent model substituting lifetime consumption of 100 or more cigarettes for methylation, cigarette consumption was not a significant predictor of mortality, while age (p<0.05) and male gender (p<0.05) remained significant predictors.

Discussion

In this study, we show that cg05575921 methylation status, a new epigenetic biomarker for smoking, is a significant predictor of early mortality in a high-risk longitudinal sample of adoptees. Cox proportional hazards regression models adjusted for major demographic and clinical risk factors for mortality continued to demonstrate the effect, and in fact the significance of the effect became even stronger. The predictive ability of the full model combining cg05575921 methylation and major clinical and demographic risk factors was particularly strong, with a Harrell’s C of over 0.8 ³⁹. Examination of the cause of death of those deceased in our sample revealed a majority of deaths were smoking-related.

Although methylation at cg05575921 has previously been linked with increased morbidity and mortality, ^40,41 to our knowledge this is the first time this predictive ability has been demonstrated with methylation sensitive digital PCR methods that are suitable for clinical translation. The use of this novel ddPCR assay for cg05575921 methylation may have significant clinical impact by either quantitating the extent or smoking or by allowing clinicians and others to detect surreptitious smoking in those who deny use, or in whom smoking is light or intermittent and are therefore less likely to have a positive serum cotinine test. ^42,43

However, it isn’t the first time DNA methylation has been shown to be a predictor of all-cause mortality. ^44–48 A series of studies from Hermann Brenner’s Heidelberg group has shown that DNA methylation status at markers associated with smoking is a significant predictor of mortality.^44–47 The first work, published in 2014, showed that methylation status at F2RL3, was a strong predictor of both smoking and all-cause mortality in Europeans. ⁴⁴ Although subsequent work that this effect at F2RL3 with respect to smoking and perhaps mortality may be confounded by ethnic specific genetic confounding in other populations, they subsequently expanded this approach to show that methylation at other smoking markers that are not confounded by ethnic specific variation, such as cg05575921, work equally if not even better than F2RL3 methylation in predicting mortality. ^45,46 Most recently, using death certificate and genome wide epigenetic data from whole blood DNA from 1954 subjects from the ESTHER cohort, they analyzed the relationship of DNA methylation to mortality and showed that DNA methylation status strongly predicted subsequent death. Most notably, this genome wide signature was strongly correlated with smoking status and disease directly linked to smoking with the strongest genome wide effects being observed at cg05575921. ⁴⁷

When considering these findings it is also important to appreciate the difference between using DNA methylation to predict mortality versus inferring biological age as well as the relationship of these epigenetic indices to smoking. Beginning with the work of Boks, Belotti and Christiansen, a number of groups have also, ^49–51 a number of groups have used methylation data to infer “epigenetic age”, which is an estimator of actual chronical age. More recently, many groups have used either indices devised by ⁵² Hannum and colleagues in 2013 or Horvath ⁵³ and associates to predict these ages. However, the link between these inferred ages and differential mortality is not well understood with only a handful of groups actually comparing their “epigenetic clock” result to actual mortality with one notable exception. In a meta-analysis using DNA methylations status from 13,089 subjects with both genome wide methylation data from the Illumina 450K platform and death status information, Horvath and colleagues analyzed the relationship of “biological age” to mortality and found that the Horvath approach was a significant, yet modest predictor of mortality (overall Hazard ratio 1.029). ⁵⁴ The amount of that risk that is secondary to smoking is unknown. Despite Horvath and colleague’s claim that the Horvath index does not load on smoking status, other groups have analyzed the relationship of the Horvath index to smoking status. In 2016, again using data from the ESTHER cohort, Brenner and colleagues showed that “biological age”, as calculated by the Horvath method, and epigenetically measured smoking status are significantly correlated even after controlling for self-reported smoking status, and speculated that the reason for their finding may be underreporting of smoking by participants in the ESTHER study.⁵⁵ Similarly, using publically available data sets, our group has also shown that methylation status at a significant fraction of both the Hannum and the Horvath indices are significantly correlated with smoking. Hence, although these and other studies ⁵⁶ show that aging indices can also clearly be used to predict mortality, the amount of variance that is predicted that is independent of smoking history is not clear.

Potential advantages of using the current approach are speed, cost and interpretability. The ddPCR method described is relatively affordable (total cost of all reagents for a 96 well plate is approximately $5 per sample), can be carried out in a matter of five hours with the output (percent methylation) being easily interpretable. In contrast, array based methods are rather costly (~$400–500 per sample) and take both considerable time (4–5 days) and expertise to perform. Perhaps most concerning, the Illumina platform for which these indices were compiled (the Illumina 450K array) is no longer commercially available. This is important because: 1) only 422,524 of the 485,577 probes from the 450K array were retained on the new Infinium MethylationEPIC BeadChip (a.k.a. Epic array), and 2) Logue and colleagues, using data from 145 whole blood sample profiled on both platforms, that the methylation from at least 55% of probes had a correlation (r) of less than 0.2. ⁵⁷ Therefore, if replicated, this ddPCR method may be an attractive approach for imputing mortality risk for investigators on a modest budget or lacking the means to implement the statistical approaches of either Horvath or Hannum.

Like all studies, this study has sources of potential error or bias. Weaknesses of this study include a relatively small sample size and short observation period. Continued follow up of this cohort, which has a high rate of smoking and age range from the 50s-70s is be expected to strengthen the effect seen in this study. Additional epigenetic measurement including ascertainment of other substance use patterns, for example alcohol, ⁵⁸ and epigenetic loci that have been linked to an “epigenetic clock” ^52,53 may also improve our ability to predict early mortality in this and other samples.

In summary, we report that the methylation status at cg05575921 as obtained by methylation sensitive ddPCR is a significant predictor of mortality. Additional research to expand and corroborate these findings is needed.