Interaction between Continuous Pack-Years Smoked and Polygenic Risk Score on Lung Cancer Risk: Prospective Results from the Framingham Heart Study

Lung cancer is the leading cause of cancer mortality in the United States (1). As the primary risk factor for lung cancer, 80% to 90% of these deaths are linked to cigarette smoking (2). The risk of lung cancer due to smoking is dose dependent such that those with a longer and/or heavier smoking history are at greater risk of developing lung cancer (3–6). A person's smoking history can be represented through pack-years: packs of cigarettes smoked per day (20 cigarettes per pack) multiplied by the years smoked. Importantly, genetic susceptibility to lung cancer is also well established, with prior studies providing strong support for the contribution of germline genetic variation to lung cancer (7–21). Many of the genetic loci that convey a greater risk of lung cancer are also associated with smoking behaviors, (9, 10, 15, 18, 22, 23) which supports the presence of a gene-by-smoking interaction for lung cancer (19, 20, 24–27).

Wang and colleagues recently evaluated gene–environment interactions on lung cancer risk using genotype data collected from the prospective UK Biobank (26). They observed that smoking was the leading risk factor for lung cancer incidence with an unweighted population attributable fraction of 63.73% and that smoking displayed a positive additive interaction with their polygenic risk score (PRS) to influence lung cancer risk (26). The interaction between smoking and high genetic risk possessed an unweighted population attributable fraction of 17.85% (26). However, they classified individuals with a history of smoking as heavy (≥40 pack-years) or non-heavy (<40 pack-years), genetic risk as “low,” “medium,” or “high,” and analyzed the categorical versions of these variables rather than using them continuously, which could potentially lead to biased results if the chosen cut-off points do not correspond to clinical decision points or a true difference in risk.

Although prior findings support a gene-by-smoking interaction, few have examined the dose-dependency by incorporating cumulative pack-years (26) and have instead simplified smoking history to current/former/never smoking status (19). Extending on these findings, we sought to evaluate the presence of a potential interaction between genetic risk and cumulative pack-years smoked, both examined continuously. Furthermore, many prior studies have used a case–control design without follow-up or multiple assessments of smoking status over time. To this end, we applied a PRS to the Original and Offspring cohorts of the Framingham Heart Study (FHS), with a median of greater than nine smoking assessments over five decades of follow-up, to examine the interplay between genetic risk and lifetime pack-years smoked on lung cancer risk.

This investigation includes data from the FHS, a longitudinal cohort of community-dwelling individuals from the town of Framingham, Massachusetts which was established in 1948 to determine the risk factors for cardiovascular disease (28). Original cohort participants presented for biennial examinations during which routine physical examinations were conducted and health-related questionnaires were completed. In 1971, children of the Original cohort participants, as well as other community dwelling individuals were enrolled into the Offspring cohort; follow-up examinations occur every 4 years (29). As FHS is community-based, the cohort characteristics reflect the town's residents at the time of enrollment in that the participants are almost exclusively middle class individuals of European ancestry. The current analysis includes participants in the Original (28) and Offspring (29) cohorts who attended their fourth (1954–1958, n = 4,541) and first (1971–1975, n = 5,122) examination cycles, respectively. Included participants were free of lung cancer at baseline and possessed complete data on smoking history and genetic information (Fig. 1). Following exclusions, our analytic sample included 2,924 individuals (454 Original cohort; 2,470 Offspring cohort) who attended a total of 29,564 examinations (referred to hereafter as “person-examinations”). Notably, DNA was extracted from stored blood samples in the late 1980s and early 1990s upon receipt of written participant consent in accordance with the Belmont Report; by this point, many of the Original cohort participants had passed away and were unable to consent. As such, genetic data were obtained on fewer than 30% of Original cohort participants (30). Of the 2,470 Offspring cohort participants in our sample, 393 have one parent in the FHS study and 83 have both parents in the cohort. Participant characteristics were assessed regularly via in-person clinic examination throughout the follow-up period: approximately every 2 years for the Original cohort (28), and every 4 years for the Offspring cohort (29). This study was approved by the Vanderbilt University Medical Center Institutional Review Board.

FHS participants were continuously followed for diagnosis of lung cancer during the follow-up period from the baseline examination [examination cycle 4 (Original cohort); examination cycle 1 (Offspring cohort)] until the end of 2013 (Original cohort) or 2018 (Offspring cohort). Follow-up for the Original cohort concluded at the end of 2013 because fewer than 40 participants remained alive and were ages 96 years on average. Lung cancer incidence was adjudicated following standardized protocols, including a review of medical records and pathology and laboratory reports (31). In brief, FHS participants (or their designated family member if the participant is unable to respond for themselves due to death or illness) and participants’ primary care physicians are contacted at the end of each calendar year and are asked to report any hospitalizations, new cancer diagnoses, or death. Study staff then review this information along with data collected during routine Study examinations (biennial for Original cohort, quadrennial for Offspring cohort) to adjudicate cancer diagnoses. If more information is needed to confirm the diagnosis or cancer location, all hospital records are obtained from the patient's physician, with consent, for review. In the presence of remaining uncertainty, a surgical oncologist reviews the information to adjudicate the diagnosis.

For this analysis, we used existing genotyping data from the Affymetrix GeneChip Human Mapping 500K Array and the 50K Human Gene Focused Panel platforms which are included as part of the FHS SNP Health Association Resource (SHARe) data available in the database of Genotypes and Phenotypes (dbGaP; ref. 32). Biospecimens for DNA extraction were collected from FHS participants between 1971 and 2002 and genotyped on the Affymetrix 500K and MIPS 50K arrays. Genotyping data were mapped to genome build 37 (32). Depending on when an individual's DNA was collected, there is potential for immortal time bias because there are potentially decades between the baseline examination and DNA extraction. A large amount of time between baseline and genetic testing will bias our results if many people die from lung cancer prior to DNA collection. Thus, we examined the baseline characteristics of individuals excluded because of a lack of genetic information versus those included in the sample; no significant differences were observed. After assessing the possibility of immortal person-time bias but before combining phenotypic and genotypic data, we performed quality control of the genotyping data (33).

Imputed genotyping data on FHS participants were downloaded from dbGaP (phs000710.v1.p1). Genotypes were previously imputed to the 1000 Genomes using MACH (version 1.00.15) and HapMap (release 22, build 36, CEU) as the imputation backbone. From the imputed data, we excluded SNPs with an average call rate <90%, in linkage disequilibrium (r² > 0.7), or minor allele frequency <5% (34).

Principal components to identify population substructure were estimated using EIGENSTRAT and were provided by FHS (35, 36).

To build a lung cancer PRS, we used summary statistics from a previously performed genome-wide association study (GWAS) in individuals of European ancestry from the OncoArray Consortium (7). The OncoArray genotyping platform covers 533,631 SNPs that passed quality control procedures (18); of these, 517,482 SNPs passed the filtering algorithm described by McKay and colleagues (adherence to Hardy–Weinberg equilibrium and call rate >95%) and were included in their analyses of lung cancer risk (7). Genotypes were then imputed using the 1000 Genomes (phase III) as the reference panel (18). The OncoArray Consortium sample used in the GWAS for lung cancer risk contains data on 85,716 participants from 26 cohorts; FHS was not among them. Smoking data were available on 50,046 OncoArray Consortium participants; of those, 40,187 (80%) had ever smoked, and 20,833 (42%) were current smokers (7). We determined the overlap between the OncoArray and the imputed FHS genotyping data and built our PRS from the shared risk variants. We used PRSice (37), a P-value thresholding method, to develop our PRS. After linkage disequilibrium pruning, we retained 83,304 SNPs. We then constructed two PRSs: one weighted by the regression coefficient (log-odds) between the SNP and lung cancer risk, that is, β, referred to as PRS_β hereafter, and the second (PRS_β/Var(β)) was weighted by the regression coefficient divided by its estimated variance, that is, β/[Var(β)] to incorporate a measure of uncertainty.

Collection and construction of smoking variables, including pack-years smoked, have been described previously (38, 39). At the baseline examination (cycle 4 for Original cohort, cycle 1 for Offspring cohort), data on current and prior smoking behaviors were collected. For current and former smokers, information was obtained on age at which the participant started smoking, usual number of cigarettes smoked per day in the past, age quit smoking (former smokers), and current number of cigarettes smoked per day. From these data, we calculated pack-years at baseline for both current and former smokers; never smokers were assigned a pack-years value of 0. Pack-years were updated at each examination in the follow-up period as described below.

At postbaseline examinations, data on current smoking status and cigarettes per day were collected, allowing us to calculate cumulative smoking exposure. For a given participant, smoking status (current, former, never) could change over time such that each participant contributed person examinations and person time to the category reflecting his or her status at each assessment. If an individual developed lung cancer, this event counted only in the group to which the individual belonged at the time of the event. The median number of examinations during which smoking was assessed was 22 for the Original cohort and 9 for the Offspring cohort.

We identified age, sex, education, genetics, population substructure, and family history as potential confounders of the association between smoking and lung cancer. Age (in years), sex (male/female), and education at the baseline examination (less than high school graduate, high school graduate, more than high school education) were available in the FHS examination data. Population substructure was identified from principal components as described above. Lung cancer family history was not available in these data and was therefore not included in models. Dynamic predictors (smoking variables and age) were updated at each examination while static predictors (PRS, population substructure, sex, education) remained constant over follow-up. An example timeline of exams, including variable updates and censoring, is displayed in Supplementary Fig. S1.

In the FHS data from in-person examinations, missingness was relatively low. In the Original cohort, 87% of person-exams had no missing data while in the Offspring cohort, 98% of person-exams had no missing data. Education level was most frequently missing in the Offspring cohort with a mere 2.1% missingness. Smoking status was not assessed during all the Original cohort exams and was missing at 13% of person-examinations. However, smoking status was collected at all Offspring examinations and was missing at <1% of person-exams in this cohort. All other variables had <5% missingness.

Missing covariate data were handled using multiple imputation by chained equations techniques to produce five complete datasets for analysis. We imputed continuous variables through the use of predictive mean matching (40) to produce imputed values that are clinically plausible. Categorical variables were imputed using the discriminate function with a noninformative Jeffrey prior (41). Results across imputed datasets were combined according to Rubin's rules (42).

We calculated baseline summary statistics by smoking status. Means and SD were reported for normally distributed continuous variables, while medians along with the 25th and 75th percentile were reported for continuous variables with skewed distributions. We report counts and percentages for categorical variables. We then plotted the distribution of pack-years smoked stratified by smoking status (former/current) and calculated the Pearson correlation between each of the PRSs and continuous pack-years to assess their association. Next, we determined how many parent-child sets existed in our data in which the child and at least one parent developed lung cancer.

Using Poisson regression with an offset term equal to the natural logarithm of follow-up time, we calculated lung cancer incidence rates per 1,000 person-years stratified by smoking status; current and former smokers were further categorized by above/below 20 pack-years as this was the median pack-years in our sample over all included person-examinations as well as the eligibility threshold for initiating lung cancer screening in the United States (43).

After confirming the proportional hazards assumption via the interaction of continuous pack-years and PRS (separately) with the natural logarithm of follow-up time, we fit Cox proportional hazards regression models with incident lung cancer as the outcome. Because the FHS includes related individuals, we used mixed-effects Cox proportional hazards regression in our analyses that incorporated the kinship matrix to adjust the variance accordingly. Models included continuous pack-years, smoking status, lung cancer PRS, principal component 1 [PC1 (representing European ancestry)], age, sex, and educational attainment as independent variables.

We first assessed whether the PRS modified the effect of the association between continuous pack-years and lung cancer risk using a test of heterogeneity. In the presence of heterogeneity, we assessed the lung cancer risk associated with each increase of 10 pack-years at the mean PRS and 1 SD above and below the mean (separate models for each PRS). We then examined the risk associated with a one SD increase in each PRS at 0, 10, 20, and 30 pack-years smoked. Note that assessing the risk at 0 pack-years is analogous to estimating risk among those who have never smoked. Finally, we calculated the area under the ROC curve (AUC) for the models incorporating the interactions between each PRS and pack-years smoked. Tests for heterogeneity, modeling, and calculation of the AUC were performed for both versions of the PRS.

A two-sided P value <0.05 was considered statistically significant except for tests for heterogeneity which considered a more liberal two-sided P value <0.1 significant as our number of events was relatively low and tests for interactions are typically underpowered (44, 45). For all regression models, missing data were handled via multiple imputation by chained equation techniques. Analyses were performed in SAS 9.4 and R 4.0.1.

Genotypic and phenotypic data for this project were obtained from dbGaP (phs000342.v20.p13: NHLBI Framingham SHARe) and are available to other researchers upon data request approval through dbGaP at https://www.ncbi.nlm.nih.gov/projects/gap/cgi-bin/study.cgi?study_id=phs000342.v20.p13.

Our sample included 2,924 FHS participants: 1,170 who had never smoked, 520 who formerly smoked, and 1,234 currently smoking individuals. At study baseline, those who had never smoked or currently smoke had an average age of 35 years compared with formerly smoking individuals, who had a baseline age of 38 years (Table 1). Women were most likely to have never smoked, while men were most likely to currently smoke. Currently smoking individuals consumed a median of one pack per day (20 cigarettes). Individuals who had formerly smoked had a median of 11 pack-years compared with currently smoking individuals, who had a median of 13.4 pack-years. Consistently, the distribution of pack-years was more heavily right-skewed among current than former smokers (Supplementary Fig. S2). While PRS distributions were similar across smoking status (Table 1), we observed a positive correlation between the PRSs and continuous pack-years (r = 0.042, P = 0.0226 for PRS_β; r = 0.040, P = 0.0312 for PRS_β/Var(β)).

Among the 476 Offspring cohort participants in our sample with at least one parent also in the sample, 393 had a single parent in the sample; both parents of the remaining 83 were in the sample. Of the 393 with one parent in the study, 4 developed lung cancer, 1 of whom had a parent diagnosed with lung cancer as well. None of the Offspring participants with both parents in the study developed lung cancer. Thus, of 476 Offspring participants with at least one parent also in the sample, there was only one parent-child set (0.2%) in which both members developed lung cancer.

Using PRSice, we determined that the highest R² value between lung cancer status and a PRS was achieved when including the 120 SNPs with P < 6 × 10⁻⁵ (Supplementary Fig. S3). When examining the PRS distributions by incident lung cancer status, we found that a greater burden of risk alleles was present in those who developed lung cancer than those who did not (Supplementary Fig. S4).

Among the 2,924 individuals in our sample, 86 were diagnosed with incident lung cancer. We observed that both formerly and currently smoking individuals had higher lung cancer incidence rates than individuals who never smoked, with the highest rates in current smokers (Table 2). When stratifying smokers’ incidence rates by above/below 20 pack-years, point estimates were slightly higher in formerly smoking individuals in both categories than currently smoking individuals, but confidence intervals overlapped, indicating no significant difference.

We then assessed the presence of a PRS-by-pack-years interaction on lung cancer risk by performing tests for heterogeneity for each PRS interacting with pack-years. For both PRSs, P values were ≤0.1 (PRS_β×pack-years P = 0.09; PRS_β/Var(β)×pack-years P = 0.098), and the beta coefficient for the interaction term was negative, indicating the presence of an antagonistic (i.e., submultiplicative) interaction. Because of this heterogeneity, we analyzed the effect of each additional 10 pack-years smoked at the mean PRS, one SD below the mean PRS, and one SD above the mean PRS. Similarly, we assessed the impact of each SD increase in the PRS at 0, 10, 20, and 30 pack-years smoked.

Using PRS_β/Var(β) as an example, each additional 10 pack-years smoked was significantly associated with: a 56% increase in lung cancer risk at one SD below the mean of PRS_β/Var(β), a 48% increase in risk at the mean of PRS_β/Var(β), and a 40% increase at one SD above the mean of PRS_β/Var(β); a similar pattern was observed for PRS_β (Table 3). Continuing to use PRS_β/Var(β) as an example, we observed that each SD increase in PRS_β/Var(β) was significantly associated with: a 55% increase in the risk of lung cancer at 0 pack-years (individuals who had never smoked), a 47% increase in risk at 10 pack-years, a 39% increase in risk at 20 pack-years smoked, and a 32% increase in risk at 30 pack-years smoked; a similar trend was observed PRS_β (Table 4). Thus, as each PRS increased, the effect of additional pack-years smoked conferred a relatively smaller risk of lung cancer, and as pack-years increased, the effect of an additional pack-years smoked was associated with a smaller, but still significant risk of lung cancer. This is consistent with our observation of a submultiplicative interaction.

The models which incorporated the interaction between each PRS and pack-years smoked possessed strong discrimination, with AUC values above 0.85 (AUC for model with PRS_β = 0.8624, AUC for model with PRS_β/Var(β) = 0.8616).

To our knowledge, this is the first investigation to observe a gene-by-smoking interaction in a prospective cohort with continuous pack-years over a 50-year period and adjudicated lung cancer incidence. We confirmed that pack-years smoked remains a strong risk factor for lung cancer, even when adjusting for genetic contribution. After accounting for the interaction between smoking and genetic risk, both the PRS and pack-years remained significantly associated with lung cancer incidence. Because the interaction was submultiplicative, the risk of lung cancer associated with increasing pack-years smoked was greatest among those with the lowest genetic risk and the risk associated with genetic risk was greatest among those who had never smoked; this is consistent with other recent findings (46). Thus, although there is an interaction between the genetic risk of lung cancer and pack-years smoked, it is not so strong as to indicate that there is any level of genetic risk at which smoking becomes “safe”—the long-term benefits of smoking cessation, including decreased mortality both pre- and post-cancer diagnosis, cannot be overstated (38, 47, 48).

While the presence of a gene-by-smoking interaction has been observed previously (19, 24–27, 46, 49), we are among the first to examine the interaction between a lung cancer PRS comprised of hundreds of SNPs and continuous pack-years smoked (25). VanderWeele and colleagues were among the first to report a significant gene-by-smoking interaction on lung cancer, but they examined only two SNPs on chromosome 15 (24). Similarly, Zhou and colleagues observed that 8092 C>A polymorphism in ERCC1 is able to modify the associations between polymorphisms rs1316298 and rs4589502 and smoking, demonstrating some of the missing heritability that exists in lung cancer susceptibility (49). More recently, in a study of 413,870 UK Biobank participants, Kachuri and colleagues observed that PRS was associated with lung cancer risk, and noted that modifiable risk factors, including smoking, appeared to be stronger risk stratifiers than the PRS, supporting the use of both the PRS and smoking history to estimate lung cancer risk (50). Dai and colleagues also constructed a PRS containing 19 SNPs for lung cancer in a prospective sample of Chinese men and women and found it to be associated with diagnosis of lung cancer (21). While they did not directly test for effect modification of this association by pack-years smoked, they did perform stratified analyses in non-smoking individuals as well as heavy (≥30 pack-years) and non-heavy smoking individuals (<30 pack-years), and observed synergistic effect modification such that those with high genetic risk and a ≥30 pack-year smoking history were at the greatest lung cancer risk (21).

Hung and colleagues recently constructed a weighted PRS with 128 SNPs using genetic data from the OncoArray Consortium as we did here. However, they validated their PRS in the UK Biobank to assess the PRS's utility in predicting lung cancer and identifying individuals eligible for lung cancer screening (19). On the whole, our results and theirs complement one another. Although only six of the SNPs in Hung's PRS overlapped with the 120 included in our PRS, they also observed a greater lung cancer risk with increasing PRS. The authors also observed an interaction between smoking status and their PRS (19) but did not assess the role of pack-years smoked. Similarly, Shi and colleagues observed a submultiplicative interaction between smoking and a PRS such that the association between the PRS and lung adenocarcinoma was stronger in individuals without a history of smoking than in those with a history of smoking (46), but as they did not have individual-level data, they were unable to assess the interaction between pack-years and their PRS.

Our observed submultiplicative interaction could be because many of the genetic loci which convey greater risk of lung cancer are also associated with smoking behaviors (9, 10, 15, 18, 22, 23), so the PRS and pack-years smoked are likely capturing some amount of shared variability in lung cancer incidence. This is also consistent with our observation that the PRSs and pack-years smoked are positively correlated. For example, three genes that are strongly associated with lung cancer which were identified in GWAS are nicotinic acetylcholine receptors, CHRNA3, CHRNB4, and CHRNA5 (9, 10, 15, 18), all of which reside on chromosome 15. These same loci are also associated with smoking behaviors, including smoking intensity, topography, duration of exposure, and successful cessation (9, 22, 23, 51). In addition, CYP2A6 and CYP2B6 are part of the P450 system and encode for enzymes governing nicotine metabolism and activating tobacco-specific nitrosamines (i.e., the carcinogens in cigarettes; refs. 52, 53) which influence cigarette consumption and smoking cessation (9, 54–56). Those who metabolize nicotine quickly tend to smoke more and are therefore exposed to more carcinogens. It is believed that CYP2A6 and CYP2B6 are associated with lung cancer because they increase an individual's propensity to smoke both longer and more heavily (57). These associations support the biologic plausibility of a submultiplicative gene-by-smoking interaction for lung cancer (19, 20, 24–27). Among never smokers, there is evidence that genetic mutations due to environmental exposures, like pollution and radon, may cause lung cancer (58, 59). For example, a recent investigation reveals that EGFR mutations, which are common among never smokers with lung cancer, are elevated among persons living in areas with high PM2.5 pollution (58). In addition, it is known that the chromosome 15q25 locus is associated with lung cancer among non-smokers (10); this association may be modified by radon exposure (59).

The decades-long follow-up and comprehensive smoking variables that are regularly collected in FHS allowed us to incorporate pack-years as a continuous time-varying covariate. We were therefore able to produce more precise estimates of the effect of pack-years smoked than is available in most other data sources and highlights the strength of our findings when compared with case–control studies. Specifically, inclusion of continuous pack-years as opposed to categorical smoking status (current/former/never) or categorical smoking intensity (heavy vs. non-heavy) allowed us to demonstrate the graded contribution of genetics to lung cancer risk as smoking duration and intensity (captured via continuous pack-years) increased. Thus, we encourage researchers to collect and model smoking history using continuous pack-years whenever possible as it allows the description of more granular associations between smoking and outcomes.

We acknowledge the lack of statistical power resulting from a relatively small sample size and low number of incident lung cancer events; the existence of relatedness among participants further decreased our effective sample size and statistical power. In particular, although FHS is a study that includes family members, there is no efficient way to adjust for family history of lung cancer with these data because participants were never asked about their family history of lung cancer and only 19% of Offspring cohort participants in our sample had one or more parent also in our sample. Therefore, the confounding effect of family history could not be accounted for. Given that only 0.2% of the American population has a current or prior diagnosis of lung cancer (58), and only one parent-child set in our data both developed lung cancer, this bias is likely small. Because of sample size concerns, we decided a priori to use an alpha level of 0.1 to assess the presence of an interaction between PRS and pack-years. This threshold is higher than a standard alpha level of 0.05, and as our P values for the interaction were between 0.05 and 0.1, we present results which incorporate this interaction; this would not have been the case had we used an alpha level of 0.05. Results should be interpreted with this choice in mind. We were also unable to include important regions of the genome, including the CYP2A6 nicotine metabolism gene, which is associated with both smoking behavior and lung cancer development, in our PRS because of this region is notoriously challenging to genotype and was not covered by our chip or imputation panel (59). Finally, FHS Original and Offspring cohort participants are predominantly of European ancestry so results may not generalize to individuals of other genetic ancestries.

In conclusion, our results support the presence of a gene-by-smoking interaction on lung cancer risk both reinforcing the negative effect of continued smoking on lung cancer regardless of genetic risk and supporting the notion that lung cancer risk attributable to genetics is highest among those who have never smoked. These results can inform future studies to quantify the value of incorporating genetic information into the linked processes of lung cancer screening and tobacco cessation treatment. Larger studies with both genetic data and longitudinal smoking information should investigate this further.

M.S. Duncan reports grants from NIH, National Center for Advancing Translational Sciences during the conduct of the study. M.C. Aldrich reports grants from NIH/NCI during the conduct of the study; personal fees from Guardant Health outside the submitted work. No disclosures were reported by the other authors.

The funders had no role in the design of the study; the collection, analysis, and interpretation of the data; the writing of the article; or the decision to submit the article for publication.

M.S. Duncan: Conceptualization, data curation, formal analysis, methodology, writing–original draft, writing–review and editing. H. Diaz-Zabala: Writing–original draft, writing–review and editing. J. Jaworski: Data curation, formal analysis, writing–review and editing. H.A. Tindle: Conceptualization, methodology, writing–review and editing. R.A. Greevy: Conceptualization, methodology, writing–review and editing. L. Lipworth: Conceptualization, writing–review and editing. R.J. Hung: Resources, writing–review and editing. M.S. Freiberg: Conceptualization, resources, methodology, writing–review and editing. M.C. Aldrich: Conceptualization, resources, methodology, writing–review and editing.

M.S. Duncan was supported by the National Center for Advancing Translational Sciences at NIH (KL2TR001996). M.C. Aldrich was supported by the NCI at NIH (R01CA251758 and U01CA253560). H. Diaz-Zabala acknowledges the support of the NCI at NIH through the VUMC Molecular and Genetic Epidemiology of Cancer training program (T32CA160056). H.A. Tindle acknowledges the support of the William Anderson Spickard, Jr, MD Chair in Medicine. M.S. Freiberg acknowledges the support of the Dorothy and Laurence Grossman Chair in Cardiology. The Framingham Heart Study is supported by contract no. 75N92019D00031 from the National Heart, Lung, and Blood Institute (NHLBI), NIH, and Department of Health and Human Services with additional support from other sources.