|Home | About | Journals | Submit | Contact Us | Français|
Nicotine dependence is moderately heritable, but identified genetic associations explain only modest portions of this heritability. We analyzed 3,369 SNPs from 349 candidate genes, and investigated whether incorporation of SNP-by-environment interaction into association analyses might bolster gene discovery efforts and prediction of nicotine dependence. Specifically, we incorporated the interaction between allele count and age-at-onset of regular smoking (AOS) into association analyses of nicotine dependence. Subjects were from the Collaborative Genetic Study of Nicotine Dependence, and included 797 cases ascertained for Fagerström nicotine dependence, and 811 non-nicotine dependent smokers as controls, all of European descent. Compared with main-effect models, SNP x AOS interaction models resulted in higher numbers of nominally significant tests, increased predictive utility at individual SNPs, and higher predictive utility in a multi-locus model. Some SNPs previously documented in main-effect analyses exhibited improved fits in the joint-analysis, including rs16969968 from CHRNA5 and rs2314379 from MAP3K4. CHRNA5 exhibited larger effects in later-onset smokers, in contrast with a previous report that suggested the opposite interaction (Weiss et al, PLOS Genetics, 4: e1000125, 2008). However, a number of SNPs that did not emerge in main-effect analyses were among the strongest findings in the interaction analyses. These include SNPs located in GRIN2B (p=1.5 × 10−5), which encodes a subunit of the NMDA receptor channel, a key molecule in mediating age-dependent synaptic plasticity. Incorporation of logically chosen interaction parameters, such as AOS, into genetic models of substance-use disorders may increase the degree of explained phenotypic variation, and constitutes a promising avenue for gene-discovery.
Progress in understanding the genetics of nicotine dependence has advanced rapidly in recent years, with numerous replications of the association between non-synonymous polymorphism on the nicotinic receptor gene CHRNA5 (rs16969968 and highly correlated proxy SNPs) and nicotine dependence (Amos et al. 2008; Berrettini et al. 2008; Bierut et al. 2008; Grucza et al. 2008; Thorgeirsson et al. 2008; Weiss et al. 2008; Caporaso et al. 2009; Chen et al. 2009; Saccone et al. 2009; Wang et al. 2009). Additional promising findings have been identified through both genome-wide and large-scale candidate gene approaches (Bierut et al. 2007; Saccone et al. 2007; Caporaso et al. 2009). However, for both nicotine dependence in particular, and complex diseases in general, a paradox has emerged that discoveries to date explain only small portions of the variance for such phenotypes (Orr and Chanock 2008; Goldstein 2009; Hirschhorn 2009; Kraft and Hunter 2009). Factors invoked to explain the “missing heritability” include polygenic effects (very large numbers of polymorphisms having very small effects), undiscovered rare variants, copy-number variants that have yet to be systematically analyzed, and interactions among genes, or between genes and environmental factors (Maher 2008; Phillips 2008; Goldstein 2009). Investigation of these potential sources of undiscovered phenotypic variation can lead to novel genetic discoveries, provide new insights into biological pathways involved in disease, and potentially provide clues to environmental interventions capable of mitigating genetic risk.
One of the key risk factors for addiction in general, as well as for nicotine dependence in particular, is early onset of substance use, as assessed by retrospectively reported age at first use or age at onset of regular use (Robins and Przybeck 1985; Breslau et al. 1993; Breslau and Peterson 1996; Grant and Dawson 1997; Grant et al. 2006; Grucza et al. 2008; Agrawal et al. 2009). Recently, advances in developmental neurobiology have pointed to adolescence as a potential critical period of vulnerability for the development of substance use disorders (Chambers et al. 2003; Crews et al. 2007). The development of adolescent animal models for nicotine dependence has enabled researchers to show that both neurobiological changes and long-term behavioral sequelae differ for adolescent as compared to adult nicotine exposure (Trauth et al. 2000; Trauth et al. 2001; Slotkin 2002; Adriani and Laviola 2004; Slotkin et al. 2008; Kota et al. 2009; Slotkin and Seidler 2009). Hence, the molecular mechanisms and neuroadaptive processes involved in adolescent exposure and adult exposure may differ considerably. The ramifications of these animal models for human genetic analysis are clear: If the etiology for the development of nicotine dependence differs by developmental stage, then the age at which smoking is initiated could be a critical moderator of SNP-associations for nicotine dependence.
In this paper, we examine genotype-phenotype associations with nicotine dependence, and investigate whether modeling statistical interactions between SNP and age-at-onset of regular smoking (AOS) can increase the amount of variation explained by genetic polymorphisms. Specifically, we posit that early and late onset smokers are likely to differ in terms of etiology for development of nicotine dependence, and that this will result in differences in the relative influence of polymorphisms as a function of AOS. We take a large-scale candidate gene approach (i.e., 3,369 SNPs from 348 genes) using data from the Collaborative Study for the Genetics of Nicotine Dependence (COGEND), a case-control study of the genetics of nicotine dependence among smokers. Because we are interested in the full explanatory power of the model (as opposed to the interaction term alone), we utilize a two degree-of-freedom test that jointly evaluates both SNP main effects and potential SNP x AOS interactions (Kraft et al. 2007).
Gene-environment interactions, including interactions with age-at-onset of smoking, have been the focus of several recent candidate gene analyses (Weiss et al. 2008; Chen et al. 2009; Schmid et al. 2009) However, we are unaware of multi-gene analyses of gene-environment interactions for smoking; i.e., analyses in which many polymorphisms are tested in interactive models with putative environmental moderators. Such exploratory approaches may prove to be fruitful for both enhancing our understanding of the etiology of nicotine dependence, as well as for gene discovery and validation (Kraft et al. 2007).
The Collaborative Genetic Study of Nicotine Dependence (COGEND), recruited participants from the St. Louis, Missouri and Detroit, Michigan, metropolitan areas. Both cases and controls were required to have smoked at least 100 cigarettes in their lifetime, a frequently used threshold to define being a smoker (e.g. CDC 2005). In addition to this criterion, subjects were ascertained based on their Fagerström Test for Nicotine Dependence (FTND) scores for their period of heaviest smoking (Heatherton et al. 1991). Cases were required to score 4 or higher, while controls were required to report never having symptoms of dependence (lifetime FTND = 0). Cases and controls were identified through a screening process in which telephone interviews on 54,644 individuals across the two sites were conducted (screening response rate= 74%). Among screened individuals, subjects were ineligible if they did not meet the requirements listed above, or were not within the targeted age range for the study (25–44 years). Among the 5,010 individuals meeting eligibility requirements, 3,137 were interviewed (response rate = 63%), and 2,949 donated a blood sample for genetic testing (response rate = 94%). The initial wave of genotyping was limited to European-Americans ascertained prior to summer 2005, who comprise the 789 cases and 811 controls analyzed here. Procedures were approved by Institutional Review Boards of participating institutions, and were carried out in accord with the Helsinki Declaration.
All subjects were personally interviewed using the Semi-Structured Assessment for Nicotine Dependence (SSAND) which was developed specifically for COGEND and is modeled after the widely use Semi-structured Assessment for the Genetics of Alcoholism (SSAGA)(Bucholz et al. 1994; Hesselbrock et al. 1999). The tobacco section is poly-diagnostic, and includes the assessment of nicotine dependence using both the FTND, and DSM-IV criteria. Alcohol and drug use, abuse, and dependence, and other psychiatric disorder sections are identical to those in the SSAGA.
In addition to case-control status, variables used in these analyses included age at first full cigarette (AFC), age at onset of regular smoking (AOS), age at first full drink (AFD) and age at onset of regular drinking (AOD). All of these variables were retrospectively assessed as single items. Regular smoking was defined as having smoked at least once a week for a period of at least two months. Regular drinking was defined as drinking at least once a month for a period of 6 months or more. Individuals who had never initiated regular drinking were excluded from the AOD analysis. To reduce the influence of extreme values on age at onset variables in regression estimates, AFC, AOS, AFD and AOD were subjected to a “Winsorization” procedure (Tukey 1960). Winsorization involves selecting an equal number of observations at both tails of a distribution, and reassigning their values to be identical to the next-highest or next-lowest value in the distribution. For example, subjects reporting AOS <8 (n=7) were assigned an AOS value of 8, those with AOS over 30 (n=13) were assigned a value of 30, and so on. This transformation is appropriate for variables that have approximately normal distribution, but exhibit kurtosis due to small numbers of outliers (Fernandez et al. 2002; Shete et al. 2004). We selected Winsoriation thresholds that reduced positive kurtosis without inducing negative kurtosis. This corresponded to highest and lowest 1–2% of the sample on each variable. Finally, all of these covariates were recoded so that the zero-value would reflect the sample median on the raw variable (AOS=16), and higher values would reflect risk, or earlier onset, rather than protection. We designate these recoded variables using a prefixed “r”. Hence, rAOS=16-AOS, after Winsorization.
The analyses presented here utilize the candidate gene panel presented in Saccone et al (Saccone et al. 2007), in addition to a number of “fine mapping” SNPs from follow-up genotyping of the nicotinic acetylcholinergic receptor (CHRN) gene family. The initial candidate gene panel consisted of SNPs from genes selected based on known neurobiological and metabolic pathways, previous reports on the genetics of nicotine dependence, and previous reports on the genetics of other psychiatric disorders (see Saccone et al. 2007) for more details on gene and SNP selection). An additional 107 SNPs from 16 CHRN genes were genotyped in a separate experiment to follow-up and fine-map initial discoveries in this gene family (Bierut et al. 2007; Saccone et al. 2007; Saccone et al. 2009).
For the initial candidate gene study, SNPs were chosen to cover physical regions as uniformly as possible, while denser coverage was used for SNPs within exons and near candidate gene promoter regions. Quality control measures are described elsewhere (Saccone et al. 2007; Saccone et al. 2009) and include reliability of genotype-calls, Hardy-Weinberg equilibrium > 0.01, and 95% genotyping call-rate. Because of the additional statistical power-requirements for gene-environment analyses, we eliminated SNPs from analysis if they exhibited a minor allele frequency (MAF) of 4% or less. This resulted in a total of 3369 SNPs tested in this report. These SNPs cover a total of 348 genes.
Our overall purpose is not solely to detect interactions, but rather to understand the genotype-phenotype association in relation to onset age of smoking. As such, it is the joint-significance of the main effect and the interaction terms that is of interest. Hence, for each SNP, we allow for the possibility of statistical interaction between age-at-onset of smoking (and other covariates) and allele count. This involves evaluation of the non-genetic base model, the main-effect model and the interaction model:
where α is the intercept, age is age at interview, sex is coded as 0 or 1, rAOS is age-at-onset of regular smoking (reversed and recoded as described), G is genotype for a given SNP, presumed to be log-additive. The final term in Equation 3 represents the product or interaction term. Note that in Equation 3, the term β4* designates the genetic main effect when rAOS=0, which corresponds to AOS=16, the sample median.
Because we are interested in the full explanatory power of the interaction model (Equation 3) as compared to the main-effect model (Equation 2), we compare each model to the non-genetic base mode (Equation 1). When the main effect model is compared to the base model using a likelihood ratio Chi-square test with one degree of freedom, the significance of the overall main effect (β4) is evaluated. To evaluate the joint significance of the main effect and interaction, the interaction model (Equation 3) is compared with the base model (Equation 1) using a two-degree of freedom likelihood ratio Chi-square test. This approach has been formally explicated by Kraft and colleagues, who showed it to be comparable in power to other approaches for detecting G x E, with little loss of power in detecting main effect association where G x E is non-significant (Kraft et al. 2007). To facilitate evaluation of the goodness-of-fit for both main effect and interaction models, we present Akaike’s information criterion (AIC) for all genetic analyses. AIC is a log-likelihood based measure of model fit that penalizes for model complexity (Akaike 1974). Lower AIC values are indicative of better-fitting models; here we present the reduction in AIC relative to the non-genetic base model (Equation 1).
In addition to rAOS, we incorporated several other variables to examine the specificity of smoking onset age and to examine whether changes in model fits might be due to modeling artifacts. These included permuted rAOS, to simulate the null hypothesis with respect to the interaction – that is, individual rAOS values were randomly re-assigned to other individuals in the data set. This is labeled “pAOS.” Other moderators, that are correlated with rAOS but which were not hypothesized a priori to have moderating effects were age at first cigarette (rAFC), age at first drink (rAFD) and age at onset of drinking (rAOD). Strong results for these alternative moderators might indicate that a more general behavioral factor is involved in the interaction, as opposed to tobacco use.
As a simple metric of whether or not the incorporation of SNP x rAOS or other interaction terms into the genetic models would result in improved model fits, we talliled the number of SNPs that met traditional significance thresholds (<0.05, <0.01 and <0.001) in the main effect and interaction models The p-values were those derived from the main effect model (equation 2) and the joint p-value for the main effect and interaction terms from the interaction model (Equation 3). Overall false-discovery rates (π0 parameters) were also computed for each set of analyses using the program QVALUE (Storey 2002). The π0 parameter is an estimate of the total proportion of false-positives in the complete distribution of p-values if the null hypothesis is rejected for all tests.
The logistic regression concordance rate is a measure of concordance between predicted and observed risk between all possible case-control pairs. Pairs are considered to be concordant if the predicted risk for the case is higher than the predicted risk for the control, and discordant if the opposite is true. We employed “leave-one-out” cross-validation as implemented in SAS Proc Logistic (SASInstituteInc 2002–03) to compute predicted risk scores and concordance rates. In this procedure, the risk for each observation was predicted from a model parameterized on all other observations. This approach reduces inflated concordance rates that result from over-fitting.
In order to evaluate whether or not modeling of rAOS interactions would improve the predictive utility of multiple SNPs, we constructed multi-locus genetic models to predict case-control status from either : 1.) multiple SNPs; or 2.) multiple SNPs paired with their respective SNP x rAOS interaction terms. For the former model, the top 30 SNPs were entered into the model, and a backward elimination procedure was applied until all remaining SNPs in the model were independently associated with p<0.15 (a commonly used threshold for multiple regression model building). A similar procedure was done for the model that included SNP x rAOS interaction except in this case, interactions were only entered into the initial model if they were significant in single-locus analyses, and SNPs were eliminated until all remaining SNPs exhibited either a main effect or interaction with p<0.15.
The analysis sample contains 797 cases and 811 controls of European descent. The average age at interview for nicotine dependent cases was 35.7 years (SD=5.5), the average age for controls was 36.9 years (SD=5.3); this was a small, but significant difference (p<0.001). There were significantly more women among controls than cases (70% vs. 56%, p<0.001). As such, age and sex were included as covariates in all analyses. To test for the possibility of population sub-structure not accounted for by self-reported race/ethnicity, STRUCTURE analyses (Pritchard et al. 2000) were conducted; no evidence of confounding substructure was detected (Bierut et al. 2007). This was subsequently confirmed using principal components analysis in Eigenstrat (Li and Yu 2008). Mean AOS was 15.4 years for cases (SD=3.6) and 18.0 years for controls (SD=3.6); a highly significant difference (ANOVA t=13.55, p<0.001).
For all 3369 SNPs, we used logistic regression to predict case-status using the main effect model (Equation 2) and rAOS interaction model (Equation 3), and tallied the number of nominally significant loci using three traditional thresholds (p<0.05, p<0.01 and p<0.001) – results of these counts are summarized in Table 1. This count is intended as a simple metric of the relative utility of each model, and does not reflect adjustment for multiple testing or linkage disequilibrium. In interaction models, the p-values reflect joint effects of SNP and SNP x rAOS interactions (2 degree of freedom test), whereas for the main-effect models, tallies reflect 1-degree of freedom tests for the effect of SNP only. For each threshold, substantially more loci met the p-value criterion under the interaction model than under the main-effect model, with 40–60% more “hits” in each case (e.g., 448 SNPs with p<0.05 in the interaction model, vs. 268 in the main-effect model). The increased number of low p-values may stem in part from overfitting, resulting from the additional degree of freedom to the model. However, the addition of alternative interaction terms to the model (Equation 3), including randomly reassigned rAOS (permuted AOS, designated pAOS; i.e., SNP x pAOS), age-at-first cigarette (SNP x rAFC), age-at-onset of drinking (SNP x rAOD) and age-at-first drink (SNP x rAFD) yielded fewer nominally significant p-values than the SNP x rAOS model and, in many cases, the main-effect model. (The age-at-onset variables rAFC, rAOD, and rAFC correlate with rAOS with R=0.67, R=0.26 and R=0.18, respectively). False-discovery rate analysis π0 parameters, derived from each of the complete p-value distributions, also suggest that incorporation of the SNP x rAOS interaction yields more overall findings than the main effect or alternative interaction models (Table 1).
Table 2 summarizes the results from the 30 lowest p-values from the SNP main-effect model (Equation 2). These findings correspond to p-values of 0.0021 and lower. An R2-binning procedure (Carlson et al. 2004) was used to identify bins of highly correlated SNPs (all R2 of 0.80 or higher), and only the top results from each bin are listed. Results are similar to those presented in Saccone et al (Saccone et al. 2007), but some discrepancies were expected because of differences in model specification, analysis sample composition, and SNP inclusion criteria. They are presented here for comparison with the interaction model results, and will not be discussed in detail except in that context.
Results from the SNP x rAOS interaction model are summarized in Table 3; the top 30 findings correspond to p-values less than or equal to 0.001. These p-values correspond to the 2-degree of freedom for the joint-significance of main effect and interaction terms. Genes and chromosomal positions are also listed for these top findings, and p-values from the main-effect model (Equation 2) for these loci are listed. A number of the genes enumerated in Table 3 correspond to findings that were identified in the initial candidate gene analysis of these data (Saccone et al. 2007); these include the nicotinic receptor gene cluster on chromosome 15 (CHRNA5-A3-B4) and its flanking regions (e.g., IREB2), DBH on chromosome 9, and KCNJ6 on chromosome 21. However, a number of new findings emerge. The most prominent new finding involves GRIN2B, for which two SNPs with rather low inter-correlation (R2=0.06) yielded p-values less than 10−4 in the interaction model. Other examples of genes containing SNPs that were highly significant in the interaction model, even though they did not reach p<0.05 in the main effect model are DBI, ADCY8, VAPA and GABRR1. Two genes, DBH and KCNB1 were among the top findings in the interaction model, despite the fact that their p-values in the main-effect model were actually lower, suggesting that these SNPs exhibit robust main effects across all values of rAOS. Finally, several SNPs exhibited relatively low p-values in the main effect model but yielded improved fits under the interaction model, suggesting statistically significant (or near-significant) interactions; these included SNPs from CHRNA5, MAP3K4, KCNJ6.
As a measure of the overall predictive power, the logistic regression concordance rate (C) was computed for each of the top findings in Tables 2 and and33 – the reported ΔC parameter is the difference in concordance rates between the genetic model (equations 2 or 3 for Tables 2 and and3,3, respectively) and the base model (equation 1). A “leave-one-out” cross-validation approach was employed to reduce overfitting (i.e, risk for each observation is predicted from model parameterized on all other observations). For the main effect models, ΔC values ranged from 0.18 to 0.50 with a median of 0.28%, whereas for the interaction model ΔC values ranged from 0.18 to 0.72, with a median value of 0.40%.
Table 4 provides parameter estimates for the top findings derived from the interaction model. In addition to presenting parameter estimates from regression models that treat AOS as a continuous variable, we also present results of post-hoc analyses in which the sample was divided based on the median value for AOS (age 16). This latter set of analyses was largely conducted for illustrative purposes; i.e., to facilitate examination of differences in odds ratios between earlier and later onset smokers. The sample split with AOS of 16 or lower comprised 562 cases and 280 controls while the split with AOS of 17 or higher comprised 219 cases and 500 controls. In addition to being the sample median, this split-value corresponds to the dichotimization threshold utilized in a previous analysis of SNP x AOS interactions (Weiss et al. 2008). Main effect odds ratios were computed separately for the two sample halves. This approach, unlike the primary analyses, makes no assumptions about the scale of the putative interaction variable (i.e., does not assume that genetic effect sizes scale multiplicatively with AOS). Parameter estimates indicate that some interactions are synergistic; i.e., the main effect estimate at the sample median for environmental risk (onset of smoking at age 16) is in the same direction as the interaction odds ratio. In other words, SNP effects are stronger in earlier onset smokers. This was the case for GRIN2B (rs890), MAP3K4 and AGPAT4 (which is in LD with MAP3K4). In the cases of CHRNA5, KCNJ6 and IREB2 (in the flanking region of CHRNA5), the main effect and interaction odds ratios were in opposite directions (SNP effects stronger in later-onset smokers). Other cases of significant interactions were ambiguous, i.e., the main effect odds ratio was not significant at the median value for AOS; including GRIN2B (rs17760877), DBI, ADCY8, VAPA and GABRR1.
The top SNP findings summarized in Tables 2 and and33 were used in a multi-locus model to predict nicotine dependence. The initial model included main effects for all 30 top SNPs. A backwards elimination procedure was applied to eliminate all SNPs that were not significant in the multiple regression at p<0.15. This was done to eliminate loci that are moderately correlated with nearby loci due to linkage disequilibrium. Thirteen loci remained in the model after the elimination procedure. For the interaction model, both main effect and interaction terms from the top 30 SNPs were initially entered, except in cases where interaction terms were not significant in the single-locus model (DBH and KCNB1). After eliminating redundant SNPs through the backwards elimination procedure, 13 SNPs remained. Predicted logistic regression risk scores and concordance rates, employing the leave-one-out cross-validation procedure were computed for each of these multi-locus models, compared with the non-genetic base model (Equation 1). Results of these parameters for both models, as well as covariates-only (non-genetic) model are presented in Table 5. Incorporation of the top genetic main effects into the model containing age, sex and the rAOS improves cross-validated concordance from 73.7% to 76.6%, and changes the predicted risk difference between cases and controls by 0.06 standard-deviations, relative to the non-genetic model (Table 5, second column, difference between rows 3 and 2). However, incorporation of top findings from the interaction-model, improves the cross-validated concordance value to 78.1%; and changes the predicted risk score difference between cases and controls by 0.09 SD, relative to the non-genetic base model. Hence, the incorporation of rAOS interactions into the genetic analyses results in a 50% relative increase in the predicted risk due genetic variables.
In this work, we demonstrate that incorporation of age-at-onset of smoking into a model that allows for moderation of genetic effects in predicting nicotine dependence results in improved SNP discovery using a variety of metrics, compared with a model that did not allow for such moderation. This effect seems to be specific to age-at-onset of regular smoking, which was selected based on the results of experimental studies in animal models (Trauth et al. 2000; Trauth et al. 2001; Rodd-Henricks et al. 2002; Rodd-Henricks et al. 2002; Slotkin 2002; Adriani et al. 2004; Slotkin et al. 2008; Slotkin et al. 2009); other age-at-onset variables did not exhibit similar effects, despite being correlated with rAOS. Finally, in a multi-locus prediction, the interaction model exhibited greater predictive power than did the main-effect model with a comparable number of loci.
Like other complex diseases with moderate heritability, substance use disorders are likely to be influenced by numerous common polymorphisms, each with only a small to moderate influence on the overall liability to the disorder (Orr et al. 2008). However, it is clear that there are also numerous environmental contributions to liability. Distinct genetic pathways may influence biological responses to specific environmental risk factors. Put simply, any given polymorphism may be important in some risk environments but not in others. It logically follows that specification of key environmental variables may aid in identifying genes that are otherwise masked under heterogeneity of the risk environment.
While concerns about multiple testing apply in this exploratory work, there are several discoveries that warrant further discussion. The first and third top findings in this paper involved two weakly correlated SNPs from the GRIN2B gene; these were not detected under the main-effect model in the initial candidate gene study. This gene encodes subunits of the N-methyl-D-aspartate (NMDA) receptor channel, which appears be involved in mediating neural-plasticity leading to drug dependence (Kelley 2004; Hyman et al. 2006). Notably, the expression and subunit composition of NMDA receptors change markedly over the course of development, and NMDA-mediated neuroplasticity may be heightened in adolescence (Carpenter-Hyland and Chandler 2007; Lau and Zukin 2007). GRIN2B SNPs have been implicated in correlated phenotypes such as ADHD and alcohol-related traits (Wernicke et al. 2003; Dorval et al. 2007) and one very recent report provides strong evidence for an association between SNPs in this gene and smoking initiation (Vink et al. 2009). The results here suggest that GRIN2B polymorphisms may be important, age-dependent factors in the development of nicotine dependence. Based on both compelling biological evidence and association studies, the SNPs identified here and correlated polymorphisms warrant more detailed studies in this and other samples.
Another noteworthy finding is that the effect of the non-synonymous coding SNP rs16969968 in CHRNA5 appears to be more important in later onset smokers. This contrasts with the findings of Weiss and colleagues, who suggested that haplotypes involving this polymorphism were more important among early onset smokers (Weiss et al. 2008). There are a number of differences in phenotype definitions, sampling approaches, and comorbidity between our sample and that of Weiss and colleagues; however there is no obvious reason for the clear discrepancy between findings. It is noteworthy that a significant interaction appeared in both samples, but further examination in larger samples and across populations will be required to discern the functional significance of this interaction, if any. The confidence intervals for this interaction parameter are close to 1.0 (0.89, 0.99), whereas interaction parameter estimates for some other SNPs were more robust (e.g., GRIN2B and ADCY8). Differences in sampling, secular trends in environmental contributions to nicotine dependence, as well as random fluctuation could all contribute to discrepancies in estimates of effect sizes.
In summary, the incorporation of AOS into genetic models of nicotine dependence seems to be a promising avenue to discover novel SNPs that may be important in the context of different environments. The age at which a person begins smoking is a complex phenomenon that is correlated with a number of behaviors, and so it would be naive to treat AOS as a simple environmental exposure in interpreting these findings. Nonetheless, it seems reasonable to pursue some of the present findings in replication attempts, and to extend this approach to genome-wide association studies.
This work was supported by NIH R21 DA026612 (RAG), R01DA25888, K02DA021237, P01 CA089392 (LJB), and DA02691 (NLS).