|Home | About | Journals | Submit | Contact Us | Français|
Results from genome-wide association studies of complex traits account for only a modest proportion of the trait variance predicted to be due to genetics. We hypothesize that joint analysis of polymorphisms may account for more variance. We evaluated this hypothesis on a case–control smoking phenotype by examining pairs of nicotinic receptor single-nucleotide polymorphisms (SNPs) using the Restricted Partition Method (RPM) on data from the Collaborative Genetic Study of Nicotine Dependence (COGEND). We found evidence of joint effects that increase explained variance. Four signals identified in COGEND were testable in independent American Cancer Society (ACS) data, and three of the four signals replicated. Our results highlight two important lessons: joint effects that increase the explained variance are not limited to loci displaying substantial main effects, and joint effects need not display a significant interaction term in a logistic regression model. These results suggest that the joint analyses of variants may indeed account for part of the genetic variance left unexplained by single SNP analyses. Methodologies that limit analyses of joint effects to variants that demonstrate association in single SNP analyses, or require a significant interaction term, will likely miss important joint effects.
A key challenge for genetic analysis today is to account for the bulk of the phenotypic variance in complex traits attributable to genetic factors. Traditional genetic analysis methods exploit the possibility that variation in a single genetic locus may result in detectable effects on a phenotype. These methods typically examine genotypes one at a time, and build additive (or log-additive) effects models from them. The approach has been successful, with genome-wide association studies (GWAS) alone identifying over 500 genetic variants contributing to disease (Hindorff et al. 2009). However, for many complex traits (e.g. obesity, smoking, diabetes), the variants identified by studies with large samples and dense genome-wide geno-typing account for only a modest fraction of the phenotypic variance estimated to be attributable to genetic contributions (Goldstein 2009; Hirschhorn 2009; Kraft and Hunter 2009). A variety of factors have been suggested to account for the “missing variance”, including rare variants, variants not surveyed by current GWAS chips, structural variants (e.g. copy number variants such as insertion/deletions or copy neutral variation such as inversions and translocations), population heterogeneity, gene–gene interactions, and gene-environment interactions (Galvan et al. 2010; Manolio et al. 2009).
While all of these likely have some role in the “missing” genetic contribution to complex diseases, examining datasets for joint effects (gene–gene or gene–environment) may prove essential in order to understand genetic associations to complex phenotypes, in spite of the computational and analytical challenges. Biology provides well-documented examples of epistasis playing an important role in phenotypes ranging from gross morphology to longevity to efficiency of reproduction (Anholt et al. 2003; Gerke et al. 2009; Mackay 2010; Vieira et al. 2000; Wolf et al. 2005). Further, many important traits of medical interest (such as heart disease, hypertension, diabetes, cancer, and infection) arise from biological systems controlled by interacting genetic factors (Churchill et al. 2004; Lander and Schork 1994; Phillips 2008; Routman and Cheverud 1995; Schork 1997; Szathmary et al. 2001).
There are a variety of definitions for epistasis, but a common hallmark is that the effect of one genetic locus differs depending on the genetic background. In such cases, modeling genetic loci jointly may explain more of the trait variance than a model that includes only main effects. In the most extreme case, it may turn out that the importance of a genetic variant may be revealed as significant only in a joint analysis (Culverhouse et al. 2002).
In this study, we have taken a systematic approach to examine joint effects of variants marking the cholinergic nicotinic receptor subunit genes on smoking behavior. Nicotine dependence is a heritable behavior, with estimates of approximately 50% for the genetic contribution to phenotypic variance (Li et al. 2003; Sullivan and Kendler 1999). The α5α3β4 cholinergic nicotinic receptor cluster on chromosome 15 is now unequivocally associated with nicotine dependence and the number of cigarettes smoked per day (Amos et al. 2008; Berrettini et al. 2008; Bierut et al. 2007; Bierut et al. 2008; Liu et al. 2010; Saccone et al. 2009b; Saccone et al. 2007; Stevens et al. 2008; Thorgeirsson et al. 2010; Tobacco and Genetics Consortium 2010; Weiss et al. 2008). For example, a recent report of a meta-analysis with 76,972 subjects (Thorgeirsson et al. 2010) achieves a p value 2.4 × 10–69 for a SNP in this cluster. Other cholinergic nicotinic receptors play a more modest role (Saccone et al. 2010b; Schlaepfer et al. 2008; Thorgeirsson et al. 2010).
Our goals in this study are to investigate the following questions:
Individuals from the Collaborative Genetic Study of Nicotine Dependence (COGEND) used in these analyses were of European descent (N = 2,062). Recruitment of subjects started with community-based telephone-screening of approximately 50,000 people aged 25–44. The screening identified individuals who had smoked at least 100 cigarettes in their lifetime (smokers). The smokers then answered further questions about their smoking behavior, including the Fagerström Test for Nicotine Dependence (FTND) (Heatherton et al. 1991). This assessment identified nicotine dependent subjects who were current smokers with a score of 4 or higher on the FTND (cases). Non-dependent control subjects smoked at least 100 cigarettes, but were required to have a lifetime FTND score of 0 or 1 even during the heaviest period of smoking. Nicotine dependent cases and non-dependent controls who smoked were invited to participate in this genetic study of smoking. COGEND recruited subjects from three urban areas in the USA: St. Louis (N = 1,434), Detroit (N = 581), and Minneapolis (N = 47). The COGEND study was carried out with approval from the appropriate institutional review boards and with informed consent from all participants.
The SNPs examined in this study represent variants in 16 cholinergic nicotinic receptor genes, with comprehensive coverage of the α5α3β4 cluster on chromosome 15. The COGEND data contain genotypes for 210 SNPs in cholinergic nicotinic receptor genes. Before analysis, genotypic data were thinned using a pair-wise threshold of r2 < 0.95 based on linkage disequilibrium in COGEND to reduce the number of redundant pair-wise tests. This resulted in an analysis dataset of 127 SNPs. The SNPs were not thinned further because a goal of these analyses was to be sensitive to epistasis or other multi-locus effects that might display little to no marginal effects, and whose effects might not be adequately captured using tags with lower correlations.
All genotype data have undergone extensive quality control measures (Bierut et al. 2007; Saccone et al. 2009a). Self-reported race was verified with an EIGENSTRAT (Price et al. 2006) analysis including HapMap data. EIGENSTRAT analysis also found no significant population differences between sites. Map positions and annotations were obtained from the Genome Reference Consortium Homo sapiens assembly GRCh37.
A subset of individuals from the American Cancer Society (ACS) Cancer Prevention Study-II Nutrition cohort (N ~ 184,000) was used as a replication dataset. DNA was collected from 109,380 individuals from this cohort. Smoking behavior information was obtained from questionnaires administered in 1982, 1992, and 1997. Based on the responses to these questionnaires, a case/control phenotype (heavy smokers/light smokers) was developed as a proxy for nicotine dependence and non-dependence (Stevens et al. 2008). Inclusion in the ACS dataset used for our analysis required subjects to report smoking at least 100 cigarettes lifetime. Cases (heavy smokers) also needed to report in at least one interview smoking at least 30 cigarettes per day for at least 5 consecutive years. Controls (light smokers), in addition to smoking more than 100 cigarettes lifetime, smoked for at least 1 year, but endorsed the lightest smoking rate available on each of the three questionnaires (<5 cigarettes/day in 1982 and 1992, <10 cigarettes/day in 1997). After ascertainment for heavy and light smoking, 1,500 heavy smokers and 1,500 light smokers were selected for genotyping. Of these, 2,847 smokers (1,452 heavy smokers and 1,395 light smokers) passed genotyping quality control and were included in this study. Both current and former smokers were included in the analysis.
SNPs genotyped in the ACS subjects included those in the neuronal nicotinic acetylcholine receptor genes with p < 0.05 in the initial univariate analysis of the COGEND study of nicotine dependence (Bierut et al. 2007) and additional haplotype-tagging SNPs chosen using the method of Gabriel et al. (2002). After filtering for genotype quality control, the ACS data used in our analyses contain a total of 61 cholinergic nicotinic receptor subunit SNPs (a subset of the 210 COGEND cholinergic nicotinic receptor subunit SNPs). Detailed descriptions of the ACS sample ascertainment and genotyping quality control are described in Stevens et al. (2008).
Summary information about age and sex in the two datasets is given in Table 1.
A variety of methods can be used to investigate how multiple genetic variants contribute jointly to a dichotomous trait. We chose to use the restricted partition method (RPM), a statistical approach designed to identify combinations of qualitative genetic and environmental factors (e.g. genotypes, categorical or dichotomized environmental exposures) contributing to a quantitative or binary trait (Culverhouse et al. 2007; Culverhouse et al. 2004) and logistic regression modeling, a more traditional statistical approach. The RPM is agnostic regarding a specific genetic model (e.g. additive, dominant, or recessive in the case of a single SNP) and was specifically designed to be sensitive even if the contribution from a combination of factors is predominantly presented as an interaction displaying little or no marginal effects. It is an exploratory method that uses the data to determine the number of distinct risk classes, with the aim of determining if modeling predictors jointly accounts for more of the variation in phenotype than summing their individual effects. It does not specifically identify interactions or test them for significance.
Intuitively, the RPM is a multi-locus ‘measured genotype’ (Boerwinkle et al. 1986) approach, assessing the mean trait values or prevalence information for different multi-locus genotypes. Genotypes (groups) are iteratively merged if a multiple comparisons test indicates that they are not significantly different. The algorithm halts when either all genotypes are merged (indicating no significant association between these genotypes and the phenotype) or when the remaining groups are significantly different from each other. In this case, the resulting model is evaluated for the proportion of trait variance explained by the groups. Significance of the full model (not just the interaction) is evaluated using permutation tests because the goal is to identify contributors to phenotypic variation, not to identify or test interactions. The benefits of focusing on the full model, rather than on specific interactions, have been demonstrated by extensive evaluation in a variety of simulated models (Marchini et al. 2005). We note that if the RPM model explains none of the variance (i.e. all the genotypes are merged into a single group), then the p value is necessarily = 1, as no permuted dataset could result in a lower explained variance. A detailed description of the RPM can be found in Culverhouse et al. (2004).
Applying the RPM to the dichotomous nicotine dependence phenotype among smokers in COGEND allowed the development of hypotheses that joint effects of specific genetic loci (tagged by SNPs genotyped in COGEND) increased the variance explained. To minimize the chance of reporting the same signal multiple times, we report two pair-wise results as independent only if at least one SNP from the first pair is not in linkage disequilibrium (r2 < 0.4) with either of the SNPs in the second pair.
Univariate and pair-wise RPM analyses were performed for the 127 nicotinic receptor SNPs that passed quality control for the combined COGEND data. Because we were analyzing a binary trait, the explained variance, VE, was defined by where N is the total sample size, p is the proportion of cases, and q is the proportion of controls, with the corresponding subscripted variables defined similarly for the resulting RPM model groups. We note that this is always less than or equal to the variance accounted for by the full model (i.e. all cells kept separate). Thus, the merging of cells by the RPM method provides some protection from over-fitting and the corresponding inflation of estimated explained variance. Sampling variability and sparseness can be expected to result in “noise” in the estimates of explained phenotypic variance. A test of the RPM using real genotypes with simulated phenotypes (Culverhouse et al. 2009) suggests that pair-wise results accounting for less than 0.5% of the trait variance may have an increased risk of being false positives. Because of this, we restricted our 2-SNP signals of interest to those explaining more than 1.0% of the trait variance.
We took two approaches to further examine the joint effects identified by the RPM in the COGEND data—testing using another analytic method, logistic regression, and replication in an independent dataset.
All the top SNP-pairs identified by the RPM were evaluated using logistic regression to determine if the signals could be identified by this alternative analytic method. The SNPs were treated as class variables. Results from univariate models and joint models with and without an interaction term were compared. Logistic regression analyses were performed using SAS® software version 9.1 (SAS Institute Inc., Cary, NC).
Replication using independent data is the gold standard. For this, we used the ACS data that had the heavy-smoking/light-smoking phenotype as proxy for the FTND derived nicotine dependence/non-dependence phenotype from COGEND. Our criteria for deciding that an RPM result is recapitulated in the independent dataset involve two factors. First, the joint RPM model should account for substantially more of the trait variance than the sum of the univariate RPM models in the independent dataset (we used an arbitrary threshold of 15% increase or greater). The second condition we require for replication is that the RPM model produced in the replication data should be consistent with the RPM model in the initial dataset. By this we mean that, in general, genotypes that are of high risk in the initial model should be high risk in the replication, and the same for low risk genotypes. We note that although this second replication condition is heuristic, (1) it is a higher dimensional analog for the routine procedure of checking univariate results to ensure that allelic effect is in the same direction in the initial and replication datasets, and (2) it provides an additional way to reject replication and cannot increase the number of false positives. Because only a limited number of the COGEND SNPs were genotyped in the ACS data, this testing for replication was only available for a few signals identified in the COGEND data. For one of the COGEND SNPs not genotyped in the ACS data, rs3743075, linkage disequilibrium information from both COGEND and HapMap suggests that rs514743 is an appropriate surrogate (r2 = 0.97 in COGEND data, r2 = 0.92 in the HapMap release 22 CEU data). The use of rs514743 as a proxy for rs3743075 allowed us to evaluate one additional signal from the COGEND data in the ACS data.
A complete list of the 127 SNPs used in this study, along with the results of the univariate RPM analyses on the hypothesis generating data (COGEND) can be found in Table S1 (supplementary information). The phenotype was case–control status for nicotine dependence. The table lists the nicotine receptor gene cluster associated with each SNP, the chromosome, the base-pair location, the proportion of variance attributable to the SNP, and an empirical p value for association. The permutation-based empirical p values resulted from a maximum of 100,000 permutations of the phenotype for each SNP. The predictors were the SNPs treated as 3-level categorical variables. The results are consistent with the primary analyses (univariate) from the first-stage analysis of the COGEND data using logistic regression methods (Bierut et al. 2007; Saccone et al. 2009a).
There were a total of 8,001 SNP-pairs analyzed in the hypothesis generating dataset (COGEND). We present our top results for each of two rankings: the first based on total explained variance, the second based on increase in explained variance by the joint analysis of the pair compared to the sum of the univariate effects. Table 2 reports the top six results in terms of total variance explained. For each result, we list the chromosomes, gene clusters, and rs numbers for the two SNPs representing the signal. This information is followed by the percentage of the phenotypic variance accounted for by each of the SNPs under univariate RPM analyses and the variance explained under RPM analyses combining both SNPs in the model. All of the pair-wise RPM models in this table had empirical p values of ≤10–6 (based on results from 1,000,000 permutations each; 10 times as many as were used for the univariate estimates).
Each of the joint effect signals in Table 2 involves the SNP rs16969968. This SNP provides the strongest signal in the univariate RPM analysis of the COGEND data, accounting for 1.22% of the trait variance. Because selection focused on maximizing explained variance, it is not surprising to find rs16969968 in each of the signals. We note that RPM models for two of the secondary SNPs accounted for none of the trait variance in the single SNP analyses; however, when jointly analyzed with rs16969968, additional trait variance is explained.
The RPM results are followed in Table 2 by the results from a logistic regression analysis. The first column under “Logistic Regression results” contains the p values associated with the 2 degree of freedom (df) univariate model for each of the two SNPs. The second column contains the p value for the 4 df interaction term in the model containing both SNPs and the interaction variable. Univariate RPM and logistic regression results are consistent (cf supplementary Table S1): the variant rs16969968, which explains the most variance in the univariate RPM analysis, is the most statistically significant SNP in the logistic regression analysis; and the SNPs that explain none of the variance in the RPM analysis all have p values >0.2 in the univariate logistic regression analysis. The first reported signal, involving rs16969968 and rs2133965 (which account for nearly 90% more of the variance when analyzed jointly), shows a significant interaction p value (p = 0.002) using logistic regression. None of the other logistic regression interaction terms in this table were significant.
Four of the signals listed in Table 2 were testable in the ACS data, and details of the replication analysis are summarized in Table 3. The format is similar to that of Table 2. Our criteria for replication are that the RPM model accounts for at least a 15% increase in explained variance and that the RPM models produced in the two datasets should be consistent. Three of the four signals replicate. All of these signals displayed >20% increase in explained variance, well above our threshold for replication. For three of the four, the RPM models in the ACS data were similar to those found for the COGEND data, meeting our criteria for replication. The fourth RPM model was not sufficiently similar to the one proposed in the COGEND data for us to consider it to be a replication.
The variant rs1696998, though still the most significant SNP in the ACS data, only accounts for 0.87% of the trait variance under a univariate RPM analysis. All of the secondary SNPs account for none of the variance under univariate RPM analysis of these data. As in the COGEND data, the univariate p values are consistent between the RPM and logistic regression analyses, and the interaction terms were not significant under the logistic regression analysis (although the final pair, rs16969968 and rs514743, displayed a nominal p value of 0.05 for the interaction term).
Our second ranking of signals, based on increased explained variance under an RPM analysis, was aimed at identifying pairs with joint effects much larger than the sum of the individual effects. We note that this is a filtered ranking because of our requirement that any reported pair-wise signal explain at least 1% of the phenotypic variance. The RPM and logistic regression results for this set of pairs are summarized in Table 4. Once again, all of the pair-wise RPM models had empirical p values of no more than 10–6 (based on results from 1,000,000 permutations each). Because this ranking was designed to find pairs with large synergistic effects, it is perhaps not surprising that the logistic regression analysis found the interaction terms for each of these signals to be nominally significant, with uncorrected p values ranging from 0.04 to 2.2 × 10–5. None of these pairs were testable in the independent ACS data because in each case, at least one of the SNPs was not genotyped in the ACS data and no good surrogate could be identified.
To evaluate how well we succeeded in increasing the explained variance in phenotype, we evaluated two logistic regression models for nicotine dependence in the COGEND dataset: one with interactions and one without. Predictors included the 11 SNPs from the 3 two-locus models that were validated in the ACS data and the 4 two-locus models that had logistic regression interaction term p values <0.01: rs16969968, rs2133965, rs3787138, rs13277524, rs3743075 from Table 2 and rs2292977, rs17483548, rs667282, rs680244, rs1500948, rs2611603 from Table 4. The logistic regression model without interactions accounted for 5.15% of the trait variance and resulted in a c statistic of 0.622. In contrast, the model including the same predictors, but adding the 7 pair-wise interaction terms listed in the tables, accounted for 9.49% of the trait variance and resulted in a c statistic of 0.662.
Our goal was to determine if examining genetic variants jointly could account for substantially more of the variance in phenotype than would be estimated by summing the results of univariate analyses. We identified several variants that jointly explain more phenotypic variance than they do individually, and three of the four tested signals replicate in independent data. Our results highlight two important lessons: (1) joint effects need to be investigated across all SNPs, not just those with main effects, and (2) it is not sufficient to look at the significance of an interaction term in a logistic regression model to determine if there is a joint effect.
The prime illustration for both of these points is the joint effect of the loci tagged by rs16969968 and rs3743075: as illustrated in Table 2, univariate analysis of the COGEND data indicates that the genetic locus tagged by rs16969968 displays a highly significant association to nicotine dependence, and the second locus, tagged by rs3743075, does not display evidence of association to nicotine dependence under univariate analyses. However, when the two genetic loci are analyzed jointly using the RPM, this pair accounts for much more of the variance in the dichotomous nicotine dependence/non-dependence pheno-type than the sum of the univariate effects. This same pattern of results is seen in the ACS data: univariate analysis finds that the locus tagged by rs16969968 displays a substantial univariate effect on the dichotomous heavy versus light smoking phenotype, and the second locus is not associated with the phenotype in a univariate analysis. Again, a joint RPM analysis of the two SNPs accounts for considerably more of the phenotype variance than the sum of the univariate results. Although analysis by logistic regression also displayed the increase in explained variance (results not shown), the interaction term was not significant. As a result, we would have missed this joint effect had we required a significant interaction coefficient.
The case for these two loci having a synergistic joint effect has been greatly strengthened by results very recently reported by the Consortium for the Genetic Analysis of Smoking Phenotypes (CGASP) (Saccone et al. 2010a). The CGASP subjects consist of current and former smokers of European ancestry (N = 38,617), including COGEND and ACS subjects. A case/control phenotype of heavy versus light smoking was analyzed. In univariate analysis, the locus tagged by rs16969968 was highly associated with the heavy smoking phenotype (p = 5.96 × 10–31), while the second locus, tagged by rs588765, was considerably less significant (p = 4.54 × 10–4). When the two loci were analyzed jointly, each became more significant: p = 3.52 × 10–36 for the locus tagged by rs16969968, and p = 6.03 × 10–9 (passing the threshold for genome-wide significance) for the locus tagged by rs588765. The logistic regression interaction term in this analysis was not significant. The explanation for these results is that the risk alleles for the two loci are negatively correlated. As a result, in a univariate analysis, the association of rs16969968 to nicotine dependence is present, but dampened, while the effect of the second locus is almost completely masked until jointly analyzed.
Although we cannot know how often this kind of joint effect might occur, we find this result, displayed in COGEND data, ACS data, and the larger CGASP data, a strong argument that tests for interactions not be limited to loci with substantial main effects, and that the consideration of joint effects should not be limited to ones giving rise to a statistically significant interaction term. It is an advantage of the RPM, focused on identifying sources of trait variance and not specifically interactions, that it identifies such pairs in an automated fashion, without the need to examine each model individually (an impractical approach when evaluating many thousands or millions of multi-locus models).
Although a replicable joint effect does not require a significant logistic regression interaction term, it appears that if the proportional increase in explained variance is great enough (e.g. over 80%), logistic regression can detect a significant interaction term. The first SNP pair in Table 2, rs16969968 and rs2133965, with nearly 90% increase in explained variance, was the only pair from this table for which the logistic regression interaction term was associated with a p value <0.05. In contrast, all five pairs from Table 4, which focused on large proportional increases in explained variance, had interaction p values <0.05.
The variants rs16969968, a non-synonymous coding SNP in CHRNA5, and rs1051730, a synonymous SNP, are of particular interest because these SNPs are strongly associated with nicotine dependence and smoking behaviors (Amos et al. 2008; Berrettini et al. 2008; Liu et al. 2010; Thorgeirsson et al. 2008; Thorgeirsson et al. 2010; Tobacco and Genetics Consortium 2010). These two SNPs are highly correlated (r2 = 0.991) in our data, and we included rs16969968 in our list of 127 to be tested. We selected this SNP because it results in an amino acid change, and in vitro studies demonstrate that it alters receptor function (Bierut et al. 2008). Because SNPs were thinned to include only those with r2 < 0.95, rs1051730 was not included in our primary analyses. We confirmed, with secondary univariate and joint analyses using rs1051730, that the results were very similar to those for rs16969968. (An analog of Table 2, replacing rs16969968 throughout with rs1061730, can be found in Supplemental Table S2.) We cannot say which, if either, of these two SNPs is biologically linked to smoking behavior. We can only say that they are representatives of a cluster of variants in tight linkage disequilibrium that display these effects. Further research in the lab will be required to identify the causative variants definitively.
We focused on cholinergic nicotinic receptor subunit genes because the protein products physically combine to form biologically active receptors that bind nicotine. For example, the α4 and α5 nicotinic acetylcholine receptor (nAChR) subunits combine with the nAChR β2 subunit to form an α42β22α51 receptor that is expressed in various brain regions including the mesolimbic reward pathway (McClure-Begley et al. 2009; Salminen et al. 2004; Zoli et al. 2002). Thus, it is plausible that variants in and around the genes that form these subunits may alter the subunit makeup of nicotinic receptors or the relative expression of the receptors. Of course, joint effects for these genes need not be limited to changes that alter an amino acid, but can include variants that alter splicing, mRNA expression, stability, or other regulatory factors. Our analytic evidence of joint effects can suggest models that can be tested biologically in the laboratory.
We have presented evidence that examining factors jointly, including SNPs displaying little to no univariate association to the phenotype, may uncover interactions and other synergistic effects that account for a sizable portion of the “missing” genetic variance remaining after univariate analyses of well-powered GWAS (Goldstein 2009; Hirschhorn 2009; Kraft and Hunter 2009). Although this study on the genetics of nicotine dependence focused on only a small portion of the genome, we found that examining nicotinic receptor variants in a pair-wise manner increased the proportion of variation explained. The challenge with such an approach, particularly if applied to data with 1,000,000 genotyped polymorphisms, is the number of tests. Even if limited to all pair-wise models, such analyses would be computationally expensive and statistically intractable. The problem becomes more extreme if analyses involve more than two factors at a time.
Two approaches for addressing this problem have been applied in this study. One approach to address both the computational and statistical problem of the large number of tests is to filter the polymorphisms to be analyzed based on biological plausibility. In this case, we limited our analyses to polymorphisms located in or in linkage dis-equilibrium with nicotinic receptor subunit genes and filtered so that no two SNPs in our analysis had r2 > 0.95. Other filtering approaches could make use of additional biological information such as evolutionary conservation, biological pathways, and results from previous studies.
An approach to address the statistical problem of the large number of tests is to use a second, independent dataset. Although filtering may provide considerable help with the computation challenges, the number of tests will likely remain very large and the effects to be detected relatively modest. As a result, the issue of statistical significance will likely remain even after judicious a priori filtering of predictors. One way to address this is to use the initial data for hypothesis generation and an independent dataset to test a small number of hypotheses. If the key biological contributors to the phenotype are identified in the initial data, we would expect that the signal would replicate in an ethnically matched, independent replication sample.
The focus of this study on pair-wise analyses is one of its limitations. The reasons for this choice include limiting the large number of statistical tests performed, the theoretical results suggesting that factors not detectible with pair-wise analyses are unlikely to display large effects (Culverhouse et al. 2002), and the fact that our hypothesis-generating sample was of modest size for examining joint effects. We recognize that higher-order joint effects with three or more variants may play important roles in the etiology of complex disease, but larger datasets are needed to evaluate such complex multi-SNP effects. A second limitation of the study is that we were not able to test all of the identified signals in independent data.
This study was also limited to two analytic methods for examining joint genetic effects. A variety of methods are available to investigate how multiple genetic variants contribute jointly to a dichotomous trait. In addition to the RPM and the traditional logistic regression, there are other partition-based methods [e.g. the Combinatorial Partition Method (CPM) (Nelson et al. 2001), multifactor dimensionality reduction (MDR) (Hahn et al. 2003), the Generalized MDR (GMDR) (Lou et al. 2007)], as well as information theory based methods [e.g. k-way interaction information (KWII) and total correlation information (TCI) (Chanda et al. 2007)], and a variety of other approaches. Because these approaches each have their advantages for detecting joint effects, an analysis using multiple methods can identify additional signals that might have been missed by a single method, as well as provide increased confidence in signals that are identified by multiple methods.
The results of this study support three ideas which we believe are important for future studies. First, this approach identified several variants that have minimal evidence of association when tested individually, but which have larger effects when examined in combination with other genetic predictors. Second, requiring a significant interaction term in logistic regression testing may not detect variants that jointly explain more of the variance. Finally, increases in explained variance can be obtained by including only a small number of interactions.
We note that as more researchers begin specifically looking for joint effects, the reports of interactions have increased. This is particularly true for genetic studies of smoking. In addition to the CGASP result mentioned above, several other studies have reported interesting results from the joint analysis of multiple loci chr15q25 in samples of European descent (Li et al. 2010a; Liu et al. 2010; Thorgeirsson et al. 2010; Tobacco and Genetics Consortium 2010), African Americans (Li et al. 2010a), and Koreans (Li et al. 2010b). Additional joint genetic effects related to smoking have been reported in other candidate regions, including interactions among variants in GABBR1 and GABBR2 related to nicotine dependence in African and European Americans (Li et al. 2009) and an interaction between CYP2A6 and MAOA affecting smoking in Chinese (Tang et al. 2009).
As researchers struggle to make the best use of the wealth of genetic data now available, a key goal is to account for more of the phenotypic variance that is expected to be attributable to genetics. The results of this study suggest that pair-wise examination of genetic data can be a useful tool for achieving this goal, uncovering substantial additional genetic contribution to phenotypic variance through joint effects of multiple loci.
EIGENSTRAT software: http://genepath.med/harvard.edu/~reich/EIGENSTRAT.htm. R software: http://www.r-project.org/.
In memory of Theodore Reich, founding Principal Investigator of COGEND, we are indebted to his leadership in the establishment and nurturing of COGEND and acknowledge with great admiration his seminal scientific contributions to the field. Lead investigators directing data collection are Laura Bierut, Naomi Breslau, Dorothy Hatsukami, and Eric Johnson. The authors thank Heidi Kromrei and Tracey Richmond for their assistance in data collection and Brian Suarez for his thoughtful comments on the manuscript. This work was supported by National Institutes of Health grants K25 GM69590 from the National Institute of General Medical Sciences, R03 DA023166 from the National Institute on Drug Abuse (NIDA), and P01 CA89392 from the National Cancer Institute, as well as IRG-58-010-50 from the American Cancer Society. Geno-typing services for this project were provided by Perlegen Sciences under NIDA Contract HHSN271200477471C. Additional genotyping services were provided by the Center for Inherited Disease Research (CIDR). CIDR is fully funded through a federal contract from the National Institutes of Health to The Johns Hopkins University, contract number HHSN268200782096. Phenotypic and genotypic data are stored in the NIDA Center for Genetic Studies (NCGS) at http://zork.wustl.edu/ under NIDA Contract HHSN271200477451C (PIs J Tischfield and J Rice).
Electronic supplementary material The online version of this article (doi:10.1007/s00439-010-0911-7) contains supplementary material, which is available to authorized users.
Conflict of interest Drs. LJ Bierut, AM Goate, and JC Wang are listed as inventors on a patent “Markers for Addiction” (US 20070258898) covering the use of certain SNPs, including rs16969968, in determining the diagnosis, prognosis, and treatment of addiction. Dr. N Saccone is the spouse of S Saccone, who is also listed on the above patent. Dr. Bierut acted as a consultant for Pfizer, Inc. in 2008.
Robert C. Culverhouse, Division of General Medical Sciences, Department of Medicine, Washington University, Saint Louis, MO 63110, USA ; Email: rculverh/at/wustl.edu. Division of Biostatistics, Washington University, Saint Louis, MO 63110, USA.
Nancy L. Saccone, Department of Genetics, Washington University, Saint Louis, MO 63110, USA.
Jerry A. Stitzel, Department of Integrative Physiology, University of Colorado, Boulder, CO 80309, USA.
Jen C. Wang, Department of Psychiatry, Washington University, Saint Louis, MO 63110, USA.
Joseph H. Steinbach, Department of Anesthesiology Basic Science Research, Washington University, Saint Louis, MO 63110, USA.
Alison M. Goate, Department of Genetics, Washington University, Saint Louis, MO 63110, USA. Department of Psychiatry, Washington University, Saint Louis, MO 63110, USA.
Tae-Hwi Schwantes-An, Department of Genetics, Washington University, Saint Louis, MO 63110, USA.
Richard A. Grucza, Department of Psychiatry, Washington University, Saint Louis, MO 63110, USA.
Victoria L. Stevens, Department of Epidemiology and Surveillance Research, American Cancer Society, Atlanta, GA 30329, USA.
Laura J. Bierut, Department of Psychiatry, Washington University, Saint Louis, MO 63110, USA.