Abstract
Standard genomewide association studies (GWASs) rely on analyzing a single trait at a time. However, many human phenotypes are complex and composed by multiple correlated traits. Here we introduce CGWAS, a method for combining GWAS summary statistics of multiple potentially correlated traits. Extensive computer simulations demonstrated increased statistical power of CGWAS compared to the minimal pvalues of multiple singletrait GWASs (MinGWAS) and the current stateoftheart method for combining singletrait GWASs (MTAG). Applying CGWAS to a metaanalysis dataset of 78 single trait facial GWASs from 10,115 Europeans identified 56 studywide suggestively significant loci with multitrait effects on facial morphology of which 17 are novel loci. Using data from additional 13,622 European and Asian samples, 46 (82%) loci, including 9 (53%) novel loci, were replicated at nominal significance with consistent allele effects. Functional analyses further strengthen the reliability of our CGWAS findings. Our study introduces the CGWAS method and makes it available as computationally efficient opensource R package for widespread future use. Our work also provides insights into the genetic architecture of human facial appearance.
Similar content being viewed by others
Introduction
Standard genomewide association studies (GWASs) rely on analyzing a single trait at a time. However, many human phenotypes are complex and composed by multiple correlated traits. A common approach in dealing with correlated traits in GWAS is to separately conduct GWASs for the different traits and use the minimum pvalues of multiple GWASs (MinGWAS) with a more stringent significance threshold. However, because MinGWAS cannot exploit genetic information underlying correlated traits, it is of limited power in detecting genetic variants with multitrait effects.
Several methods were previously developed to enable integrated analysis of multiple traits based on singletrait GWAS summary statistics^{1,2,3}, which are more flexible than dimension reduction of individuallevel data. Such methods have shown increased power than MinGWAS in detecting genetic associations, but have limitations. The most prominent example is the current stateoftheart method for integrated GWAS named multitrait analysis of GWAS (MTAG)^{3}. MTAG distinguishes true effect and estimate error and obtains higher power than other competing methods. However, MTAG relies on the key assumption that all SNPs share the same variance–covariance matrix of effect sizes across all of the traits tested. Both, the statistical power and the typeI error rate (α) of MTAG are sensitive to the parameters related to this assumption. Moreover, accurate estimation of these parameters requires highpowered input GWASs. When the assumption is satisfied but the input GWASs do not have sufficiently high statistical power, which often is barely achievable with tens or even hundreds of thousands of samples, MTAG is expected to have limited power and unstable α. Furthermore, the key assumption of MTAG is violated as soon as different SNPs are associated with different traits of different effect sizes, which is likely the case for many if not all complex traits. When this assumption is violated, MTAG also has limited power and unstable α.
Aiming to overcome the limitations of previous methods, we developed CGWAS, a method to combine GWAS summary statistics of multiple potentially correlated traits. We implemented CGWAS as an R package that can effectively handle a large number of traits in a parallel manner by optimizing load splitting and memory management. We performed extensive computer simulations under a variety of scenarios to assess the performance of CGWAS in comparison with MinGWAS and MTAG. To empirically exemplify its practical suitability, we applied CGWAS to a previously published summary statistic dataset of 78 GWASs of correlated facial traits with mediocre statistical power (N = 10,115 Europeans)^{4} and compared it with MTAG and other competing methods. We used new and previously published data (N = 13,622 Europeans and Asians)^{5,6} for replicating the CGWAS findings and accumulated functional evidence to further strengthen the reliability of the novel genetic face loci we discovered with CGWAS.
Results
Overview of CGWAS
CGWAS represents a complete solution for investigating potential multitrait effects of SNPs based only on the summary statistics of K GWASs (Supplementary Fig. 1). For optimizing statistical power, it is necessary to partition the correlation matrix of the T statistics from K GWASs into two matrices that differentiate the ‘effect correlation’ Π caused purely by true allelic effects from the ‘background correlation’ Ψ under the null. Both Ψ and Π are K by K symmetric matrices, where Ψ_{ij}, i ∈ {1, 2,…, K} and j ∈ {1, 2,…, K} (in brief ψ), indicating the correlation between the T statistics of ith GWAS and jth GWAS caused only by nongenetic effects and similarly Π_{ij} (or π) indicating the correlation caused only by true allelic effects. In standard metaanalysis, the offdiagonal elements of Ψ would be zero. In CGWAS, however, the offdiagonal of Ψ may deviate from zero in presence of sample overlap and/or nongenetic phenotype correlations caused by shared environmental or unknown factors. CGWAS and MTAG have an important similarity in distinguishing Π from Ψ as key parameters in combining GWASs. Note that the true effect Ω and the estimate error Σ in MTAG are also K by K symmetric matrices and ideally, \({\varPi }_{{ij}}={\varOmega }_{{ij}}/\sqrt{{\varOmega }_{{ii}}{\varOmega }_{{jj}}}\) and \({\varPsi }_{{ij}}={\varSigma }_{{ij}}/\sqrt{{\varSigma }_{{ii}}{\varSigma }_{{jj}}}\). Different from MTAG that generates genomewide pvalue vectors with the same number of input GWASs, CGWAS generates a single vector of genomewide pvalues testing for each SNP if the null is deviated (H0: absence of any effect on all traits, H1: deviation from zero for at least one trait). In addition, different from MTAG that performs joint analysis of all GWASs at once, CGWAS use twostep design where separately employs an iterative decisionmaking algorithm and truncated test strategy to take into account that SNPs may have different effects in different subsets of GWASs.
Twostep design
Full details regarding formulation, derivation, and simulation can be found in the “Methods” section. In brief, CGWAS uses a twostep design to achieve high power in detecting multitrait effects under various complex scenarios with the emphasis that SNPs may have different effects in different subsets of potentially correlated GWASs. The first step is to use iterative Effectbased Inversed Covariance Weighting (iEbICoW) to combine subsets of K GWASs in which Π can be distinguished from Ψ. EbICoW is originated from traditional Inversed Variance Weighting (IVW) based metaanalysis with three extensions. For each SNP, let \({{{{{\bf{\upsigma}}}}}}={({\sigma}_{1},\, {\sigma}_{2},\ldots,\, {\sigma}_{K})}^{{{{{{\rm{T}}}}}}}\), \({{{{{\bf{n}}}}}}={({n}_{1},\, {n}_{2},\ldots,\,{n}_{K})}^{{{{{{\rm{T}}}}}}}\) and \({{{{{\bf{t}}}}}}={({t}_{1},{t}_{2},\ldots,{t}_{K})}^{{{{{{\rm{T}}}}}}}\) be the vectors of standard errors, sample sizes, and T statistics from K GWASs for one SNP, the combined T statistic of IVW is expressed as \({t}_{{meta}}={{{{{{\bf{w}}}}}}}^{{{{{{\bf{T}}}}}}}{{{{{\bf{t}}}}}}/\sqrt{{{{{{{\bf{w}}}}}}}^{{{{{{\bf{T}}}}}}}{{{{{\bf{w}}}}}}}\), where w is a vector of weights, which are usually chosen from \({{{{{\bf{w}}}}}}\,{{{{=}}}}\,{(\sqrt{{n}_{1}},\sqrt{{n}_{2}},\ldots,\sqrt{{n}_{K}})}^{{{{{{\rm{T}}}}}}}\) or \({{{{{\bf{w}}}}}}\,{{{{=}}}}\,{(1/{\sigma }_{1},1/{\sigma }_{2},\ldots,1/{\sigma }_{K})}^{{{{{{\rm{T}}}}}}}\). The first extension of EbICoW guarantees the standard normal distribution under the null for combining correlated GWASs, i.e., the presence of nonzero offdiagonal elements of Ψ. The revised combined T statistic can be derived as a function of Ψ and w and t from the original IVW.
The second extension optimizes the statistical power when elements in the vector of expected true effects \({{{{{\bf{{\updelta}}}}}}}={({\delta}_{1},{\delta}_{2},\ldots,{\delta}_{K})}^{{{{{{\rm{T}}}}}}}\) are different. Let b be a vector of 1 with length K, and Η be a K by K covariance matrix of δ among the genomewide SNPs, where \({{ H}}_{{ij}}={{{{{\rm{E}}}}}}({\delta }_{i}{\delta }_{j})\).
The third extension accounts for different effect directions of the expected δ. For k∈{1, 2,…, K},
here the function sign for a variable x is defined as sign(x) = 1 when x ≥ 0 and sign(x) = −1 when x < 0, so that for a matrix X, \({[{{{{{\rm{sign}}}}}}\left({{{{{\bf{X}}}}}}\right)]}_{{ij}}={{{{{\rm{sign}}}}}}({X}_{{ij}})\). Note that Eqs. (1), (2), and (3) are all equivalent to the original form of IVW when all GWASs are uncorrelated and all expected δ are the same (see “Methods”).
CGWAS then iteratively applies a decisionmaking algorithm to find and combine subsets of K GWASs that are suitable for EbICoW, i.e., when offdiagonal elements of Π are difficult to be distinguished those from Ψ. It iteratively combines a pair of GWASs with the max π–ψ until none of the GWAS pairs is suitable for EbICoW.
The second step is to use Truncated Wald Test (TWT) to combine the revised GWASs in a SNPspecific manner. For each SNP, TWT identifies its most significant subset and produces a single combined pvalue by applying the Wald test^{1} (Wald) to all subsets satisfying a series of preset thresholds. Different from iEbICoW, in which the T statistics asymptotically follow the standard normal distribution, TWT results obtained without calibration do not follow any known distribution due to procedures for selecting SNPspecific subset.
Parameter estimation and distribution calibration
Coming with CGWAS, three methods were additionally developed as functions for estimating inflation (getI), ψ (getPsi), and for calibration of unknown distributions (getCoef). The getI removes potential inflations of input GWASs utilizing the relationship between the distribution of the true effectexcluded GWAS and the standard normal distribution. The getPsi partitions the correlation matrix of GWASs into Ψ and Π in such a way that ψ is estimated as the correlation of two GWASs by only focusing on nonsignificant SNPs in the joint distribution of the two GWASs. The getCoef function calibrates TWT pvalues according to quantile specific coefficients between an empirically simulated null and the uniform distribution. This guarantees that final pvalues from CGWAS are directly comparable with those from any standard GWAS.
CGWAS performance via computer simulations
Extensive computer simulations were carried out to assess the performance of CGWAS in terms of statistical power (scenarios 1–5), parameter estimation (scenarios 67), and typeI error (scenarios 8–10), in comparison with MTAG and MinGWAS whenever appropriate.
Scenario 1 is designed to compare the power between two intermediate forms of EbICoW (Eqs. 1 and 2) for combining summary statistics of two GWASs (Supplementary Fig. 2). The simulations confirm that (1) is a special case of (2) when the true effects and the sample sizes are the same between the two GWASs. When they are different, the power of (2) is almost always higher than that of (1). The larger the π–ψ is, and the larger the difference of true effects, the larger the power of (2), particularly when the sample sizes are the same. These results demonstrate that (2) has a higher power than (1) in most circumstances.
Scenario 2 compares the power between an intermediate (2) and the final form of EbICoW (3) for combining summary statistics of two GWASs (Supplementary Fig. 3). The simulations showed that when π>ψ, the power of (3) is the same as (2), and when the π<ψ, the power of (3) is the same as negative signed (2), which has a higher power than (2). The power gain is increasingly considerable when the true effects become large. These results demonstrate that (3) has the same or higher power than (2).
Scenario 3 explores the power relationship between EbICoW and Wald under different π and ψ when combining summary statistics of two GWASs (the central panel of Fig. 1 and Supplementary Fig. 4). Simulations showed that the power of both EbICoW and Wald increase with π–ψ, but the power of EbICoW increases in a faster rate than that of Wald. Hence, the power of EbICoW is larger than that of Wald when π–ψ is large. In contrast, the power of Wald is higher than that of EbICoW when π–ψ is small, especially when two GWASs have the same true effect. The statistical nature of EbICoW and Wald when combining two GWASs is illustrated in panels (a–g) of Fig. 1 and Supplementary Fig. 4. Notably, Wald only uses the joint distribution of null to determine its detection range, where the null is determined only by ψ. In contrast, EbICoW uses the joint distributions of both null and true effects to determine its detection range, which are determined by both π and ψ. This explains the power gain of EbICoW relative to Wald when π–ψ increases.
Scenario 4 compares the power between TWT, Wald, and MinGWAS for combining multiple GWASs (here we simulated ten GWASs) under the same α (Supplementary Fig. 5). Simulations showed that TWT has a similar or higher power than Wald and MinGWAS in all scenarios investigated. One extreme scenario is the presence of only one GWAS having a true effect and the absence of background correlations. Under this scenario, the power of TWT is almost the same as MinGWAS. Here the loss of power due to the extra burden of distribution calibration for TWT is negligible. Another extreme scenario is the presence of equally distributed true effect for all GWASs and the absence of background correlations. Under this scenario, the power of TWT is almost the same as Wald. For all other simulated scenarios, TWT has the optimal power. A general pattern is that compared with Wald, the gain of power in TWT decreases with the increase of π and increases with the increase of ψ.
Scenario 5 compares the power between EbICoW, TWT, MTAG and MinGWAS for combining two GWASs (Supplementary Fig. 6) as well as for combining multiple GWAS (here we simulated ten GWASs, Fig. 2) under the same α. For combining two GWASs, a general pattern is that when π–ψ = 0, the power of TWT > MTAG = MinGWAS > EbICoW, and when π–ψ becomes large, the power of EbICoW > MTAG > TWT > MinGWAS. The exception is that when the difference of true effect between two GWASs is large (\({{{\rm{E}}}}({\chi }_{12}^{2})=1.11.6\)), the power of EbICoW is always the highest, regardless of π–ψ. For combining multiple GWASs, we observed the following patterns. First, MinGWAS almost always performs the worst. Second, the power of MTAG is not the highest at any time, either lower than EbICoW or lower than TWT. Third, when π is large, EbICoW performs the best and when π is small, TWT performs the best. Fourth, when ψ ≠ 0, the relative power of TWT increases with the proportion of GWASs with true effects. The exception is the pattern 1, i.e., when only one GWAS has a true effect, in which case EbICoW always performs the best regardless of other parameters. These results convincingly demonstrate that a combined use of EbICoW and TWT as done in CGWAS outperforms MTAG and MinGWAS, which also likely holds true when comparing CGWAS with any statistical approach that is originated purely from the IVW or the Wald.
Scenario 6 assesses the performance of our function for estimating inflation (getI, Supplementary Fig. 7). Simulations showed that getI always produces unbiased estimation, although the estimation error negatively correlates with the size of true effect and positively correlates with the proportion of SNPs having true effects. On the other hand, it is obvious that the genomic control method produces an increasing bias with the increase of the true effect and with the increase of the proportion of SNPs having true effects.
Scenario 7 assesses the performance of the function for estimating ψ (getPsi, Supplementary Fig. 8). Simulations showed that getPsi always produces unbiased estimation of ψ with very high accuracies. Simply computing the correlation between two GWASs using all test statistics^{1} is not only biased but also has large estimation errors, especially when true effects are large. Computing the correlation between two GWASs by only focusing on nominally nonsignificant SNPs in their marginal distribution^{7} also produces biased results, i.e., it tends to overestimate ψ when π < 0 and underestimate ψ when π > 0.
Scenario 8 explores the α of all intermediate results from CGWAS, i.e., results from EbICoW and TWT, under three different levels of estimation errors of getI and getPsi (Supplementary Fig. 9). These levels were ascertained according to our simulation results from Scenario 6 and Scenario 7 to cover the whole range of estimation errors. Simulations showed that the studywide α of both EbICoW and TWT are stable at 5% or slightly higher than 5%. Larger estimation errors of inflation and ψ indeed led to an increased α, with a slightly pronounced effect for EbICoW. In the absence of estimation errors, α is always stable at 5%, while in the presence of estimation errors, the increase of the number of phenotypes slightly increased the studywide α. These results demonstrate that the intermediate results of CGWAS (EbICoW and TWT) have highly stable studywide α.
Scenario 9 evaluates the performance of getCoef in adjusting for the final pvalues from CGWAS (Fig. 3). Simulations confirmed that the final pvalues produced by getCoef always follow the uniform distribution under the null, regardless to the Ψ and regardless to the shape of the initial distribution. In multipletrait studies, a commonly used approach for multiple testing correction is to adjust for the number of independent tests, which can be derived either analytically using the matrix spectral decomposition analysis^{8} or empirically using simulations^{5}. Our simulations showed that using the number of independent tests guarantees a uniform distribution under the null only for correcting MinGWAS without background correlations, but is slightly biased at least in the presence of background correlations leaving alone other types of unknown initial distributions.
Scenario 10 assesses the performance of CGWAS and MTAG in combining GWASs of overlapping samples under extreme settings, i.e., we attempt to combine three GWASs of the same phenotype, which are conducted in fully overlapping, partially overlapping, or nonoverlapping subsamples. Our simulations showed that under all investigated settings, the test statistics from CGWAS for combining the three subsample GWASs tightly followed those from the GWAS conducted in the union of three subsamples (Supplementary Fig. 10). Instead of providing a single vector of combined test statistics as done in CGWAS, MTAG provides three sets of combined test statistics for combining the three GWASs. In fully overlapping samples, the results of MTAG also tightly followed those from the GWAS. When the sample overlapping substantially reduces, MTAG results slightly deviated from GWAS results, as explained by increased errors in estimating ψ and π (Supplementary Fig. 10).
CGWAS R package and computational efficiency
CGWAS was implemented as a userfriendly, publicly accessible, OSindependent, and parallel R package called “CGWAS”, https://github.com/FunGene/CGWAS. We assessed the performance of CGWAS and compared it with MTAG using simulated data of 6 million SNPs; 5, 10, 20, 30, 40, and 80 GWASs; and 1, 2, 4, 8, and 16 paralleled threads. CGWAS appeared more computationally effective than MTAG as indicated by several patterns (Supplementary Fig. 11). First, CGWAS utilizes multithreads in parallel to reduce computational time, while MTAG only runs in serial. Second, the peak memory usage of CGWAS becomes much smaller than that of MTAG when the number of GWASs increases e.g., exceeding 20. Third, the computational time of CGWAS become much less than MTAG when the number of GWASs exceeds 20, already under the singlethread setting. Fourth, MTAG becomes impractical when the number of GWASs exceeds 30 due to out of memory (>256 GB), while CGWAS completed the analysis of 80 GWASs in 1.3 hours with 16 threads in parallel and peak memory usage of 32 GB.
CGWAS outperforms MinGWAS in real data of the human face
To empirically demonstrate the power gain of CGWAS indicated in the simulations, we applied CGWAS to real data for which MinGWAS have been conducted previously^{4}, thereby allowing direct method comparison in the very same dataset. Human facial variation represents a multidimensional set of correlated and mostly symmetric traits with high heritability. The previous MinGWAS was conducted in 10,115 samples from four European cohorts^{4}, which involved 78 GWASs of different facial traits representing pairwise distances between 13 anatomically meaningful facial landmarks (Fig. 4a). The previous MinGWAS had identified 6 studywide significant loci and 24 studywide suggestively significant loci^{4}.
For a fair comparison, we adjusted the previous MinGWAS, which is by definition inflated under the null, using our getCoef function. A simulation analysis imbedded in CGWAS confirmed that under the null, both CGWAS and the adjusted MinGWAS tightly followed the uniform distribution (Fig. 4b, c). This means that the comparison among the two approaches is very strict under the same α. Therefore, the traditional genomewide significance threshold of (5 × 10^{−8}) used in a standard GWAS corresponds to our studywide significance threshold. Note that when applying the threshold of 5 × 10^{−8} in the adjusted MinGWAS, which corresponded to an unadjusted pvalue of 7.52 × 10^{−10}, i.e., the same number of SNPs survived 7.52 × 10^{−10} before and 5 × 10^{−8} after the getCoef calibration, is in fact 1.6 folds more stringent than the studywide significance threshold of 1.2 × 10^{−9} in the previous MinGWAS. This observation is consistent with our simulation scenario 9 showing that adjusting for MinGWAS using the estimated number of independent traits is prone to inflation under stringent thresholds. Below, MinGWAS refers to our getCoef calibrated MinGWAS unless otherwise specified.
Under the studywide significance threshold (5 × 10^{−8}), CGWAS identified 1200 SNPs from 23 distinct loci, while MinGWAS only identified 155 SNPs from 6 distinct loci. Not unexpected, the 6 loci identified by MinGWAS were the same 6 loci previously identified in the original study of Xiong et al.^{4}. Moreover, they were included in the 23 loci identified by CGWAS. Among the 17 loci exclusively identified by CGWAS, 3 (18%) were novel and 14 (82%) were reported in previous singletrait face GWASs^{5,9,10,11,12,13,14,15,16,17,18,19,20,21} including the study of Xiong et al.^{4}.
Moreover, under the studywide suggestively significant threshold estimated for CGWAS at 3 × 10^{−6}, CGWAS identified a total of 2591 SNPs from 56 distinct loci compared to only 418 SNPs from 17 distinct loci identified by MinGWAS (Fig. 4d–f), of which the majority (13) overlapped with CGWAS findings. Similar to the studywide significant threshold of 5 × 10^{−8}, the threshold of 3 × 10^{−6} in our MinGWAS is 1.15 folds more stringent than the studywide suggestive significance threshold of 5 × 10^{−8} in the study of Xiong et al.^{4}. The 13 loci overlapping between CGWAS and MinGWAS all showed an order of magnitude enhanced significance in CGWAS compared to MinGWAS. A prominent example is rs55674676 near 12q24.21 TBX3 which demonstrated pvalue enhancement from 2.58 × 10^{−8} in MinGWAS to 3.11 × 10^{−19} in CGWAS. The 4 MinGWASonly identified loci were all reported in previous GWASs. Of the 43 CGWASonly identified loci, 17 (40%) were novel and 26 (60%) were reported in previous GWASs^{5,9,10,11,12,13,14,15,16,17,18,19,20,21} including the study of Xiong et al.^{4}.
We doublechecked our results using the LDscore regression method^{22}, which estimated the LDscore intercepts close to 1.0 for both CGWAS and MinGWAS. This analysis confirmed that the signals from both CGWAS and MinGWAS cannot be explained by potential population substratifications and other unknown factors. Compared with MinGWAS, CGWAS gained 54% extra statistical power as estimated using the increase in the mean χ^{2} statistic method described in Turley et al.^{3}.
We looked up in our CGWAS results the 327 previously established faceassociated SNPs that were identified in traditional singleSNP GWASs^{5,9,10,11,12,13,14,15,16,17,18,19,20,21}. For these SNPs as well, CGWAS pvalues deviated further from the null than those from MinGWAS (Fig. 4g). The observation that CGWAS boosted the significance of MinGWAS findings and reidentified a much larger number of previously established loci than did MinGWAS empirically confirmed a substantial gain in statistical power and strengthened the reliability of the CGWASidentified novel loci, in line with our simulation results.
Next, we performed replication analyses for the 56 regional lead SNPs identified by CGWAS with studywide suggestively significance in a total of 13,622 individuals not overlapping with Xiong et al.^{4}. These replication data included the Rotterdam Study (N = 1174 Dutch Europeans), 2774 individuals of European descent from two cohorts^{5}, and 9674 Chinese from three cohorts^{6}. Combining replication evidence from all 6 cohorts revealed an overall highly significant deviation from the expectation under the null (KS test p = 9.7 × 10^{−16}, Fig. 4h). Moreover, among the 56 SNPs, 91% (51/56) showed consistent allele effects with the discovery cohort and 82% (46/56) was replicated at nominal significance (p < 0.05). Focusing on the 17 novel SNPs identified by CGWAS also showed an overall significant deviation from the expectation under the null (KS test p = 8.6 × 10^{−3}, Fig. 4h), with 82% (14/17) showing consistent allele effects with the discovery cohort, and 53% (9/17) replicating at the nominal significance (p < 0.05).
CGWAS outperforms MTAG and other methods in real data of the human face
Next, we compared the performance of CGWAS and MTAG in the real face data. To this end, because MTAG could not complete the analysis of all 78 GWASs within our computational capacity, we focused on a subset of 30 GWASs largely covering the left side of the strongly symmetrical human face. Although the H0 of CGWAS and MTAG are different, a fair comparison was achieved by adjusting the minimal pvalue of MTAG using the same getCoef function as in CGWAS. In addition, we compared CGWAS with four competing statistics, including mixAda, mixFisher and mixTippett proposed by Liu et al.^{23} and S_{Het} proposed by Zhu et al.^{2} (Supplementary Table 1). These four statistics test the same H0 as CGWAS, but do not distinguish Π and Ψ as CGWAS and MTAG. Both CGWAS and S_{Het} consider GWAS subsets whereas others do not.
Under the studywide significance threshold of 5 × 10^{−8}, CGWAS identified 649 SNPs from 16 distinct loci, all (100%) of which represent previously established facial loci. MTAG only identified five SNPs from two loci, one (50%) of which has been previously established (Supplementary Fig. 12). The mixAda appeared the best performing among the three statistics proposed by Liu et al.^{23}, which identified 18 SNPs from 12 loci, one (8%) of which has been previously established. The S_{Het} identified 252 SNPs from 16 loci, 12 (75%) of which overlapped with previously established loci. The 2q31.1 MTX2 was the only locus overlapping between all methods.
Considering the studywide suggestive threshold of 1.35 × 10^{−6}, CGWAS identified 1,043 SNPs from 32 distinct loci, 22 (69%) of which represent previously established facial loci. MTAG only identified 55 SNPs from 15 loci, 3 (20%) of which has been previously established (Supplementary Fig. 12). The mixAda again was the best performing statistic among those proposed by Liu et al.^{23}, which identified 160 SNPs from 61 loci, 9 (15%) of which have been previously established. The S_{Het} identified 880 SNPs from 56 loci, 18 (32%) of which overlapped with previously established loci. The 2q31.1 MTX2 and 12q24.21 TBX3 were the only loci overlapping between all methods.
The LDscore intercept of CGWAS was very close to 1.0 as expected, whereas the MTAG showed severe deflation (mean of 30 LDscore intercepts = 0.83; sd = 0.04). In contrary, the S_{Het} showed severe inflation (1.07) and the three statistics proposed by Liu et al. also showed nonnegligible inflation (>1.04) (Supplementary Table 1).
Finally, we looked up the 327 previously established faceassociated SNPs and it was obvious that CGWAS pvalues deviated much further from the null than did MTAG and mixAda pvalues, the latter two were largely nonsignificant (Supplementary Fig. 12). CGWAS replicated considerably more SNPs (49) than did S_{Het} (33) after Bonferroni correction (p = 2 × 10^{−4}), although S_{Het} pvalues also significantly deviated from the null (Supplementary Fig. 13).
The observation that MTAG showed severe deflations and reidentified only a small number of previously established loci appears surprising and suggests that the GWAS dataset of Xiong et al. is not suitable for MTAG. Note that the extra multiple testing burden posed to MTAG by forcing it to test our H0 had negligible impact given that MTAG performed similarly poor even without adjusting for the 30 facial traits. Complex patterns of SNP effects across different facial traits likely led to violation of the assumption in MTAG. The fact that S_{het} outperformed MTAG and mixAda indicates that considering GWAS subsets is important when the number of GWASs is large.
CGWAS outcome increases proportion of facial phenotype variance explained
Next, we compared CGWAS and MinGWAS in their outcomes to genetically explain the facial phenotype variance in RS based on a polygenic risk score (PRS) analysis. The PRS consisting of the 57 suggestively significant SNPs from CGWAS (56 regional lead SNPs plus one in low LD) and explained on average 2.28% and up to 4.51% sex and ageadjusted facial variance (Fig. 4i). In contrast, the PRS based on MinGWAS consisting of 17 suggestively significant lead SNPs only explained on average 0.84% and up to 2.37% of facial variance (Fig. 4j). The PRS based on MinGWAS consisting of 57 topranked SNPs explained on average 2.03% and up to 4.33% sex and ageadjusted facial variance (Supplementary Fig. 14), but is significantly lower than the PRS based on 57SNP suggestively significant SNPs from CGWAS (twoside paired Wilcoxon ranksum test p = 9 × 10^{−4}).
We then compared face PRSs between Europeans (EUR), SubSaharan Africans (AFR) and East Asians (EAS) in totally 1668 samples from the 1000Genomes Project. Although the face PRSs were solely derived from Europeans, when applied to nonEuropean populations, the CGWAS PRS largely assembled the facial features in major continental groups^{6,24,25}, which appeared more consistent with anthropological knowledge than the MinGWAS PRS (Supplementary Fig. 14). For example, genetically, AFR mouth appeared larger and AFR face appeared wider than those of EUR and EAS, EUR mouth appeared smaller and EUR nose appeared narrower and more protuberant than those of AFR and EAS, and EAS face was overall less distinct, largely in the middle between AFR and EUR. Overall, these trends appeared more obvious in CGWAS PRS than MinGWAS PRS (Supplementary Fig. 14).
Multitrait facial effects of CGWAS outcome
To empirically test whether CGWAS indeed identified loci with multitrait effects, as expected based on the method design, we developed a statistic to empirically estimate multitrait effect (MTE, see “Methods”). We then estimate MTE of the 56 studywide suggestively significant SNPs identified by CGWAS, including 13 suggestively significant also in MinGWAS. Overall, 26 eigen vectors were derived to represent the 78 GWASs after removing background correlations (see “Methods”). The 43 CGWASonly identified SNPs were more enriched with larger MTE (>0) values than the 13 significant SNPs identified by both CGWAS and MinGWAS (twoside Fisher’s exact test p = 0.026, odds ratio = 5.06). In addition, the 17 SNPs from novel loci identified with CGWAS were significantly more enriched with larger MTE (>0) values than the 39 rediscovered SNPs reported in previous GWASs (twoside Fisher’s exact test p = 0.025, odds ratio = 8.4, Fig. 5a).
Eight examples were examined in more detail, including two novel loci, EIF4G3 and CDK2AP1 with suggestively significant association and large MTE as well as six reidentified previously established loci with studywide significant association and varying MTE: PAX3, SFRP2 (large MTE), TBX3, SIX2 (moderate MTE), INTU and RPE65 (small MTE, Fig. 5b). Their allelic effects of the regional lead SNPs on the 26 eigen vectors displayed two clear patterns. First, the further a SNP effect deviated from the null, the more significant its association was in CGWAS (example, PAX3 vs EIF4G3, both with large MTE). Second, the earlier a SNP effect deviated from the null, the larger its MTE value was (example, PAX3 vs TBX3 vs INTU, all established loci with studywide significant association). Consistently, SNPs with large MTE values were generally associated with a larger number of facial traits with different effects, which was seen not only for the examples mentioned here (Fig. 5c), but also for other faceassociated SNPs from CGWAS (Supplementary Fig. 15). These results empirically demonstrate the multitrait effects of CGWAS findings and illustrate the power of CGWAS in identifying genetic variants with various degrees of multitrait effects.
Biological annotations of CGWAS and MinGWAS findings
Gene ontology analyses were conducted for comparing the CGWAS (83 genes in 56 loci) and MinGWAS (24 genes in 17 loci) findings in Gene Ontology (GO), Human Phenotype (HP), and Mouse Phenotype (MP) databases. Overall, CGWAS loci harbored genes significantly enriched with a larger number of facial morphogenesis relevant terms relative to MinGWAS loci (Fig. 6a), with the term “abnormality of external nose” being the most significant in HP (p = 3.36 × 10^{−9}). For the significant biological terms found by both CGWAS and MinGWAS, those from CGWAS were generally more significant than those from MinGWAS. The vast majority of terms from MinGWASidentified loci were covered by terms from CGWASidentified loci. The unique terms from CGWASidentified loci not only revealed the facial morphogenesis pathway, but also revealed the nervous development pathway including ‘central nervous system development’ and ‘nervous differentiation’, which is consistent with the recent observation of overlapping findings between GWASs of facial traits and those of brain traits^{26}.
Cranial neural crest cells (CNCC) reflect a migratory cell population at early human development involving peripheral nervous system, skin pigment cells, and craniofacial bones. Previously, CNCC regulatory activity in faceassociated loci has been used to elucidate their functional roles in facial morphogenesis^{4,5,17}. Here, we applied the recently published regulatory network of CNCC^{27} to annotate our CGWAS and MinGWAS findings. Among 56 CGWASidentified loci, 26 (46%) were found to contain CNCCassociated regulatory elements (REs), while only 5 (29%) out of 17 MinGWASidentified loci contained CNCCassociated REs (ISTAT^{28} enrichment p = 2.16 × 10^{−4} for CGWAS and p = 0.04 for MinGWAS). Furthermore, the subnetwork obtained by CGWASidentified loci regulates more target genes (TGs) than the one obtained by MinGWASidentified loci (54 vs. 8 TGs, Fig. 6b, c). For example, ALX1 and TCF4, which are involved in CNCC development^{29} and identified in regulatory network of CNCC^{27} as the core transcription factors^{30}, are exclusively from CGWAS. These results demonstrated that CGWAS loci were more biologically interpretable than MinGWAS loci.
Finally, we investigated the colocalization of genetic association at 56 CGWASidentified loci with eQTL signals by using expression data from 22 classified tissues potentially relevant to facial morphogenesis. Previously, a smaller scaled colocalization analysis had been conducted for 20 faceassociated loci in 6 tissues^{31}, which highlighted EEFSEC with evidence of colocalization. Our study confirmed this finding and additionally revealed many novel loci with evidence of colocalization in various tissues. Overall, 38 genes in 17 loci were identified with large posterior probabilities (PP4>0.7) of colocalization in at least one tissue (Fig. 7a). These included 23 genes in 7 novel loci identified by CGWAS, i.e., 1q42.3, 2q32.1, 2q34, 4p16.3, 11q12.2, 12q24.31, and 13q14.3. Two novel (RP11410E4.1 at 2q32.1 and DLEU7AS1 at 13q14.3) and two previously established faceassociated genes (RPGRIP1L at 16q12.2 and KCTD15 at 19q13.11, Fig. 7b–e) colocalized with eQTLs in a large number of different tissues (≥9). Thirteen out of the 17 loci contained at least one colocalized genes that were different to the genes closest to the regional lead SNPs. These results provided direct evidence for functional involvement of some of the faceassociated SNPs, proposed a different set of functional candidate genes, and strengthened the reliability of 7 novel loci from CGWAS.
Novel loci highlighted by CGWAS
Among the 17 novel loci identified by CGWAS, 13 are worth highlighting as they were supported by the replication study, the CNCC network analysis, or the colocalization analysis (Supplementary Data 1), of which we mention the most prominent locus here and the other 12 in Supplementary Note 1. The strongest supportive evidence was observed at the novel locus 12q24.31, where the lead SNP rs10773002 is an intron variant of CDK2AP1. The Aallele of rs10773002, minor in Europeans (f = 0.24) and Asians (f = 0.25) but major in Africans (f = 0.79), was nominally significantly associated with a wide spread of facial phenotypes, mainly contributing to reduced facial width and increased facial length (Fig. 5c). The multitrait effect of this SNP was successfully replicated (p = 6.9 × 10^{−4}). The REs in the vicinity of the faceassociated SNPs at this locus regulate several nearby genes, including CDK2AP1 in the CNCC subnetwork. Faceassociated SNPs in this locus colocalized with eQTLs of CDK2AP1 and nearby genes in testis (PP4 = 0.97) and other tissues. The protein encoded by CDK2AP1 is a cyclindependent kinase 2 associated protein, which plays a role in cellcycle, embryonic stem cell differentiation and epigenetic regulation. SNPs in this gene have been associated with educational attainment and waisttohip ratio in previous GWAS^{32,33}.
Discussion
In this study, we introduce CGWAS, a method to integrate GWAS results of multiple correlated traits with an increased statistical power in detecting multitrait effects. EbICoW is originated from the inverse variance weighting method with three extensions, which allows using the joint distributions of both null and true effects to detect multitrait effects. This is particularly powerful when π can be clearly differentiated from ψ, e.g., when the genetic determinants shared between traits are substantial and/or when samples do not overlap or partially overlap. When it is difficult to distinguish π from ψ, Wald, which does not rely on the joint distribution of true effects, is more powerful, e.g., when traits are mainly determined by their own unique genetic factors and/or when samples overlap completely or mostly. Utilizing the statistically complementary properties of EbICoW and Wald, CGWAS provides excellent power in a variety of complex situations, as we have demonstrated via simulations and in real face data.
In CGWAS, the decision on whether or not using EbICoW to combine a given pair of GWASs in each iteration is made via a selfadaptive datadriven approach. Such decisions, optimal regarding data fitness, may not always fit the true underlying genetic architectures of the studied traits. This may affect the power in detecting multitrait effects, but not the pvalue distributions under the null, as double guaranteed by both the statistical nature of EbICoW and our approach for distribution calibration.
Coming with CGWAS, three methods were additionally developed for inflation estimation, background correlation estimation, and distribution calibration, respectively. The first two methods provide accurate estimates of inflation and background correlation, which is critical for CGWAS to work properly. Compared with previous LD scorebased methods^{22,34}, our methods abandon LDrelated calculation for computational efficiency, while maintaining similarly levels of accuracy and estimation error. More importantly, our methods have less limitations in the presence of GWASs with different genomic backgrounds. The third method for distribution calibration uses a quantilebased function that can convert any given distribution to a uniform distribution, which guarantees unbiased estimates as long as the null can be obtained via simulations, which allows fair comparisons between CGWAS, MinGWAS, and any given standard GWAS. These methods are generic and flexible (requiring only summary statistics), so that they can be widely applied to genetic studies of complex traits.
CGWAS has the flexibility to be conducted as a whole or by parts considering only EbICoW or TWT. We recommend to carry out CGWAS as a whole when the expected differences between Π and Ψ among a large number of GWASs are largely unknown, so that our selfadaptive decisionmaking algorithm can assist in optimizing statistical power. In cases of nonoverlapping GWASs where all traits are expectedly highly correlated, we recommend using EbICoW only. Generalization of CGWAS to other omics data requires the null following multidimensional standard normal distributions.
Methods
Ethnic statement
This study complies with all relevant ethical regulations. The Rotterdam Study has been approved by the Medical Ethics Committee of the Erasmus MC (registration number MEC 02.1015) and by the Dutch Ministry of Health, Welfare and Sport (Population Screening Act WBO, license number 1071272–159521 PG). The Rotterdam Study has been entered into the Netherlands National Trial Register (NTR; www.trialregister.nl) and into the WHO International Clinical Trials Registry Platform (ICTRP; www.who.int/ictrp/network/primary/en/) and under shared catalog number NTR6831. All participants provided written informed consent to participate in the study and to have their information obtained from treating physicians. Other data used here come from previous publications^{4,5,6}, where the respective ethics statements are provided.
Definition of background and effect correlations
Consider K GWAS of M SNPs, under the null, for each SNP, its T statistics from different GWASs follow a sum of two multidimensional normal distributions including the estimate error and the inflation as \({{{{{\bf{t}}}}}}={({t}_{1},{t}_{2},\ldots,{t}_{K})}^{{{{{{\rm{T}}}}}}}\). The covariance matrix of the estimate error has a diagonal of 1 and an offdiagonal of zero, which may deviate from zero in presence of sample overlap and nongenetic phenotype correlations (ρ), e.g., caused by shared environmental or unknown factors. The mean squared inflation is zero and may deviate from zero in the presence of cryptic relatedness and population stratification. In the absence of inflation or when inflation is well controlled, the correlation ψ (K = 2) or the correlation matrix Ψ (K > 2) between the summary statistics of GWAS T statistics, T∈(T_{1}, T_{2},…, T_{K}), is defined as ‘background correlation’. For two GWASs of sizes N_{1} and N_{2} with an overlapping sample of size N_{c}, ideally \(\psi=\frac{{N}_{c}}{\sqrt{{N}_{1}{N}_{2}}}\rho\). Therefore, when the sample does not overlap at all or in the absence of nongenetic phenotype correlations, Ψ is an identity matrix.
Under the polygenicity, a proportion of SNPs additionally has true effects in at least one GWAS. Similar to ψ or Ψ, the correlation π (K = 2) or the correlation matrix Π (K > 2) between the true effects of T is defined as ‘effect correlation’. With these definitions, the relationship between any given numbers of GWASs can be described using Ψ and Π.
CGWAS flowchart
The CGWAS flowchart is shown in Supplementary Fig. 1. Consider K GWASs G∈(G_{1}, G_{2},…, G_{K}), where G_{i} contains effect sizes, standard errors, and T statistics of the GWAS i, CGWAS takes G as the input, which is passed to iEbICoW followed by TWT and distribution calibration. The important parameters for iEbICoW include the vector of inflation factors \({{{{{\bf{s}}}}}}={({s}_{1},\,{s}_{2},\ldots,\,{s}_{K})}^{{{{{{\rm{T}}}}}}}\) for all GWASs, estimated using the getI function; the background correlation matrix Ψ, estimated using the getPsi function; the effect correlation matrix Π, estimated using the getPi function, and the effect vector h, estimated using the getH function. iEbICoW iteratively combines GWASs in pairs, and the order of combination is determined by the matrix D, which is the difference between Π and Ψ. The core functions of iEbICoW involve EbICoW, optimize, and evaluate. EbICoW combines two GWAS based on an effectbased inversed covariance weighting method. Optimize optimizes the result of EbICoW within three possible choices of h. Evaluate makes the decision whether or not accepting the result from Optimize by comparing the resultant pvalues with those from Wald and MinGWAS. D is updated in each iteration until all elements of D are sufficiently small. In such a way, iEbICoW generates a set of new G so that no result from Optimize for any pair of G are accepted by Evaluate. These newly derived G are further passed to TWT and distribution calibration. TWT analyzes iEbICoW results in a SNPspecific manner. For each SNP, TWT takes the subset of G with the minimal pvalue from iEbICoW as the initial subset, then extends to a series of subsets under a gradient (n = Q) of preset pvalue thresholds \({{{{{\bf{r}}}}}}=({r}_{1},\,{r}_{2},\ldots,\,{r}_{Q})\), and then conducts the Wald test on all subsets to obtain the combined pvalues of all subsets. The analysis for each SNP under Q thresholds is realized through a nested loop. The distribution calibration is achieved in an empirical manner via simulations through two progressive steps. The first step is to correct all intermediate results to make them comparable with each other. It uses getNtest to estimate the number of independent tests from a larger number of dependent hypothesis testing results via simulations, then uses the Tippett function to correct the pvalue for each TWT combinations using the number of independent tests. The second step is to guarantee the final results from CGWAS under the null follow the uniform distribution. It uses getCoef to build a correction model by fitting the observed distribution under the null (via simulations) with the uniform distribution for all quantiles. Then the minimal pvalue of all results from the first step is further corrected based on the quantile specific coefficients from getCoef. In this way, CGWAS produces the final results as a single vector of combined pvalues, which follows the uniform distribution under the null, so directly comparable with the pvalues from any standard GWAS.
Effectbased inversed covariance weighting (EbICoW)
We propose EbICoW to allow joint analysis of multiple correlated phenotypes with overlapping samples. Let β be the vector of effect sizes from K GWASs for a SNP, \({{{{{\bf{\upbeta }}}}}}={({\beta }_{1},{\beta }_{2},\ldots,{\beta }_{K})}^{{{{{{\rm{T}}}}}}}\). The null of EbICoW is the absence of any effect i.e., H_{0}: β = 0, and the alternative is that at least one element of the effect vector deviates from 0. Below, we formulate EbICoW and show that the test statistics from EbICoW asymptotically follow the standard normal distribution. We prove that the traditional Inversed Variance Weighting (IVW) based metaanalysis is a special case of EbICoW when the offdiagonal of Ψ is 0 and the offdiagonal of Π is 1. We also prove that the S_{Hom} proposed by Zhu et al.^{2} is a special case of EbICoW when the true effect is the same for all GWAS and the offdiagonal of Π is 1.
For each SNP, the traditional IVW based metaanalysis combines results from K nonoverlapping GWASs of the same phenotype, requiring that the offdiagonal of Ψ is 0 and the offdiagonal of Π is 1. The combined effect β_{c} from the metaanalysis represents the weighted sums of K effects, β_{c} = w^{T}β, where β~N(0,V), V is a K by K diagonal matrix, and the vector of weights \({{{{{\bf{w}}}}}}={({w}_{1},{w}_{2},\ldots,{w}_{K})}^{{{{{{\rm{T}}}}}}}\) is inversely proportional with the variance of β, i.e., diagonal elements of V, and the sum of elements in w^{T} is set to 1. In practice, w^{T} can be constructed using the vector of reciprocal standard errors \({{{{{\bf{\upnu }}}}}}=(1/{\sigma }_{1},\,1/{\sigma }_{2},\ldots,\,1/{\sigma }_{K})\) or the samples sizes \({{{{{\bf{n}}}}}}=({n}_{1},\,{n}_{2},\ldots,\,{n}_{K})\) from K GWASs. Let b be a vector of 1 with length K,
For each SNP, the test statistics of IVW (t_{V}) can be described using β and ν or using t and n as
Under the null, t_{V} asymptotically follows the standard normal distribution.
When the offdiagonal elements of Ψ is not 0, β~N(0,C), where C is the covariance matrix of β under the null, w^{T} can be derived under the same frame of weighted sums using the Lagrange multiplier method (Supplementary Method A). Thus, we have
The revised statistics t_{C} is expressed as follows,
where e is the vector of weights only relying on n, i.e., ν^{T} or \({(\sqrt{{n}_{1}},\sqrt{{n}_{2}},\ldots,\sqrt{{n}_{K}})}^{{{{{{\rm{T}}}}}}}\), but not on Ψ or Π. Under the null, the revised t_{C} in (6) also asymptotically follows the standard normal distribution. The S_{Hom} proposed by Zhu et al. has the same null as ours thus directly comparable with ours. S_{Hom} is based on a chisquared statistic with 1 df, which is in fact equivalent to the t_{C}^{2} in (6) when it uses ν^{T} or \({(\sqrt{{n}_{1}},\sqrt{{n}_{2}},\ldots,\sqrt{{n}_{K}})}^{{{{{{\rm{T}}}}}}}\) as weights. The only difference here is that the effects can be directly estimated as w^{T}β using our revised weight but not available from S_{Hom}. Although this t_{C} statistic can be used for joint analysis of multiple GWASs when the offdiagonal of Ψ is not 0, it has limited power when expected true effect \({{{{{\bf{\updelta}}}}}}={({\delta }_{1},{\delta }_{2},\ldots,{\delta }_{K})}^{{{{{{\rm{T}}}}}}}\) is not the same for all GWAS or the offdiagonal of Π deviates from 1. The loss of power becomes obvious when the Kbased weights are the same across all GWAS but the true effects are different. The loss of power also become obvious when the offdiagonal of Π is negative, and particularly pronounced when it is approaching 1. To overcome these limitations, we made two improvements. First, we introduce an effect vector, \({{{{{\bf{h}}}}}}\,{{{{=}}}}\,{({h}_{1},{h}_{2},\ldots,{h}_{K})}^{{{{{{\rm{T}}}}}}}\), which is a projection of the true effect covariance matrix Η as h = b^{T}Η. Η is a K by K covariance matrix of δ among the genomewide SNPs, where \({{H}}_{{ij}}={{{{{\rm{E}}}}}}({\delta }_{i}{\delta }_{j})\), i ∈ {1, 2,…, K} and j ∈ {1, 2,…, K}. We use h to revise w^{T} as
A nice feature of (7) is that h only affect power but not α. In practice, h can be estimated using the summary statistics from GWAS (details see estimation of h section). It can be shown that (5) is a special case of (7) when δ is same across all GWAS,
It can also be empirically shown that when the δ is different across GWASs, the test based on (7) has higher power than that based on (5) in most circumstances.
The second improvement is that we take the direction of combination into consideration, i.e., by changing the direction of combination if π is smaller than ψ while keeping w^{T}sign(ΠΨ)_{i} equals to b^{T}, where the function sign is defined as sign(x) = 1 when x ≥ 0 and sign(x) = −1 when x < 0 for a variable x so that \({[{{{{{\rm{sign}}}}}}\left({{{{{\bf{X}}}}}}\right)]}_{{ij}}={{{{{\rm{sign}}}}}}({X}_{{ij}})\) for a matrix X. For k\(\in\){1, 2,…, K},
where \({{{{{\bf{h}}}}}}={{{{{{\rm{sign}}}}}}\left({{{{{\mathbf{\Pi }}}}}}{{{{{\boldsymbol{}}}}}}{{{{{\mathbf{\Psi }}}}}}\right)}_{k}{{{{{\bf{H}}}}}}\). In this way, the direction of elements in the effect covariance matrix is changed by comparing the effect and background correlations. It can be empirically shown that when π is smaller than ψ, the test based on (8) has higher power than that based on (7) in combining two GWASs. However, (8) does not generate unique combined results for combining more than two GWASs, meaning that for obtaining onedimensional summary statistics in the form of traditional GWAS, an iterative procedure is preferred (see Combining multiple GWASs section). Summing up, the revised t_{E} of our proposed EbICoW is expressed as
Note that a statistically desired feature of EbICoW is that t_{E} asymptotically follows the standard normal distribution under the null regardless to h.
Wald test, Wald(G)
EbICoW uses additional information provided by effect correlation Π and effectbased factor to improve the power of detecting SNPs with true effects. However, the increase of power depends on ΠΨ. For example, considering two GWAS, the smaller π–ψ is, less obvious is the power improvement of EbICoW. To overcome this limitation, we used the multitrait chisquared statistic proposed previously by Bolormaa et al.^{1} when all elements of ΠΨ are small. The nature of the multitrait chisquared test is Wald’s test. For each SNP, the test statistic of the Wald test for combining multiple test statistics from K GWAS asymptotically follows the chisquared distribution with K df, and can be estimated by considering only the background correlation,
The fact that the Wald test does not rely on Π compensate the power loss of EbICoW when all elements of ΠΨ is small.
Combining multiple GWASs
CGWAS uses both EbICoW and Wald to improve the power of detecting SNPs with multitrait effects. Here the key concepts involve the decisionmaking, subset identification, and distribution calibration. Below we present an algorithm with two consecutive modules to achieve these purposes. The first module is iterative EbICoW (iEbICoW), which iteratively apply a decisionmaking algorithm to select subsets suitable for EbICoW, resulting in a set of EbICoW combined GWAS and potentially some individual GWAS, in such a way that all GWAS pairs from iEbICoW are no longer suitable for applying EbICoW. The second module is a truncated Wald test (TWT) in combination with distribution calibration. TWT further analyze iEbICoW results in a SNPspecific manner, which produce all possible SNPspecific subsets using a series of predefined thresholds. Then a series of multiple testing corrections, using getNtest and Tippett, are carried out for all intermediate results to ensure all thresholdbased subsets have the same level of α. The choice of best subset is made according to the levels of significance from TWT of all subsets after adjusting. Finally, getCoef guarantees that the final results from CGWAS are fully adjusted for the number of GWASs and for all intermediate iterative steps, thus directly applicable using the traditional genomewide significant threshold as our studywide significance threshold.
The first module: iEbICoW
Theoretically, π–ψ\(\in\)[0, 2]. The more π–ψ is deviated from 0, the higher the power of EbICoW than that of Wald and vice versa. We introduce a decisionmaking algorithm for the choice of using EbICoW by taking π–ψ as one of the key parameters into consideration. In addition, when combining a large number of GWAS, the relationships between Π and Ψ tend to be complex. SNPs unlikely share the same Π across all GWAS and π–ψ may be sufficiently large only in a subset of GWAS. This requires finding the subsets best suitable for EbICoW, which is achieved using a datadriven approach by iteratively applying our decisionmaking algorithm.
Consider K GWAS \({{{{{\bf{G}}}}}}\in ({{{{{{\bf{G}}}}}}}_{{{{{{\bf{1}}}}}}},\,{{{{{{\bf{G}}}}}}}_{{{{{{\bf{2}}}}}}},\ldots,\,{{{{{{\bf{G}}}}}}}_{{{{{{\boldsymbol{K}}}}}}})\), where G_{i} contains all summary statistics of the GWAS k, including the test statistic vector T_{k}, the effect vector Β_{k}, and the pvalue vector P_{k} of genomewide SNPs, derived Ψ, Π, and h using functions Ψ = getPsi(G), Π = getPi(G), and h = getH(G). The getPsi, getPi and getH functions are described in details in the Estimation of ψ and Estimation of π and h sections. Define F as minimum difference of π and ψ allowed for combined analysis in each iteration. Default of F is set to 0.05. Set U=0, U is a count of iterations in which evaluate returns TRUE, and D=ΠΨ, D is a K by K symmetric matrix of diagonal of 0 to reflect difference of π and ψ for each pair of GWASs. The process of iEbICoW is described below:
while any element in D > F {
(i, j) = which(Max(D)); # get the index of the maximal element of D
G_{E} = optimize(EbICoW(G_{i}, G_{j}, Ψ, Π, h));
P_{W} = Wald(T_{i}, T_{j}); # using (9) for t of each SNP
P_{A} = Min(P_{i}, P_{j})\(;\) # get the minimal pvalue of each SNP
if evaluate(G_{E}, P_{W}, G_{A}) is TRUE {
Remove(G_{i}, G_{j}) from G; G = (G, G_{E});
Ψ = getPsi(G); Π = getPi(G); D = ΠΨ; U = U+1;
} else {D_{ij}=0}};
The optimize function (Supplementary Method B) is to select the optimal weight for EbICoW (5) to combine two GWAS where the effect correlation π and effect vector h is chosen from three our proposed ones, \({{{{{{\bf{G}}}}}}}_{{{{{{\bf{all}}}}}}}={{{{{\rm{EbICoW}}}}}}({{{{{{\bf{G}}}}}}}_{{{{{{\boldsymbol{i}}}}}}},\,{{{{{{\bf{G}}}}}}}_{{{{{{\boldsymbol{j}}}}}}},\,{{{{{\mathbf{\Psi }}}}}},{{{{{{\mathbf{\Pi }}}}}}}_{{{{{{\bf{all}}}}}}},{{{{{{\bf{h}}}}}}}_{{{{{{\bf{all}}}}}}})\); \({{{{{{\bf{G}}}}}}}_{{{{{{\bf{sig}}}}}}}={{{{{\rm{EbICoW}}}}}}({{{{{{\bf{G}}}}}}}_{{{{{{\boldsymbol{i}}}}}}},\,{{{{{{\bf{G}}}}}}}_{{{{{{\boldsymbol{j}}}}}}},\,{{{{{\mathbf{\Psi }}}}}},{{{{{{\mathbf{\Pi }}}}}}}_{{{{{{\bf{sig}}}}}}},{{{{{{\bf{h}}}}}}}_{{{{{{\bf{sig}}}}}}})\); \({{{{{{\bf{G}}}}}}}_{{{{{{\bf{stb}}}}}}}={{{{{\rm{EbICoW}}}}}}({{{{{{\bf{G}}}}}}}_{{{{{{\boldsymbol{i}}}}}}},\,{{{{{{\bf{G}}}}}}}_{{{{{{\boldsymbol{j}}}}}}},\,{{{{{\mathbf{\Psi }}}}}},{{{{{{\mathbf{\Pi }}}}}}}_{{{{{{\bf{stb}}}}}}},{{{{{{\bf{h}}}}}}}_{{{{{{\bf{stb}}}}}}})\). The proposal of the three π and h is described in detail in the Estimation of π and h section. The evaluate function make decision whether accept combined result from optimize via evaluating the power and robustness of G_{E} (Supplementary Method C). iEbICoW returns KU newly obtained GWAS \(({{{{{{\bf{G}}}}}}}_{{{{{{\bf{1}}}}}}},\,{{{{{{\bf{G}}}}}}}_{{{{{{\bf{2}}}}}}},\ldots,\,{{{{{{\bf{G}}}}}}}_{{{{{{\boldsymbol{K}}}}}}{{{{{\boldsymbol{}}}}}}{{{{{\boldsymbol{U}}}}}}})\). All pairs of updated G are no longer suitable for applying EbICoW and these results are further passed to the second step analysis.
The second module: TWT and distribution calibration
TWT analysis is carried out on the KU newly obtained GWAS \(({{{{{{\bf{G}}}}}}}_{{{{{{\bf{1}}}}}}},\,{{{{{{\bf{G}}}}}}}_{{{{{{\bf{2}}}}}}},\ldots,\,{{{{{{\bf{G}}}}}}}_{{{{{{\boldsymbol{K}}}}}}{{{{{\boldsymbol{}}}}}}{{{{{\boldsymbol{U}}}}}}})\) from iEbICoW. Set a vector of thresholds as \({{{{{\bf{r}}}}}}=({r}_{1},\,{r}_{2},\ldots,\,{r}_{Q})\), where r\(\in\)[1, 0) in the descending order. In default \({{r}_{q}=10}^{{{{{{\boldsymbol{q}}}}}}/3}\) and Q = 18, where q = (1, 2, …, Q). For each SNP m\(\in\){1, 2,.., M}, the vector of pvalue and T statistics are \({{{{{\bf{p}}}}}}={({P}_{m,1},\, {P}_{m,2},\ldots,\, {P}_{m,KU})}^{{{{{{\rm{T}}}}}}}\) and \({{{{{\bf{t}}}}}}={({T}_{m,1},\, {T}_{m,2},\ldots,\, {T}_{m,KU})}^{{{{{{\rm{T}}}}}}}\). The vector of combined pvalues \({{{{{\bf{cp}}}}}}\,{{{{=}}}}\,{({{cp}}_{1},\, {{cp}}_{2},\ldots,\, {{cp}}_{Q})}^{{{{{{\rm{T}}}}}}}\) from a series of rbased subsets is derived as cp = TWT(p,t), where TWT is a function of p and t as defined below:
For each element r_{q} in r {
id = which\(({{{{{\bf{p}}}}}}{{{{{\rm{\le }}}}}}{r}_{q})\);
if (length(id)=0) {\({{cp}}_{q}={{{{{\rm{NA}}}}}}\)};
else {\({{cp}}_{q}={{{{{{\rm{F}}}}}}}_{{\chi }_{{{{{{\rm{length}}}}}}\left({{{{{\bf{id}}}}}}\right)}^{2}}({{{{{{\bf{t}}}}}}}_{{{{{{\bf{id}}}}}}}^{{{{{{\bf{T}}}}}}}{({{{{{{\mathbf{\Psi }}}}}}}_{{{{{{\bf{id}}}}}}{{{{{\boldsymbol{,}}}}}}{{{{{\bf{id}}}}}}})}^{1}{{{{{{\bf{t}}}}}}}_{{{{{{\bf{id}}}}}}})\)}};
Return cp;
\({{{{{{\rm{F}}}}}}}_{{\chi }_{Y}^{2}}\left(x\right)\) denotes the cumulative distribution function of a chisquared random variable x with Y degrees of freedom. The minimal pvalue of p can be regarded as a oneelement subset. Therefore, \({{cp}}_{Q+1}={{{{{\rm{Min}}}}}}({{{{{\bf{p}}}}}})\). After applying TWT to all SNPs on the genome, we have a pvalue matrix CP with M rows (the number of SNPs) and Q+1 columns (the number of preset thresholds add one of minimum), where \({{{{{\bf{CP}}}}}}\,{{{{=}}}}\,{{{{{{\boldsymbol{[}}}}}}{{{{{{\bf{CP}}}}}}}_{{{{{{\bf{1}}}}}}},\, {{{{{{\bf{CP}}}}}}}_{{{{{{\bf{2}}}}}}},\ldots,{{{{{{\bf{CP}}}}}}}_{{{{{{\boldsymbol{M}}}}}}}{{{{{\boldsymbol{]}}}}}}}^{{{{{{\bf{T}}}}}}}\).
The next idea is for each row CP_{m} to find the best subset under the threshold r_{q}, which corresponds to the minimal pvalue in this row. However, the results from this selection procedure deviate from any known distribution. Multiple testing correction and distribution calibration is necessary to guarantee all columns of CP and final bestsubsetbased results are comparable at the same α. To achieve this we firstly adjusted all columns empirically in CP using NS rounds of simulations as below:
For ns in 1 to NS {\({{{{{{\bf{t}}}}}}}_{{{{{{\bf{simu}}}}}}} \sim N(0,{{{{{\mathbf{\Psi }}}}}})\); p_{simu}\(={{{{{{\rm{F}}}}}}}_{{\chi }_{1}^{2}}({{{{{{\bf{t}}}}}}}_{{{{{{\bf{simu}}}}}}} \circ {{{{{{\bf{t}}}}}}}_{{{{{{\bf{simu}}}}}}})\); \({{{{{{\bf{SP}}}}}}}_{{{{{{\boldsymbol{ns}}}}}}}={{{{{\rm{TWT}}}}}}\left({{{{{{\bf{p}}}}}}}_{{{{{{\bf{simu}}}}}}},\,{{{{{{\bf{t}}}}}}}_{{{{{{\bf{simu}}}}}}}\right)\)};
For q in 1 to Q+1 {\({{{{{{\bf{CP}}}}}}}_{{{{{{\boldsymbol{q}}}}}}}={{{{{\rm{Tippett}}}}}}({{{{{{\bf{CP}}}}}}}_{{{{{{\boldsymbol{q}}}}}}},{{{{{\rm{getNtest}}}}}}({{{{{{\bf{SP}}}}}}}_{{{{{{\boldsymbol{q}}}}}}}))\)};
t_{simu} is a random vector of length KU generated from multivariable standard normal distribution with updated Ψ (after iEbICoW) in each round of simulations, and SP is pvalue matrix with NS rows and Q+1 columns containing all simulation results. Multiple testing correction methods Tippett and getNtest function used here is detailed in Supplementary Method D. With the adjustment described above, the best subset can be obtained from the adjusted CP. Take the minimum of each CP_{m} as the final output \({{{{{\bf{min }}}}}}{{{{{{\bf{CP}}}}}}}=(\min ({{{{{{{{\bf{CP}}}}}}}}}_{{{{{{{{\bf{1}}}}}}}}}),\,\min \left({{{{{{{{\bf{CP}}}}}}}}}_{{{{{{{{\bf{2}}}}}}}}}\right),{{\ldots }},\,\min ({{{{{{{{\bf{CP}}}}}}}}}_{{{{{{{{\boldsymbol{M}}}}}}}}}))\) with studywide significance, we further adjust minCP to guarantee its null distribution and α are same as any conventional GWAS, i.e., uniform distribution and α = 5 × 10^{8} respectively. The selection of minimal pvalue among the adjusted pvalues in all combinations obtained by TWT inevitably results in a nonuniform null and the inflation of this nonuniform null is quantile specific in nature. In order to make all final pvalues from CGWAS follow the uniform distribution under the null, we obtain the nonuniform null via the same NS rounds of simulations and then calibrate the simulated nonuniform null to a uniform distribution using a quantilebased method (getCoef function, Supplementary Method D) as below:
For q in 1 to Q+1 {\({{{{{{\bf{SP}}}}}}}_{{{{{{\boldsymbol{q}}}}}}}={{{{{\rm{Tippett}}}}}}({{{{{{\bf{SP}}}}}}}_{{{{{{\boldsymbol{q}}}}}}},\, {{{{{\rm{getNtest}}}}}}({{{{{{\bf{SP}}}}}}}_{{{{{{\boldsymbol{q}}}}}}}))\)};
This method guarantees a uniform null for any arbitrarily defined statistics as long as the true null can be obtained via simulations. Therefore, the final minCP from CGWAS is directly comparable to the pvalues from a standard GWAS where the conventional 5 × 10^{8} significance threshold is applied. The getCoef is generic and we also use it for correcting the minimal pvalues from MinGWAS. The empirical calibration is sufficiently accurate when NS is large. We suggest use at least 100 times of independent SNP number for NS. In CGWAS, the default of NS is set to 10^{8}.
Estimation of Inflation, getI(G)
Accurate estimation of inflation is a prerequisite for accurate estimation of π and h. We propose a fast and accurate method for estimating inflation. Compared with the genomic control method, our method is less likely affected by the presence of true effects, thus less conservative. Compared with the LD score method described by BulikSullivan et al.^{22}, our method has the same level of estimate error with higher computation efficiency because it does not require LD information, as detailed below.
Following the polygenicity model described in BulikSullivan et al.^{22}, we modeled the test statistics T from a GWAS using a mixed distribution of three normal distributions, including the estimated error N(0, e), inflation induced by genetic drift assumed to be normal distributed N(0, I), and true effect assumed to be normal distributed N(0, Λ). We additionally assume that p is the proportion of the SNPs (including those in LD) having true effects. For a SNP m, where m\(\in\){1, 2,…, M}, we have
The moments of T can be expressed by I, Λ and p,
where T is known for individual GWASs, and I, Λ, p are unknown. Our expression of \({{{\rm{E}}}}({T}_{m}^{2})\) here corresponds to the expression of \(N{h}^{2}/M+{Na}+1\) as described in BulikSullivan et al.^{22}. Note that the inflation term Na is simplified as I in (10) and true effect term \(N{h}^{2}/M\) is simplified as pΛ in (10). Next, I (or inflation factor s=I+1) can be determined using an iterative search algorithm with the moments of T and a series of quantiles of T \(\circ\) T (Supplementary Method E). All GWAS can be preadjusted using GWASspecific s (\({{{{{{\bf{T}}}}}}}_{{{{{{\bf{adj}}}}}}}{{=}}{{{{{\bf{T}}}}}}{{/}}\sqrt{s}\)) before CGWAS analysis.
Estimation of ψ, getPsi(G)
Accurate estimation of ψ is a prerequisite for combining correlated traits. We propose a fast and accurate method to estimate ψ. Compared with the methods which directly compute correlations using GWAS statistics, such as Zhu et al.^{2} and Bolormaa et al.^{1}, our method is less likely affected by the true effect. Compared with MTAG, which estimate covariance of the estimate error using LDSC^{34}, our method has a similar level of estimate error but does not require LD information, thus more computationally efficient, especially for the analysis of a large number of GWAS.
The idea of our method is similar to that of Park et al.^{7}, which select a set of marginally insignificant SNPs from two GWAS. Our approach selects a set of insignificant SNPs from the joint distribution of two GWAS, thus more tightly fitting the true null. This is achieved using an iterative optimization algorithm. Consider T statistics of two GWAS for genomewide SNPs T_{1} and T_{2}, let ψ_{0} be the correlation between T_{1} and T_{2} as \({\psi }_{0}={{{{{\rm{cor}}}}}}({{{{{{\bf{T}}}}}}}_{{{{{{\bf{1}}}}}}},\,{{{{{{\bf{T}}}}}}}_{{{{{{\bf{2}}}}}}})\), ψ_{1} = 0, and ε = 10^{4}. The process of the getPsi is described as below,
while ψ_{0}ψ_{1} > ε {
ψ_{0} = ψ_{1}; \({{{{{{\mathbf{\Psi }}}}}}}_{{{{{{\bf{0}}}}}}}=[(1,{\psi }_{0}),({\psi }_{0},1)]\);
for m in 1 to M {\({{CT}}_{m}=({T}_{m,1},\,{T}_{m,2}){{{{{{{\mathbf{\Psi }}}}}}}_{{{{{{\bf{0}}}}}}}}^{{{{{{\boldsymbol{}}}}}}{{{{{\bf{1}}}}}}}{({T}_{m,1},{T}_{m,2})}^{{{{{{\rm{T}}}}}}}\)};
id = which\(\left({{{{{{\rm{F}}}}}}}_{{\chi }_{2}^{2}}{{{{{\bf{CT}}}}}} \, > \, 0.5\right)\);
\({\psi }_{1}={{{{{\rm{cor}}}}}}({{{{{{\bf{T}}}}}}}_{{{{{{\bf{id}}}}}},{{{{{\bf{1}}}}}}},\,{{{{{{\bf{T}}}}}}}_{{{{{{\bf{id}}}}}},{{{{{\bf{2}}}}}}})\)};
ψ is estimated as ψ_{1}. In a CGWAS analysis, Ψ is derived only once and does not change. Therefore, Ψ is the same for all SNPs throughout the CGWAS analysis.
Estimation of π getPi(G,Ψ) and h getH(G,Ψ,Π)
Let B and Z both be a K by K covariance matrix of β and t among the genomewide SNPs, where \({{{{{{\rm{B}}}}}}}_{{ij}}={{{\rm{E}}}}({\beta }_{i}{\beta }_{j})\) and \({{{{{{\rm{Z}}}}}}}_{{ij}}={{{\rm{E}}}}({t}_{i}{t}_{j})\). H and C denote the covariance matrix of δ and β_{null} (β under the null) among the genomewide SNPs as described above. The basic idea of estimating π and h is similar to the estimation of covariance of true effect in MTAG, i.e., to estimate the expected overall difference (Π_{all} and h_{all}) between observed effects and the estimation error, i.e., Η = BC. For each SNP, we assume Ψ and Π is fixed but h can vary with different σ to keep a unified null distribution while accounting for potential heterogeneity of the statistical power of SNPs in different GWASs. Therefore,
These estimators fits best under polygenicity, i.e., all SNPs share the same Π. In deviation of polygenicity, e.g., major gene effect, or when dealing with lowpowered GWAS, the power may not be optimal. To overcome this limitation, we provide an alternative estimator (Π_{sig} and h_{sig}) which focus on the SNPs with more significant effects rather than genomewide SNPs,
Where Ψ_{nosig} is the covariance matrix of the nonsignificant SNPs among K GWASs. Here the ‘significant’ is defined as pvalue of Wald test using (9) below a predefined threshold L. These estimators is expected to perform better in the presence of major gene effect and when dealing with lowpowered GWAS.
Alternatively, to improve the robustness of estimation, when the estimation error of Η is large, setting Π to a square matrix of 1 (Π_{stb}) or assuming the same δ as usually done in the most of previous multitrait analysis method may provide stable or higher power. In this case, h_{sig} can be derived as \({\sum }_{k=1}^{K}({\sigma }_{k}\sqrt{{Z}_{{kk}}1}){{{{{{\bf{h}}}}}}}^{{{{{{\boldsymbol{*}}}}}}}\), where \({{{{{{\bf{h}}}}}}}^{{{{{{\boldsymbol{*}}}}}}}\,{{{{=}}}}\,({\sigma }_{1}\sqrt{{Z}_{11}1},\, {\sigma }_{2}\sqrt{{Z}_{22}1},\ldots,\, {\sigma }_{K}\sqrt{{Z}_{{KK}}1})\). Because the choice of Π and h does not influence α but has an effect on power, the selection of Π and h is dataspecific and can be achieved empirically using a datadriven approach (details see combining multiple GWASs section). Similar to Ψ, Π is the same for all SNPs. However, different from Ψ, Π is continuously updated in each iteration of iEbICoW so that different GWAS subsets may have different Π. For example, consider 5 GWASs with estimated Ψ and Π, the Ψ of the GWAS subset 1,2,3 used in the full process of CGWAS is Ψ_{[1~3, 1~3]} while the Π of this subset after updated in iEbICoW may be different from initial Π_{[1~3, 1~3]}.
Simulations
We conducted extensive simulations to investigate the performance of CGWAS under a variety of scenarios. For simulation scenarios 1–9, the core of our simulations is to simulate the test statistics of N independent SNPs on K GWAS (simT, a matrix with N rows and K columns) under the null by generating K null vectors of length N, which follow the multivariate standard normal distribution with a predefined background correlation matrix Ψ, so that for one simulated vector of test statistics simt, we have simt ~ N(0,Ψ). For a GWAS simT_{k}, a vector of test statistics which follow normal distribution with variance I is added to simulate inflation, so for a SNP simT_{nk} in GWAS simT_{k}, we have \({{simT}}_{{nk}}\, \sim {N}\left({{{{\mathrm{0,1}}}}}\right)+N\left(0,I\right)\). To add true effects, a proportion (p) of the N SNPs were set to have effects by adding a matrix of test statistics which follow multivariate normal distribution with covariance matrix Π∘θθ^{T}, where θ is a preset vector of standard deviation of each column of the added effect matrix, so that we have
The mean χ^{2} statistics (\({{{\rm{E}}}}\left({\chi}^{2}\right)\)) of simulated GWAS k can be derived as \(1+p{\theta }_{k}^{2}\). The ratio of variance of true effect δ can be set via multiplying preset ratio between elements of θ \(\circ\) θ and vectors of squared standard errors se ∘ se, i.e., \(\frac{{{{{{\rm{Var}}}}}}\left({\delta }_{i}\right)}{{{{{{\rm{Var}}}}}}\left({\delta }_{j}\right)}=\frac{{\theta }_{i}^{2}{{se}}_{i}^{2}}{{\theta }_{j}^{2}{{se}}_{j}^{2}}\). The parameter grids in different nine scenarios are specified in Supplementary Method F, and the core was replicated 1,000 times for every cell of the grid in 1–9 scenarios.
For scenario 10 assess the performance of CGWAS and MTAG in combining GWASs of overlapping samples, we used the real chromosome 1 genotype data (526,822 SNPs) of 10,000 participants randomly selected from the Rotterdam Study (total 14,926 participants). We simulated a normally distributed variable as the phenotype in such a way that 10% of the phenotypic variance is explained by 1% of all SNPs, whose true effects follow a normal distribution. We created three subdatasets according to different configurations of sample overlapping, (1) nonoverlapping with sizes of 2000, 3000, and 5000; (2) partially overlapping with sizes of 5000, 7000, and 8000; and (3) almost complete overlapping of 9950 each. We compared the results of CGWAS and MTAG in combining the three GWASs in subdatasets with the GWAS in all samples.
CGWAS specifications in real applications
The getI and getPsi functions use unlinked SNPs via multiple rounds of systematic sampling. For input GWASs with \({{{\rm{E}}}}\left({\chi }^{2}\right)\, < \,1.001\), CGWAS forces E(χ^{2}) = 1.001 to avoid failure of getH. To avoid abnormal values when solving large Ψ in the presence of strong collinearity, GWAS pairs with ψ^{2} > 0.5 (default) are forced combined using optimize function regardless to evaluate function.
CGWAS implementation and performance
CGWAS was implemented as a userfriendly, publicly accessible, operating system independent and parallel R package called “CGWAS”, https://github.com/FunGene/CGWAS. The idea of improving computational efficiency is to locate all independent parts from the whole serial calculations and parallelize them using R package “foreach” and “doParallel”. In general, all independent loops are parallelized. For the computational hotspot getPsi, since it is nested in the iterative decisionmaking structure of iEbICoW, we designed finegrained parallelization within each iteration for calculating ψ between multiple GWASs. For the hotspot getNtest of the TWT module, because NS simulations are independent, the parallelization is designed at a coarsegrained level. For controlling peak memory during parallelization, 1) we minimized the duplicated data between the subthreads and main thread, optimized the balance between the overall computational efficiency and the loads of subthreads based on a hierarchical load distribution; 2) reduced the number of slave threads for simple calculations but requesting large memory.
We assessed the performance of CGWAS and MTAG on a machine with 72 cores Intel Xeon CPU at 2.30GHz and 256 GB RAM, using simulated data (6 million SNPs; 5, 10, 20, 30, 40, and 80 GWASs; 1, 2, 4, 8, and 16 paralleled threads). Scenarios in each configuration were replicated testing for 3 times, and the mean performance was reported in Supplementary Table 2.
CGWAS analyses in 78 facial shape GWASs
We applied CGWAS to combine the summary statistics of 78 GWAS of facial traits, which were conducted in 10,115 individuals of European decent as described in Xiong at al^{4}. Prior to CGWAS, we dropped SNPs with sample sizes smaller than 60% of the total sample sizes from 7,029,494 to 5,308,962. This is because a reduced sample size in a subset of SNPs may increase the estimation error of Π and thus hamper the power of EbICoW. A simulation analysis imbedded in CGWAS was conducted to obtain the true null of our CGWAS application. In brief, the null summary statistics of 78 GWAS were simulated, each consisting of 1,000,000 unlinked SNPs without true effect, where the background correlations were simulated according to the Ψ estimated from the real GWAS data using the getPsi function. The same GWAS sets and corresponded w^{T} of each combination in iEbICoW of real GWAS data were used to combine simulation data. This step derived the same EbICoW GWAS with the real application but using simulated null. Next, CGWAS obtain the true null by applying TWT and distribution calibration as described above and finally adjust the application pvalues using the outcome of getCoef from simulation to ensure that CGWAS results are directly comparable to a standard GWAS. To calculate genomewide significant threshold for each iEbICoW combination in CGWAS final pvalues, we derive multiple testing and calibration burden estimated in getNtest and getCoef function as 27.13 and 2.22 respectively (Supplementary Method D), which results the threshold as \({5\times 10}^{8}\times 27.13\times 2.22={3\times 10}^{6}\). Under this threshold, CGWAS and MinGWAS results are provided in Supplementary Data 2. Regional lead SNPs (56) were selected based on the FUMA clumping algorithm^{35} with default parameter as detailed in Supplementary Data 3. Univariate LD score regression^{22} was applied to ztransformed GWAS pvalues, \({{{{{\bf{z}}}}}}=\sqrt{{{{{{{\rm{F}}}}}}}_{{\chi }_{1}^{2}}^{1}(1{{{{{\bf{p}}}}}})}\). The LD score intercepts of CGWAS and MinGWAS were estimated as 0.985 (0.007) and 0.998 (0.006) respectively. The gain of power for CGWAS was estimated according to the increase in the mean χ^{2} statistic method described in Turley et al.^{3}. In brief, the power ratio between two GWASs can be derived as \(\frac{{{{{{\rm{E}}}}}}({\chi }_{{GWAS}1}^{2}){{{{{{\rm{E}}}}}}({\chi }_{{null}}^{2})}_{{GWAS}1}}{{{{{{\rm{E}}}}}}({\chi }_{{GWAS}2}^{2}){{{{{{\rm{E}}}}}}({\chi }_{{null}}^{2})}_{{GWAS}2}}\). In our study, \({{{{{\rm{E}}}}}}\left({\chi }_{C{GWAS}}^{2}\right)=1.142\) and \({{{{{\rm{E}}}}}}\left({\chi }_{{MinGWAS}}^{2}\right)=1.1\), so that \(\frac{1.1420.985}{1.10.998}=1.54\).
Replication of CGWAS findings and previous GWAS findings
The RS is a populationbased cohort study of 14,926 participants aged 45 years and older, living in the same suburb of Rotterdam, the Netherlands^{36}. The present study includes 1,174 participants of Dutch European ancestry, for whom highresolution 3dMDface digital photographs were taken. Note that these samples have not been used in the previous GWAS^{4}. Genotyping was carried out using the Infinium II HumanHap 550K Genotyping BeadChip version 3 (Illumina, San Diego, California USA). Microarraybased genotyping according to the manufacturer’s instructions was performed at Erasmus MC^{37}. All SNPs were imputed using MACH software (www.sph.umich.edu/csg/abecasis/MaCH/) based on the 1000Genomes Project reference population information^{38}. After all quality controls, the current study included a total of 6,886,439 autosomal SNPs (MAF > 0.01, imputation R2 > 0.8, SNP call rate > 0.97, HWE > 0.0001).
The Rotterdam Study has been approved by the Medical Ethics Committee of the Erasmus MC (registration number MEC 02.1015) and by the Dutch Ministry of Health, Welfare and Sport (Population Screening Act WBO, license number 1071272–159521 PG). The Rotterdam Study has been entered into the Netherlands National Trial Register (NTR; www.trialregister.nl) and into the WHO International Clinical Trials Registry Platform (ICTRP; www.who.int/ictrp/network/primary/en/) and under shared catalog number NTR6831. All participants provided written informed consent to participate in the study and to have their information obtained from treating physicians.
The raw 3D facial images of participates were acquired using a 3D photographic scanning system manufactured by 3dMD (http://www.3dmd.com/). Participants were asked to keep their mouths closed and adopt a neutral expression during the acquisition of the 3D scans. Software package MeshMonk^{39} was then used to derive the 13 facial landmarks. The description of these 13 landmarks are in Supplementary Table 3. Association tests were conducted between the 78 3D facial distance measurements (after scaled GPA) and 6,886,439 SNPs. GWAS was performed using linear regression under the additive genetic model while adjusting for sex, age, and the first four genomic PCs. CGWAS analyses was then conduct use default parameters.
In addition, we reached out and obtained the GWAS summary statistics^{5} from 2,774 individuals of European descent from the Pennsylvania State University (PSU) cohort (N = 1990) and the Indiana UniversityPurdue University Indianapolis (IUPUI) cohort (N = 784), which have not been used in our previous study^{4} and are thus not included in our CGWAS discovery dataset used here. The data contains 63 sets of GWAS summary statistics corresponding to the 63 facial segments. In addition, we downloaded the summary statistics from a recent facial variation GWAS of Chinese population^{6}, which included 9674 Chinese individuals from 3 cohorts, i.e., the National Survey of Physical Traits (NSPT), the Northern Han Chinese (NHC) and the Taizhou Longitudinal Study (TZL). The data also contains 63 sets of GWAS summary statistics for 63 facial segments. We then carried out our replication analysis for the 56 SNPs in RS, PSU+IUPUI, and NSPT+NHC+TZL.
We select 56 lead SNPs CGWAS pvalue in RS CGWAS result as replicated evidence from RS. We use Bonferroni and Fisher’s combined test to combine the evidence of replication and account for the trait differences between our landmarksbased study and the previous facial segmentbased studies. In brief, we first adjust the minimal pvalue of the 63 segmentbased pvalues using Bonferroni correction of the numbers of independent traits as reported in corresponding studies. We then use Fisher’s combined test to combine the RS pvalue derived from CGWAS, the adjusted minimal pvalue from PSU+IUPUI, and the adjusted minimal pvalue from NSPT+NHC+TZL. In short, the test statistic of the Fisher’s combined test T_{F} of N independent tests with pvalue p_{n} follows a chisquared distribution of 2N degrees of freedom as \({T}_{F}=2{\sum }_{n=1}^{N}\log \left({p}_{n}\right)\, \sim \,{{{{{{\rm{\chi }}}}}}}_{2N}^{2}\). Replication results are detailed in Supplementary Data 3. For 56 independent observations, the pvalue of uniform null distribution at their expected quantile were generated from beta distribution. For 327 SNPs that have been associated with facial variation in previous GWASs and also available in our GWAS data, we looked up their pvalues from our CGWAS and MinGWAS analysis of 10,115 individuals.
Comparison with other multitrait analysis methods using facial data
CGWAS and MTAG have an important similarity in distinguishing the effect correlation Π (true effect Ω in MTAG) from the background correlation Ψ (estimate error Σ in MTAG) with the advantage in better estimating true null, whereas other methods do not. In addition, when Π and Ψ are different, both methods provide additional statistical power, while the other methods do not. It is therefore desirable to compare the statistical performance of the two methods. CGWAS and MTAG analyses were carried out using GWAS summary statistics generated previously, which consisted of 10,115 individuals of European decent. Because MTAG cannot complete the analysis of 78 facial GWASs within our computational capacity, we focused on a subset of 30 GWAS covering the left side of the face. Because MTAG generated 30 GWAS outcomes (one outcome per trait), we selected the minimal pvalues and adjusted them using the getCoef function. CGWAS setting was the same with those used in all 78 GWAS application. To calculate genomewide significant threshold for each iEbICoW combination in CGWAS final pvalues, we derive multiple testing and calibration burden estimated in getNtest and getCoef function as 12.44 and 2.17 respectively (Supplementary Method D), which results the threshold as \(5\times 10^{8}\times 12.44\times 2.17={1.35\times 10}^{6}\). Under this threshold, CGWAS and MTAG results are provided in Supplementary Data 4. We additionally compared CGWAS with four competing statistics, including mixAda, mixFisher and mixTippett proposed by Liu et al. using R package ‘MPAT’^{23} and S_{Het} proposed by Zhu et al. using R script ‘CPASSOC’^{2}. We also looked up the 327 previously established SNPs in CGWAS and results from all competing methods described above. Note that due to SNP filtering of MTAG, only 271 SNPs retained for comparison.
Polygenetic score and facial variance explained
Multivariable and polygenic score analyses were conducted in the same 1,174 participates from RS which described above. 57 lead SNPs in CGWAS, 57 and 17 lead SNPs in MinGWAS were involved in analyses below. We calculated PRS using the selected n SNPs and the effect sizes estimated from the GWAS summary of 10,115 individuals, that is \({\sum }_{i=1}^{n}{\beta }_{i}{A}_{i}\). Then proportion of phenotype explained for 78 facial traits were assessed by r^{2} between MinGWAS based and CGWAS based PRS and sex, age adjusted face phenotypes. 95% confidence intervals of r^{2} were obtained from 1000 times randomly sampling with replacement. PRS z score test were then conducted by cooperating r^{2} value and standard deviation of sampling results. Next, we calculated PRS for AFR, EUR, EAS individuals of the 1000 Genomes Project using 57 lead SNPs in CGWAS, 17 lead SNPs in MinGWAS and effect sizes as used above.
Estimation of multitrait effects
We developed a statistic to quantify multitrait effect (MTE) for the positive findings from CGWAS. The idea is to use the skewness of the independent phenotypic effects of each SNP to represent its multitrait effect, i.e., the more independent phenotypes a SNP is associated with (less skew), the larger its multitrait effect is. The skewness statistic of a vector X can be estimated using the skew function, \({{{{{\rm{Skew}}}}}}\left({{{{{\bf{X}}}}}}\right)={{{\rm{E}}}}{\left({{{{{\bf{X}}}}}}{{{\rm{E}}}}\left({{{{{\bf{X}}}}}}\right)\right)}^{3}/{({{{\rm{E}}}}{\left({{{{{\bf{X}}}}}}{{{\rm{E}}}}\left({{{{{\bf{X}}}}}}\right)\right)}^{2})}^{3/2}\), as proposed previously^{40}. However, because the phenotypes are correlated due to background correlations, steps to remove background correlations are necessary. To achieve this, we firstly construct an uncorrelated null space of eigen vectors set V on Ψ, so that all vectors in this space have no background correlations under the null. In our application, we derived top 26 eigen vectors with decreased eigen value from 78 eigen vectors of Ψ to form V for eliminating the noise from the eigen vectors with small eigen value. Number of derived eigen vectors was ascertained by keeping the accumulated eigen value over 95% of total 78. Next, for each significant SNP, its phenotypic effects \({{{{{\bf{t}}}}}}\,{{{{=}}}}\,{({t}_{1},\, {t}_{2},\ldots,\, {t}_{78})}^{{{{{{\rm{T}}}}}}}\) is projected to the null space as \({{{{{\bf{pt}}}}}}\,{{{{=}}}}\,{({{pt}}_{1},\, {{pt}}_{2},\ldots,\, {{pt}}_{26})}^{{{{{{\rm{T}}}}}}}\) and then is scaled using the eigen value g_{(1~26)} to ensure that they are comparable (same variance) under the null, i.e., \({{pt}}_{i}={{{{{{\bf{t}}}}}}}^{{{{{{\bf{T}}}}}}}{{{{{{\bf{V}}}}}}}_{{{{{{\boldsymbol{i}}}}}}}/\sqrt{{g}_{i}}\), where i\(\in\){1, 2, …,26}. Finally, we calculate the skewness of squared projected effects (eliminate the impact of sign, focus on absolute effect) for each significant SNP, then centralize these skewness values to that under the null. Noted that the squared projected effects under the null follow the chisquared distribution with one degree of freedom. Therefore, the skewness value under the null can be directly calculated as \({{{\rm{E}}}}{\left({\chi }_{1}^{2}{{{\rm{E}}}}\left({\chi }_{1}^{2}\right)\right)}^{3}/{({{{\rm{E}}}}{\left({\chi }_{1}^{2}{{{\rm{E}}}}\left({\chi }_{1}^{2}\right)\right)}^{2})}^{3/2}= ({{{\rm{E}}}}{({\chi }_{1}^{2})}^{3}3{{{\rm{E}}}}({\chi }_{1}^{2}){{{{{\rm{Var}}}}}}({\chi }_{1}^{2}){({{{\rm{E}}}}({\chi }_{1}^{2}))}^{3})/{{{{{{\rm{Var}}}}}}({\chi }_{1}^{2})}^{3/2}=(1561)/2 \sqrt{2}= 2\sqrt{2}\). The inverse of the centralized skewness of each SNP is considered as a quantification of its multitrait effect, MTE=2√2Skew(pt∘pt). MTE < 0 indicates insufficient evidence supporting the presence of multitrait effect. MTE values for each CGWAS lead SNP are in Supplementary Fig. 15.
Gene ontology analysis
To reveal and compare the potential functional roles of the CGWAS and MinGWAS findings, we investigate the enrichment of nearby genes from MinGWAS and CGWAS loci in three otology databases, including Gene Ontology (GO), Human Phenotype Ontology (HPO) and Mouse Phenotype Ontology (MPO). The gene sets used for CGWAS (total 83 genes from 56 loci) and MinGWAS (total 24 genes from 17 loci) represented the unions of three nonexclusive sets, including the closest genes to regional lead SNPs (57 for CGWA, 17 for MinGWAS), annotated genes from ‘SNP2GENE’ function in FUMA^{35} (64 for CGWAS, 12 for MinGWAS), and cisregulated genes proposed by GREAT^{41} based on lead SNPs (70 for CGWAS, 22 for MinGWAS). The enrichment analysis was performed using the R package’clusterProfiler’^{42} in GO and using GREAT in HPO and MPO. All Bonferroni adjusted significant (p<0.05) enriched terms are in Supplementary Data 5. The resultant GO terms were classified into 4 groups according to functionality, ‘morphogenesis’, ‘development’, ‘differentiation’, and ‘regulation’, whereas all the resultant HP and MP terms could be classified into one group ‘abnormality’.
Colocalization analysis
We performed Bayesian colocalization analysis between CGWAS loci and eQTL dataset from GTEx^{43} including 22 tissues from brain, skin, adipose, muscle skeletal and glands using R package ‘coloc’^{44}. This method evaluates whether a shared causal variant is responsible in both GWAS and gene expression in a genomic region of interest. Each CGWAS locus was defined as a region with upstream and downstream 500kb from the regional lead SNP. The prior probabilities are set by default. A posterior probability was obtained for every gene around a region and a high posterior probability (PP4>0.7) suggested strong evidence of colocalization. Colocalizations in all 22 tissues of the loci with high posteriori probability (PP4>0.7) in at least one tissues are in Supplementary Data 6. Colocalization plots were generated using R package ‘locuscomparer’^{45}.
Analysis of CNCC regulation network
The regulatory network of CNCC was downloaded at https://github.com/AMSSwanglab/hRegCNCC. We used ISTAT^{28} to compute the enrichment of detected loci in the regulatory element of CNCC. The loci associated regulatory was obtained according to their intersection with regulatory elements proposed in regulatory network of CNCC^{27}. For every locus, we checked every TFRETG triplet and extracted it if the RE was intersected with this locus. Extracted triplet based on CGWAS and MinGWAS loci are in Supplementary Data 7. All loci associated TFRETG triplets were pooled to form the loci associated regulatory subnetwork of CNCC.
Reporting summary
Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.
Data availability
CGWAS pvalues of the studywide suggestively significant SNPs are provided in Supplementary Data 2. Full CGWAS summary statistics are publicly available via figshare at https://doi.org/10.6084/m9.figshare.21559086. Full GWAS summary statistics of the 78 facial traits used in discovery phase of CGWAS application are available on GWAS Catalog with the access number 31763980. The summary statistics of the Pennsylvania State University (PSU) and the Indiana UniversityPurdue University Indianapolis (IUPUI) are available on GWAS Catalog with the access number 33288918. The summary statistics of the National Survey of Physical Traits (NSPT), the Northern Han Chinese (NHC) and the Taizhou Longitudinal Study (TZL) are available on the National Omics Data Encyclopedia with NODE number OEP002283. The ciseQTL results in 22 tissues were downloaded from GTEx V7 database at https://www.gtexportal.org/home/datasets. The regulatory network of CNCC is publicly available at https://github.com/AMSSwanglab/hRegCNCC.
Code availability
CGWAS is implemented as an opensource R package available at https://github.com/FunGene/CGWAS.
References
Bolormaa, S. et al. A multitrait, metaanalysis for detecting pleiotropic polymorphisms for stature, fatness and reproduction in beef cattle. PLoS Genet. 10, e1004198 (2014).
Zhu, X. et al. Metaanalysis of correlated traits via summary statistics from GWASs with an application in hypertension. Am. J. Hum. Genet. 96, 21–36 (2015).
Turley, P. et al. Multitrait analysis of genomewide association summary statistics using MTAG. Nat. Genet. 50, 229–237 (2018).
Xiong, Z. et al. Novel genetic loci affecting facial shape variation in humans. Elife 8, e49898 (2019).
White, J. D. et al. Insights into the genetic architecture of the human face. Nat. Genet. 53, 45–53 (2021).
Zhang, M. et al. Genetic variants underlying differences in facial morphology in East Asian and European populations. Nat. Genet. 54, 403–411 (2022).
Park, H., Li, X., Song, Y. E., He, K. Y. & Zhu, X. Multivariate analysis of anthropometric traits using summary statistics of genomewide association studies from GIANT Consortium. PLoS ONE 11, e0163912 (2016).
Li, J. & Ji, L. Adjusting multiple testing in multilocus analyses using the eigenvalues of a correlation matrix. Heredity (Edinb) 95, 221–227 (2005).
Liu, F. et al. A genomewide association study identifies five loci influencing facial morphology in Europeans. PLoS Genet. 8, e1002932 (2012).
Paternoster, L. et al. Genomewide association study of threedimensional facial morphology identifies a variant in PAX3 associated with nasion position. Am. J. Hum. Genet. 90, 478–485 (2012).
Adhikari, K. et al. A genomewide association scan implicates DCHS2, RUNX2, GLI3, PAX1 and EDAR in human facial variation. Nat. Commun. 7, 11616 (2016).
Cole, J. B. et al. Genomewide Association Study of African Children identifies association of SCHIP1 and PDE8A with facial size and shape. PLoS Genet. 12, e1006174 (2016).
Pickrell, J. K. et al. Detection and interpretation of shared genetic influences on 42 human traits. Nat. Genet. 48, 709–717 (2016).
Shaffer, J. R. et al. Genomewide association study reveals multiple loci influencing normal human facial morphology. PLoS Genet. 12, e1006149 (2016).
Lee, M. K. et al. Genomewide association study of facial morphology reveals novel associations with FREM1 and PARK2. PLoS ONE 12, e0176566 (2017).
Cha, S. et al. Identification of five novel genetic loci related to facial morphology by genomewide association studies. BMC Genomics 19, 481 (2018).
Claes, P. et al. Genomewide mapping of globaltolocal genetic effects on human facial shape. Nat. Genet. 50, 414–423 (2018).
Crouch, D. J. M. et al. Genetics of the human face: Identification of largeeffect single gene variants. Proc. Natl Acad. Sci. USA 115, E676–E685 (2018).
Qiao, L. et al. Genomewide variants of Eurasian facial shape differentiation and a prospective model of DNA based face prediction. J. Genet. Genomics 45, 419–432 (2018).
Bonfante, B. et al. A GWAS in Latin Americans identifies novel face shape loci, implicating VPS13B and a Denisovan introgressed region in facial variation. Sci. Adv. 7, eabc6160 (2021).
Hoskens, H. et al. 3D facial phenotyping by biometric sibling matching used in contemporary genomic methodologies. PLoS Genet. 17, e1009528 (2021).
BulikSullivan, B. K. et al. LD Score regression distinguishes confounding from polygenicity in genomewide association studies. Nat. Genet. 47, 291–295 (2015).
Liu, Z. & Lin, X. Multiple phenotype association tests using summary statistics in genomewide association studies. Biometrics 74, 165–175 (2018).
Noback, M. L., Harvati, K. & Spoor, F. Climaterelated variation of the human nasal cavity. Am. J. Phys. Anthropol. 145, 599–614 (2011).
Guo, J. et al. Variation and signatures of selection on the human face. J. Hum. Evol. 75, 143–152 (2014).
Naqvi, S. et al. Shared heritability of human face and brain shape. Nat. Genet. 53, 830–839 (2021).
Feng, Z. et al. hRegCNCC reconstructs a regulatory network in human cranial neural crest cells and annotates variants in a developmental context. Commun. Biol. 4, 442 (2021).
Sarmashghi, S. & Bafna, V. Computing the statistical significance of overlap between genome annotations with iStat. Cell Syst. 8, 523–529 e4 (2019).
SimoesCosta, M. & Bronner, M. E. Establishing neural crest identity: a gene regulatory recipe. Development 142, 242–257 (2015).
Uz, E. et al. Disruption of ALX1 causes extreme microphthalmia and severe facial clefting: expanding the spectrum of autosomalrecessive ALXrelated frontonasal dysplasia. Am. J. Hum. Genet. 86, 789–796 (2010).
Liu, C. et al. Genome scans of facial features in East Africans and crosspopulation comparisons reveal novel associations. PLoS Genet. 17, e1009695 (2021).
Lee, J. J. et al. Gene discovery and polygenic prediction from a genomewide association study of educational attainment in 1.1 million individuals. Nat. Genet. 50, 1112–1121 (2018).
Christakoudi, S., Evangelou, E., Riboli, E. & Tsilidis, K. K. GWAS of allometric bodyshape indices in UK Biobank identifies loci suggesting associations with morphogenesis, organogenesis, adrenal cell renewal and cancer. Sci. Rep. 11, 10688 (2021).
BulikSullivan, B. et al. An atlas of genetic correlations across human diseases and traits. Nat. Genet. 47, 1236–1241 (2015).
Watanabe, K., Taskesen, E., van Bochoven, A. & Posthuma, D. Functional mapping and annotation of genetic associations with FUMA. Nat. Commun. 8, 1826 (2017).
Hofman, A. et al. The Rotterdam Study: 2014 objectives and design update. Eur J Epidemiol 28, 889–926 (2013).
Kayser, M. et al. Three genomewide association studies and a linkage analysis identify HERC2 as a human iris color gene. Am. J. Hum. Genet. 82, 411–423 (2008).
Genomes Project, C. et al. An integrated map of genetic variation from 1,092 human genomes. Nature 491, 56–65 (2012).
White, J. D. et al. MeshMonk: Opensource largescale intensive 3D phenotyping. Sci. Rep. 9, 6085 (2019).
Joanes, D. N. & Gill, C. A. Comparing measures of sample skewness and kurtosis. J. R. Statistical Soc. Ser. D (The Statistician) 47, 183–189 (1998).
McLean, C. Y. et al. GREAT improves functional interpretation of cisregulatory regions. Nat. Biotechnol. 28, 495–501 (2010).
Yu, G., Wang, L. G., Han, Y. & He, Q. Y. clusterProfiler: an R package for comparing biological themes among gene clusters. OMICS 16, 284–287 (2012).
Consortium, G. T. The GTEx Consortium atlas of genetic regulatory effects across human tissues. Science 369, 1318–1330 (2020).
Giambartolomei, C. et al. Bayesian test for colocalisation between pairs of genetic association studies using summary statistics. PLoS Genet. 10, e1004383 (2014).
Liu, B., Gloudemans, M. J., Rao, A. S., Ingelsson, E. & Montgomery, S. B. Abundant associations with gene expression complicate GWAS followup. Nat. Genet. 51, 768–769 (2019).
Acknowledgements
The authors are grateful for the dedication, commitment and contribution of the study participants, the general practitioners, pharmacists, and the staff from the Rotterdam Study. The Rotterdam Study is supported by the Erasmus MC; the Erasmus University Rotterdam; the Netherlands Organization for Scientific Research (NWO); the Netherlands Organization for Health Research and Development (ZonMw); the Research Institute for Diseases in the Elderly (RIDE); the Netherlands Genomics Initiative (NGI); the Ministry of Education, Culture and Science; the Ministry of Health Welfare and Sport; the European Commission (DG XII); and the Municipality of Rotterdam. The generation and management of GWAS genotype data for the Rotterdam Study were executed by the Human Genotyping Facility of the Genetic Laboratory of the Department of Internal Medicine, Erasmus MC. We thank Susan Walsh and Mark Shriver for sharing with us for replication purposes in this study their published data^{5} from the Indiana UniversityPurdue University Indianapolis (IUPUI) and Pennsylvania State University (PSU) cohorts. The authors of Zhang et al.^{6} for making their GWAS summary statistics publicly available allowing us to use them here for replication purposes. FL was supported by the Strategic Priority Research Program of Chinese Academy of Sciences (Grant Nos. XDB38010400, XDC01000000), Shanghai Municipal Science and Technology Major Project (Grant No. 2017SHZDZX01), National Natural Science Foundation of China (NSFC) (81930056), Science and Technology Service Network Initiative of Chinese Academy of Sciences (KFJSTSQYZD202108001, KFJSTSZDTP079).
Author information
Authors and Affiliations
Contributions
Z.X., M.K., and F.L. designed and drafted the work. Z.X., X.G., and Y.C. developed mathematical formulations, performed data analyses, developed software package, visualized, and interpreted results. Z.F., S.P., and H.L. contributed to data analysis. A.G.U., T.N., A.I., F.R., M.G., Y.W., M.K., and F.L. provided the data resource. Y.W. critically revised all mathematical formulas. All authors approved the submission of the final version of the article for publication.
Corresponding authors
Ethics declarations
Competing interests
The authors declare no competing interests
Peer review
Peer review information
Nature Communications thanks the anonymous reviewers for their contribution to the peer review of this work. Peer reviewer reports are available.
Additional information
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons license, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons license and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this license, visit http://creativecommons.org/licenses/by/4.0/.
About this article
Cite this article
Xiong, Z., Gao, X., Chen, Y. et al. Combining genomewide association studies highlight novel loci involved in human facial variation. Nat Commun 13, 7832 (2022). https://doi.org/10.1038/s41467022353289
Received:
Accepted:
Published:
DOI: https://doi.org/10.1038/s41467022353289
This article is cited by

Revisiting informed consent in forensic genomics in light of current technologies and the times
International Journal of Legal Medicine (2023)
Comments
By submitting a comment you agree to abide by our Terms and Community Guidelines. If you find something abusive or that does not comply with our terms or guidelines please flag it as inappropriate.