Genome-wide association studies (GWAS) have broadened our knowledge on architectures of disease susceptible loci for many common disease of public health importance. A general approach for GWAS follows a strategy to investigate the correlations between single genetic variants and single traits within a univariate framework. The GWAS have not considered complicated genetic nature such as pleiotropy that occurs due to potential genetic correlation between different traits [1
]. Thus, it tends to be restricted to identify pleiotropic genes that situated at common etiologic pathways of correlated human diseases.
Patterns of pleiotropic effects have been observed more with an increasing number of variants identified through GWAS [3
]. For instances, Winkler and colleagues identified a variant of TCF2 (Transcription Factor 2) associated with T2D [4
], while a different variant in the same gene was associated with an increased risk of prostate cancer [5
]. These two studies indicate that the risk allele for prostate cancer protects from T2D with an odds ratio of 0.91. In addition, two studies [3
] showed that the same variant in GDF5 associated with greater height also was associated with reduced risk of osteoarthritis [6
As we have mentioned the examples above, previous work has shown that ignoring pleiotropic effects may cause imprecise phenotype definition of heterogeneous samples or even spurious associations. A bias in sampling cases and controls characterizing single traits might be propagated since the sampling errors tend to be correlated if the single traits were correlated. This may confound the interpretation of results. Although any loss of power occurred by selection of samples can be recovered by increasing the sample size, the sample size of the GWAS has cost constraints. With large sample sizes of several thousand cases and controls there has been usually limited study power to detect alleles of modest effect sizes (e.g., an odds ratio of 1.20) [8
In this regard, incorporation of the multiple phenotypes to the GWAS can be an alternative way to unravel missing heritability in the GWAS and to find pleiotropic genes. Even though the GWAS of multivariate phenotypes are known to enhance the power of the GWAS such an approach has not been well established.
To perform multiple phenotypes based GWAS, application of traditional GWAS approaches has suffered penalties from multiple testing problems caused by testing multiple genome-wide scans of single traits separately. This may diminish the power of GWAS due to elevating heterogeneity and bias in samples. Statistically classic multivariate methods have been applied to GWAS of multivariate phenotypes to tackle in an effective manner. Such methods are likelihood-based mixed effects model (LME) [9
] and generalized estimating equations (GEE) methods. Liu et al. suggested an extension of the GEE to test association analysis for a mixture of continuous and binary traits [11
]. Their work manifested statistical power of bivariate association analysis with two continuous traits, i.e. obesity and osteoporosis. Their method is limited to bivariate traits and applicable to independent samples.
O’Brien model [12
] and its extension [13
], which suggested the integration of results from association tests of single traits of a multivariate phenotype, can work well for a homogeneous mean among individual tests of single traits but not for heterogeneous ones. To overcome this limitation, Yang et al. [2
] improved O’Brien method by use of a sample splitting method and a cross validation method as a screening tool for detecting pleiotropic effects. Previous work has contributed to addressing association tests for multivariate phenotypes. However, there is still no standard method to be free from multiple test problems and be accepted for multivariate phenotypes [11
Much work have not investigated what types of single traits can be correlated to induce multivariate phenotypes. In this context, we aimed to discover novel multivariate phenotypes from large scale epidemiological data by a data mining approach and develop a scheme to GWAS of multivariate phenotypes. In our previous work [14
], we reported the discovery of multivariate phenotypes by applying association rule mining over 52 anthropometric and biochemical traits in Korea Association Resource (KARE)[15
] population. We showed an analytical scheme for GWAS of the multivariate phenotype lowLDLhighTG
, which means a negative relation between low levels of LDL and high levels of TG. Our preliminary results revealed that effect sizes (odds ratios=1.44-2.38) of genetic loci associated with the multivariate phenotype were higher than genetic loci identified in the initial GWAS, while their p-values were less significant than those in the initial GWAS. Those loci cannot be detected within a single trait based framework.
Here, we present a more sophisticated scheme for refining association rules to extract patterns of phenotypic associations and to visualize them graphically. As a case study, we describe the results of GWAS for multivariate phenotype highLDLhighTG combining elevated low density lipoprotein cholesterol (LDL-C) levels and elevated triglyceride (TG) levels, which have an important clinical implication in metabolic syndrome (MS).
An association rule which expresses patterns of multivariate phenotypes encoding partial correlations between single traits specifies quantitative descriptions of the single traits. Association rules can provide explicit boundaries of the single traits of multivariate phenotypes for optimal selection of both cases and controls. This work contributes a methodology for exploration in GWAS analysis of multiple phenotype highLDLhighTG.