Abstract
Background
Currently, most methods for detecting genegene interaction (GGI) in genomewide association studies (GWASs) are limited in their use of single nucleotide polymorphism (SNP) as the unit of association. One way to address this drawback is to consider higher level units such as genes or regions in the analysis. Earlier we proposed a statistic based on canonical correlations (CCU) as a genebased method for detecting genegene coassociation. However, it can only capture linear relationship and not nonlinear correlation between genes. We therefore proposed a counterpart (KCCU) based on kernel canonical correlation analysis (KCCA).
Results
Through simulation the KCCU statistic was shown to be a valid test and more powerful than CCU statistic with respect to sample size and interaction odds ratio. Analysis of data from regions involving three genes on rheumatoid arthritis (RA) from Genetic Analysis Workshop 16 (GAW16) indicated that only KCCU statistic was able to identify interactions reported earlier.
Conclusions
KCCU statistic is a valid and powerful genebased method for detecting genegene coassociation.
Keywords:
Genomewide association study (GWAS); Genegene coassociation; Genegene interaction (GGI); Kernel canonical correlation analysis (KCCA)Background
Genomewide association studies (GWASs), which may involve a large number of single nucleotide polymorphisms (SNPs) on many individuals, are widely used to identify genetic variants underlying complex diseases or other types of traits. Although a primary interest in a GWAS is to identify SNPs associated with a trait of interest, it is important to consider the associate genes and their coassociation as well. One form of coassociation is epistasis, which was introduced approximately 100 years ago and generally defined as interactions among genes [1]. These are linked to genegene interactions (GGIs) which are often characterized to be functional, compositional and statistical [2]. The statistical definition was given by Fisher [3] and developed further by Cockerham [4] and Kempthorne [5], whereby the effect of GGIs is treated as deviation from additive genetic effects of single genes [6].
Methods to detect GGIs on the basis of the statistical definition include but are unlimited to logistic regression, multifactor dimensionality reduction [7], linkage disequilibrium (LD)based [8,9] and entropybased statistics [10,11], together with others implemented in PLINK[12], Tuning ReliefF [13], Random Jungle[14], BEAM[15] and BOOST[16]. However, most of these consider SNP as the unit of association, which has limitations and are insufficient for interpretation of GGI [17] which calls for consideration of higher level units such as genes or regions in the analysis. Genebased analysis can account for multiple independent functional variants within genes with a potential increase of power to identify GGI. Earlier, Peng et al. [17] proposed a genebased statistic (CCU statistic) for detecting genegene coassociation based on canonical correlation analysis (CCA) in a case–control study, which was defined as joint effect of genes contributing to a binary trait and proved to have good performance on detecting genegene coassociation or GGI. However, CCA can only detect linear correlation, and may be inappropriate for genomic data containing nonlinear structure. Recent years have witnessed considerable work and successes on kernel CCA (KCCA) as a nonlinear generalization of the classical CCA in machine learning, face recognition, data classification [1820], and notably genomic data analysis by Yamanishi et al. [19]. We here construct a kernel CCU (KCCU) statistic for detecting genegene coassociation and evaluate its performance via simulations and data analysis.
Methods
CCA
CCA is a classical multivariate method which concerns about linear dependencies between
sets of variables. Let X_{i}, Y_{i} (i = 1, . . ., m) denote samples of measurements on m objects, We assume the data
to be column centred. Let A be any m × n matrix then L(A) = {Aαα ∈ R^{n}} will be referred to as the columnspace and L(A^{T}) = {A^{T}αα ∈ R^{m}} the rowspace of A. The aim of canonical correlation analysis is to determine vectors ν_{j} ∈ L(X^{T}) and ω_{j} ∈ L(Y^{T}) such that a_{j} = Xν_{j} and b_{j} = Yω_{j} are maximally correlated.
KCCA
KCCA generalizes CCA as follows: Objects x_{i} and y_{i} are first mapped to some Hilbert spaces H_{x} and H_{y} through mapping Φ_{x}(.) and Φ_{y}(.), CCA is then performed on images {Φ_{x}(x_{i})}_{i = 1}^{m} and {Φ_{y}(y_{i})}_{i = 1}^{m}. Let K_{x} and K_{y} denote m × m kernel inner product matrices (also known as kernel gram matrices), constructed elementwise
as
Explicit form for the mapping Φ_{x}(.) and Φ_{y}(.) are not always required but the kernel K_{x} and K_{y} need to be fixed. Common kernel functions include linear, polynomial, radial basis function (RBF), sigmoid [21], identicalbystate and weighted identicalbystate kernels [22]. It is worthwhile to note that these kernel functions generally have similar performance with parameters that are appropriately chosen.
Test statistic
Strategy analogous to CCU statistic was used to construct the KCCU statistic except
that the maximum kernel canonical coefficient of the two genes, rather than the maximum
canonical coefficient, was taken as a measure of genegene coassociation in cases
and controls. Let genotyped data of case–control study be (X_{1}^{D}, X_{2}^{D}, …, X_{P}^{D}) and (Y_{1}^{D}, Y_{2}^{D}, …, Y_{q}^{D}) for gene A and gene B for cases, and (X_{1}^{C}, X_{2}^{C}, …, X_{P}^{C}) and (Y_{1}^{C}, Y_{2}^{C}, …, Y_{q}^{C}) for controls. The maximum kernel canonical coefficient κr_{D} between (X_{1}^{D}, X_{2}^{D}, …, X_{P}^{D}) and (Y_{1}^{D}, Y_{2}^{D}, …, Y_{q}^{D}) obtained through KCCA could be considered as a measurement of genebased gene–gene
coassociation in cases, and κr_{C} between (X_{1}^{C}, X_{2}^{C}, …, X_{P}^{C}) and (Y_{1}^{C}, Y_{2}^{C}, …, Y_{q}^{C}) be a measurement of gene–gene coassociation in controls. The transformation analogous
to Fisher’s simple correlation coefficient transformation was done to κr_{D} and κr_{C}, i.e.
The KCCU statistic for detecting statistical significance of the difference of genebased
genegene coassociation between cases and controls can be defined as
Data simulation
Simulation studies were conducted to assess the performance of KCCU relative to CCU under both the null (H_{0}) and alternative hypotheses (H_{1}), which were based on the HapMap data in the following steps:
Step 1. Phased haplotype (Phases 1 & 2 of CEU) data were downloaded from the HapMap web site (http://snp.cshl.org webcite) on two unlinked genome regions for generating the simulated genotypes. The GNPDA2 region is on Chr 4: 44401210.44410098 involving six SNPs while FAIM2 region is on Chr 12: 48571829.48583937 involving seven SNPs. Their LD patterns where shown in Figures 1 and 2 together with pairwise r^{2}.
Figure 1. Pairwise r^{2 }among the six SNPs in the first region. The six SNPs are rs16857402, rs2709, rs10020551, rs4484337, rs12643262, and rs7670601. The values to the right of the 6 dbSNP IDs (rs# IDs) are the corresponding minor allele frequencies.
Figure 2. Pairwise r^{2 }among the seven SNPs in the first region. The seven SNPs are rs17201502, rs905619, r637871, rs1027711, rs956864, rs640081, and rs706795. The values to the right of the seven dbSNP IDs (rs# IDs) are the their minor allele frequencies.
Step 2. Based on data above, large samples with 100, 000 cases and 100, 000 controls were generated using software gs2.0 [25] under a twolocus interaction multiplicative effects model (see Additional file 1), treating the 2^{nd} SNP of the first region and the SNP of the other as the causal variants and they were removed in the simulation to assess genegene coassociation. The interaction odds ratio was set as 1.0 under H_{0} and 1.1, 1.2, 1.3, 1.4, 1.5 under H_{1}. The SNPs in the regions were coded according to an additive genetic model. To further investigate the performance on causal SNPs with respect to minor allele frequency and LD, different SNP pairs from the two gene regions were defined as the casual variants.
Additional file 1. Twolocus interaction multiplicative effects model.
Format: DOCX Size: 1.3MB Download file
Step 3. From the remaining SNPs, simulated data were sampled and CCU and KCCU performed under various sample sizes N (N/2 cases and N/2 controls, N = 1000…5000) with R package kernlab (http://cran.rproject.org/web/packages/kernlab/index.html webcite). 500 simulations were repeated each with a significant level of 0.05.
Applications
The proposed KCCU statistic was applied to rheumatoid arthritis (RA) data from GAW16 Problem 1, consisting of 2,062 Illumina 550k SNP chips from 868 RA patients and 1,194 normal controls collected by the North American Rheumatoid Arthritis Consortium [26]. Three genes (C5, ITGAV, and VEGFA) on three different chromosomes were selected to detect genegene coassociation in this work, involving eight, eight and four SNPs, respectively. Logistic regression test and the CCU statistic were also used. For each pair of genes, the statistic which yielded the minimum p value was recorded from all pairs of SNPs one on each gene. The significance of the statistic was compared to its empirical distribution generated from 1,000 permutations by permuting case–control labels [27] which is relatively easy compared to the “BY” method [28] for multiple testing adjustment.
Results
Simulation
Shown in Table 1 are simulation results under H_{0}. The KCCU statistic is normally distributed according to the one sample KolmogorovSmirnov test with the type I error rates of KCCU statistic being close to given nominal value (α = 0.05) for different sample sizes. This indicates that the proposed statistic performs well under the null hypothesis.
Table 1. Performance of CCU and KCCU under the null hypothesis
Results on various interaction odds ratios and a sample size of 3,000 are shown in Figure 3, as with different sample sizes with an interaction odds ratio of 1.4 in Figure 4. It is clear that power of KCCU is a monotonically increasing function of sample size and interaction odds ratio. Figure 5 shows results with different SNP pairs defined as causal SNPs with an interaction odds ratio of 1.3. The power of KCCU statistic was higher than that of CCU statistic. Power as a function of interaction odds ratio for different sample size is provided as Additional file 1.
Figure 3. Power of CCU and KCCU statistics given different interaction odds ratios and a sample size of 3,000.
Figure 4. Power of CCU and KCCU statistics given an interaction odds ratio of 1.4 and different sample sizes.
Figure 5. Power of CCU and KCCU statistics when SNP pairs from two regions are defined as casual variants at an interaction odds ratio of 1.3 and a sample size of 3,000.
Application
The performance of logistic regression test, CCU and KCCU statistics on pairwise genegene coassociation of three genes is shown in Table 2, which also contains results on the Gaussian RBF kernels with various parameter values (σ=0.05, 0.5, 5 and 50). Through KCCU the three genes were shown to have coassociation with each other at significance level 0.05 regardless the parameter value, in contrast to the CCU statistics showing no significant coassociation and none of the SNP pairs were significant under logistic regression test with correction for multiple testing.
Table 2. Pvalues of genegene coassociation among C5, ITGAV and VEGFA
Discussion
We have extended the CCU statistic to a new statistic KCCU, which can extract nonlinear correlation between two genes. Simulation studies show that both CCU and KCCU statistics performed well under null hypothesis with KCCU being more powerful than CCU with respect to significant level, sample size and relative risk. As results vary with userdefined kernel parameter, various parameters were used (the bandwidth parameter in RBF kernel) to RA data in GAW16 Problem 1, showing that the logistic regression test and CCU statistic failed to detect any interaction but KCCU statistics identified the pairwise interactions among the three genes under various parameters. The interaction between ITGAV and VEGF genes has been identified by a rank method [29]. As suggested by a reviewer, it is critical to consider timeefficiency in genomewide association studies to make the proposed methods practical. In our case, the computing time as required for KCCU was about 2.5 times slower than CCU, but nevertheless will still be feasible with the development as well as the extensive applications of multiprocessor and multithreading computational technique.
A reviewer has also suggested us to reiterate the relationship between genegene coassociation and GGI which is readily available. GGI generally refers to the synergetic or antagonistic effect of two genes in addition to the summation of their independent effects on an outcome. To represent the interaction between two genes A and B in a case–control association study, a product term is customarily added to the logistic regression model Logit(P) = β_{0} + β_{1}A + β_{2}B + γA × B so that γ reflects both the direction and size of the interaction. This model implicitly assumes that gene A and gene B are independent so as to infer interaction (γ). However, it might well be that genes are correlated with each other in genetic networks to contribute to disease susceptibility, so the independence assumption is rarely ratified. Genegene coassociation extends the concept of GGI in that it describes the generic joint distribution of two gene effects on disease or trait without assuming either independence or linear relationship. Here the measurement of the coassociation between genes is based on the correlation between genes (such as CCU statistic and KCCU statistic), provides a measure of the contribution of two genes. As for two unlinked genes, their relationship can be described as either coassociation or interaction. The reviewer has also brought to our attention to earlier work by Song and Nicolae [30] on imposing natural restrictions for the parameter space and discussion on the definition of “no interaction” between two unlinked loci as two loci being independent conditioned on the subject having the disease. In this paper, the null hypothesis of the proposed test is that there is no genegene coassociation (i.e. GGI for two unlinked genes), the data under the null hypothesis are generated from the gs software with the interaction odds ratio parameter to be one.
Several issues remain to be resolved: the uncertainty to set the kernel function with appropriate parameters for each data, the undesirable performance of both CCU and KCCU with small interaction odds ratio (e.g. 1.1), and the possible failure of maximum kernel canonical correlation coefficient to represent genegene coassociation.
Conclusions
KCCU statistic is a valid and powerful genebased method for detecting genegene coassociation compared to CCU and logistic regression test. Further work is needed to make its use in GWAS more practical.
Competing interests
The authors declare that they have no competing interests.
Authors’ contributions
ZSY, QSG, YGH, XSZ, FYL, JHZ and FZX conceptualized the study, acquired and analyzed the data and prepared for the manuscript. All authors approved the final manuscript.
Acknowledgements
This work was supported by the grant from National Natural Science Foundation of China (30871392) and Young Talents Innovation Foundation of School of Public Health, Shandong University. We thank GAW16 and the North American Rheumatoid Arthritis Consortium for the RA data and two anonymous reviewers for suggestions which led to substantial improvement and clarification of the paper.
References

Mendel's principles of heredity. 1909. 1902. [Mendel's principles of heredity] PubMed Abstract  Publisher Full Text

Phillips PC: Epistasisthe essential role of gene interactions in the structure and evolution of genetic systems.
Nat Rev Genet 2008, 9:855867. PubMed Abstract  Publisher Full Text  PubMed Central Full Text

Fisher RA: The correlation between relatives on the supposition of Mendelian inheritance.

Cockerham CC: An extension of the concept of partitioning hereditary variance for analysis of covariances among relatives when epistasis is present.
Genetics 1954, 39(6):859. PubMed Abstract  Publisher Full Text  PubMed Central Full Text

Kempthorne O: The correlation between relatives in a random mating population.
Proc R Soc Lond B Biol Sci 1954, 143(910):103. Publisher Full Text

Cordell HJ: Epistasis: what it means, what it doesn't mean, and statistical methods to detect it in humans.
Hum Mol Genet 2002, 11(20):2463. PubMed Abstract  Publisher Full Text

Ritchie MD, Hahn LW, Roodi N, Bailey LR, Dupont WD, Parl FF, Moore JH: Multifactordimensionality reduction reveals highorder interactions among estrogenmetabolism genes in sporadic breast cancer.
Am J Hum Genet 2001, 69(1):138147. PubMed Abstract  Publisher Full Text  PubMed Central Full Text

Wu X, Jin L, Xiong M: Composite measure of linkage disequilibrium for testing interaction between unlinked loci.
Eur J Hum Genet 2008, 16(5):644651. PubMed Abstract  Publisher Full Text

Zhao J, Jin L, Xiong M: Test for interaction between two unlinked loci.
Am J Hum Genet 2006, 79(5):831845. PubMed Abstract  Publisher Full Text  PubMed Central Full Text

Dong C, Chu X, Wang Y, Jin L, Shi T, Huang W, Li Y: Exploration of genegene interaction effects using entropybased methods.
Eur J Hum Genet 2007, 16(2):229235. PubMed Abstract  Publisher Full Text

Kang G, Yue W, Zhang J, Cui Y, Zuo Y, Zhang D: An entropybased approach for testing genetic epistasis underlying complex diseases.
J Theor Biol 2008, 250(2):362374. PubMed Abstract  Publisher Full Text

Purcell S, Neale B, ToddBrown K, Thomas L, Ferreira MA, Bender D, Maller J, Sklar P, de Bakker PI, Daly MJ, Sham PC: PLINK: a tool set for wholegenome association and populationbased linkage analyses.
Am J Hum Genet 2007, 81:559575. PubMed Abstract  Publisher Full Text  PubMed Central Full Text

Moore J, White B: Tuning reliefF for genomewide genetic analysis.
Lect Notes Comput Sci 2007, 4447:166175. Publisher Full Text

Schwarz D, Ko’ Nig I, Ziegler A: On safari to random jungle: A fast implementation of random forests for high dimensional data.
Bioinformatics 2010, 26:17521758. PubMed Abstract  Publisher Full Text  PubMed Central Full Text

Zhang Y, Liu JS: Bayesian inference of epistatic interactions in case–control studies.
Nat Genet 2007, 39:11671173. PubMed Abstract  Publisher Full Text

Wan X, Yang C, Yang Q, Xue H, Fan X, Tang N, Yu W: BOOST: A Fast Approach to Detecting GeneGene Interactions in Genomewide Case–control Studies.
Am J Hum Genet 2010, 87:325340. PubMed Abstract  Publisher Full Text  PubMed Central Full Text

Peng Q, Zhao J, Xue F: A genebased method for detecting genegene coassociation in a case–control association study.
Eur J Hum Genet 2009, 18(5):582587. PubMed Abstract  Publisher Full Text  PubMed Central Full Text

Zheng W, Zhou X, Zou C, Zhao L: Facial expression recognition using kernel canonical correlation analysis (KCCA).
Neural Networks IEEE Trans 2006, 17(1):233238. Publisher Full Text

Yamanishi Y, Vert JP, Nakaya A, Kanehisa M: Extraction of correlated gene clusters from multiple genomic data by generalized kernel canonical correlation analysis.
Bioinformatics 2003, 19(suppl 1):i323. PubMed Abstract  Publisher Full Text

Zepeda JAY, Davoine F, Charbit M:
Face tracking using canonical correlation analysis. 2007, 396402. [Visapp 2007: Proceedings of the Second International Conference on Computer Vision Theory and Applications, Volume Iu/Mtsv]

Liu Z, Chen D, Bensmail H: Gene expression data classification with Kernel principal component analysis.

Wu MC, Kraft P, Epstein MP, Taylor DM, Chanock SJ, Hunter DJ, Lin X: Powerful SNPset analysis for case–control genomewide association studies.
Am J Hum Genet 2010, 86(6):929942. PubMed Abstract  Publisher Full Text  PubMed Central Full Text

Nguyen VH, Golinval JC: Fault detection based on Kernel Principal Component Analysis.
Eng Struct 2010, 32(11):36833691. Publisher Full Text

Zhang DQ, Zhou ZH, Chen SC: Adaptive kernel principal component analysis with unsupervised learning of kernels.
Proceedings of the 6th IEEE International Conference on Data Mining (ICDM) 2006.

Li J, Chen Y: Generating samples for association studies based on HapMap data.
BMC Bioinforma 2008, 9(1):44. BioMed Central Full Text

Plenge RM, Seielstad M, Padyukov L, Lee AT, Remmers EF, Ding B, Liew A, Khalili H, Chandrasekaran A, Davies LRL, et al.: TRAF1C5 as a risk locus for rheumatoid arthritis  A genomewide study.
New Engl J Med 2007, 357(12):11991209. PubMed Abstract  Publisher Full Text  PubMed Central Full Text

Westfall PH, Young SS: Resamplingbased multiple testing. New York: Wiley; 1993.

Benjamini Y, Yekutieli D: The control of the false discovery rate in multiple testing under dependency.
Ann Stat 2001, 29:11651188. Publisher Full Text

Huang CH, Cong L, Xie J, Qiao B, Lo SH, Zheng T: Rheumatoid arthritisassociated genegene interaction network for rheumatoid arthritis candidate genes.
BMC Proc 2009, 3(Suppl 7):S75. PubMed Abstract  BioMed Central Full Text  PubMed Central Full Text

Song M, Nicolae D: Restricted parameter space models for testing genegene interaction.
Genet Epidemiol 2009, 33:386393. PubMed Abstract  Publisher Full Text