Volume 6 Supplement 1
Genetic Analysis Workshop 14: Microsatellite and singlenucleotide polymorphism
A Bayesian approach for applying HasemanElston methods
 Seungtai Yoon^{1}Email author,
 Young Ju Suh†^{2},
 Nancy Role Mendell†^{1} and
 Kenny Qian Ye†^{3}
DOI: 10.1186/147121566S1S39
© Yoon et al; licensee BioMed Central Ltd 2005
Published: 30 December 2005
Abstract
The main goal of this paper is to couple the HasemanElston method with a simple yet effective Bayesian factorscreening approach. This approach selects markers by considering a set of multigenic models that include epistasis effects. The markers are ranked based on their marginal posterior probability. A significant improvement over our previously proposed Bayesian variable selection methodology is a simple MetropolisHasting algorithm that requires minimum tuning on the prior settings. The algorithm, however, is also flexible enough for us to easily incorporate our hypotheses and avoid computational pitfalls. We apply our approach to the microsatellite data of Collaborative Studies on Genetics of Alcoholism using the coded values for the ALDX1 variable as our response.
Background
The HasemanElston method is an effective method for studying linkage between markers and diseases. Given the genotypes and phenotypes of a set of sibpairs, it regresses a statistic measuring similarity of quantitative traits in the sibling pair on the number of alleles identical by descent (IBD) at each marker. The original HasemanElston methods used the squared difference between the trait values of the siblings as the measure of similarity. More recently, Elston et al. [1] proposed the crossproduct of the trait values as the response. Several recent papers suggested other statistics as the response. Feingold [2] gave a comprehensive review of these regressionbased linkage methods.
The simple regression setup of the HasemanElston method allows one to easily study simultaneous effects of several loci and epistasis effects. It also allows one to apply model selection methods developed for regression analysis. In the Genetic Analysis Workshop 12, Suh et al. [3] explored these possibilities by applying stochastic search variable selection (SSVS), a Bayesian variable selection method proposed by George and McCulloch [4] for linear regression models, to the HasemanElston method. In the Genetic Analysis Workshop 13, Oh et al. [5] took a further step to consider 399 markers as well as their interactions using the same SSVS algorithm. In addition, they proposed using the marginal posterior of each variable, both main effects and interaction effects, to rank their importance. A known deficiency of SSVS is that its results are sensitive to specification of a number of parameters in prior settings. The effects of the choice of different prior settings are also not well understood. Oh et al. [5] also adopted a method developed by Chipman [6] to impose a hierarchical prior structure on the model space. Such a hierarchical prior reflects an a priori belief in certain types of models but further complicates the algorithm and adds computation burdens.
The goal of this work is to propose a simple and effective Bayesian variable selection method that requires minimum specification of prior settings and leaves flexibility to impose hypotheses on the model space. Another notable difference in our new approach from our previous approaches [3, 5] is that we make inference only on the markers but not specific interaction effects. However, when we evaluate a marker, its interactive effects with other markers are considered in additional to its main effects. The goal of our approach is to identify markers that are potentially related to the phenotype but not to find a genetic model that explains the phenotype.
Methods
We chose to analyze ALDX1 measured in the Collaborative Study on the Genetics of Alcoholism study, using microsatellite genotype information. We only took those individuals classified in ALDX1 as "purely unaffected", coded 1, "unaffected with some symptoms", coded 3, "affected", and coded 5. Among those individuals, we used GENEHUNTER to obtain IBD values for all sibpairs. We used the cross product CP_{ i } = (Y_{1i} m)(Y_{2i} m) as our measure of similarity of phenotype in the pair. Here m is the mean of all individuals in the sib group. We also used squared difference for comparison. We also included the sex as an explanatory variable. Sex was coded "0" if the sib pair was of the same gender and "1" if not. For simplicity, we assumed the errors to be independent and normally distributed, but the correlation structure could be implemented into our method in a straightforward way. Therefore, without loss of generality, we have
where ε~N(0, σ^{2}) are assumed to be independent, and X_{ j }values are either IBD values at markers, or measures of similarity for some covariate. We refer to these marker loci and covariates as factors in our discussion. The above model includes main effects as well as twofactor interaction effects and threeway interaction effects but can be extended to include higher order interaction effects.
Oh et al. [5] used SSVS to obtain a Markov chain Monte Carlo (MCMC) sample from the posterior of the model space of 2^{ p }models, where p is the total number of effects (i.e., main effects, epistasis effects) in the above equation. Each model corresponds to a subset of p effects being active. The marginal posterior of each effect is then obtained to evaluate the importance of the effect. When interaction effects are considered, the model space becomes extremely large so that the MCMC algorithm is less effective. Therefore, we propose an approach that reduces the model size dramatically while considering the interaction effects. Our model space only contains models, where m is the total number of factors (genetic markers and/or environment factors), often much smaller than p, and l is the number of factors involved in a model. The choice of l reflects the number of factors believed to be associated with the disease. Each model in our model space corresponds onetoone to a subset of l elements from n marker loci and covariates. Each model contains the main effects of l factors and interaction effects up to a certain order among those factors. The effects involving other factors are not included in the model, i.e., their corresponding coefficients in Equation (1) are set to be zero. Because all models in our model space have the same number of parameters, they are assigned the same prior probability. A model in our model space can be represented by a binary vector γ= (γ_{1}, γ_{2}, ..., γ_{ n }), where γ_{ i }= 1 indicates the i^{th} factor being active and . For a model γ, we denote the active coefficients as β_{ γ }in which the first element always corresponds to the constant term, and the model matrix as X_{ γ }. The prior of β_{ γ }is given as N(0, σ^{2}Γ^{1}) where . The variance of the error term is assigned a noninformative prior f(σ) ∞ 1 / σ. By integrating out β_{ γ }, the posterior of model γ can be shown to be
We adopt the MetropolisHasting algorithm to sample from the model space and obtain a sequence γ^{(0)}, γ^{(1)}, ..., γ^{(j)}, .... At the (j + 1) th step, a model γ* is selected by replacing a random active factor in γ^{(j)}by an random inactive factor not in γ^{(j)}. Set γ^{(j + 1)}= γ* with probability min {f(γ*y) / f(γ^{(j + 1)} y),1}; otherwise, γ^{(j + 1)}= γ^{(j)}. The marginal posterior of all factors are estimated from the MCMC sample to measure the importance of those factors. It is important to note here that the posterior probability of a factor not only captures its main effects but also its interaction effects with other factors.
Our study is conducted with two settings, one with square of difference, D^{2}, and the other with cross product, CP. There were 224 nuclear families and 1,499 sib pairs. The mean in CP was the average of each family. There are total of 315 candidate markers treated as factors. An additional factor is sex.
Results and Discussion
The methods described in the previous section were implemented using JAVA programming language for its portability. The program runs successfully under Linux and Mac OS X operating systems.
As model size increases, the overall frequency was higher, as we expected. Factors with high frequency are amplified and new factors with higher frequency appear. Marker GATA62F03 in chromosome 9 shows relatively high marginal probability in most of the plots, as does marker D1S226 on chromosome 1. These two markers also show evidence of linkage with ALDX1, but not as strong as D3S4542 and sex. We see an interesting region around the D6S493. For D^{2}, the plot shows some spikes and for CP, the plot shows relatively wide discontinuity in this area. It could be due to strong correlation between the IBD values or something associated with ADXL1. We would need to do further investigation of this region to address this issue.
Although one can visually inspect those plots and determine which markers are important, more formal statistical inferences are desired. However, this topic is not in the scope of this paper and will be addressed in the future.
Conclusion
In this paper, we applied a Bayesian method to select important markers using HasemanElston methods. The likelihood of a marker associated with a phenotype is evaluated by a posterior probability that captures not only main effects but also all interaction effects of the marker. The method is much simpler than previously proposed Bayesian variable selection methods and does not require fine tuning the prior settings. The result also shows the consistency and robustness of our method. Our approach is similar to the Bayesian factor screening method proposed by Box and Meyer [7]. A major difference between our approach and theirs is that we only consider models with same number of active factors, and they consider all 2^{ m }models. An important advantage of our approach is that we can assign equal priors to all models in our model space. Assigning priors to models of different sizes has many complications as discussed by Chipman et al. [8] in great detail.
Notes
Abbreviations
 IBD:

Identical by descent
 MCMC:

Markov chain Monte Carlo
 SSVS:

Stochastic search variable selection
Declarations
Acknowledgements
The research by SY is supported in part by Simons Foundation and Cold Spring Harbor Laboratory. The research of KQY is supported in part by National Science Foundation (DMS0306306).
Authors’ Affiliations
References
 Elston RC, Buxbaum S, Jacobs KB, Olson JM: Haseman and Elston revisited. Genet Epidemiol. 2000, 19: 117. 10.1002/10982272(200007)19:1<1::AIDGEPI1>3.0.CO;2E.View ArticlePubMedGoogle Scholar
 Feingold EL: Regressionbased quantitativetraitlocus mapping in the 21^{st} century. Am J Hum Genet. 2002, 71: 217222. 10.1086/341964.PubMed CentralView ArticlePubMedGoogle Scholar
 Suh YJ, Finch SJ, Mendell NR: Application of a Bayesian method for optimal subset regression to linkage analysis of Q1 and Q2. Genet Epidemiol. 2001, 21 (Suppl 1): 706711.Google Scholar
 George EI, McCulloch RE: Variable selection via Gibbs sampling. J Am Stat Assoc. 1993, 88: 881889. 10.2307/2290777.View ArticleGoogle Scholar
 Oh C, Ye KQ, He Q, Mendell NR: Locating disease genes using Bayesian variable selection with the HasemanElston method. BMC Genet. 2003, 4 (Suppl 1): S6910.1186/147121564S1S69.PubMed CentralView ArticlePubMedGoogle Scholar
 Chipman H: Bayesian variable selection with related predictors. Can J Stat. 1996, 24: 1736. 10.2307/3315687.View ArticleGoogle Scholar
 Box GEP, Meyer RD: Finding the active factors in fractioned screening experiments. J Quality Technol. 1993, 25: 94105.Google Scholar
 Chipman H, George E, McCulloch R: The practical implementation of Bayesian model selection. Model Selection, Institute of Mathematical Statistics Lecture Notes, Monograph series. Edited by: Lahiri, P. 2001, Beachwood, OH: IMS, 38: 67116.Google Scholar
Copyright
This article is published under license to BioMed Central Ltd. This is an open access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.