Multivariate genome wide association for continuous and discrete responses using multivariate Bernoulli prior

• Published in: Journal of animal science 2018-12, Vol.96, p.126-126
• Main Authors: Rekaya, R, Toghiani, S, Sumreddee, P, Ling, A, Aggrey, S
• Format: Article
• Language: English
• Subjects: Bayesian analysis; Computer simulation; Genetic transformation; Genomes; Genomics; Genotypes; Identification methods; Markers; Multivariate analysis; Simulation; Single-nucleotide polymorphism; Statistical analysis; Statistical methods; Statistics
• ISSN: 0021-8812; 1525-3163

Genome selection is becoming the tool of choice for genetic evaluation. When the SNP effects are directly fitted, the association models are generally implemented univariately. This is largely due to statistical convenience. Several approximations were presented including the assumption of independence of residual terms, estimation of (co)variances under a null model (no SNPs included) followed by data transformation, or through a set of single marker analyses after previous adjustment for the background. In all these cases, major assumptions were made about data generating process and/or the genetic determination of traits. In presence of discrete responses, these approximations are no longer appropriate. Respecting the distributional form of discrete traits will further complicate the statistical analysis. Consequently, there have been only a few examples of joint association analyses of continuous and discrete responses. Extending methods that includes all markers into the association model (e.g., GBLUP) will further accentuate the problem of lack of statistical power. Methods that prioritize SNPs (e.g. BayesC) are more appropriate; however identifying the set of markers to be included in the association model for each trait is a major challenge. In this study, we present a flexible general method for the multivariate GWAS analysis of continuous and discrete traits based on the assumption of a multivariate Bernoulli distribution as a prior for the indicator variables of trait specific markers to be included in the association model. Three correlated traits (two continuous and one discrete) and 50k SNP marker genotypes were simulated. Several scenarios were simulated with varying heritabilities (0.1 to 0.4) and genetic and residual (co)variances. Except for one scenario, the multivariate implementation resulted in a small to moderate gain in accuracy for the continuous traits (2 -6%). For the discrete response, the gain in accuracy was substantial and ranged between 3 -14% depending on the genetic (co)variances.