Heritability Estimation in Matrix-Variate Mixed models -- A Bayesian Approach

Abstract: Since the emergence of genome-wide association studies (GWASs), estimation of the narrow sense heritability explained by common single-nucleotide polymorphisms (SNPs) via linear mixed model approaches became widely used. As in most GWASs, most of the heritability analyses are performed using univariate approaches i.e. considering each phenotype independently. In this study, we propose a Bayesian matrix-variate mixed model that takes into account the genetic correlation between phenotypes in addition to the genetic correlation between individuals which is usually modelled via a relatedness matrix. We showed that when the relatedness matrix is estimated using all the genome-wide SNPs, our model is equivalent to a matrix normal regression with matrix normal prior on the effect sizes. Using real data we demonstrate that there is a boost in the heritability explained when phenotypes are jointly modelled (25-35% increase). In fact based on their standard error, the joint modelling provides more accurate estimates of the heritability over the univariate modelling. Moreover, our Bayesian approach provides slightly higher estimates of heritability compared to the maximum likelihood method. On the other hand, although our method performs less well in phenotype prediction, we note that an initial imputation step relatively increases the prediction accuracy.