Estimation & Prediction

Bayesian variance component estimation for GBLUP using hybrid variational inference

---

Abstract

In this study we compared the performance of a natural gradient hybrid variational inference (VI) algorithm (NGHVI) and the Gibbs sampler (GS) for variance component estimation (VCE) in genomic models. Estimation of variance components (VC) is an essential task in genetic and genomic evaluations. Bayesian inference of VC typically relies on sampling from the posterior distribution of the parameters via the GS. While accurate, the GS requires long computational times, especially for complex and dense models involving genomic data. On the other hand, VI relies on minimizing the Kullback-Leibler divergence, or evidence lower bound (ELBO), between the posterior distribution of the parameters and an approximating density, e.g., a multivariate normal, accomplished by stochastic gradient ascent. Thus, VI may provide a computationally efficient alternative to the GS. However, the performance of these approaches for VCE has not been assessed. Two simulated data sets were used: (i) a bivariate GBLUP model, and (ii) a single-trait random regression GBLUP model. For (i), the two simulated traits had a genetic correlation of 0.95 and a residual correlation of 0.50. Genomic data were simulated for either 500 or 5000 individuals such that the number of random effect equations was 1000 and 10000, respectively. For (ii), an intercept, linear, quadratic, and exponential random regression coefficient was simulated for both the permanent-environmental and additive genetic effects of the simulated 6400 genotyped individuals such that the number of random effect equations was 51200. Simulated fixed effects included a mean, and fixed regression coefficients for (i) and (ii), respectively. In all scenarios, 5500 samples were obtained for the while the NGHVI method was run until convergence was achieved. Convergence of the GS was determined by posteriori analysis of samples while for the NGHVI was assessed through changes in the ELBO. For the bivariate model with 5000 genotyped individuals, computing times were 66.3 h for GS and 14.5 h for NGHVI, while for the random regression model they were approximately 17d for GS and 3d for NGHVI. For the bivariate model with 500 genotyped individuals, the posterior distribution of the genetic variances was skewed such that posterior variances from NGHVI were underestimated. However, when the number of individuals genotyped increased to 5000, both the posterior means from GS and NGHVI methods resulted in a high correlation > 0.99, and a correlation of 0.92 between posterior standard deviations. For the random regression model, correlation between the posterior means of the two methods > 0.99, and correlation between the posterior standard deviations of the two methods > 0.98. In conclusion, for models where more information is available, NGHVI inference was accurate. Thus, for large models such as those involving random regression, VI may be an efficient alternative to the GS.

Keywords: 2026

How to Cite:

Vargas Jurado, N., Strandén, I. & Lidauer, M., (2026) “Bayesian variance component estimation for GBLUP using hybrid variational inference”, World Congress on Genetics Applied to Livestock Production Digital Archive 2026(1): 2284535. doi: https://doi.org/10.31274/wcgalp.23579

Rights: 1