Software

Estimation of variance components in single-step models using MiX99

---

Abstract

In this study we investigated the computational performance of Monte Carlo (MC) Expectation Maximization (EM) REML variance component estimation for different single-step model representations. The MiX99 software supports the estimation of variance components using REML. To efficiently analyze large data sets, a MC approach has been implemented that uses the EM-REML algorithm. The variance component estimation has been implemented for most linear mixed effect models used for breeding value evaluation, including the standard single-step GBLUP (ssGBLUP), the fully component-wise ssGTABLUP (ssGTABLUP), and the single-step SNPBLUP (ssSNPBLUP) models. To assess the computational performance of these approaches, variance components were estimated using a simulated small data set having 44,280 individuals in the pedigree with phenotypic records for a simple single trait model. Genotypes were available for either 10,000, 20,000, or 30,000 individuals from the most recent generations. The model included a fixed generation number and random genetic and residual effects. Every EM-REML iteration had one original data and five sampled data breeding value estimations. In all cases, ssSNPBLUP required the least amount of RAM memory (at most 3.2 GB) and ssGBLUP needed the most (about 15.5 GB), which is the case for breeding value estimation as well. The number of EM-REML iterations until convergence varied between the models, affecting to the overall computing time. The largest difference was observed in the 30,000 genotyped case for ssGBLUP to require 370 EM-REML iterations, when the other models managed with about 300 EM-REML iterations. Despite these differences in the EM-REML convergence behavior, the estimated variance components were similar by the models when the number of genotyped individuals was the same. Computationally, ssGBLUP was the fastest. The ssSNPBLUP model was always the slowest. This may be due to the convergence criterion of BLUP (norm of difference between iterates) signaling convergence differently than in ssGBLUP and ssGTABLUP. These differences in computing times are similar as observed in the estimation of breeding values. Larger pedigrees with more genotyped individuals may favor ssSNPBLUP due to its lowest memory requirements.

Keywords: 2026

How to Cite:

Gao, H., Taskinen, M., Pitkänen, T., Ylitalo, A., Vargas Jurado, N., Lidauer, M. & Strandén, I., (2026) “Estimation of variance components in single-step models using MiX99”, World Congress on Genetics Applied to Livestock Production Digital Archive 2026(1): 2283934. doi: https://doi.org/10.31274/wcgalp.23537

Rights: 1