Estimation & Prediction

Estimation of heritabilities and genetic correlations by time predictivity in large genomic models

---

Abstract

Estimating genetic parameters in large genomic models is computationally challenging, as accurate estimation requires complete datasets with genomic information, and most methods, e.g. REML-based or Bayesian-based methods, do not scale well. Under genomic selection, parameters may change rapidly, and must be estimated in time slices. Here we (1) present a method called Genetic Parameters via Predictability (GPP) to estimate heritabilities based on predictivity, and genetic correlations between traits based on across-trait predictivity, (2) apply GPP to national data sets with large genomic data, and (3) examine the efficiency of the GPP by comparing its performance with a Bayesian model via Gibbs sampling (GIBBS). Heritability was estimated as hÌ‚2 ＝ ½(c2 ± sqrt(c4 + 4c2Me/N), where c is predictive ability, equal to corr(y - Xb,à»), Me is the number of independent chromosome segments, and N is the number of genotyped animals with phenotypes in the reference population. Genetic correlation between traits i and j was calculated as corr(ui,uj) = c/Accjhi= corr(yi - Xbi, Ì‚uj)/Accjhi, where yi - Xbi is a vector of observations for selection candidates (validation population) adjusted for fixed effects, uÌ‚j is a vector of GEBVs of trait j for animals calculated from N observations excluding the validation population, Accj is the accuracy in GEBVs à»j, and hi is the square root of heritability for trait i. We tested GPP using 18 linear type traits in US Holsteins using data from 692,040 genotyped animals and 6.17 million pedigree animals, using 11 training datasets and 11 validation populations each spanning 3 years from 2009 to 2020. Heritability and correlation corr(ui,uj) estimates were tracked over time. We compared these to estimates by GIBBS using validation populations only and ignoring genomic information. We then compared Spearman correlations between GEBVs from GPP and GIBBS. Heritability estimates by GPP were generally lower than by BM. While trends from both were similar for many traits, they diverged for stature and feet & leg. Genetic correlations by GPP were generally higher than by GIBBS. Differences in correlations were particularly high for stature but low for rear teat placement. The largest change in correlations (0.33) was between rear udder width and udder depth, from 0.36 in 2010 to 0.03 in 2020. This may be attributable to recent selection for longevity or structural changes toward more compact body. Spearman rank correlations for GEBVs by GIBBS and GPP ranged between 0.96 and 1. GPP proved computationally efficient, with computing for one time slice taking about 7 hours (with genomics) compared to 14 days by GIBBS (without genomics). GPP allows estimating genetic parameters using complete data with genomic information within a reasonable time. Methods ignoring genomic information are subject to preselection bias, particularly for stronger selected traits.

Keywords: 2026

How to Cite:

Ben Zaabza, H., Tsuruta, S., Hidalgo, J., Lourenco, D. & Misztal, I., (2026) “Estimation of heritabilities and genetic correlations by time predictivity in large genomic models”, World Congress on Genetics Applied to Livestock Production Digital Archive 2026(1): 2286913. doi: https://doi.org/10.31274/wcgalp.24162

Rights: 1