Estimation & Prediction

Genetic parameters via predictivity in large scale multi-trait genomic selection models

---

Abstract

Estimation of accurate genetic parameters require using all data that were available upon selection. Existing methods for parameter estimation become computationally unfeasible with large genomic data. With higher intensity of selection, the genetic parameters change rapidly over time and their real time estimation is essential for accurate estimation of genetic gain. Our study aimed to test a new method called 'GPP' (Genetic Parameters via Predictivity), which combines within and across traits predictivity formulas with a deterministic formula to predict the accuracy of genomic breeding values. The GPP method can handle large data size and can estimate parameters across time slices. GPP method was tested using simulated datasets under a two-trait model with negatively and positively correlated traits: primary production (h2=0.40), secondary production (h2=0.10) and a fitness (h2=0.10) trait. Three scenarios included 5,000 (A), 40,000 (B) and 100,000 (C) genotyped animals per generation for 10 generations. Genomic selection was based on the primary production trait, whereas the secondary production and fitness traits were composite traits, with their genetic correlation to primary production trait increasing or decreasing by 0.1 unit per generation, respectively. Genetic parameters for time slices were estimated by REML and GREML (1-3 generations) in scenario A and B, whereas by REML using one generation in scenario C. GPP used 3-generation windows, with the first 2 generations treated as a reference population. True parameters from the simulation were used as benchmark. Analyses covered generations 3-10. Across the three scenarios, GPP estimates closely matched the realized values, with a slight, variable bias. Using incorrect variances to compute GEBV prior to GPP had non-significant impact on estimates. With the smaller data sets, estimates by REML using only one generation were highly variable, and estimates by GREML had lower standard errors. For large data, REML estimates were slightly biased. Genetic correlations obtained with GPP were non-symmetric, meaning they varied depending on which trait adjusted phenotype was used as a benchmark in the formula. We observed lower bias when predictivity involved adjusted phenotypes for the secondary trait. The GPP method performed equally well across both positively and negatively correlated traits, as well as for weak and strong genetic correlation scenarios, in small as well as large datasets. Computations took 55 min and 11 s for a one-time slice run of GPP in the large dataset (scenario C). GPP had an approximately linear cost with the number of genotyped animals. GPP is a fast, flexible, and accurate approach for estimating dynamic genetic parameters for positively or negatively correlated traits and also for weak and strong genetic correlation scenarios across time slices in large datasets, subject to genomic evaluation, where predictivity computation is feasible.

Keywords: 2026

How to Cite:

Gowane, G., Hidalgo, J., Misztal, I. & Lourenco, D., (2026) “Genetic parameters via predictivity in large scale multi-trait genomic selection models”, World Congress on Genetics Applied to Livestock Production Digital Archive 2026(1): 2208921. doi: https://doi.org/10.31274/wcgalp.23377

Rights: 1