Genotype-by-environment interaction with high-dimensional environmental data: an example in pigs

Abstract

Background: In traditional genetic prediction models, environments are typically treated as uncorrelated effects, either fixed or random. Environments can be correlated when they share the same location, management practices, or climate conditions. The temperature-humidity index (THI) is often used to address environmental effects related to climate or heat stress. However, it does not fully describe the complete climate profile of a specific location. Therefore, it is more appropriate to use multiple environmental covariates (ECs), when available, to describe the weather in a specific environment. This raises the question of whether publicly available weather information (such as NASA POWER) is useful for genomic predictions. Genotype-by-environment interaction (GxE) can be modeled using multiple-trait models or reaction norms. However, the former requires a substantial number of records per environment, while the latter can result in over-parametrized models when the number of ECs is large. This study investigated whether using ECs is a suitable strategy to correlate environments (herds) and to model GxE in the genomic prediction of purebred pigs for production traits.

Results: We evaluated different models to account for environmental effects and GxE. When environments were correlated based on ECs, we observed an increase in environmental variance, which was accompanied by an increase in phenotypic variance and a decrease in heritability. Furthermore, including environments as an uncorrelated random effect yielded the same accuracy of estimated breeding values as treating them as correlated based on weather information. All the tested models exhibited the same bias, but the predictions from the multiple-trait models were under-dispersed. Evidence of GxE was observed for both traits; however, there were more genetically unconnected environments for backfat thickness than for average daily gain.

Conclusions: Using outdoor weather information to correlate environments and model GxE offers limited advantages for genomic predictions in pigs. Although it adds complexity to the model and increases computing time without improving accuracy, it does enhance model fit. Including environment information (e.g. herd effect) as an uncorrelated random effect in the model could help address GxE and environmental effects.

Fig. 1

Averages and standard deviations of environmental covariates within each environment (farm) for 30 and 100 days. T temperature at two meters (°C), Td dew/frost temperature (°C), Tw wet-bulb temperature, Ts hearth-skin temperature (°C), H relative humidity (%), R precipitation/rainfall (mm), Ws wind speed at two meters (m/s), and Wd wind direction at two meters (°)

Fig. 2

Correlations among environmental covariates (ECs) over 30 days (above diagonal) or 100 days (below diagonal). T temperature at two meters (°C), Td dew/frost temperature (°C), Tw wet-bulb temperature, Ts hearth-skin temperature (°C), H relative humidity (%), R precipitation/rainfall (mm), Ws wind speed at two meters (m/s), and Wd wind direction at two meters (°)

Fig. 3

Correlations among environments (farms) based on environmental covariates (ECs) over 30 days (above diagonal) or 100 days (below diagonal)

Fig. 4

Validation statistics for all tested models by trait, considering all animals in a given environment (farm) as focal individuals (environment-validation). MG traditional genomic best linear unbiased predictor (GBLUP) model, ME₃₀ GBLUP considering environmental effects correlated based on 30 days of weather information, ME₁₀₀ GBLUP considering environmental effects correlated based on 100 days of weather information, MGE₃₀ GBLUP considering genotype by environment interaction (GxE) based on 30 days of weather information, MGE₁₀₀ GBLUP considering GxE based on 100 days of weather information, MTM multiple-trait GBLUP model, ADG average daily gain, BFT backfat thickness