Missing comparability: When genomic selection faces field variability. A case study in soybeans

Area of Science:

• Plant breeding
• Quantitative genetics
• Agricultural science

Background:

• Phenotypic data processing is vital for genomic selection (GS) model training.
• Existing methods often fail to fully separate genetic signals from field variability.
• Statistical metrics for model comparison can be inconclusive due to unknown true breeding values.

Purpose of the Study:

• To evaluate spatial models for separating genetic and field variability components.
• To assess the impact of these models on genomic selection (GS) implementation.
• To investigate the influence of field variability correction on breeding value prediction.

Main Methods:

• Analysis of real soybean (Glycine max L. Merr.) data.
• Simulation study under controlled conditions.
• Implementation of three spatial models (M1: block, M2: block + row + column, M3: block + row + column + row × column).

Main Results:

• Real data: Accounting for field variability reduced the predictive ability of isolated genetic signals.
• Simulated data: Field variability corrections enhanced the predictive ability of breeding values.
• The choice of spatial model impacts the effectiveness of genetic signal isolation.

Conclusions:

• Training GS models with isolated genetic signals enhances breeding value predictability.
• Standard benchmarks (correlation between predicted and observed values) can be misleading.
• Accurate separation of genetic and field variability is essential for reliable GS.