Genomic prediction based on data from three layer lines using non-linear regression models.

Area of Science:

• Animal genetics
• Quantitative genetics
• Genomic prediction

Background:

• Linear models are commonly used for genomic prediction in multi-population reference sets.
• Data heterogeneity in multi-population datasets can violate linear model assumptions.

Purpose of the Study:

• To compare the performance of non-linear models and a multi-trait genomic best linear unbiased prediction (GBLUP) model against conventional linear models for genomic prediction.
• To evaluate models using data from three lines of layer hens (B1, B2, W1).

Main Methods:

• Employed multi-trait GBLUP and non-linear kernel learning models.
• Compared these with conventional linear models for genomic prediction.
• Utilized training datasets (1004-1023 animals) and validation datasets (238-240 animals) for three hen lines.

Main Results:

• Non-linear models performed similarly to linear models when using data solely from the evaluated line.
• Non-linear models maintained accuracy when distantly related lines were added, unlike linear models which sometimes decreased.
• Multi-trait GBLUP leveraged genetic correlations between lines.
• Combining linear and non-linear models enhanced multi-line genomic prediction accuracy.

Conclusions:

• Non-linear radial basis function (RBF) models showed similar performance to linear models for genomic prediction, contrary to expectations for handling heterogeneous data.
• Modeling trait-by-line combinations as separate, correlated traits effectively addressed data heterogeneity.
• Complementary information from non-linear models in multi-line datasets suggests underlying data distribution differences.