Predicting complex traits using a diffusion kernel on genetic markers with an application to dairy cattle and wheat...

Area of Science:

• Quantitative genetics
• Statistical learning
• Genomic prediction

Background:

• Linear additive models in quantitative genetics may not fully capture complex genotype-phenotype relationships.
• Developing statistical learning models is crucial for predicting phenotypic values from molecular data and understanding genetic networks.
• Bayesian kernel ridge regression utilizes kernels (relationship matrices) for non-parametric prediction.

Purpose of the Study:

• To investigate the performance of a diffusion kernel for predicting phenotypic values.
• To compare the diffusion kernel with Gaussian and additive genomic relationship kernels.
• To assess the impact of embedding genetic markers in a non-Euclidean metric space on prediction accuracy.

Main Methods:

• Employed Bayesian kernel ridge regression, a non-parametric statistical learning model.
• Utilized a diffusion kernel, designed for discrete marker inputs, and compared it with Gaussian and additive kernels.
• Applied the models to Holstein cattle and wheat datasets for whole-genome prediction.

Main Results:

• The diffusion kernel's predictive ability was comparable to non-spatial additive genomic relationship kernels in Holstein cattle.
• The diffusion kernel outperformed additive kernels in wheat data.
• The performance difference between the diffusion kernel and the Gaussian kernel was negligible across datasets.

Conclusions:

• The diffusion kernel's ability to capture total genetic variance is not superior to the Gaussian kernel for the studied data.
• Embedding genetic markers in a non-Euclidean space had minimal impact on prediction accuracy.
• The use of the Gaussian kernel is justified due to its similar predictive performance and connection to the diffusion kernel.