Prediction performance of linear models and gradient boosting machine on complex phenotypes in outbred mice

Area of Science:

• Animal Genetics
• Quantitative Genetics
• Bioinformatics

Background:

• Genomic prediction utilizes marker data to predict breeding values.
• Linear models like GBLUP, BayesB, and elastic net are common for genomic prediction.
• Nonparametric methods, such as gradient boosting machine, offer alternative approaches.

Purpose of the Study:

• To compare the performance of linear genomic prediction models against a gradient boosting machine for complex traits in mice.
• To evaluate the effectiveness of these methods when epistatic effects are present.
• To assess the impact of data size and population structure on model performance.

Main Methods:

• Genomic data from 835 mice across 6 generations with 50,112 SNP markers.
• Phenotypic data for bone mineral density, body weight, fat percentage, cholesterol, glucose, insulin, triglycerides, and urine creatinine.
• Comparison of GBLUP, BayesB, elastic net, and gradient boosting machine models using a multi-generation dataset.

Main Results:

• Linear models outperformed gradient boosting machine for 7 out of 10 complex traits.
• Gradient boosting machine showed superior prediction accuracy for bone mineral density, cholesterol, and glucose, traits potentially influenced by epistasis.
• Using top markers selected by gradient boosting machine improved prediction accuracy in both linear and gradient boosting machine models for some traits.

Conclusions:

• Linear models are generally more robust for polygenic traits in this dataset.
• Gradient boosting machine is a competitive method for genomic prediction of complex traits, especially when epistatic effects are suspected.
• Model performance is sensitive to data size and the connectedness between reference and validation populations.