Application of high-dimensional feature selection: evaluation for genomic prediction in man

Area of Science:

• Genomics
• Statistical Genetics
• Quantitative Genetics
• Bioinformatics

Background:

• Genomic prediction models are crucial for estimating breeding values and predicting complex traits.
• Comparing different prediction methods and feature selection approaches is essential for optimizing genomic selection strategies.
• Understanding the performance of models like G-BLUP and Bayes C across diverse traits and populations is vital.

Purpose of the Study:

• To evaluate the impact of five feature selection methods on the predictive performance of G-BLUP and Bayes C.
• To compare the accuracy of G-BLUP and Bayes C for predicting height, HDL, and BMI in Croatian and UK populations.
• To assess the generalizability and computational efficiency of feature selection within G-BLUP.

Main Methods:

• Utilized genome-wide SNP data from 2,186 Croatian individuals for within-population prediction.
• Employed G-BLUP (Genomic Best Linear Unbiased Prediction) and Bayes C (Bayesian method) prediction models.
• Investigated five distinct feature selection approaches, including supervised methods, to identify informative SNPs.

Main Results:

• G-BLUP and Bayes C demonstrated comparable predictive performance using all SNPs across traits in the Croatian dataset.
• Both methods showed similar accuracy for highly polygenic traits (height, BMI) when predicting into the UK population.
• Bayes C outperformed G-BLUP for HDL prediction in the UK data, suggesting an advantage for traits influenced by moderate-effect loci.

Conclusions:

• Supervised feature selection within G-BLUP offers a flexible, generalizable, and computationally efficient alternative to Bayes C.
• The choice of prediction method and feature selection strategy impacts accuracy, particularly for traits with varying genetic architectures.
• Careful evaluation of predictive performance is necessary when employing feature selection in genomic prediction models.