Using visual scores for genomic prediction of complex traits in breeding programs.

Area of Science:

• Plant breeding and genetics
• Quantitative genetics
• Bioinformatics

Background:

• Genomic prediction methods often assume normally distributed data.
• Visual scores in plant/animal breeding are frequently categorical, violating this assumption.
• This violation can impact breeding value prediction and genetic parameter estimation.

Purpose of the Study:

• Evaluate methods for handling visual scores in genomic prediction and genetic parameter estimation.
• Address challenges posed by subjectivity and errors in visual scoring.
• Improve decision-making for breeders using recurrent selection schemes.

Main Methods:

• Compared Linear Mixed Models, Bayesian Linear Regression, Bayesian Ordinal Regression Models (BORM), and Random Forest Classification.
• Utilized simulated and real breeding data sets, including autotetraploid blueberry.
• Assessed strategies for data collection (number of categories) and phenotype type (continuous vs. categorical).

Main Results:

• Collecting data with 1-5 categories is optimal, even with score errors.
• BORM and Random Forest Classification offer marginal gains over robust methods like Linear Mixed Models and Bayesian Linear Regression.
• BORM provides superior estimation of genetic parameters.
• Investing in 600-1000 low-error categorical data points can improve predictive abilities when continuous phenotypes are infeasible.

Conclusions:

• Bayesian Ordinal Regression Models (BORM) and Random Forest Classification are effective for visual scores in genomic prediction.
• Data collection with intermediate categories (1-5) and high-quality phenotyping are crucial.
• Findings are applicable to real breeding data, aiding breeder decision-making and improving predictive abilities.