Graphical Model Selection to Infer the Partial Correlation Network of Allelic Effects in Genomic Prediction With an...

Area of Science:

• Genomics
• Statistical Genetics
• Network Analysis

Background:

• Genomic prediction models often assume independence of marker effects, which can limit accuracy.
• Estimating complex relationships between genetic markers is crucial for understanding trait architecture.

Purpose of the Study:

• To develop and evaluate novel statistical methods for genomic prediction that account for partial correlation of marker effects.
• To estimate partial correlation networks (PCN) and precision matrices from genomic data.
• To assess the performance of these methods in predicting breeding values and identifying biologically relevant genomic regions.

Main Methods:

• Extended canonical model selection for Gaussian concentration and directed acyclic graph models.
• Developed frequentist methods (Glasso-EM, Concord-EM, CSCS-EM) integrating the EM algorithm.
• Developed Bayesian hierarchical models (Bayes G-Sel, Bayes DAG-Sel).
• Applied methods to a bull fertility dataset and conducted simulation studies.

Main Results:

• Methods accurately recovered PCN (up to 0.98 accuracy with Concord-EM).
• Concord-EM provided the best precision matrix estimation.
• Glasso-EM achieved the highest breeding value prediction reliability (0.85).
• Identified biologically relevant genomic regions based on high connectivity in the PCN.

Conclusions:

• The developed methods effectively estimate partial correlation networks and precision matrices.
• Identifying highly connected genomic regions in the PCN provides biological insights.
• These advanced statistical approaches enhance genomic prediction accuracy and genetic discovery.