Improving predictive ability in sparse testing designs in soybean populations.

Area of Science:

• Agricultural Science
• Genetics
• Plant Breeding

Background:

• High-dimensional genomic data and genome-based prediction (GP) models accelerate genetic gains in soybean breeding.
• GP-based sparse testing optimizes genotype evaluation capacity and reduces costs.
• This study pioneers GP-based sparse testing implementation in soybeans.

Purpose of the Study:

• To evaluate the impact of training set composition and size on GP predictive ability in soybean.
• To investigate the effectiveness of overlapping (O-RILs) and non-overlapping (NO-RILs) recombinant inbred lines (RILs) in training sets.
• To assess methodologies for maximizing genetic diversity within fixed-size training sets.

Main Methods:

• Utilized 1,755 RILs from 39 Soybean Nested Association Mapping (NAM) populations tested across nine environments.
• Assessed predictive abilities using various GP models with different training set sizes and compositions (O-RILs vs. NO-RILs).
• Employed methods to maximize or minimize genetic diversity in fixed-size training samples.

Main Results:

• Reduced training set size generally decreased predictive ability across most compositions.
• Maximizing genetic diversity and including O-RILs improved prediction accuracy for a fixed training set size.
• The most complex GP models showed less sensitivity to training set composition and diversity factors.

Conclusions:

• GP-based sparse testing is effective for soybean breeding, with training set composition significantly impacting accuracy.
• Incorporating genetic diversity and overlapping RILs are key strategies for enhancing predictive performance.
• Increased environmental testing early in the breeding pipeline aids in selecting stable, broadly adapted genotypes.