Partial least squares regression, support vector machine regression, and transcriptome-based distances for prediction...

Area of Science:

• Plant breeding
• Genomics
• Bioinformatics

Background:

• Hybrid performance prediction using parental gene expression data offers potential for increased breeding efficiency.
• Developing accurate prediction models is crucial for optimizing hybrid crop development.

Purpose of the Study:

• To compare the prediction accuracy of multiple linear regression (MLR), partial least squares regression (PLS), support vector machine regression (SVM), and transcriptome-based distances (D(B)).
• To evaluate model performance for predicting hybrid performance in scenarios with (type 2) and without (type 0) parental testcross data.

Main Methods:

• Gene expression profiling of parental maize inbred lines using a 56k microarray.
• Assessing grain yield of factorial hybrids (7 flint x 14 dent lines).
• Comparing prediction accuracies using two cross-validation schemes for MLR, PLS, SVM, and D(B).

Main Results:

• MLR, SVM, and PLS demonstrated high prediction accuracy for type 2 hybrids.
• For type 0 hybrids, D(B) showed greater prediction accuracy compared to regression methods.
• Regression methods were robust with a few hundred genes, while D(B) required 1,000-1,500 genes and was sensitive to gene set selection.

Conclusions:

• MLR is a promising approach for hybrid performance prediction within a single genetic material set.
• D(B) is most effective for transferring prediction models between related genetic materials.