A comparison of random forests, boosting and support vector machines for genomic selection

Area of Science:

• Quantitative genetics
• Bioinformatics
• Machine learning in animal and plant breeding

Background:

• Genomic selection (GS) estimates breeding values using genome-wide molecular markers.
• Accurate prediction of genomic breeding values (GEBVs) is crucial for modern breeding programs.
• Evaluating diverse marker-based prediction approaches is essential for identifying optimal methods.

Purpose of the Study:

• To compare the predictive accuracy of random forests (RF), stochastic gradient boosting (boosting), and support vector machines (SVMs) for GEBVs.
• To explore the use of RF for marker importance ranking and quantitative trait loci (QTL) discovery.

Main Methods:

• GEBVs were predicted for a simulated quantitative trait using dense single nucleotide polymorphism (SNP) markers.
• Five-fold cross-validation was employed to measure predictive accuracy via Pearson correlation.
• Marker importance was assessed using RF and visualized against chromosomal positions and QTLs.

Main Results:

• Stochastic gradient boosting yielded the highest predictive accuracy (r=0.547).
• Support vector machines (r=0.497) and random forests (r=0.483) showed intermediate and lower accuracies, respectively.
• Boosting outperformed SVMs and RF in predicting breeding values.

Conclusions:

• Stochastic gradient boosting demonstrated superior performance in predicting GEBVs among the tested machine learning methods.
• The predictive accuracies of RF, SVMs, and boosting were comparable and similar to ridge regression BLUP (RR-BLUP).
• RF can be a valuable tool for marker-based QTL discovery and pre-screening in genomic selection.