Genomic-enabled prediction with classification algorithms | JoVE Visualize

Area of Science:

Plant Breeding and Genetics
Bioinformatics and Computational Biology
Statistical Genomics

Background:

Pearson's correlation coefficient (ρ) is widely used to assess prediction success in genomic selection (GS), but may not accurately reflect performance in selecting elite individuals.
Real-world breeding programs require methods that effectively identify superior individuals from the distribution tails, where traditional metrics might be insufficient.

Purpose of the Study:

To evaluate the effectiveness of various regression and classification algorithms for genomic selection in identifying elite individuals in maize and wheat.
To compare the predictive accuracy of different models, focusing on their ability to select individuals within specific top percentiles (e.g., 15%) of the distribution.

Main Methods:

Cross-validation was employed on 14 maize and 16 wheat datasets, using 90% of individuals for training and 10% for testing.
Six regression models (Bayesian LASSO, Ridge Regression, RHKS, RFR, SVR-lin, SVR-rbf) and three classification models (RFC-lin, SVC-lin, SVC-rbf) were assessed.
Predictive accuracy was measured using Cohen's kappa coefficient (κ) and relative efficiency (RE), focusing on selection percentiles like α=15%.

Main Results:

Support Vector Classification with linear kernel (SVC-lin) demonstrated the highest predictive accuracy (κ and RE) in most maize and wheat datasets, particularly for the 15% selection threshold.
Regression models showed varied performance, with Ridge Regression (RR) and Reproducing Kernel Hilbert Spaces (RHKS) performing well in some instances.
The performance of both regression and classification algorithms was highly dependent on the selection percentile, highlighting the importance of choosing appropriate metrics.

Conclusions:

Classification methods, especially SVC-lin, offer a promising alternative to traditional regression approaches for genomic selection in plant breeding.
The study underscores the need to evaluate GS methods based on their ability to select elite individuals, not just overall predictive accuracy.
The effectiveness of GS algorithms is contingent on the specific selection threshold, necessitating careful consideration of the breeding program's goals.