(Quasi) multitask support vector regression with heuristic hyperparameter optimization for whole-genome prediction of...

Area of Science:

• Animal Genetics and Breeding
• Genomic Prediction Models
• Machine Learning in Agriculture

Background:

• Genomic prediction models are crucial for improving livestock breeding.
• Multitrait (MT) models can enhance prediction accuracy by leveraging information from related traits.
• Support Vector Regression (SVR) offers a nonlinear approach to genomic prediction.

Purpose of the Study:

• To investigate nonlinear kernels for MT genomic prediction using SVR.
• To propose and evaluate a novel (quasi) multitask SVR (QMTSVR) approach.
• To compare QMTSVR with established single-trait (ST) and MT models for carcass trait prediction in broiler chickens.

Main Methods:

• Implemented QMTSVR with hyperparameter optimization via genetic algorithm.
• Benchmarked QMTSVR against ST and MT Bayesian shrinkage models (GBLUP, BayesC, RKHS).
• Assessed predictive ability using prediction accuracy (ACC), standardized RMSE*, and inflation factor (b) across two validation designs (CV1, CV2).

Main Results:

• QMTSVR-CV2 achieved the highest prediction accuracy (ACC) and lowest standardized RMSE* for both carcass traits.
• The proposed QMTSVR model showed higher predictive accuracy than MT-GBLUP and MT-BayesC.
• QMTSVR performance was comparable to MT-RKHS and competitive with conventional MT Bayesian regression models.

Conclusions:

• The QMTSVR approach offers a competitive and effective nonlinear method for MT genomic prediction.
• QMTSVR enhances predictive ability for carcass traits in broiler chickens.
• The choice of validation design and accuracy metric can influence model selection, highlighting the importance of comprehensive evaluation.