Benchmarking Parametric and Machine Learning Models for Genomic Prediction of Complex Traits

Area of Science:

• Agricultural Science
• Genetics
• Computational Biology

Background:

• Genomic prediction is crucial for crop and livestock breeding.
• Advanced algorithms like artificial neural networks (ANNs) and gradient tree boosting are being developed.
• Systematic comparisons of these algorithms across diverse datasets are lacking.

Purpose of the Study:

• To systematically compare the performance of various genomic prediction algorithms.
• To evaluate linear and non-linear models across different plant species and traits.
• To identify optimal strategies for improving algorithm performance.

Main Methods:

• Compared six linear and six non-linear algorithms using data from 18 traits across six plant species.
• Investigated the impact of hyperparameter tuning and feature selection.
• Evaluated individual algorithms and ensemble predictions.

Main Results:

• No single algorithm consistently outperformed others across all traits and species.
• Ensemble predictions demonstrated robust and consistent performance.
• Non-linear algorithms showed more variable performance compared to linear models.
• Feature selection and specific strategies improved ANN performance.

Conclusions:

• Algorithm selection is critical for accurate trait value prediction in genomic prediction.
• Ensemble methods offer a reliable approach for genomic prediction.
• Further research into optimizing non-linear algorithms, including ANNs, is warranted.