MDNN: memetic deep neural network for genomic prediction.

Area of Science:

• Plant breeding
• Genomics
• Artificial intelligence

Background:

• Genomic prediction (GP) is crucial for crop and livestock improvement.
• Traditional linear models struggle with complex traits, necessitating advanced methods.
• Deep learning (DL) offers potential but requires manual architecture design.

Purpose of the Study:

• To develop an automated framework for optimizing deep learning architectures in GP.
• To address the limitations of manual neural network design in breeding applications.

Main Methods:

• Proposed a novel framework, MDNN (Memetic Deep Neural Network), integrating memetic algorithms for automated neural architecture search (NAS).
• Applied MDNN to genomic prediction tasks, focusing on complex trait analysis.
• Compared MDNN performance against traditional deep neural network genomic prediction (DNNGP).

Main Results:

• MDNN demonstrated superior performance in genomic prediction accuracy.
• Achieved a 36.49% improvement in Pearson correlation on the wheat599 dataset.
• Showcased a 12.28% improvement on the wheat2000 dataset, highlighting efficiency with larger datasets.

Conclusions:

• MDNN offers an effective solution for automated deep learning architecture optimization in GP.
• The framework enhances prediction accuracy for complex traits, advancing breeding strategies.
• Automated NAS via memetic algorithms represents a significant step forward for DL in agriculture.