Mul-PheG2P: decoupled learning and prediction-space fusion enables robust and interpretable multi-phenotype genomic...

Area of Science:

• Genomics
• Plant Breeding
• Machine Learning

Background:

• Genomic prediction of multiple phenotypes is vital for plant breeding.
• Current methods face challenges with negative transfer and lack interpretability, especially with high-dimensional, small-sample data.

Purpose of the Study:

• To introduce Mul-PheG2P, a novel paradigm for multi-phenotype genomic prediction.
• To address limitations of existing methods in terms of negative transfer and interpretability.

Main Methods:

• Mul-PheG2P utilizes a two-stage decoupled learning approach with predictive space fusion.
• Phenotype-specific encoders are trained on genetic data, followed by cross-phenotype aggregation via an interpretable prediction layer.

Main Results:

• Mul-PheG2P demonstrates superior performance over existing methods in maize, wheat, and tomato datasets.
• The model provides multi-scale interpretability, quantifying phenotypic contributions and revealing genetic bases using Integrated Gradients.
• It successfully identified key genetic motifs (CCT, SQUAMOSA promoter) related to plant traits.

Conclusions:

• Mul-PheG2P offers high performance and interpretability for multi-phenotype genomic prediction.
• Its ability to capture functional biological mechanisms advances precision breeding through cost-effective, large-scale screening.