MtCro: multi-task deep learning framework improves multi-trait genomic prediction of crops.

Area of Science:

• Agricultural Science
• Genetics
• Computational Biology

Background:

• Genomic Selection (GS) utilizes genome-wide markers to predict traits, accelerating genetic progress in plant breeding.
• Deep learning models have shown promise for enhancing GS prediction accuracy.
• Current deep learning approaches often overlook the crucial inter-correlations among diverse plant phenotypes.

Purpose of the Study:

• To introduce MtCro, a multi-task learning framework designed to simultaneously model multiple plant phenotypes within a shared parameter space.
• To evaluate the performance of MtCro against existing deep learning models for genomic prediction.
• To demonstrate the impact of capturing inter-phenotype correlations on prediction accuracy.

Main Methods:

• Developed MtCro, a multi-task learning model for genomic prediction.
• Compared MtCro's performance against mainstream models like DNNGP and SoyDNGP using multiple plant datasets (Wheat2000, Wheat599, Maize8652).
• Analyzed the contribution of inter-phenotype correlations to prediction accuracy.

Main Results:

• MtCro achieved performance gains of 1-9% on Wheat2000, 1-8% on Wheat599, and 1-3% on Maize8652 compared to mainstream models.
• A consistent 2-3% improvement in multi-phenotype predictions was observed, highlighting the importance of inter-phenotype correlations.
• MtCro demonstrated enhanced model training efficiency and prediction accuracy.

Conclusions:

• Multi-task learning, as implemented in MtCro, effectively captures diverse plant phenotypes and their inter-correlations.
• MtCro significantly improves genomic prediction accuracy and breeding efficiency.
• The findings underscore the potential of multi-task learning to accelerate advancements in plant genetic breeding.