Cropformer: An interpretable deep learning framework for crop genomic prediction.

Area of Science:

• Agricultural Science
• Genomics
• Computational Biology

Background:

• Genomic selection (GS) uses machine learning to accelerate crop breeding, but current deep learning models lack robustness and interpretability.
• Identifying superior genotypes and understanding genetic contributions to traits are crucial for crop improvement.

Purpose of the Study:

• To develop and evaluate Cropformer, a novel deep learning framework for enhanced genomic selection and trait-associated gene discovery.
• To improve the accuracy, robustness, and interpretability of predictive models in crop breeding.

Main Methods:

• Cropformer integrates convolutional neural networks with multiple self-attention mechanisms.
• The framework was evaluated on over 20 phenotypic traits across five major crops: maize, rice, wheat, foxtail millet, and tomato.
• Performance was benchmarked against existing genomic selection methods.

Main Results:

• Cropformer demonstrated superior prediction accuracy and robustness compared to other GS methods, with up to a 7.5% improvement.
• The framework successfully identified numerous single nucleotide polymorphisms (SNPs) associated with maize phenotypic traits.
• Key genetic variations underlying trait differences were revealed, enhancing gene mining capabilities.

Conclusions:

• Cropformer offers a significant advancement in predictive performance and gene identification for crop breeding.
• This general-purpose tool provides a powerful approach for improving genomic design and accelerating the development of superior crop varieties.
• The Cropformer framework is publicly accessible for broader research application.