Coformer: a deep learning-based framework for cross-environment and multi-year cotton phenotype prediction and...

Area of Science:

• Genomics
• Plant Breeding
• Computational Biology

Background:

• Accurate prediction of cotton agronomic traits is vital for genetic improvement and breeding-by-design.
• High-dimensional genomic data (SNP markers) presents challenges like overfitting and poor generalization in traditional prediction models.

Purpose of the Study:

• To introduce Coformer, an innovative deep learning model for robust cotton phenotype prediction.
• To enhance the accuracy and interpretability of genomic prediction models in cotton breeding.

Main Methods:

• Developed Coformer, a hybrid Transformer-autoencoder deep learning model.
• Utilized a self-attention Transformer encoder for SNP dependency capture and genotype compression.
• Integrated a normalization module and linear projection layer for adaptive input processing and end-to-end training.

Main Results:

• Coformer demonstrated outstanding predictive robustness on a multi-environment, multi-year dataset, even without explicit environmental factor modeling.
• The model effectively improved generalization across varying data dimensionalities.
• Coformer successfully pinpointed key genetic loci influencing target traits, offering interpretability.

Conclusions:

• Coformer provides a powerful and interpretable deep learning framework for genomic prediction in cotton.
• The developed Cotton Phenotype Prediction System (CPPS) lowers the barrier for applying genomic prediction in breeding practices.
• This approach bridges genomics research and breeding practice, facilitating precision breeding in cotton.