GPTransformer: A Transformer-Based Deep Learning Method for Predicting Fusarium Related Traits in Barley

Area of Science:

• Agricultural Science
• Genetics
• Plant Pathology

Background:

• Fusarium head blight (FHB) and deoxynivalenol (DON) contamination pose significant threats to barley production and grain quality.
• Developing resistant barley cultivars is crucial, but limited resistant parent lines and labor-intensive breeding methods hinder progress.
• Genomic prediction offers a path to accelerate breeding for FHB and DON resistance.

Purpose of the Study:

• To evaluate machine learning algorithms, particularly a Transformer deep learning model, for predicting FHB and deoxynivalenol content in barley.
• To compare the performance of machine learning approaches against traditional statistical genomic prediction models.
• To investigate the impact of different single nucleotide polymorphic (SNP) marker encoding methods on prediction accuracy.

Main Methods:

• A diverse panel of 400 two-row spring barley lines was phenotyped for FHB and DON across multiple environments and genotyped using a 50K SNP array.
• Machine learning models, including a Transformer and a Residual Fully Connected Neural Network (RFCNN), were applied for genomic prediction.
• Single nucleotide polymorphic (SNP) markers were encoded categorically or using Hardy-Weinberg probability frequencies for model input.

Main Results:

• The Transformer-based deep learning model demonstrated comparable or superior performance to existing statistical and machine learning methods in predicting FHB and DON.
• Hardy-Weinberg encoding of SNP markers generally improved prediction correlation for both FHB (6.9%) and DON (9.6%) when used with the Transformer network.
• While including all markers showed marginal improvement, specific marker selection and encoding strategies significantly impacted prediction accuracy.

Conclusions:

• The Transformer deep learning algorithm presents a viable and potentially more effective alternative to traditional BLUP models for genomic prediction of complex traits like FHB and DON in barley.
• Machine learning approaches, especially with optimized marker encoding, hold significant potential for accelerating breeding programs aimed at enhancing crop resistance to diseases and mycotoxin contamination.
• This research provides a foundation for developing advanced genomic prediction tools to improve barley resilience and grain safety.