Multimodal deep learning improves cross-environment prediction of durum wheat yield components.

Area of Science:

• Agricultural Science
• Genetics
• Plant Breeding

Background:

• Durum wheat breeding faces challenges in predicting genotype performance under Mediterranean dryland conditions characterized by heat and drought.
• Accurate prediction of yield-related traits is crucial for developing climate-resilient durum wheat varieties.

Purpose of the Study:

• To evaluate a multimodal deep learning (MM-DL) framework for improving prediction accuracy of five key yield-related traits in durum wheat.
• To assess the contribution of genomic markers, environmental covariates (ECs), near-infrared spectroscopy (NIRS), and phenology to prediction accuracy.

Main Methods:

• A multimodal deep learning (MM-DL) framework was developed integrating genomic markers, ECs, NIRS, and phenology data.
• Multi-environment trial data from three contrasting seasons were used for model training and validation.
• Two prediction scenarios were employed: predicting an unseen year (PoY) and predicting new sowing environments within a season (PoSY).

Main Results:

• Integrating multiple data sources significantly improved prediction accuracy for all five traits (GN, GW, NS, SL, SW) compared to genomics alone.
• Spike length (SL) and number of spikelets per spike (NS) were the most predictable traits.
• Environmental covariates (ECs) were key for cross-season transferability, while NIRS and phenology provided trait-specific improvements.

Conclusions:

• MM-DL framework provides more stable and biologically informed predictions for durum wheat breeding.
• This approach offers a practical strategy for enhancing selection decisions in the face of increasing climate variability.
• The study highlights the potential of integrating diverse data types for advancing crop improvement programs.