Enhancing prediction accuracy of key biomass partitioning traits in wheat using multi-kernel genomic prediction...

Achieving significant genetic gains in grain yield (GY) in wheat (Triticum aestivum L.) requires optimization of the key biomass partitioning traits such as spike partitioning index (SPI) and fruiting efficiency (FE). However, traditional manual phenotyping of these traits is labor-intensive and destructive, making it unsuitable for evaluating large germplasm panels. This study developed genomic prediction models to estimate these traits using diverse statistical methods while enhancing predictive ability (PA) by integrating environmental covariates (ECs) and secondary traits. A panel of 341 soft wheat elite lines was evaluated for biomass partitioning and yield-related traits from 2022 to 2024 in Citra, FL. Genomic best linear unbiased predictor (GBLUP) and Bayesian methods performed similarly or better than machine learning models for SPI, harvest index (HI), and GY. On the other hand, random forest models performed better in predicting effective tillers m^-2 (ET), 1000-grain weight (TGW), and grain numbers per m² (GN). Multi-kernel models incorporating ECs and secondary traits, such as plant height (PH) and aboveground biomass, substantially improved PA compared to genomics-only approaches. For 1000-grain weight, PA increased from 18% to 78%, with similar enhancements varying across other traits. Validations performed on separate breeding trial confirmed the reliability of the multi-kernel models, even though they showed a slightly lower PA compared to within-panel validations. These findings highlight the potential of integrating diverse data types or omics to enhance the prediction of biomass partitioning traits, speeding up genetic advancements, and the development of high-yield wheat varieties to address future food security challenges.