Near-infrared spectroscopy-based phenomic prediction is competitive with genomic prediction in sorghum

Abstract Genomic selection is well established in many crop improvement programs, and the inclusion of phenomic data to prediction models is increasing. Applications of phenomic data for predictive methods lag because genotypes must be grown whereas genomic prediction is applied to untested genotypes. To address this drawback, we propose a low-cost approach of phenomic selection strategies using Near Infrared Spectroscopy (NIRS) data collected on inbred parental lines. For this study, 366 grain sorghum hybrids, developed from a partial factorial, were evaluated across four total environments between 2019 and 2020. The inbred parents of these hybrids were evaluated in the 2020 environments. Both hybrids and inbreds were phenotyped for grain yield, days to anthesis, and plant height and grain samples were scanned using NIRS bandwidths from 400–2500 nm to estimate compositional traits such as starch, protein, fat, fiber, and ash. Parents were genotyped using genotyping by sequencing which produced 38,234 high-quality SNP markers. Various prediction models incorporating these genomic, phenomic and both data were trained. Phenomic prediction models trained using NIRS obtained on parents ( M2 ) and mid-parent heterosis ( M3 ) of NIRS were similar, and in most cases competitive to genomic prediction ( M1 ) in predicting hybrids from un-testcrossed lines ( CVa ) and untested hybrids ( CVb ). Linear discriminant analysis using NIR spectra of grain samples explained majority of variance and was able to characterize environments and heterotic groups accurately. NIRS-based phenomic prediction could be implemented as low-cost alternative to genomic prediction to predict sorghum hybrid performance across environments.