Embeddings from standardized sorghum leaf images capture variation in disease response that human scoring misses
- 1University of Nebraska-Lincoln
Abstract
Ordinal scoring of plant disease severity by human raters compresses variation in lesion color, size, and number. Inter-rater variability further complicates comparisons and integrated analyses across environments. We developed a low-cost portable imaging chamber to rapidly image large numbers of leaves under standardized lighting, orientation, and backdrops in the field and employed this system to image more than 11,000 leaves across three states. Embeddings from vision encoders predicted human-assigned disease severity scores. No significant GWAS hits were identified using human-assigned or vegetation-index-based disease severity scores, but GWAS using embeddings identified twelve genomic hotspots controlling leaf appearance. Nine were linked to variation in disease symptom severity. Five hotspots corresponded to previously characterized sorghum genes: all three hotspots not linked to disease and two of the nine that were. Roughly one-third of tested embedding--hotspot associations replicated across at least two states, and twenty replicated across all three. eQTL, PheWAS, and large-effect variant analyses identified single candidate genes with plausible mechanistic links to disease symptom severity for six of the seven hotspots not mapping to characterized genes. These results demonstrate the power of combining scalable, standardized leaf imaging with pretrained image encoders to capture genetically controlled variation in diverse disease symptoms that human ordinal scoring misses.
Related
- Comparative evaluation of methods for measuring standardized ileal digestibility of amino acids for different sorghum varieties in broilers.
- The Sorghum Lipid Database (SoLD): population-scale lipidomics linking environmental and genetic variation in the Sorghum Association Panel
- Pangenomics and machine learning reveal genetic variation to optimize carotenoids in sorghum grain
- Arbuscular mycorrhizal fungal communities in Jiang-flavor baijiu sorghum rhizosphere: environmental correlates and variation among production bases
- Investigating GERMs : how genotype, environment, and rhizosphere microbiome interactions underlie heat response in maize and sorghum
BibTeX
@misc{davis2026embeddings,
title = {{Embeddings from standardized sorghum leaf images capture variation in disease response that human scoring misses}},
author = {J. M. Davis and J. Turkus and S. Arora and K. M. Cuellar-Perez and L. F. Gomez-Trejo and R. Ruvalcaba-Ramirez and C. Ullagaddi and X. Kuang and S. Punnuri and S.-B. Kim and J. C. Schnable},
howpublished = {bioRxiv},
year = {2026},
month = sep,
doi = {10.64898/2026.09.21.753202},
url = {https://www.biorxiv.org/content/10.64898/2026.09.21.753202v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - Embeddings from standardized sorghum leaf images capture variation in disease response that human scoring misses
AU - J. M. Davis
AU - J. Turkus
AU - S. Arora
AU - K. M. Cuellar-Perez
AU - L. F. Gomez-Trejo
AU - R. Ruvalcaba-Ramirez
AU - C. Ullagaddi
AU - X. Kuang
AU - S. Punnuri
AU - S.-B. Kim
AU - J. C. Schnable
PY - 2026
DA - 2026/09/21
PB - bioRxiv
DO - 10.64898/2026.09.21.753202
UR - https://www.biorxiv.org/content/10.64898/2026.09.21.753202v1
AB - Ordinal scoring of plant disease severity by human raters compresses variation in lesion color, size, and number. Inter-rater variability further complicates comparisons and integrated analyses across environments. We developed a low-cost portable imaging chamber to rapidly image large numbers of leaves under standardized lighting, orientation, and backdrops in the field and employed this system to image more than 11,000 leaves across three states. Embeddings from vision encoders predicted human-assigned disease severity scores. No significant GWAS hits were identified using human-assigned or vegetation-index-based disease severity scores, but GWAS using embeddings identified twelve genomic hotspots controlling leaf appearance. Nine were linked to variation in disease symptom severity. Five hotspots corresponded to previously characterized sorghum genes: all three hotspots not linked to disease and two of the nine that were. Roughly one-third of tested embedding--hotspot associations replicated across at least two states, and twenty replicated across all three. eQTL, PheWAS, and large-effect variant analyses identified single candidate genes with plausible mechanistic links to disease symptom severity for six of the seven hotspots not mapping to characterized genes. These results demonstrate the power of combining scalable, standardized leaf imaging with pretrained image encoders to capture genetically controlled variation in diverse disease symptoms that human ordinal scoring misses.
ER -