Machine Learning Enables High-Throughput Phenotyping for Analyses of the Genetic Architecture of Bulliform Cell Patterning in Maize

Abstract

Bulliform cells comprise specialized cell types that develop on the adaxial (upper) surface of grass leaves, and are patterned to form linear rows along the proximodistal axis of the adult leaf blade. Bulliform cell patterning affects leaf angle and is presumed to function during leaf rolling, thereby reducing water loss during temperature extremes and drought. In this study, epidermal leaf impressions were collected from a genetically and anatomically diverse population of maize inbred lines. Subsequently, convolutional neural networks were employed to measure microscopic, bulliform cell-patterning phenotypes in high-throughput. A genome-wide association study, combined with RNAseq analyses of the bulliform cell ontogenic zone, identified candidate regulatory genes affecting bulliform cell column number and cell width. This study is the first to combine machine learning approaches, transcriptomics, and genomics to study bulliform cell patterning, and the first to utilize natural variation to investigate the genetic architecture of this microscopic trait. In addition, this study provides insight toward the improvement of macroscopic traits such as drought resistance and plant architecture in an agronomically important crop plant.

Keywords: GWAS; Machine learning; bulliform cell; development.