A high-throughput pipeline for phenotyping, object detection and quantification of leaf trichomes

Abstract

We developed a new high-throughput device and AI image detection model capable of rapidly collecting phenotype data for a population of wild grass, facilitating identification of genomic regions associated with trichome density. Access to increasing amounts of high-quality genomic sequence data for many plant species is allowing for faster, more accurate gene identification. To maximize the use of this sequence data for association genetics, gene discovery, and validation, it must be coupled with phenotype data. However, phenotype data acquisition can represent a bottleneck in studies requiring many datapoints, such as large diversity panels for genome-wide association studies. Here we developed a portable handheld imaging device-the Tricocam-and method for image capture and semi-automated quantification of leaf edge trichomes in a grass species. Trichomes have been implicated in abiotic and biotic stress tolerance in grasses, but so far, no trichome genes have been cloned in this plant family. We also refined and implemented the AI detection processes underpinning the web-based image quantification platform from Thya Technology® to rapidly quantify leaf edge trichomes. We used the phenotype acquisition method in the wild wheat progenitor Aegilops tauschii in combination with k-mer-based Genome-Wide Association Study to validate a trichome-controlling genomic region on chromosome arm 4DL and discover a new one on 4DS. By making the Tricocam 3D print design and AI visual detection model public, we aim to deliver useful resources for the plant science community to use or adapt for other large-scale phenotyping projects on diversity panels.

Fig. 1

High-throughput and non-destructive leaf imaging in Aegilops tauschii using the Tricocam. a Diagram of device structure. b Tricocam assembly process. c Assembled Tricocam. d Phenotyping process with Tricocam

Fig. 2

Development of object detection algorithm and augmentation methods. a Automated detection of leaf edge trichomes in Aegilops tauschii. Detected trichomes are indicated by red boxes. b Comparative analysis of the two different AI detection models, Faster R-CNN and YOLOv8. mAP(0.5) considers a single threshold for the IoU set to 0.5, mAP(0.5:0.95) averages over 10 stricter IoU threshold ranging from 0.5 to 0.95, with intermediate steps of 0.05. c Analysis of effects of image augmentation on detection accuracy with Faster R-CNN. d Comparative analysis of image augmentation with YOLOv8

Fig. 3

Leaf trichome density distribution in Aegilops tauschii and k-mer-based GWAS for association genetics. a Leaf edge trichome density distribution in 358 Lineage 1 and 203 Lineage 2 accessions. Scoring was done using the Tricocam and AI-based phenotyping system with manual correction. b Leaf edge trichome density distribution in datasets generated by Gaurav et al. (2022) and using the Tricocam and AI-based phenotyping system with manual correction. c Leaf edge trichome density distribution in 203 Lineage 2 accessions based on uncorrected and manually corrected datasets generated using the Tricocam and AI-based phenotyping system. d k-mer association mapping using uncorrected trichome density data for 140 non-redundant L2 accessions generated with the Tricocam and AI-based phenotyping system. e k-mer association mapping using manually corrected trichome density data for 140 non-redundant L2 accessions generated with the Tricocam and AI-based phenotyping system. f k-mer association mapping using trichome density data for 116 non-redundant L2 accessions taken from Gaurav et al. (2022). g k-mer association mapping using corrected trichome density data generated with Tricocam and AI-based phenotyping for the same set of 116 non-redundant L2 accessions used in Gaurav et al. (2022). For the Manhattan plots in d-g, the x-axis shows the seven chromosomes of the Aegilops tauschii genome and the y-axis shows the significance values of the associated k-mers mapped to the genome. Blue peaks indicate a positive correlation with the phenotype, while red peaks indicate a negative correlation. The horizontal dashed line indicates the significance threshold value based on Bonferroni correction. The association mapping was executed against the genome assembly of accession TOWWC0023, displaying high leaf edge trichome density (indicated by arrows in the histograms in a, b, and c)