APTES: a high-throughput deep learning–based Arabidopsis phenotypic trait estimation system for individual leaves and siliques

TL;DR

APTES is a deep learning system that automatically estimates plant traits from images, improving accuracy and enabling high-throughput analysis in plant science.

Contribution

APTES introduces a high-throughput deep learning system for precise leaf and silique trait estimation in Arabidopsis with improved segmentation accuracy.

Findings

APTES achieved high precision and recall scores for leaf and silique segmentation using enhanced deep learning models.

Genome-wide association study identified 1,042 SNPs significantly associated with leaf and silique traits in Arabidopsis.

APTES demonstrated applicability across diverse datasets and plant species beyond Arabidopsis.

Abstract

High-throughput phenotyping of growth kinetics and organ size in the model plant Arabidopsis thaliana requires rapid and precise methods for trait estimation. To address this need, we developed the Arabidopsis Phenotypic Trait Estimation System, APTES, an open-access, high-throughput program that uses computer vision and deep learning to extract 64 leaf traits and 64 silique traits from photographs. The enhanced segmentation model Cascade Mask Region-based Convolutional Neural Network (Mask R-CNN) achieved precision (measure of positive prediction accuracy), recall (sensitivity in detection), and F1 score values (harmonic mean of precision and recall) of 0.965, 0.958, and 0.961, respectively, for individual leaf segmentation. These metrics demonstrated a consistent improvement of approximately 1 percentage point over the baseline model. For silique segmentation, our enhanced DetectoRS model for silique segmentation attained precision, recall, and F1 scores of 0.954, 0.930, and 0.942, respectively. Notably, precision increased by 1%, while the F1 score improved by 2 percentage points. Trait parameters were automatically calculated with coefficient of determination values for leaf and silique traits ranging from 0.776 to 0.976 and mean absolute percentage error values from 1.89% to 7.90%. We phenotyped 166 Arabidopsis accessions, using APTES, and subjected the resulting values to a genome-wide association study (GWAS), revealing 1,042 single-nucleotide polymorphisms (SNPs) as being significantly associated with 18 leaf and silique traits, and one significant SNP on chromosome 3 linked to silique number. Furthermore, we validated APTES across other public Arabidopsis databases and other plant species, with segmentation results demonstrating its applicability across diverse datasets. In conclusion, APTES is a valuable automated tool for leaf and silique segmentation and trait estimation, which should offer benefits to the broader plant science community.

The online version contains supplementary material available at 10.1007/s42994-025-00239-y.

Figures