nf-root: a best-practice pipeline for deep learning-based analysis of apoplastic pH in microscopy images of developmental zones in plantroot tissue

Abstract

Here we report nextflow-root (nf-root), a novel best-practice pipeline for deep learning-based analysis of fluorescence microscopy images of plant root tissue, aimed at studying hormonal mechanisms associated with cell elongation, given the vital role that plant hormones play in the development and growth of plants. This bioinformatics pipeline performs automatic identification of developmental zones in root tissue images, and analysis of apoplastic pH measurements of tissue zones, which is useful for modeling plant hormone signaling and cell physiological responses. Mathematical models of physiological responses of plant hormones, such as brassinolide, have been successfully established for certain root tissue types, by evaluating apoplastic pH via fluorescence imaging. However, the generation of data for this modeling is time-consuming, as it requires the manual segmentation of tissue zones and evaluation of large amounts of microscopy data. We introduce a high-throughput, highly reproducible Nextflow pipeline based on nf-core standards that automates tissue zone segmentation by implementing a deep-learning module, which deploys deterministically trained (i.e. bit-exact reproducible) convolutional neural network models, and augments the segmentation predictions with measures of prediction uncertainty and model interpretability, aiming to facilitate result interpretation and verification by experienced plant biologists. To train our segmentation prediction models, we created a publicly available dataset composed of confocal microscopy images of A. thaliana root tissue using the pH-sensitive fluorescence indicator, and manually annotated segmentation masks that identify relevant tissue zones. We applied this pipeline to analyze exemplary data, and observed a high statistical similarity between the manually generated results and the output of nf-root. Our results indicate that this approach achieves near human-level performance, and significantly reduces the time required to analyze large volumes of data, from several days to hours.