YeastMate: neural network-assisted segmentation of mating and budding events in Saccharomyces cerevisiae

Abstract Summary Here, we introduce YeastMate, a user-friendly deep learning-based application for automated detection and segmentation of Saccharomyces cerevisiae cells and their mating and budding events in microscopy images. We build upon Mask R-CNN with a custom segmentation head for the subclassification of mother and daughter cells during lifecycle transitions. YeastMate can be used directly as a Python library or through a standalone application with a graphical user interface (GUI) and a Fiji plugin as easy-to-use frontends. Availability and implementation The source code for YeastMate is freely available at https://github.com/hoerlteam/YeastMate under the MIT license. We offer installers for our software stack for Windows, macOS and Linux. A detailed user guide is available at https://yeastmate.readthedocs.io. Supplementary information Supplementary data are available at Bioinformatics online.


Introduction
An important experimental approach when working with the budding yeast Saccharomyces cerevisiae is to examine the effects of mutations on morphological features or protein localization either by brightfield or fluorescent microscopy (Ohya et al., 2015). Such experiments can also be performed in a systematic manner by combining automated microscopy with strain libraries comprising thousands of yeast strains that carry gene deletions or express fluorescently tagged proteins (Giaever et al., 2002;Huh et al., 2003;Weill et al., 2018), which, however, calls for robust automated image analysis pipelines. In recent years, tools based on convolutional neural networks (CNNs) have become state-of-the-art for many tasks in biomedical image analysis (von Chamier et al., 2019), including segmentation of individual S.cerevisiae cells (Dietler et al., 2020;Lu et al., 2019;Salem et al., 2021). Many experimental strategies facilitated by the yeast system also make use of specific transitions in the yeast lifecycle like budding and mating to study things like organelle inheritance and mitochondrial quality control (Jakubke et al., 2021;Rafelski et al., 2012). Detection of matings and buddings is often done by hand or through dedicated postprocessing routines on the output of a single-cell segmentation tool (e.g. tracking of cells in a time series of images), highlighting the need for easy-to-use end-to-end solutions for these more complex tasks.
Here, we present YeastMate, a novel deep learning-based tool for end-to-end segmentation of single cells and detection of transitions in the lifecycle of S.cerevisiae in single transmitted light images.
YeastMate performs three tasks: instance segmentation of single cells, object detection of zygotes and budding events and automatic assignment of mother and daughter cells involved in a mating or budding event. The detection backend is based on Mask R-CNN (He et al., 2017) and is complemented by a user-friendly frontend using modern web technologies as well as a Fiji plugin. In the task of detecting mating and budding events, we achieve accuracies comparable to manual human reannotation and we can also perform singlecell segmentation robustly across various datasets. YeastMate is already being used in ongoing research in our lab (Jakubke et al., 2021). In addition to the software, we also provide a new dataset of images with manually annotated cells and mating and budding events.

Materials and methods
CNN architecture and training: Our network architecture builds upon Mask R-CNN with a modified mask segmentation head producing multiclass semantic segmentations ( Supplementary Fig. S1) for each detected object. Instead of individually detecting mother and daughter cells, we first detect the whole mating and budding events as well as single cells. In a postprocessing step, we resolve the roles of the cells involved in budding or mating events based on the multiclass segmentation masks (Supplementary Fig. S2). We used 80% of our images for training our network and cross-validation and 20% as a hold-out test set for the final performance assessment.
Software architecture: YeastMate is implemented in a modular way: the detection backend can be used as a Python library but also runs as a webservice to provide its capabilities to client applications. As clients, we provide a Fiji plugin that directly interfaces with the server via HTTP requests as well as a standalone GUI Desktop application ( Fig. 1 and Supplementary Figs S3-S5). We provide the whole YeastMate stack as a single installable file for use on a local workstation.
For more details, please refer to the Supplementary Methods.

Results
Dataset: For training YeastMate, we collected 147 brightfield and differential interference contrast (DIC) images of S.cerevisiae acquired on two different microscopes at various imaging conditions and generated curated single-cell segmentation masks and mating and budding annotations. In total, our data contain 17 058 individual cells, 3615 buddings and 2380 zygotes (Supplementary Table  S1). Object detection performance: Our network achieves a mean average precision (mAP) of 0.878 as well as favourable APs for the individual classes of objects when applied to our test set ( Fig. 1B and Supplementary Fig. S6). We also assessed interhuman reproducibility by comparing the annotations for mating and budding events of four different annotators (with the most experienced one chosen as the reference). On our test dataset, mean interhuman precision and recall are (0.908, 0.946) for mating events and (0.774, 0.763) for budding events. These values lie close to the PRcurves of our network, indicating that YeastMate can achieve object detection performance comparable to manual human annotation.
Single-cell segmentation performance: YeastMate also compares favourably to existing CNN-based solutions in single-cell segmentation, showing consistent performance not only on our own test dataset, but also two publicly available datasets (Fig. 1C, Supplementary Results, Supplementary Fig. S7 and Table S2). Based on the robust Mask R-CNN architecture, YeastMate achieves the highest singlecell detection performance of all tools as well as competitive results in segmentation.

Conclusion
With YeastMate, we introduce an easy-to-use application to not only perform single-cell segmentation in images of S.cerevisiae with high robustness across datasets, but also detect transitions in the cell cycle in single images with accuracies comparable to human annotation. YeastMate is implemented in a modular way and can be run on local workstations but the detection server can also be run on a remote compute server. Additionally, we provide two user-friendly frontends to make the tool available without the need to write code. We can envision YeastMate to be expanded to detect other life cycle states, such as meiotic asci, similar to existing work in Arabidopsis thaliana (Lim et al., 2020). YeastMate can provide a considerable improvement to high-throughput studies of yeast, facilitating not only the automated analysis of large image datasets of single cells, but also enabling the study the complex interplay of cellular components during both sexual and asexual reproduction of S.cerevisiae.