DetecDiv, a generalist deep-learning platform for automated cell division tracking and survival analysis
Curation statements for this article:-
Curated by eLife
Evaluation Summary:
This article introduces "DetecDiv", a high throughput, deep learning method to perform automated cell-division tracking in yeast. The performance of the method, estimated to be 100 times faster than manual annotation, overcomes current time processing limitations that are inherent to large single cell datasets. In particular, DetecDiv allows to automatically get quantitative measurements of replicative life span in yeast. The method is of broad interest for quantitative biology as it can be used to study yeast cells dynamics across their lifespan.
(This preprint has been reviewed by eLife. We include the public reviews from the reviewers here; the authors also receive private feedback with suggested changes to the manuscript. Reviewer #3 agreed to share their name with the authors.)
This article has been Reviewed by the following groups
Listed in
- Evaluated articles (eLife)
- Cell Biology (eLife)
Abstract
Automating the extraction of meaningful temporal information from sequences of microscopy images represents a major challenge to characterize dynamical biological processes. So far, strong limitations in the ability to quantitatively analyze single-cell trajectories have prevented large-scale investigations to assess the dynamics of entry into replicative senescence in yeast. Here, we have developed DetecDiv, a microfluidic-based image acquisition platform combined with deep learning-based software for high-throughput single-cell division tracking. We show that DetecDiv can automatically reconstruct cellular replicative lifespans with high accuracy and performs similarly with various imaging platforms and geometries of microfluidic traps. In addition, this methodology provides comprehensive temporal cellular metrics using time-series classification and image semantic segmentation. Last, we show that this method can be further applied to automatically quantify the dynamics of cellular adaptation and real-time cell survival upon exposure to environmental stress. Hence, this methodology provides an all-in-one toolbox for high-throughput phenotyping for cell cycle, stress response, and replicative lifespan assays.
Article activity feed
-
-
Author Response
Reviewer #3 (Public Review):
In this work, the authors describe a novel method, based on deep learning, to analyze large numbers of yeast cells dividing in a controlled environment. The method builds on existing yeast cell trapping microfluidic devices that have been used for replicative lifespan assay. The authors demonstrate how an optimized microfluidic device can be coupled with deep learning methods to perform automatic cell division tracking and single cell trajectories quantification. The overall performance of the method is impressive: it allows to deal with large image datasets generated by timelapse microscopy several order of magnitudes faster than what manual annotation would require. The method has been carefully tested on several microscopy settings and datasets and compared with known results from the …
Author Response
Reviewer #3 (Public Review):
In this work, the authors describe a novel method, based on deep learning, to analyze large numbers of yeast cells dividing in a controlled environment. The method builds on existing yeast cell trapping microfluidic devices that have been used for replicative lifespan assay. The authors demonstrate how an optimized microfluidic device can be coupled with deep learning methods to perform automatic cell division tracking and single cell trajectories quantification. The overall performance of the method is impressive: it allows to deal with large image datasets generated by timelapse microscopy several order of magnitudes faster than what manual annotation would require. The method has been carefully tested on several microscopy settings and datasets and compared with known results from the literature in a convincing manner. In addition, the authors show how the analysis pipeline can be enriched with semantic segmentation to quantify cellular physiology and gene expression during their lifespan, creating high quality, high throughput measurements of single cell trajectories. The software, its documentation and related datasets are available through public repository. Taken together, the author succeeded in setting up a method that can be a game changer for high throughput longitudinal analysis of yeast cells.
Overall, the method seems robust and powerful but some aspects need to be clarified and/or extended.
- The authors chose MATLAB to develop DetecDiv. This is a valid choice but as Python is becoming the standard for deep learning developments it is important to 1/ better justify the use of MATLAB and 2/ discuss how this can be "translated into" and/or linked with Python. This would facilitate adoption by other research teams.
Using MATLAB as a prototyping language was instrumental for us in establishing the proof of principle of the method reported here since we have a long-standing experience in MATLAB programming. Yet, we fully agree with the reviewer that Python may appear as a more legitimate choice, especially in the field of deep learning. For future work, we are considering moving our code to Python to make it more widely accessible and to more quickly benefit from the latest development in the field. We also envision that Python developers could transpose our methods for their own research interests.
Last, we note that MATLAB has bidirectional communication with a number of programming languages, including Python. Therefore, it is currently possible to use Python scripts to fully control the DetecDiv pipelines by calling its low levels functions at the command line. Obviously, it may be more cumbersome to use than native Python code and it restricts the possibility to use the graphical user interface that we have developed.
- A critical aspect of deep learning methods is their potential ability to be used on a different datasets and/or experimental setup (transfer learning). The authors explained that a "generalist" model, trained using several datasets perform comparably (or even better) than "specialist" models that are independently trained on a specific dataset. Yet, they do not discuss how accurate would an already trained generalist model perform on a novel dataset made with a different imaging setup and/or a different yeast strain?
We thank the reviewer for this comment, which is somewhat related to point #2 raised by reviewer #1, regarding the ability of a model to generalize its prediction to various contexts. In the revised version, we now provide clear evidence that the model designed for division counting and RLS analysis, which is trained on WT data only, can successfully predict the lifespan of mutants such as fob1delta and sir2delta (new Figure 2 - Figure supplement 5) and the onset of cell death during stress response assays (Figure 6 - Figure supplement 1).
However, changing the imaging conditions (e.g. magnification, illumination, etc) would quickly deteriorate the performance of the model, unless it has been exposed to these new conditions upon training. Hence, the purpose of the ‘generalist’ approach we use is to demonstrate that the models we use have the capacity to deal with various imaging conditions when appropriately trained.
We have added a sentence in the discussion to explain what determines the potential of a model to be successfully employed in different contexts.
-
Evaluation Summary:
This article introduces "DetecDiv", a high throughput, deep learning method to perform automated cell-division tracking in yeast. The performance of the method, estimated to be 100 times faster than manual annotation, overcomes current time processing limitations that are inherent to large single cell datasets. In particular, DetecDiv allows to automatically get quantitative measurements of replicative life span in yeast. The method is of broad interest for quantitative biology as it can be used to study yeast cells dynamics across their lifespan.
(This preprint has been reviewed by eLife. We include the public reviews from the reviewers here; the authors also receive private feedback with suggested changes to the manuscript. Reviewer #3 agreed to share their name with the authors.)
-
Reviewer #1 (Public Review):
Microfluidics-based live-cell imaging is a powerful technique that can reveal detailed quantitative insights on for example cell growth, cell cycle, and - if coupled with fluorescent markers - molecular processes. Especially for fast growing unicellular organisms such as yeast, high-throughput imaging of multiple strains or conditions is possible over many generations. This allows biologists to quickly obtain hundreds of videos in a relatively short time-span, making the image analysis to extract useful information the bottleneck. Recent progress on convolutional neural networks such as UNet has made a strong impact on the quality of automated segmentation. However, to extract useful information, additional time-consuming steps are still necessary, which limits high-throughput experiments. With the present …
Reviewer #1 (Public Review):
Microfluidics-based live-cell imaging is a powerful technique that can reveal detailed quantitative insights on for example cell growth, cell cycle, and - if coupled with fluorescent markers - molecular processes. Especially for fast growing unicellular organisms such as yeast, high-throughput imaging of multiple strains or conditions is possible over many generations. This allows biologists to quickly obtain hundreds of videos in a relatively short time-span, making the image analysis to extract useful information the bottleneck. Recent progress on convolutional neural networks such as UNet has made a strong impact on the quality of automated segmentation. However, to extract useful information, additional time-consuming steps are still necessary, which limits high-throughput experiments. With the present manuscript, Aspert et al. now make an important step towards filling this gap by establishing a fully automated approach to extract biological information on the replicative life-span of yeast cells from experiments performed with dedicated microfluidics devices that retain mother cells over multiple generation while 'washing out' newborn daughter cell.
In their work, Aspert et al. take an innovative approach of using convolutional neural networks to classify images of single traps according to whether the mother cell is in G1, early budded phase, late budded phase, or dead. In addition, two classes of empty and crowded traps are used to clean up the results. This initial classification is then combined with an LSTM to predict cell cycle transitions over complete life-times of mother cells. As a proof-of-principle, they then also combined this approach with semantic segmentation to extract cellular features of the mother cells. In addition to the computational developments, the study also suggests a cheap experimental setup that makes using this novel image analysis routine affordable.
Overall, this is an interesting and well-executed study that opens new territory of using AI for yeast live-cell imaging approaches. The main focus of this study is clearly to develop a functional assay, all the way from experiment to data analysis. The authors put effort into providing a tool that works for diverse optical setups. To achieve this aim, at some points pragmatic decisions were made, in particular with regards to the neural networks used. While these decisions are reasonable, it still leaves open the possibility that even better performance could be achieved with other state-of-the art approaches. -
Reviewer #2 (Public Review):
The authors propose a pipeline for investigation and analysis of cell devision which incorporates both hardware and software solutions. The described microfluidic image acquisition system allows for imaging single cell divisions in a very efficient way, using different types of imaging devices. The classification and analysis of the division events are done by employing modern neural network solutions. Here, the authors present a DetecDiv package, capable of automated classification of events and creation of survival curves for individual cells. This drastically speeds up the research compared to the manual analysis. The strong points include adaptation of the proposed pipeline to different imaging systems and microfluidic device geometries. The comparison of several neural network architectures is also …
Reviewer #2 (Public Review):
The authors propose a pipeline for investigation and analysis of cell devision which incorporates both hardware and software solutions. The described microfluidic image acquisition system allows for imaging single cell divisions in a very efficient way, using different types of imaging devices. The classification and analysis of the division events are done by employing modern neural network solutions. Here, the authors present a DetecDiv package, capable of automated classification of events and creation of survival curves for individual cells. This drastically speeds up the research compared to the manual analysis. The strong points include adaptation of the proposed pipeline to different imaging systems and microfluidic device geometries. The comparison of several neural network architectures is also convincing. The considered biological application support the conclusions of the initial research questions, even though those are not strongly motivated.
-
Reviewer #3 (Public Review):
In this work, the authors describe a novel method, based on deep learning, to analyze large numbers of yeast cells dividing in a controlled environment. The method builds on existing yeast cell trapping microfluidic devices that have been used for replicative lifespan assay. The authors demonstrate how an optimized microfluidic device can be coupled with deep learning methods to perform automatic cell division tracking and single cell trajectories quantification. The overall performance of the method is impressive: it allows to deal with large image datasets generated by timelapse microscopy several order of magnitudes faster than what manual annotation would require. The method has been carefully tested on several microscopy settings and datasets and compared with known results from the literature in a …
Reviewer #3 (Public Review):
In this work, the authors describe a novel method, based on deep learning, to analyze large numbers of yeast cells dividing in a controlled environment. The method builds on existing yeast cell trapping microfluidic devices that have been used for replicative lifespan assay. The authors demonstrate how an optimized microfluidic device can be coupled with deep learning methods to perform automatic cell division tracking and single cell trajectories quantification. The overall performance of the method is impressive: it allows to deal with large image datasets generated by timelapse microscopy several order of magnitudes faster than what manual annotation would require. The method has been carefully tested on several microscopy settings and datasets and compared with known results from the literature in a convincing manner. In addition, the authors show how the analysis pipeline can be enriched with semantic segmentation to quantify cellular physiology and gene expression during their lifespan, creating high quality, high throughput measurements of single cell trajectories. The software, its documentation and related datasets are available through public repository. Taken together, the author succeeded in setting up a method that can be a game changer for high throughput longitudinal analysis of yeast cells.
Overall, the method seems robust and powerful but some aspects need to be clarified and/or extended.
- The authors chose MATLAB to develop DetecDiv. This is a valid choice but as Python is becoming the standard for deep learning developments it is important to 1/ better justify the use of MATLAB and 2/ discuss how this can be "translated into" and/or linked with Python. This would facilitate adoption by other research teams.
- A critical aspect of deep learning methods is their potential ability to be used on a different datasets and/or experimental setup (transfer learning). The authors explained that a "generalist" model, trained using several datasets perform comparably (or even better) than "specialist" models that are independently trained on a specific dataset. Yet, they do not discuss how accurate would an already trained generalist model perform on a novel dataset made with a different imaging setup and/or a different yeast strain? -
Reviewer #4 (Public Review):
The paper describes the development of cutting edge microfluidics, imaging, and image analysis tools to allow the uninterrupted observation of many yeast life trajectories under well controlled conditions.
The authors improve the traps to avoid several problems that other systems had, such as the loss of cells (especially the larger cells) and clogging. They then developed a pipeline to automatically extract the patterns of cell division (start of cell division, end of lifespan) using a convolutional neural network to extract representative image features (unbudded, small or large budded, dead, clogged or empty) and taking into account the sequence of these features using a long short-term memory network. They thoroughly compare their new pipeline to several other previously published.
The method presents an …
Reviewer #4 (Public Review):
The paper describes the development of cutting edge microfluidics, imaging, and image analysis tools to allow the uninterrupted observation of many yeast life trajectories under well controlled conditions.
The authors improve the traps to avoid several problems that other systems had, such as the loss of cells (especially the larger cells) and clogging. They then developed a pipeline to automatically extract the patterns of cell division (start of cell division, end of lifespan) using a convolutional neural network to extract representative image features (unbudded, small or large budded, dead, clogged or empty) and taking into account the sequence of these features using a long short-term memory network. They thoroughly compare their new pipeline to several other previously published.
The method presents an important methodological innovation that goes far beyond previous developments in this field. If widely implemented - which seems possible due to the transparent and detailed methods sections, the availability of software developed, and the relatively low costs - the methods described can be transformative to many research fields that need high-throughput phenotyping of processes related to cell cycle, stress response, and replicative ageing.
I find that the work is of outstanding quality.
-
-
-
-