Automated workflow for the cell cycle analysis of non-adherent and adherent cells using a machine learning approach

Reviewer #1 (Public Review):

Summary:

The manuscript proposes a series of steps using the FIJI environment, the authors have created a plugin for the initial steps of the process, merging images into an RGB stack, conversion to HSV, and then using brightness for reference and hue to distinguish the phases of the cycle. Then, the well-known Trackmate plugin was used to identify single cells and extract intensities. The data was further post-processed in R, where a series of steps, smoothing, scaling, and addressing missing frames were used to train a random forest. Hard-coded values of hue were used to distinguish G1, S, and G2/M. The process was validated with a score comparing the quality of the tracks and the authors reported the successful measure of the cell cycles.

Strengths:

The implementation of the pipeline seems easy, although it requires two separate platforms: Fiji and R. A similar approach could be implemented in a single programming environment like Python or Matlab and there would not be any need to export from one to the other. However, many labs have similar setups and that is not necessarily a problem.

Weaknesses:

I found two important weaknesses in the proposal:

(1) The pipeline relies on a large number of hard-coded conditions: size of Gaussian blur (Gaussian should be written in uppercase), values of contrast, size of filters, levels of intensity, etc. Presumably, the authors followed a heuristic approach and tried values of these and concluded that the ones proposed were optimal. A proper sensitivity analysis should be performed. That is, select a range of values of the variables and measure the effect on the output.

(2) Linked to the previous comments. Other researchers that want to follow the pipeline would have either to have exactly the same acquisition conditions as the manuscript or start playing with values and try to compensate for any difference in their data (cell diameter, fluorescent intensity, etc.) to see if they can match the results of the manuscript.

Reviewer #2 (Public Review):

Summary:

This paper presents an automated method to track individual mammalian cells as they progress through the cell cycle using the FUCCI system and applies the method to look at different tumor cell lines that grow in suspension and determine their cell cycle profile and the effect of drugs that directly affect the cell cycles, on progression through the cell cycle for a 72 hour period.

Strengths:

This is a METHODS paper. The one potentially novel finding is that they can identify cells that are at the G1-S transition by the change in color as one protein starts to go up and the other one goes down, similar to the change seen as cells enter G2/M.

Weaknesses:

They did not clearly indicate whether the G1/S cells are identified automatically or need to be identified by the person reviewing the data. In Figures 1 and S1, the movie shows cells with no color at a time corresponding to what is about the G1/S transition. Their assigned cell cycle phase is shown in Figure 1 but not in Figure S1. None of these pictures show the G1/S cells that they talk about being able to detect with a different color.

Author Response:

We greatly appreciate the insightful feedback provided by the reviewers and the editor on our manuscript titled "Automated workflow for the cell cycle analysis of non-adherent and adherent cells using a machine learning approach". We will provide a revised version of the manuscript aiming to address the comments and recommendations provided by the reviewers to enhance the quality and clarity of our work. In detail:

Reviewer #1 (Public Review):

Summary:

The manuscript proposes a series of steps using the FIJI environment, the authors have created a plugin for the initial steps of the process, merging images into an RGB stack, conversion to HSV, and then using brightness for reference and hue to distinguish the phases of the cycle. Then, the well-known Trackmate plugin was used to identify single cells and extract intensities. The data was further post-processed in R, where a series of steps, smoothing, scaling, and addressing missing frames were used to train a random forest. Hard-coded values of hue were used to distinguish G1, S, and G2/M. The process was validated with a score comparing the quality of the tracks and the authors reported the successful measure of the cell cycles.

Strengths:

The implementation of the pipeline seems easy, although it requires two separate platforms: Fiji and R. A similar approach could be implemented in a single programming environment like Python or Matlab and there would not be any need to export from one to the other. However, many labs have similar setups and that is not necessarily a problem.

Weaknesses:

I found two important weaknesses in the proposal:

(1) The pipeline relies on a large number of hard-coded conditions: size of Gaussian blur (Gaussian should be written in uppercase), values of contrast, size of filters, levels of intensity, etc. Presumably, the authors followed a heuristic approach and tried values of these and concluded that the ones proposed were optimal. A proper sensitivity analysis should be performed. That is, select a range of values of the variables and measure the effect on the output.

(2) Linked to the previous comments. Other researchers that want to follow the pipeline would have either to have exactly the same acquisition conditions as the manuscript or start playing with values and try to compensate for any difference in their data (cell diameter, fluorescent intensity, etc.) to see if they can match the results of the manuscript.

We thank Reviewer #1 for the insightful comments. We acknowledge the importance of ensuring the reproducibility and robustness of our pipeline among different sample types, acquisition conditions and, consequently, image S/N ratio and resolution. To address the concerns regarding the reliance on hard-coded conditions and the impact of varying parameter values on the output, we will complete the Methods section of the manuscript and the “Usage” section of the README file in the Github repository (https://github.com/ieoresearch/cellcycle-image-analysis) providing a summary of best practices that should be applied in the pre-processing part of the analysis. As an example, the usable image filters types and their settings related to cells with different size, fluorescence intensities and acquisition conditions will be analysed in detail and general guidelines will be provided.

Moreover, we will provide detailed documentation on the acquisition conditions required for reproducibility in the README file and Methods section.

For the Tracking Analysis part, we will refer to the well documented TrackMate tutorial to adapt the tracking analysis to different cell types, image resolution and intensities.

Reviewer #2 (Public Review):

Summary:

This paper presents an automated method to track individual mammalian cells as they progress through the cell cycle using the FUCCI system and applies the method to look at different tumor cell lines that grow in suspension and determine their cell cycle profile and the effect of drugs that directly affect the cell cycles, on progression through the cell cycle for a 72 hour period.

Strengths:

This is a METHODS paper. The one potentially novel finding is that they can identify cells that are at the G1-S transition by the change in color as one protein starts to go up and the other one goes down, similar to the change seen as cells enter G2/M.

Weaknesses:

They did not clearly indicate whether the G1/S cells are identified automatically or need to be identified by the person reviewing the data. In Figures 1 and S1, the movie shows cells with no color at a time corresponding to what is about the G1/S transition. Their assigned cell cycle phase is shown in Figure 1 but not in Figure S1. None of these pictures show the G1/S cells that they talk about being able to detect with a different color.

Thank you for your valuable feedback regarding the identification of G1/S cells in our pipeline. To clarify, the G1/S phase identification process is entirely automated within our pipeline. We apologize for any confusion caused by the lack of explicit indication in our manuscript. We will ensure to update the manuscript to clearly state that the identification of G1/S cells is performed automatically by our algorithm, eliminating the need for manual intervention.

Regarding the visualization of G1/S cells in Figures 1 and S1, we will revise the figures to include all the available frames referred to the G1/S transition. It's important to note that during this transition, fluorescence intensities for both the green and the red channels, are dimmer in comparison with their intensity levels during the G2/M transitions. This can result in frames that may seem visually darker, despite both colors coexisting at the same time point. In our revised figures, we will ensure to include all available frames relevant to the G1/S transition and provide a clearer representation of this phenomenon.

In response to Reviewer #2's recommendation, we plan to conduct additional experiments to further validate our observations. We will utilize the EdU technology to highlight the S-phase in FUCCI cells, allowing for better discrimination between the red and green fluorescence of the FUCCI reporter during the initial S-phase.

Additionally, we acknowledge that the link to the Docker container (https://hub.docker.com/r/emanuelsoda/rf_semi_sup) was not included in the manuscript. We apologize for this oversight, and it will be included in the revised version of the paper.