Protocol for AI-assisted quantitative analysis and setup of tumor spheroid invasion into tissue

Summary

Recent work has revealed an active mechanical role for the tissue barrier in collective invasion, offering a new perspective on tissue invasion and morphogenesis. Here, we present a protocol for AI-assisted quantitative analysis and setup of tumor spheroid invasion into mesothelial tissue. We describe steps for co-culture setup, time-lapse imaging, AI training, and segmentation. We then detail procedures for cell tracking and quantitative analysis to study the fracturing process of the mesothelium.

For complete details on the use and execution of this protocol, please refer to Wu et al.¹

Graphical abstract

Highlights

• •3D model system to examine cancer spheroid invasion dynamics in the mesothelium
• •Simulation of mesothelial clearance during peritoneal metastatic tumor invasion
• •Using AI tools to segment nuclei for spheroid cell dynamics analysis

Publisher’s note: Undertaking any experimental protocol requires adherence to local institutional guidelines for laboratory safety and ethics.

Recent work has revealed an active mechanical role for the tissue barrier in collective invasion, offering a new perspective on tissue invasion and morphogenesis. Here, we present a protocol for AI-assisted quantitative analysis and setup of tumor spheroid invasion into mesothelial tissue. We describe steps for co-culture setup, time-lapse imaging, AI training, and segmentation. We then detail procedures for cell tracking and quantitative analysis to study the fracturing process of the mesothelium.

Before you begin

Visualizing the motility of spheroids in a three-dimensional model helps us understand how collective migration is organized.¹ Here, we use a live heterotypic culture model system to observe real-time interactions between tumor spheroids (human ovarian adenocarcinoma OVCA433 epithelial cells) and mesothelial cells (ZT human benign pleural effusion mesothelial cells),² mimicking mesothelial clearance during ovarian tumor metastasis.

Live 3D imaging was performed on spheroid cells expressing mNeon GFP in the nuclei.¹ Data obtained from tracking these spheroids using time-lapse imaging can be used to analyze cell motility dynamics by measuring several parameters, including the mean squared displacement of cells over time and their velocity patterns. AI training and segmentation based on StarDist³ were used to segment and recognize nuclei in cells within cell collectives. This helps to detect and segment individual nuclei within each spheroid in 3D image stacks more accurately, drastically reducing the time needed for user operation and errors.

Since this protocol is optimized for use with ZT mesothelial and OVCA433 epithelial cell lines, incubation times may require further adjustments based on the adherence level of the cell lines used.

Key resources table

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Biological samples
Human ovarian adenocarcinoma OVCA433 epithelial cells (RRID:CVCL_0475)	Laboratory of Michael Sheetz	N/A
ZT human benign pleural effusion mesothelial cells	Laboratory of Tan A. Ince	N/A
Chemicals, peptides, and recombinant proteins
10× Phosphate-buffered saline (PBS)	1st BASE	Cat# BUF-2040-10X1L
Fetal bovine serum, heat inactivated, certified, One Shot, United States	Gibco	Cat# A3840001
Bovine serum albumin (BSA)	Sigma-Aldrich	Cat#A9647
Glycine	1st BASE	Cat#BIO-2085
Triton X-100	Sigma-Aldrich	Cat# 1.08603
Human plasma fibronectin purified protein	Sigma-Aldrich	Cat#FC010
MCDB 105 Medium	Sigma-Aldrich	Cat#117-500
Medium 199, no phenol red	Thermo Fisher Scientific	Cat#11043023
TrypLE Express Enzyme (1×), no phenol red	Gibco	Cat# 12604039
Cell dissociation buffer, enzyme-free, PBS	Thermo Fisher Scientific	Cat#13151014
Geltrex LDEV-free, hESC-qualified, reduced growth factor basement membrane matrix	Gibco	Cat#A1413301
Software and algorithms
Imaris 9.7 or higher	Oxford Instruments	RRID: SCR_007370
Fiji (ImageJ)	Schindelin et al.4; Schneider et al.5	RRID:SCR_002285https://imagej.net/software/fiji/downloads
3D StarDist	Martin Weigert et al.	https://github.com/stardist/stardist
MetaMorph Microscopy 7.10.5.476	Molecular Devices	RRID:SCR_002368
Other
Nikon ECLIPSE Ti-2 Inverted Microscope Systems	Nikon	Eclipse Ti2-U
Iwaki 12 mm Glass Base dishes	AGC Techno Glass (Iwaki)	Cat#3911-035
Falcon 15 mL High Clarity PP centrifuge tube, conical bottom, with dome seal screw cap, sterile	Falcon	Cat#352096

Materials and equipment

Image acquisition system

• •The Nikon Eclipse Ti2-E Inverted Microscope W1 Spinning Disk rapidly acquired a wide range of Z-stacks within the stipulated time intervals. Image acquisition was performed with the Nikon Perfect Focus System, which maintains focus by detecting and tracking the position of the coverslip surface in real-time. This system corrects potential focal drifts throughout multidimensional acquisition over the time-lapse.
• •A 40× silicone immersion oil objective, numerical aperture 1.406, was selected. 40× silicone objective allows for a greater capture range, resulting in a full 3D view of the spheroids.
• •
  Dichroic and emitters used in this study are:

  • ○
    GFP filters (525/40 nm single-band bandpass filter).
  • ○
    RFP filters (607/36 nm single-band bandpass filter).
  • ○
    Analyzer (for DIC and brightfield).

Step-by-step method details

Seeding mesothelial monolayer

Timing: 18–48 h

This step aims to seed a confluent mesothelial monolayer for co-culture spheroid invasion imaging.

• 1.
  Glass base dish preparation for cell seeding.

  • a.
    Prepare 50 μg/mL of fibronectin solution with 1 × PBS.
  • b.
    Coat the glass base of the dishes with fibronectin solution.
  • c.
    Incubate glass base dishes with fibronectin solution at 4°C for 12–48 h.

Note: For TFM analysis, prepare glass base dishes coated with PDMS substrate by incubating them with fibronectin solution on top of the PDMS. Further optimization will be necessary when using other protein or ECM coatings or alternative thin hydrogels.

• 2.Culture mesothelial cells with a 1:1 ratio of MCDB and Medium 199 supplemented with 10% FBS.

Note: We use the culture media recipe according to the published experimental setup by Iwanicki et al.⁶

• 3.
  Wash cells with 1 × PBS and trypsinize.

  • a.
    For next-day imaging, trypsinize mesothelial cells with trypsin for 1 min.
  • b.
    For same-day imaging, dissociate mesothelial cells using cell dissociation buffer for 3–5 min.
• 4.Neutralize using media supplemented with FBS and centrifuge at 500 × g for 2 min in a 15 mL Falcon tube.
• 5.Remove the supernatant and resuspend the pellet with a P20 micropipette.

Note: Ensure that the pellet is thoroughly broken up into a single-cell suspension.

• 6.Remove the fibronectin solution from the glass base dish.
• 7.Add 1 mL of media to the cell suspension and seed cells onto a glass base dish, ensuring that a confluent monolayer forms in 18–24 h (for next-day imaging) or 5–6 h (for same-day imaging).

Note: Tap the dish immediately after adding cells to distribute them evenly. For ZT mesothelial cells used in our model, approximately 1.2 cells/mL is ideal. Seed 200 μL of this single-cell suspension onto a 12 mm glass base dish for 24 h of incubation to achieve a confluent monolayer.

• 8.Incubate for 3–5 min to allow cell attachment before adding 2 mL of FBS-supplemented media (troubleshooting problem 1).

Note: For experiments that require adding exogenous fibronectin to the mesothelial monolayer, remove the current media after cells have attached and replace it with 200 μL fresh media supplemented with 20 μg/mL of human fibronectin plasma. Incubate in a humidity chamber at 37°C for 18–20 h before washing with fresh media and replacing with 2 mL of FBS-supplemented media.

• 9.Incubate for 18–24 h at 37°C for next-day imaging, or 5–6 h at 37°C for same-day imaging.

Culturing cancer cell spheroids

Timing: 2–3 h

This step aims to culture whole ovarian cancer cell spheroids for co-culture spheroid invasion imaging.

• 10.Culture ovarian cancer cells with a 1:1 ratio of MCDB and Medium 199 supplemented with 10% FBS.
• 11.Wash ovarian cancer cells twice with 1 × PBS.
• 12.Dissociate with cell dissociation buffer at 60%–80% confluency.
• 13.Incubate for 5–10 min at 37°C (troubleshooting problem 2).

CRITICAL: Ensure that cells remain attached and lift off together as cell collective sheets after incubation.

• 14.Neutralize with FBS-supplemented media.
• 15.Centrifuge at 500 × g for 2 min in a 15 mL Falcon tube.
• 16.Remove the supernatant and add 2 mL of FBS-supplemented media.
• 17.Gently resuspend the pellet and incubate it in the Falcon tube for 15–30 min at 37°C, keeping the tube upright.

Note: Cells should congregate at the bottom of the Falcon tube after each incubation.

• 18.Remove the supernatant and replenish with 2 mL of FBS-supplemented media.
• 19.Incubate for another 15–30 min at 37°C (troubleshooting problem 3).

Co-culture setup with cancer spheroids and mesothelium

• 20.Remove the supernatant. Gently pick up 100 μL of media containing cell clusters at the bottom using a P1000 pipette and add it drop by drop to the seeded mesothelial monolayer.

Note: The P1000 pipette is preferred because its wider, larger tips are less likely to disturb spheroid formation during aspiration.

• 21.Incubate at 37°C for 15–45 min to allow for attachment of spheroids before imaging.

Expected outcomes

After successfully seeding the mesothelial cell monolayer on the glass base dish, there should be no visible gaps between cells. Cancer cells that lift off in sheets after incubation with cell dissociation buffer should form spheroids in suspension with a grape-like cluster shape⁷ after incubation in media.

Cancer cells that lift off in sheets after incubation with cell dissociation buffer should be able to fold into spheroids after incubation in media for 30 min. Using this spheroid-making protocol, spheroids will be compact and ready for experiments within a day, compared to spheroid-making strategies that involve growing from smaller cell collectives to large, compact spheroids.

Quantification and statistical analysis

Nuclei annotation for machine training

Note: Labkit plugin is required.

• 1.
  Open TIF images in Fiji.

  • a.
    Select [Hyperstacks] under the [Image] menu.
  • b.
    Select [Reorder hyperstacks].
  • c.
    Change ‘z’ to ‘t’, and vice versa.
• 2.Select [Open image with Labkit] and remove all backgrounds.
• 3.Select [Settings] and set the contrast to 500 and 1,500.

Note: Adjust settings to enable overlapping labels.

• 4.Add label.
• 5.
  Outline and fill the regions of interest.

  • a.
    Use the Draw Tool to draw an outline of one spheroid nucleus on the current z-slice.
  • b.
    Ensure that the outline is connected.
  • c.
    Fill the outline with the Flood Fill Tool.

Note: Do not label mesothelial nuclei.

• 6.Repeat step 5a–5c for the rest of the z-stack.
• 7.Drag the slider tool across the z-slices to review the image stack from the top of the spheroid to the bottom.
• 8.Stop labelling at the z-slice where the nucleus is out of focus.
• 9.Proceed with the next visible nucleus by selecting a new label to outline.
• 10.Repeat steps 4–8 for multiple nuclei within the spheroid.

Note: Adjust the image's brightness for a clearer view of the nuclei.

• 11.Select [Labelling] > [Open labeling in ImageJ].
• 12.Save the labeled image as a TIFF file. Use the same name as the original image stack.
• 13.Repeat steps 1–12 with the next z-stack image until there are at least 10 annotated samples before moving on to machine training.

Note: Use time points that are evenly spaced at regular intervals.

Nuclei annotation checking

• 14.Open the TIF image and its corresponding mask generated from “nuclei annotation for machine training” in ImageJ.
• 15.Select [Image] > [Color] > [Merge channels].
• 16.Allocate the image as C1 (red) and its corresponding mask as C2 (green).

Note: Select “Keep source images.”

• 17.Adjust the brightness of both red and green colors to maximum.
• 18.Scroll through the image stack to identify any unlabeled nuclei. Accurately labeled nuclei will be marked yellow (Figure 1A).
• 19.
  Locate any missing annotated nuclei.

  • a.
    Select [Analyze] > [Tools] > [Synchronize windows] to synchronize the image and mask windows.

Figure 1.

Nuclei annotation of spheroid cells and machine training for predicting nuclei segmentation

(A) Composite image stack showing spheroid nuclei and labeled mask. Spheroid nuclei were outlined and filled in Fiji ImageJ using the marker tool.

(B) Threshold graphs of the trained model using labeled spheroid nuclei image stacks obtained through imaging. The trained model was further optimized to improve the accuracy of nuclei detection by providing a higher volume of labeled image datasets and fine-tuning the number of epochs.

(C) Ideal threshold graph statistics obtained after machine training conducted in Jupyter with the StarDist 3D code.³

Machine training for AI prediction of nucleus segmentation

• 20.
  Create the necessary folders in the server. Name the new folders as follows:

  • a.
    “∼ StarDist 3D” > “data” > “train” > “masks” and “images”.
  • b.
    “∼ StarDist 3D” > “models”.
• 21.Upload the original image stacks into the “train” > “images” folder.

Note: Image stacks must be in .tif format.

• 22.Upload the corresponding labeled images into the “masks” folder.

Note: The number of masks should match the number of images. Files must be in .tif format.

• 23.Ensure that the server used is “DL4Mic Notebook”.

Note: To change servers, select [File] > [Hub Control Panel] > [Stop My Server] > [Start My Server] > [Launch Server] > [DL4Mic Notebook] > [Start].

• 24.Open “2_training with stats.ipynb” in Jupyter Notebook.
• 25.Change the directory in cell 3.
• 26.Adjust the patch size to maintain the same ratio and set the batch size to 1.
• 27.Run the program by selecting [Kernel] > [Restart Kernel and Run all cells…].
• 28.Duplicate the program to record the data.
• 29.Record the threshold graphs (Figures 1B and 1C).

Running AI prediction

• 30.Upload the “∼ StarDist 3D” folder, including models, into Google Drive.
• 31.Create the following additional folders: “prediction” > “images” > “masks”.
• 32.Upload “batch prediction.ipynb” into a new notebook.
• 33.Upload a maximum of 3 image stacks into the “prediction” > “images” folder.

Note: Ensure that the images are in “.tif” format (troubleshootingproblem 4).

• 34.Mount the drive.
• 35.Edit the directory to (/content/drive/MyDrive/prediction/images/∗.tif).
• 36.Select [Runtime] > [Restart and run all].
• 37.Once completed, the generated labels should be saved in “prediction” > “labels.”
• 38.Load the generated labels into ImageJ to view the labels side-by-side with the original unlabeled image.
• 39.Identify inaccuracies in labeling and adjust the training as necessary.

Note: AI prediction should identify and label all nuclei in the image stack.

Importing labels into Imaris 3D workspace

• 40.Open the ‘.nd’ file in ImageJ and select [View virtual stack].
• 41.Open relevant channels.
• 42.Identify suitable datasets for analysis.
• 43.
  Save the selected datasets as a ‘.tif’ file.

  • a.
    Select [File] > [Save as…] > [.tiff].
  • b.
    Name the datasets accordingly.
• 44.Convert the ‘.tif’ files to an ‘Imaris Image’ file type using Imaris Image Converter. Input the appropriate voxel sizes during conversion.

Note: Voxel size can be identified through metadata.

• 45.Open Imaris.
• 46.Select [Arena] > [Observe folder] and select the folder with the converted Imaris images (.ims format) (Figure 2A).
• 47.Import corresponding image labels by selecting [File] > [Open Segmentation/label] (Figure 2B).

Figure 2.

3D tracking of spheroid cells’ movement

(A) 3D image view of spheroid. Image stacks were captured with a Z-step size of 2 over the necessary number of steps to image the entire spheroid from top to bottom.

(B) Image labels were imported using Imaris. AI-segmented labels and tracks were reviewed against the 3D spheroid image file and corrected when necessary.

3D cell tracking in Imaris

• 48.Select [Surfaces] > [Creation] > [Tracks].
• 49.Adjust the maximum distance and gap size parameters accordingly.
• 50.Proceed with the track creation to finalize automated track generation (troubleshooting problem 5).
• 51.To manually edit tracks, select the [Edit Tracks] tab.
• 52.Select a nucleus from timeframe 1 that requires manual correction.
• 53.Select [Rebuild around Selection].
• 54.Proceed to the next timeframe and select the nucleus that needs to be connected to the same nucleus from the previous timeframe.
• 55.Select [Add to Track] and [Rebuild around Selection] to organize the track.
• 56.Repeat steps 52–55 for all timeframes.

Note: To ensure that all the cells are tracked, deselect [Focus Tracks] to view all tracks and check that all tracks are connected.

Limitations

Due to machine limitations, AI prediction may produce inaccurate tracking results. User intervention through manual corrections will be required to obtain the most accurate tracking outcomes.

Troubleshooting

Problem 1

Mesothelial cells detach from the glass base dish after seeding and cannot form a confluent monolayer (related to step 8).

Potential solution

• •Ensure that the glass base dish has been sufficiently coated and incubated with a protein solution that can facilitate adherence to the glass base dish (e.g. fibronectin, collagen).
• •Try using glass base dishes from different manufacturers.
• •Treat cells with 25 μm Blebbistatin for 2 h and wash thoroughly before use. Blebbistatin treatment reduces cell contractility, helping cells spread out and form a complete monolayer. It is crucial to fully wash out the drug and optimize the concentration and duration to prevent harming cell viability and affecting experimental results.

Problem 2

Cancer cells dissociate into individual cells instead of sheets (related to step 13).

Potential solution

• •Ensure that cells are washed twice before incubating with the cell dissociation buffer.
• •Incubate cells with the cell dissociation buffer at 37°C for an additional 3–5 min.
• •Avoid using high passage cells.

Problem 3

Spheroids fail to form during incubation while in suspension (related to step 19).

Potential solution

• •Incubate cells in suspension for at least 25–30 min.
• •Incubate cells in suspension in a sterile bacterial dish at 37°C for an additional 15–20 min to promote spheroid formation.
• •Avoid using high passage cells.

Problem 4

There is an error while running the code (related to step 33).

Potential solution

• •Ensure that the uploaded training images have been renamed to .tif (lowercase). Uppercase .TIF files will not work with the code.

Problem 5

Imaris automated tracking after segmentation produces inaccurate results (related to step 50).

Potential solution

• •Manually correct tracks after generating the automated tracking results.
• •Switch between time points to verify that the correct cells are connected to the same track.

Resource availability

Lead contact

Requests and further information for resources reported in this study should be directed to the lead contact, Selwin K. Wu (selwin_wu@mail.dfci.harvard.edu).

Technical contact

Requests and further information for resources reported in this study should be directed to the technical contact, Celestine Z. Ho (celestineho@u.nus.edu).

Materials availability

Information on equipment used and reported in this study should be directed to Celestine Z. Ho (celestineho@u.nus.edu) or Boon Chuan Low (dbslowbc@nus.edu.sg).

Data and code availability

• •The scripts generated in this study can be downloaded from the GitHub repository: https://github.com/selwinwu/morphogenesis.git.
• •The specific version of this protocol is also archived in Zenodo under https://doi.org/10.5281/zenodo.16737385.

Acknowledgments

We thank Wei Jia Goh for the discussions and suggestions to use 3D StarDist to segment spheroids and Ince Tan’s lab for providing the ZT-GFP cells. We also appreciate Eunice Lee, Hui Ting Ong, Joel Tian, and Nandini Dhakaan for their input and support during this project. This research was supported by a Young Individual Research Grant (MOH grant no: MOH-000653) from the National Medical Research Council of Singapore to S.K.W. Additionally, this work received support from a Research Scholarship Block Research Fellow Scheme and MOE Tier 1 funding from the Singapore Ministry of Education to S.K.W. and B.C.L., as well as a National Research Foundation Grant (NRF-MSG02023-0001) to B.C.L. I.Y. thanks the financial support from Schmidt Sciences through the Eric and Wendy Schmidt AI in Science Postdoctoral Fellowship.

Author contributions

Conceptualization, S.K.W., C.Z.H., and C.B.-X.H.; investigation, C.Z.H., C.B.-X.H., E.W.P., P.S.L., K.W.L., S.W.B., D.B.L., S.K.W., and S.B.; resources, I.Y.; funding acquisition, B.C.L. and S.K.W.; writing, C.Z.H., D.B.L., and S.C.T.; editing, S.K.W.; supervision, S.K.W., J.L.Y., and B.C.L.

Declaration of interests

The authors declare no competing interests.