APPLICATION NOTE

Monitoring of T-cell invasions assay using a 3D spheroid model Zhisong Tong | Research Scientist | Molecular Devices

Introduction T-cell therapies are designed to help our immune system eliminate cancer cells. Those include CAR T-cells (Chimeric Antigen Receptor engineered T-cells), tumor infiltrating lymphocytes (TIL), and other genetically modified T-cells. In recent years, the field of cell therapy has started to expand, including the launch of the first CAR T-cell therapies to treat blood cancer in 2017, which was a critical milestone in this field.1 Despite its boom, the discovery of novel immunotherapies that specifically enhance T-cell response against cancer cells remains a challenging task limited by the absence of robust in vitro models to evaluate these immunotherapies throughout their development. In the past, these models have been limited to the use of suspension cells and 2D cell monolayers.2 The use of CAR T-cells on solid tumors has been lagging due to challenges that include tumor heterogeneity, immunosuppressive microenvironments, and the lack of unique tumor antigens that can be recognized by the CAR-T cells. As such, the ability to screen for CAR T-cells (e.g. with CRISPR) that effectively target and kill tumors is an area of active research.3 Current approaches to measure CAR T-cell induced cytotoxicity includes flow cytometry or image-based techniques. In addition, several groups have published studies on in vitro evaluation of the interaction of T-cells

Benefits • Easily acquire 4D (3D + timelapse) datasets with high-content imager • Monitor phenotypic changes of 3D spheroid model live • Leverage machine-learning approach to classify spheroids in a robust manner • Analyze key features of phenotypic evolution over time with tumoroids.4 Here we describe a workflow for the generation of 3D tumor spheroids, co-cultured with T-cells as a proof-of-concept model for CAR T assays. Activated peripheral blood mononuclear cells (PBMCs) were added to spheroids and their activity monitored over time using high-content imaging. To optimize the workflow, we developed an image analysis approach that uses deep learning to accurately segment biological objects of interest and machine learning to quantify the T-cell induced phenotypic changes in the spheroids using only brightfield images.

moleculardevices.com | © 2022 Molecular Devices, LLC. All rights reserved.

Our results show the feasibility of using AI-based analysis workflows to predict the efficacy of T-cells. A robust deeplearning model was first trained to recognize spheroids and generate masks of the whole spheroids and their edges. Extracted measurements from the masks were then used to classify the spheroids using machine-learning approaches, where distinct phenotypic changes of the spheroids were observed compared to controls, allowing for analysis of the T-cell efficacy. To further elucidate which features dominate the classification, we quantified and compared features such as area, form factor, total intensity, and grey level non-uniformity between the different treatment groups.

The interaction of T-cells with spheroids was recorded by time-lapse high-content imaging every 2 hours over a 72 hours course of culture.

Cell culture 3D Spheroid: Hela cell line was passaged and 3D spheroids were formed by seeding Hela cells into a 96-well, round-bottom, ultra-low attachment plate (Corning cat # 4515) with a density of 2000 cells-per-well and cultured at 37°C for 2 days. T-Cell: T-cells were thawed from cryopreserved PBMC vials (ALLCELLS). T-cells were stimulated with 25ng/mL PMA and 1ug/mL ionomycin for 6 hours5 and labeled with CellTracker Green (Thermo Fisher Scientific) before seeding to the 96-well spheroid plate at a ratio 10:1 of the number of Hela cells.

Methods Experiment round-up In the study we compared the development of untreated spheroids (control), spheroids mixed with un-stimulated PBMCs, and spheroids mixed with stimulated PBMCs. Spheroids treated with staurosporine that is known to kill cells were included as another control (Figure 1A, 1B).

A

HeLa spheroid

Single HeLa cells

B

Cell culture (2 days)

Stimulated T cells

Co-culture (3 days) & Time-lapse imaging

Single T cells

PMA/i stimulation (6 days)

4

5

6

7

8

9

10

11

12

A

Stimulate T

Stimulate T

Stimulate T

Stimulate T

Stimulate T

Stimulate T

Stimulate T

Stimulate T

Stimulate T

B

Stimulate T

Stimulate T

Stimulate T

Stimulate T

Stimulate T

Stimulate T

Stimulate T

Stimulate T

Stimulate T

C

Stimulate T

Stimulate T

Stimulate T

Stimulate T

Stimulate T

Stimulate T

Stimulate T

Stimulate T

Stimulate T

D

Stimulate T

Stimulate T

Stimulate T

Stimulate T

Stimulate T

Stimulate T

Stimulate T

Stimulate T

Stimulate T

E

Non-Stimulate T

Non-Stimulate T

Non-Stimulate T

Non-Stimulate T

Non-Stimulate T

Non-Stimulate T

Non-Stimulate T

Non-Stimulate T

Non-Stimulate T

F

Staurosporine

Staurosporine

Staurosporine

Staurosporine

Staurosporine

Staurosporine

Staurosporine

Staurosporine

Staurosporine

G

Negative

Negative

Negative

Negative

Negative

Negative

Negative

Negative

Negative

Figure 1. (A) Workflow. Hela cells were seeded in 96-well round-bottom plate to form spheroids for 2 day. After 2 days, the thawed PBMC/T cells were stimulated in PMA/i for 6 hours before adding to the spheroids for co-culture. The time-lapse live imaging were then performed on spheroids and T cells every two hours. (B) Plate map of the assay. moleculardevices.com | © 2022 Molecular Devices, LLC. All rights reserved.

Spheroid imaging Images were acquired with a transmitted light (TL) channel and fluorescent channels on the ImageXpress® Micro Confocal High-Content Imaging System with MetaXpress® software. The 3D stack and 2D projection images were acquired with 10X objective and 10um focus step with time-lapse imaging every 2 hours with environmental control at 37 °C and 5% CO2. The 6.7.2 version of MetaXpress software allows us to directly save 3D stacks under time-lapse (4D) mode into a format compatible with IN Carta® Image Analysis Software.

Image analysis

Data analysis Measurements extracted for each spheroid were later used to generate a model in Phenoglyphs, a machinelearning based classification tool, to classify spheroids into 5 classes. The trends of key features under different treatments were plotted with Python matplotlib.

Staurosporine

Negative Control

Unstimulated T Cells

Stimulated T Cells

SINAP, a deep learning-based segmentation module, was used to train a custom model for accurate segmentation of spheroids across all timepoints and treatments. Model

has been trained using over 50 annotated images to incorporate all the conditions and using a previously trained organoid base model under Fine-tune mode with 100 epochs. Please note that if there is no similar model available, one may use default models as base model and retrain by incorporating more training sets. Masks were generated based on 2D projection of the TL channel to enable label-free analysis.

T=0

T = 68hr

Figure 2. TL image of Hela spheroids treated with stimulated T cells (A) at early stage; (B) at middle stage; (C) at late stage; treated with non-stimulated T cells (D) at early stage; (E) at middle stage; (F) at late stage; untreated with T cells (G) at early stage; (H) at middle stage; (I) at late stage; treated with staurosporine (J) at early stage; (K) at middle stage; (L) at late stage.

moleculardevices.com | © 2022 Molecular Devices, LLC. All rights reserved.

Results T-cell induced phenotypic changes of Hela spheroid To address the obstacle of in vitro evaluation of T-cells, we developed an image analysis method to study the interaction of T-cells with tumor cells in 3D assay format by using high-content imaging and analysis (Figure 1). Hela cells were then used to form spheroids. Forty-eight hours later, activated PBMCs were added to the spheroids and then imaged every 2 hours for 3 days (Figures 3C–3K).

Stimulated T Cells Treated

Stimulated T Cells Treated

Nonstimulated T Cells Treated

Nonstimulated T Cells Treated

After 3 days of co-culture, we observed several phenotypic changes in the spheroids that were imaged in TL (Figure 2). Spheroids from all treatment groups (except those treated with staurosporine) show an increase in size (area) and optical density (become darker). We also observed the presence of more stimulated T-cells within the spheroids compared to non-stimulated T-cells (Figures 3C–3J), along with deterioration of spheroid edges (Figure 2C and Figure 5).

Machine-learning based classification of spheroids To improve the workflow for the T-cell based assay, we developed a custom analysis pipeline to assess phenotypic changes in spheroids using only images acquired in brightfield. This approach eliminates the need to label cells before culture, saving both time and reagent cost. The 2D projection images in the TL channel (Best Focus Plane) were used to train and generate a SINAP model to mask the spheroid region. Due to the variability among different treatment groups of stimulated T-cells, non-stimulated T-cells and no T-cells, and different time points (Figure 2), images from all conditions were annotated to create a balanced training set. We created separate segmentation masks for spheroids and the edges because we observed that untreated spheroids form smoother boundaries compared to the treated spheroids (Figures 3A–3C). Measurements extracted from the resulting masks were then used in Phenoglyphs (Figure 4D) to generate a model for spheroid classification. Like other AI-based analyses, the pipeline in Phenoglyphs is outlined below.

Figure 3. (A) Primary Target, Cells and Organelles models used in IN Carta to generate the masks for spheroids and the edge of the spheroids; (B) The mask of the edge (red) and the spheroid (blue) treated with stimulated T cells and nonstimulated T cells at late stage; The fluorescent image blending of mitotracker (spheroid, red) and celltracker (T cells, green) treated with stimulated T cell (C) at 0hr; (D) at 18hr; (E) at 48hr; (F) at 68hr; treated with unstimulated T cell (G) at 0hr; (H) at 18hr; (I) at 48hr; (J) at 68hr. moleculardevices.com | © 2022 Molecular Devices, LLC. All rights reserved.

Step 1: Label the measurements to generate training sets using clustering tool. In Phenoglyphs, one starts with a clustering tool to label the appropriate classes before training a classifier model. The clustering tool relies on the selected measurements, which may be same as or different from the measurements used in the next step to train the classifier model. The fact that spheroids become darker and less transparent (Figure 2) suggests that the features related with intensity may play an important role in spheroid clustering, which also was reflected in the maturation of spheroids. We thus compared the clusters that are subject to all the features and the clusters that are subject to the intensity-related features and found that the latter case gives better clustering among different conditions—especially the condition of stimulated T-cells at late stage (Figure 4A pink and cyan label) and that of unstimulated T-cells at late stage (Figure 4B). We then labeled the training sets into 5 categories, including early stage, middle stage, stimulated T-cells late stage, unstimulated T-cells late stage and no T-cells late stage with reassignment of the images if necessary and proceeded to the next step. Step 2: Choose top 20 measures and train the classifier model using training sets.

Ranking tool sorts all measurements with significance scores after the clusters were manually labeled. We chose top 20 measurements with 95% correlation threshold to avoid overfitting of the resulting classifier model and proceeded with training of the model using a subset of measurements. Please note that the measures used at this step are different from those used in the clustering step. The trained model is applied to entire dataset and new set of exemplars, i.e., the images, are shown and we have options to reassign any of the images into a different class or ignore the outliers. Please note that the measurements are all related with each mask from images and when we reassign the images to different classes, we literally reassign the related measurements to different classes. This reassignment-training process may need to be repeated multiple times until satisfactory results obtained (Figure 4C). Step 3: Examine the classified images. To reach our goal that it classifies the spheroids into correct treatment categories, we used our trained model to predict all the spheroids at the final time point. The model accurately predicted 63 out of 65 wells (Figure 4E), with prediction accuracy of about 97%. Thus, artificial intelligence (AI)-based label-free prediction of spheroid evolution stemmed from T-cell infiltration is possible with high accuracy.

Figure 4. (A) The clustering generated from intensity related features with 50 clusters (only a subset shown); (B) The clustering generated from all features with 50 clusters (only a subset shown); (C) The final classification generated from the trained phenoglyph model; (D) Screenshot of IN Carta Phenoglyphs classifier; (E) The projected spheroid classification from Phenoglyphs model

moleculardevices.com | © 2022 Molecular Devices, LLC. All rights reserved.

Measurements of phenotype related features Above, we have the multi-parametric classifier based on machine learning, which is necessary due to the complexity of an assay environment where not a single measurement should dominate the classification. For example, even the intensity measurement may help distinguish the images in one scenario with meticulously chosen threshold, it may not work in another scenario when we change the light exposure time. Therefore, the machine learning based multi-parametric classification approach is a generic approach in phenotypical analysis. However, the single measurement-based trend may still give insight into the biological phenomenon and help choose the measurements in clustering step for the future similar assays. Thus, to further investigate how the treatments impact the phenotypes of the spheroids over time, we plotted 4 of the key features from the top 20 features: area, form factor, total intensity, and grey level non uniformity throughout the time point (Figure 5). Here we focus on the divergence of the stimulated curve (blue) from the rest of the curves. Overall, stimulated T-cell curves began to diverge as early as 22 hours for the form factor feature, at approximately 50 hours for total intensity, and at 38 hours for grey level non-uniformity features. The area feature was the least affected between stimulated

and non-stimulated groups. The holistic trends align with the penetration of T-cells which started at around 18 hours (Figures 3C-3J). Separately, the area of spheroids treated with T-cells (stimulated and unstimulated) were similar to the negative controls in the first 40 hours. After that, the area of T-cell treated spheroids were significantly larger than the control spheroids (Figure 5A). It is possible that the presence of T-cells (stimulated and unstimulated) affected the integrity of the spheroid structure. We have previously observed that, in some cases, a larger area is correlated with a disintegration of the 3D spheroid structure, resulting in a loose cluster of cells in the well. Spheroids treated with staurosporine do not show significant change in size over time, which suggests that the growth of the spheroid is inhibited. Form factor designates an object’s roundness index with value ranges from 0 to 1, where 1 is a circle. Figure 5B indicates that the stimulated group shows modest but statistically significant change of form factor compared to negative group and staurosporine group with p value .001 and .001 respectively, using student’s t-test. This trend is expected aligning with our observation that the edges of the simulated T cells treated spheroids are bumpier than the other groups making them form bulge structures around the edges (Figure 3B).

Spheriod area under different treatments

Spheriod mask form factor under different treatments

Spheriod total intensity under different treatments

Spheriod gray level non uniformity under different treatments

Figure 5. (A) The trend of the spheroid area under different treatments; (B) The trend of the spheroid mask form factor under different treatments; (C) The trend of the spheroid total intensity under different treatments; (D) The trend of the spheroid grey level non uniformity under different treatments. moleculardevices.com | © 2022 Molecular Devices, LLC. All rights reserved.

Total intensity is another feature worth mentioning here. First, the total intensity is a sum of all the pixel values of the object and thus one may expect larger total intensity with larger object size. Our results in Figure 5C aligned with that theory, as shown in Figure 5A, in terms of their relative values. Second, all the measurements here are performed in a TL channel, more compact and less transparent structures block more light resulting in smaller pixel values, coinciding with our observation that spheroids become opaquer after longer culture time— probably due to increased cell density.

Last, grey level non uniformity (GLNN) is one of the texture measures and evaluates similarity of grey-level intensity values in an object, where a lower GLNN value correlates with a greater similarity in intensity values. Figure 5D shows that T-cell treated spheroids tend to have the least similarity while staurosporine treated spheroids tend to have the most similarities.

Conclusions • We used time-lapse high-content imaging to monitor the growth and phenotypic changes of T-cell treated 3D spheroids. • We successfully generated SINAP models to apply masks to the whole spheroid and the edge of the spheroid. • We also trained a model in Phenoglyphs to classify the spheroids into 5 classes. •

Four key features are outlined over time suggesting useful information for future similar experiments.

moleculardevices.com | © 2022 Molecular Devices, LLC. All rights reserved.

References 1. CAR T-Cell Therapy Approved for Children, Young Adults with Leukemia - NCI (cancer.gov) 2. Jensen C. el al. Is it time to start transitioning from 2D to 3D cell culture? Front. Mol. Biosci., 7:33 (2020). 3. Wang D. et al. CRISPR Screening of CAR T Cells and Cancer Stem Cells Reveals Critical Dependencies for Cell-Based Therapies. Cancer Discov. 2021 May;11(5):1192-1211. doi: 10.1158/2159-8290.CD-20-1243. Epub 2020 Dec 16. PMID: 33328215; PMCID: PMC8406797. 4. Arno A. et al. Development of an innovative 3D cell culture system to study tumor—stroma interactions in non-small cell lung cancer cells. Plos One. 9 (3), e92511 (2014). 5. Wenchao A. et al. Optimal method to stimulate cytokine production and its use in immunotoxicity assessment. Int J Environ Res Public Health. 10(9): 3834-3842 (2013).

Contact Us

Regional Offices

Phone: +1.800.635.5577 Web: www.moleculardevices.com Email: [email protected] Check our website for a current listing of worldwide distributors.

USA and Canada United Kingdom Europe* China

+1.800.635.5577 +44.118.944.8000 00800.665.32860 +86.4008203586

Taiwan/Hong Kong Japan South Korea India

+886.2.2656.7585 +81.3.6362.9109 +82.2.3471.9531 +91.73.8661.1198

*Austria, Belgium, Denmark, Finland, France, Germany, Iceland, Ireland, Italy, Luxembourg, Netherlands, Portugal, Spain, Sweden, Switzerland and United Kingdom

The trademarks used herein are the property of Molecular Devices, LLC or their respective owners. Specifications subject to change without notice. Patents: www.moleculardevices.com/productpatents FOR RESEARCH USE ONLY. NOT FOR USE IN DIAGNOSTIC PROCEDURES.

moleculardevices.com | © 2022 Molecular Devices, LLC. All rights reserved.

©2022 Molecular Devices, LLC 11/22 2519A Printed in USA