Download PDF

Zhisong Tong | Research Scientist | Molecular Devices

Introduction

Immunotherapy is increasingly popular as a type of cancer treatment. Chimeric antigen receptor (CAR) T-cell therapy is an approach to arm the immune system to treat cancer by genetically altering T-cells so they can target and destroy cancer cells. Although much success has been achieved with immunotherapy for treatment of blood cancers, the application in solid tumors remains challenging. One of barriers is intrinsic to the solid tumor microenvironment (TME) where suppressive cytokines limit the tumor killing ability of T-cells. Thus, TME plays a critical role in the response to cancer treatment and understanding the role the TME plays is essential to enhance CAR T-cell efficacy.

The benefit of using 3-dimensional (3D) patient-derived organoids (PDO) is that they model more of the physical and chemical cues present within the TME that are lacking in traditional 2D monolayer cultures. Studies show that PDOs show similar responses to drugs as original tumors, suggesting the value of using PDOs to improve therapeutic outcomes. Thus, PDOs represent a superior preclinical model system compared to 2D models through their inherent heterogeneity, long-term stability, applicability for high-throughput screens, and enhanced capacity to capture tumor characteristics.

Despite the benefits associated with the use of PDOs, there are significant barriers to their widespread adoption in drug discovery due to costly and highly labor-intensive growth and maintenance. To address the challenges associated with the use of PDOs in large-scale applications, a semi-automated bioreactor was developed for the large-scale expansion of assay-ready organoids. Here, we developed a workflow to quantify the efficacy of T-cell in solid tumors using PDOs. Using bioreactorexpanded patient-derived colorectal cancer organoids (CRCs), activated human peripheral blood mononuclear cells (PBMCs) were added to CRCs in a 384-well microtiter plate and monitored every 4 hours for 3 days using highcontent imaging. To quantify the morphological changes induced by T-cells, we first segmented each organoid with the transmitted light (TL) channel using a deeplearning U-net model and extracted features from each organoid. The resulting features were then classified using a trained Random Forest model that differentiates the morphologically modified organoids from the intact organoids. Using this approach, we found that comparing to control wells, the percentage of modified organoids in stimulated PBMC-treated wells increased rapidly— within around 2 days—and then reached a fluctuated plateau, suggesting the potential of a live T-cell efficacy assessment approach with or without labeling. The results demonstrate the utility of the bioreactor-expanded organoids with the analysis approach in largescale T-cellbased screens.

Instruments and methods

Workflow

Bioreactor-expanded patient-derived CRCs were first mixed with 80% Matrigel and grown for 48 hours before collection. The collected large organoids were mixed with 3% Matrigel before seeding into a 384-well flat-bottom plate with ultra-low attachment (Corning). The thawed PBMC/T-cells were stimulated in PMA/i for 6 hours before adding to the CRC organoids for co-culture. We used the ImageXpress® Confocal HT.ai High-Content Imaging System (Molecular Devices) equipped with spinning disk confocal and sCMOS camera to perform timelapse live imaging every 4 hours.

Automation setup

An automated workcell consisting of an incubator and a high-content imager (Figure 2, green frame) was used for monitoring the co-culture of organoids and T-cells. The Genera scheduling software (RETISOFT) was used to execute routine monitoring of the organoids in culture. The protocol involved the retrieval of the plate from the incubator, transport of the plate to the ImageXpress Confocal HT.ai to image the organoids every 4 hours (z-stack acquisition with step size 10um, 10X), and placement of the plate back in the incubator using the PreciseFlex400 robotic arm (Brooks).

Figure 1. T-cell and CRC PDO interaction workflow.

Analysis workflow

The timelapse images taken with HT.ai were first exported into a format compatible with IN Carta® Image Analysis Software (v. 2.4, Figure 3A, 3B). The organoids were segmented using a custom deep learning model trained to detect organoids in SINAP module (Figure 3C, 3D). A set of over 50 features were measured and extracted for each organoid. Analysis results were then loaded into Phenoglyphs—a machine learning-based classifier module that allows users to label each set of features by annotating the images of the organoids, and then classifies organoids into custom clusters individually (Figure 3E, 3F). In our case, organoids were split into morphologically intact or modified organoids. The result was then exported into a .csv file and the percentage of the organoids of morphological change was graphed across timepoint. Both SINAP and Phenoglyphs are available as seamlessly integrated modules in IN Carta image analysis software.

Figure 2. Layout of the automated workcell. This standardized workflow included a liquid handler (Hamilton), a robotic arm (Brooks), incubator (LiCONiC), the ImageXpress Confocal HT.ai system (Molecular Devices), ImageXpress Pico Automated Imaging System (Molecular Devices), SpectraMax® iD5 MultiMode Microplate Reader (Molecular Devices), AquaMax® Microplate Washer (Molecular Devices), a plate hotel, a centrifuge, and a barcode scanner. The curved arrows show an example of the process to monitor cells in culture where plates are moved from the incubator to the ImageXpress Confocal HT.ai system for imaging and then back to the incubator

Figure 3. A, B. The collected images were first exported and opened in IN Carta. C, D. Single organoids were segmented using custom SINAP model in TL images. E, F. Extracted single object features were loaded into Phenoglyphs and organoids were classified into morphologically intact or modified organoids.

Results

Segmentation in SINAP

To describe the T cell-induced morphological change of organoids and how fast these changes happen, we used a custom deep learning model trained in SINAP that specifically segments CRC organoids in TL channel. SINAP is established on the convolutional neural network (CNN) model and orchestrated in a way that even a data analytics beginner can easily grasp and do deep learningbased segmentation (Figure 3C). To quickly build the training image set, we first used the Brush tool to annotate the initial training set (5 to 10 images) for an initial model, which was finetuned from existing standard model and then used for roughly segmenting other organoids. With further refinement by Brush tool, a larger training set of over 50 images was built. An interesting point worth mentioning is that there is usually more than one approach to segment the objects and we ended up with segmenting the organoids surrounded by the solid edge together with the organoid fragments attached onto the solid edge (Figure 4E and 4F, green arrows).

The training set should include annotated images of all available morphologies so that the resulting CNN model can robustly segment all the organoids. In the current case, we divided the morphological types of organoids into two categories: I. intact organoids with bubbly structure which included those at the early timepoints treated with stimulated or non-stimulated T-cells with light appearance in TL channel (Figure 4A–4B) and those at the late timepoints treated with non-stimulated T-cells but with dark appearance in TL channel (Figure 4C–4D), II. modified organoids that were losing their bubbly structure at the late timepoints treated with stimulated T-cells with dark appearance in TL channel (Figure 4E–4F). This implies that stimulated T-cells impact the overall structure of the CRC organoids, while non-stimulated T-cells have little effect. Thus, we were able to quantify the percentage of organoids of stimulated T cell-induced morphological change as a measure of T-cell efficacy.

Figure 4. A. Representative TL image and B. the corresponding segmentation of intact organoids at the first timepoint. C. Representative TL image and D. the corresponding segmentation of intact organoids at the last timepoint treated with non-stimulated T-cells. E. Representative TL image and F. the corresponding segmentation of modified organoids at the last timepoint treated with stimulated T-cells.

Classification in Phenoglyphs

To classify thousands of organoids in TL channels from 22 timepoints into two classes of morphologically modified organoids and intact organoids, we loaded object-level features into Phenoglyphs module. Phenoglyphs classifier uses Random Forest as its underlying algorithm and enables image-based labeling to build the training set (Figure 3E). To quickly build the training set, we generated 50 clusters with the built-in k-means++ algorithm, where clusters of interest were labeled as Changed and Unchanged classes with minor reassignment of images (Figure 5A). The model was first trained with the labeled classes (ground truth) and the resulting classification can be improved with the reassignment of the organoids of classification error, which would be added into training set (expanded ground truth, Figure 5B). The iterations of the cycle of Class Reassignment to correct the classification error and Training using the now expanded ground truth was then repeated (Figure 5C). This “human-in-the-loop” approach enables training models more efficiently. We stopped the loop when the model quality metric F1 score reaches 1 and the exemplars images are satisfactorily classified in both classes (Figure 5D). Total 125 objects were labeled as ground truth when we stopped the training loop.

Figure 5E shows that for stimulated T cell-treated wells, the percentage of morphologically modified organoids increases rapidly at an average growth rate of 1.38%/hour within 52 hours after treatment and then begins to saturate at an average growth rate of 0.09%/hour (Figure 5E, purple dotted line). This implies that stimulated T-cells have more effect on organoids’ morphology within 52 hours. Since the manually labeled ground truth is mostly picked from later timepoints (for example, compare Figure 4C and Figure 4E), the classification of organoids at earlier timepoints may produce more inevitable errors, resulting in a bump in the percentage curve (Figure 5E, orange line) at earlier timepoints.

Figure 5. A. Example of Exemplars section when labeling training datasets begins. B. Example of Exemplars section when one round of training is completed. C. Phenoglyph training workflow. D Model quality metric F1 score graph. E. The representative images of classified organoids showing the increased number of modified organoids throughout the time (Pink: intact organoids, Green: modified organoids). F. Percentage of morphologically modified organoids summarizing all corresponding wells that are either treated with stimulated T-cells (10 wells) or non-stimulated T-cells (5 wells).

Conclusions

• We successfully implemented a novel T-cell workflow to automate time-lapse high-content imaging.
• CRC organoid segmentation that included various morphologies was successfully achieved using a customized segmentation model.
• CRC organoids were successfully classified according to the morphological characteristics using Phenoglyphs.
• The whole system—along with the percentage analysis of morphologically modified organoids across time—can be used as a successful live T-cell screening approach.

Download PDF