Download PDF

Zhisong Tong | Research Scientist | Molecular Devices
Angeline Lim | Senior Application Scientist | Molecular Devices
Maria Clapés | Product Manager | SUN bioscience
Marine Meyer | Product Manager | Doppl SA
Aline Roch | Head of R&D | Doppl SA

Introduction

Pancreatic ductal carcinoma (PDAC) is one of the most aggressive forms of cancer. In the United States, PDAC is the 4th leading cause of cancer mortality. Despite advances in understanding the mechanisms in PDAC using pancreatic cancer cell lines, xenografts and mouse models, the development of effective therapies has been lagging. To this end, patient-derived organoids (PDOs) have been proposed as viable and efficient models for ex vivo testing. PDOs show long-term expansion potential while retaining tumor histopathology as well as cancer gene mutations. However, the translation of organoids in screening applications has been hampered by challenges such as the highly manual cell culture process, time consuming imaging protocols and the lack of automation capabilities.

To overcome these challenges, we set up a compound screening workflow with PDOs using the Gri3D® platform (SUN bioscience), which is comprised of plates with micro-cavities suitable for high-throughput and reproducible organoid culture. Based on a standard 96-microtiter plate, each well contains a microwell array patterned in cellrepellent hydrogel. On Gri3D, organoids are generated in the microwells and located in the same imaging plane. This improves 3D image acquisition and quantitative analyses in high-content, image-based screens. Furthermore, the pipetting port enables automation of cell seeding, media exchange, and compound incubation with liquid handlers, thus increasing assay reproducibility.

In this study, we exposed human pancreatic cancer PDOs to a panel of anti-cancer compounds at different doses and followed their response to the drugs using viability dyes calcein AM (live stain) and ethidium homodimer-1 (dead stain) with high-content confocal imaging. Using an AI-based approach, we efficiently detected each organoid and extracted phenotypic features from all three channels (transmitted light, live stain, and dead stain) that correlated with cytotoxicity. The extracted features (more than 100) were first dimensionally reduced to 3 components using either PCA or UMAP, and then visualized in 3D scatter plots, ranked, and clustered using machine learning. The data suggests that the compound treatment palbociclib 50 µM has significant cytotoxicity effects similar to the positive control, which is consistent with the traditional live/dead analysis. The AI approach demonstrates the feasibility of performing drug screening using a robust and unbiased data analysis approach.

Materials and methods

Gri3D

Gri3D is a ready-to-use platform for high-throughput and reproducible organoid culture. Based on an array of ultradense U-bottom microwells (Figure 1) in hydrogel, single organoids were generated in each microcavity and grown in a suspension-like culture without a solid extracellular matrix (ECM).

Workflow

Standardized pancreatic cancer PDO arrays were generated in Gri3D 96-well plate plastic-bottom 500 μm microwells (Figure 1) and exposed to anti-cancer drugs for 72 hrs. Organoid line was provided and cultured by Doppl SA scientists who also carried out the experiments. Organoid response to drugs was followed over time with transmitted light (TL) on the ImageXpress® Micro Confocal High-Content Imaging System (Molecular Devices). Then, a live/dead assay was performed on treated organoids. Images were analyzed using IN-Carta® Image Analysis Software (Figure 2). Forty organoids were segmented per well and more than 100 metrics were extracted from each. Finally, the features related to each organoid were analyzed using StratoMineR® data analysis software (Core Life Analytics).

Figure 1. Gri3D technology. Gri3D is a 96-wellplate format plate for 3D cell culture. At the bottom of each well sits a cell-repellent polyethylene-glycol hydrogel where microcavities are imprinted. A pipetting port allows safe media exchange without organoid loss. Left: top view of Gri3D 500 μm microwells. Right: schematic of Gri3D well.

Figure 2. Schematic of the organoid drug efficacy workflow. Using Gri3D in combination with a high-content imager and AI-based image analysis software.

Results

Phenotypic effects of compounds

Upon palbociclib exposure, the live/dead assay showed a viability decrease of organoids with increasing drug concentrations, whereas no response is observed in trametinib-treated organoids (Figure 3B).

Using a deep-learning, image-based approach on TL images, we efficiently detected each single organoid and quantified different parameters over time. Briefly, SINAP— a deep learning-based tool in IN Carta image analysis software—was used for image segmentation. The highly variable nature of brightfield images makes it challenging to use conventional analysis methods to effectively segment objects based on a predefined set of parameters.

Using SINAP, images were first annotated to define the ground truth. Next, the set of annotated images was broken into a training set and a validation set which were used to train or refine a base model, then used as part of the image analysis workflow to extract features from each segmented organoid.

Grey level non-normality factor (GLNN) is an indicator of the similarity of grey values within an organoid. This value decreases with increasing doses of palbociclib and trametinib (Figure 3C), indicating organoid growth defects. Thus, TL analyses reveal mechanisms unnoticed on Live/Dead.

Figure 3. Response of pancreatic cancer PDOs exposed to anti-cancer compounds for 72 hours. A. TL images before (0h) and after (72h) exposure and maximum projection images of organoids after live/dead assay at 72h. Green: calcein AM, live; red: EthD-1, dead. B. Ethidium homodimer-1 (EthD-1) to calcein AM intensity ratio. C. Grey levels non-normality factor (GLNN) from TL images. Error bars show standard deviation. Each dot represents an organoid. One-way ANOVA Dunnett’s multiple comparisons, **P < 0.01, P**** < 0.0001, ns: non-significant. Scale bar: 250 μm.

Hit selection via dimensionality reduction

IN Carta analysis returns hundreds of features associated with each organoid from the TL, EthD-1, and calcein AM channels. To visualize and cluster the treated and control wells, we first perform dimensionality reduction through principal component analysis (PCA) using three components, i.e., PCA01, PCA02, and PCA 03 (Figure 4A, 4B, 4C). We used PCA as dimensionality reduction tool with rationale to reduce to 3 dimensions for the convenience of visualization and classification purpose. For details about PCA, please refer to Wikipedia. We then visualize the treated and control wells in a 3D scatter plot, where the negative controls and positive controls form their own clusters, respectively, and palbocilib 50 µM is the only treatment that clusters closer to the positive controls (Figure 4D). This observation is aligned with the live/dead analysis (Figure 3B). Plus, the ranking of each treatment (Chebyshev maximum distance) from the average negative control shows that palbociclib 50 µM ranks number one, while the four positive control wells rank the top six. Thus, palbociclib 50 µM has the greatest compound effect on human pancreatic cancer organoids.

Figure 4. Clustering and ranking via the principal component analysis. A., B., C. The three PCA components with weighted association with original features. D. 3D scatter plots of the three PCA components. E. Ranking of treatments and controls through the Chebyshev maximum distance from the average negative control.

UMAP

To validate our analysis, we use Uniform Manifold Approximation and Projection (UMAP) to reduce feature dimension. With all features considered, the clustering plot also shows palbociclib 50 µM clustered close to the positive controls (Figure 5A). To determine if we could distinguish phenotypic changes with only brightfield images, we ran the analysis using only features derived from the TL channel. The resulting 3D scatter plot shows no obvious clustering characteristics (Figure 5B), which suggests the need for additional features.

Figure 5. UMAP dimensionality reduction and clustering. A. 3D scatter plot of UMAP components of all features. B. 3D scatter plot of UMAP components of the features of TL channel only

Summary

• Gri3D plate is a beneficial platform for high-throughput and reproducible organoid culture.
• The ImageXpress Micro Confocal system successfully captured the 3D structure of organoids in a high-throughput and high-resolution approach.
• IN Carta image analysis software generated masks for each organoid to measure and hundreds of features that are associated with each organoid.
• Using AI, StratoMineR data analysis software clustered these organoid features to rank the efficacy of each drug treatment.
• The organoid model was obtained from Doppl SA biobank, additional models are available.

Download PDF