Osheen Sharma[PhD Student], Greta Gudoityte[PhD Student], Karolinska Institute, Stockholm, Sweden. Advisor: Brinton Seashore-Ludlow
Title: Morphmap: Systematic Morphology Mapping to Deconvolute the Cell-Cell Interaction
Problem and Rationale
Cell-cell interaction is a complex phenomenon playing a big part in various processes such as cell proliferation, migration, homeostasis etc [1]. However, cell communication can be overlooked when studying various diseases, such as cancer. Emerging evidence suggests that tumour cell interactions with its’ surrounding cells is important for disease progression and response to the treatment [1–3]. High content screening enables research into cell interactions at a single-cell level [2]. Image-based phenotypic profiling (e.g., Cell Painting [4]) can reveal cell structural changes reflecting cell proliferation and state, as well as provide mechanistic information on response to drugs [5]. Typically, feature extraction from high throughput imaging data is achieved using open-source software like CellProfiler, ImageJ, etc [6]. These softwares have shown promising results in extracting phenotypic variations in cells. However, they require handcrafting to each assay variation, which is time consuming and leads to bias in the results. To improve the existing feature extraction methods and make it more robust, recent studies have implemented deep neural networks that can capture phenotypic variations in cells after drug treatment [7,8]. Ljosa et al., uses CellProfiler to extract phenotypic features from breast cancer cells treated with 113 compounds [6]. The work of Janssens et. al., and Nikita et al., apply fully unsupervised and weakly supervised neural network to cluster cells with similar phenotypes together [6,8]. All of these studies have been performed on single cell type in the presence of single compound to understand the effect of drugs on cell morphology. Therefore, it would be valuable to merge high content screening with machine learning to investigate cell-cell interactions at single-cell level. Systematic phenotypic profiling of cell changes from the innate to activated state through intracellular communication can help to build a tool that captures and predicts cell state (Fig. 1A). By layering drug induced perturbations on this dataset we will learn which pathways control these interactions (Fig. 1B).
Details of suggested approach
To investigate drug induced morphological changes a co-culture system will be established. For initial study ovarian cancer cell lines will be in our focus. This gynaecological malignancy is known for high heterogeneity at the genomic, phenotypic, and histological levels, as well as high relapse rate to a first line treatment, which makes it an interesting disease for search of new treatment possibilities. Co-cultures will be treated with an annotated drug library containing compounds with different mechanism of actions (MoA). Cancer and fibroblast cells, as well as co-cultures will be examined initially. Morphological changes of cancer and fibroblast cells with and without treatment will be attained using Cell Painting method [4].Subsequently, high-throughput microscopy will be used to acquire the images (Fig. 2A).
To transform our high throughput images that are rich in information into biologically meaningful data we will implement deep neural network. Since supervised training requires huge, annotated dataset which is trained to answer only known questions, we will use weakly supervised and unsupervised learning approaches to associate treatments with images thereby leading to data driven findings. First, baseline transfer learning model will be used to extract cellular features from network embeddings. This strategy will enable us to validate our hypothesis of extracting features from multiple cell population treated with combination compounds. Further, to make the models more generalized, we will train the pre-trained networks on single cell images with a weak auxiliary task to generate binary classification of categorizing cells as negative or positive controls without any MoA/compound knowledge. This strategy will force the neural network to learn cell specific features associated with combinational drug effect and will unveil unknown MoA. The features from both baseline model and trained neural network will be then used for downstream analysis and model evaluation. The results will be used to uncover new biological insights into compound MoA and cell-type specific drug response across cell lines with distinct innate morphologies (Fig. 2B).
Impact to the Field
In the proposed application we aim to establish framework for systematic analysis of cell interactions empowered by Artificial Intelligence. Using high-throughput Cell Paining dataset of co-cultures we will be able to uncover phenotypic changes and predict cell state. Previous studies have shown promising results of understanding cell morphology changes on one cell type. We believe our workflow will enable us to uncover cell communication. Additionally, evaluating drug induced perturbations we will systematically target cell interactions to pinpoint effective connections.
These insights will provide new opportunities to accelerate drug discovery process. Our workflow first applied for cancer-fibroblast interactions will serve as a foundation for studying cancer cell communication with other cell types (adipocytes, immune cells etc) to build the interaction profile of the tumour environment (Fig. 3). While this approach only is discussing 2D co-culture datasets, it would be valuable to transfer the pipeline to a 3D environment. Finally, this could result in a recourse containing described framework to examine cell interactions and be share with a global scientific community.
References
1. Kumar MP, Du J, Lagoudas G, et al. Analysis of Single-Cell RNA-Seq Identifies Cell-Cell Communication Associated with Tumor Characteristics. Cell Reports. 2018; 25: 1458-1468.e4.
2. Jin M-Z, Jin W-L. The updated landscape of tumor microenvironment and drug repurposing. Signal Transduct Target Ther. 2020; 5: 166.
3. Farc O, Cristea V. An overview of the tumor microenvironment, from cells to complex networks (Review). Exp Ther Med. 2021; 21: 96.
4. Bray M-A, Singh S, Han H, et al. Cell Painting, a high-content image-based assay for morphological profiling using multiplexed fluorescent dyes. Nat Protoc. 2016; 11: 1757–74.
5. Rietdijk J, Aggarwal T, Georgieva P, Lapins M, Carreras-Puigvert J, Spjuth O. Morphological profiling of environmental chemicals enables efficient and untargeted exploration of combination effects. Sci Total Environ. 2022; 832: 155058.
6. Ljosa V, Caie PD, Horst R ter, et al. Comparison of Methods for Image-Based Profiling of Cellular Morphological Responses to Small-Molecule Treatment. J Biomol Screen. 2013; 18: 1321–9.
7. Moshkov N, Bornholdt M, Benoit S, et al. Learning representations for image-based profiling of perturbations. Biorxiv. 2022; 2022.08.12.503783.
8. Janssens R, Zhang X, Kauffmann A, Weck A de, Durand EY. Fully unsupervised deep mode of action learning for phenotyping high-content cellular images. Bioinformatics. 2021; 37: 4548–55.