(1259-D) Determination of Cell Viability using High-Dimensional Morphology Analysis
Tuesday, February 6, 2024
2:00 PM – 3:00 PM EST
Location: Exhibit Halls AB
Abstract: Morphological changes associated with alterations in cell state, health, and fitness have been reported in studies of drug response, cell differentiation, and disease progression. However, the ability to profile cell health at scale can be costly, time consuming, and labor intensive. Furthermore, late stages of death can be mistaken as debris by some methods. Here, we describe label-free morphological profiling of cells to determine cell viability. Deepcell’s REM-I is a research platform for phenotyping and sorting cells of morphologically similar profiles based on real-time deep learning interpretation of high-content morphology data without biomarker labels. Healthy cell line samples were subjected to various chemical and physical methods to induce cell death. The varied resulting cell health states were imaged on the REM-I platform as single cell suspensions. The captured brightfield images of single cells were characterized using the Deepcell Human Foundation Model (HFM) to infer morphological features associated with respective cell health states. The HFM is a hybrid architecture that combines self-supervised learning and morphometrics (computer vision) to extract 115 embedding vectors representing cell morphology from the high-resolution REM-I images. High-dimensional morphological profiling showed that untreated, apoptotic, and necrotic cells exhibited distinct morphotypes, with respective cell death pathways localizing to distinct areas of the morphometric embedding space (UMAP). Morphological profiling exhibited concordance with traditional cell death detection techniques for cell viability assessment. Furthermore, high-dimensional morphology profiling could detect subtle morphological changes during distinct stages of apoptosis and necrosis which are not apparent via traditional flow cytometry. This data suggests that cellular profiling with deep learning and morphometrics can be used to determine cell viability and state of health without label-associated bias. Next, we applied respective embeddings extracted from images of viable and non-viable cells to train a random forest classifier to infer features linked to cell viability. A total of 1.5 million images of viable and non-viable cells acquired from various cell lines composed the training set. Model performance evaluation of the ‘Cell Viability Classifier’ on the hold-out dataset showed over 97% specificity and 96% sensitivity. Combined with the REM-I platform, the ‘Cell Viability Classifier’ can be used to determine the viability of cells in a sample without labels, stains, or dyes. Importantly, while traditional cell viability analyses are endpoint experiments, cells of interest can be sorted and collected on the REM-I platform for downstream assays requiring high-quality, viable cells, such as RNA sequencing, cell culture expansion, and functional assays.