(1197-B) A Graphical User Interface for ML-Based Modeling of Single-Cell and Well-Level Imaging Data
Monday, February 5, 2024
2:00 PM – 3:00 PM EST
Location: Exhibit Halls AB
Abstract: We developed an ML-based platform targeted towards biologists with no background in machine learning to empower them to analyze cellular imaging data with artificial intelligence. Specifically, we built three modules that are integrated into one software application, which we term PhenoSorter, PhenoFinder, and Supervised Learner. PhenoSorter allows users to train ML models on observed single-cell phenotypes. PhenoFinder gives users the ability to harness unsupervised clustering to explore morphologically distinct cell classes in an unbiased manner. Finally, Supervised Learner allows users to train classification models on well-level data. The models trained using these three modules are then applied to the whole dataset, the results of which can be visualized with a suite of analysis tools. Here, we demonstrate the utility of these modules on a cell death dataset in which we stimulated primary human PBMCs and U2OS cells towards either apoptotic or pyroptotic death with control compounds at multiple doses, and stained them with a set of morphological and death type-specific labels. We used PhenoSorter to label single cells according to their death stain labels, and to predict the labels with only the morphological stains. PhenoFinder allowed us to group single cells into morphological clusters with statistical associations to experimental conditions. Finally, we trained classifiers to identify control compound-mediated effects at the well level using Supervised Learner. With these analytical tools, we characterized the cellular morphologies induced by each control compound, and how they relate to established ground-truth death stains.