(1073-B) From Pixels to Phenotypes: Interpreting Cell Painting Profiles in Cell Health Space

Abstract: The Cell Painting assay generates morphological profiles that are versatile descriptors of biological systems, which have been used to predict drug toxicity at cellular, mitochondrial, and organ levels. Although these high-dimensional cell morphology readouts are generated in a hypothesis-free manner, they can still capture information about specific Cell Health phenotypes and biological processes, such as DNA fragmentation and heterogeneity in mitochondrial content. However, Cell Painting features are based on image statistics and parameters like pixel count, making them less interpretable in a biological context. In this study, we introduce an algorithm that maps specific Cell Painting features to describe the shared role of these cell morphological features in various Cell Health terms. Consequently, the Cell Health space facilitates mechanistic analysis when examining feature importance in machine learning models. To demonstrate its utility, we optimized Random Forest classifiers using Cell Painting features to predict nine broad biological activities from the ToxCast assay, such as apoptosis, ER stress, and oxidative stress, for 658 unique compounds. The most contributing Cell Painting features from these models were mapped into the Cell Health space which revealed effects of biological activity (such as cell cycle arrest in the G2 phase for ER stressors) as well as for individual compounds (such as the compound emetine, a known protein synthesis inhibitor, also inhibits DNA replication in the early S phase of the cell cycle). In summary, the Cell Health space offers biologists an alternative, more biologically relevant way to interpret cell morphological features and generate hypotheses for experimental validation compared to the image statistics-based features in the Cell Painting assay, where information on the cell cycle of biological pathways is not directly measured.