Cell death quantification is one of the most common readouts in drug development pipelines and cell-based experiments, and it serves as a critical first pass in assessing compound toxicity. Traditional cell death assays are typically run independently from screening assays, and require specialized reagents that can be costly when scaled. 

To quantify cell death in conjunction with functional assay measurements, we developed a computer vision-based single-cell death predictor using morphological Cell Painting stains, which enables low-cost, high-throughput cell death quantifications alongside a myriad of other high-content imaging readouts. Our computer vision models were verified on a combination of established cell death stains (Annexin V and Zombie Red) and Cell Painting stains in U2OS cells and primary human PBMCs. The cell death stains allowed accurate model training (up to 100% in U2OS and 98.5% in PBMCs) with biological ground truths on a single-cell level. We then predicted two general death categories, inflammatory and non-inflammatory death, using only Cell Painting stains as inputs with model accuracies of up to 90% in U2OS and 91% in PBMCs. 

By analyzing concentration-response curves for several death-inducing control compounds such as apoptosis and pyroptosis inducers, we highlight specific death-related cell morphologies that occur without death stain labeling, potentially revealing cell death pathway activation earlier than in traditional death assays. Importantly, these death predictions and associated cell features can be used to filter out compounds that induce specific types of cell death and toxicity at the initial screening stage, saving valuable time and resources early in the drug discovery process.