Group-based learning on label-free phase-contrast images across dose and exposure time improves bioactive compound classification

High-throughput phenotypic screening enables large-scale analysis of compound effects on cells using microscopy images. Typical approaches to phenotypic screening rely on fluorescence images, which are relatively straightforward to analyze but require fluorescence staining—a time-consuming process that can perturb cells. In this work, we focus instead on digital phase-contrast imaging. This imaging alternative avoids staining and is therefore faster and non-invasive, but it produces visually complex images that are more challenging to analyze. We address classifying 1 280 diverse compounds solely from digital phase-contrast images. To account for a priori unknown drug-specific potency and temporal effects, we propose a method that classifies groups of images acquired at varying compound concentrations and exposure times, rather than processing individual images. Our approach achieves 50% top-1 accuracy on the classification problem of 1 280 compounds. Moreover, we demonstrate that the learned image encoder transfers effectively to related tasks, including mechanism-of-action prediction and control-versus-treated classification. The presented results highlight the potential of combining digital phase-contrast imaging with appropriate machine learning models to enable scalable, cost-effective, and minimally invasive phenotypic screening.