CondensateGym data resource and Phase2Face algorithm for quantitative phenotyping of biomolecular condensates in cells

Biomolecular condensates are membraneless compartments that respond rapidly to their environment to regulate crucial biological processes. They are commonly analysed by fluorescence microscopy imaging, but their large heterogeneity across cells makes it difficult to define and compare condensates across experimental systems in a quantitative way. Here, we present Phase2Face, a tool for unbiased condensate segmentation and quantitative phenotyping at scale, which was trained on CondensateGym, a benchmarking resource of ~1,000 expert-annotated cells. Phase2Face is an adaptable machine-learning algorithm that generalizes across condensate types, cell lines and fluorescent labels. By analyzing cells with varying expression levels, Phase2Face reconstructs in-cell phase diagrams, revealing concentration-dependent condensate formation. We applied Phase2Face to >50,000 images of six PML isoforms and uncovered distinct phase behaviors missed by binary condensate classifications. Phase2Face extracts interpretable features that distinguish condensate classes and enables quantitative comparison of condensate phenotypes across sequence perturbations and conditions in health and disease.