Mapping human microglial morphological diversity via handcrafted and deep learning-derived image features
Microglia play critical roles in brain health and disease by adopting a spectrum of dynamic activation states. However, capturing this continuous heterogeneity in a scalable way remains a major challenge. To address this, we developed an imaging and data analysis framework to map the activation landscape of human iPSC-derived microglia (iMG) at single-cell resolution. We combined high-content imaging using two complementary strategies: a hypothesis-driven immunofluorescence (IF) panel targeting key activation markers (NF-{kappa}B, ASC, CD45) and a discovery-oriented, pan-morphological Cell Painting (CP) assay. Diverse phenotypes were captured through handcrafted and representation learning-based features. To classify cells, we applied Gaussian Mixture Models (GMMs) to image-derived features, enabling soft probabilistic assignments that capture transitional states between phenotypes. Compared to graph-based methods like the Leiden algorithm, GMMs provided comparable classification performance while offering a more nuanced and biologically interpretable view of microglial heterogeneity. We demonstrate that deep learning features from the targeted IF panel are most powerful, achieving high classification accuracy and strong correlation with biological states such as functional NLRP3 inflammasome activation. Our model system provides a robust and scalable platform for quantifying microglial heterogeneity, offering a new tool to identify novel disease-associated states and compounds that precisely modulate microglial phenotypes for therapeutic discovery.