A machine-learning framework that combines conditional variational autoencoders with spatial graph analysis to learn unsupervised phenotypic embeddings of complex tissue architectures from microscopy images.