Counterfactual Hypothesis Testing Of Tumor Microenvironment Scenarios Through Semantic Image Synthesis
Recent multiplexed protein imaging technologies make it possible to characterize cells, their spatial organization, and interactions within microenvironments at unprecedented resolution. Although observational data can reveal spatial associations, it does not allow users to infer biologically causative relationships and interactions between cells. To address this challenge, we develop a generative model that allows users to test hypotheses about the effect of cell-cell interactions on protein expression through in silico perturbation. Our Cell-Cell Interaction GAN (CCIGAN) model employs a generative adversarial network (GAN) architecture to generate biologically realistic multiplexed cell images from semantic cell segmentations. Our approach is unique in considering all imaging channels simultaneously, and we show that it successfully captures known tumor-immune cell interactions missed by other state-of-the-art GAN models, and yields biological insights without requiring in vivo manipulation. CCIGAN accepts data from multiple imaging technologies and can infer interactions from single images in any health or disease context.