(721e) Reconstructing Signaling History and Spatial Organization of Single Cells with Generalizable Statistical Inference

Proper cell-cell communication is the linchpin for the development of multicellular organisms. Mapping these communication networks is crucial for understanding the logic of embryonic development and for directing embryonic stem cells differentiating into desired fates. In the past, cell-cell communication has been primarily mapped through time-consuming animal genetics. Here, by fitting conditional variational autoencoders (CVAE) to scRNA-seq data, we created IRIS (Intracellular Response to Infer Signaling state), a semi-supervised deep learning method for annotating signaling state of individual cells only using its gene expression. Compared to other cell communication prediction algorithms based on ligand-receptor expression, IRIS relies on the induced gene expression changes in signal-receiving cells, which serves as the most accurate measure for functional cell-cell communication. However, such inference is complicated by the perceived context-dependent nature of gene expression changes. We demonstrate that many developmental signaling pathways induce cell type independent signatures of gene expression. This enables us to train our model on limited datasets in which the signaling states of the cells are known, and predict the signaling state of previously unseen cell types. The predictions are highly accurate, as quantified by orthogonal datasets we generated, in which human embryonic stem cells were differentiated upon sequential combinatorial ligand stimulation. We validated that the predictions successfully recovered existing knowledge of cell communication in diverse biological contexts, including gastrulation, early endoderm organogenesis, and various mesoderm lineages in mouse embryos. In contrast to other signal prediction algorithms, IRIS requires no observation of ligand sending cells or hard labels on the clusters for prediction, allowing heterogeneous ligand stimulation to be accurately annotated at the single-cell level. Furthermore, we show that these predictions can be used to annotate signaling history and reconstruct the spatial relationship amongst cells. We anticipate that this tool will be generalizable to solve a wide class of cell-cell communication inference problems and provide better defined signaling conditions for stem cell differentiation.