Graph Learning of Cell-cell Communications in Spatial Transcriptomics

PROJECT SUMMARY Research Project: Many mechanisms of intercellular communications in tissue depend on the physical proximity between cells. Spatial barcoding-based transcriptomic data provide important and essential information to understand intercellular communications in intact tissue by measuring transcriptomic data and spatial locations of cells simultaneously. However, analysis of these data to understand intercellular communications faces the following challenges. First, spatial barcoding-based transcriptomic data lacks single-cell resolution, and cell type deconvolution is needed to infer cell-cell communication networks (CCCNs). Second, it is critical but challenging to integrate spatial information with prior knowledge of ligand-receptor interactions and downstream regulated genes for CCCN inference. Finally, difference in the distribution of CCCNs between disease and control groups needs to be assessed to identify disease associated CCCN perturbations. Current methods are not able to account for spatial correlation of cell type compositions between neighboring cells in cell type deconvolution nor to integrate prior knowledge of ligand-receptor pairs and downstream regulated genes in CCCN inference. There is no existing method developed to compare CCCNs between two groups of subjects. The goal of this application is to develop accurate, robust, and efficient bioinformatic and computational tools to deconvolve spatial barcoding-based transcriptomic data, infer CCCNs using spatial transcriptomic data, and assess differences in CCCNs between two groups of subjects. Our long-term objective is to identify disease associated intercellular communication changes from spatial transcriptomic data beyond what has been discovered by investigating individual cell types or cells. To achieve this goal, we propose to 1) develop a graph Laplacian regularized model to deconvolve spatial barcoding-based transcriptomic data using scRNA-seq data from same tissue type with integration of spatial information; 2) develop a regularized graph attention network model to infer CCCNs by integrating spatial information, prior knowledge of ligand-receptor pairs and corresponding downstream regulated genes; and 3) develop a graphical generative model that compares CCCNs between disease and control samples to identify disease associated perturbations in intercellular communications. Research design and methods: Drs. Xiting Yan and Zuoheng Wang will jointly lead the proposed research together with collaborator Dr. Naftali Kaminski, a team of experienced, committed experts in the fields of bioinformatics, statistics, genomics and genetics, biology, translational research and precision medicine. Regularized graph learning models will be developed. The datasets for our main study populations will come primarily from Dr. Kaminski’s lab, which will also execute discovery validations and downstream functional studies. R and python packages will be developed and freely distributed as open-source software in a light weight, portable and self-sufficient way using containers.