SpaChat: Integrating Single-Cell Foundation Model with Cell Graph Network for Spatially Resolved Cell-Cell Communication Inference

Currently, single-cell foundation models, which are based on extensive single-cell sequencing data, are highly effective in extracting crucial biological information about genes and cells. They consistently demonstrate exceptional performance across a wide range of downstream applications. However, when it comes to spatially resolved transcriptomic data (ST), the availability of methods for inferring cell-cell communications (CCCs), in conjunction with foundation models, remains limited. In this work, we present SpaChat, which integrating a fine-tuned single-cell foundation model with a cell graph network for spatially resolved CCCs inference. SpaChat defines two scoring strategies to infer significant ligand-receptor (LR) pairs, including intracellular score based on gene-level attention of fine-tuned single-cell foundation model and inter-cellular score based on the KNN algorithm in a cell-cell graph network. On this basis, SpaChat employs the law of molecular diffusion and mass action in chemistry and the strategy of permutation test to calculate communication strengths and filter out communications with low specificity. The benchmarked performance of SpaChat on public spatial transcriptomic dataset is superior to that of existing inference methods. Furthermore, SpaChat subsequently identifies the communication patterns of specific cell types and provides a variety of options for visualizing the results of CCC analyses. In summary, SpaChat enables the inference of spatially resolved CCCs from spatial transcriptomic data, providing valuable insights into understanding CCCs in tissues.