SPCC: Inferring Spatial Cell-Cell Interaction by Integrating Single-cell and Spatial Transcriptomics.

Study of single-cell spatial biology reveals the importance of integrating single-cell and spatial data for capturing spatial structure at individual cell resolution in various fields. With the lack of cellular-level information in most spatial data, it is necessary to integrate single-cell and spatial data. Here, we developed a deep learning computational framework for alignment and mapping of unpaired single-cell and the spatial data by using adversarial joint-variational autoencoder and Random Forest model (SPCC). SPCC will generate a mapping matrix for single-cell and spatial positions that can transfer spatial location to individual cells. SPCC can be used to perform downstream analysis, such as cell-cell communication and spatial variable gene identification with pseudo-space information. Then, we validated the performance with current popular integration methods on three datasets with different spatial sequencing techniques, including the mouse somatosensory cortex data, breast cancer data and melanoma brain metastasis data. Our findings suggest that SPCC exhibits greater robustness against sequencing noise compared to previous methods. SPCC has successfully captured the precise spatial structure in multiple datasets. For example, our results indicated that PECAM1 can interact with SOX4 between cancer cells and endothelial cells, which was rarely identified by using previous tools. The PECAM1-SOX4 interaction can regulate the vascular adhesion in melanoma and further contribute to tumor cell metastasis. These results show that SPCC is highly sensitive and accurate to identify unique spatial cell interactions. SPCC is freely accessible at https://github.com/ploughhh/SPCC.