Spatial domain identification based on variational autoencoder and single sample network

Due to its ability to reveal tissue heterogeneity, spatial analytic transcriptomic data has been used to decipher the spatial domain of complex diseases for precise treatment. At present, the problems of high dimensional, sparse and noisy data in the application process are awaiting to be solved. However, it has not fully leveraged the intercellular molecular interactions, which affect the accuracy of spatial domain identification. In order to solve this problem, the stMVC model is improved by integrating the single sample network method and variational autoencoder, and the spatial domain(SDI-VASSN) is extracted from the transcriptome data of multimodal spatial decomposition. Specifically, the model uses multimodal data of the human dorsolateral prefrontal cortex obtained through 10X Genomics Visium technology. Firstly, from the samples, we used cell specific molecular interaction network (CSN) to calculate the gene interaction network of each cell and extract key genes; Then we encode key genes by one-hot coding; Finally, we use variational autoencoder (VAE) instead of autoencoder (AE) to maximize the probability of the input data to learn the probability distribution of the data, thus input the resulting data into the stMVC model for identifying the spatial domain.