The proposed SCGDL owns a two-levels architecture, as shown in Figure 1. The feature and adjacent matrices are regarded as two incentives in the higher-level framework. Based on the built-in gene expression profiles, the feature matrix is calculated. It indi-cates the inclusion relation of spots and genes. The adjacent matrix is derived according to the positional information of spots. A spa-tial neighbor graph (SNG) is capable of being delineated during the generation of an adjacent matrix. These two inputs are conducted by a DGI module with four layers of RGGCNN. Finally, the low-dimensional latent embeddings are acquired to imply the spots representation at the higher-level.
running SCGDL_Tutorial.ipynb to see the simulation results step by step
Distributed under the MIT License. See LICENSE.txt
for more information.
data files
Please down load the spatial transcriptomics data from the provided links.
Porch_pyg
Please follow the instruction to install pyG and geometric packages.