DSSC model architecture. The inputs of DSSC are the gene expression matrix and the cell coordinates. The outputs of DSSC are the low-dimension latent space (32D) and the predicted labels. Briefly, DSSC learns a low-dimensional representation of the gene expression matrix while simultaneously leveraging the prior knowledge from the spatial coordinates of cells/spots and the marker genes. Clustering is performed on latent space. Constraint loss, reconstruction loss, and clustering loss are optimized simultaneously. ML loss and CL loss are optimized alternately. (BN) Batch normalization, (ELU) ELU activation, (ML) must-links constraints, (CL) cannot-link constraints, (ZINB) zero-inflated negative binominal.
