Figure 1.

Overview of our method

(A) The computational framework. Our method utilized a framework with two key components: an encoder and a projection head. Both components were implemented as MLP. During training, we harnessed spatial transcriptomics (ST) samples as the training dataset. For each sample X, we selected a positive pair from its nearest neighbors and chose n − 2 random samples as negative pairs. The model’s neural network takes log-normalized gene expressions as input and is optimized through the infoNCE loss, which minimizes the distance between positive pair projections and maximizes the separation of negative pairs.

(B) Spatial reconstruction workflow. The gene profiles of SC samples are processed by the trained encoder to generate representations. These representations enable the identification of cell neighbors and global cell-cell relationships by comparing pairwise SC-SC representation similarities. Moreover, the SC-ST representation similarities was included to locate the most matched ST sample’s coordinates, effectively predicting the SC sample’s spatial location based on the reference dimension.