Fig. 1: Overview of the HERGAST model.
a We divide the spatial transcriptomics data into patches and iteratively train HERGAST on them; inference is then conducted on the entire slice. b To avoid over-smoothing, we use the heterogeneous graph neural network as the core model, introducing gene expression profile similarity between different spots as a means of establishing connections through which information can flow during training. By linking central spots to border spots within a patch (s14 to s16 and s8 to s11) and implicitly connecting border spots to border spots in adjacent patches (s16 to s11) based on expression similarities, implicit connections between spots across different patches are created (e.g., the orange dash-lined circle). Spatial neighborhood relationship is added as another connection between spots to make the output of model locally-aware (e.g., s2 to s3 and s4, the blue dash-lined circle). Each spot’s profile is transformed into a latent embedding by an encoder and reconstructed using a linear decoder. c By considering gene expression similarity and spatial proximity, HERGAST generates low-dimensional embeddings that enable fine-scale spatial clustering. The reconstructed expression profile serves as amplified gene expression signal.
