Fig. 1: Overview of the SIMVI framework.
From: SIMVI disentangles intrinsic and spatial-induced cellular states in spatial omics data
a We consider gene expression in spatial omics data to be generated by a (noisy) function of intrinsic (e.g., cell type) and spatial-induced (e.g., spatial gradient, cellular interaction) variations. Intrinsic variation may also exhibit spatial patterns by cell-type-specific spatial organization. b SIMVI workflow. Spatial omics data are transformed into graph-structured format. Intrinsic and spatial variations are estimated by multilayer perceptron (MLP) and graph attention network (GAT) variational posteriors. The loss function consists of the evidence lower bound (ELBO) and an asymmetric regularization term minimizing information encoded by intrinsic variation. Downstream applications of SIMVI include batch integration, niche identification, spatial effect estimation, and spatial process interpretation. It can also be used to infer multi-omics spatial effects in spatial multi-omics data.
