Investigation of pair-wise single-cell interactions by statistically interpreting spatial cell state correlation learned by self-supervised graph inductive bias transformer

Abstract

Image-based spatial transcriptomics (ST) offers spatial gene expression profile at the single-cell resolution and provides information to understand intercellular communication that is critical for maintaining tissue development and organ function. Disruption of normal cell-cell interactions (CCI) can lead to disease onset and progression. Current CCI analysis methods face several limitations, including subjection to the number of measured ligand-receptor genes in image-based spatial transcriptomics, limited graph encoding power, inadequate use of spatial information, and low interpretability. Here, we present GITIII, an interpretable self-supervised graph transformer-based language model that treats cells as words (nodes) and their cell neighborhood as a sentence to explore the communications among cells. Enhanced by multilayer perceptron-based distance scaler, physics-informed attention mechanism, and a state-of-the-art, expressive, and lightweight graph transformer model, GITIII infers CCI by investigating how the state of a cell is influenced by the spatial organization, ligand expression, cell types and states of neighboring cells. With its interpretable architecture, GITIII can be used to understand how the sender cell influences target genes in the receiver cell, visualize the spatial pattern and utility of CCI, identify significant CCI networks, perform CCI-informed cell subtyping, and compare CCI strength between disease groups. Applications to four ST datasets from several species, organs, and platforms, GITIII effectively identified and quantitatively interpreted key CCI patterns driving within-sample heterogeneity and disease progression, thus improving our understanding of brain structures, tumor microenvironments, and the interplay among different cell types responding to neighboring CCIs.