STGAT: Graph attention networks for deconvolving spatial transcriptomics data

Background and Objective: Spatially resolved gene expression profiles are crucial for understanding tissue structure and function. However, the lack of single-cell resolution in these profiles demands their integration with single-cell RNA sequencing data for accurate dataset deconvolution. We propose STGAT, an innovative deconvolution method that leverages graph attention networks to enhance spatial transcriptomic (ST) data analysis. Methods: STGAT generates pseudo-ST data that more comprehensively represents the cell-type composition within real-ST data by using three different sampling probabilities. A comprehensive combined graph is then constructed to capture the complex relationships both across pseudo- and real-ST data and within each dataset. Moreover, integrating a graph attention network further enables STGAT to dynamically assign the weights to the connections between spots, significantly enhancing the accuracy of cell-type composition predictions. Results: Extensive comparative experiments on simulated and real-world datasets, demonstrate the superior performance of STGAT for cell-type deconvolution. The method outperforms six established methods and is robust across various biological contexts. Conclusion: STGAT exhibits more precise results in cell-type composition inference that are more consistent with the known knowledge, suggesting its potential utility in improving the resolution and accuracy of spatial transcriptomics data analysis.