iHAST: Integrating Hybrid Attention for Super-Resolution in Spatial Transcriptomics

Abstract

Spatial transcriptomics (ST) technologies have transformed genomic research by facilitating spatially-resolved gene expression profiling, offering unprecedented opportunities to explore cellular communication and organization. However, most existing ST assays are constrained to resolving cell clusters or multicellular structures, resulting in the amalgamation of signals from multiple cells and obscuring the spatial dynamics of diverse cell populations. To address this limitation, we conceptualize ST data as a specific type of image, where cells represent pixels and genes are akin to channels. Consequently, we propose a novel Hybrid Attention Transformer framework, named iHAST, to enhance the spatial resolution of ST data. In support of this framework, we have curated an extensive ST dataset comprising over 800 examples from diverse cell types and sequencing technologies for training and testing purposes. Subsequently, we conducted rigorous evaluations of our approach and consistently observed performance enhancements surpassing those achieved by state-of-the-art methods, underscoring the robustness and generalizability of our iHAST model across a diverse array of biological contexts and experimental conditions.