Spatial Gene Expression Prediction Using Hierarchical Sparse Attention

Spatial Transcriptomics (ST) quantitatively interprets human diseases by providing the gene expression of each fine-grained spot (i.e., window) in a tissue slide. This paper focuses on predicting gene expression at windows on a tissue slide image of interest. However, gene expression related to image features usually exhibits diverse spatial scales. To spatially model these features, we newly introduced Hierarchical Sparse Attention Network (HSATNet). The core idea of HSATNet  is to employ a two-level sparse attention mechanism, namely coarse (i.e., area) and fine attention. Each HSAT Block consists of two main modules: i) An adaptive sparse coarse attention filters out the most irrelevant areas, resulting in the acquisition of adaptive sparse areas. ii) An adaptive sparse fine attention module filters out the most irrelevant windows, leading to the acquisition of adaptive sparse windows. The first module aims to identify similarities among windows within the same area, while the second module captures the differences between different windows. Particularly, after fusing these two modules together, without any additional training data or pre-training, experiments conducted on 10X Genomics breast cancer data show that our HSATNet  achieves an impressive PCC@S of 7.43 for gene expression prediction. This performance exceeds the current state-of-the-art model. Code is available at https://github.com/biyecc/HSATNet .