Spatial Gene Expression Prediction Using Coarse and Fine Attention Network

Spatial Transcriptomics (ST) quantitatively interprets human diseases by providing the gene expression of each fine-grained spot (i.e., window) on a tissue slide. This paper focuses on predicting gene expression in specific windows on a tissue slide image. However, gene expression related to image features typically exhibits diverse spatial scales. To spatially model these features, we propose the Coarse and Fine Attention Network (CFANet). At the coarse level, we employ a coarse-to-fine strategy to acquire adaptable global features. Through coarse-gained areas (i.e., area) guiding to realize sparse external window attention by filtering out the most irrelevant feature areas. At the fine level, using dynamical convolutions realizes internal window attention to obtain dynamic local features. By iterating our CFAN Block, we construct features for different gene types within the slide image windows to predict gene expression. Particularly, without any pre-training, on 10X Genomics breast cancer data, our CFANet  achieves an impressive PCC@S of 81.6% for gene expression prediction, surpassing the current SOTA model by 5.6%. This demonstrates the potential of the model to be a useful network for gene prediction. Code is available ( https://github.com/biyecc/CFANet ).