Interpretable high-resolution dimension reduction of spatial transcriptomics data by DeepFuseNMF

Abstract

Spatial transcriptomics (ST) technologies have revolutionized tissue architecture studies by capturing gene expression with spatial context. However, high-dimensional ST data often have limited spatial resolution and exhibit considerable noise and sparsity, posing significant challenges in deciphering subtle spatial structures and underlying biological activities. Here, we introduce DeepFuseNMF, a multi-modal dimension reduction framework that enhances spatial resolution by integrating ST gene expression with high-resolution histology images. DeepFuseNMF incorporates non-negative matrix factorization into a neural network architecture, enabling the identification of interpretable, high resolution embeddings. Furthermore, DeepFuseNMF can simultaneously analyze multiple samples and is compatible with various types of histology images. Extensive evaluations on synthetic and real ST datasets from various technologies and tissue types demonstrate that DeepFuseNMF can effectively produce highly interpretable, high-resolution embeddings, and detects refined spatial structures. DeepFuseNMF represents a powerful approach for integrating ST data and histology images, offering deeper insights into complex tissue structures and functions.