Abstract
Spatial transcriptomics (ST) technologies revolutionize biomedical research by providing unprecedented insights into tissue architecture and disease mechanisms. While imaging-based ST technologies achieve single-cell spatial resolution, they face inherent limitations in gene detection capacity and measurement accuracy of expression profiles. Although computational approaches make notable progress, current methods remain challenged by insufficient integration of spatial context and systematic biases toward the single-cell RNA sequencing distribution. To address these limitations, EDGES is developed a spatially constrained non-negative matrix factorization framework that simultaneously predicts undetected gene expression and denoises measured transcriptional profiles. EDGES incorporates spatial information through graph Laplacian regularization while synergistically integrating cellular representations with gene-specific representations, thereby ensuring that the predicted gene expression aligns closely with the real ST distribution. Comprehensive evaluations demonstrate that EDGES achieves superior predictive performance and outperforms existing denoising methods. The framework's versatility further facilitates the identification of novel biological markers and spatially resolved expression patterns. With its innovative design, EDGES provides an advanced tool to enhance the reliability of the imaging-based ST data, facilitating more accurate and biologically meaningful interpretation of downstream discoveries.