An Integrated and Robust Deep Learning Framework for Denoising and Analyzing Single‐Cell Spatial Transcriptomics

Spatial transcriptomics provides an unprecedented opportunity to study gene expression in its spatial tissue context, facilitating deeper insights into cellular organization and function. However, spatial transcriptomics data are often affected by high dropout noise, high dimensionality, and complex structure, posing significant challenges for downstream analysis. Here, we present Single‐cell Spatial Transcriptomics Analysis and Denoising Engine (scSTADE), an unsupervised deep learning framework for simultaneous denoising, clustering, and identifying functionally variable genes (FVGs). scSTADE employs a dual‐channel architecture that combines a linear denoising module with a graph convolutional network‐based nonlinear module, capturing both linear and nonlinear representation embeddings. It adaptively models gene distributions across spatial domains and leverages tissue‐level spatial context to identify FVGs at multiple spatial resolutions. Extensive benchmark analyses on diverse spatial transcriptomics datasets generated by different platforms demonstrate that scSTADE consistently outperforms eight popular methods across multiple clustering metrics. In brain tissue datasets, scSTADE achieves up to a 15% performance improvement and its inferred domains align closely with known neuroanatomical structures. In human breast cancer datasets, scSTADE reveals immune‐related tertiary lymphoid structures, with predicted FVGs supported by gene ontology enrichment, cell–cell communication analysis, and survival outcomes. scSTADE also exhibits high robustness against random initialization and varying levels of synthetic dropout noise.