Abstract
Spatial transcriptomics provides a unique opportunity for understanding cellular organization and function in a spatial context. However, spatial transcriptome exists the problem of dropout noise, exposing a major challenge for accurate downstream data analysis. Here, we proposed DenoiseST, a dual-channel unsupervised adaptive deep learning-based denoising method for data imputing, clustering, and identifying functionally variable genes in spatial transcriptomics. To leverage spatial information and gene expression profiles, we proposed a dual-channel joint learning strategy with graph convolutional networks to sufficiently explore both linear and nonlinear representation embeddings in an unsupervised manner, enhancing the discriminative information learning ability from the global perspectives of data distributions. In particular, DenoiseST enables the adaptively fitting of different gene distributions to the clustered domains and employs tissue-level spatial information to accurately identify functionally variable genes with different spatial resolutions, revealing their enrichment in corresponding gene pathways. Extensive validations on a total of 19 real spatial transcriptome datasets show that DenoiseST obtains excellent performance and results on brain tissue datasets indicate it outperforms the state-of-the-art methods when handling artificial dropout noise with a remarkable margin of ~ 15%, demonstrating its effectiveness and robustness. Case study results demonstrate that when applied to identify biological structural regions on human breast cancer spatial transcriptomic datasets, DenoiseST successfully detected biologically significant immune-related structural regions, which are subsequently validated through Gene Ontology (GO), cell-cell communication, and survival analysis. In conclusion, we expect that DenoiseST is a novel and efficient method for spatial transcriptome analysis, offering unique insights into spatial organization and function.