stGRL: spatial domain identification, denoising, and imputation algorithm for spatial transcriptome data based on multi-task graph contrastive representation learning

• Published in: BMC biology Vol. 23; no. 1; pp. 177 - 23
• Main Authors: Lu, Xin, Zhou, Murong, Gao, Bo, Wang, Fang, Jin, Shuilin, Liu, Qiaoming, Wang, Guohua
• Format: Journal Article
• Language: English
• Published: London BioMed Central 01.07.2025; BioMed Central Ltd; Springer Nature B.V; BMC
• Subjects: Algorithms; Analysis; Benchmarks; Binomial distribution; Biomedical and Life Sciences; Breast cancer; Carcinoma; Care and treatment; Clustering; Contrastive learning; Data analysis; Datasets; Denoising; Diagnosis; Domains; Encoders-Decoders; Gene expression; Gene Expression Profiling - methods; Graph neural network; Graph neural networks; Graphical representations; Health aspects; Heterogeneity; Humans; Identification; Immune system; Imputation; Life Sciences; Machine learning; Methods; Neural networks; Neural Networks, Computer; Noise reduction; Ovarian cancer; Risk factors; Sparsity; Spatial analysis; Spatial data; Spatial discrimination learning; Spatial domain identification; Spatial heterogeneity; Spatial transcriptomics; Therapeutic targets; Transcriptome; Transcriptomes; Transcriptomics; China; Imputation; Spatial transcriptomics; Spatial domain identification; Graph neural network; Contrastive learning; Denoising
• ISSN: 1741-7007; 1741-7007
• DOI: 10.1186/s12915-025-02290-z

Background Spatial transcriptomics now enables sequencing while preserving the spatial location of cells. This significantly enhances researchers' understanding of cellular and tissue functions in their spatial context. However, due to current technical limitations, spatial transcriptomics data often exhibit high dropout rates and noise, posing challenges for downstream analysis, like spot clustering, differential gene analysis, and spatial domain identification. To address those challenges, we propose stGRL, a novel deep multi-task graph neural network model tailored for spatial transcriptomics. stGRL employs an encoder-decoder architecture with a zero-inflated negative binomial (ZINB) distribution to reconstruct input data while effectively addressing dropout events. Additionally, it integrates graph contrastive representation learning to enhance the consistency of node embeddings, thereby improving clustering performance. Results Through benchmark experiments on various spatial transcriptomics datasets, stGRL demonstrated a superior ability to identify spatial features compared to current mainstream methods. In-depth analyses reveal that the denoised data generated by stGRL not only preserves the spatial hierarchy of tissues but also accurately identifies differentially expressed genes. When applied to breast cancer datasets, stGRL effectively analyzed the differences between cancerous regions and carcinoma in situ areas, uncovering that carcinoma in situ regions are predominantly regulated by the immune system, which limits cancer cell development through inflammatory responses. Additionally, in the spatial transcriptomics analysis of ovarian cancer, stGRL successfully annotated cell types, accurately identified B cell-enriched regions, and discovered a novel target gene, MZB1, with potential therapeutic value. Conclusions stGRL is an effective method for integrating multiple tasks in spatial transcriptome analysis. Our study highlights its broad applicability and outstanding performance in analyzing spatial transcriptome data. This method offers a powerful analytical tool for uncovering the spatial heterogeneity of complex tissues and identifying potential therapeutic targets for disease.