A collaborative graph–transformer network for spatial transcriptomics data clustering with semantic graph induction

Jiangkai Long

doi:10.1016/j.engappai.2026.115097

A collaborative graph–transformer network for spatial transcriptomics data clustering with semantic graph induction

2026-06-13

0

PRE

AI

DOI：10.1016/j.engappai.2026.115097

Original

Original request for help

Save

Abstract

En 中文

Spatial transcriptomics enables the simultaneous capture of gene expression profiles and spatial localization, offering critical insights that facilitate the deciphering of tissue architecture and cell–cell interactions. To fully exploit spatial transcriptomic data, effective modeling of spatial relationships is essential. However, current methods exhibit limitations when constructing graph structures. Traditional static topological relationships built on spatial proximity struggle to delineate cell-layer boundaries and fail to adequately identify the cross-regional distribution patterns of cells within the same layer. Meanwhile, those based on raw gene expression similarity are susceptible to noise and cellular heterogeneity, making it difficult to capture the underlying semantic relationships. To address these challenges, we propose GTFST, a graph–Transformer network for spatial transcriptomics clustering. GTFST employs a dual-branch architecture with designed Reciprocal Contrastive Alignment (RCA) strategy to comprehensively model both local microenvironmental interactions and global contextual semantics. Notably, we introduce a Semantic Graph Induction (SGI) module that automatically learns and induces graph topology from gene semantic relationships. This design overcomes the limitations of fixed-neighborhood approaches, enabling more insightful modeling of complex biological interactions. Extensive experiments on spatial transcriptomic datasets demonstrate the superiority of GTFST over state-of-the-art methods in spatial domain clustering. Its enhanced precision provides a robust foundation for subsequent biological analyses such as differentially expressed gene identification, functional enrichment, and cell–cell interaction inference.

A collaborative graph–transformer network for spatial transcriptomics data clustering with semantic graph induction

Abstract

Journal

Organization