STCF: Multi-View Clustering for Spatial Transcriptomics Based on Cross-View Fusion

Spatial transcriptomics has revolutionized the ability to investigate transcriptional patterns within tissue morphology. However, many ST clustering pipelines operate on a single preselected gene set, typically prioritizing either highly variable genes (HVGs) or spatially variable genes (SVGs), and therefore may not directly model how genes with different levels of global variability provide complementary cues for spatial domain identification. Although non-HVG signals can be partially captured through SVG selection and spatial graph modeling, a dedicated two-view formulation that disentangles high-variance and low-variance gene subsets and fuses them under a unified objective remains underexplored. To this end, we propose a Spatial Transcriptomics clustering framework for Cross-view information Fusion, termed STCF, which casts HVGs and low-variability genes (LVGs) as two gene-expression views and integrates them via a plug-and-play cross-view fusion strategy. Specifically, STCF introduces a cross-view fusion mechanism that employs reverse-scaled cosine error loss (R-SCE) to balance alignment and separation of gene embeddings, ensuring robust representation learning while preserving spatial coherence, which enhances the model's ability to resolve fine-grained spatial structures. Extensive experiments on three benchmark datasets (DLPFC, HBC, and MBA) demonstrate the superiority, effectiveness, and transferability of STCF. Case studies further validate its ability to uncover latent spatial patterns and improve clustering precision.