Biased multi-view contrastive learning with attentive masking for spatial transcriptomic analysis

Abstract Spatial transcriptomics (ST) enables the simultaneous measurement of gene expression and spatial context, offering unprecedented insights into tissue architecture and cellular communication. However, existing approaches often fail to jointly capture spatial topology and transcriptional heterogeneity, leading to suboptimal representations and limited biological interpretability. To address this limitation, we propose stCAMBL, a biased multi-view contrastive framework that integrates spatial graph structure modeling with attentive feature masking and partial contrastive regularization. Built upon a variational graph autoencoder backbone, stCAMBL learns biologically informed and noise-robust embeddings by adaptively emphasizing informative molecular features while mitigating confounding patterns across spatial domains. Comprehensive evaluations on multiple 10$\times$ Visium datasets demonstrate that stCAMBL substantially improves clustering accuracy, gene ontology enrichment, and signal restoration, demonstrating strong generalizability for high-fidelity ST analysis.