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

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Spatial transcriptomics (ST) offers insights into tissue architecture and cellular communication by measuring gene expression with spatial context.
• Current ST methods struggle to simultaneously capture spatial topology and transcriptional heterogeneity, limiting biological interpretability.
• There is a need for advanced computational frameworks to enhance the analysis of ST data.

Purpose of the Study:

• To introduce stCAMBL, a novel biased multi-view contrastive framework for high-fidelity spatial transcriptomics analysis.
• To improve the joint modeling of spatial topology and transcriptional heterogeneity in ST data.
• To develop a robust method for learning biologically informed and noise-resistant embeddings from ST data.

Main Methods:

• stCAMBL utilizes a variational graph autoencoder backbone.
• The framework integrates spatial graph structure modeling with attentive feature masking.
• Partial contrastive regularization is employed to emphasize informative molecular features and mitigate confounding patterns.

Main Results:

• stCAMBL demonstrated substantial improvements in clustering accuracy on multiple 10x Visium datasets.
• The framework enhanced gene ontology enrichment analysis, providing deeper biological insights.
• stCAMBL showed significant improvements in signal restoration and demonstrated strong generalizability for ST analysis.

Conclusions:

• stCAMBL effectively addresses limitations in existing ST approaches by jointly modeling spatial and transcriptional information.
• The proposed framework learns robust, biologically relevant embeddings, enhancing the interpretability of ST data.
• stCAMBL represents a significant advancement for high-fidelity spatial transcriptomics analysis, applicable across diverse datasets.