SpaBiT: enhancing spatial transcriptomics resolution via bidirectional attention transformers.

Area of Science:

• Computational biology
• Genomics
• Bioinformatics

Background:

• Spatial transcriptomics (ST) offers precise gene expression mapping within tissue architecture but faces limitations in spatial resolution and sampling density.
• Existing deep learning methods struggle to fully capture the interplay between histological context, spatial topology, and local neighborhood relationships.

Purpose of the Study:

• To develop a novel multimodal framework, SpaBiT, for enhancing ST resolution and generating high-density gene expression maps.
• To address limitations in current ST analysis by improving the modeling of histological context and spatial topology.

Main Methods:

• SpaBiT utilizes a bidirectional attention mechanism, specifically a bidirectional cross-attention module, for information exchange between image features and graph attention network-derived neighborhood representations.
• The framework explicitly models synergistic constraints between local morphology and spatial graph topology.

Main Results:

• SpaBiT successfully generates high-fidelity, high-density gene expression maps.
• The framework demonstrates competitive performance, outperforming benchmark models in reconstructing complex spatial gene expression patterns across various quantitative metrics.

Conclusions:

• SpaBiT provides a robust tool for enhancing ST resolution and deciphering complex tissue microenvironments.
• The proposed method offers significant improvements in spatial gene expression analysis.