SpaBiT: Enhancing Spatial Transcriptomics Resolution via Bidirectional Attention Transformers

Spatial transcriptomics (ST) enables the simultaneous measurement of gene expression and spatial locations within tissue sections, providing a powerful framework for studying spatial heterogeneity and tissue microenvironments. However, current spot-based ST platforms (e.g., 10x Genomics Visium) suffer from limited spatial resolution, sparse sampling density, and high experimental costs, which restrict their widespread use in large-scale biomedical studies.To address these challenges, we introduce SpaBiT, a multimodal deep learning framework that integrates H&E-stained histology images with spot-level ST data to predict and enhance spatial gene expression. SpaBiT leverages a general-purpose pathology foundation model (UNI) to extract high-dimensional image features, a graph attention network (GAT) to learn neighborhood-aware expression embeddings, and a bidirectional cross-attention + Transformer architecture to fuse multimodal information and infer high-density spatial gene expression maps.Experimental evaluations across multiple datasets from different ST platforms demonstrate that SpaBiT outperforms representative image-driven and expression-driven baselines in predicting gene expression at unmeasured locations. Moreover, SpaBiT enhances spatial gene expression patterns, preserves underlying spatial structure, and facilitates downstream analyses such as spatial domain identification, tumor region detection, and functional enrichment.空间转录组学（Spatial transcriptomics, ST）能够同时检测组织切片内的基因表达水平与空间位置信息，为研究空间异质性及组织微环境提供了强有力的研究框架。然而，当前基于斑点的ST平台（例如10x Genomics Visium）存在空间分辨率有限、采样密度稀疏及实验成本高昂的问题，这限制了其在大规模生物医学研究中的广泛应用。为解决上述挑战，本研究提出SpaBiT——一种整合苏木精-伊红（H&E）染色组织学图像与斑点级ST数据的多模态深度学习框架，用于预测并增强空间基因表达。SpaBiT采用通用病理基础模型（UNI）提取高维图像特征，借助图注意力网络（Graph Attention Network, GAT）学习邻域感知的表达嵌入，并通过双向交叉注意力+Transformer架构融合多模态信息，进而推断出高密度的空间基因表达图谱。针对来自不同ST平台的多组数据集开展的实验评估结果表明，在预测未检测位点的基因表达方面，SpaBiT的性能优于主流的图像驱动与表达驱动基线模型。此外，SpaBiT能够优化空间基因表达模式，保留原始空间结构，并可辅助下游分析任务，例如空间域识别、肿瘤区域检测以及功能富集分析。