variational spatiotemporal gaussian processを用いたマルチモーダル・マルチスライス空間トランスクリプトミクスからのコヒーレント機能ランドスケープの再構築

Spatial transcriptomics (ST) technologies are revolutionizing our ability to investigate the spatial organization of complex tissues. While ST has significantly advanced our understanding of tissue architecture, most analytical approaches remain restricted to 2D sections, limiting insights into the full 3D spatial context. Here, we introduce stVGP, a variational spatial Gaussian process framework designed to align, integrate, and reconstruct spatial coherent domains from multi-modal, multi-slice ST datasets. By integrating spatial Gaussian processes with spatially hierarchical transformers, stVGP enables accurate cross-slice alignment, robust batch effect correction, and the identification of biologically meaningful spatial domains. Critically, a key innovation of stVGP is its support for virtual tissue slices generation, allowing for continuous 3D reconstruction and interpolation of gene expression in unsampled regions. Comprehensive evaluations across diverse datasets demonstrate that stVGP consistently outperforms state-of-the-art methods in alignment accuracy, domain detection, and gene expression prediction. Furthermore, stVGP enables cross-modal generation of gene expression from histological images in human breast cancer samples, facilitating virtual transcriptomic reconstruction with high fidelity. Collectively, stVGP offers a unified, scalable framework for modeling 3D landscapes in complex tissues and developmental systems, bridging the gap between discrete 2D sections and continuous 3D biological insights.