Spatial Transcriptomic Alignment, Integration, and de novo 3D Reconstruction by STAIR

Spatial transcriptomics (ST) has emerged as a transformative approach for comprehending tissue architecture with molecular profiles. However, amalgamating discrete two-dimensional (2D) ST snapshots into a unified 3D atlas remains an outstanding challenge. To this end, we introduce STAIR, an end-to-end solution for ST alignment, integration, and de novo 3D reconstruction. STAIR uses a heterogeneous graph attention network with spot-level and slice-level attention mechanisms to obtain a unified embedding space and guide 3D space reconstruction in an unsupervised manner. We demonstrate STAIR's marked improvements in slice alignment and integration across samples and platforms over the previous methods. Furthermore, STAIR shows first-of-its-kind performance in de novo 3D reconstruction, with demonstrations in mouse hypothalamus preoptic area, mouse brain, and breast tumor tissue, which provides precise delineation of brain regions and reveals tumor progression in 3D space. Additionally, STAIR integrates additional slices into the existing 3D atlas incorporating both molecular features and physical coordinates. STAIR is the first to address the core obstacles limiting 3D positioning and harnessing alignments for atlas construction and assimilation. It lays a computational foundation to construct unified tissue maps and provides novel biological insights from a 3D perspective.