Spatial transcriptomic alignment, integration, and 3D reconstruction by STAIR

Merging multiple slices into a unified 3D atlas is a significant challenge in spatial transcriptomics. Here, we introduce STAIR, an end-to-end solution for alignment, integration, and 3D reconstruction. STAIR employs a heterogeneous graph attention network with spot-level and slice-level attention mechanisms to achieve a unified embedding space and guide unsupervised 3D reconstruction. We demonstrate STAIR's marked improvements in feature integration and 2D alignment across samples and platforms compared to previous methods. Furthermore, STAIR shows first-of-its-kind performance in z-axis reconstruction of parallel slices and seamlessly integrates new slices into existing 3D atlases, providing novel biological insights from a 3D perspective.