THiRST: Computer Vision Empowered Three-dimensional High-Resolution Spatial Transcriptomics Analytics

Abstract

High-resolution spatial transcriptomics (ST) preserves gene expression within tissue architecture but remains limited to individual two-dimensional (2D) sections, restricting analysis of continuous three-dimensional (3D) structures and cross-section consistency. We present THiRST, a computational framework for reconstructing multi-resolution 3D spatial transcriptomic representations from serial sections. THiRST combines coarse-to-fine morphology-guided registration, spatial matching of ST units, cross-section association of segmented structures, and integration of Visium V2 and high-definition ST. In serial kidney sections, patch-level refinement increased the same-cluster proportions of matched units at adjacent offsets from 63.8% and 65.6% to 70.7% and 70.0%, respectively, while improving expression concordance. The reconstructed correspondences enabled 3D glomerular and object-level expression analysis, with 54.1% of differential genes detected only in 3D. Cross-platform integration further aligned V2 and high-definition measurements in a shared expression space. THiRST demonstrates the feasibility of object-level 3D ST analysis from serial 2D sections and underscores the importance of accurate local registration.