Space-map is a machine learning based method for aligning and reconstructing three-dimensional multiplexed spatial data at single-cell resolution. The method combines non-rigid deformation models with multi-step alignment to accurately register sequential tissue sections. Using unsupervised learning, SpaceMap performs well even with limited samples. Benchmarks show improved alignment accuracy and speed, allowing detailed analysis of gene expression patterns across tissue sections. This approach advances our ability to study tissue organization and cell relationships in development and disease.

Key Features

• Multi-modal Registration: Combines cell coordinates, cell types, gene expression, and histological images for robust alignment
• Two-stage Registration Approach: Efficient coarse alignment followed by precise fine registration
• Advanced Feature Matching: Combines deep learning (LoFTR) with traditional computer vision methods (SIFT)
• GPU-accelerated LDDMM: Optimized for handling large-scale cellular data from multiple tissue sections
• Global Consistency: Ensures structural coherence between non-adjacent sections
• High Performance: ~2-fold more accurate than PASTE and STalign while running on a standard laptop

Code and usage

We provide a repository with our code and instructions on how to install and run it.

Data Availability

• The 3D Xenium Spatial Transcriptomics input data for the Space-Map study are available through the HTAN Data Portal:
  Link
• The 3D CODEX input data as well as the output data for the Space-Map study are provided via Globus:
  Link to folders using Globus file explorer (requires authentication)
  Link for direct download of archive without authentication (411GB)

The raw imaging data used as input to Space-map were generated in the Snyder lab at Stanford. They have been deposited through the HuBMAP and HTAN data portals and will be available shortly. 

Our first research article demonstrating the application of SpaceMap on 3D data produced with the Phenocycler / CODEX and Xenium assays has been submitted for publication.