CASTLE method improves spatial domain detection in tissue by incorporating cell-type information
Researchers at PLOS Genetics describe CASTLE, a contrastive learning approach that uses cell-type composition to sharpen the delineation of spatial domains in spatially resolved transcriptomics data.
Published in PLOS Genetics on 17 November 2026, a study by Lai Wei, Aoqi Xie, and Yuehua Cui introduces CASTLE (Cell-type Aware SpaTial domain detection via contrastive Learning Embedding), a computational method for identifying spatial domains in spatially resolved transcriptomics datasets.
Spatially resolved transcriptomics technologies measure gene-expression profiles while preserving information about where cells are located within a tissue section. Identifying coherent spatial domains — regions of tissue with similar gene-expression programmes — is a central analytical task, but accurately delineating domain boundaries remains difficult, particularly where cell-type composition changes gradually rather than sharply.
CASTLE addresses this by leveraging cell-type proportion estimates alongside gene-expression similarity within a graph neural network (GNN) framework, using contrastive learning to improve the separation of domain boundaries. The authors benchmark CASTLE against existing GNN-based approaches across multiple platforms and resolutions, reporting improved accuracy in domain boundary delineation.
The method is a computational tools paper of primary interest to researchers developing or applying spatial transcriptomics pipelines. It has no direct clinical application at present but is relevant to those studying the tumour microenvironment, neural tissue architecture, or developmental biology contexts where spatial organisation is biologically meaningful.
Sources
Read the original reporting — these are the public sources this summary draws from.
-
Primary source PLOS Genetics · 2026-11-17CASTLE: Cell-type aware SpaTial domain detection via contrastive learning embedding