Path2ST: Hierarchical Cross-Modal Translation for Spatial Transcriptomics
A novel AI framework named Path2ST has been developed to forecast spatial gene expression using hematoxylin and eosin (H&E)-stained images, presenting a budget-friendly alternative to spatial transcriptomics (ST). This method, outlined in an arXiv paper (2608.14710), approaches H&E-to-ST prediction as a task of cross-modal semantic translation, overcoming the shortcomings of current methods that overlook the biological hierarchy within tissues. Path2ST features a Hierarchical Cell-Tissue Conditioning mechanism, which merges both explicit and implicit cellular attributes with tissue-level semantic insights, along with a Scale-Adaptive Autoregressive Generation process that utilizes a hierarchical semantic vocabulary, facilitating biologically consistent predictions from coarse to fine. This framework aims to harness the natural organization of cell types in functional tissue microenvironments that dictate local gene expression patterns, making it significant for computational pathology and genomics by potentially lowering the costs and complexities associated with spatial transcriptomics research.
Key facts
- Path2ST is a new AI framework for predicting spatial gene expression from H&E-stained images.
- It is described in a paper on arXiv with identifier 2608.14710.
- The method treats H&E-to-ST prediction as a cross-modal semantic translation task.
- It uses a Hierarchical Cell-Tissue Conditioning mechanism.
- It employs a Scale-Adaptive Autoregressive Generation process.
- The framework aims to capture the biological hierarchy of tissues.
- It offers a cost-effective alternative to spatial transcriptomics.
- The paper was announced as a cross type on arXiv.
Entities
Institutions
- arXiv