PaSTel: Hierarchical Pretraining Framework for Spatial Transcriptomics
A new hierarchical multimodal pretraining framework called PaSTel has been developed by researchers to enhance spatial transcriptomics (ST) by incorporating biological priors across various scales. This approach tackles two significant challenges in current methods: the prevalence of housekeeping genes in gene selection, which results in poorly discriminative representations, and the inability of independent spot-patch alignment to recognize spatial dependencies. PaSTel functions on three tiers: spot level, where TF-IDF reweighting pinpoints spatially relevant genes; functional level, using curated KEGG pathways as anchors for overarching biological semantics; and regional level, where spatial clustering combines adjacent spots to represent meso-scale tissue structure. This framework is detailed in a paper available on arXiv (ID: 2608.14924), aimed at improving the correlation between histology images and gene expression data, thereby enhancing insights into tissue organization and molecular programs.
Key facts
- PaSTel is a hierarchical multimodal pretraining framework for spatial transcriptomics.
- It integrates biological priors at three levels: spot, functional, and regional.
- TF-IDF reweighting is used at the spot level to identify spatially informative genes.
- Curated KEGG pathways serve as anchors at the functional level.
- Spatial clustering aggregates neighboring spots at the regional level.
- The paper is available on arXiv with ID 2608.14924.
- The announcement type is 'cross'.
- The paper addresses limitations in existing ST pretraining methods.
Entities
Institutions
- arXiv