MSGR: Gene Ontology-Guided Hierarchical Spatial Gene Expression Prediction
Researchers have introduced a novel artificial intelligence model known as MSGR (Multi-Scale Gene Refiner) to forecast spatial gene expression using histopathology images. This model is detailed in a preprint on arXiv (2608.00405) and employs the Gene Ontology (GO) as a structured prior to enhance prediction precision. In contrast to current techniques that view target genes as a flat vector, MSGR arranges genes within a four-tier GO tree, utilizing a GO-guided decoder that incrementally sharpens predictions from broad functional categories to specific genes through residual corrections with scale-weighted supervision. This method effectively tackles the issue of deducing inter-gene relationships from limited paired datasets. The model exclusively functions on the gene aspect, with its GO-guided decoder being a crucial element.
Key facts
- MSGR (Multi-Scale Gene Refiner) is a new AI model for predicting spatial gene expression from histopathology images.
- The model incorporates Gene Ontology (GO) as an explicit structural prior.
- MSGR organizes target genes into a four-level GO tree.
- The GO-guided decoder refines predictions from coarse functional domains to fine individual genes.
- Residual corrections and scale-weighted supervision are used in the refinement process.
- The method aims to overcome limitations of flat, unstructured gene decoding in existing methods.
- The research is available as a preprint on arXiv with identifier 2608.00405.
- The announcement type is 'new'.
Entities
Institutions
- arXiv