FmH2ST: foundation model-based spatial transcriptomics generation from histological images.

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Spatial transcriptomics (ST) provides tissue-level gene expression data.
• Predicting spatial gene expression from histology faces challenges like limited data and capturing complex tissue heterogeneity.

Purpose of the Study:

• To introduce FmH2ST, a novel foundation model-based method for enhanced spatial gene expression prediction.
• To address limitations in current ST data and improve the accuracy of gene expression mapping.

Main Methods:

• FmH2ST utilizes a dual-branch framework integrating foundation models and spot-specific learning.
• It employs multilevel feature extraction, dual-graph strategies, multiscale convolutions, Transformers, and graph neural networks.
• Adaptive fusion of features from both branches enhances prediction accuracy.

Main Results:

• FmH2ST significantly outperforms state-of-the-art methods on benchmark datasets.
• The model effectively denoises gene expression data and reveals spatial heterogeneity in cancer.
• It identifies gene co-expression patterns and supports gene regulatory network inference.

Conclusions:

• FmH2ST is a highly effective method for spatial gene expression prediction.
• The approach has potential applications in clinical diagnostics and personalized medicine through improved ST analysis.