Spatial transcriptomics prediction from histology jointly through Transformer and graph neural networks.

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Spatial transcriptomics enables high-resolution RNA abundance measurement alongside histology images.
• Predicting gene expression from accessible histology images is promising but challenging.
• Existing methods struggle to capture complex 2D vision features and spatial dependencies.

Purpose of the Study:

• To develop a deep learning model, Hist2ST, for predicting RNA-seq gene expression from histology images.
• To improve the integration of visual and spatial information for more accurate predictions.
• To generate spatial transcriptomics data from histology for tissue molecular signature elucidation.

Main Methods:

• Hist2ST utilizes a convolutional module for 2D vision feature extraction from image patches.
• Transformer and graph neural network modules capture global and local spatial relationships.
• A self-distillation mechanism is employed to address limitations of small spatial transcriptomics datasets.

Main Results:

• Hist2ST outperforms existing methods in predicting gene expression from histology images.
• The model demonstrates superior performance in spatial region identification.
• Pathway analyses confirm the preservation of biological information by Hist2ST.

Conclusions:

• Hist2ST effectively predicts spatial gene expression using histology images.
• The model advances the integration of imaging and transcriptomics data.
• Hist2ST facilitates the generation of spatial transcriptomics data for biological discovery.