THItoGene: a deep learning method for predicting spatial transcriptomics from histological images

Area of Science:

• Computational Biology
• Genomics
• Pathology

Background:

• Spatial transcriptomics provides insights into cellular regulation but is expensive.
• Current AI methods for predicting spatial gene expression from histology lack deep information extraction capabilities.

Purpose of the Study:

• To develop an affordable and effective method for predicting spatial gene expression from histological images.
• To explore the relationship between high-resolution pathology image phenotypes and gene expression regulation.

Main Methods:

• Developed THItoGene, a hybrid neural network combining dynamic convolutional and capsule networks.
• Utilized deep learning to adaptively sense molecular signals within histological images.
• Evaluated performance on human breast cancer and cutaneous squamous cell carcinoma datasets.

Main Results:

• THItoGene demonstrated superior performance in spatial gene expression prediction compared to existing methods.
• The model successfully deciphered spatial context and enrichment signals within specific tissue regions.
• Validated on diverse human cancer datasets, showcasing robust predictive power.

Conclusions:

• THItoGene offers a cost-effective and accurate solution for spatial gene expression prediction from histology.
• The AI tool can reveal complex gene regulation patterns and spatial tissue information.
• This approach enhances the utility of pathological images for genomic research.