Integrating spatial gene expression and breast tumour morphology via deep learning

Area of Science:

• Computational biology
• Pathology
• Genomics

Background:

• Spatial transcriptomics enables linking cellular morphology to gene expression.
• Histopathology images contain rich information about tissue architecture.
• Predicting gene expression from images could advance biomarker discovery.

Purpose of the Study:

• To develop a deep learning algorithm for predicting local gene expression from histopathology images.
• To identify genes whose expression can be predicted from tissue morphology.
• To assess the generalizability of the algorithm across different datasets.

Main Methods:

• Utilized a dataset of 30,612 spatially resolved gene expression data matched to histopathology images from 23 breast cancer patients.
• Developed and trained a deep learning algorithm to predict gene expression from haematoxylin-and-eosin-stained images.
• Validated the algorithm's performance on The Cancer Genome Atlas and other breast cancer datasets.

Main Results:

• Identified over 100 genes, including breast cancer biomarkers, whose expression is predictable from histopathology images at 100 µm resolution.
• Demonstrated the co-localization of tumor growth and immune activation markers with predicted gene expression patterns.
• Showed that the algorithm generalizes well to external datasets without re-training.

Conclusions:

• Deep learning can accurately predict spatially resolved gene expression from histopathology images.
• This approach facilitates the discovery of molecular biomarkers based on spatial variation within tumors.
• Image-based prediction of the transcriptome holds promise for high-throughput biomarker screening.