Breast cancer histopathology image-based gene expression prediction using spatial transcriptomics data and deep...

Area of Science:

• Oncology
• Computational Biology
• Medical Imaging

Background:

• Tumor heterogeneity in breast cancer complicates treatment and prognosis.
• Spatial transcriptomics provides detailed gene expression data but is costly for large studies.
• Hematoxylin and eosin (H&E) stained histology images offer an affordable alternative for gene expression prediction.

Purpose of the Study:

• To develop a deep learning framework, BrST-Net, for predicting gene expression from histopathology images.
• To evaluate the performance of various deep learning architectures for this task.
• To improve the prediction of gene expression for better clinical oncology applications.

Main Methods:

• Developed BrST-Net, a deep learning framework utilizing spatial transcriptomics data.
• Trained and evaluated four architectures: ResNet101, Inception-v3, EfficientNet, and vision transformer, without pretrained weights.
• Incorporated an auxiliary network to enhance the generalization performance of the main prediction network.

Main Results:

• Successfully predicted gene expression from histopathology images for 250 genes.
• Achieved positive correlation for 237 genes, significantly outperforming previous methods.
• Identified 24 genes with a median correlation coefficient greater than 0.50, a substantial improvement over prior studies.

Conclusions:

• BrST-Net provides a powerful and cost-effective method for predicting gene expression from H&E stained breast cancer images.
• This approach can facilitate large-scale clinical oncology studies by overcoming the cost limitations of spatial transcriptomics.
• The enhanced prediction accuracy holds promise for improving breast cancer outcome and therapy response prediction.