Spatial-context-aware RNA-sequence prediction from head and neck cancer histopathology images

Area of Science:

• Computational pathology
• Genomics
• Artificial intelligence in oncology

Background:

• Molecular profiling complements histological analysis for targeted cancer therapies.
• RNA-sequence expression (RNA-seq) data is often inaccessible due to processing delays, cost, or availability issues.
• Accurate RNA-seq prediction from histopathology is crucial for advancing personalized cancer treatment.

Purpose of the Study:

• To develop a deep learning framework for predicting RNA-seq from Hematoxylin and Eosin whole-slide images (H&E WSI) in head and neck cancer.
• To overcome limitations of patch-by-patch methods by preserving spatial-contextual relationships.
• To enable cost-effective and timely discovery of genetic biomarkers from histopathology slides.

Main Methods:

• Proposed a novel framework utilizing a neural image compressor to maintain spatial relationships and generate a compressed whole-slide image representation.
• Employed a customized deep-learning regressor to predict RNA-seq from the compressed representation, learning both global and local features.
• Validated the method on the TCGA-HNSC dataset, predicting RNA-seq for 10 oncogenes in 43 patients.

Main Results:

• The proposed method achieved a 4.12% higher mean correlation compared to a state-of-the-art baseline.
• Successfully predicted 6 out of 10 genes with improved correlation.
• Demonstrated interpretability through pathway analysis and activation maps highlighting salient image regions for prediction.

Conclusions:

• The deep learning framework effectively predicts RNA-seq from H&E WSI, outperforming existing methods.
• This approach offers a potential solution for discovering genetic biomarkers directly from histopathology images.
• The method could facilitate pre-screening of patients, reducing costs and time associated with genetic testing.