A deep learning model to predict RNA-Seq expression of tumours from whole slide images

Area of Science:

• Computational pathology
• Genomics
• Artificial intelligence in medicine

Background:

• Deep learning in digital pathology aids clinical questions, including diagnosis and treatment outcome prediction.
• Previous studies explored predicting gene mutations from pathology images, but comprehensive evaluation for molecular feature extraction is lacking.

Purpose of the Study:

• To evaluate the potential of deep learning for extracting molecular features from histology slides.
• To develop and validate a model for predicting RNA-Seq profiles from whole-slide images.

Main Methods:

• HE2RNA, a multi-modal data integration model, was trained to predict RNA-Seq profiles from whole-slide images without expert annotation.
• The model's interpretable design allowed for virtual spatialization of gene expression, validated by CD3 and CD20 staining.
• Transcriptomic representations learned by HE2RNA were transferable to other datasets to enhance prediction of molecular phenotypes.

Main Results:

• HE2RNA successfully predicted RNA-Seq profiles from whole-slide images alone.
• Virtual spatialization of gene expression was achieved and validated.
• The model demonstrated transferability of learned transcriptomic representations for improved molecular phenotype prediction.

Conclusions:

• HE2RNA offers a systematic approach to predict RNA-Seq profiles from digital pathology images.
• The model facilitates virtual spatialization of gene expression and aids in clinical diagnosis, such as identifying tumors with microsatellite instability.