Cancer drug sensitivity prediction from routine histology images

Area of Science:

• Computational Biology
• Digital Pathology
• Oncology

Background:

• Drug sensitivity prediction is crucial for personalized cancer therapy but often relies on expensive clinical trial data.
• Existing models face challenges due to the time and cost associated with acquiring survival data.

Purpose of the Study:

• To demonstrate the feasibility of using deep learning to link histological patterns in whole slide images (WSIs) with drug sensitivities.
• To develop a model that predicts patient drug sensitivity directly from routine Haematoxylin & Eosin (H&E) stained breast cancer sections.

Main Methods:

• Utilized patient-wise drug sensitivities imputed from gene expression data of cancer cell lines.
• Trained a deep learning model to predict drug sensitivity profiles from WSIs.
• Employed histological patterns from H&E stained breast cancer sections as input features.

Main Results:

• Successfully linked histological patterns in WSIs to drug sensitivities.
• Demonstrated the ability to predict patient sensitivity to multiple approved and experimental drugs using routine WSIs.
• Identified specific cellular and histological patterns associated with drug sensitivity profiles.

Conclusions:

• Routine whole slide images can be leveraged for predicting cancer patient drug sensitivity profiles.
• Deep learning offers a novel, cost-effective approach to drug sensitivity prediction in oncology.
• This method holds potential for biomarker discovery and personalized cancer treatment strategies.