Predicting anticancer drug sensitivity on distributed data sources using federated deep learning.

Area of Science:

• Bioinformatics
• Computational Biology
• Precision Medicine

Background:

• Drug sensitivity prediction is vital for personalized cancer treatment.
• Data privacy regulations hinder collaborative prediction studies.
• Centralized models face challenges due to data silos.

Purpose of the Study:

• To develop a privacy-preserving federated drug sensitivity prediction model.
• To enable collaboration across medical institutions without compromising data protection.
• To enhance the generalization of drug sensitivity prediction models.

Main Methods:

• A novel horizontal federated deep learning framework with focal loss (HFDL-fl) was proposed.
• Cell lines were classified into three categories using the waterfall method.
• The model was applied to both homogeneous (HFDL-Within) and heterogeneous (HFDL-Cross) data.

Main Results:

• HFDL-fl demonstrated superior performance compared to private, local models.
• The focal loss function effectively improved classification accuracy for sensitive and resistant cell lines.
• HFDL-fl showed robustness against data heterogeneity across institutions.

Conclusions:

• HFDL-fl offers a privacy-preserving solution for collaborative drug sensitivity prediction.
• This approach overcomes data sharing barriers in biomedical research.
• It facilitates advancements in cancer precision medicine and other sensitive data research.