Data imbalance in drug response prediction: multi-objective optimization approach in deep learning setting

Area of Science:

• Computational biology
• Machine learning in oncology
• Drug discovery and development

Background:

• Drug response prediction (DRP) links patient genetics to drug effectiveness, crucial for personalized cancer therapy.
• Anti-cancer DRP is complex due to broad pathogenic mechanisms and limited data depth compared to other AI domains.
• Existing DRP models struggle with data imbalance, hindering generalizability and clinical application.

Purpose of the Study:

• To develop strategies addressing data imbalance in DRP datasets.
• To enhance the generalizability and performance of deep learning-based DRP models.
• To reframe DRP as a multi-objective optimization problem across multiple drugs.

Main Methods:

• Implemented a Multi-Objective Optimization Regularized by Loss Entropy (MOORLE) loss function.
• Integrated the MOORLE loss function into a deep learning model architecture.
• Evaluated the approach on anti-cancer drug screening datasets.

Main Results:

• Demonstrated improved performance of DRP models by addressing data imbalance.
• The multi-objective optimization strategy enhanced model generalizability.
• The proposed method shows utility for advancing drug discovery and personalized medicine.

Conclusions:

• The MOORLE approach effectively tackles data imbalance in DRP.
• This strategy enhances deep learning model performance for anti-cancer drug response prediction.
• The work offers a pathway for improved drug discovery and healthcare outcomes.