SHIFT-DRP: Dynamic Multi-Scale Active Learning for Drug Response Prediction.

Area of Science:

• Computational biology
• Pharmacogenomics
• Machine learning in drug discovery

Background:

• Deep learning models for cancer drug response prediction face challenges with novel drug-cell line combinations due to limited chemical space coverage.
• Comprehensive experimental screening is impractical, and uniform sampling is inefficient for drug-cell line pair selection.

Purpose of the Study:

• To develop SHIFT-DRP, an active learning framework for intelligent selection of drug-cell line pairs for experimental validation.
• To maximize model improvement in drug response prediction under resource constraints.

Main Methods:

• Employs a dynamic sampling strategy transitioning from diversity exploration to uncertainty refinement.
• Utilizes a pretrained model for molecular representation and a cross-attention mechanism for drug-cell line interactions.

Main Results:

• SHIFT-DRP outperforms existing active learning methods in prediction performance.
• Achieved a 24% reduction in experimental resources compared to uniform sampling.
• Demonstrated ability to identify structurally similar compounds with divergent responses.

Conclusions:

• SHIFT-DRP provides an efficient solution for guided experimental screening and data collection in drug response prediction.
• Offers significant implications for advancing precision medicine development.