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

Deep learning models show promise for drug response prediction in personalized cancer treatment, but exhibit limited prediction capability for novel drug-cell line combinations due to insufficient coverage of the chemical spaces in training data. The vast combinatorial space of drug-cell line pairs makes comprehensive experimental screening impractical, while uniform sampling is inefficient, as a large proportion of randomly selected experiments contain redundant information. To address this challenge, we present SHIFT-DRP, an active learning framework that intelligently selects drug-cell line pairs for experimental validation to maximize model improvement under limited resources. The core approach is a dynamic sampling strategy that transitions from diversity-focused exploration to uncertainty-driven refinement, ensuring both broad feature space coverage and targeted improvement where the model lacks adequate learning. The framework employs a pretrained model for molecular representation and a cross-attention mechanism to model drug-cell line interactions. Evaluation on four data sets demonstrates SHIFT-DRP's superiority over existing active learning methods, achieving better prediction performance while reducing the required experimental resources by 24% compared with uniform sampling. Ablation studies confirm the effectiveness of the dynamic sampling strategy, and case studies reveal SHIFT-DRP's ability to identify structurally similar compounds that exhibit divergent responses in the same cell line. SHIFT-DRP offers an efficient solution for guided experimental screening and data collection in drug response prediction with significant implications for precision medicine development.