Synthetic Assessment of Transfer Learning Method From Cell Lines to Single-Cell on Drug Response Prediction

Drug response prediction is of critical importance in precision medicine and novel drug development, yet it remains highly challenging due to the complexity, high cost, and low success rate of the process. With the rapid advancement of single-cell sequencing technologies, researchers are now able to gain deeper insights into intratumoral clonal heterogeneity and drug resistance mechanisms, thereby the necessity of drug response prediction is underscored at the single-cell level. However, despite the growing availability of single-cell expression data, corresponding drug sensitivity annotations remain extremely scarce, significantly hindering the development and generalization of supervised learning models. To bridge this data gap, transfer learning has emerged as a key strategy, enabling the migration of drug response knowledge learned from bulk cell line data to single-cell contexts. However, most existing studies focus on model development without conducting systematic, comparative evaluations. To address this gap, this study presents the first comprehensive assessment of representative single-cell drug response prediction models based on transfer learning. We perform a multidimensional analysis of these methods-including transfer mechanisms, feature alignment strategies, and predictive performance-using publicly available datasets for empirical benchmarking. It provides valuable methodological guidance for the future selection and design of predictive models, and establishes a foundation for advancing toward clinically actionable single-cell pharmacogenomics.