scHDR: A Heterogeneous Network Transfer Learning Model forPredicting Single-Cell Drug Responses

Introduction: Single-cell RNA sequencing (scRNA-seq) generates expression data from individual cells, and drug response prediction based on these data can aid in drug therapy at the cell level. Existing methods for predicting single-cell drug responses primarily focus on gene expression, neglecting the complex interactions when drugs act on cells and inadequately integrating crossdomain information. This study proposes scHDR, which integrates multiple types of information based on heterogeneous networks and uses transfer learning to achieve cross-domain prediction of cell drug responses. Methods: By integrating drug, cell, and gene information from both bulk and single-cell levels into heterogeneous networks, and employing message passing and structure-preserving transfer learning, scHDR predicts single-cell drug responses while maintaining high performance in both domains, with labels in the target domain by default completely unknown during training. Results: Comparison experiments across six datasets demonstrate that scHDR outperforms other representative models at both the individual cell and cell cluster levels. Ablation and interpretability experiments confirm the critical role of the message passing and transfer learning components, while domain difference analysis and sensitivity experiments examine the effects of domain discrepancy and network size on model performance, respectively. Additionally, scHDR successfully screens drugs for gastric cancer cells, stratifies drug responses in breast cancer cells over time, and captures the overall response of patient cells, identifying corresponding drug response biomarkers and cell response biomarkers. Key chemical structures of drugs and important genes in cells are also calculated based on gradients. Discussion: This model effectively leverages the strengths of heterogeneous networks and transfer learning to improve the accuracy of single-cell drug response prediction. Its components are well coordinated, enabling cross-domain information transfer. Case study results align with existing evidence, demonstrating excellent performance across multiple tasks. Conclusion: scHDR provides a novel method for applying complex network modeling to single-cell drug research. Not only does it improve prediction accuracy, but it also offers valuable insights for drug research and precision therapy.