Accurate prediction of cancer drug responses is crucial for personalized therapy. Single-cell RNA sequencing (scRNA-seq) captures cellular heterogeneity and rare resistant populations, offering valuable insights into treatment responses. However, the distinct distributions of bulk RNA-seq and scRNA-seq data hinder the transfer of drug response knowledge from large-scale cell line datasets. To address this, single-cell Pathway Drug Sensitivity (scPDS) model is developed, a Transformer-based deep learning method that predicts drug sensitivities from scRNA-seq data through pathway activation transformation. By integrating bulk RNA-seq data from extensive cell line datasets, scPDS improves accuracy and computational efficiency in scRNA-seq analysis. It is demonstrated that scPDS outperforms state-of-the-art methods in both time and memory consumption. When applied to breast cancer cells treated with bortezomib, scPDS showed that resistance increases initially but diminishes with prolonged exposure. The method also identifies drug-sensitive populations in bortezomib-resistant cells and predicts the efficacy of combination therapies, including docetaxel, gemcitabine, and irinotecan. Furthermore, scPDS successfully distinguishes between sensitive and resistant patients, predicting significantly different survival outcomes. In summary, scPDS offers a robust tool for predicting cellular responses, providing insights to optimize cancer treatment strategies.