Cancer is a highly heterogeneous disease, with different patients’ cancers potentially having varying genetic mutations, phenotypic characteristics, and molecular mechanisms. Therefore, accurately predicting drug responses is crucial in precision personalized medicine. However, existing research usually focuses solely on the basic information of the drugs themselves, potentially overlooking the impact of interactions between genes and gene pathway-specific combinatorial implications on biological processes or drug responses. In this work, we propose a multi-module hybrid neural framework for drug response prediction (DRP) that learns global and local feature information, called Graph-Pathway-Transformer Drug Response Prediction (GPT-DRP). In GPT-DRP, drugs are represented by molecular graphs using two types of graph neural networks to capture drug structural information, while cell lines are described by gene pathway activity scores employing a fully connected network and a convolutional neural network to capture cell line features. In addition, Transformer is used for the extracted drug and cell line representations to integrate the features learned from the drugs and the cell lines. Experimental results on the CCLE/GDSC datasets show that GPT-DRP outperforms the state-of-the-art models.