In view of the flaw that tumor heterogeneity causes patients to respond differently to the same drug therapy, thus leading to tumor relapse or metastasis, a tumor drug-resistant cell recognition method GCDrug has been proposed based on graph structure and unsupervised domain adaptation. Firstly, similarity maps are to be constructed between source domain (batch transcriptome sequencing) and target domain (single-cell transcriptome sequencing) samples based on k-nearest neighbors, thus extracting features using graph convolutional networks. Subsequently, Deep CORAL loss is introduced to achieve cross-domain distribution adaptation. Experimental results on 10 single-cell drug-annotated datasets demonstrate that GCDrug outperforms current mainstream methods across three classification metrics. Notably, on the Etoposide dataset, GCDrug achieves an F1-score of 0.956, significantly outperforming existing methods. Ablation experiments further verify that the synergistic effect of the graph structure module and the domain adaptation module helps to improve the discriminative performance and generalization ability of the model. The experimental results indicate that the proposed method can accurately identify drug-resistant individual cells in tumors.