This paper proposes a novel approach for predicting drug-protein interactions (DPIs) by integrating global and local features in a drug-protein bipartite graph. The model employs pre-trained models to obtain initial features of drugs and proteins, then uses MinHash and HyperLogLog algorithms to estimate similarity and set cardinality between drug and protein subgraphs, serving as local features. An energy-constrained diffusion mechanism is integrated into the transformer architecture to capture interdependencies between nodes and extract global features. The local and global features are then fused, and multilayer perceptrons are used to predict the likelihood of potential DPIs. The model is validated on multiple datasets, demonstrating its accuracy and reliability in predicting unknown DPIs. The results show that the proposed model outperforms existing methods in terms of AUC and AUPR, indicating its effectiveness in identifying potential DPIs not present in existing databases. The model's contributions include the use of large-model pre-training, a global diffusion mechanism, and local subgraph extraction technologies to efficiently and accurately infer potential DPIs. The model's performance is evaluated across various public datasets, and molecular docking experiments confirm its ability to identify unvalidated DPIs. The study highlights the importance of node representation in DPI prediction and suggests future research directions to enhance model interpretability and incorporate multimodal information.This paper proposes a novel approach for predicting drug-protein interactions (DPIs) by integrating global and local features in a drug-protein bipartite graph. The model employs pre-trained models to obtain initial features of drugs and proteins, then uses MinHash and HyperLogLog algorithms to estimate similarity and set cardinality between drug and protein subgraphs, serving as local features. An energy-constrained diffusion mechanism is integrated into the transformer architecture to capture interdependencies between nodes and extract global features. The local and global features are then fused, and multilayer perceptrons are used to predict the likelihood of potential DPIs. The model is validated on multiple datasets, demonstrating its accuracy and reliability in predicting unknown DPIs. The results show that the proposed model outperforms existing methods in terms of AUC and AUPR, indicating its effectiveness in identifying potential DPIs not present in existing databases. The model's contributions include the use of large-model pre-training, a global diffusion mechanism, and local subgraph extraction technologies to efficiently and accurately infer potential DPIs. The model's performance is evaluated across various public datasets, and molecular docking experiments confirm its ability to identify unvalidated DPIs. The study highlights the importance of node representation in DPI prediction and suggests future research directions to enhance model interpretability and incorporate multimodal information.