Joint inference based on pre-trained language model (LM) and knowledge graph (KG) has not achieved better results in the biomedical domain due to its diverse terminology representation, semantic ambiguity and the presence of large amount of noise in the knowledge graph. This paper proposes an interpretable inference method DF-GNN for biomedical field, which unifies the entity representation of text and knowledge graph, denoises the subgraph constructed by a large biomedical knowledge base, and further improves the information interaction mode of text and subgraph entities by increasing the direct interaction between corresponding text and subgraph nodes, so that the information of the two modes can be deeply integrated. At the same time, the path information of the knowledge graph is used to provide interpretability for the model reasoning process. The test results on the public dataset MedQA-USMLE and MedMCQA show that DF-GNN can more reliably leverage structured knowledge for reasoning and provide explanatory properties than existing biomedical domain joint inference models.