Drug-Target Interaction (DTI) prediction faces two problems. Question 1: Existing methods either only consider the interaction of the drug’s multi- modality fusion features and the protein’s features. Or they only consider the interaction form between one modality of the drug and the target, i.e., they do not consider the interaction in the form of protein-SMILES and the protein-molecule simultaneously. Question 2: how can we simulate the 3D pocket data using sequence data? For the question 1, we propose a DTI prediction framework based on MoE that includes two experts. One expert is related to the interaction between the protein and the drug for DTI prediction, in which the drug is represented by the fusion of SMILES and the molecular structure. The other expert considers two types of interaction forms, namely the protein-SMILES interaction form and the protein- molecule interaction form, for DTI prediction. For question 2, We propose utilizing the heterogeneous graph network of drug-target to dynamically learn the implicit pocket information. It employs a node feature aggregation method based on edge weight-based cross-attention. Additionally, we propose a lightweight MoE based on parameter sharing and decoupled feature representation and feature fusion. Experimental results demonstrate that our method achieves the best performance.