Drug-target interactions (DTIs) prediction plays a crucial role in drug development, impacting areas such as virtual screening, drug repurposing, and the identification of potential drug side effects. Although existing methods are effective, they still face limitations in handling the sparsity of current datasets and the complexity of heterogeneous networks. To address these challenges, we propose a unified framework for DTIs prediction based on a semantics-guided meta-path walk. Specifically, we first extract the semantic information of drugs and proteins based on their structural information. This semantic information is then employed to guide a meta-path-based random walk on the biological heterogeneous network for generating sequences. Finally, these sequences are used to compute embedding features by heterogeneous skip-gram, which are input into the downstream task to predict DTIs. SGMDTI achieves substantial performance improvement over other state-of-the-art methods for drug-target interaction prediction. Moreover, it excels in the cold-start scenario, which is often a challenging case in DTIs prediction. These results indicate the effectiveness of our approach in predicting DTIs.