Understanding drug–target interactions (DTI) is crucial for drug discovery and repositioning. Although experimental methods are more reliable, they tend to be costly and time-consuming, highlighting the urgent need for efficient computational approaches. Existing deep learning models have advanced DTI prediction by leveraging molecular structures, protein sequences, and heterogeneous biological networks. However, challenges remain due to the computational overhead and noise introduced by large-scale knowledge graphs (KGs), as well as limited generalizability in cold-start scenarios involving novel drugs or targets. To address these issues, we propose KGsteeredDTI, a knowledge graph-guided framework that optimizes molecular representations through biologically informed guidance. KGsteeredDTI constructs drug-centric guidance graphs from biological entity information within KGs and progressively refines drug representations through multi-level information steering. Extensive experiments on multiple public DTI datasets demonstrate that KGsteeredDTI significantly outperforms existing mainstream methods, particularly exhibiting stronger robustness and generalization in cold-start scenarios.