Accurate drug–target affinity (DTA) prediction is critical for drug discovery and repurposing. However, existing models often struggle with generalizing to unseen drug–target pairs, lack interpretability, and fail to integrate heterogeneous biological features effectively. To overcome these challenges, we introduce KANPM-DTA, a deep learning framework designed to capture richer biochemical interactions and improve prediction reliability. Specifically, an ESM-guided protein graph construction strategy incorporates evolutionary and structural information to overcome underexplored protein representations. A gated fusion mechanism was employed to integrate drug–protein graph features, while linear attention captures cross-modal dependencies that enhance discriminative power. For the final affinity prediction, a Kolmogorov–Arnold network was used, offering a stronger nonlinear approximation and improved interpretability. Comprehensive experiments on benchmark datasets demonstrate that KANPM-DTA significantly outperforms state-of-the-art methods. On the Davis, KIBA, Metz, and BindingDB datasets, we achieved significant performance improvements under warm setting, with MSE reductions of 6.42%, 4.86%, 4.44%, and 5.46%, CI increases of 0.45%, 0.34%, 0.48%, and 0.80%, and $r_{m}^{2}$ gains of 1.85%, 0.90%, 0.84%, and 1.05%, respectively. Moreover, a case study on the epidermal growth factor receptor further highlights the effectiveness of KANPM-DTA in predicting DTAs for unknown drug–target pairs, emphasizing its potential for real-world applications in drug discovery. However, wet-lab validation is required to assess the applicability of the results.