Precise drug-target affinity (DTA) prediction plays a pivotal role in streamlining drug discovery processes and facilitating drug repurposing strategies. Although deep learning approaches have made significant progress in this task, existing models commonly rely on multilayer perceptrons (MLPs) with fixed and uniform activation functions to model nonlinearity. This limits the network’s adaptability to variations in input distributions and hinders its ability to fully capture the complex structural features and nonlinear interactions between drugs and targets. To overcome these limitations, we develop an enhanced framework incorporating an optimized Kolmogorov-Arnold Network variant (FastKAN) architecture. By incorporating FastKAN, the model enhances the capability of Graph Isomorphism Networks (GINs) in capturing the topological characteristics of molecular and target graphs, and accurately fits the nonlinear relationships in the prediction module. Additionally, we integrate semantic features extracted from two pretrained models to further enhance representation power. Our method has been verified, showing consistent improvements over previous models across several metrics.