Abstract: Drug-Target Interaction (DTI) prediction is a key task in drug discovery and repurposing. The challenge lies in how to integrate multi-source heterogeneous features to comprehensively characterize the complex relationships between drugs and targets. To address the shortcomings of traditional methods that rely on a single data source and model complex nonlinear relationships, a DTI prediction method based on Structure-Network collaborative features and grid-attention enhanced Kolmogorov-Arnold network (SNKDTI) was proposed. First, a feature extraction strategy based on structure and network collaboration was designed: for drug representation, molecular fingerprints were fused with graph embedding methods to quantify chemical structures; for target representation, traditional physicochemical encoding was combined with pre-trained models to extract sequence features. Meanwhile, heterogeneous networks such as drug-disease associations and protein-protein interactions were introduced, and network topological features were extracted by using the Random Walk with Restart (RWR) algorithm. And these features were compressed by using a Denoising AutoEncoder (DAE) to integrate structural and network information of drugs and targets. Second, a Heterogeneous Biological Information Network (HBIN) was constructed, with feature propagation carried out using a Graph Convolutional Network (GCN). Additionally, a Grid-Attention enhanced KAN (GA-KAN) was proposed, which introduced multiple learnable B-spline basis function grids and an attention mechanism to achieve adaptive combinations of multiple nonlinear mapping modules, thereby enhancing the model’s expressive power and input adaptability. Finally, a Gradient Boosting Decision Tree (GBDT) was used to build an end-to-end prediction framework. The proposed method and benchmark methods are compared on public datasets. The results show that SNKDTI achieves relative improvements of 0.81%, 1.36%, and 3.29% in Area Under the receiver operating characteristic Curve (AUC), Area Under the Precision-Recall curve (AUPR) and F1-score, respectively, over the best-performing benchmark methods for each corresponding metric. Therefore, the SNKDTI significantly enhances accuracy, robustness, and generalization ability, providing an efficient tool for new drug target screening.

Key words: Drug-Target Interaction (DTI) prediction, structure-network collaborative feature, KAN (Kolmogorov-Arnold Network), Graph Neural Network (GNN), attention mechanism

摘要： 药物-靶标相互作用(DTI)预测是药物发现与再利用的关键任务，它的难点是融合多源异构特征以全面表征药物与靶标间的复杂关联。针对传统方法依赖单一数据源和建模复杂非线性关系方面的不足，提出一种基于结构-网络协同特征与网格注意力增强的Kolmogorov-Arnold网络(KAN)的DTI预测方法SNKDTI(DTI prediction method based on Structure-Network collaborative features and grid-attention enhanced Kolmogorov-Arnold network)。首先，设计结构与网络协同的特征提取策略：在药物表示方面，融合分子指纹与图嵌入方法以量化化学结构；在靶标表示方面，结合传统物理化学编码与预训练模型提取序列特征。同时，引入药物-疾病关联、蛋白质相互作用等异构网络，基于随机游走重启(RWR)算法提取网络拓扑特征，并利用去噪自编码器(DAE)压缩特征，从而融合药物与靶标的结构与网络信息。其次，构建异质生物信息网络(HBIN)，使用图卷积网络(GCN)传播特征，并提出一种网格注意力增强的KAN(GA-KAN)，通过引入多组可学习的B样条基函数网格与注意力机制，实现多个非线性映射模块的自适应组合，从而增强模型的表达能力与输入适应性。最后，结合梯度提升决策树(GBDT)构建端到端预测框架。在公开数据集上的对比实验结果表明，SNKDTI的特征曲线下面积(AUC)、精确召回曲线下面积(AUPR)和F1-score相较于对应指标的最优基准方法分别提升了0.81%、1.36%和3.29%。验证了SNKDTI在准确性、鲁棒性和泛化能力方面均有显著提升，可为新药靶点筛选提供高效工具。

关键词: 药物-靶标相互作用预测, 结构-网络协同特征, KAN, 图神经网络, 注意力机制