Abstract
The accurate prediction of drug-target interactions is essential for drug discovery and development. However, current models often struggle with two challenges. First, they fail to model the directional flow and positional sensitivity of protein sequences, which are critical for identifying functional interaction regions. Second, they lack mechanisms to integrate multi-scale information from both local binding sites and broader structural context. To overcome these limitations, we propose CAMF-DTI, a novel framework that incorporates coordinate attention, multi-scale feature fusion, and cross-attention to enhance both the representation and interaction learning of drug and protein features. Drug molecules are represented as molecular graphs and encoded using graph convolutional networks, while protein sequences are processed with coordinate attention to preserve directional and spatial information. Multi-scale fusion modules are applied to both encoders to capture local and global features, and a cross-attention module integrates the representations to enable dynamic drug-target interaction modeling. We evaluate CAMF-DTI on four benchmark datasets: BindingDB, BioSNAP, C.elegans, and Human. Experimental results show that CAMF-DTI consistently outperforms seven state-of-the-art baselines in terms of AUROC, AUPRC, Accuracy, F1-score, and MCC. Ablation studies further confirm the effectiveness of each module, and visualization results demonstrate the model's potential interpretability.