CF-DTI: coarse-to-fine feature extraction for enhanced drug-target interaction prediction

Accurate prediction of drug-target interactions (DTIs) is crucial for improving the efficiency and success rate of drug development. Despite recent advancements, existing methods often fail to leverage interaction features at multiple granular levels, resulting in suboptimal data utilization and limited predictive performance. To address these challenges, we propose CF-DTI, a coarse-to-fine drug-target interaction model that integrates both coarse-grained and fine-grained features to enhance predictive accuracy. To extract fine-grained features, an information filtering module incorporating focal and cross-attention mechanisms is designed to dynamically identify key protein and drug sites, reducing non-contributory noise and improving data quality. Then, a multi-granularity learning strategy is employed to effectively capture interactions at different levels, followed by an adaptive fusion learning approach to combine these features and maximize predictive performance. Experimental results on four benchmark datasets demonstrate that CF-DTI consistently outperforms existing baselines across multiple evaluation metrics. On large-scale datasets such as BindingDB and BioSNAP, CF-DTI achieves an average improvement of around 2% over the second-best models. In the more challenging unseen-pair setting, it attains significantly higher AUROC and AUPRC scores, with a notable 6% AUROC gain on BioSNAP. Visualization examples further demonstrate CF-DTI's ability to reduce prediction ambiguity and capture complex interaction patterns. Additionally, three case studies provide interpretability into molecular interactions, showcasing CF-DTI's potential to support more effective drug discovery.