HAFI-DTA: an interpretable multimodal framework for drug-target affinity prediction based on hierarchical attention guidance and feature iterative interaction

Accurate prediction of drug-target affinity (DTA) is a core technical step in high-throughput drug screening and targeted drug discovery, which directly affects the efficiency of candidate compound mining and the control of research and development costs. Traditional experimental methods are difficult to meet the requirements of high-throughput screening, while existing deep learning-based models generally suffer from bottlenecks such as shallow fusion of multi-source heterogeneous features, insufficient mining of interactive features, and inadequate generalization robustness. To address these issues, we propose HAFI-DTA (Hierarchical Attention-Guided and Feature Iterative Interaction for Multimodal Drug-Target Affinity Prediction). Specifically, our framework employs a hierarchical attention mechanism to guide the fusion of multi-source heterogeneous features, and combines a two-layer GRU with iterative interaction to capture the deep correlations among features. Experimental results on the KIBA and Davis datasets demonstrate that the proposed HAFI-DTA model achieves overall performance comparable to the current state-of-the-art, while delivering stable and modest improvements in both CI and MSE metrics. Moreover, under the most challenging drug–target double cold-start scenario, our model attains the best performance, exhibiting excellent generalization robustness. In addition, interpretability analysis shows that the model can automatically focus on the key residues of the real binding pockets of proteins, providing an important basis for subsequent drug design and binding mode research.