CHAMS-DTA: cross hybrid attention with multi-stage sampling for drug-target binding affinity prediction

Accurate prediction of drug-target binding affinity is crucial for computational drug discovery. In this work, we evaluate CHAMS-DTA on the standard kinase-centric benchmarks Davis and KIBA. We propose CHAMS-DTA, a three-stage coarse-to-fine framework that integrates cross-hybrid attention and adaptive gated fusion to progressively refine feature selection. At each stage, a cross-hybrid attention is used to capture both global interactions between the protein and drug molecules and local contextual information within the input sequences. This enables more accurate identification of key functional sites in both the protein and drug. Furthermore, an adaptive gated fusion mechanism is designed to selectively integrate cross-stage representations using learnable gates. This allows the model to dynamically determine the relative contribution of features from each stage. The proposed framework achieves not only high predictive accuracy but also provides initial insights into the model’s attention patterns, offering a degree of interpretability. Experimental results demonstrate that CHAMS-DTA achieves improved performance on specific evaluation metrics: the Concordance Index (CI) on the Davis dataset and the $$\\:{{\\text{r}}_{\\text{m}}}^{2}\\:$$ metric on the KIBA dataset.