Abstract
Drug-target interaction (DTI) prediction is crucial in drug discovery. Recent advances in multiview learning have made it possible to automatically extract complex features from multiple perspectives. Multiview models, which integrate diverse biological data sources, have demonstrated improved prediction accuracy and robustness. However, current approaches still face major limitations: (1) reliance on single-scale sequence tokenizers that fail to capture biological information across different granularities and (2) shallow, single-layer integration of data views that overlook the hierarchical nature of biological systems. To tackle these challenges, we propose the concept of "bio-token" and design a multiscale biological tokenizer that captures biological features at varying resolutions. We also introduce a novel hierarchical multi-bio-view learning (HMBV) approach, implemented in an end-to-end DTI prediction network termed HMBVIP. The hierarchical multiview fusion enriches hidden representations with multidimensional biological context, thereby enhancing both prediction accuracy and biologically meaningful interpretability. The results on benchmark data sets demonstrate that HMBVIP consistently outperforms current state-of-the-art models.
