HiF-DTA: Hierarchical Feature Learning Network for Drug-Target Affinity Prediction
Minghui Li, Yuanhang Wang, Peijin Guo, Wei Wan, Shengshan Hu, Shengqing Hu
TL;DR
The paper tackles sequence-based prediction of drug–target affinity (DTA) by addressing limitations in capturing local binding-site cues and multi-scale interactions. It introduces HiF-DTA, a dual-pathway encoder that retrieves global sequence semantics (BiLSTM for drugs and Mamba for proteins) and local topology (PNA-based MPNN), while modeling drugs at atomic, substructural, and molecular levels and fusing them via multi-scale bilinear attention. Evaluations on the Davis, Metz, and KIBA benchmarks show state-of-the-art performance, with ablations confirming the benefits of global–local integration and multi-scale fusion, including a log-scale transformation for Davis’ $K_d$ values using $pK_d=-\log_{10}(K_d/10^{9})$. The results demonstrate that a sequence-based, multi-scale, bilinear-attention framework can match or exceed structure-based approaches while improving efficiency for large-scale virtual screening. This approach has potential to accelerate early drug discovery by providing accurate DTA predictions without requiring costly 3D structural data.
Abstract
Accurate prediction of Drug-Target Affinity (DTA) is crucial for reducing experimental costs and accelerating early screening in computational drug discovery. While sequence-based deep learning methods avoid reliance on costly 3D structures, they still overlook simultaneous modeling of global sequence semantic features and local topological structural features within drugs and proteins, and represent drugs as flat sequences without atomic-level, substructural-level, and molecular-level multi-scale features. We propose HiF-DTA, a hierarchical network that adopts a dual-pathway strategy to extract both global sequence semantic and local topological features from drug and protein sequences, and models drugs multi-scale to learn atomic, substructural, and molecular representations fused via a multi-scale bilinear attention module. Experiments on Davis, KIBA, and Metz datasets show HiF-DTA outperforms state-of-the-art baselines, with ablations confirming the importance of global-local extraction and multi-scale fusion.