Drug‑target affinity (DTA) prediction plays a pivotal role in computer‑aided drug discovery by prioritizing compound‑protein pairs before experimental validation. The main challenge lies in mining multi‑level molecular information from drug structures while accurately modeling the binding‑site environment of target proteins.
HierDTA addresses these issues with a hierarchical structure‑aware design. Each drug is represented as an atom‑motif heterogeneous graph, and an augmentation module propagates functional‑motif semantics to the corresponding atoms. For proteins, especially kinase‑domain structures, residues are converted into interaction graphs. Cross‑attention integrates atom and residue representations to capture pair‑specific relationships, while molecular fingerprints and pretrained ESM‑2 embeddings supply complementary global context.
Compared with state‑of‑the‑art methods, HierDTA achieves superior performance on the Davis dataset across all metrics and attains the highest $r_m^2$ of 0.791 on KIBA. In cold‑start settings, particularly the unseen‑protein scenario on KIBA, the mean squared error is reduced by 21.7%. Ablation studies and interpretability analyses confirm the contributions of the hierarchical augmentation and interaction components. Case studies further demonstrate early enrichment of bioactive candidate compounds.
These findings indicate that HierDTA is an effective framework for affinity prediction and target‑specific virtual screening.
Review