Attending on Multilevel Structure of Proteins enables Accurate Prediction of Cold-Start Drug-Target Interactions

TL;DR

ColdDTI tackles cold-start drug–target interactions by explicitly modeling multi-level protein structures and learning cross-level interactions with drugs. It uses ProtBERT-based representations for protein levels and ChemBERTa-2 for drug features, with a hierarchical attention mechanism and adaptive fusion to fuse multi-level representations. The approach demonstrates superior performance on four benchmark datasets across cold-start settings, suggesting stronger biological priors and generalization. This framework offers a scalable path to more reliable DTI predictions for novel drugs and proteins, aiding early-stage drug discovery.

Abstract

Cold-start drug-target interaction (DTI) prediction focuses on interaction between novel drugs and proteins. Previous methods typically learn transferable interaction patterns between structures of drug and proteins to tackle it. However, insight from proteomics suggest that protein have multi-level structures and they all influence the DTI. Existing works usually represent protein with only primary structures, limiting their ability to capture interactions involving higher-level structures. Inspired by this insight, we propose ColdDTI, a framework attending on protein multi-level structure for cold-start DTI prediction. We employ hierarchical attention mechanism to mine interaction between multi-level protein structures (from primary to quaternary) and drug structures at both local and global granularities. Then, we leverage mined interactions to fuse structure representations of different levels for final prediction. Our design captures biologically transferable priors, avoiding the risk of overfitting caused by excessive reliance on representation learning. Experiments on benchmark datasets demonstrate that ColdDTI consistently outperforms previous methods in cold-start settings.

Figures (4)

• Figure 1: A real example of the Human COX-1 Crystal Structure with drug FLC (A) Protein exhibits hierarchical structures from primary to quaternary levels. (B)Drug–protein interactions vary in granularity, with global interactions involving broader binding surfaces and local interactions reflecting weaker contacts.
• Figure 2: Overview of ColdDTI. (A) Overall architecture with hierarchical attention and fusion. (B) Inter-level feature fusion aggregates protein levels. (C) Intra-level feature fusion combines features within the same level. (D) Hierarchical attention captures drug–protein cross-level interactions. (E) Protein multi-level structure feature extraction with ProtBERT encoding.
• Figure 3: Ablation study on the DrugBank dataset under cold drug, cold protein, and cold pair settings.
• Figure 4: Case study of -COO- group of DB00945 interacts with -OH group of Ser-530 in P23219.