ColdstartMHDTI: integrating biomolecular pretraining and attention-based heterogeneous graph learning for drug-target interaction prediction.

Motivation: Accurate drug-target interaction (DTI) prediction remains difficult for underexplored drugs and targets, especially when available interaction evidence is sparse. Existing approaches often focus either on pairwise molecular representations or on heterogeneous biomedical graph topology, making it difficult to effectively integrate structure-derived representations with multi-relational contextual evidence. Result: We propose ColdstartMHDTI, a two-stage framework for heterogeneous-graph-based DTI prediction that integrates sequence-derived structural representations with local and global relational information. Specifically, drug SMILES and target sequences are encoded by pretrained transformer models, while one-hop heterogeneous relations are captured through self-supervised DistMult embeddings. These representations are then fused through a meta-path-guided module that models ordered meta-path instances and aggregates them with cross-attention for interaction scoring. Across two benchmark datasets, ColdstartMHDTI shows consistent improvements under warm-start and entity-disjoint settings, with particularly strong performance for underexplored drugs and targets. It also remains robust under more imbalanced evaluation protocols with 1:5 and 1:10 positive-to-negative ratios. In addition to standard classification performance, ColdstartMHDTI supports candidate prioritization for downstream screening and evidence-guided hypothesis generation. Case studies on ESR1, EGFR, and Parkinson's disease further demonstrate its practical utility, with the Parkinson's disease analysis additionally highlighting strong per-drug target ranking performance.