HierHGT-DTI: a multiscale relation-aware heterogeneous graph transformer for cold-start drug-target interaction prediction

Cold-start drug–target interaction (DTI) prediction requires transferring structural and sequence evidence to drugs or proteins with no interaction labels in the training set. This challenge is especially relevant for newly implicated or poorly characterized protein targets that lack established ligands. Existing models encode useful local features, but few coordinate fine-, intermediate- and global-scale representations of both interaction partners within one typed graph. HierHGT-DTI represents each candidate pair as a bilateral multiscale graph containing atom, substructure, drug, residue, sequence-derived residue-community and protein nodes. A heterogeneous graph transformer coordinates local, adjacent-scale, direct fine-to-global and drug–protein relations within this graph. Across five seeds, we compared HierHGT-DTI with six matched baselines on three DTI benchmarks: DrugBank, BioSNAP and BindingDB. HierHGT-DTI ranked first in all six DrugBank and BioSNAP random/cold-start settings. The largest margins occurred under cold-protein evaluation, where AUROC improved by 8.8 percentage points on DrugBank and 5.8 points on BioSNAP over the strongest baseline. On BindingDB, HierHGT-DTI achieved the highest cold-protein AUPR, whereas the strongest baseline, HiGraphDTI, led the random and cold-drug evaluations. HierHGT-DTI provides an effective multiscale relation-aware framework for cold-start DTI prediction. Its consistent gains on DrugBank and BioSNAP, together with the highest cold-protein AUPR on BindingDB, support its utility for prioritizing candidate interactions involving previously unseen protein targets. Source code, processed splits and configurations are publicly available.