Answers are grounded in this paper and cite their sources.
Drug-target interaction (DTI) prediction is important for drug discovery and repurposing, but many existing heterogeneous graph methods rely on manually designed meta-paths. We propose SGAN-DTI, a dual-channel subgraph attention framework that independently models drug-centered and target-centered heterogeneous subgraphs without explicitly constructing meta-paths. Learnable edge -type embeddings are introduced to capture relation-specific information, and a drug-target pair graph with graph convolution is used to model higher-order dependencies among candidate pairs. Warm-start experiments on Hetero-A, Hetero-B, and Zeng's dataset demonstrate competitive predictive performance against the evaluated baselines, while entity-level cold-start evaluations reveal substantial variation across datasets and held-out entity types. Ablation studies confirm the contributions of heterogeneous subgraphs, edge -type embeddings, and the drug-target pair graph. The model is relatively robust to the tested hidden dimensions and learning rates, whereas batch composition can affect prediction-score stability, ranking performance, and threshold-dependent predictions. The SGAN-DTI implementation and the available evaluation resources are available at https://github.com/DTI-sketch/SGAN-DTI.
No takes yet. Share an insight, caveat, or question.
Guo et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: