Predicting drug-target interactions (DTIs) is a fundamental task in computational drug discovery, where reliable generalization to novel compounds and protein targets is essential for practical virtual screening and drug repurposing. However, most deep learning models are evaluated using random or single-split settings that fail to reflect the ligand and protein novelty conditions encountered in real-world discovery pipelines. This work proposes a dual-encoder fusion framework for robust and interpretable DTI prediction. The model combines pretrained ESM protein embeddings with two complementary ligand representations: a ChemBERTa-based molecular language encoder and a graph-based structural encoder. Predictions from both branches are integrated through decision-level fusion. To rigorously evaluate generalization, we adopt a novelty-aware protocol that isolates ligand novelty, protein novelty, and their joint occurrence. A ligand-centered, gradientbased interpretability analysis is also employed to examine how molecular substructures contribute to binding predictions. Experiments on the largescale BindingDB dataset show that ligand novelty induces minimal performance degradation, with strong results under warm and cold-drug settings (F1 = 0.84-0.87; AUC = 0.87-0.89). In contrast, protein novelty emerges as the dominant generalization bottleneck, producing the largest performance drops under cold-protein and double-cold conditions. Across all scenarios, the fusion model exhibits the most stable behavior, achieving an AUC of 0.90 in the warm setting. External validation on the Davis and KIBA datasets further demonstrates consistent AUC values (0.60-0.64) and high recall despite substantial biochemical domain shift. These results highlight the importance of novelty-aware evaluation and show that decision-level fusion provides a practical pathway toward reliable and interpretable DTI prediction.
Abou-Abbas et al. (Wed,) studied this question.