Pre-trained language model-based similarity relationship network approach for Drug-Target Interaction prediction

The rapid progress of drug and protein language models has enabled their application across a broad spectrum of biological prediction tasks, where they have consistently demonstrated remarkable predictive power. Recently, these models have been introduced into the Drug-Target Interaction (DTI) prediction task, yielding substantial improvements. Nevertheless, most existing language model-based DTI prediction methods rely solely on sequence representations of drugs and proteins, thereby emphasizing sequence-level features while overlooking the relational dependencies between drugs as well as those among proteins. Such dependencies provide complementary perspectives that are critical for enhancing predictive accuracy. To address this limitation, we propose a novel method called pre-trained language model-based similarity network approach for drug-target interaction prediction. This method combines pre-trained features from language models with drug and protein similarity networks to enrich relational representations. Subsequently, a cross-fusion attention mechanism integrates similarity features to suppress redundant or ineffective features while highlighting key ones. Finally, similarity-based features are combined with structural features to form a multi-view, comprehensive feature representation, thereby improving interaction prediction accuracy. Extensive experiments on four benchmark datasets-Human, C. elegans, BioSNAP, and DrugBank-demonstrate that our model consistently surpasses competitive baselines. Moreover, the results demonstrate its robustness in both imbalanced and cold-start scenarios, highlighting the benefits of incorporating relational information into language model-based DTI prediction.