Scholarlink

<p>Drug-target interactions (DTIs) are crucial biological activities that refer to the binding of a given drug molecule to a specific target for exerting targeted influence, which plays a crucial role in disease therapy. Traditional methods rely mainly on in vitro experiments to identify DTIs effectively. However, it is a time-consuming and laborious procedure. With the advent of artificial intelligence technology, computational methods, especially graph representation learning, have significantly accelerated the novel discovery of DTIs. Nevertheless, several challenges persist: (1) These methods focus only on the topological information of the higher-order neighboring nodes (i.e., drugs, targets or drug-protein pairs) in biological networks and neglect the interactive semantic information of drug-protein pairs (DPPs). (2) Existing methods generally treat non-interactive DPPs as negative samples through negative sampling during training, resulting in degraded performance due to potential false-negative samples. To address these challenges, we propose a novel semantic-enhanced representation learning method with a negative training (NT) strategy, i.e., SENT-DTI. To augment DPP features, we integrate inter-sequence molecule interactive information into the drug-protein pair network (DPPN) and enhances DPP knowledge representations via novel semantic-aware graph convolutional network (GCN). On this basis, to mitigate the impact of false-negative associations caused by negative sampling of DPPs, we are inspired by the advantage of negative training (NT) strategy on identification of false-negative samples and design a novel NT strategy that adaptively learns the probability distribution of known DPP features by incorporating a unified high-confidence false-negative association filtering (UHCF) mechanism into a negative loss objective function. Extended experiments are conducted on five public datasets, and the proposed SENT-DTI method demonstrates consistent improvements over the current state-of-the-art (SOTA) baselines. Specifically, SENT-DTI achieves average improvements of 2.6% in AUROC, 3.1% in AUPRC, 2.6% in Accuracy, 2.6% in Precision, and 2.1% in Specificity than the second-best baselines, particularly in addressing false-negative associations. Our method can be accessed at <a href="https://github.com/AlexCostra/SENT-DTI.">https://github.com/AlexCostra/SENT-DTI.</a></p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

SENT-DTI: Semantic-enhanced drug-target interaction prediction with negative training strategy

  • 期刊论文

  • 发表日期：

    2026年7月17日

查看全文
  • Weiyu Shi,
  • Zihao Liu,
  • Yuehui Zhang,
  • Hai Cui,
  • Yijia Zhang

摘要

Drug-target interactions (DTIs) are crucial biological activities that refer to the binding of a given drug molecule to a specific target for exerting targeted influence, which plays a crucial role in disease therapy. Traditional methods rely mainly on in vitro experiments to identify DTIs effectively. However, it is a time-consuming and laborious procedure. With the advent of artificial intelligence technology, computational methods, especially graph representation learning, have significantly accelerated the novel discovery of DTIs. Nevertheless, several challenges persist: (1) These methods focus only on the topological information of the higher-order neighboring nodes (i.e., drugs, targets or drug-protein pairs) in biological networks and neglect the interactive semantic information of drug-protein pairs (DPPs). (2) Existing methods generally treat non-interactive DPPs as negative samples through negative sampling during training, resulting in degraded performance due to potential false-negative samples. To address these challenges, we propose a novel semantic-enhanced representation learning method with a negative training (NT) strategy, i.e., SENT-DTI. To augment DPP features, we integrate inter-sequence molecule interactive information into the drug-protein pair network (DPPN) and enhances DPP knowledge representations via novel semantic-aware graph convolutional network (GCN). On this basis, to mitigate the impact of false-negative associations caused by negative sampling of DPPs, we are inspired by the advantage of negative training (NT) strategy on identification of false-negative samples and design a novel NT strategy that adaptively learns the probability distribution of known DPP features by incorporating a unified high-confidence false-negative association filtering (UHCF) mechanism into a negative loss objective function. Extended experiments are conducted on five public datasets, and the proposed SENT-DTI method demonstrates consistent improvements over the current state-of-the-art (SOTA) baselines. Specifically, SENT-DTI achieves average improvements of 2.6% in AUROC, 3.1% in AUPRC, 2.6% in Accuracy, 2.6% in Precision, and 2.1% in Specificity than the second-best baselines, particularly in addressing false-negative associations. Our method can be accessed at https://github.com/AlexCostra/SENT-DTI.