Scholarlink

<p>Accurate prediction of drug-target binding affinity (DTA) is crucial for drug discovery and drug repositioning. Although graph neural networks (GNNs) have achieved significant progress in DTA prediction, existing GNN-based methods still face two major challenges: (1) how to effectively integrate the complex relationships between multimodal data, such as sequences and graphs, and (2) how to more comprehensively and accurately represent the features of drugs and proteins. This paper proposes a multimodal information fusion model, MDEBert-DTA, which combines a dynamic graph aggregation network with the ProtBert pre-trained model for DTA prediction. Specifically, MDEBert-DTA employs both molecular graphs and molecular fingerprints to represent drugs, and balances similar functional groups or residues across different drugs through a multimodal contrastive learning method. It represents each drug molecule as a graph structure of interactions between nodes (atoms) and edges (bonds). employs a dynamic attention scoring mechanism to identify the most critical atoms and bonds for DTA prediction, and integrates a graph aggregation network (GEN) to hierarchically aggregate neighbor information adaptively, forming multi-level feature representations from atoms to molecular scaffolds. Additionally, the ProtBert pre-trained model is utilized to comprehensively capture the sequential dependencies among amino acids, thereby more accurately representing protein features. Experimental results demonstrate that compared to state-of-the-art models, MDEBert-DTA reduces the mean squared error (MSE) by 1.7%, improves the concordance index (CI) by 0.3%, and enhances performance by 3% on the Davis dataset; on the KIBA dataset, the model achieves a 2.1% reduction in MSE, a 1.4% improvement in CI, and a 1.9% enhancement in performance.</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

MDEBert-DTA: drug-target affinity prediction using multimodal contrastive learning and dynamic graph aggregation

  • 期刊论文

  • 发表日期：

    2026年4月8日

查看全文
  • Huaizhou Yang,
  • Xiaohan Tong

摘要

Accurate prediction of drug-target binding affinity (DTA) is crucial for drug discovery and drug repositioning. Although graph neural networks (GNNs) have achieved significant progress in DTA prediction, existing GNN-based methods still face two major challenges: (1) how to effectively integrate the complex relationships between multimodal data, such as sequences and graphs, and (2) how to more comprehensively and accurately represent the features of drugs and proteins. This paper proposes a multimodal information fusion model, MDEBert-DTA, which combines a dynamic graph aggregation network with the ProtBert pre-trained model for DTA prediction. Specifically, MDEBert-DTA employs both molecular graphs and molecular fingerprints to represent drugs, and balances similar functional groups or residues across different drugs through a multimodal contrastive learning method. It represents each drug molecule as a graph structure of interactions between nodes (atoms) and edges (bonds). employs a dynamic attention scoring mechanism to identify the most critical atoms and bonds for DTA prediction, and integrates a graph aggregation network (GEN) to hierarchically aggregate neighbor information adaptively, forming multi-level feature representations from atoms to molecular scaffolds. Additionally, the ProtBert pre-trained model is utilized to comprehensively capture the sequential dependencies among amino acids, thereby more accurately representing protein features. Experimental results demonstrate that compared to state-of-the-art models, MDEBert-DTA reduces the mean squared error (MSE) by 1.7%, improves the concordance index (CI) by 0.3%, and enhances performance by 3% on the Davis dataset; on the KIBA dataset, the model achieves a 2.1% reduction in MSE, a 1.4% improvement in CI, and a 1.9% enhancement in performance.