Scholarlink

<p>Drug-target binding affinity (DTA) describes the strength of drug-target binding. Accurate prediction of DTA enables identify the potential drug candidates that effectively bind to target proteins, is a crucial step in drug discovery and development. Most existing deep learning methods employ Convolutional Neural Networks (CNNs) models to extract features from drug Simplified Molecular Input Line Entry System (SMILES) sequences while encoding protein sequences as one-hot vectors. The final prediction is generated through fully connected layers. However, these methods face two significant challenges. The first point is that the SMILES sequence, as a linear representation, cannot directly encode specific chemical properties of molecules. The second is single connection method cannot effectively predict which protein binding sites contribute most to the binding of a specific drug. To address these issues, we propose a novel method—CLAT-DTA. The method uses molecular fingerprint as drug representation to improve the encoding capability for drug molecules and a pre-trained protein model fine-tuned by Bidirectional Long Short-Term Memory(Bi-LSTM) to extract protein features. A lightweight collaborative cross-attention module is proposed to better aggregate important information and predict the binding affinity between drugs and proteins. Extensive experiments on Davis and Kinase inhibitor Bioactivity (KIBA) datasets demonstrate the effectiveness of the proposed model. On the Davis dataset, CLAT-DTA achieves a Concordance Index (CI) of 0.911, outperforming the best baseline model by 0.006, corresponding to a relative improvement of 0.66%. In terms of Mean Squared Error (MSE), CLAT-DTA attains a value of 0.163, which is 0.005 lower than the best baseline model, representing a relative reduction of 2.98% and demonstrating smaller prediction errors. On the KIBA dataset, CLAT-DTA achieves a CI of 0.901, surpassing the best baseline model by 0.003, with a relative improvement of 0.33%. For MSE, CLAT-DTA reaches 0.132, which is 0.004 lower than the best baseline model, corresponding to a relative reduction of 2.94%, further validating the model’s superiority in prediction accuracy. Our code is available at <a href="https://github.com/zhanggit001/CLAT-DTA/tree/master/CLAT-DTA">https://github.com/zhanggit001/CLAT-DTA/tree/master/CLAT-DTA</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

CLAT-DTA: a collaborative attention-based approach for drug-target binding affinity prediction

  • 期刊论文

  • 发表日期：

    2026年5月28日

查看全文
  • Yu Zhang,
  • Shengbo Fu,
  • Hai Cui,
  • Yuehui Zhang,
  • Meiyu Duan,
  • Yijia Zhang

摘要

Drug-target binding affinity (DTA) describes the strength of drug-target binding. Accurate prediction of DTA enables identify the potential drug candidates that effectively bind to target proteins, is a crucial step in drug discovery and development. Most existing deep learning methods employ Convolutional Neural Networks (CNNs) models to extract features from drug Simplified Molecular Input Line Entry System (SMILES) sequences while encoding protein sequences as one-hot vectors. The final prediction is generated through fully connected layers. However, these methods face two significant challenges. The first point is that the SMILES sequence, as a linear representation, cannot directly encode specific chemical properties of molecules. The second is single connection method cannot effectively predict which protein binding sites contribute most to the binding of a specific drug. To address these issues, we propose a novel method—CLAT-DTA. The method uses molecular fingerprint as drug representation to improve the encoding capability for drug molecules and a pre-trained protein model fine-tuned by Bidirectional Long Short-Term Memory(Bi-LSTM) to extract protein features. A lightweight collaborative cross-attention module is proposed to better aggregate important information and predict the binding affinity between drugs and proteins. Extensive experiments on Davis and Kinase inhibitor Bioactivity (KIBA) datasets demonstrate the effectiveness of the proposed model. On the Davis dataset, CLAT-DTA achieves a Concordance Index (CI) of 0.911, outperforming the best baseline model by 0.006, corresponding to a relative improvement of 0.66%. In terms of Mean Squared Error (MSE), CLAT-DTA attains a value of 0.163, which is 0.005 lower than the best baseline model, representing a relative reduction of 2.98% and demonstrating smaller prediction errors. On the KIBA dataset, CLAT-DTA achieves a CI of 0.901, surpassing the best baseline model by 0.003, with a relative improvement of 0.33%. For MSE, CLAT-DTA reaches 0.132, which is 0.004 lower than the best baseline model, corresponding to a relative reduction of 2.94%, further validating the model’s superiority in prediction accuracy. Our code is available at https://github.com/zhanggit001/CLAT-DTA/tree/master/CLAT-DTA.