ColdDTA: Utilizing data augmentation and attention-based feature fusion for drug-target binding affinity prediction.

Area of Science:

• Computational chemistry
• Bioinformatics
• Drug discovery

Background:

• Accurate drug-target affinity (DTA) prediction is vital for efficient drug discovery.
• Deep learning models excel on standard datasets but struggle with real-world cold-start problems where novel drugs or targets appear.
• Improving the generalization ability of DTA prediction models is a significant challenge.

Purpose of the Study:

• To develop a novel deep learning framework, ColdDTA, to enhance the generalization performance of drug-target affinity prediction.
• To address the limitations of current methods in cold-start experimental setups.
• To provide interpretable insights into DTA prediction.

Main Methods:

• ColdDTA employs data augmentation by generating new drug-target pairs through drug subgraph removal.
• An attention-based feature fusion module is utilized to better capture complex drug-target interactions.
• The model was evaluated using cold-start experiments on three benchmark datasets (Davis, KIBA, BindingDB).

Main Results:

• ColdDTA demonstrated superior performance over five state-of-the-art baseline methods on the Davis and KIBA datasets, measured by consistency index (CI) and mean square error (MSE).
• On the BindingDB dataset, ColdDTA achieved better performance in the classification task, indicated by the area under the receiver operating characteristic curve (ROC-AUC).
• Model weight visualization provided interpretable insights into the prediction process.

Conclusions:

• ColdDTA effectively improves the generalization ability for drug-target affinity prediction, particularly in realistic cold-start scenarios.
• The proposed data augmentation and attention-based fusion strategies enhance model robustness and predictive accuracy.
• The publicly available code facilitates further research and application in drug discovery.