ReduMixDTI: Prediction of Drug-Target Interaction with Feature Redundancy Reduction and Interpretable Attention...

Area of Science:

• Computational chemistry
• Bioinformatics
• Drug discovery

Background:

• Accurate drug-target interaction (DTI) prediction is crucial for efficient drug discovery.
• Current deep learning models for DTI prediction suffer from feature redundancy and noise due to holistic representations, and oversimplified interaction modeling.
• This limits predictive accuracy and the interpretability of binding mechanisms.

Purpose of the Study:

• To develop an end-to-end deep learning model, ReduMixDTI, that addresses feature redundancy and explicitly models complex local interactions for improved DTI prediction.
• To enhance the interpretability of DTI prediction models.

Main Methods:

• Drug and target features are encoded using graph neural networks (GNNs) and convolutional neural networks (CNNs), respectively.
• Feature refinement is performed from channel and spatial perspectives.
• An attention mechanism is employed to model pairwise interactions between drug and target substructures.

Main Results:

• ReduMixDTI significantly outperforms seven state-of-the-art methods on three benchmark datasets and external test sets.
• Ablation studies and visualization of protein attention weights confirm the model's robustness and interpretability.
• The model effectively reduces feature redundancy and captures complex binding interactions.

Conclusions:

• ReduMixDTI provides a robust and interpretable approach for DTI prediction by addressing limitations in existing deep learning models.
• The model's ability to reduce feature redundancy and capture local interactions advances the field of DTI prediction.
• This contributes to more efficient and effective drug discovery and development pipelines.