PSDTA: An Approach to Drug-Target Binding Affinity Prediction by Integrating Physicochemical and Structural...

Area of Science:

• Computational chemistry
• Structural biology
• Drug discovery

Background:

• Accurate drug-target binding affinity (DTA) prediction is crucial for drug repurposing.
• Current deep learning methods struggle with integrating physicochemical properties and structural data, leading to feature redundancy and poor generalization.

Purpose of the Study:

• To propose PSDTA, a novel deep learning method for DTA prediction.
• To enhance DTA prediction by integrating physicochemical properties and explicit amino acid structural information.
• To reduce feature redundancy and improve model generalization.

Main Methods:

• Developed PSDTA, incorporating physicochemical properties into initial feature representations.
• Explicitly integrated amino acid structural information, avoiding coordinate-based information leakage.
• Employed two complementary channels to identify binding-relevant residues at residue and group levels.

Main Results:

• PSDTA achieved superior performance on three benchmark datasets (PDBBind v2016, PDBBind v2020, Davis) compared to state-of-the-art methods.
• Interpretability analyses confirmed consistent and complementary binding-related regions identified by the two channels.
• The method demonstrated enhanced generalization capability by avoiding direct coordinate usage.

Conclusions:

• PSDTA offers a robust and effective approach for DTA prediction.
• The integration of physicochemical and structural information significantly improves prediction accuracy and model generalization.
• The dual-channel residue identification effectively reduces feature redundancy and enhances biological relevance.