Improving binding affinity prediction by emphasizing local features of drug and protein

Area of Science:

• Computational chemistry
• Bioinformatics
• Drug discovery

Background:

• Binding affinity prediction is crucial for drug discovery.
• Previous models often overlook local structural features of drugs and proteins, potentially limiting prediction accuracy.
• Macro-level features capture overall characteristics but miss fine-grained interactions.

Purpose of the Study:

• To develop a deep learning model for accurate binding affinity prediction by comprehensively extracting local features.
• To address the limitations of prior work that focused solely on macro-level features.
• To investigate the significance of local structural information in predicting drug-target interactions.

Main Methods:

• Proposed a novel deep learning model comprising Multi-Stream Convolutional Neural Network (CNN) and Multi-Stream Graph Convolutional Network (GCN).
• Multi-Stream CNN extracts local features from target protein sequences.
• Multi-Stream GCN captures local features from drug molecule subgraphs.
• Utilized multiple streams with varying layers to preserve micro-level characteristics.

Main Results:

• The proposed model demonstrated superior performance over baseline models on the Davis and KIBA datasets.
• Evaluation confirmed that incorporating local features significantly enhances binding affinity prediction accuracy.
• The model effectively captures micro-level characteristics from both protein and drug structures.

Conclusions:

• Local features are critically important for accurate binding affinity prediction in drug discovery.
• The proposed Multi-Stream CNN and GCN model offers a powerful approach for leveraging local structural information.
• This work advances computational methods for predicting drug-target interactions.