StackCPA: A stacking model for compound-protein binding affinity prediction based on pocket multi-scale features.

Area of Science:

• Computational chemistry
• Drug discovery
• Bioinformatics

Background:

• Accurate prediction of compound-protein binding affinity is vital for efficient drug discovery.
• Computational models offer a faster, safer, and more cost-effective alternative to traditional methods.
• Protein pockets are critical determinants of binding affinity, offering insights for drug design and repositioning.

Purpose of the Study:

• To develop an advanced ensemble learning model, StackCPA, for predicting compound-protein binding affinity.
• To integrate multi-scale features from protein pockets and compounds using transfer learning.
• To evaluate the efficacy of StackCPA against existing state-of-the-art deep learning models.

Main Methods:

• Proposed an ensemble learning model named StackCPA.
• Integrated multi-scale features (atomic, residue, subdomain levels) of protein pockets with compound features.
• Employed a transfer learning strategy for feature integration.
• Evaluated performance on three benchmark binding affinity datasets.

Main Results:

• StackCPA demonstrated superior performance compared to other state-of-the-art deep learning models across all three tested datasets.
• Ablation studies confirmed that multi-scale protein pocket features significantly enhance prediction accuracy.
• The model's effectiveness was validated through a case study on epidermal growth factor receptor erbB1 (EGFR) for drug repurposing.

Conclusions:

• StackCPA provides a highly accurate computational tool for predicting compound-protein binding affinity.
• The integration of multi-scale protein pocket features is crucial for improving prediction performance.
• StackCPA shows promise as an effective platform for drug repurposing and accelerating drug development.