Predicting compound-protein interaction using hierarchical graph convolutional networks.

Area of Science:

• Computational chemistry
• Bioinformatics
• Machine learning in drug discovery

Background:

• Convolutional neural networks (CNNs) show promise beyond computer vision, particularly in drug discovery.
• Predicting compound-protein interactions is crucial for identifying potential drug candidates.

Purpose of the Study:

• To develop an advanced computational method for predicting compound-protein interactions using compound structure and protein sequence.
• To leverage hierarchical graph convolutional networks (HGCNs) for effective molecular representation.

Main Methods:

• A hierarchical graph convolutional network (HGCN) was developed to encode small molecules by aggregating atom and substructure embeddings.
• The HGCN was integrated with a 1D convolutional network to create a comprehensive model for interaction prediction.
• The Grad-CAM technique was employed for visualizing the contribution of individual amino acids in predictions.

Main Results:

• The proposed HGCN-based model demonstrated superior performance in predicting compound-protein interactions compared to existing GCN-based and sequence-based methods (e.g., DeepDTA).
• The model's predictions are explainable, allowing for the identification of key amino acid residues involved in compound-protein interactions.

Conclusions:

• The developed HGCN model offers a powerful and explainable approach for predicting compound-protein interactions.
• This method has significant implications for accelerating drug discovery and development by improving the accuracy and interpretability of interaction predictions.