A novel method for predicting the interaction between Chinese herbal compounds and targets

The active components and target points of traditional Chinese medicine are highly complex and difficult to ascertain. In recent years, computational methods have become an effective approach for predicting compound-target interactions. However, effective utilization of the topological structural features of the network remains an urgent challenge. Traditional prediction models tend to focus excessively on the relationships between nodes while overlooking the features of the edges. To address these issues, we propose a TCMCPI model to predict the interactions between herbal compounds and target points. First, we transform the CPI network into a line graph by establishing a topological structure of the edges. We then used GCN and CNN to extract features from the compounds and targets. These features are fused together to represent the nodes of the line graph, and GAT is utilized to learn the node features. Finally, we map the node features of the line graph back to the edge features in the CPI network by merging the adjacent edge features into node features. The experimental results demonstrate that TCMCPI outperforms other baseline methods in predicting interactions between herbal compounds and target proteins, and high AUROC and AUPR evaluation metrics attest to its superior predictive performance.