Graph convolution network based on meta-paths and mutual information for drug-target interaction prediction.

BACKGROUND: Predicting drug-target interactions (DTIs) plays a pivotal role in accelerating drug repositioning by prioritizing candidate drugs and reducing experimental costs. Despite advancements in deep learning, several challenges still require further exploration, including sparsity and inadequate representation of feature relationships. RESULTS: We propose GCNMM, a novel graph convolutional network based on meta-paths and mutual information, to predict latent DTIs in drug-target heterogeneous networks. Our approach begins by constructing a fused DTI network based on meta-paths and a graph attention network. We compute multiple similarity networks by using Jaccard coefficients and integrate them into the fused drug and target similarity networks through entropy-based fusion. These networks are then jointly processed by graph convolutional auto-encoder to generate low-dimensional feature representations. To preserve the topological structure of the original network in the embedding space and strengthen the relationship between the input and latent representations, we incorporate spatial topological consistency and mutual information maximization as dual optimization objectives. CONCLUSIONS: The experimental results illustrate that GCNMM exhibits superior performance to existing baseline models in DTI prediction. Furthermore, case studies validate the practical effectiveness of GCNMM, highlighting its potential in DTI prediction and drug repositioning.