Efficient substructure feature encoding based on graph neural network blocks for drug-target interaction prediction.

Background: Predicting drug-target interaction (DTI) is a crucial phase in drug discovery. The core of DTI prediction lies in appropriate representations learning of drug and target. Previous studies have confirmed the effectiveness of graph neural networks (GNNs) in drug compound feature encoding. However, these GNN-based methods do not effectively balance the local substructural features with the overall structural properties of the drug molecular graph. Methods: In this study, we proposed a novel model named GNNBlockDTI to address the current challenges. We combined multiple layers of GNN as a GNNBlock unit to capture the hidden structural patterns from drug graph within local ranges. Based on the proposed GNNBlock, we introduced a feature enhancement strategy to re-encode the obtained structural features, and utilized gating units for redundant information filtering. To simulate the essence of DTI that only protein fragments in the binding pocket interact with drugs, we provided a local encoding strategy for target protein using variant convolutional networks. Results: Experimental results on three benchmark datasets demonstrated that GNNBlockDTI is highly competitive compared to the state-of-the-art models. Moreover, the case study of drug candidates ranking against different targets affirms the practical effectiveness of GNNBlockDTI. The source code for this study is available at https://github.com/Ptexys/GNNBlockDTI.