Using the Heterogeneous Network's Self-Adaptive Topological Structure for Drug-Protein Interaction Prediction

The prediction of drug-protein interactions (DPIs) has been the subject of many robust computer approaches using graph neural networks (GNNs). The burden in laboratories and the costs associated with drug discovery and repurposing may be significantly decreased with its help. The unreported indications of many proteins and medications render their clinical activities unclear, nonetheless. Therefore, using the data that is currently available, it is challenging to construct a trustworthy drug-protein heterogeneous network that can characterize the interactions between medicines and proteins. We provide TAG, a DPI prediction approach that may self-adaptively alter the topological framework of the heterogeneous networks, as a solution to this challenge. Graph Neural Networks (GNNs) have recently gained a lot of interest due to their exceptional compatibility with viral proteins and pharmacological substances. Building such a model that accurately depicts the molecular structures of living molecules is, however, no easy feat. Also, newly created GNNs have been overfitting on Coronavirus datasets as they are still not widely used. Next, we propose a new model called Topology-enhanced Graph Neural Network using Multi-hop Gating Mechanism to fix those problems. Firstly, our model is able to learn more specific properties of proteins and chemicals. Conversely, we provide a novel gating approach for improving atom representation using non-neighbor data. In order to implement the drug-protein heterogeneous network, TAG sets up a representation learning module that is based on graph attention network. The network's topology may be fine-tuned in response to the training loss, and it can learn the connections between nodes using their properties. Lastly, TAG uses the embeddings of medicines and proteins to forecast the likelihood of their interactions. The experimental findings demonstrate that TAG is capable of significantly enhancing the network's topological structure. Across a range of performance criteria, TAG surpasses other state-of-the-art DPI prediction approaches. These examples demonstrate the practicality of TAG in handling partial data and unstable networks. Additional evidence that TAG is effective for finding new DPIs may be found in the case studies included at the top of the prediction results.