Link Prediction of biomedical data using Hetrosim GCN-GAT

With the emergence of large number of Electronic Medical Records (EMRs), the researchers turned their attention towards the prediction, recommendation systems with these available data. This paper deals with the input of the patient details and the drugs administered to them for the different health conditions. The Drug to Drug Interaction (DDI) prediction is essential in order to avoid the unintended adverse reactions and making the drugs less effective. It is represented in the form of graphs with nodes and edges/links. Link prediction is often used to detect the existence of link or edges present in between the specified nodes. Graph Neural Network based algorithm HetrosimGCN-GAT is proposed to address the underlying link prediction tasks. A heterogeneous network is initially built using the drug-drug, protein-protein similarity matrices and drug protein matrices. The embedded representation of the drug and protein is then obtained using the graph convolutional neural network (GCN). Graph attention mechanism provides a weighting factor for every connect in order to assess the significance or highlight the contribution of distinct neighbourhood nodes to the central node. The unknown DDI and Drug Protein Interaction is predicted by applying the inner product decoder. The performance of HetrosimGCN-GAT is evaluated against the baseline methods and its outperforms the baseline methods.