Drug-Drug and Drug-Protein Link Prediction on DTINet Dataset

In the pharmaceutical industry, predicting drug-drug interactions (DDIs) and drug-protein interactions (DPIs) is crucial for ensuring security of patient and improving treatment outcomes. Classical laboratory strategies for identifying these interactions are often costly and requires too much time. This study uses computational methods to predict DDIs and DPIs with graph convolutional networks (GCNs) applied to the DTINet dataset. Two different methods were used: the first focuses on drug-drug link prediction using drug-drug interaction data, while the second combines multiple types of drug-drug interactions, drug-protein interactions, drug-disease associations, and drug-side effect relationships to improve drug-protein link prediction. The results indicate that these methods are efficient, with the drug-drug link prediction model's accuracy is 0.8724 and AUROC score was calculated as 0.9428, and the drug-protein link prediction model achieved an accuracy level of 0.9150 and an AUROC of 0.9680. These findings shows the potential of GCNs in accurately predicting new interactions, which can help in drug repurposing and combination therapy development.