Dcomg: Drug Combination Prediction by Applying Gnns on DDI Node2vec Features

Recent studies have been indicating that many clinical drug combinations surpass single-drug therapy efficacy. Machine learning, deep learning, network analysis, and search algorithms have been considered to facilitate the discovery of synergistic drug combinations, and two of the best state-of-the-art models in this area are under the deep learning category. In this paper, we present DComG, a Graph Auto Encoder method to predict synergistic drug combinations. Using the dataset provided in DCDB, our analysis shows tremendous improvement in the performance of predicting new drug combinations over previously introduced state-of the-art models by an average of 4% in ROC-AUC scores. We highlight the importance of drug-drug interactions (DDI) in the form of node2vec features of DComG graph inputs for predicting new drug combinations. Finally, we address the results of our model in terms of biological interpretations of drug combinations based on recent medical drug combination papers in the literature.