An improved hybrid graph method for drug side-effects prediction

Severe side effects resulting from drug treatments can directly impact human health and are one of the common reasons for drug development failures. Therefore, identifying the associations between drugs and side effects can help reduce medication risks for patients and lower development costs for pharmaceutical companies. As ongoing pharmacovigilance and medication monitoring have shown to be costly and time-consuming, computational techniques have become attractive substitutes. But the majority of current computational techniques ignore the importance of capturing local features and global features within drug-side effect associations. To address this issue, we propose an improved hybrid graph model called IHDSP, which constructs drug similarity and side effect similarity networks. A hybrid graph, primarily composed of GCN and GAT, is then used to explore the deep connections between drugs and side effects. Finally, a random forest classifier serves to train and obtain the final prediction results. IHDSP performs better than state-of-the-art prediction techniques, according to experimental results.