Metabolite-disease Associations Prediction Based on Multi-kernel Learning and Graph Attention Networks

A growing number of scholars have confirmed through research that there is a strong correlation between changes in metabolite levels in the human body and various complex diseases. In this study, we propose a new framework (MKLGAT-MDA) that integrates multi-kernel learning and graph attention network (GAT) for enhanced metabolite-disease associations prediction. First, the study employs multi-kernel learning to combine different similarity matrices to capture feature information from different perspectives of metabolites and diseases. Then a GAT network with a path-learning graph space operator is employed to generate meaningful spatial convolutions from the paths in GAT to improve the prediction performance. The average AUC at 5-cv is 0.9562 and the average AUPR is 0.9575. The experimental findings indicate that our model surpasses current methods regarding prediction accuracy and reliability.