Prediction of human gut microbiota-drug associations based on multiple graph convolutional networks

Abstract Motivation Many associations among gut microbiota, diseases, and therapeutic drugs have been reported in various clinical settings, providing a new layer of clues for understanding the variations of therapeutic features among patients. However, experimental strategies heavily suffered from the technic challenges of the isolation and in vitro culture of individual microbiota species. Results In this study, we developed a multiple graph convolutional networks model (GutMDA) for predicting gut micro-biota-drug associations to complement experimental strategies and existing prediction methods. We constructed gut microbiota species-species based on phylogenetical-based similarity, drug-drug network based on chemical structural similarity, and disease-disease association networks based on MESH-derived semantic similarity. These networks were subsequently integrated into a microbiota-disease-drug triple network and trained on graph convolutional network models to predict microbiota-drug associations. The results showed that GutMDA demonstrates superior performance with the average AUC and AUPR over 0.99 on MASI dataset in five-fold cross validations. The case studies showed that the most of top-20 predicted associations have been validated experimentally in recent publications, which further demonstrated the effectiveness of GutMDA. In summary, GutMDA provided an effective option for predicting gut microbiota-drug associations. Availability and implementation Original data and codes are available at https://github.com/Wang-Shuaiqi/GutMDA.