GRL–PUL: predicting microbe–drug association based on graph representation learning and positive unlabeled learning

Abstract

Extensive research has confirmed the widespread presence of microorganisms in the human body and their crucial impact on human health, with drugs being an effective method of regulation. Hence it is essential to identify potential microbe–drug associations (MDAs). Owing to the limitations of wet experiments, such as high costs and long durations, computational methods for binary classification tasks have become valuable alternatives for traditional experimental approaches. Since validated negative MDAs are absent in existing datasets, most methods randomly sample negatives from unlabeled data, which evidently leads to false negative issues. In this manuscript, we propose a novel model based on graph representation learning and positive-unlabeled learning (GRL–PUL), to infer potential MDAs. Firstly, we screen reliable negative samples by applying weighted matrix factorization and the PU-bagging strategy on the known microbe–drug bipartite network. Then, we combine muti-model attributes and constructed a microbe–drug heterogeneous network. After that, graph attention auto-encoder module, an encoder combining graph convolutional networks and graph attention networks, is introduced to extract informative embeddings based on the microbe–drug heterogeneous network. Lastly, we adopt a modified random forest as the final classifier. Comparison experiments with five baseline models on three benchmark datasets show that our model surpasses other methods in terms of the AUC, AUPR, ACC, F1-score and MCC. Moreover, several case studies show that GRL–PUL could capably predict latent MDAs. Notably, we further verify the effectiveness of a reliable negative sample selection module by migrating it to other state-of-the-art models, and the experimental results demonstrate its ability to substantially improve their prediction performance.