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

Area of Science:

• Microbiology
• Pharmacology
• Computational Biology

Background:

• Microorganisms in the human body significantly impact health, and drugs can regulate these interactions.
• Identifying microbe-drug associations (MDAs) is crucial for understanding and managing human health.
• Existing computational methods for MDAs face challenges due to the lack of validated negative samples, leading to inaccurate predictions.

Purpose of the Study:

• To develop a novel computational model, Graph Representation Learning and Positive-Unlabeled Learning (GRL-PUL), for inferring potential microbe-drug associations (MDAs).
• To address the issue of false negatives in MDA prediction by developing a reliable negative sample screening method.
• To enhance the accuracy and reliability of computational prediction of MDAs.

Main Methods:

• Screening reliable negative samples using weighted matrix factorization and a PU-bagging strategy on a microbe-drug bipartite network.
• Constructing a microbe-drug heterogeneous network by integrating multi-model attributes.
• Employing a graph attention auto-encoder, combining graph convolutional networks and graph attention networks, for informative embedding extraction.
• Utilizing a modified random forest classifier for the final prediction of MDAs.

Main Results:

• The proposed GRL-PUL model demonstrated superior performance compared to five baseline models across three benchmark datasets, evidenced by higher AUC, AUPR, ACC, F1-score, and MCC.
• Case studies confirmed the capability of GRL-PUL in predicting latent MDAs.
• The reliable negative sample selection module significantly improved the prediction performance when integrated into other state-of-the-art models.

Conclusions:

• The GRL-PUL model offers a robust and accurate computational approach for identifying potential microbe-drug associations (MDAs).
• The developed negative sample screening strategy effectively mitigates false negative issues, enhancing prediction reliability.
• This work provides a valuable tool for researchers in microbiology, pharmacology, and computational biology for discovering novel MDAs and their health implications.