UMMAN: Unsupervised Multi-Graph Merge Adversarial Network for Disease Prediction Based on Intestinal Flora

Area of Science:

• Microbiology
• Computational Biology
• Bioinformatics

Background:

• Intestinal flora abundance correlates with human diseases, but diseases arise from complex microbial interplays, not single microbes.
• Predicting diseases from gut microbiome data is difficult due to the intricate and implicit associations among microbes across hosts.
• Existing prediction methods struggle to capture these complex microbial relationships, limiting their performance.

Purpose of the Study:

• To propose a novel architecture, Unsupervised Multi-graph Merge Adversarial Network (UMMAN), for unsupervised learning of multiplex and implicit associations among gut microbes.
• To enhance intestinal flora disease prediction by effectively learning inter-microbial associations from host data.
• To introduce a new approach combining Graph Neural Networks with gut microbiome disease prediction.

Main Methods:

• Developed UMMAN, an unsupervised architecture utilizing Graph Neural Networks to learn node embeddings from a Multi-Graph.
• Constructed an Original-Graph with multiple relation types and a Shuffled-Graph by node disruption.
• Introduced a Node Feature Global Integration (NFGI) module and a joint loss function (adversarial and hybrid attention loss) for graph embedding alignment.

Main Results:

• UMMAN successfully learns multiplex and implicit associations among gut microbes in an unsupervised manner.
• The method demonstrates effectiveness and stability across five benchmark OTU gut microbiome datasets.
• Achieved improved performance in intestinal flora disease prediction by capturing complex microbial interdependencies.

Conclusions:

• UMMAN represents a significant advancement in leveraging Graph Neural Networks for gut microbiome disease prediction.
• The proposed method effectively models complex microbial interactions, overcoming limitations of previous approaches.
• UMMAN offers a robust and stable solution for predicting diseases based on intestinal flora composition.