A microbial knowledge graph-based deep learning model for predicting candidate microbes for target hosts.

Area of Science:

• Microbial Ecology
• Computational Biology
• Genomics

Background:

• Predicting microbe-host interactions (MHIs) is crucial for understanding microbiome dynamics, microbial evolution, and global health.
• Characterizing complex microbe-host signaling mechanisms presents a significant challenge.
• Computational methods offer a cost-effective alternative to experimental approaches for identifying MHIs.

Purpose of the Study:

• To develop a novel computational model for predicting candidate microbes interacting with specific hosts.
• To leverage heterogeneous microbial network (HMN) and knowledge graph embedding for MHI prediction.
• To enhance the understanding of fundamental regulatory mechanisms in microbe-host relationships.

Main Methods:

• Constructed a heterogeneous microbial network (HMN) integrating human proteins, viruses, and pathogenic bacteria with their attributes.
• Employed a knowledge graph embedding strategy to capture global network topology.
• Utilized natural language processing (NLP) to extract local biological attribute information.
• Integrated local and global information into a blended deep neural network (DNN) for prediction.

Main Results:

• The KGVHI model achieved excellent performance on three MHI datasets, outperforming state-of-the-art methods.
• Case studies involving pathogenic bacteria demonstrated KGVHI's strong predictive capability for potential MHI pairs.
• The model effectively combines network structure and biological attributes for accurate MHI prediction.

Conclusions:

• KGVHI provides a powerful computational tool for predicting microbe-host interactions.
• The findings facilitate insights into microbial ecology and the development of targeted therapies for microbial infections.
• This approach advances the systematic characterization of microbe-host signaling.