Automatic disease prediction from human gut metagenomic data using boosting GraphSAGE.

Area of Science:

• Microbiome research
• Computational biology
• Machine learning

Background:

• The human microbiome is crucial for health, with distinct profiles linked to various diseases.
• Advances in sequencing enable large-scale microbiome data analysis for host phenotype classification.

Purpose of the Study:

• To develop a novel deep learning framework for automatic disease prediction from metagenomic data.
• To enhance the accuracy of classifying host phenotypes based on microbiome profiles.

Main Methods:

• A boosting GraphSAGE deep learning framework was developed.
• The framework includes a Metagenomic Disease graph (MD-graph) construction module and a Disease Prediction Network (DP-Net) module.
• Graph construction represents metagenomic samples as nodes, capturing inter-sample relationships.

Main Results:

• The framework achieved high performance on real and synthetic datasets for inflammatory bowel disease and colorectal cancer.
• For inflammatory bowel disease, the model reached an AUC of 93%, Accuracy of 95%, F1-score of 95%, and AUPRC of 95%.
• For colorectal cancer, the model achieved an AUC of 90%, Accuracy of 91%, F1-score of 87%, and AUPRC of 93%.

Conclusions:

• The proposed deep learning framework significantly outperforms existing machine and deep learning models.
• The model demonstrates superior classification accuracy, AUC, F1-score, and AUPRC for both synthetic and real metagenomic data.