Multiclass Disease Classification from Microbial Whole-Community Metagenomes

Area of Science:

• Microbiome research
• Computational biology
• Machine learning applications in medicine

Background:

• The human microbiome plays a crucial role in health and disease.
• Developing non-invasive diagnostic tools is a significant challenge in modern medicine.
• Microbiome analysis offers a promising avenue for disease screening.

Purpose of the Study:

• To develop the first large-scale, multiclass microbiome disease classifier.
• To discriminate between 18 different diseases and healthy states using metagenomic data.
• To compare the performance of different machine learning models for microbiome-based diagnostics.

Main Methods:

• Utilized 5643 aggregated and annotated whole-community metagenomes.
• Implemented and compared three machine learning models: random forests, deep neural nets, and a graph convolutional architecture.
• The graph convolutional architecture leveraged the phylogenetic tree structure of microbial communities.

Main Results:

• The graph convolutional model achieved 75% average test-set accuracy, outperforming deep neural nets.
• Achieved 92.1% average area under the ROC curve (AUC) and 50% average area under the precision-recall curve (AUPR).
• The graph convolutional network and random forest models showed complementary error profiles, with combined top-3 accuracy exceeding 90%.

Conclusions:

• Predictive, disease-specific signatures exist within the microbiome.
• These signatures can be leveraged for accurate and non-invasive disease diagnostics.
• Advanced machine learning, particularly graph convolutional networks, shows significant promise for microbiome-based health applications.