Incorporating metabolic activity, taxonomy and community structure to improve microbiome-based predictive models for...

Area of Science:

• Microbiome research
• Computational biology
• Host-microbiome interactions

Background:

• The human gut microbiome plays a crucial role in host health and disease.
• Predicting host phenotypes from microbiome data is challenging due to complexity.
• Existing computational methods often lack interpretability.

Purpose of the Study:

• To develop a novel interpretable neural network, MicroKPNN, for microbiome-based human host phenotype prediction.
• To integrate prior biological knowledge into a machine learning model for improved accuracy.
• To provide interpretable insights into the relationship between the microbiome and host phenotypes.

Main Methods:

• Developed MicroKPNN, a prior-knowledge guided shallow neural network.
• Incorporated bacterial metabolic activities, phylogenetic relationships, and community structure as prior knowledge.
• Applied MicroKPNN to seven gut microbiome datasets across five human diseases.

Main Results:

• MicroKPNN significantly improved host phenotype prediction accuracy compared to fully connected neural networks across all datasets.
• MicroKPNN outperformed the deep-learning approach DeepMicro in all tested cases.
• The model provided interpretable importance scores for hidden nodes, explaining microbiome contributions to predictions.

Conclusions:

• Prior knowledge integration enhances microbiome-based phenotype prediction accuracy.
• MicroKPNN offers a powerful and interpretable tool for studying host-microbiome interactions.
• The approach can validate existing findings and suggest new research avenues in microbiome science.