A permutable MLP-like architecture for disease prediction from gut metagenomic data

Area of Science:

• Microbiology
• Bioinformatics
• Computational Biology

Background:

• Metagenomic data is vital for understanding microbe-disease links but presents challenges like high dimensionality and sparsity for deep learning.
• Transforming metagenomic data using phylogenetic trees improves classification with Convolutional Neural Networks (CNNs).

Purpose of the Study:

• To introduce Metagenomic Permutator (MetaP), a novel deep learning model inspired by Permutable MLP-like architectures.
• To leverage phylogenetic information within a 2D matrix for enhanced metagenomic data classification.

Main Methods:

• Developed Metagenomic Permutator (MetaP) utilizing a Permutable MLP-like network structure.
• Applied MetaP to 2D matrices derived from metagenomic abundance data and phylogenetic trees.
• Employed SHAP (SHapley Additive exPlanations) for model interpretability.

Main Results:

• MetaP achieved competitive performance against existing deep neural networks and traditional machine learning methods.
• The model demonstrates strong potential for multi-classification tasks and analysis of large sample sizes.
• SHAP analysis successfully identified microbial features linked to specific diseases.

Conclusions:

• Metagenomic Permutator (MetaP) offers a promising deep learning framework for metagenomic data analysis.
• The model effectively captures phylogenetic information, improving classification accuracy and interpretability.
• MetaP shows significant potential for advancing microbiome-disease research, particularly with large-scale datasets.