Deep in the Bowel: Highly Interpretable Neural Encoder-Decoder Networks Predict Gut Metabolites from Gut Microbiome

Area of Science:

• Computational Biology
• Systems Biology
• Bioinformatics

Background:

• Next-generation sequencing (NGS) and LC-MS enable large-scale microbiome and metabolome profiling.
• Predicting metabolite abundances from microbe abundances is a key challenge in multi-omics research.

Purpose of the Study:

• To develop a novel sparse neural encoder-decoder network for predicting metabolite abundances from microbe abundances.
• To investigate the interpretability and clinical relevance of the learned microbe-metabolite relationships.

Main Methods:

• A sparse neural encoder-decoder model was proposed.
• The model was trained and evaluated on paired microbiome and metabolome data from inflammatory bowel disease (IBD) patients.
• Compositional data analysis methods were used for data pre-processing.

Main Results:

• The neural encoder-decoder model demonstrated superior accuracy, sparsity, and stability compared to linear methods.
• The model's latent space captured clinically meaningful microbe-metabolite relationships.
• Learned latent features accurately predicted IBD diagnosis and treatment status.

Conclusions:

• The constrained neural network forms an interpretable directed graph, highlighting key microbe-metabolite axes.
• The multi-omics workflow is generalizable to various -omics data types.
• This represents the first application of neural encoder-decoders for interpretable multi-omics integration.