Learning Microbial Community Structures with Supervised and Unsupervised Non-negative Matrix Factorization.

Area of Science:

• Microbiology
• Bioinformatics
• Computational Biology

Background:

• Understanding microbial community structure is crucial for discerning microbial functions in individuals.
• Microbial communities are conceptualized as interconnected subcommunities of functionally dependent microbes.
• This research focuses on extracting these subcommunities from microbial data.

Purpose of the Study:

• To present methods for extracting microbial subcommunities from data, primarily using Non-Negative Matrix Factorization (NMF).
• To apply both unsupervised and supervised NMF approaches for analyzing microbial community data.
• To extract interpretable information from classification problems using NMF.

Main Methods:

• Utilized Non-Negative Matrix Factorization (NMF), an unsupervised method, for subcommunity extraction.
• Developed a novel supervised NMF method tailored for classification tasks.
• Applied methods to Operational Taxonomic Unit (OTU) data and functional metagenomic data.

Main Results:

• NMF-identified subcommunities demonstrated high accuracy in classification tasks.
• Analysis of mammalian metagenomes revealed macrolide biosynthesis pathways in herbivores, aligning with veterinary findings.
• Time-series microbiome data showed a synchronized shift in tongue and gut microbiomes for an individual, indicating a biological trait.

Conclusions:

• NMF is a powerful tool for identifying key microbial community features, enabling accurate classification and biological interpretation.
• NMF acts as a dimension-reduction technique, simplifying complex microbial data for analyses like temporal pattern searching.
• Supervised NMF offers an improved approach for comparing community structures between groups while maintaining interpretability.