A comparative study of supervised and unsupervised machine learning algorithms applied to human microbiome.

Area of Science:

• Microbiology
• Bioinformatics
• Computational Biology

Background:

• The human microbiome comprises diverse microbial species influencing health and disease.
• Analyzing complex microbiome data presents challenges, necessitating advanced computational tools.
• Machine learning algorithms are increasingly employed for microbiome data interpretation.

Purpose of the Study:

• To evaluate dimensionality reduction methods for 16S rRNA gene sequencing data.
• To assess the predictive performance of supervised machine learning on reduced microbiome datasets.
• To identify key microbial features for predicting patient conditions.

Main Methods:

• Analysis of 16S rRNA gene sequencing data from healthy controls and patients with adenoma or colorectal cancer.
• Application of network-based (graph) and projection (NMF, PCA) methods for dimensionality reduction.
• Implementation of supervised machine learning algorithms for predictive modeling.

Main Results:

• Graph-based methods reduced data from 255 to 78 features with a modularity score of 0.73.
• Projection methods reduced data to 7 key features.
• Supervised machine learning achieved comparable predictive performance on original, 78-feature, and 7-feature datasets.

Conclusions:

• Graph-based and projection methods are effective for interpreting 16S rRNA gene sequencing data.
• Machine learning on refined features provides robust predictive performance.
• Specific microbes like Bacteroides, Prevotella, and Fusobacterium are critical predictors of patient status.