Machine learning-based feature selection to search stable microbial biomarkers: application to inflammatory bowel...

Area of Science:

• Microbiome research
• Bioinformatics
• Machine learning applications in health

Background:

• Machine learning, particularly recursive feature elimination (RFE), is used for microbiome biomarker discovery.
• Existing RFE methods face challenges with feature selection stability.
• This study addresses the need for more stable and reliable feature selection in microbiome analysis.

Purpose of the Study:

• To enhance the stability of feature selection in microbiome biomarker discovery.
• To identify robust biomarkers for inflammatory bowel disease (IBD) using improved machine learning techniques.
• To evaluate different data transformation methods for optimizing stability and performance.

Main Methods:

• Utilized gut microbiome abundance matrices from 1,569 samples (IBD patients and healthy controls).
• Applied data transformation, including Bray-Curtis similarity mapping, prior to recursive feature elimination (RFE).
• Evaluated feature selection stability using multiple metrics and compared 8 machine learning algorithms (Multilayer Perceptron, Random Forest).

Main Results:

• Data transformation, specifically Bray-Curtis similarity mapping before RFE, significantly improved feature selection stability.
• The developed pipeline maintained high classification performance for IBD detection.
• Identified 14 species-level biomarkers for IBD, with roles analyzed using Shapley additive explanations.
• Random Forest outperformed other algorithms for generalizability with a limited biomarker set.

Conclusions:

• The study presents a method to improve biomarker discovery stability in metataxonomics without compromising classification performance.
• The findings offer valuable insights for future comparative microbiome studies.
• The identified IBD biomarkers and optimized pipeline contribute to advancing microbiome-based diagnostics.