Interpretable Log Contrasts for the Classification of Health Biomarkers: a New Approach to Balance Selection.

Area of Science:

• Biotechnology and Bioinformatics
• Microbiome Research
• Computational Biology

Background:

• High-throughput sequencing enables cost-effective molecular profiling of tissues, including microbial, RNA, and metabolite abundances.
• Omics data, particularly microbial composition, are valuable biomarkers for disease prediction but present analytical challenges due to their relative nature.
• Conventional analysis methods often struggle with high-dimensional compositional data, relying on complex normalization procedures.

Purpose of the Study:

• To highlight the relative nature of health biomarkers derived from omics data.
• To review classification methods for relative data in biological and medical research.
• To benchmark data transformations for regularized logistic regression across various biomarker types.

Main Methods:

• Utilized log contrasts, termed balances, to prepare compositional data for classification.
• Developed and proposed discriminative balance analysis (DBA) for selecting bacterial groups.
• Benchmarked transformations for regularized logistic regression using different biomarker types.

Main Results:

• Balances effectively prepare relative omics data for classification tasks.
• Discriminative balance analysis (DBA) efficiently identifies key bacterial groups (pairs and trios) for discrimination.
• DBA provides a fast, accurate, and interpretable alternative to traditional data normalization methods.

Conclusions:

• Discriminative balance analysis offers a normalization-free approach to microbiome data classification.
• This method enhances interpretability and reduces feature space without compromising classifier performance.
• DBA is a valuable tool for disease biomarker discovery using high-throughput sequencing data.