Multi-class boosting for the analysis of multiple incomplete views on microbiome data.

Area of Science:

• Microbiome research
• Computational biology
• Machine learning

Background:

• Microbiome dysbiosis is linked to various diseases.
• Machine learning (ML) can identify patterns and build predictive models from microbiome data.
• Traditional ML methods struggle with multi-view and incomplete microbiome datasets from varied processing pipelines.

Purpose of the Study:

• To develop an advanced multi-view learning method capable of handling incomplete datasets.
• To improve disease prediction accuracy using diverse microbiome data views.
• To address limitations of existing multi-view learning algorithms in microbiome analysis.

Main Methods:

• Proposed irBoost.SH, an extension of the rBoost.SH multi-view boosting algorithm.
• Incorporated multi-armed bandits to dynamically select the most informative data view at each iteration.
• Enabled analysis of incomplete multi-view datasets and multi-class classification tasks.

Main Results:

• irBoost.SH consistently outperformed single-view models, rBoost.SH, and feature concatenation methods across 5 microbiome datasets.
• Achieved significant F1-score improvements: 11.8% for Autism Spectrum Disorder prediction and 114% for Colorectal Cancer prediction.
• Demonstrated superior performance in multi-class classification and handling of incomplete data views.

Conclusions:

• irBoost.SH shows outstanding performance in microbiome data analysis.
• The method effectively leverages multiple feature sets from different data processing pipelines.
• irBoost.SH offers a powerful tool for advancing microbiome-based disease prediction and research.