Interpretable and accurate prediction models for metagenomics data.

Area of Science:

• Microbiome research
• Metagenomics
• Machine learning in healthcare

Background:

• Microbiome biomarkers are crucial for diagnosing and evaluating diseases like cancer and cardio-metabolic conditions.
• Current machine learning models lack interpretability, hindering clinical trust and routine use.
• There is a need for novel methods that offer both accuracy and biological insight.

Purpose of the Study:

• Introduce predomics, a novel machine learning approach for microbiome data analysis.
• Develop a method that provides interpretable and accurate predictive signatures.
• Facilitate reliable diagnostic decisions in the microbiome field.

Main Methods:

• Predomics is an original machine learning approach inspired by microbial ecosystem interactions.
• The method is tailored for metagenomics data, discovering predictive signatures.
• Model decisions are based on a simple score derived from microbiome abundance measurements.

Main Results:

• Predomics models demonstrate simplicity and high interpretability across over 100 datasets.
• The models achieve accuracy comparable to state-of-the-art methods.
• Successfully predicted body corpulence and metabolic improvement post-bariatric surgery using pre-surgery microbiome data.

Conclusions:

• Predomics provides a reliable and trustworthy approach for diagnostic decisions in microbiome research.
• The algorithm aligns with requirements for explainable artificial intelligence in medicine.
• Offers biological insights and decipherable predictability signatures for various conditions.