Best practices for developing microbiome-based disease diagnostic classifiers through machine learning.

Area of Science:

• Microbiome Research
• Computational Biology
• Machine Learning Applications

Background:

• The human gut microbiome plays a critical role in various diseases.
• Machine learning (ML) offers potential for developing diagnostic models from microbiome data.
• Model performance and generalizability depend heavily on preprocessing, batch effect removal, and algorithm selection.

Purpose of the Study:

• To establish a generally applicable machine learning workflow for constructing diagnostic models from human gut microbiome data.
• To optimize and benchmark different tools and parameters for each step of the ML process.
• To provide a comprehensive guideline for future microbiome-based disease diagnostics.

Main Methods:

• Sequentially optimized three key ML steps: data preprocessing, batch effect removal, and algorithm selection.
• Utilized 83 gut microbiome cohorts across 20 diseases for optimization and benchmarking.
• Tested 156 tool-parameter-algorithm combinations, evaluating performance using internal- and external- AUCs.

Main Results:

• Identified optimal data preprocessing methods for both regression and non-regression algorithms.
• "ComBat" function from the sva R package was identified as an effective batch effect removal method.
• Ridge and Random Forest algorithms were found to be the top-performing ML algorithms.

Conclusions:

• The developed optimized workflow is generally applicable across various diseases.
• The workflow demonstrates comparable performance to previous disease-specific optimization methods.
• This provides a robust guideline for developing microbiome-based diagnostic models, advancing medical diagnostics.