Detecting and mitigating doppelgänger bias in microbiome data: impacts on machine learning and disease classification

Area of Science:

• Microbiome research
• Computational biology
• Bioinformatics

Background:

• Microbiome sample similarity can impact analytical outcomes.
• Highly similar samples, termed 'doppelgänger pairs,' are often overlooked in studies.
• These pairs can introduce bias in machine learning and statistical analyses.

Purpose of the Study:

• To investigate the impact of doppelgänger pairs on microbiome data analysis.
• To quantify the inflation of machine learning performance and bias in statistical tests caused by these pairs.
• To evaluate the effect of removing doppelgänger pairs on analytical stability and biological interpretability.

Main Methods:

• Simulated and real-world microbiome datasets (16S, shotgun metagenomic) were analyzed.
• Machine learning models (KNN, SVM, Random Forest) were trained and evaluated with and without doppelgänger pairs.
• Statistical association tests and microbial network analyses were performed.
• The impact of removing doppelgänger pairs on classification accuracy, false-positive rates, effect size stability, and network topology was assessed.

Main Results:

• Even a small proportion (1-10%) of doppelgänger pairs significantly inflated machine learning performance (15-30% accuracy boost).
• Doppelgänger pairs biased statistical tests, increasing false-positive rates and reducing effect size stability.
• Removal of doppelgänger pairs reduced bootstrap variance by up to 28.3% and yielded more stable microbial networks.
• These effects were consistent across different data types and disease cohorts (CRC, IBD, CDI, obesity).

Conclusions:

• Doppelgänger pairs are a significant source of analytical noise and false discoveries in microbiome research.
• Accounting for highly similar samples is essential for robust and biologically meaningful microbiome insights.
• Removing doppelgänger pairs improves the accuracy and reliability of machine learning and statistical analyses in microbiome studies.