A realistic simulation-based benchmark of microbiome normalization in sample stratification and taxa-level analysis

Area of Science:

• Microbiome research
• Bioinformatics
• Computational biology

Background:

• Normalization is essential in microbiome studies due to sequencing depth and sparsity.
• Assessing normalization method performance is challenging because true biological signals are unknown in real datasets.

Purpose of the Study:

• To develop a simulation-based framework for evaluating microbiome normalization methods.
• To quantitatively compare normalization methods using realistic datasets with known ground truth.

Main Methods:

• Developed a simulation framework informed by real microbiome data.
• Generated realistic datasets with known ground truth for quantitative comparison.
• Evaluated normalization methods at both sample and taxa levels.

Main Results:

• Method performance varied with taxonomic resolution and confounding of sequencing depth with group structure.
• Model-based methods (edgeR-TMM, DESeq2) showed better recovery of taxa-level differences and sample separation.
• Sequencing depth differences alone could create false positives; rarefaction maintained Type I error control.
• Sample separation alone is insufficient for judging normalization performance.

Conclusions:

• A simulation-based framework provides a coherent strategy for evaluating microbiome normalization methods.
• Model-based normalization-factor methods generally outperform others in preserving true biological signals.
• The framework enables realistic, data-dependent evaluation of existing and new normalization approaches.