Omics data in relative values are almost subcompositionally coherent

Area of Science:

• Bioinformatics
• Statistical genomics
• Data science

Background:

• Omics data are compositional and often normalized to relative abundances.
• Subcompositional incoherence is a known issue, where relative abundances change upon data subsetting.
• This problem's impact on large omics datasets was previously uninvestigated.

Purpose of the Study:

• To evaluate subcompositional coherence in large omics datasets.
• To compare statistical outputs from full omics compositions versus abundance-weighted subcompositions.
• To assess the implications for using relative abundances versus logratio transformations.

Main Methods:

• Analysis of five diverse omics datasets (metagenomics, transcriptomics, metabolomics).
• Generation of 100 random subcompositions (one-third of original features) using an abundance-weighted scheme.
• Comparison of statistical outputs (unsupervised and supervised learning) between full and subcomposed data.

Main Results:

• Raw omics data exhibited near-perfect subcompositional coherence (concordance ≥0.98-0.99).
• Supervised models (linear regression, PLS, random forest, linear mixed models) also showed high coherence.
• Abundance-weighted subcompositions closely mirrored full composition results.

Conclusions:

• Large omics datasets normalized to relative abundances demonstrate substantial subcompositional coherence.
• The criticism regarding incoherence of relative omics data may be overstated under specific subcomposition schemes.
• This supports the use of relative abundances, potentially simplifying omics data analysis.