The accuracy of absolute differential abundance analysis from relative count data

Area of Science:

• Microbiology
• Genomics
• Bioinformatics

Background:

• Next-generation sequencing (NGS) generates relative abundance data, often used as a proxy for absolute abundances.
• Compositional effects in relative abundance data can lead to spurious differences (false positives) in differential abundance analyses.
• Renormalization strategies are employed to mitigate these compositional effects, but their effectiveness remains unclear.

Purpose of the Study:

• To explore the consistency of differential abundance calling between renormalized relative abundances and absolute abundances using simulated data.
• To assess the impact of widespread changes in feature abundance on the accuracy of differential abundance analysis.
• To evaluate the predictive utility of relative abundance data features for identifying discrepancies in differential abundance calling.

Main Methods:

• Utilized simulated data to compare differential abundance calling on renormalized relative abundances versus absolute abundances.
• Analyzed real-world datasets from 16S metabarcoding, expression arrays, bulk RNA-seq, and single-cell RNA-seq experiments.
• Evaluated summary features of relative abundance data, such as sparsity and prevalence of feature-level change, for predictive power.

Main Results:

• Overall consistency between renormalized relative and absolute abundances is high (median sensitivity 0.91, specificity 0.89).
• Consistency decreases significantly when there are widespread changes in feature abundance across conditions.
• Real data sets with greater changes between conditions exhibit higher false positive rates.
• Sparsity and feature-level change prevalence in relative abundance data predict discrepancies in differential abundance calling.

Conclusions:

• While renormalization improves consistency, relative abundance data can still yield false positives, particularly with substantial biological variation.
• Careful consideration of data characteristics like sparsity and feature-level change is crucial for interpreting differential abundance results.
• Summary features of relative abundance data can offer insights into potential inaccuracies in differential abundance analyses.