IRS: iterative reference selection improves normalization of microbiome sequencing data

Microbiome studies often seek to determine how the absolute abundances of individual taxa change across biological conditions, yet sequencing read counts are sample-specific scaled representations of those abundances. Because sampling depth can differ across samples, fold changes calculated directly from sequencing read counts do not generally represent absolute-abundance fold changes. Normalization methods attempt to account for these between-sample differences in sampling depth, but their accuracy depends on the reference used. In particular, total-sum scaling uses all taxa as the reference and can introduce compositional bias. Reference-based methods instead rely on taxa that are stable across conditions, but contamination of the reference set by differentially abundant (DA) taxa can distort sampling-depth estimation and downstream inference. Here, we present iterative reference selection (IRS), a robust normalization method that iteratively screens and refines a candidate reference set to exclude DA taxa. By deriving a clean reference set, IRS accurately captures between-sample differences in sampling depth and recovers absolute-abundance fold changes. Benchmarking using simulations and datasets with experimental absolute quantification shows that IRS outperforms standard scaling and existing reference-based methods in controlling false discovery rates while maintaining power.