Taxonomic classification cost tracks neither sequencing depth nor community richness at single-sample scale: a measured resource protocol for 16S rRNA amplicon pipelines

Abstract

Marker-gene amplicon workflows are routinely run on shared compute, yet the cores, memory and wall time they are given are chosen by convention and not by measurement. We present a protocol for measuring them, applied to the two dominant stages of a QIIME 2 16S rRNA pipeline, DADA2 denoising and Naive Bayes taxonomic classification, across nine upper-respiratory samples from a paediatric otitis media cohort. The two stages do not consume the same input: denoising reads every sequence, classification only those surviving it. Subsampling one library across a 27-fold range of sequencing depth, denoising wall time rose 14.3-fold while classification changed by 1% and its peak memory not at all (3.11 GiB). Amplicon sequence variant (ASV) richness rose 2.8-fold over that range, so this is not richness saturating: the stage is dominated by a fixed per-invocation cost. Across a body-site gradient of 5 to 70 ASVs, denoising followed read count (exponent 0.75) while classification followed neither: a 5-ASV effusion and a 70-ASV adenoid community cost 40.81 s and 40.79 s. One ASV took 36.20 s and 218 took 37.27 s, 97% fixed cost. Thread-level parallelism offered little benefit. Denoising peaked at 1.18x near 8 threads and then declined; classification was slower at every setting above one job, consuming 10.5 times the CPU at 40. Representative sequences and their taxonomic assignments were identical at 1, 4 and 40 threads, so a reduced allocation changes what the analysis costs, not what it reports. Extending the query set to 10,000 sequences located two distinct boundaries: eight jobs first beat one at roughly 5,000 queries, and fitted fixed and per-query costs become equal at 15,248. Both lie roughly two orders of magnitude above the richest single sample measured. Practically: size denoising by read count, calibrate classification once against the reference in use, request one job for classification below a few thousand sequences, and take throughput from sample-level parallelism. Protocol, data and analysis code are released with the pipeline.