Meta-analysis of the agreement and accuracy of functional inference from 16S amplicon studies in arsenic-affected soils and aquifers

Abstract Background. Almost all archived environmental sequence data are 16S amplicon, so gene content is inferred from lineage rather than measured, yet inferred profiles are routinely read as measurements. Whether the reliability of inference differs between genes, and whether the predictor’s own confidence metric identifies the unreliable ones, has not been tested at archive scale. Here, 3,041 archived 16S amplicon runs from 78 studies of rice-paddy soil, arsenic-contaminated soil and arsenic-affected aquifers were re-analysed through one pipeline. Gene content was inferred by a curated marker-taxa bridge and by PICRUSt2, their agreement was quantified per gene and tested against candidate predictors, and both routes were checked against paired shotgun metagenomes. Compartments were compared under depth adjustment and study-level permutation. Results. Agreement between the routes varied widely by gene, from ρ = 0.906 (mcrA) to ρ = −0.118 (arsM), around a weak median of ρ = 0.286 across 698 gene-study comparisons in 57 studies. PICRUSt2’s confidence metric did not explain that range, whereas agreement tracked how consistently the marker taxa carried each gene in PICRUSt2’s reference genomes (ρ = +0.66). FAPROTAX, a published lookup, matched PICRUSt2 no better than the bridge, whereas Tax4Fun2, a second genome-based predictor, matched it more closely (ρ = +0.43 against ρ = +0.29). Sequencing depth lowered inferred abundance in 15 of 20 genes, while its effect on detection was positive for every gene. Against shotgun metagenomes from 267 biosamples in 16 studies, PICRUSt2 tracked measured abundance weakly (median ρ = +0.21), Tax4Fun2 no better, and the bridge barely (ρ = +0.05). Agreement between the bridge and PICRUSt2 predicted the accuracy of the bridge but not of PICRUSt2. No compartment difference was resolved after adjustment. Conclusions. Agreement between amplicon-based inference routes differs by gene within a single panel. Neither the predictor’s confidence metric nor phylogenetic conservation identifies the discordant genes, while carriage within the marker taxa partly anticipates them. Disagreement between independent routes flags method-dependent genes, but agreement does not certify accuracy. The routes disagree on arsM, one of only two arsenic-cycling genes both score. Compartment differences in arsenic-cycling potential cannot be resolved from public amplicon data without shotgun metagenomes.