Answers are grounded in this paper and cite their sources.
Soil antibiotic-resistome studies increasingly rely on computational annotation of metagenomic data, yet the sensitivity of published ecological conclusions—global patterns, temporal trends, risk rankings, and soil–human connectivity—to the choice of annotation pipeline (database × classifier × threshold × quantification convention) has never been systematically measured. This audit preregisters a measurement-science study of two flagship studies (Zheng et al. 2022, Science Advances; Zhao et al. 2025, Nature Communications), using a six-ruler matrix (ARGs-OAP, DIAMOND/CARD, RGI, Resfams, FunGeneTyper, LARGE) plus one calibrated correction ruler with an abstention mechanism. Pre-registered endpoints: (E1) Rank-I risk ranking, (E2) ecological structure, (E3) environmental drivers, (E4) habitat connectivity. For each endpoint we froze operational definitions of "conclusion flip" (direction change with 95% CI excluding zero, reproduced in two independent data scenarios) before analysis. A frozen sentinel set (726 ResFinderFG holdout + 297 functionally verified records) and a failure-mode atlas of 136 curated false reports lock the evaluation layer; sentinel sequences never enter training or threshold selection. Hypotheses: H1 (at least one endpoint flips reproducibly), H2 (calibrated measurement with unknown imputation reduces pipeline sensitivity), H3 (failure-mode-locked hard negatives reduce false reports at matched low-homology recall; query–reference residual features provide independent external gain). Falsification conditions, SESOIs, cluster-bootstrap CIs, multiple-testing control, leakage isolation rules, and deviation policies are specified in the attached protocol (G0 v1.0, frozen 2026-11-16, SHA-256 recorded in-document).
No takes yet. Share an insight, caveat, or question.
Lei Zhou (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: