The investigators began from a simple, unsettling observation: most low-biomass disease–microbe associations measure biomass differences rather than disease biology. They reported that identical input communities produced genus differences that span 98% of published effect sizes. Patients and controls, the team observed, can differ in biomass by as much as 153-fold — a scale that makes contamination a dominant signal unless explicitly modelled. The authors used dilution experiments to quantify this: fixed samples diluted 64,000-fold moved 82% of the way toward their extraction blanks, an empirical anchor the study team used to set expectations for contaminant impact.
Short statement: contamination can mimic biological effects. Longer context: when background reads are large relative to true load, genus-level shifts commonly reported in the literature can be driven by reagent or process noise rather than site-specific pathogens.
The study assessed background-to-load ratios across 12 specimen types. Data indicated that contamination could erase a majority of reproduced directional claims: 61–76% of 354 reproduced directions were lost when background signals were considered. By contrast, 27 directions from high-biomass brush and oral samples were unaffected. This contrast points to a practical rule — when background-to-load ratios are high, observed signals are vulnerable. The authors noted that own-blank predictions (predictions based on a sample’s matching blank) matched published directional claims more often than permuted-label selections — 48.6% versus 39.3% as the source reports — which the team used to argue that many published directions follow contaminant signatures.
The authors reviewed 6,175 claims drawn from 175 studies spanning 29 sites. Of 3,813 claims that the team judged decidable under their framework, 3,487 (91%) failed; 1,664 failed even before any contamination was permitted in their tests. Recurrent, significant claims that had undergone decontamination still failed more often: 169 of 182 such claims (93%) did not hold under the authors’ criteria. Where differences persisted, the retained signals were enriched for the site’s known pathogens.
The team traced downstream effects. Failed differences were not confined to a single paper: 655 of 703 subsequent papers used those failed premises, and the failed signals extended into higher-level clinical outputs. The source reports that one patient stratification reversed significance (P changed from 0.025 to 0.75) after applying the authors’ approach; the manuscript also notes a failed signal reached a practice guideline and a registered clinical study. The authors used these examples to illustrate how contamination-driven claims can propagate into clinical decisions.
The centerpiece of the manuscript is TRUEDA, the authors’ tolerance-based decontamination method. Instead of attempting to remove reads labelled as contaminants, TRUEDA quantifies the largest contaminant share at which each genus-level difference still holds, testing every plausible contaminant composition. The authors explain why simple read removal can fail: a genus can act both as a contaminant and as a true pathogen in different contexts. TRUEDA returns a per-effect tolerance metric that communicates how robust a directional claim is to contamination.
Short technical note: the source describes TRUEDA’s operational logic and contrasts it with deletion-based strategies, but specific algorithmic parameters, implementation code paths, or step-by-step procedures are detailed in the manuscript and supplementary material; practitioners should consult those files for implementation details.
On SCRuB’s Minich benchmark reconstructions, TRUEDA retained all 17 true input-identity contrasts that were present uncorrected and withdrew 33 of 35 false positives, producing a false-positive rate of 4% where SCRuB’s reported false-positive rate was 55%. The source also reports TRUEDA ranked first on five additional benchmarks. In four independent urine cohorts totaling 195 patients, TRUEDA produced an external AUROC of 0.592; comparator methods ranged from 0.49 to 0.56 in those same cohorts, per the manuscript. The authors further report that TRUEDA’s candidate genera were known pathogens 2.8-fold as often as those nominated by ALDEx3.
The authors emphasise that TRUEDA’s tolerance output provides richer interpretive context than a single corrected p-value: it tells clinicians and researchers how much contamination a claimed difference can absorb before it disappears.
The manuscript includes an absolute-scale re-analysis of bladder-cancer urine directions. According to the source, TRUEDA fixed 193 of 240 directions on the absolute scale and reversed 37 directions. Where relative correction methods yielded four directions indicating higher abundance in cancer, TRUEDA replaced those with 30 directions distributed across 19 genera. The authors use this example to show how tolerance-informed correction can materially change reported disease-associated genera and their clinical interpretation.
This work is a preprint and has not been peer reviewed, as the source explicitly states. The authors declare no competing interests. The manuscript provides supplementary material; specific computational parameters, code availability and full implementation details are described there and should be consulted for replication. The study flags several unresolved clinical questions: which reported associations should guide diagnostic decisions, how to standardise tolerance thresholds across specimen types, and when a tolerance value is sufficient for clinical action. The authors propose reporting a per-effect tolerance alongside each claimed association, but the source does not prescribe a single universal cutoff for clinical use. These open points — and the preprint status — are the clear next steps before clinical adoption can be considered.