Why do submitted antimicrobial susceptibility letters disagree with their own MICs? A preregistered cause-attribution audit of NCBI BioSample antibiograms for Klebsiella pneumoniae and Escherichia coli

Objectives: Susceptibility letters (S, I, R) submitted alongside public bacterial genomes serve as training labels for genome-based resistance prediction, yet many conflict with their own MICs. We investigated how much of this conflict earlier breakpoint editions explain, and whether the unexplained remainder is concentrated among a few submitters. Methods: Preregistered audit of NCBI BioSample antibiograms for Klebsiella pneumoniae and Escherichia coli. Every row with an MIC and a letter was classified against each CLSI (2010–2026) and EUCAST (v1.0–v16.1) edition of its recorded standard within a date window. Shares were averaged across submitting groups with t-based 95% CIs. A BioProject with transposed S/R letters served as a positive control. Results: The positive control placed 99.7% of classifiable rows in the unexplained class. Among 17 organism–drug pairs with at least 10 submitting groups, 74%–97% of letters agreed with the current edition and 0.1%–9.4% (median 2.7%) matched no edition. Unexplained letters clustered by submitter (latent-scale ICC 0.54, 95% CI 0.36–0.90; 3 of 54 groups accounted for 53%), as did downgrades (72 rows; 69% in three groups). Within a submitter-year, letters followed one side of each identifiable CLSI revision in 94% (95% CI 87%–100%) of units, versus 3% for a fictitious revision. Older editions explained letters beyond chance in 2 of 9 testable pairs. Excluding censored MICs raised unexplained shares by up to 16 percentage points. Conclusions: Among classifiable letters, most disagreement stemmed from an older breakpoint edition. The unexplained remainder is concentrated in a few submitters, supporting submitter-level screening of public susceptibility labels. This is the author's original version, submitted to the Journal of Antimicrobial Chemotherapy. It has not been peer reviewed. Note added 19 October 2026: after this version was submitted, the submitters of BioProject PRJNA1322038 had NCBI correct its S/R interpretations (13 October 2026). This study analysed the records as downloaded, so its results are unchanged.