Microbiology · surveillance data · a line between two doublings

The Gene Below the Line

A bacterium can carry a resistance gene and still test susceptible, and in NCBI's Salmonella records the gene is rarely idle. Of 847 isolates whose only known quinolone-resistance change is one gyrA mutation, 818 have ciprofloxacin MICs of 0.12 or 0.25 mg/L, against 0.03 mg/L or less for most isolates without one; CLSI's pre-2012 rule calls all 847 susceptible, its 2012 rule calls 832 intermediate, and FDA's surveillance convention counts 834 as decreased susceptibility. The records are a convenience sample, so no count here says how common resistance is: lab results for about 1.1% of the snapshot's 908,609 isolates, half of them from one FDA surveillance project and nearly all of those from meat, with 682 of the 847 in one NCBI SNP cluster.

A laboratory measures how much of a drug a bacterium can stand by growing it in a row of wells, each holding twice the concentration of the one before. The minimum inhibitory concentration, the MIC, is the lowest concentration in that row with no visible growth: EUCAST's reading rule is to "read MICs at the lowest concentration of the agent that completely inhibits visible growth", and the row itself is a convention, since "MIC dilution series are based on twofold dilutions up and down from 1 mg/L". So an MIC read from such a row is never a point. It is a step on a ladder, and between two wells the row measures nothing.

A word is then attached to the step. Susceptible, intermediate, resistant: a committee decides which steps get which word, and the committee can move. That is the whole subject of this page, because the same measurement, unchanged, can carry any of the three.

Below are the Salmonella isolates in one NCBI Pathogen Detection snapshot (PDG000000002.4281, built 1 November 2026) that have both a genome and a laboratory antibiogram: 9,976 of the snapshot's 908,609 isolates, about 1.1%. For each one this page ships the MIC rows of its BioSample antibiogram for twelve of the drugs tested, as submitted (for the two combination drugs, the first number of each pair), 115,704 of them, beside the resistance genes NCBI's own AMRFinderPlus found in its genome. Nothing here is a sample of Salmonella in the world. The isolates that get tested are the ones a surveillance programme chose to test: 5,037 of the 9,976 come from one FDA project, "CVM NARMS Salmonella", and 4,953 of those 5,037 name chicken, turkey, pork, beef or veal as the thing they were isolated from. And the 847 isolates at the centre of the argument below are not 847 independent events: 682 of them fall in one NCBI SNP cluster. No percentage on this page is a prevalence of resistance.

One gene, one drug, one line

Pick a resistance gene and the drug it is supposed to act on. The upper histogram is the isolates whose genomes carry that gene; the lower one is the isolates with no known gene for that drug at all. The axis is the dilution ladder. The line is where a committee put the word "resistant", and the counter says how many carriers that line calls susceptible; the laboratories' own words for the same carriers are listed beneath it.

Start with tet(A), a tetracycline efflux pump. Of the 2,902 isolates that carry it and were tested against tetracycline, 22 come out susceptible, fewer than one in a hundred. That is the ordinary case, and it was reported ten years ago. McDermott and colleagues reported in 2016 on 640 NARMS Salmonella isolates they had sequenced, and found that "resistance genotypes and phenotypes correlated in 99.0% of cases", with correlations that "approached 100% for most classes of antibiotics but were lower for aminoglycosides and beta-lactams". Nothing on this page overturns that. What the MICs add is where the disagreements sit.

Two pairs in that table are worth staring at. aac(3)-VIa and aac(3)-IVa are both gentamicin-modifying enzymes, and the catalogue maps both to gentamicin. The first behaves: 481 of 508 carriers resistant, 11 susceptible. The second does not: of 523 carriers, 147 are resistant, 295 are intermediate and 81 are susceptible. Look at the histogram and the reason is visible. The aac(3)-IVa carriers are not scattered among the wild type; they pile up at 4 and 8 mg/L, two steps and one step below the line at 16, where 295 of them land on the single word "intermediate". The same shape appears for blaCMY-2 and cefoxitin, where 65 carriers sit at exactly 16 mg/L, one step under a line drawn at 32. McDermott's paper described this for its own 21 cefoxitin disagreements: "In 20 of the 21 instances, MICs were a single dilution below the clinical breakpoint for cefoxitin resistance, resulting in intermediate phenotypes."

"So the gene did nothing"

Here is the sophisticated reading of all this. If a quarter of the carriers of some gene test susceptible, either the gene is silent in those isolates, or sequencing overcalls resistance, and in both cases the genome is the weaker instrument. It is a fair objection, and the data answer it, because the MIC says how far the gene moved the bug before the word was assigned.

Ciprofloxacin is the clearest case. A single amino-acid change in gyrA, a topoisomerase gene, raises the MIC a few steps without reaching the concentration a clinical laboratory calls resistant. Take the 847 isolates in this snapshot whose only known quinolone-resistance determinant of any kind is one gyrA point mutation: 766 carry D87Y, 28 S83F, 24 D87N, 15 D87G, 12 S83Y and 2 D82N. Their ciprofloxacin MICs are not wild type. 818 of the 847 read 0.12 or 0.25 mg/L, while 8,549 of the 8,756 isolates with no known quinolone determinant read 0.03 or below. The gene did something: the commonest reading among the 847, 0.25 mg/L, is at least four doublings above the commonest reading without one, which is "≤0.015". Whether that counts as "resistance" is a separate question, and three documented rules, from two bodies, answer it three different ways.

Drag the line and watch the same measurements change their names. The instrument opens on those 847, and at each documented line it reads them three ways. The line can only sit between two dilutions, because a doubling panel measures nothing in between.

One set of measurements, three words

Those 847 single-mutation isolates have one set of MICs. Read them under the rule CLSI gave all Enterobacteriaceae before 2012, susceptible at 1 mg/L or below, and every one of the 847 is susceptible. Read them under the rule CLSI introduced in January 2012 for extraintestinal Salmonella, susceptible at 0.06 and below, resistant at 1 and above (applied here to every isolate, as CDC's Medalla and colleagues applied it to all of theirs), and 832 become intermediate, 2 resistant, 8 susceptible, with 5 censored readings the rule cannot place. Read them under the convention FDA's NARMS table uses, where 0.12 and above is decreased susceptibility, a category that table lists under both intermediate and resistant, and 834 fall in it.

The pre-2012 numbers are a two-source chain, because CLSI's standard, M100, was not read for this page (CLSI sells it, and also offers a free read-only view on its MicroFree portal, behind a sign-in). Sjölund-Karlsson and colleagues, writing at CDC in 2014, record that "Prior to 2012, the Clinical and Laboratory Standards Institute (CLSI) provided MIC and disk diffusion breakpoints for ciprofloxacin, levofloxacin, and ofloxacin that were common for all Enterobacteriaceae", and that the 2012 change "lowered the resistance breakpoint for ciprofloxacin and extraintestinal Salmonella from ≥4 μg/ml to ≥1.0 μg/ml". A Wisconsin State Laboratory of Hygiene slide from 2019 prints the previous Enterobacteriaceae broth-microdilution values as ≤1, 2, ≥4; reading its susceptible limit of 1 as the pre-2012 one assumes that the Enterobacteriaceae values did not change between 2012 and 2019, which neither source says. The 2012 Salmonella values are in the same CDC paper: "new Salmonella-specific breakpoints were implemented (susceptible [S], ≤0.064; intermediate [I], 0.12 to 0.5; and resistant [R] ≥1.0 μg/ml)". Medalla and colleagues, also at CDC, date the change: "In January 2012, CLSI established new ciprofloxacin MIC interpretive criteria for extraintestinal Salmonella spp." They were unchanged in 2024: a slide from the same Wisconsin series prints the Salmonella ciprofloxacin row of CLSI M100-Ed34 as ≤0.06, 0.12 to 0.5, ≥1. And the third line is FDA's own: "We now use decreased susceptibility to ciprofloxacin (DSC, MIC >= 0.12 ug/ml) as a marker for emerging fluoroquinolone resistance".

What the 2012 change was for is in the same paper: with the old rule, first-step mutants "would have remained undetected and would have been classified along with fully susceptible isolates free of QRDR mutations and/or PMQR mechanisms". The MICs here are what that sentence looks like as a histogram. The authors also measured the gap directly in 100 isolates: "The ciprofloxacin MICs of the isolates lacking a resistance mechanism ranged from ≤0.016 to 0.064 μg/ml, whereas the ciprofloxacin MICs for the 40 isolates with topoisomerase mutations ranged from 0.125 to ≥16 μg/ml."

Where would you put the line?

If a line can be moved, it can be chosen, and the genotype gives one way to choose: put the line where the fewest isolates disagree with their own genomes. Count two kinds of mismatch, a carrier the line does not call resistant and a non-carrier it does, and sweep the line along the ladder. Switch the instrument above to streptomycin and the curve has a floor.

For the 4,857 isolates collected in 2014 or later and tested against streptomycin, 2,067 carry a gene the catalogue maps to streptomycin and 2,726 carry none. Resistant at 8 mg/L or above leaves 1,890 disagreements, at 16 leaves 591, at 32 leaves 192, at 64 leaves 594, at 128 leaves 1,306. The minimum sits at 32, which is where NARMS moved its line, for these isolates, in this snapshot. Their stated reason is in the 2014 NARMS report:

In 2014, the testing range for streptomycin was expanded in order to determine the full range of MIC values. These data, along with WGS information, were used to lower the resistance breakpoint from ≥ 64 µg/ml to ≥ 32 µg/ml. Because the historical testing range only included two streptomycin concentrations of 32 µg/ml and 64 µg/ml, this new breakpoint will be applied to all testing beginning with the 2014 isolates.

NARMS Integrated Report: 2014, FDA, p. 3

Read that quotation twice, because it disarms the result it seems to confirm. The line was lowered using genotype data: Tyson and colleagues sequenced 463 Salmonella and E. coli isolates and found that "using a streptomycin resistance breakpoint of ≥ 64 μg mL(-1) classified over 20% of strains possessing aadA or strA/strB resistance genes as wild-type", and proposed 32 "to improve the concordance between genotypic and phenotypic data". The agreement runs the other way too: NCBI validated AMRFinder against NARMS phenotypes, measuring "the consistency between predicted AMR genotypes from AMRFinder and resistance phenotypes of 6,242 isolates from the National Antimicrobial Resistance Monitoring System (NARMS)". So the floor of this curve landing on 32 is not independent confirmation that 32 is right. It is the same method, run again on more isolates, agreeing with itself. What it does show is that the choice was not arbitrary, and that it is the kind of choice a histogram can argue about.

Now switch to the isolates collected before 2014, and most of the curve goes flat and grey. The sentence above explains why: the old panel held only two streptomycin wells, 32 and 64 mg/L. 2,499 of those 3,878 isolates are recorded as "≤32", which means somewhere at or below the bottom of the panel, and a line at 32 cannot decide a single one of them. The instrument marks 2,417 of them undecidable there and refuses to guess (the other 82 belong to 8 isolates whose only streptomycin hit is partial and 74 with an empty genotype field, which the sweep leaves out). This is the same limit McDermott's team ran into: "Because of the design of the NARMS susceptibility testing panel, we could not evaluate the concordance between genotype and MIC (a high proportion of isolates were at the extremes of the dilution range)." A measurement that was never made cannot be reinterpreted later, and that is why NARMS applied its new streptomycin line only to isolates from 2014 on. Where the old readings do reach a new line, they can be read again: for ciprofloxacin, Medalla and colleagues applied the 2012 rule to isolates from 1996 to 2009, writing "We used the new criteria for all isolates".

The word is a convention, and the records show it

Every reading here carries the laboratory's own word beside the number, and all but 21 of the 115,704 name the standard the laboratory says it followed. So take one measurement, ciprofloxacin at exactly 0.25 mg/L, and read the words.

One concentration, three words, no disagreement about the measurement. All 572 resistant calls under the label "CLSI" were submitted by FDA, and they match FDA's surveillance convention, which puts 0.12 and above in both its intermediate and its resistant column; the 37 intermediate calls are CLSI's clinical rule read literally. Read that way, nobody mislabelled anything. The word is downstream of a choice, and the records disagree because the choices do.

The dilution labels carry the same lesson in miniature. EUCAST writes the 2^-4 dilution as 0.06 and sets its Salmonella ciprofloxacin breakpoint at susceptible 0.06 or below: "EUCAST has decided to use the following format (in bold): 0.125→0.125, 0.0625→0.06, 0.03125→0.03". In this snapshot 70 records write that dilution as "0.06" under the label CLSI and call it susceptible, while 27 write it as "0.0625" under the label EUCAST and call it resistant. All 27 come from one submitter, SENASICA in Mexico, in 2019. By EUCAST's own notation those 27 sit on the susceptible side of EUCAST's own line, in its 2019 tables (version 9.0) as in its current ones. Nothing in the records says why, and this page does not guess.

Why the genotype's line is not the right line

It would be easy to end with the moral that committees are behind the data. Gentamicin says otherwise. Sweep the line for the 1,368 carriers and 8,416 non-carriers of a gentamicin determinant and the minimum is at 4 mg/L, with 87 disagreements. The resistant line FDA's 2021 table takes from CLSI is at 16, which leaves 507; CLSI's 2023 revision for the Enterobacterales, M100-Ed33, moved it to 8, which leaves 168. Two doublings apart, then one, and neither line is wrong, because a clinical breakpoint answers a different question. The genotype's line asks whether the organism carries a mechanism. A clinical line asks whether a drug at an achievable dose is likely to work in a patient, and CLSI redrew this one in 2023 after a review of pharmacokinetic and pharmacodynamic data. For Salmonella that question has a blunter answer still: the CLSI newsletter announcing the change says that "Although aminoglycosides may appear active against Salmonella and Shigella in vitro, they are ineffective against these genera clinically", so a gentamicin line in a Salmonella surveillance table is the Enterobacterales line, borrowed. The CDC authors of the ciprofloxacin paper put the general point in one sentence:

The primary purpose of clinical breakpoints is to predict clinical outcomes and not to detect resistance mechanisms.

Sjölund-Karlsson, Howie, Crump and Whichard, Journal of Clinical Microbiology 52(3):877–884, 2014, Discussion

Microbiology already has a name for the other line. An epidemiological cutoff value, as a Wisconsin State Laboratory of Hygiene slide summarises CLSI's definition, gives "MIC values that separate bacterial populations into those that have (don't have) resistance mechanisms strictly on basis of phenotype", and is "Strictly in vitro; no clinical correlation". The floor of the curve in this instrument is an argument about an epidemiological cutoff, never about treatment. This page has nothing to say about how any infection should be treated.

So the honest summary of all of it: the genome is a good predictor of the laboratory's word (99.0% of the time in McDermott's 640 isolates), and when the two disagree here it is usually because the gene lifted the MIC a few doublings and stopped below a line drawn for another purpose. The gene did something. The line could not see it, because the line was not built to.

curl -sO https://artwaste.land/checks/verify-the-gene-below-the-line.mjs
node verify-the-gene-below-the-line.mjs