The Antibiotic Dosage of Fastest Resistance Evolution: Gene Amplifications Underpinning the Inverted-U

Abstract To determine the dosage at which antibiotic resistance evolution is most rapid, we treated Escherichia coli in vitro, deploying the antibiotic erythromycin at dosages ranging from zero to high. Adaptation was fastest just below erythromycin’s minimal inhibitory concentration (MIC) and genotype-phenotype correlations determined from whole genome sequencing revealed the molecular basis: simultaneous selection for copy number variation in three resistance mechanisms which exhibited an “inverted-U” pattern of dose-dependence, as did several insertion sequences and an integron. Many genes did not conform to this pattern, however, reflecting changes in selection as dose increased: putative media adaptation polymorphisms at zero antibiotic dosage gave way to drug target (ribosomal RNA operon) amplification at mid dosages whereas prophage-mediated drug efflux amplifications dominated at the highest dosages. All treatments exhibited E. coli increases in the copy number of efflux operons acrAB and emrE at rates that correlated with increases in population density. For strains where the inverted-U was no longer observed following the genetic manipulation of acrAB, it could be recovered by prolonging the antibiotic treatment at subMIC dosages.

OD 600 is shown on the y-axis whilst the concentration of erythromycin is represented in a logarithmic scale on the x-axis.
The IC 99 and its 95% confidence intervals (n = 8) are indicated as horizontal bars. (IC x is the dosage at which the population density of a strain is reduced by x%.)

E. coli eTB108
Rate of Adaptation (h S6. Daily changes in bacterial density at di erent drug concentrations for 3 strains. Beginning on day two, each line in A and B represents the di erence in bacterial density (orange) or relative GFP fluorescence (green) between a given day and the preceding day's data. For strain eTB108, comparing data in A and B suggests a correlation between changes in bacterial density and e ux pumps per cell, where a proxy for the latter in B is the daily change in GFP units per OD signal. This correlation is a rmed and quantified in Figures S13 and S2. F S7. Dynamics of GFP per OD for eTB108 shown here are a proxy for mean AcrB-GFP per cell for 7 daily treatments. Note the within-season changes from lag to stationary phases at all of the antibiotic dosages used, note also the 'logistic shape' of each curve and the increasing AcrB-GFP levels through time each day. By the end of treatment only the very-highest dosages have not seen a rise in AcrB-GFP levels. The grey curve is the antibiotic-free control dataset for which AcrB-GFP levels remain similar, forming a logistic curve each day.   Raw Data Mean ± 95% CI MIC ± 95% CI Spline interpolant F S13. Population density rates of adaptation (defined using r e , Methods) peak below the MIC. Rates of adaptation (ROA) determined using r e (see Methods) vary with erythromycin concentration. A and B use population density (OD) data whereas C uses the mean relative abundance of AcrB (GFP per OD). The thick cyan line interpolates ROA data using a spine to guide the eye. The thick grey line is a baseline determined by propagation of the strains in liquid media without antibiotic and the dashed line denotes a ROA of zero. Significant di erences with respect to the baseline are green dots based on two-sided t-tests (mean ± s.e. with n = 8).
E. coli AG100 (using OD data) Erythromycin (µg/mL) Rate of Adaptation (h Raw Data Mean ± 95% CI MIC ± 95% CI Spline interpolant F S14. Rates of population density adaptation depend on drug concentration. Population density rates of adaptation (ROA) for strain AG100 measured using r auc and applied to OD data. Like the analysis of ROA that uses r e (Figure S13A) these vary with drug concentration and peak near the MIC.    F S17. Analogous to Figure S16 but for AG100. The MIC is around 30µg/ml erythromycin (Ery) but population growth is observed in some replicates up to 40µg/ml following 5 days of treatment.

C) whitened image (mask boundaries in red)
F S20. Cell size of eTB108 can decrease during erythromycin treatment. A) Anova to compare sizes of ancestral strain (eTB108) and its drug-treated descendant (eeTB108) isolated after the former was treated at sub-MIC erythromycin (Methods). Both were imaged following isolation in mid-exponential phase during which eeTB108 was exposed to erythromycin and eTB108 was not: the median size of eeTB108 is approximately 82% the area (in pixels) of demonstrates the GFP-AcrB in TB108 has highest density towards the membrane of E.coli. B) Fluorescence microscope image data taken before (eTB108 data) and after (eeTB108 data) erythromycin treatment show eTB108 can double the expression of AcrB-GFP when treated at sub-MIC levels (strains described in Methods). GFP levels were measured using fluorescence microscopy and were quantified using edge detection in Matlab where algorithms report mean green channel pixel intensity per 2d projected cell area of approximately 70 cells pre-and post-treatment (respective eTB108 and eeTB108 medians are 18.0 and 33.3, interquartile ranges [14.9,25.4] and [27.6,42.7] relative pixel units, see Figure   S20 for typical images). This represents a median change of circa 184%, corroborating analogous spectrophotometry data from the main text (e.g. the increases in GFP per OD reported in Figure 6A Figure 3D. B) The same information as A is shown but for non-inoculated microtitre plates and these also show a peak near 3/4h. We conclude that oscillations in GFP per OD data (also see Figure S7) do not have a biological basis but are a feature of the spectrophotometer device. No drug (± SE) 10µg/mL (± SE) 15µg/mL (± SE) 20µg/mL (± SE) 25µg/mL (± SE) 30µg/mL (± SE) 40µg/mL (± SE) F S23. Cumulative novel SNPs per genome in each treatment on days 1, 3 and 5. Most SNPs (above 5% frequency, see Methods) were observed in the absence of erythromycin where population densities were highest.

Contents of the 25Kb amplified region containing rrlB in Figure 4
Figure 4 highlights a 25Kb region (see the black band) containing the drug target rrlB that has increased coverage in populations exposed to erythromycin dosages around 10-20µg/ml. The rrlB datapoint in Figure 8A and also Figure 7A) confirm this region is amplified at rates that are commensurate with other rrl operons in the bacterial population. The only novel SNP detected in this 25Kb region resides in argE (acetylornithine deacetylase in the arginine biosynthetic process). The following genes are also amplified when rrlB is: (1) transfer RNA genes gltT, thrU, tyrU, glyT, thrT; (2) ribosomal genes: rplK, rplA, rplJ, rplL; (3) RNA-polymerase: rpoB (beta subunit), rpoC (beta prime subunit); (4) putative gene (yjaZ): adaptation to atypical conditions (heat shock protein); (5) sRNA sroH of unknown function; (6) cell wall (peptidoglycan) synthesis murB; (7) coA (coenzyme A), birA (biotin synthesis), thiH (thiamin synthesis), thiS (sulphur carrier). treatment for most of the genome but an amplified region is significantly above twice that by the end of treatment; these raw coverage data are for DNA data extracted after 120h of treatment at 30µg/ml erythromycin.

Supplementary Tables
The following 6 tables show novel SNPs detected above 5% frequency in their respective populations sampled from each of the di erent antibiotic treatments that are also not present in any of the drug-free control treatment replicates. The value of s j is a selection proxy determined from the frequency dynamics of each SNP (Methods), where the s-coe cient is determined from the logistic model (5) after fitting it to frequency data, j denotes replicate 1, 2 or 3.