Systematic mapping of antibiotic cross-resistance and collateral sensitivity with chemical genetics

Abstract

By acquiring or evolving resistance to one antibiotic, bacteria can become cross-resistant to a second antibiotic, which further limits therapeutic choices. In the opposite scenario, initial resistance leads to collateral sensitivity to a second antibiotic, which can inform cycling or combinatorial treatments. Despite their clinical relevance, our knowledge of both interactions is limited. We used published chemical genetics data of the Escherichia coli single-gene deletion library in 40 antibiotics and devised a metric that discriminates between known cross-resistance and collateral-sensitivity antibiotic interactions. Thereby we inferred 404 cases of cross-resistance and 267 of collateral-sensitivity, expanding the number of known interactions by over threefold. We further validated 64/70 inferred interactions using experimental evolution. By identifying mutants driving these interactions in chemical genetics, we demonstrated that a drug pair can exhibit both interactions depending on the resistance mechanism. Finally, we applied collateral-sensitive drug pairs in combination to reduce antibiotic-resistance development in vitro.

Fig. 1. Chemical genetics allow for systematic XR and CS assessment.

a, Schematic illustration of the conventional way XR and CS drug interactions are assessed by experimental evolution. Resistant mutants selected by drug 1 are tested for susceptibility to drug 2. The MIC, or 90% inhibitory concentration (IC₉₀), of drug 2 is compared with that of the ancestral strain. b, Schematic illustration of chemical genetic screens with arrayed libraries. Several drugs (drug 1, 2 and so on) are profiled across genome-wide gain-of-function or loss-of-function mutant libraries. The fitness of each mutant is evaluated independently—for example, by measuring colony size. c, XR and CS are associated with chemical genetics profile similarity and dissimilarity, respectively. The s-score (used as a proxy for fitness) of each deletion mutant is plotted for two drugs involved in either XR or CS. If the same mutations make cells more resistant or sensitive to two drugs, cells are more likely to evolve mechanisms that inhibit or promote these exact processes during evolution and become XR to both drugs, whereas the opposite is true for CS.

Fig. 2. Chemical genetics-derived metric separates well-known XR and CS interactions and infers new ones.

a, The overlap between published XR and CS interactions from four existing datasets^,– is low, even when directionality is not taken into account. b, A devised metric derived from chemical genetics profile similarity, OCDM, can robustly discern between known XR, CS and neutral interactions. False detection rate-adjusted P values were obtained from a two-sided Mann–Whitney U-test. The box boundaries represent the first and third quartiles, with the median indicated. The whiskers extend to the furthest data points within 1.5× the interquartile range. c, ROC curves for the classification of XR (positive class) versus non-XR (negative class) and CS (positive class) versus non-CS (negative class). Each OCDM cutoff represents a point on the curve and is associated with a true-positive rate and a false-positive rate. The OCDM cutoffs chosen for XR and CS interactions are depicted with a closed circle. d, New XR, neutral and CS pairs inferred by chemical genetics using the OCDM cutoff expand the currently known XR and CS interactions in E. coli by two- and fourfold, respectively. This difference further increases if we take into account drug pairs for which the interaction is inferred differently from previous studies (Extended Data Fig. 2). Note that known interactions (n = 420 total) include drug pairs for which there is no available chemical genetics data. Source data

Fig. 3. Inferred XR and CS interactions are validated with high accuracy by experimental evolution.

a, Schematic of benchmarking conducted for 70 drug pairs by experimental evolution and IC₉₀ measurements. Twelve lineages were evolved in parallel for five passages in increasing concentrations of 23 antibiotics. At each passage, the culture growing at the highest concentration was transferred to a new antibiotic gradient. The IC₉₀ of the final resistant population was then measured for all lineages in the relevant antibiotics. b–d, Heatmaps of 70 new and known drug-pair interactions, split depending on whether they were inferred as CS (b), neutral (c) or XR (d). Interactions were tested in both directions, with the drug for which selection occurred shown first and the drug for which MIC/IC₉₀ was tested shown second. In each interaction, all tested lineages are shown (n = 9–12). Coloured boxes denote the interaction observed for a given lineage. The three columns on the right of the lineage results represent the summary for all lineages. We considered an interaction as validated if the log₂-transformed IC₉₀ fold change was >1 for XR and <−1 for CS in any direction tested for at least one lineage compared with the wild type. An interaction of a drug pair was deemed to be XR if there was at least one lineage showing XR despite any CS for other lineages. Interaction monochromaticity (that is, whether the interaction is exclusively CS or XR; neutral lineages do not affect this call and were labelled as not applicable (N/A)) and directionality (drug pair interacting consistently in both directions) are shown. Interactions referred to as reclassified in the text are those for which our inference and validation agree but previous reports have reported differently. The interaction in red (least monochromatic interaction) is used in Fig. 5 to understand the mechanisms in play. The interactions in bold are used later in Fig. 6 to test resistance evolution in drug combinations. The interaction in italics (drug pair 14), which was conflicting across studies (XR in one study and CS in another), has been inferred and validated to be CS. Source data

Fig. 4. CS and XR interactions between and within antibiotic classes.

a, Interactions between members of the same antibiotic class (within class) are exclusively inferred as XR. The within-class group includes classes with more than one member probed—that is, β-lactams, aminoglycosides, quinolones, macrolides, tetracyclines and sulfonamides. b, Overview of all inferred and known drug interactions in E. coli at the class level. When a class has only one representative the antibiotic is named and shown in grey. The heat map sums XR and CS interactions across drug classes inferred by the OCDM metric. Within-class interactions are not displayed in the plot but are all exclusively classified as XR. Antibiotics are grouped according to their modes of action. Dot size represents the count of interactions between classes (or single antibiotics). c, Coherency of interactions of each class with all other classes—that is, if all members of the class interact the same with other classes—calculated as the sum of the absolute differences between the number of XR and the number of CS interactions with each other class normalized to the number of drugs in the class. The higher the number, the more coherently the class is behaving. d, Interaction preference of each class (single- or multi-membered), calculated as the log₂-transformed ratio of the number of CS and XR interactions with all other antibiotics from other classes. Antibiotic classes with a ratio of >0 are considered predominantly CS (n = 8), whereas those with a ratio of <0 as predominantly XR (n = 12). Antibiotic classes in bold are classes with more than one antibiotic tested. Source data

Fig. 5. Chemical genetics recapitulate the dynamics and explain the mechanisms of non-monochromatic interactions.

a, Changes in azithromycin susceptibility during the evolution of 12 lineages in tetracycline (100 generations, Methods). Resistance levels of 12 lineages to both antibiotics are shown for days 2, 3, 5, 7 and 10. Lineages are grouped according to whether they exhibited CS, neutrality or XR on day 5 (same as Fig. 3d). Dashed lines indicate the neutral threshold. b, Chemical genetic profiles of the E. coli deletion library in tetracycline and azithromycin. Mutants with concordant (XR-related) and discordant (CS-related) profiles are highlighted. Dots in grey represent mutants that do not have s-scores within the 3% extreme cutoff for both drugs. Lines at x = 0 and y = 0 are shown to separate concordant and discordant zones of the plot. c, Mutations of lineage 11 during evolution. Genome sequencing of the lineage population reveals a succession of two point mutations in genes that both lead to CS—first in hldE, which is then replaced by mutations in waaF, a slightly less detrimental gene for azithromycin resistance according to chemical genetics data in b. For the other 11 lineages see Extended Data Fig. 5. d, The fold change in tetracycline and azithromycin IC₉₀ of knockout mutants compared with the wild type confirms that both hldE and waaF contribute to resistance to tetracycline and sensitivity to azithromycin, whereas ompF deletion leads only to resistance to tetracycline; n = 6 biological replicates. e, Tetracycline uptake is reduced in a waaF deletion (ΔwaaF) mutant. Tetracycline fluorescence was measured in cell pellets and the signal was normalized to the optical density at 600 nm (OD_600nm); n = 3–6 biological replicates. d,e, Data are the mean ± s.e.m. f, OmpF, a major tetracycline importer, is the most downregulated protein in ΔwaaF. FC, fold change. Source data

Fig. 6. Combinations of reciprocal CS antibiotic pairs reduce resistance evolution.

a, Experimental design. After evolving resistance to single antibiotics or their combination (seven lineages for each, passaged every 24 h for 7 d; 70 generations in total), the IC₉₀ of both antibiotics was determined for the evolved mutants. In each passage mutants growing (coloured yellow) at the highest concentration (well marked by a thick circle) were transferred (Methods). b, The measured IC₉₀ values were used to calculate the evolvability index (equation (2), Methods; data using slightly different original evolvability indices (equation (3), Methods) in Extended Data Fig. 6a). The red line represents the cutoff (log₂(evolvability index) = 0; the evolvability index was log₂-transformed to make data symmetrical) below which the antibiotic pair is considered to reduce resistance evolution compared with single antibiotics. Red dots on the violin plots represent the median. The box boundaries represent the first and third quartiles, with the median indicated. The whiskers extend to the furthest data points within 1.5× the interquartile range. Non-XR antibiotic combinations led to lower collective resistance and in the case of reciprocal CS to lower evolvability indices and lower resistance to each of the antibiotics combined (Extended Data Fig. 6b). Source data

Extended Data Fig. 1. Performance of different metrics and models in capturing XR and CS antibiotic interactions from chemical genetics data.

a, Receiver operating characteristic (ROC) curves for classification of XR (positive class) vs non-XR (negative class), and CS (positive class) vs non-CS (negative class), using simple linear and non-linear correlation metrics. AUC is the area under the curve. b, The performance of the decision tree model on balanced classes shows that both XR and CS interactions can be well classified. c, Decision tree with classes CS (class 1) versus the rest (class 0), where a maximum depth of 3 is shown for visualization, illustrates the hierarchy of decisions to discriminate classes. Each node in the tree represents a decision point based on the value of a particular feature, and branches represent the outcome of the decision. The root node divides the data based on the ‘concordant_negative_w’ feature, which is the sum of s-scores (as weights) of hits on the negative concordant site of a scatterplot. The tree branches out to ‘discordant_w’ feature, which is the sum of s-scores (as weights) of hits on the discordant site of a scatterplot, while ‘discordant_w_m’ is the sum of products of s-scores (as weights) of hits on the discordant site of a scatterplot. d, P values from a paired Mann–Whitney U-test (two-sided) are depicted across quantile cutoffs for extreme s-scores to differentiate XR/CS/neutral interactions based on OCDM values. Q3 and Q97 perform the best. e, Confusion matrix of results based on Q3 and Q97. Most interactions inferred as non-XR/non-CS were previously reported neutral. For more information, see also Extended Data Fig. 2a-c. Source data

Extended Data Fig. 2. Chemical genetics metric captures well prior information and can be used to reclassify a subset of prior interactions.

a,b, Comparison of previously reported XR (a) and CS (b) interactions with our inferences based on our chemical genetics metric (OCDM) show an agreement of 67–68% for CS (n=17) and XR (n=47) - 10 such interactions were validated experimentally during our benchmarking (Fig. 3b,d). The rest is inferred as neutral or the opposite interaction by OCDM, including seven interactions (4 CS, 2 neutral & 1 XR) that we experimentally validated that OCDM inference was correct (Fig. 3b,d). c, In contrast to CS or XS, there is less agreement for neutral interactions with previous studies. This is consistent with the high false negative rates when comparing prior studies between them (Fig. 2a). The majority of previously reported neutral interactions (76.6%, n=85) are inferred as CS/XR by chemical genetics. 11/13 we included in the benchmarking set were confirmed as inferred by OCDM. The other two were inferred CS, but although most lineages exhibited CS, a single lineage exhibited XR, and hence called XR (Fig. 3b–d). d, New XR, neutral, and CS pairs inferred by chemical genetics and the OCDM cutoff are 2.8- and 6.4-fold more than currently known XR and CS antibiotic interactions in E. coli, after reclassifying interactions (n = 116) we infer differently than previously reported. The plot includes known interactions for which chemical genetics data is not available. e, Resistance against 12 antibiotics was evolved again for up to ~100 generations in 12 lineages. The MIC of evolved populations was measured at ~50 and ~100 generations for the same lineages in different antibiotics, allowing us to assess XR/CS for 17 drug pairs in both directions. Data are represented and drug pairs are numbered as in Fig. 3b,d. All inferred interactions were validated at both ~50 and ~100 generations. The length of experimental evolution affected the XR/CS of individual lineages and to a lower degree the cumulative call of the drug pair. Source data

Extended Data Fig. 3. Chemical genetics can uncover the biological processes that drive interactions between antibiotic classes.

a, Clustered heatmap of discordant mutants that are part of CS interactions between sulfonamides and macrolides (blue) or beta-lactams (green). Genes in bold are involved in LPS or nucleotide biosynthesis. b, Clustered heatmap of concordant mutants that are part of XR interactions between tetracyclines (violet), macrolides (blue), and other protein synthesis inhibitors. Genes in bold regulate or are part of the major efflux pump in E. coli (AcrAB-TolC). Source data

Extended Data Fig. 4. Chemical genetics can infer mechanisms and monochromaticity of XR and CS drug interactions.

a, Scatter plot of chemical genetic profiles of the E. coli deletion library in tobramycin and nalidixic acid. Mutants with concordant (XR-related) and discordant (CS-related) profiles are highlighted. Dots in grey represent mutants that do not have s-scores within the 3% extreme values for both drugs. The underlined knockout mutants are known causal genes of this CS interaction^,. b, Chemical genetic profiles for novobiocin and cefoxitin, presented as in a. Underlined knockout mutants indicate that the changes in polarity of the lipopolysaccharide (LPS) core can drive resistance to cefoxitin while providing sensitivity to the large and non-polar novobiocin. c, Non-monochromatic XR interactions (n=11) have higher absolute discordance scores than their monochromatic counterparts (n=27) (two-sided Mann–Whitney U-test; P = 3.758e-07) - monochromaticity was defined in the validation experiment. This means that chemical genetics can infer the monochromaticity of XR interactions. The box boundaries represent the first and third quartiles, with the median indicated. The whiskers extend to the furthest data points within 1.5 times the interquartile range (IQR). d, The highest discordance score of -133.8481 based on the 11 non-monochromatic XR interactions from c was used to separate the remaining inferred XR interactions (excluding the 38 validated) into monochromatic (n=225) or non-monochromatic (n=168). e–g. Scatter plots of chemical genetic profiles of the E. coli deletion library for examples of other pairs of drugs with both high concordance and discordance (in addition to azithromycin and tetracycline shown in Fig. 5b). As the azithromycin-tetracycline pair, those are expected to be non-monochromatic. Data are depicted as in a,b. For all data, see the relevant Shiny app at https://shiny-portal.embl.de/shinyapps/app/21_xrcs. Source data

Extended Data Fig. 5. Genome sequencing of lineage populations evolved in tetracycline.

Results of the remaining 11 lineages from days 3, 5, and 7. Results are shown as in Fig. 5c, and lineages grouped in XR, CS, and neutral according to classification in Fig. 5a.

Extended Data Fig. 6. CS antibiotic combinations constrain resistance evolution to one or both compounds.

a, log₂ transformed evolvability Index as originally proposed in the literature (equation (3), Methods) confirms observations made using the slightly modified Evolvability Index (equation (2), Methods) – Fig. 6b. Red dots on the violin plots represent the median. The box boundaries represent the first and third quartiles, with the median indicated. The whiskers extend to the furthest data points within 1.5 times the interquartile range (IQR). b, The log₂ of MIC (IC90) of the evolved population in both drugs compared evolved population of the drug itself is used to identify whether and how well combining drugs reduces resistance to each drug compared to single-drug treatments. Reciprocal CS drug pairs do this efficiently. The red dashed line shows the no-effect when combining drugs does not change resistance evolution to single drug treatments. The bars in the violin plots represent the distributions of log₂ MIC ratios for each antibiotic combination. Source data