Chemogenomic Screen for Imipenem Resistance in Gram-Negative Bacteria

Gram-negative carbapenem-resistant bacteria are a major threat to global health. The use of genome-wide screening approaches to probe for genes or mutations enabling resistance can lead to identification of molecular markers for diagnostics applications. We describe an approach called Mut-Seq that couples chemical mutagenesis and next-generation sequencing for studying resistance to imipenem in the Gram-negative bacteria Escherichia coli, Klebsiella pneumoniae, and Pseudomonas aeruginosa. The use of this approach highlighted shared and species-specific responses, and the role in resistance of a number of genes involved in membrane biogenesis, transcription, and signal transduction was functionally validated. Interestingly, some of the genes identified were previously considered promising therapeutic targets. Our genome-wide screen has the potential to be extended outside drug resistance studies and expanded to other organisms.

IMP MICs also showed decreased susceptibilities to MEM ( Table 3). The MEM MICs increased by 2-fold, 4-fold, and up to 16-fold in IMP-selected K. pneumoniae, E. coli, and P. aeruginosa, respectively ( Table 3). The MEM resistance in P. aeruginosa reached levels consistent with clinical resistance. Clusters of orthologous groups of mutated proteins. The genomes of 145 IMP-resistant clones (45 K. pneumoniae, 50 E. coli, and 50 P. aeruginosa clones) were analyzed by NGS. The genomes of the parent WT strains were also sequenced. A total of 3,810 single nucleotide (nt) variants (SNVs) were identified in the E. coli mutants (Fig. 1A), with a majority of the clones having at least 40 SNVs (Table S2). In K. pneumoniae, 1,379 SNVs were detected (Fig. 1A), with an average of 22 mutations (range, 5 to 39) per genome (Table S3). In P. aeruginosa, 654 SNVs were identified (Fig. 1A), with an average of 10 mutations (range, 2 to 18) found per mutant (Table S4). SNVs were more prevalent in coding regions than in intergenic regions for all species, and nonsynonymous SNVs predominated among coding mutations (Fig. 1A). While we cannot exclude the possibility that some SNVs resulted from IMP selection alone, it is highly likely that most were instead genuinely induced by EMS treatment. First, resistant clones were obtained using experimental conditions optimized such that no clone would grow upon IMP selection if the culture was not initially mutagenized by EMS. Second, most SNVs consisted of G-to-A and C-to-T transitions (Fig. 1B), which is consistent with the mode of action of EMS (29), and these were widespread among the 145 mutants sequenced. Some small insertions and deletions (InDels) were also observed in coding regions, and these were more frequent in P. aeruginosa than in E. coli or K. pneumoniae (Fig. 1A). We hypothesized that functional recurrence among mutated genes between species or clones would help in pinpointing the mutations that are the most relevant to IMP resistance. To ease comparisons among the three species, we relied classification using the Clusters of Orthologous Groups of proteins (COGs). The COG database is composed of over 4,600 specific functional COG descriptions grouped into 26 general category letter associations (30). We initially focused on the part of the COG descriptions corresponding to the mutated genes common to the three species, then on those shared by at least two species, and finally on the genes that are species specific but that were mutated in a higher number of clones. Thirty-five functional COG descriptions were found in common among the three species (Table S5). A third of these belonged to two general categories: (i) transcription mechanisms (K) and (ii) signal transduction mechanisms (T) (Fig. 2). The rpoD gene was the most prevalent in the COG category transcription, with mutations detected in a total of 10 mutants in the three species ( Table S5). A C58T transition occurred in rpoD for one mutant each of K. pneumoniae and P. aeruginosa, leading to G20S and E20K amino acid substitutions, respectively, while eight E. coli mutants harbored mutations leading to an A72V, A444V, A444T, or I457L substitution (Fig. 3A). Additional genes coding for DNA-binding transcriptional regulators, sensor histidine kinases (HK), major facilitator superfamily transporters, and multidrug transporters were also mutated in the three species, but the mutations were often seen in a single mutant for each species and thus they were not further studied (Table S5). The E. coli and K. pneumoniae mutants shared 275 COG descriptions. Of these, 11 consisted of one-to-one matches of E. coli and K. pneumoniae proteins that shared at least 70% sequence identity (implying genuine functional similarity) and that were also mutated in at least two mutants in each species (Table 4). These 11 genes belonged to 7 general functional categories, the cell wall/membrane/envelope biogenesis category (M) being the most prevalent, with 4 mutated genes ( Table 4). One of the 11 genes was amiC coding for an N-acetylmuramoyl-L-alanine amidase mutated in 9 E. coli and 26 K. pneumoniae mutants (Table 4). A majority of mutations localized to the AmiC domain of the protein, and several were nonsense mutations (see Fig. S1A in the supplemental material). Another gene was nlpD, mutated in 5 E. coli and 18 K. pneumoniae mutants ( Table 4). This gene codes for the activator of AmiC (31). Similarly to amiC, several nonsense mutations were observed in nlpD (Fig. S1B). The third gene was wecA, coding for an undecaprenyl-phosphate alpha-N-acetylglucosaminyl 1-phosphate transferase mutated in 5 mutants of each species (Table 4). The last gene from the M category was slt, which codes for a soluble lytic murein transglycosylase. The slt gene was mutated in 5 mutants, a majority harboring nonsense mutations (Fig. S1C). The remaining 7 genes were not part of the cell wall biogenesis category and had unrelated functions (Table 4). These were shared by 4 to 9 E. coli and K. pneumoniae mutants (Table 4). For gene clbJ (COG1020, category Q), the E1684G mutation found in 15 mutants of K. pneumoniae was present also in the IMP-sensitive E. coli WT strain, making it less likely to contribute to IMP resistance.
P. aeruginosa mutants were fairly distinct compared to E. coli and K. pneumoniae mutants, and these were thus not directly compared to E. coli and K. pneumoniae mutants in 2-way comparisons.
Species-specific mutations. Recurrence among clones was previously shown to ease the identification of SNVs contributing to the resistance phenotype (26), so, to retrieve gene candidates specifically mutated in E. coli, we thus focused on those mutated in at least 3 clones ( Table 1). The gene yceG encoding a cell division protein was the most prevalent among E. coli clones, with 15 SNVs (including 5 nonsense mutations) detected in 20 mutants ( Table 1). The second most prevalent gene was DR76_2948 coding for a trehalose-6-phosphate synthase, with two mutations (D376N and P380L) found in a total of 14 E. coli mutants (Table 1). Seven different mutations (5 of which were nonsense) occurred in the rne gene for six mutants. This gene codes for RNase E. Four different mutations occurred in five mutants for gene DR76_475, encoding a glutamate racemase. Finally, the tolA gene, which is part of the Tol-Pal cell envelope complex, was mutated in three E. coli mutants, including mutants M24 and M45, which also hold a mutated amiC gene (Table 1). A more exhaustive list of mutations can be found in Table S6.
The mutational landscape of P. aeruginosa was more limited than that determined for E. coli (Fig. 1A), so we lowered our cutoff for candidate genes to correspond to those mutated in at least two mutants (Table S7). All mutants had a mutation either in the porin OprD or in two-component systems (TCSs) consisting of a sensor HK and a response regulator (Table 2). For oprD, 20 different mutations were found in 28 P. aeruginosa clones, 16 of which led to a stop codon ( Table 2). Every mutant with a nonsense mutation in oprD had an IMP MIC of 16 g/ml, representing an 8-fold increase compared to WT P. aeruginosa (Table 2). This was also the case for mutant M34 Outer membrane receptor for ferric coprogen and ferric-rhodotorulic acid acrB a Gene accession numbers are indicated for E. coli ATCC 25922 and K. pneumoniae ATCC 13883. The number of mutants in which the genes were found mutated is indicated within parentheses. b Gene designations in bold represent orthologues sharing at least 70% sequence identity at the protein level. c Gene designations in italics represent genes that have been functionally tested for their role in resistance to IMP by individual transformation in E. coli ATCC 25922 WT (see Table 5).
harboring a Y343N substitution in OprD, while other oprD coding mutations (G183D, G402D, and S325F) were associated with lower IMP resistance levels ( Table 2). For TCSs, 15 mutants had a mutation in the sensor HK A4W92_13070 (Table 2). This HK is part of a two-component signal transduction system for which A4W92_13065, mutated in 4 independent P. aeruginosa clones (Table 2), is the response regulator. Interestingly, mutants M6 and M9 had a R419H mutation in the HK A4W92_04840 (Table 2), the same change that occurred in mutants M1, M3, and M11 for the HK described above (A4W92_13070). Lastly, the sensor HK PhoQ (A4W92_05675) was mutated in P. aeruginosa clones M5 (Q258*) and M36 (E198K) ( Table 2). Phenotypic validation of mutations highlighted by Mut-Seq. The role of specific mutations in resistance to IMP was tested in E. coli ATCC 25922 using a knock-in approach. This involved the transformation of WT cells with a DNA cassette made of a PCR fragment containing the mutation fused to a kanamycin (Kan) resistance gene flanked by FLP recombination target (FRT) sequences used for the removal of the selection marker by the FLP/FLPe recombinase. For the rpoD gene mutated in all species, we tested the G1331A transition that was detected in E. coli and that led to the A444V amino acid substitution (Table 5). This mutation conferred a 4-fold increase in IMP resistance in the rpoD G1331A ::kan transformants compared to the rpoD WT ::kan cells used as a control (we failed to remove the kan resistance marker in rpoD G1331A ::kan despite several attempts) ( Table 5). The transformant rpoD WT ::kan had the same IMP MIC (0.25 g/ml) as the E. coli ATCC 25922 WT. The A444V mutation allowed cells to sustain higher IMP concentrations than the control both in liquid medium (Fig. 3B) and on solid agar (Fig. 3C) but conferred a slight fitness cost in the absence of IMP (Fig. 3B). Interestingly, the rpoD G1331A ::kan mutant had decreased susceptibility also for MEM (Table 5).
Single knock-in, mutations in genes were transformed in an individual fashion; Double knock-in, mutations in two genes were transformed into the same E. coli ATCC 25922 cells. Gene IDs in bold had mutations in at least two E. coli and K. pneumoniae mutants (see Table 4). The gene in italics was mutated in E. coli, K. pneumoniae, and P. aeruginosa. b The mutant whose genomic DNA was used to amplify the mutation by PCR to generate the knock-in cassettes. c MICs were monitored with at least three biological replicates. For all differences of 2-fold or higher, there was no variability in the observed MICs. The kan resistance marker was removed from all transformants using the FLP/FLPe recombinase except for rpoD. ND, not determined.
Using the same knock-in approach, we also investigated the role in IMP resistance of mutations in genes shared by E. coli and K. pneumoniae mutants ( Table 4). The G402R substitution in AmiC, a position mutated in two E. coli mutants and one mutant of K. pneumoniae (Fig. S1A), decreased the susceptibility to IMP by 2-fold when introduced into E. coli ATCC 25922 (Table 5). Similarly, the nonsense mutations Q161* and R475* in NlpD and Slt, respectively, increased the IMP MIC by 2-fold (Table 5). E. coli mutant M14 and K. pneumoniae mutant M17 had mutations in both amiC and slt, and the AmiC G402R and Slt R475* mutations were indeed additive in increasing the IMP MIC by 4-fold under conditions of cotransformation into E. coli ATCC 25922 (Table 5). This amiC and slt double knock-in also showed decreased susceptibility to MEM (Table 5). Despite being mutated in 5 mutants each of E. coli and K. pneumoniae (Table 4), the role of wecA mutations was less clear as the transformant for the R40C mutation (detected in E. coli mutant M12) was not conclusively altered for its IMP MIC (Table 5). Regarding the mutations in genes that are unrelated to cell wall biogenesis but that have at least 70% sequence identity between E. coli and K. pneumoniae (Table 4), we tested mutations A138V in SpoT, A429T in DR76_3362, and A539V in GidA, but, similarly to the results seen with wecA, none had a significant impact on the IMP MIC (Table 5).
We also tested some representatives of the most prevalent genes specifically detected in E. coli. The mutations in yceG and rne leading to the Q92* and Q775* nonsense mutations, respectively, as well as the P380L substitution in DR76_2948 and the M67I mutation in TolA, were independently transformed in E. coli ATCC 25922. The IMP MIC was increased by 2-fold (0.5 g/ml) in the case of the rne and tolA knock-in (Table 5).
The knock-ins for mutations in wecA and spoT described above had an ambiguous phenotype by MIC measurements, so we looked for subtler phenotypes by monitoring their growth by serial dilution on solid medium in the presence of IMP. As expected, the E. coli M14 mutant (harboring an A138V mutation in SpoT) grew until it reached the highest cell dilution and IMP concentration tested. In contrast, E. coli ATCC 25922 WT grew only to the 10 Ϫ1 dilution on an agar plate supplemented with 0.25 g/ml of IMP (its MIC) (Fig. 4). As a positive control for transformation, we used the M67I mutation in TolA. This mutation allowed cells to grow until they reached the 10 Ϫ2 dilution at 0.5 g/ml (Fig. 4), consistent with the MIC of this transformant (Table 5). Mutations R40C in WecA and A138V in SpoT had an intermediate phenotype in growing with more dilutions than the WT cells at 0.12 g/ml IMP and at the 10 Ϫ1 dilution at 0.25 g/ml IMP (Fig. 4). Given its recurrence, we also tested the yceG gene but the Q92* mutation did not increase growth in the presence of IMP and was even detrimental under the conditions tested (Fig. 4).

DISCUSSION
The use of whole-genome screens is now allowing holistic views of mechanisms of action and mechanisms of resistance against antimicrobial agents. We applied here a Mut-Seq screen (25,26,32) combining chemical mutagenesis and NGS to obtain clones of E. coli, K. pneumoniae, and P. aeruginosa with decreased susceptibility to IMP. The main advantage of Mut-Seq in comparison to other approaches such as step-by-step drug selection is the rapidity with which resistant mutants are obtained (24 h versus months), hence accelerating studies on the mode of action or mechanisms of resistance to antimicrobials. We posited that using diverse species and analyzing a large set of independent mutants would pinpoint shared pathways of resistance as well as speciesspecific traits. As expected from phylogeny, the response of E. coli to IMP shared more similarity with K. pneumoniae than with P. aeruginosa. Some responses were shared among the three species, and while we concentrated on the genes mutated in the greatest number of mutants, it is quite possible that mutant-specific genes are also important for IMP responses.
The two main categories of orthologous proteins shared among the three species were those corresponding to transcription and signal transduction mechanisms. The most prevalent gene common to the three species was rpoD, which codes for the 70 factor that is associated with the core RNA polymerase complex for initiation of specific transcription (33). Mutations in rpoD were in distinct domains, and we validated experimentally the role of the A444V substitution detected in several independent E. coli mutants. This substitution occurred in a region of RpoD that is highly conserved, as it is part of the Ϫ10 promoter recognition helix binding (34). One current limitation of our work is the lack of validation of the role of rpoD directly in K. pneumoniae or P. aeruginosa. This type of effort could form the basis of further studies. While rpoD has never been shown to be involved with IMP resistance, sigma factors are well known to regulate a number of genes often associated with responses to stress (35). Further work may also provide insights into the downstream target gene(s) responsible for the decrease in IMP susceptibility. Many other genes (62 in E. coli and 39 in K. pneumoniae), including several that are strain specific, that are involved in transcription were mutated, and some may indeed help in the response to IMP. Among many of the TCSs mutated in each of the three species, evgS and phoQ were common (see Table S5 in the supplemental material). Knock-in of evgS or phoQ was not possible because of the gene arrangement and the close proximity of their respective regulators evgA and phoP.
IMP inhibits bacterial cell wall synthesis by binding to and inactivating the PBPs, with the highest affinity to PBP-1a, PBP-1b, and PBP-2 (36,37), leading to rapid cell lysis and death (38). MEM also targets PBPs, with the highest affinity for Pseudomonas PBP-2 and PBP-3. While less active than IMP against enterococci, it is more active against P. aeruginosa (39). Our highly IMP-resistant P. aeruginosa mutants and our E. coli knock-ins with at least 4-fold resistance to IMP were all less susceptible to MEM (Table 3 and  Table 5), demonstrating that our findings with IMP can be applied to MEM as long as the IMP MIC has reached a minimum threshold. This corroborates results of other studies demonstrating that IMP-resistant P. aeruginosa strains are usually cross-resistant to MEM as well (19,20). Similarly, E. coli strains selected for MEM resistance showed decreased susceptibility to IMP (40). Consistent with IMP mode of action, genes from the category of cell wall and membrane biogenesis were among those most highly represented in E. coli and K. pneumoniae. Amidases such as AmiC split the peptidoglycan (PG) of daughter cells during cell division (37). These amidases are autoinhibited and AmiC is activated by NlpD, a lipoprotein anchored in the outer membrane (31). Mutations in both amiC and nlpD contributed to a decrease in susceptibility to IMP (Table 5). It was demonstrated that the Tol-Pal system is implicated in regulating cell wall cleavage during cell division by activating NlpD (41) as well as in the modulation of PG synthesis (42). TolA, part of the Tol-Pal complex, was found mutated in three mutants of E. coli, and the M67I mutation decreased the susceptibility of the WT strain by 2-fold (Table 5). Mutations in amiC, nlpD, and tolA are likely to be loss-of-function mutations, since nlpD and amiC have nonsense mutations or InDels in many mutants (see Fig. S1A and B in the supplemental material). Our observations are consistent with previous studies where mutants lacking nlpD or amidases and several lytic transglycosylases (see below) displayed a delayed lysis response to ampicillin (43,44).
The PG is made of alternating N-acetylglucosamine (GlcNAc) and N-acetylmuramic acid (MurNAc) connected by a ␤-1,4-glycosidic bond (37). Lytic transglycosylases (LTs) cleave this glycosidic bond (37). E. coli has eight membrane-bound LTs (MLTs) and one soluble LT named Slt70 (37). Several of these LTs (MltA, MltB, MltD, and RlpA) were mutated in single E. coli or K. pneumoniae mutants and were not studied (Tables S2 and  S3), but the yceG (MltG) gene was mutated in 20 E. coli mutants, including 5 with nonsense mutations ( Table 1). The transformation of the mutation Q92* in the E. coli ATCC 25922 WT did not change the susceptibility of the strain to IMP, and it did not provide a growth advantage in the presence of IMP (Table 5) (Fig. 4). While we could not find a specific role of yceG in IMP resistance, its frequent mutation in E. coli suggests a role possibly in combination with other mutations. Indeed, MltG was shown to interact with PBP1b and in the absence of PBP1b, MltG was shown to be toxic (45). Since IMP inhibits PBP1b, it is possible that loss-of-function mutations in yceG are selected to limit its toxicity. Inhibition of PBPs by ␤-lactams leads to an accumulation of non-cross-linked PG, and Slt70 is the main enzyme responsible for destroying this nascent PG (46). Slt inactivation produced differential effects on ␤-lactam sensitivity depending on the genetic background (46)(47)(48). Here, we showed that a slt nonsense mutation (R475*) in E. coli ATCC 25922 decreased the susceptibility to IMP by 2-fold.
The response of P. aeruginosa to IMP differed extensively from the responses seen with the two Enterobacteriaceae species. The outer membrane permeability of P. aeruginosa is about 12-to-100-fold lower than that of E. coli (49), and this is probably due to a reduced number of general diffusion porins and the presence of a variety of specific porins such as OprD (16,50). IMP penetrates the pseudomonal membrane through OprD (51). Resistance to IMP in P. aeruginosa can occur through loss of OprD, which has been reported to take place at the levels of transcription and translation (14,52). Mutations resulting in a premature stop codon were found in a number of clinical isolates (53)(54)(55)(56), several of which (W6*, Q19*, W65*, W138*, Q158*, W277*, Q295*, W339*, and W415*) were also detected in this study together with new ones (Q30*, Q67*, Q79*, and Q235*) ( Table 2). Stop codons occurred at tryptophan or glutamine residues, representing the only two codons that can be changed to a stop codon through a single transition (along with one of the six codons for arginine) induced by EMS. TCSs are known to downregulate OprD and contribute to IMP resistance (57,58). Among the P. aeruginosa ATCC 27853 mutants selected in this study, 19 mutants had a mutation either in the sensor HK gene (A4W92_13070) or in the gene encoding its response regulator (A4W92_13065) without having a mutated oprD gene (Table 2). Two independent mutants had a mutation in another sensor HK gene (A4W92_04840) ( Table 2). These mutated HKs belong to the same clade as the ones known to regulate OprD (59). Two mutations (E198K and Q258*) in the sensor HK PhoQ (A4W92_05675) were detected in two independent clones (Table 2). A PhoQ-null mutant is resistant to polymyxin B and to aminoglycoside antibiotics (60)(61)(62), and PhoQ mutants were reported in P. aeruginosa clinical isolates resistant to polymyxin B (63) or colistin (polymyxin E) (64). Current understanding would suggest that the mutated HKs activate their respective response regulators and that they in turn downregulate oprD. Mutated HKs were also observed in E. coli and K. pneumoniae strains resistant to IMP (Table S5), highlighting the potential key role of HKs in IMP responses in Gram-negative bacteria.
Our chemogenomic screen performed with three bacterial species highlighted shared and species-specific responses to IMP. The most highly mutated genes encoded proteins involved in transcription, signal transduction, and membrane/cell envelope biogenesis. The number of mutants investigated allowed a holistic view of the response to IMP and enabled concentrating our functional work on the most frequently recurrent genes. Most mutations tested in E. coli were associated with a 2-fold difference in IMP susceptibility. This relatively low level of resistance may reflect more-subtle roles for the mutations, such as facilitating resistance emergence or compensating for fitness cost. Many other mutations are likely to be involved in response to IMP, and our data set can be useful to better understand IMP and to find strategies to restore carbapenem susceptibility in Gram-negative bacteria.

MATERIALS AND METHODS
Bacterial strains, plasmids, and growth conditions. Bacterial strains and plasmids used in this study are listed in Table S8 in the supplemental material. Unless otherwise specified, E. coli (ATCC 25922), K. pneumoniae (ATCC 13883), and P. aeruginosa (ATCC 27853) were cultured on Luria-Bertani (LB) or nutrient agar (1.5%) and incubated at 37°C for 18 to 24 h. Liquid cultures were grown in LB for E. coli and in brain heart infusion (BHI) medium for K. pneumoniae and P. aeruginosa. Plasmid pRedET and the FRT-flanked PGK-gb2-kan cassette (catalogue number K006) and the enhanced FLP expression plasmid 707-FLPe (catalogue number A104) with a tetracycline resistance marker were obtained from Gene Bridges GmbH. IMP-monohydrate and MEM were purchased from Santa Cruz Biotechnology; all other chemicals were purchased from Sigma-Aldrich.
Antibacterial susceptibility testing. Analyses of the IMP or MEM MICs for the WT strain of each of the three species and for their respective mutants and optimization of the EMS concentrations (Table S8) were performed by microdilution in 96-well plates according to recommendations from the CLSI. The IMP or MEM MICs for the single or double knock-ins were determined by macrodilution. All MICs were determined from at least three independent biological replicates, each replicate being further assessed in technical duplicates.
Chemical mutagenesis and selection of antibiotic-resistant mutants. Optimization was performed by testing different concentrations of EMS equivalent to 4ϫ, 8ϫ, or 16ϫ its MIC for each of the three species for either 10 or 20 min. We selected conditions that allowed treated cells to reach an optical density at 600 nm (OD 600 ) of 0.5 in less than 6 h. For K. pneumoniae and P. aeruginosa, this represented an EMS concentration equivalent to 4 times their MIC (0.024 g/ml and 0.012 g/ml, respectively) maintained for 10 min. For E. coli, we used an EMS concentration equivalent to 8 times its MIC (0.048 g/ml) maintained for 20 min. The minimum concentration of IMP used for selection was determined as the concentration at which growth occurred in the presence of IMP for the mutagenized populations but not for the nonmutagenized control populations. This represented IMP concentrations equivalent to 16ϫ and 20ϫ the MIC in the case of E. coli (4 and 5 g/ml) and between 2ϫ and 4ϫ the MIC for both K. pneumoniae (2 and 4 g/ml) and P. aeruginosa (4 and 8 g/ml). No clones survived beyond these concentrations. The detailed protocol was as follows: the overnight (ON) cultures of the strains were diluted and incubated at 37°C with shaking (220 rpm) until they reached an OD 600 of 0.5. Each culture was separated into two tubes of 10 ml. We added EMS to one of the tubes at the appropriate concentration. Cultures were incubated for 10 min (K. pneumoniae and P. aeruginosa) or 20 min (E. coli) at 37°C. Cultures were then diluted by half using ice-cold medium and then further diluted by 1/10 before being incubated until an OD 600 of 0.5 was reached. The cultures were centrifuged at 4,000 rpm for 5 min, and then the pellet was resuspended in 200 l of the culture medium. Ten-fold dilutions (from 10 Ϫ3 to 10 Ϫ7 ) were spread on agar plates to allow colony counting.
In order to select mutants resistant to IMP, 100-l volumes of the mutagenized cultures were spread on agar plates containing an increasing concentration of the antibiotic (between 2ϫ and 20ϫ the MIC depending on the species). Agar plates were incubated overnight at 37°C, and colonies were counted the next day. This protocol was performed for each strain in duplicate.
Extraction and quantification of DNA. Genomic DNA (gDNA) was extracted using a Wizard Genomic DNA purification kit (Promega) according to the manufacturer's protocol. The purity of the gDNA was analyzed using a NanoDrop spectrophotometer. Quantification was performed by fluorescence detection using a QuantiFluor One double-stranded DNA (dsDNA) system (Promega).
DNA sequencing. Libraries were produced from 0.8 ng of gDNA using a Nextera XT DNA Library Prep kit (Illumina) according to the manufacturer's protocol. Libraries were verified by the use of model 2100 Bioanalyzer high-sensitivity DNA chips (Agilent) and quantified by the use of a QuantiFluor One dsDNA system (Promega). Libraries were sequenced on an Illumina HiSeq 2500 system (101-nt paired-end sequencing) at a final concentration of 8 pM.
Mut-Seq data analysis. Sequence reads were aligned to the E. coli ATCC 25922 (BioProject accession no. PRJNA244551), K. pneumoniae ATCC 13883 (PRJNA244567), and P. aeruginosa ATCC 27853 (PRJNA316664) genomes using bwa-mem software (65). The seed length was 32, and 2 mismatches were allowed within the seed; the maximum number of mismatches allowed was 4. Read duplicates were marked using Picard (http://broadinstitute.github.io/picard/), and calling of SNVs and InDels was done using GATK software (66). Mutations in common with the WT strain sequence examined in parallel were excluded, and the remaining mutations (i.e., mutant-specific mutations) were annotated. To ease the comparisons among the three species, the common COGs (30) were determined using the workflow of COGsoft (https://sourceforge.net/projects/cogtriangles/files/).
Generation of knock-ins in E. coli. We first transformed E. coli ATCC 25922 with the expression plasmid pRedET as recommended by the manufacturer (Gene Bridges GmbH). Mutated PCR fragments of the genes of interest were amplified from the appropriate E. coli mutant and fused to the FRT-flanked PGK-gb2-kan cassette before being transformed in the E. coli strain containing the pRedEt plasmid. This strategy was previously described by Sukhija et al. and Pyne et al. (67,68). L-Arabinose was added to induce the expression of the Red/ET recombination proteins, and transformants were selected with Kan (40 g/ml). Colonies were analyzed by colony PCR and sequencing. All primers used to generate the single knock-ins and to check the right integration of the gene replacement cassette are listed in Table S8. The enhanced FLP conditional expression plasmid 707-FLPe was transformed into the strain containing the gene replacement cassette so that recombination would occur at the FRT sites. PCR was used to confirm both the removal of the Kan cassette by FLP recombination and the presence of the knock-in mutation.
Testing the effect of mutations on the growth of E. coli. Cultures (OD 600 ϭ 0.9) were serially diluted (10 Ϫ1 to 10 Ϫ6 ) in LB, and 5-l volumes were spotted on freshly prepared LB agar plates supplemented with different concentrations of IMP. Plates were incubated at 37°C overnight, and they were photographed the next day using an AlphaImager system (Alpha Innotech).
For determination of the growth curve in liquid medium, 5 l of 1 ϫ 10 7 CFU/ml was inoculated in 200 l of LB medium in the absence or presence of IMP at concentrations of 0.03 to 4 g/ml in 96-well plates. The plate was incubated at 37°C, and the OD 600 was read each 30 min after shaking for 10 s using a Cytation 5 multimode reader. Each assay was done in technical triplicate and biological triplicate.
Data availability. The WGS data have been deposited in the SRA database, and the accession numbers are listed in Table S9.

ACKNOWLEDGMENTS
We thank Michel G. Bergeron for providing bacterial strains ATCC 13883 and ATCC 27853.