Adaptive and maladaptive consequences of deregulation in a bacterial gene regulatory network

The archetypal PhoQP two-component system from Enterobacteria regulates crucial pathways like magnesium homeostasis in Escherichia coli and virulence factor expression in Salmonella enterica. Previously we had reported that a laboratory strain of E. coli rapidly accumulated loss-of-function mutations in the mgrB gene, a negative feedback regulator of PhoQP, when evolved in the presence of the antibiotic trimethoprim. Hyperactive PhoQP enhanced the expression of dihydrofolate reductase (folA), target of trimethoprim, resulting in antibiotic tolerance. Here we ask, firstly, how important are mutations in mgrB for trimethoprim resistance? Using laboratory evolution, we show that trimethoprim resistance evolves by different mutational trajectories under condition of high and low PhoQP activity. Mutations in mgrB are only fixed when PhoQP is active. Importantly, loss of functional MgrB, though itself only mildly beneficial, enhances the fixation probability of trimethoprim-resistant bacteria under selection and this can be explained by epistasis between mgrB and folA loci. As a result, the activation status of PhoQP directly impacts how fast resistance is acquired by evolving populations of E. coli. Secondly, we investigate why negative feedback may be needed in the PhoQP system. We show that under drug-free conditions MgrB is required to mitigate the fitness costs of pervasive gene dysregulation by hyperactive PhoQP. Using RNA-seq transcriptomics and genetic analyses, we demonstrate that PhoQP-hyperactivation perturbs the balance of RpoS and RpoD-regulated transcriptional programs, and spontaneous mutations in rpoS rectify this imbalance. We propose that deregulation can be adaptive or maladaptive depending on the environmental context and this explain the evolution of negative feedback in bacterial gene regulatory networks.


Introduction
Signal transduction pathways perform the vital function of altering cellular physiology in response to the environment. Signaling cascades often culminate in gene expression states that dictate the behavior of cells and facilitate adaptation to stressful environments. Biochemical and genetic experiments have elucidated the roles of individual signaling cascades in detecting stimuli at the cell surface, transducing this signal intracellularly and modulating gene expression by changing activation states of transcription factor/s. Building on this information, systems-level approaches have revealed that signaling proteins are organized into networks, often referred to as gene regulatory networks [1,2]. Organization of regulatory proteins into networks not only brings in robustness and exquisite control, but also allows for crosstalk between signaling cascades, ultimately shaping which genes are expressed and to what extent [1][2][3].
A ubiquitous feature of gene regulatory networks, seen from bacteria to humans, is the presence of feedback [1,2,[4][5][6][7]. Positive feedback in regulatory networks ensures rapid 'switch-like' behavior in response to an activating signal [7,8]. It allows for bistability, i.e., the existence of two relatively stable transcriptional states [9]. In Gram positive bacteria such as Bacillus subtilis, bistability plays an important role in governing cell cycle and developmental programs like sporulation [9]. Negative feedback on the other hand dampens, restricts, or resets activation levels of a signaling network. Negative feedback is thought to be corrective and is required to calibrate levels of gene expression and buffer noise [3,10,11]. Corrective feedback has been shown to facilitate adaptation in bacterial signaling as well as linearize gene expression responses [12].
The major signaling proteins used by bacteria are two-component systems [13][14][15]. The first component, referred to as 'sensor kinase', auto-phosphorylates at a conserved histidine residue upon activation. It then transfers its phosphate group to a conserved aspartate residue on a second protein, called the 'response regulator', which is usually a transcription factor itself. Phosphorylation status of the response regulator directly alters its ability to bind to gene promoters and modify gene expression [13,15]. Several variations on this basic schema are observed in nature, including dual kinase-phosphatase activity of the sensor kinase [16], multiple phosphorelay steps [13,17] and crosstalk between non-cognate kinases and response regulators [18,19]. Two-component systems across bacteria orchestrate responses to a wide range of cues such as metal ions [20], pH [21], nutrients [22], antimicrobial peptides [23], quorum sensing molecules [24] and cell envelope stress [25,26]. More recently, twocomponent sensor kinases have been proposed as novel targets for the design of antimicrobials owing to their association with drug resistance in several clinically relevant bacterial species [27][28][29].
Two-component signaling pathways are important substrates for adaptive evolution in bacteria and mutations in two-component signaling proteins are responsible for adaptation to adverse environments. For instance, sequence variation in the EvgAS two-component system is implicated in modulating the survival of Escherichia coli strains in low pH [30]. Similarly, mutations in the DosR-DosS-DosT system in the Beijing lineage of Mycobacterium tuberculosis have been associated with hypervirulence [31,32]. We showed previously that loss-of-function mutations in the mgrB gene are rapidly enriched in E. coli bacteria that were exposed to the antibiotic trimethoprim [33]. The mgrB gene codes for a small membrane protein that inhibits the PhoQP two-component system by binding to the sensor kinase PhoQ [34,35]. Expression of mgrB is itself activated by PhoP, setting up a negative feedback loop

Mutations in mgrB, folA and rpoS genes drive adaptation to trimethoprim in laboratoryevolved E. coli
In our laboratory evolution experiments loss of functional mgrB was the first mutational event in E. coli adapting to trimethoprim [33]. Long term antibiotic exposure led to adaptive sweeps involving mutations in 2 other genes i.e., folA, which codes for dihydrofolate reductase (DHFR) and rpoS, which codes for the enterobacterial stationary phase sigma factor [33]. While mutations in mgrB and folA directly enhanced trimethoprim resistance, rpoS-mutations enhanced fitness without altering drug IC50 [33]. To test whether these 3 loci reproducibly accumulated mutations during the evolution of trimethoprim resistance, we sequenced the genomes of 3 randomly picked resistant isolates from 3 independent lineages of E. coli after 350 generations of evolution in trimethoprim. Mutations in folA were found in isolates from 2 of the 3 lineages, while mgrB harboured mutations in all 3 lineages (Figure 1, Supplementary File 1). Like mgrB, rpoS mutations too were found across all 3 lineages (Figure 1, Supplementary File 1). Though mutations in several other genes were also found (Supplementary File 1), folA, mgrB and rpoS were the only loci that harboured mutations consistently across lineages, reinforcing their role as the primary hotspots for adaption to trimethoprim.

Mutational loss of mgrB facilitates evolution of trimethoprim resistance
By itself, loss of mgrB led to a mild enhancement in drug IC50 [33]. Despite this, mgrB was consistently mutated across independently evolved trimethoprim-resistant bacteria ( Figure 1). This observation suggested that though mgrB-deficiency alone had only a small effect on drug resistance, it may serve a facilitatory role during adaptation to trimethoprim. To empirically test this idea, we asked, firstly, whether it was possible for E. coli to evolve resistance without implicating mgrB and secondly, if resistance did evolve without mutations in mgrB, how it would impact the rate of resistance evolution.
In strain competition experiments, E. coli mgrB showed a dose-dependent increase in relative fitness in trimethoprim-supplemented media (Figure 2A). High concentration of Mg 2+ (5-10 mM) in growth media, which is known to inhibit PhoQ activity [21,41] completely neutralized this fitness advantage of mgrB-deficiency ( Figure 2A). Thus, loss of MgrB was only advantageous for E. coli under conditions in which PhoQP was active. Based on this result, we established 6 evolving lineages (1, 2, 3: High Mg 2+ ; 4, 5, 6: Low Mg 2+ ) that were exposed to sub-MIC trimethoprim (100 ng/mL) under conditions of high or low PhoQP activity for ~210 generations ( Figure 2B). We argued that, since PhoQP would be active only in low Mg 2+ , these lineages alone would evolve mutations in mgrB, allowing us to dissect out its role in the evolution of resistance. In all three Low Mg 2+ lineages, trimethoprim-resistant bacteria were rapidly fixed (i.e. frequency >0.9) within the first 50 generations of evolution ( Figure 2C). In contrast, High Mg 2+ lineages evolved resistance at much slower rates, and resistant bacteria were unable to reach fixation over the duration of the experiment ( Figure 2C). To ensure that this effect was due to the activity of PhoQP and not an independent effect of Mg 2+ , we established 3 independent evolving lineages starting with an isogenic E. coli phoP strain in low Mg 2+ media supplemented with trimethoprim ( Figure 2B). In these populations too, trimethoprim-resistant bacteria did not get fixed and remained at low frequencies even after 200 generations of evolution ( Figure 2C). Thus, low/no active PhoQP severely impeded the establishment of trimethoprim-resistance in E. coli populations.
Genome re-sequencing of 3 randomly picked resistant isolates from each of the lineages confirmed that, as expected, mutations in mgrB occurred in all Low Mg 2+ isolates ( Figure 2D, Supplementary File 2). In contrast, none of the isolates from High Mg 2+ and phoP lineages harboured mgrB-mutations ( Figure 2D, Supplementary File 2). Importantly, all 27 sequenced trimethoprim-resistant isolates harboured mutations at the folA locus, either in the gene promoter or in the coding sequence ( Figure 2D, Supplementary File 2). Genome sequencing of the entire population further confirmed these findings (Supplementary File 2). These results demonstrated that mutations in folA were sufficient for the evolution of trimethoprimresistance. However, mgrB-mutations, though not necessary for the evolution of resistance, facilitated fixation of resistant bacteria under trimethoprim selection.

Epistasis between mgrB and folA facilitates the fixation of trimethoprim resistant bacteria
We next investigated the mechanism underlying the facilitatory role played by mgrB-loss. Intergenic epistasis is a well-known determinant of the mutational landscape of resistant bacteria evolving at sub-MIC drug [42]. Therefore, we first assessed whether mutations in mgrB were epistatic with drug-resistant folA alleles. For these analyses we chose representative strains that harboured mutations in mgrB alone, folA alone (coding region and its promoter) or at both loci and calculated fold-IC50 values over wild type. Loss of mgrB alone (E. coli mgrB) resulted in a marginal enhancement in IC50 of ~3-fold over wild type ( Figure 3A). On the other hand, isolates harbouring either a missense mutation in the folA gene (Isolate Tmp R -folA-Trp30Arg) or a promoter-up mutation (Isolate Tmp R -folA-C-35T) had significantly higher values of fold-IC50 ( Figure 3A). Based on a non-epistatic/additive model [42], we expected ~152-fold or ~41-fold increase in IC50 of strains harbouring mgrB-mutations in addition to the folA-C-35T and folA-Trp30Arg alleles respectively ( Figure 3A). Deviation from these expected values would indicate epistasis between mgrB and folA. We selected resistant isolates Tmp R -A and Tmp R -B, derived from our long-term evolution lines to test this prediction. Tmp R -A harboured mutant mgrB and folA-C-35T, while Tmp R -B harboured mutant mgrB and folA-Trp30Arg. Interestingly, IC50 values of both isolates deviated from expectation, but in opposite ways. Though both strains had higher IC50 values than mgrB/folA mutants alone, Tmp R -A showed "less-than-additive" effect or magnitude epistasis, while Tmp R -B showed "greater-thanadditive" or synergistic epistasis ( Figure 3A). To verify that the observed effects could be attributed to hyperactive PhoQP we deleted the phoP gene from Tmp R -A and Tmp R -B. Indeed, deleting phoP from Tmp R -A led to only a ~10 % reduction in fold-IC50 while Tmp R -B showed close to ~80 % reduction in fold-IC50 ( Figure 3A), corroborating our results. These experiments showed that mutations in mgrB enhanced the IC50 values of bacteria harbouring folA mutations, but this effect was more pronounced for coding mutations than promoter mutations.
In parallel, we also analysed the expression level of folA in these resistant strains. Loss of functional MgrB alone led to ~2 fold higher levels of folA transcript than wild type ( Figure  3B). Likewise, Isolate Tmp R -B also showed ~2 fold higher folA transcript, while Tmp R -folA-Trp30Arg had comparable folA expression as wild type ( Figure 3B). On the other hand, the Tmp R -folA-C-35T showed massive overproduction of folA, which was not detectably higher in strain Tmp R -A ( Figure 3B). Thus, mgrB-mutations enhanced the expression of wild type or missense alleles of folA. However, the effect of mgrB-mutation was marginal if cis-regulatory changes in the folA promoter were present as the latter masked the effects of hyperactive PhoP. The strongly correlated trends between IC50 values and folA expression demonstrated that the molecular mechanism behind the observed epistasis was MgrB's influence on folA expression levels.
Could this genetic interaction explain why mutations in mgrB facilitated fixation of resistant bacteria? To answer this question, we calculated the frequency at which resistant strains harboring different mutation combinations could establish over a large excess of a drugsensitive competitor in the presence and absence of sub-MIC trimethoprim (100 ng/mL) ( Figure 3C). Competitions (144 replicates for each strain combination) were set up with wild type E. coli at a starting ratio of 1:10 7 in favour of wild type to mimic initial stages of evolution when a resistant mutant first emerges in an ancestrally sensitive population ( Figure 3C). After 24 hours of competition, 0.5 % of the mixed culture was passaged into LB media containing trimethoprim at the MIC of wild type (1 g/mL). Growth at this concentration of trimethoprim would indicate that the resistant mutant had been enriched >10-fold during the competition ( Figure 3C). Wild type bacteria alone and competitions in the absence of trimethoprim were used as controls. Indeed, strains Tmp R -A and B had higher frequencies of establishment than Tmp R -folA-C-35T and Tmp R -folA-Trp30Arg demonstrating the impact of mgrB mutations as facilitators of resistance evolution ( Figure 3D). Further, mgrB-mutation enhanced the frequency of establishment of the folA-Trp30Arg allele to a greater extent than folA-C-35T, in line with the differential epistatic effects between these mutations ( Figure 3D). Thus, we concluded that mgrB-mutations facilitated the fixation of bacteria with mutations in folA under drug pressure, which could be explained mechanistically by epistasis between mgrB and folA loci.

MgrB is retained by E. coli to prevent costs of PhoQP hyperactivation
Since loss of functional mgrB was frequent and beneficial under antibiotic pressure, we wondered why negative feedback may be needed in the PhoQP system in the first place. To address this question, we first assessed the evolutionary conservation of mgrB across PhoQPexpressing bacteria. The PhoQP system is restricted to the Order Enterobacterales [43]. Within this order, the phoQ gene was found in all bacterial families except Budviciaceae, indicating that PhoQP may have emerged after the divergence of Budviciaceae from the other 6 families ( Figure 4A). The mgrB gene was present in representative members of 4 of the 6 families that harboured phoQ, namely Morganellaceae, Yersiniaceae, Pectobacteriaceae and Enterobacteriaceae, but not in Erwiniaceae and Hafniaceae ( Figure 4A). Phylogenetic relationships between these families [44] suggested that Erwiniaceae and Hafniacieae may have independently lost mgrB during evolution. We next examined the distribution of phoQ and mgrB among members of family Enterobacteriaceae, to which E. coli belongs ( Figure 4B). Among 61 query species examined by us, we found that most harboured both, phoQ and mgrB. A few exceptions were also identified, such as Izhakiella, Rosenbergiella, Limnobaculum and some species of Candidatus, which code for PhoQ but not MgrB ( Figure 4B). Once again, their phylogenetic relationships suggested that these genera represented independent gene-loss events ( Figure 4B). Taken together, these analyses showed that multiple independent events of loss of the mgrB gene have occurred during bacterial evolution. However, most extant bacterial species preserved a functional mgrB gene, indicating that negative feedback in PhoQP is dispensable but desirable.
In line with this idea, there was no detectable difference in the growth characteristics of wild type E. coli and its isogenic mgrB knock-out strain under standard laboratory conditions ( Figure 5A). However, when E. coli mgrB was directly competed against lacZ-marked wild type, a significant fitness cost (wΔmgrB = 0.83±0.03) was observed ( Figure 5B, C). The fitness cost of mgrB-deficiency was due to hyperactivation of PhoQP since addition of high Mg 2+ to growth media or deletion of phoP/Q genes restored relative fitness to wild type levels ( Figure  5B, C). Our earlier work has shown that loss-of-function mutations in rpoS that occurred spontaneously during long term evolution in trimethoprim improved the fitness of mgrBdeficient E. coli in the drug-supplemented media without enhancing IC50 [33]. Since RpoS is itself an indirect target of PhoQP signaling and known to be overproduced in mgrB-knock out bacteria [45], we asked whether loss of rpoS could compensate for the fitness costs of mgrBdeficiency in drug-free media as well. This was indeed the case and deletion of rpoS restored the relative fitness of an mgrB-knockout strain ( Figure 5B, D). Like rpoS, deletion of iraM, which links rpoS to PhoQP signaling [45], also rescued the fitness cost of mgrB-deficiency ( Figure 5B, D). Thus, we concluded that the costs of mgrB-deficiency could be mechanistically traced to hyperactivation of PhoQP and over-production of RpoS, explaining why negative feedback may be needed in the PhoQP system.

Pervasive gene dysregulation in mgrB-deficient E. coli is rectified by mutations in rpoS
We next sought to understand the molecular mechanistic basis for the cost of mgrB-deficiency and the compensatory role of rpoS-mutations. To do this, we compared the whole transcriptomes of trimethoprim-resistant isolates from early and late time points of our longterm evolution lines using RNA-seq. The isolates Tmp R -A and Tmp R -B were chosen from early time points since they both harboured loss-of-function mutations at the mgrB locus and had wild type rpoS ( Figure 6A). Since both isolates also harboured other mutations, comparing the transcriptomes of Tmp R -A and Tmp R -B would help to identify mgrB-specific gene regulatory effects. The isolate Tmp R -C was chosen from a later time point as it harboured the same mutation in mgrB as Tmp R -B and an inactivating mutation in rpoS ( Figure 6A). All three isolates also harboured mutations in the folA gene (Tmp R -B, C) or its promoter (Tmp R -A) ( Figure 6A).
At the global level, the differential gene expression profiles of Tmp R -A and Tmp R -B were very similar and strongly correlated (linear correlation coefficient R 2 = 0.59) ( Figure 6B). On the other hand, Tmp R -B and Tmp R -C showed significant differences in their transcriptomes (linear correlation coefficient R 2 = 0.19) ( Figure 6C). Next, we turned our attention specifically to genes of the PhoP-regulon. A total of 58 genes are known direct targets of the PhoQP system in E. coli [46][47][48]. We noted that despite the presence of mgrB-mutations, majority of the PhoPregulon remained unaltered in all 3 isolates ( Figure 6D, Supplementary File 3). For instance, the levels of PhoP-targets such as the acrAB efflux pump and the glg glycogen metabolism operon were unaffected by the loss of mgrB. These genes are regulated by other transcription factors in addition to PhoP, including master regulators such as CRP [46][47][48], and hence may be less sensitive to the activation status of PhoP. Hyperactivation of PhoQP was reflected most dramatically in the overexpression of genes that are solely regulated by PhoP, such as the phoQP operon itself or the magnesium transporter mgtA ( Figure 6D, F, Supplementary File 3).
A subset of PhoP-targets is also regulated by the RpoS sigma factor [46][47][48]. Targets of PhoP and RpoS include genes that form the 'acid-resistance regulon' such as the 'gad' and 'hde' operons [45]. In Tmp R -A and B most of the PhoP-RpoS co-regulon showed mild to high overexpression compared to wild type ( Figure 6E, Supplementary File 3). However, in Tmp R -C, several of these genes were significantly down-regulated relative to wild type, indicating that mutation in rpoS overrides the effects of PhoQP hyperactivation and compensates for the hyperactivity of PhoP ( Figure 6E, Supplementary File 3). We confirmed these findings using qPCR for representative genes and found the results to be in line with the RNA-seq transcriptomics data ( Figure 6F).
Curiously, we noticed that there were several differentially expressed genes between isolates that are not direct targets of PhoP or RpoS (Supplementary File 3). Particularly striking among them were genes that were significantly down-regulated in Tmp R -A and Tmp R -B but restored to wild type levels in Tmp R -C ( Figure 6C, Supplementary File 3). A closer look at this sub-set of genes revealed that a majority of them were transcribed in an RpoD-dependent (i.e. Sigma 70-dependent) manner and were involved in growth and metabolism such as lpxT (lipid A metabolism), ccmA-D (cytochrome maturation) and rcnA-B (metal ion homeostasis) ( Figure  6C, G, Supplementary File 3). Similarly, several RpoD-regulated tRNA genes were also expressed to a higher level in Tmp R -C, compared to Tmp R -A and Tmp R -B ( Figure 6G, Supplementary File 3). It is well-established that RpoS and RpoD compete for the same binding site on RNA polymerase and competition between these two sigma factors dictates whether E. coli expresses genes required for growth and division (RpoD-regulated) or stress-response and stationary phase (RpoS-regulated) [49]. The dysregulation of RpoD-target genes in Tmp R -A and B thus suggested that precocious RpoS activation due to mgrB-deficiency tilted the balance in favour of the RpoS-transcriptional program, which was compensated at later stages in evolution by mutations in rpoS. Based on these results, we concluded that loss of mgrB led to pervasive perturbation of gene expression beyond the PhoP-regulon, its primary target, to secondary and tertiary effects on the expression of RpoS-and RpoD-regulated genes respectively. Mutations in rpoS restored several of these secondary and tertiary gene regulatory effects.

RpoS-RpoD imbalance explains the fitness costs of MgrB-deficiency in E. coli and justifies the need for negative feedback in the PhoQP system
Based on the above result, we hypothesized that the cost of mgrB-loss may arise either due to overproduction of RpoS-regulated genes or the repression of RpoD-regulated genes. To test which of these two possibilities was true, we first generated knockouts of the hdeD, gadW and gadE genes (co-targets of RpoS and PhoP) in an mgrB-deficient background and asked whether these gene deletions could rescue the fitness cost of the mgrB knockout. These specific genes were selected since they are master regulators of the acid-resistance regulon and were upregulated in Tmp R -A and B, but significantly repressed in Tmp R -C. However, the fitness of the mgrB knockout was unaffected by deletion of these genes ( Figure 7A) ruling them out as the source of the observed fitness cost. Next, to test whether the imbalance between RpoS and RpoD could explain the costs of mgrB-deficiency, we deleted the rsd or crl genes from the mgrB-knockout and measured relative fitness. Rsd is an inhibitor of RpoD [50,51], while Crl is potentiator of RpoS [52], and loss of either protein would tilt the balance in favour of RpoDregulated transcription. Indeed, deletion of rsd or crl rescued the costs of the mgrB-knockout strain ( Figure 7A). Finally, we traced how the competitive ability of E. coli mgrB changed with growth phase. We found that in the logarithmic phase, when RpoD is known to be higher, there was no detectable difference between mgrB-deficient and wild type bacteria in a mixed culture. However, the competitive disadvantage of E. coli mgrB became evident at the onset of stationary phase when the shift from RpoD to RpoS-mediated transcription is known to occur ( Figure 7B). This effect too was compensated by deletion of rpoS ( Figure 7B). These results confirmed that pervasive transcriptional dysregulation of gene expression due to hyperactive PhoQP explained the costs of mgrB deficiency and provided a mechanistic explanation for why negative feedback is retained by this two-component system during evolution.

Discussion
Loss of functional MgrB occurs spontaneously under antibiotic pressure and is associated with resistance to trimethoprim in E. coli [33,38] and colistin in K. pneumoniae [39]. Being a negative feedback regulator of the PhoQP two-component system, absence of MgrB hyperactivates PhoP and leads to overexpression of several of its targets [37,45]. In this study, we have used the PhoQP/MgrB two-component system from E. coli to ask what the contribution of gene regulatory evolution is to antimicrobial resistance in bacteria. Our results demonstrate that loss of checks and balances such as negative feedback in gene regulatory pathways can be adaptive for bacteria under antibiotic pressure. Further, deregulating gene expression has the potential to amplify the phenotypes of other resistance-conferring mutations and hence facilitate the evolution of high level antibiotic resistance. Importantly, we have also shown that deregulation of signalling pathways has pervasive effects on the wider gene regulatory network of cells, which are maladaptive. As a result, it is accompanied by a fitness cost that must be compensated through additional evolutionary adaptation ( Figure 8).
Increased expression of drug-targets, most commonly due to mutations in gene promoters or transcription factors, is reported in many antibiotic-resistant bacterial pathogens. For instance, mutations in the promoter of the pbp4 gene in beta-lactam resistant Staphylococcus aureus [53], missense mutations in mgrB or phoQ in colistin-resistant Klebsiella pneumoniae [54] and inhA promoter mutations in isoniazid-resistant Mycobacterium tuberculosis [55] are all frequently-encountered, clinically-relevant mechanisms of antibiotic resistance. The main role of regulatory mutations is assumed to be to enhance bacterial fitness by increasing expression of the drug target. In our study, loss of functional MgrB, and associated overexpression of folA did indeed confer a fitness advantage to E. coli challenged with trimethoprim. By itself, this benefit was incremental and would normally be considered clinically irrelevant. However, mgrB-loss contributed significantly to the evolution of high-level trimethoprim resistance by synergistic epistasis with missense mutations in folA. Consequently, mgrB-mutations were the most frequent genetic change in laboratory-evolved trimethoprim resistant bacteria despite the small stand-alone impact on drug IC50/MIC. Thus, our study throws light on another, perhaps more significant role of regulatory mutations as facilitators of antimicrobial resistance evolution through epistasis. Phenomenologically similar ideas have been proposed by a few other studies, though in these cases the mechanistic bases were unclear. For instance, antibioticinduced alterations in global gene expression promoted the fixation of resistance-conferring mutations in E. coli challenged with amoxicillin and tetracycline [56]. Similarly, expressionmodifying mechanisms have been shown to serve as "latent defences" that bacteria can exploit in order to adapt to antibiotics [57]. Taken together, the role played by regulatory mutations in facilitating the evolution of antimicrobial resistance is likely to be more wide-spread than understood so far and requires greater attention.
Based on our previous work, as well as work from other groups, we explain the mechanism of epistasis between mgrB and folA as follows. Missense mutations in DHFR can structurally destabilise the protein and reduce its steady state levels [58,59]. Hyperactive PhoP is likely to compensate for this effect by increasing the transcription of mutant DHFR in the bacterial cell, resulting in high levels of drug resistance. Interestingly, we have shown earlier that loss of Lon protease activity also shows synergy with unstable drug-resistant DHFR mutants by increasing their in vivo half-life [60]. Thus, increasing expression level of mutant DHFRs, regardless of mechanism, can facilitate high-level resistance evolution. This mechanism is consistent with our observation that coding mutations in DHFR synergize with mgrB-mutations, while mutations in the folA promoter don't do so. It is important to note here that the number of possible resistance-conferring coding-region mutations in folA is much higher than promoter mutations [33,[61][62][63]. Thus, by-and-large mgrB is more likely to synergistically enhance the resistance level of folA-mutant bacteria than not.
We propose that characterising gene regulatory evolution in drug-resistant bacteria may help to identify new points of intervention for AMR pathogens. Our study shows that "switchingoff" PhoQP signalling sensitized resistant E. coli to trimethoprim as well as retarded resistance evolution. Sensitization strategies such as this one are potential solutions to the current AMR crisis. A few approaches to achieve sensitization to antibiotics that have been explored so far have used adjuvant molecules that enhance the concentration of antibiotics in bacterial cells [64]. For instance, efflux pump inhibitors like verapamil have been explored as a possible mode of re-sensitising multi-drug resistant bacteria [65]. Similarly, cell wall perturbing polymers can also potentiate the activity of some antibiotics by enhancing their penetration into bacterial cells [65,66]. Our work shows that modifiers of two-component system-regulated gene expression may serve as an additional or alternative sensitization strategy. Though our results have been limited to trimethoprim resistant E. coli, several two-component systems are associated with resistance to antibiotics in different bacteria, often due to direct activation by the antibiotic [27]. Thus, this strategy for sensitization is likely to have wide application. Indeed, there are also potential advantages to targeting two-component sensor kinases for therapeutic purposes. Since they are not found in human cells, pharmacologically inhibiting two-component may produce fewer off-target effects in the host. Further, being cell envelope proteins with known ligands, inhibitor design may be facilitated. Finally, modulation of bacterial gene expression has not yet been targeted therapeutically and hence is unlikely to be confounded by pre-existing resistant mutants in clinical strains. These interventions may also serve the purpose of "evolution-proofing" of new antibiotics since they are likely to slow down the evolution of resistance as shown by us in this study and could be thought of as prospectively-used adjuvants.
Beyond the evolution of drug-resistance, our study contributes to the understanding of how gene regulatory networks evolve, particularly in the context of the emergence and maintenance of negative feedback. Mathematical modelling coupled with experiments using natural or synthetic genetic circuits have shown that negative feedback can reduce noise in gene regulatory networks [3,5,10,11,67]. Despite this knowledge, the importance of negative feedback from an evolutionary perspective remains poorly investigated. The only empirically validated basis for the evolution of negative feedback is to enhance mutational tolerance. Marciano et al. used the LexA transcription factor of E. coli to demonstrate that negative feedback canalises phenotypes and provides greater tolerance to mutational perturbation [68]. Similar observations were made for the Rox1 protein from Saccharomyces cerevisiae where negative feedback stabilised gene expression levels and enhanced mutational robustness [69]. For the PhoQP system, it has been speculated that MgrB may have evolved to optimise the signalling output of PhoQP in environments with varying Mg 2+ concentrations [34,36,37]. Our study provides experimental evidence for an alternative explanation for the evolution of negative feedback in PhoQP, i.e., to prevent pervasive dysregulation of the larger bacterial genetic network. In other words, we propose that negative feedback serves to insulate pathways that can potentially cross-activate, like PhoQP and RpoS transcriptional networks.
We note that wider application of the above idea to other bacteria is contingent on the pervasive gene regulatory effects of PhoQP in different bacterial species. Indeed, the PhoQP regulon has been worked out in a few species other than E. coli. We find that in all characterised systems, like in E. coli, PhoQP activation leads to gene expression changes that extend beyond the PhoP regulon. For example, in Pectobacterium versatile and Yersinia pestis, genome-wide transcriptomics and ChIP-seq have revealed that PhoP influences the expression of several genes beyond its direct targets, though the mechanisms aren't yet clear [70][71][72]. For E. coli, our study shows that RpoS overproduction and exclusion of RpoD from RNA polymerase are responsible for activating secondary and tertiary transcriptional programs in response to PhoQP hyperactivity. Interestingly, while Pectobacterium and Yersinia code for RpoS, they lack IraMlike proteins and hence in these species PhoQP may produce pervasive dysregulation of gene expression by other mechanisms [43]. Regardless, a broad gene-regulatory influence of the PhoQP pathway seems to be consistent across bacteria, supporting the idea that a key selection pressure for the maintenance of MgrB may be to limit cross-activation of other transcriptional programs by PhoP. It is important to note here that the requirement for negative feedback is highly contextual, at least in pathways like PhoQP that directly respond to the extracellular medium. Indeed, deregulation of PhoQP by loss of MgrB is itself beneficial in many environments such as acid stress and antibiotics [33,39,45,54]. The multiple instances of loss of mgrB across bacterial phylogenies observed by us in the present study may reflect different selection pressures driving evolution of the PhoQP system in different bacteria. Similarly, cross-activation of other regulatory pathways, though costly under the conditions tested by us, may have advantages in more complex environments. For the PhoQP/MgrB system from K. pneumoniae this idea has been recently proposed. Bray et al. (2021) [73] demonstrated that loss of MgrB in colistin resistant-K. pneumoniae is accompanied by compromised gastrointestinal colonisation efficacy. Like our results, this study too reported the overproduction of RpoS in mgrB-deficient K. pneumoniae. However, the authors reported that RpoS overproduction was beneficial in a model for pathogen transmission [73]. Thus, lifehistory and growth context are both likely to have a strong influence on how deregulation of gene expression translates to organismal fitness.
Curiously, the evolution of negative feedback proteins in two-component signalling pathways appears to be the exception rather than the rule. In E. coli only 2 systems, i.e. PhoQP and CpxAR have known negative feedback regulator proteins [74]. Our results may throw light on why this is the case. Based on the idea that pervasive gene dysregulation drives the evolution of negative feedback, we propose that the following criteria necessitate the evolution of negative feedback in two-component pathways. Firstly, the system must have a large regulon which increases the chance of its activity translating to changes in organismal fitness across environments. Secondly, the system should have a positive feedback loop, i.e., it should activate its own expression. This is known to be true for the PhoQP system and results in rapid amplification of signalling after activation. Finally, the regulon of the two-component system should be connected to other global regulatory networks, such as RpoS in the case of PhoQP. For E. coli, only 3 two-component systems satisfy all these criteria, namely PhoQP, CpxAR and ArcAB, of which 2 have known negative feedback regulators (Figure 9). In the case of ArcAB, we are not aware of negative feedback systems, however we believe that it may be reasonable to look for them in the future. We cannot rule out the role of regulatory RNAs here, several of which are known to be activated by two-component signalling [75]. Further investigation would be needed to test whether their contribution is similar to that of MgrB.
In conclusion, despite extensive mechanistic insights into the functioning of gene regulatory networks, there is relatively less known about how they evolve and are rewired in response to environmental perturbation. Using the PhoQP-MgrB system, we have analysed the immediate and distal effects of deregulation by loss of negative feedback on the cellular transcriptional network. By linking these changes with organismal fitness across relevant environments we have shown how evolution in a gene regulatory network proceeds in response to selection and can drive and modulate evolutionary adaptation.

Bacterial strains and culture conditions
E. coli wild type and its mutants were cultured in Luria-Bertani Broth (LB) or on Luria-Bertani Agar (LA plates). Media were supplemented with trimethoprim and MgSO4 at required concentrations as needed. Kanamycin or chloramphenicol for selection of genetically manipulated strains were added at 30 g/mL each as needed.
The strains used in this study are shown in Table 1.

Relative fitness measurements
Relative fitness (w) of various mutant strains of E. coli was measured by direct competition with an E. coli lacZ strain under appropriate growth conditions. Neutrality of the lacZ genetic marker was established by competitions with unmarked wild type E. coli under every growth condition tested. The detailed methodology followed is described in Patel and Matange (2021) [33].

Monoculture and mixed culture growth curves
Bacterial strains to be characterised were initially grown overnight to saturation. From saturated cultures, bacteria were passaged (0.1%) into 5 mL LB broth and grown at 37°C, with shaking at 180 rpm. For competitive growth curves, 2.5 μL each of the competing strains were inoculated in 5 mL LB.
Aliquots were taken for measuring bacterial growth periodically until 20-24 hours of growth. Bacterial growth was monitored using optical density (OD) at 600 nm for monocultures and by viable counts (CFU/mL) for mixed culture. For viable counts, aliquots of bacterial cultures were serially diluted and plated on LA supplemented with IPTG (50 g/mL) and X-Gal (50 g/mL). Plates were incubated for 18-24 hours at 37 ºC and blue and white colonies were counted.

Laboratory evolution of trimethoprim resistance
The detailed methodology for laboratory evolution of trimethoprim resistance is described in Patel and Matange (2021) [33] and Vinchhi et al. (2023) [76]. Briefly, bacterial populations were grown in trimethoprim supplemented LB (low Mg 2+ ) or LB + 10 mM MgSO4 as required for 15-18 hours before passaging (1%) into fresh media. Bacteria were passaged for ~210 generations (6-7 generations per growth cycle) and aliquots periodically were frozen at -80 ºC for further analyses. Trimethoprim was used at a concentration of 100 ng/mL which corresponds to ~MIC/9.

Genome sequencing of laboratory-evolved trimethoprim resistant isolates
Genome sequencing and variant calling for evolved bacteria was carried out as described in Patel and Matange (2021) [33] and Vinchhi et al. (2023) [76]. Sequencing services were provided by Eurofins, India.

Genetic manipulations and gene knockout in E. coli
All gene knockouts were generated using P1 transduction by moving kanamycin-resistance marked gene deletions from donor strains taken from the Keio Collection [77,78] into appropriate recipient strains. Knockouts were confirmed using gene specific PCR on genomic DNA extracted from transductants. Detailed methodology is described in Patel and Matange (2021) [33].

Measuring trimethoprim resistance
Trimethoprim resistance of E. coli strains or trimethoprim-resistant isolates was measured using a broth-dilution assay. Briefly, appropriate strains of E. coli were grown to saturation and then inoculated into serially diluted trimethoprim-containing media. Growth was monitored after 18-20 hours of growth and IC50 values were determined by fitting experimental data to a variable-slope, 4 parameter model using Graphpad Prism (version 9.1.4). Detailed methodology is described in Patel and Matange (2021) [33] and Vinchhi et al. (2023) [76].

Establishment propensity of trimethoprim-resistant isolates at sub-MIC drug pressures
To estimate the propensity of establishment of trimethoprim-resistant strains, their ability to out-compete a large excess of wild type E. coli under sub-MIC trimethoprim selection was evaluated. To set up the primary competitions, trimethoprim-evolved mutant isolate and wildtype E. coli were mixed (1:10 7 ), and the mixture was diluted 100 times in a final volume of 15 mL LB broth supplemented with 100 ng/mL trimethoprim. This diluted culture was then dispensed into the wells of a sterile 96-well polystyrene plates (200 L per well), and incubated at 37°C for 24 hours, with shaking at 180 rpm. Next, 1 μL from each well was passaged into 200 μL of LB broth containing 1 μg/mL trimethoprim (i.e. MIC of wild type [33]) in a fresh 96-well plate. This plate was incubated at 37°C for 24 hours with shaking at 180 rpm. Only those wells in which mutant bacteria were enriched by >10-fold would show turbidity. The number of wells showing visible growth were noted, and OD600 was measured using a plate reader. Two controls were also set up in parallel. For the first control, primary competition was performed in the absence of trimethoprim such that there would be no enrichment of resistantmutants. The second control had similar number of wild-type cells without the addition of mutant bacteria. This control would indicate the frequency of spontaneous resistant mutants emerging during the competition. OD600 values of 144 replicate competitions for each condition were plotted and compared.

Sequence and phylogenetic analyses
To analyse the distribution of homologs of PhoQ and MgrB from E. coli across different bacterial species, the Pfam database [43] was initially mined. Since both proteins were restricted to the Order Enterobacterales further analyses were restricted to this group of bacteria. The phylogenetic relationships of different families under Order Enterobacterales was taken from Adelou et al., 2016 [44]. Type strains from each family were identified using the LPSN (List of Prokaryotic names with Standing in Nomenclature; https://lpsn.dsmz.de/) database [79,80]. For family Enterobacteriaceae, a maximum likelihood phylogenetic tree was constructed using 61 bacterial species in MEGA-X [81], by using 16s rRNA gene sequences of the type strain of each genus as listed in LPSN [79,80]. The presence or absence of PhoQ and MgrB in bacterial strains under consideration was determined by three methods: first, directly analysing genome annotation data available on NCBI, second, by performing BLASTn against the genome of the query organism (nucleotide similarity with wild-type E. coli mgrB-NCBI Gene ID: 946351 and phoQ-NCBI Gene ID: 946326), third, by performing NCBI BLASTp against the proteome of the query organism (amino acid similarity with wild-type E. coli MgrB-Uniprot: P64512 and PhoQ-Uniprot: P23837). A positive hit in at least one of the above searches was taken to mean that PhoQ/MgrB were present in the query organism.

Transcriptome analyses using RNA-sequencing
For RNA-seq, 1% of a saturated culture of wild type and trimethoprim resistant isolates Tmp R -A, Tmp R -B and Tmp R -C was inoculated into 3 mL LB in duplicate and incubated at 37 ºC for 3 hours with shaking at 180 rpm. Bacteria were pelleted down using centrifugation and resuspended in 1 mL of RNAlater (Invitrogen) for storage. Subsequent RNA extraction, sequencing and preliminary data analyses were performed by Redcliff Life Sciences (India). Total RNA was extracted using RNAeasy spin columns and quantitated using Qubit and Bioanalyzer. Bacterial rRNA was depleted using Ribozero kit. Bacterial mRNA was then reverse transcribed, library was prepared and quality control was performed using Tapestation platform. Paired end sequencing was performed on Illumina platform with read lengths of 150 bp read length. Processed reads were aligned to the reference genome (NZ_CP025268) using HISTAT2(version 2.1.0). Abundance estimation was done using featureCounts( version1.34.0). Differential gene expression (DGE) analysis was done by comparing the expression of individual genes trimethoprim-resistant strains to wildtype using DESeq2. The output of the DGE analyses were log2(fold change) values for each gene and P-values to test for statistical significance. Differentially expressed genes were classified into PhoP-regulated, RpoS-regulated and RpoD-regulated based on information available in RegulonDB [47,48] and Ecocyc [46]. Sigma factors regulating differentially expressed genes were also obtained from RegulonDB [47,48] and Ecocyc [46].

Gene specific quantitative RT-PCR
Quantitative RT-PCR was used to validate the findings of RNA-seq. For RNA extraction, appropriate bacterial strains were grown for 3 hours at 37 ºC and pelleted by centrifugation. Total RNA was extracted using TRIzol Reagent (Invitrogen, USA) and quantified spectrophotometrically. RNA quality was evaluated by electrophoresis on a 1% agarose gel and staining with ethidium bromide. Extracted RNA (20µg) was treated with recombinant DNase I (RNAse free) (Takara, Japan) and then cleaned-up using RNeasy spin column (Qiagen, Japan). Prepared RNA was stored at -80 ºC until further use. Reverse transcription reaction was set up using the cleaned-up RNA ( ̴ 1.6µg) using PrimeScript™ RT reagent Kit (Takara, Japan). The RT reaction was carried out with buffering temperature at 25 ºC for 2 mins, cDNA synthesis at 37 ºC for 30 mins. The RT enzyme was inactivated at 85 ºC for 90 seconds. Prepared cDNA was serially diluted (10, 100 and 1000-fold) and semi-quantitative PCR were set-up using gene-specific primers ( Table 2) to ensure the absence of contaminating genomic DNA contamination and checking for priming efficiency. cDNA was stored at -20 ºC until further use. Quantitative RT-PCR (qPCR) for target genes was performed using TB Green®Premix Ex Taq™ II (Tli RNaseH Plus) (Takara, Japan) and using gene-specific primers listed in Table 2.
For each PCR, 20 µL reaction contained 10 µL of 2X TB Green Premix, 0.5 µl each of forward and reverse primers (10 µM stock), 8µl of nuclease-free water and 1 µL of appropriately diluted cDNA. The qPCR reactions were performed on an Eppendorf Realplex2 Mastercycler (Eppendorf, Germany) using a two-step protocol with an initial denaturation at 95 ºC for 30 seconds, 40 cycles of denaturation at 95 ºC for 5 seconds and annealing and extension for 30 seconds at 60 ºC. A melt curve analysis was performed after 40 cycles of qPCR. 16S rRNA was used as a normalizing internal control to calculate change in expression for all other genes. A known concentration of genomic DNA and its dilutions were used to construct a standard graph for each gene. Fold changes in gene expression for each gene were calculated with respect to the wild type using the standard graph method. Table 2. List of primer, cDNA dilution and expected product sizes for qPCR     Representation of the phylogenetic relationships between 7 bacterial families that make up the Order Enterobacterales, taken from Adelou et al., 2016 [44]. The presence of mgrB and phoQ genes in representative members of these families are shown as yellow and purple circles respectively. Most probable events of gain of the PhoQP system and loss of MgrB are indicated. B. Maximum Likelihood phylogeny of 61 representative species belonging to Family Enterobacteriaceae constructed using 16S rRNA sequences is shown. The presence of phoQ and mgrB are indicated by purple and yellow circles next to each species. Lack of annotated genome data is indicated by a star. Possible events of mgrB-loss are shown as red arrows.

Figure 5. Loss of negative feedback is costly for E. coli due to PhoQP hyperactivity A.
Growth curves of wild type E. coli and an isogenic mgrB-knockout strain (mgrB) under standard growth conditions in LB. Growth was measured using OD at 600 nm. Mean ± SD from 3 independent experiments is plotted. B. Diagrammatic representation of the PhoQP signaling pathway (not to scale). Direct and indirect positive regulation by PhoP is shown as solid and dashed green arrows respectively. Inhibitory interactions are shown in red. AlphaFold models/crystal structures are used to represent proteins in the pathway that are relevant to this study. Source of the fitness cost of mgrB-deficiency based on our study is shown. C and D.  represent mgrB and triangles represent rpoS. Gray symbols represent wild type alleles, while coloured symbols represent mutant alleles as shown in the key. Comparisons between strains reveal the effects of mgrB and mgrB+rpoS mutations on the global transcriptome of E. coli as shown. B. Comparison of differential gene expression (DGE) between trimethoprim resistant strains Tmp R -A and Tmp R -B. For each strain DGE was estimated as log2 of fold change compared to wild type (log2FC). Each point on the scatter plot represents expression level of a single gene. C. Comparison of differential gene expression (DGE) between trimethoprim resistant strains Tmp R -B and Tmp R -C. For each strain DGE was estimated as log2 of fold change compared to wild type (log2FC). Each point on the scatter plot represents expression level of a single gene. For genes showing greater than 4-fold down-regulation in Tmp R -B or Tmp R -C compared to wild type, data points are colored by the sigma factor responsible for transcription as indicated based on data available on Ecocyc database [46]. D. Expression level of 58 genes of the PhoP regulon (gene list obtained from RegulonDB [47,48]) in Tmp R -A, B and C strains represented as pie-charts. Genes with at least 2-fold higher or lower expression level compared to wild type were classified as "up-regulated" or "down-regulated" respectively. E. Expression level of 14 genes co-regulated by PhoP and RpoS in Tmp R -A, B and C strains shown as a box plot (min to max). The median expression level is shown by the line. F. Validation of RNA-sequencing based transcriptomics using quantitative RT-PCR (qPCR) for selected genes. Mean ± SD of fold change compared to the wild type from three independent biological replicates are plotted. No change in expression level compared to wild type corresponds to a value of '1' and is shown by a dotted line. Gene names are indicated on the X-axis. Different colours represent different strains from which measurements were made as shown in the key below the graph. G. Heat map showing expression level of RpoD-target genes in Tmp R -A, B and C strains based on RNA-seq. For each strain, expression levels were compared to wild type and log2(FC) is represented on a continuous red-green colour scale as shown.

Figure 7. Loss of RpoS-RpoD balance explains the fitness costs of mgrB-deficiency A.
Relative fitness (w) of E. coli wild type and indicated gene knockout strains compared to an isogenic lacZ-knock out reference strain in antibiotic-free media. Mean value from three independent experiments are plotted as bars. Error bars represent standard deviation. A value of '1' indicates no change in fitness compared to the reference strain and is shown as a dotted line. Yellow bars represent strains that were used to test the contribution of RpoS-targets to the costs of mgrB-deficiency, while green bars represent strains that were used to test the contribution of RpoS-RpoD imbalance to the costs of mgrB-deficiency. Statistical significance between the relative fitness of double mutants and mgrB was tested using a Student's t-test (p-values shown above relevant bars). B. Growth of wild type, mgrB or mgrBrpoS (black) in competition with an isogenic lacZ strain (gray) over time expressed as CFUs/mL. Mean ± SD from three independent biological replicates is plotted. Student's t-test was used to compare statistical significance of the differences in bacterial densities of the competing strains (pvalues are shown next to relevant data points).  Green plus sign (+) indicates that the two-component system is known to have positive feedback, i.e. stimulate its own transcription. Global regulators that these systems are known to cross-activate are in crimson next to respective response regulators, while known negative feedback regulatory proteins are in black. All information collated from Ecocyc [46].