Potentiating antibiotic efficacy via perturbation of non-essential gene expression

Selecting gene targets and antibiotic concentrations

We utilized the ample evidence in literature of genes associated with stress response and/or adaptation processes to narrow down a set of promising candidates for potentiating antibiotic activity²¹. Many studies have investigated E. coli response to various antibiotics, including the use of whole genome library knockouts^22,23,24. Of the 4000+ genes in E. coli, these studies have identified a set of ~100-300 genes that influence the bacteria’s sensitivity to multiple antibiotics. This includes the tolC and acrA genes comprising the TolC-AcrAB efflux pump of E. coli, which has been frequently explored as a hub for targeting antibiotic resistance in these and many other studies^{25,26,27,28,29}. Similarly, the transcriptional regulator marA, soxS, and rob have each been implicated in increasing E. coli antibiotic resistance^30,31,32. Mutants of genes involved in the SOS response such as recA, dinB, polB, and dam are known to strongly influence E. coli response to antibiotics^{33,34,35,36,37,38}. Additionally, a set of genes have been show to exhibit increased transcriptional activity during exposure to stressful environments in general, including rpoS, mutS, hfq, and cyoA^{39,40,41,42,43,44}.

We have previously performed our own transcriptional analysis of bacteria adapting to stress response. Several genes were found to impact adaptive resistance in a transcriptome-level analysis of adapted versus unadapted strains (fiu, tar, wzc, yjjZ were differentially expressed while ybjG, ydhY, ydiV, and yehS were differentially variable)^19,20. In these studies, we also investigated the transcriptional regulators that control the genes exhibiting differential expression and narrowed down a set of novel targets (including bglG, crp, csgD, flhC, flhD, fnr, fur, gadX, and phoP). We reasoned that because these genes (or the downstream genes they regulate) exhibited particularly high gene expression variability during adaptation, they would serve as interesting candidates for targeting expression interference therapies towards.

Combining these two approaches, we developed a final set of thirty selected genes to explore as potential targets for antibiotic combination therapy (Table 1). These genes represent diverse functionalities, including transport (acrA, tolC, and fiu), mutagenesis (mutS, dam, polB, dinB), motility (tar, flhC, flhD, and ydiV), uncharacterized function (yjjZ and yehS), or general global regulation (the aforementioned others). Knockouts of gene were obtained from the Keio collection¹⁸. As E. coli is viable upon deletion of these genes, we reasoned that reducing expression of these non-essential genes is less likely to impose an inherent fitness cost compared to reducing expression of essential genes.

We chose to test these knockouts’ growth in a set of nine antibiotics representing a diverse set of common antimicrobial therapies (Table 2). Growth assays were first performed to determine a suitable dosage for each antibiotic agent below the minimum-inhibitory concentration (MIC) (Fig. S1). Antibiotic concentrations resulting in a 10–50% fold reduction in growth were identified and used for all subsequent testing⁴⁵.

Gene knockout synergy with antibiotic treatment

We characterized BW25113 growth in 270 combinations of 30 Keio knockouts with these nine antibiotics. Synergy (S) between gene knockout and antibiotics was determined using the Bliss Independence model, S = W_X*W_Y – W_XY, based on bacterial fitness in the presence of antibiotic (W_X), gene knockout (W_Y), or a combination of both (W_XY). An example of this is presented in Fig. 1a. Here, ampicillin (AMP) exhibited no significant impact on BW25113’s fitness (W_X = 1.04 ± 0.06), nor did deletion of acrA (W_Y = 0.93 ± 0.03). However, BW25113-ΔacrA exposed to AMP demonstrated poor fitness (W_XY = 0.07 ± 0.02), indicating potentiation of AMP activity by removing acrA (S = 0.89 ± 0.12). This process was repeated for all 269-remaining gene–drug pairs (Fig. 1b). Drug independent (additive) and drug dependent (synergistic or antagonistic) interactions were classified using a two-sided t-test, after log transformation of the data, as described by Demidenko et al.⁴⁶. If the interaction was statistically significant (P < 0.05) then a one-sided t-test further determined synergism (P < 0.05) or antagonism (P > 0.05).

We identified 90 gene–drug interactions which elicited synergistic interactions, and another 67 gene–drug interactions that resulted in antibiotic antagonism (Fig. 1b). Of the thirty genes explored, five (acrA, fur, recA, rpoS, and tolC) were found to potentiate activity of at least six out of nine antibiotics tested, suggesting that these genes are promising targets for co-therapies. Many of these gene knockout drug synergies, to the best of our knowledge, are novel discoveries, such as that between fnr and five of the antibiotics tested. Conversely, two gene knockouts (gadX and hfq) caused antagonism of at least seven of nine antibiotics and are thus poor candidates for antimicrobial potentiation purposes.

The goal of this systematic investigation was two-fold, the first of which was to identify strong gene–drug synergistic interactions. The second goal was to ensure that these knockouts imposed minimal fitness costs on cell growth, to protect against the possibility of natural selection working against sequence-specific therapies designed to target these genes. In order to avert development of gene knockdown therapies that pose a direct impact on cellular fitness in the absence of antibiotic exposure, we avoided gene knockouts that had particularly strong impacts on growth. This includes four gene deletions: Δdam (W_Y = 0.49 ± 0.28), Δrob (W_Y = 0.48 ± 0.05), Δhfq (W_Y = 0.38 ± 0.04), and Δtar (W_Y = 0.41 ± 0.05). Δhfq and ΔgadX demonstrated consistent antagonism and were thus already not candidates for further screening.

Collectively, these results point to promising targets for designing fitness-neutral gene expression perturbations that enhance antibiotic efficacy. We explore the development of such therapies utilizing CRISPR-Cas9 in a latter portion of the manuscript. However, this knockout-drug synergy screen also provides interesting conclusions regarding the general mechanisms of gene–drug synergy that we will now explore.

Unraveling mechanistic insights from knockout-drug synergy

We first explored the applicability of a hypothesis raised in our previous work⁴⁷ stating that the degree of gene–drug synergy is influenced by the epistatic interactions the targeted gene is involved with. In particular, we have found that as more protein–protein interactions are disrupted by targeted genetic perturbations, a greater than expected detrimental fitness impact emerges. We applied a similar approach in this study to explore the known protein interaction networks of each of the thirty genes here to unravel similar underlying correlations. From the STRING database⁴⁸, we extracted information of all the known and predicted protein–protein interactions for each gene to construct and quantify their protein interaction networks. We explored the total number of proteins interactions, i.e., the amount of nodes that are present (Fig. S2). We found that significantly greater synergy was observed with increased number of proteins interacting (P = 0.006). While correlation in itself does not prove a meaningful connection, its existence suggests that targeting nodes involved in broad protein interaction networks can lead to enhanced levels of gene–drug synergy. Additionally, two other studies from our lab have also found a similarly significant correlation between synergy and protein interactions^17,47. Our previous work demonstrated that perturbing genes involved in broad protein interaction networks lead to strong negative epistatic effects. The data presented here supports this observation and lends further credence to the notion that epistasis plays a significant role in influencing bacterial fitness response towards co-therapies.

To further explore the mechanistic underpinnings of gene–drug interactions, antibiotics and genes were grouped into mechanisms of action and pathways respectively (Fig. 2). Notably, the one knockout directly affecting metabolism, Δwzc, represented one of the top three synergistic knockouts in ceftriaxone (CRO), erythromycin (ERY), and ciprofloxacin (CIP), but was also one of the three most antagonistic knockouts in sulfadimidine (SDI) and trimethoprim (TMP). Whole genome RNA-sequencing showed that wzc, a colanic acid biosynthesis gene, was overexpressed during AMP exposure²⁰, although no significant synergy was observed between Δwzc and AMP in this experiment. Though the classes of antibiotics in which synergy was observed were diverse, clear antagonism emerged in the antibiotics related to DNA/RNA synthesis (sulfadimidine and trimethoprim).

The TolC-AcrA efflux pump knockouts demonstrated some of the highest levels of gene–drug synergy. Knockouts of at least one of these genes was always one of the three most synergistic genetic changes for all the antibiotics tested, apart from the 50S targeting antibiotics ERY and chloramphenicol (CM). However, even in these antibiotics, both knockouts resulted in significant synergy. The remaining standout knockouts include ΔrecA, Δdam, and ΔrpoS, which demonstrated high synergy with four, two, and two antibiotics, respectively. There was no clear relationship between the cellular processes these genes are involved in and the antibiotics’ modes of action.

Introducing gene–drug synergy using CRISPRi

If knocking out these genes resulted in significantly amplifying antibiotic potency, we hypothesized that lowering their expression without completely removing them from the genome might engender similar results while also demonstrating a more therapeutically viable application of gene–drug synergy. To this end, we developed CRISPR interference (CRISPRi) constructs to knockdown expression of genes showing significant antibiotic synergy. For this, catalytically dead Cas9 (dCas9) was employed to reduce mRNA production of the targeted gene, which was subsequently exposed to a variety of antibiotics to determine its potential for inducing gene–drug synergy (Fig. 3a).

We specifically focused on six of the genes showing the greatest degree of synergy with each of the antibiotics tested, while also maximizing the diversity of genetic pathways targeted for synergy with each antibiotic (Table S1). We utilized a dual-plasmid system based on the original CRISPRi system to deliver gene knockdown constructs to BW25113⁴⁹. One plasmid encoded expression of dCas9 under the anhydrous tetracycline (aTc)-inducible promoter, while the other encoded a unique single guide RNA (sgRNA). sgRNA targets were designed using criteria set forth in previous studies for successfully eliciting inhibition^13,50. This includes targeting either within the first ~50 nucleotides of the gene’s open reading frame, or within the −35 to +1 site of the gene’s promoter sequence. The ability of six of these constructs (marA-i, recA-i, acrA-i, tolC-i, soxS-i, and wzc-i) and other similarly designed constructs to inhibit gene expression have been verified in our previous studies using real-time quantitative PCR^11,19,47. We also analyzed all potential off-targets for each of these constructs based on what constitutes most likely off-targets in E. coli, and outline each of them in Table S2^51,52. Of these, only four constructs (soxS-i, tolC-i, ydhY-i, and phoP-i) had potential off-targets of genes whose deletions are known to cause fitness defects. As the two plasmids used to express the CRISPRi constructs rely on AMP and CM selection markers, we excluded exploration of these antibiotics going forward. Additionally, due to the general low degree of synergy demonstrated by gene knockouts with SDI, this antibiotic was not included.

The gene-antibiotic synergy experiments were again repeated, with CRISPRi employed in the place of gene knockouts. A strain expressing a sgRNA targeting the coding sequence of red fluorescent protein (rfp) (which is not present in the strain) was used in lieu of wildtype BW25113 as the control. All fitness values were normalized to the growth of the control strain in the absence of any antibiotic. Fitness of the control strain was measured during exposure to each antibiotic (W_X), each individual perturbation strain in the absence of antibiotic (W_Y), and each perturbation strain during antibiotic exposure (W_XY). Experiments were again performed in M9 minimal media supplemented with 0.4% glucose and antibiotics as appropriate.

Notably, the majority of our CRISPRi constructs appeared to slightly improve growth over the rfp targeting control in the absence of antibiotics, as indicated by W_Y values greater than 1.0. This phenomenon likely stems from a fundamental slight growth impact caused by the rfp targeting construct. One plausible explanation of this is that the rfp targeting construct was more prone to off-targeting effects due to the lack of an on-target sink to draw the dCas9-sgRNA complex towards. A recent report systematically examined the off-targeting potential of dCas9 in E. coli and found that binding of as little as five bases in the seed sequence is sufficient to have a measurable impact on transcription⁵². We performed a systematic exploration of the most likely off targets for this rfp-i control construct (Table S2). The most likely off-targets with potential fitness impacts include cysG, ftsI, and metL. We also investigated the off-targets of our other CRISPRi constructs. However, as each of these constructs has an actual on-target, and none presented a W_Y lower than 1.0, the influence of off-targets is likely trivial. These results suggest that a better control for future studies utilizing CRISPRi would utilize a DNA sequence present in the genome but known to be uninvolved with fitness.

Despite this caveat for our control, the majority of CRISPRi perturbation co-therapies resulted in statistically significant potentiation of antibiotic treatment, suggesting that this was an effective strategy for replicating revealed knockout-drug synergy (Fig. 3b). Four perturbations improved efficacy of ERY (acrA-i, tolC-i, cyoA-i, and wzc-i), three synergized with TMP (acrA-i, tolC-i, and recA-i) and CIP (rpoS-i, recA-i, and fnr-i), and two synergized with puromycin (PURO) (acrA-i and tolC-i) and tetracycline (TET) (acrA-i and tolC-i). Three combinations with TET resulted in clear antagonism: rpoS-i, yehS-i, and crp-i. A few CRISPR perturbations did stand out from the rest in the clear antibiotic synergy they induced. Most notable is the degree of synergy induced by inhibitions of the tolC-acrA efflux pump in ERY (acrA-i = 0.36 ± 0.16, tolC-i = 0.36 ± 0.15), PURO (acrA-i = 0.69 ± 0.40, tolC-i = 0.64 ± 0.34), TET (acrA-i = 0.62 ± 0.41, tolC-i = 0.49 ± 0.25), and TMP (acrA-i = 0.67 ± 0.18, tolC-i = 0.60 ± 0.16). Inhibitions of recA and fnr additionally showed significant improvements in CIP efficacy (S = 0.85 ± 0.18 and 0.96 ± 0.17, respectively). CRISPRi largely replicated gene knockout synergy with antibiotic treatment. A direct comparison of the synergy levels is presented in Fig. S3. A few perturbations produced greater synergy in the perturbation context than in the knockout context, including both acrA-i and tolC-i combined with ERY treatment, and both fnr-i and recA-i combined with CIP treatment. Comparing synergy values between gene knockouts and CRISPRi gene perturbations showed a strong correlation (P = 0.003) Additionally, the average synergy across all perturbations was lower than the average synergy of corresponding gene knockouts during exposure to CRO (CRISPRi = −0.02, Δ = 0.23), TET (CRISPRi = 0.10, Δ = 0.38), ERY (CRISPRi = 0.15, Δ = 0.36), PURO (CRISPRi = 0.16, Δ = 0.53), and TMP (CRISPRi = 0.23, Δ = 0.49). Lower levels of perturbation-antibiotic synergy relative to knockout-antibiotic synergy are unsurprising given that the target gene can still be expressed (albeit at a lower level) in the former case.

To have a better understanding of these perturbations on antibiotic efficacy, we also performed growth curve analysis of each CRISPRi strain with their corresponding antibiotics in an alternative environment of LB media (Figs. S4–9). In most cases, synergy between CRISPRi and antibiotics is made apparent in the early stages (5–10 h) of growth. Significant growth inhibition was caused by CRISPRi synergism with the associated antibiotic in the following conditions: four with CRO (tolC-i, ydhY-i, phoP-i, and marA-i), five with CIP (tolC-i, recA-i, rpoS-i, wzc-i, and fur-i), four with TET (tolC-i, rpoS-i, crp-i, and csgD-i), two when targeting ERY (tolC-i and acrA-i), three when targeting TMP (tolC-i, acrA-i, and ydiV-i), and three when targeting PURO (tolC-i, recA-i, and acrA-i). While the majority of perturbations induced antibiotic synergy in at least one of the two environments tested, a few perturbations provided either no synergism or was antagonistic and are therefore not useful candidates for potential therapies. This includes targeting crp during CRO treatment, yehS during TET treatment, and both wzc and blgG during PURO treatment. These data sets provide quantification of gene-drug synergies in two environments. Taken together, these results highlight that CRISPRi can effectively potentate antibiotic treatment.

Multiplexing CRISPRi exacerbates antibiotic synergy

An advantage of CRISPRi is the relative ease in which individual perturbations can be combined into a single cell by including multiple sgRNAs in a CRISPR array. Furthermore, we have previously shown that multiplexing perturbations tends to exacerbate detrimental fitness impacts by inducing negative epistatic interactions between the perturbed genes⁴⁷. This suggests that multiplexing synergistic CRISPRi perturbations could exacerbate the potentiation of antibiotic efficacy. We took advantage of this by combining the six perturbations designed for each antibiotic into one construct and testing their impacts on BW25113 growth during antibiotic exposure.

We first ensured that expanding the number of perturbations did not have an inherent impact on growth by testing the growth of a control strain harboring six tandem rfp perturbations (Fig. 4a). This strain exhibited no significant shift in basal fitness, nor did it show antagonism or synergy with any antibiotic.

In stark contrast to this, every multiplexed CRISPRi perturbation strains designed for inducing synergy showed significant potentiation of antibiotic efficacy, apart from PURO perturbations (Fig. 4a). Synergy was particularly pronounced with TMP (0.25 ± 0.14, P = 4E−12), CIP (0.22 ± 0.48, P = 0.01), and ERY (0.20 ± 0.16, P = 3E−8). The strong synergy observed by multiplexed TET perturbations (0.13 ± 0.09, P = 7E−10) is particularly notable given the varied levels of synergy observed when perturbations were applied individually.

To further elucidate multiplexed perturbations’ impacts on BW25113 growth, we examined each strain’s growth profile over 20 h in both the presence and absence of antibiotic and compared these profiles to the multiplexed control perturbation strain (Fig. 4b). All strains grew identically to the control in the absence of antibiotic exposure, apart from a slight lag-time shift in the multiplexed perturbation under TET treatment. In the presence of antibiotics, multiplexed perturbation strains demonstrated diminished growth capacity compared to the control strain, apart from multiplexed PURO perturbations. For PURO multiplexed perturbations, one possibility for why multiplexing was less effective than individual gene targeting is the dilution of dCas9 protein when targeting multiple genomic loci simultaneously, leading to less gene repression. We have previously demonstrated that multiplexed gene perturbations resulted in similar levels of gene repression using individual gene perturbations^11,47. Furthermore, this potential problem would be alleviated in a therapeutic context by utilizing direct delivery of dCas9-sgRNA pre-assembled complexes. Regardless, these results support the conclusion that the designed multiplexed perturbations largely potentiate antibiotic treatment without imposing a substantial, direct impact on fitness, and provided direction as to which perturbation-antibiotic pairs to focus on for the remainder of our experiments.

To further confirm fitness impacts of multiplexed perturbations, we employed a competition assay on the two multiplexed perturbations exhibiting the greatest degree of synergy and strongest impact on growth: the strains designed for synergizing with TET and TMP. In this assay, perturbed strains were co-cultured with the control strain, and the relative abundance of each strain was determined at the beginning and end of competition. For this, the control CRISPRi strain harboring rfp perturbation was modified to constitutively express mCherry. The relative abundance of each multiplexed perturbed strain was equivalent to that of the control strain in the absence of antibiotic exposure, confirming that these perturbations have no direct impact on fitness (Fig. 4c, d). The results noticeably changed when antibiotics were included; the relative abundance of each multiplexed perturbed strain dropped significantly, indicating a substantial impact on bacterial fitness. Synergy was calculated using the same equation as before, replacing ODs with direct measurements of viable colony-forming units (CFUs). This revealed statistically significant potentiation of antibiotic efficacy (W_XY = 0.64 ± 0.22, P = 1E−4, and W_XY = 0.61 ± 0.10, P = 1E−4 for TET and TMP multiplexed perturbations, respectively) (Fig. 4c, d). Collectively, these results demonstrate that multiplexed perturbations further potentiate antibiotic efficacy while minimizing direct fitness impacts.

CRISPRi potentiates antibiotic efficacy in infection models

A benefit of the CRISPRi strategy for enacting sequence-specific gene therapies is the relative ease with which it can be applied to a vast array of organisms. For instance, many of these CRISPRi constructs can be directly applied to a pathogenic relative of E. coli, the bacteria Salmonella enterica (Fig. 5a). An analysis of the genome of S. enterica serovar Typhimurium SL1344 shows that six sgRNAs designed for BW25113 have complete (acrA, cyoA, and fnr) or near-complete (crp, rpoS, and tolC) homology to SL1344’s genome, and therefore are likely to maintain efficacy in this organism (Fig. 5b). SL1344 is a model organism for studying bacteria in intracellular infections due to the relative ease in which it infects human cell lines⁵³. To explore the potential for gene expression perturbations to potentiate antibiotic efficacy in a therapeutic context, we created two new sgRNA plasmids: one harboring all six targets, and another harboring just the three with perfect homology. These plasmids, as well as the single rfp-targeting control plasmid, were co-transformed into SL1344 with the Cas9 expression plasmid, and antibiotic synergy was again explored.

No detrimental impact on basal SL1344 fitness was observed in the absence of antibiotics for either of these strains (W_Y = 1.11 ± 0.18 and W_Y = 0.97 ± 0.07 for the three and six perturbations respectively) (Fig. 5c, d). Significant synergy was again observed in several instances. Both multiplexed perturbed strains showed significant synergy with CRO and TMP, the latter of which appeared to be particularly potentiated (S = 0.12 ± 0.30, P = 0.01 and 0.21 ± 0.15, P = 3E−10 for three and six perturbations respectively). Strong synergy was also observed between ERY and the six-perturbation strain (0.29 ± 0.13, P = 1E−8). Additionally, antagonism was observed between the three and six perturbations and CIP (−0.15 ± 0.31, P = 0.01 and −0.07 ± 0.13, P = 5E−3, respectively), which could be due to the antagonism demonstrated by ΔacrA and ΔcyoA in BW25113. Antagonism was also seen in the three-perturbation strain combined with ERY (−0.21 ± 0.31, P = 3E−3).

Going forward, we focused our efforts on characterizing these perturbations’ impacts on TET and TMP, as these antibiotics showed high levels of synergy. The growth profiles of each strain were characterized in the presence of no antibiotic, TET, or TMP (Fig. 5e). No impact on growth was observed in the absence of antibiotic exposure, while detrimental impacts were observed for the perturbed strains during antibiotic exposure. This again indicates that perturbations imposed no direct fitness cost while still potentiating antibiotic treatment of SL1344.

To investigate the ability of perturbations to potentiate antibiotic clearance of intracellular infections, HeLa epithelial cells were infected with each SL1344 strain. Infected HeLa were subjected to no antibiotic, 0.5 µg/mL TET, or 0.5 µg/mL TMP for 18 h of post infection. HeLa were subsequently lysed, and CFUs of intracellular SL1344 were determined. The surviving SL1344 in the presence of antibiotic were compared to the relative surviving Salmonella in the absence of antibiotic. Significant reductions in viable SL1344 were observed in the presence of TMP for both the three-gene (P = 0.04) and six-gene perturbation (P = 0.03) strains (Fig. 5f, g). This was also true of the six-gene perturbation strain’s growth in presence of TET (P = 0.008). These results indicate that the targeted multiplexed CRISPRi constructs successfully potentiated intracellular antibiotic treatment, supporting the therapeutic viability of fitness-neutral gene perturbation treatments.

PNA knockdown of gene expression potentiates antibiotic treatment of MDR clinical isolates

To further explore the therapeutic potential of fitness neutral gene perturbations, we utilized an alternative gene expression knockdown approach based on PNA. The structure of PNA and DNA are similar, and the ability of PNA to bind to RNA has been well established¹². When conjugated to CPPs, PNAs can readily cross bacterial membranes and enter the cell. When these PNAs enter the cell, they form tight bonds with complementary mRNA, preventing ribosome translation of these genes into proteins (Fig. 6a)¹³. This approach can be readily applied to a wide array of bacteria to induce gene expression knockdown in a plasmid-independent fashion.

We applied PNA knockdown of gene expression to four clinically isolated, MDR bacteria obtained from the University of Colorado Anschutz Medical Campus. Each strain has been previously sequenced and found to exhibit a wide range of antibiotic resistances^17,54,55,56. This includes two strains of MDR E. coli, one of which exhibits a carbapenem-resistant Enterobacteriaceae (CRE) phenotype, and two strains of Klebsiella pnuemoniae (KPN) producing either an extended spectrum β-lactamase (ESBL) or a New Delhi metallo-β-lactamase 1 (NDM-1). These strains have been found to survive a wide range of antibiotic concentrations significantly above the resistance breakpoint levels established by the Clinical & Laboratory Standards Institute (CLSI), including AMP, CIP, CM, TET, kanamycin (KAN), rifampicin, streptomycin, gentamicin (GEN), and clindamycin⁸. Additionally, we found each MDR strain to resist TMP levels well above the CLSI breakpoint of 2.0 µg/mL, while wildtype BW25113 exhibited sensitivity at 0.25 µg/mL (Fig. 6b and Fig. S10).

We chose to focus specifically on synergy with TMP, as we achieved the greatest success in engineering synergy in a fitness-neutral fashion with this antibiotic in our CRISPRi approach. We first screened the genomes of the four MDR isolates for homology with each of the six TMP related gene perturbations tested with CRISPRi. Low homology was found for ydiV, so we chose to exclude testing of this gene. The remaining five genes showed significant homology with the four MDR strains. However, we chose to exclude testing tolC as well, as multiple off-targets were found. Additionally, the TolC-AcrAB efflux pump has been targeted by similar expression interference techniques^25,57. Excluding a tolC PNA minimizes redundancy, while allowing us to focus on other more novel targets. We designed 12 nucleotide (nt) long PNA molecules to inhibit expression of the remaining four genes (acrA, csgD, fnr, and recA) by targeting them to overlap the start codon of each gene’s open reading frame. Target sequences are presented in Table S3, and any potential off-targets identified in a genome-wide screen are noted in Table S4. A control PNA targeting a nonsense sequence not present in any of the genomes was also designed. Testing the impact of this nonsense PNA on the growth of each MDR bacterial strain revealed that it had minimal impact on growth in the presence of TMP, indicating that PNA-CPP molecules have no inherent detrimental impact on MDR bacteria growth independent of the effects caused by targeted gene repression (Fig. S11).

We next examined the ability of the four targeted PNAs to synergize with TMP treatment in each of the four MDR bacterial strains (Fig. 6c, d). Again, TMP treatment alone demonstrated no effect on growth. Three PNAs exhibited impacts on MDR bacterial growth in the absence of TMP; the recA and fnr targeting PNA minimally reduced growth (<10% reduction) in both E. coli strains, while the csgD PNA minimally reduced growth in both E. coli strains and significantly reduced growth (>10% reduction) in ESBL KPN.

For four out of the five instances where PNA showed a small (<10% reduction) impact on growth there is a greater than 50% reduction in growth for the combined PNA and antibiotic treatment condition as compared to PNA treatment alone for two out of the three biological replicates. Only one biological replicate showed greater than 50% reduction in growth for PNA targeting recA combined with antibiotic treatment in ESBL KPN. Due to the variability seen in these clinical isolates no combination showed statistically significant synergy at three biological replicates. The five cases with positive S values above 0.25 correspond with the clinical isolates that showed at least one biological replicate having greater than 50% reduction in growth in combination treatment as compared to PNA treatment alone.

While these findings substantiate the perturbation-drug synergy discovery pipeline outlined throughout this study, the large variability in synergy (caused by one replicate) suggests the potential for escape. Further research is necessary to understand the underlying mechanisms for this phenomenon.