Phenotypic convergence and collateral susceptibility development in Pseudomonas aeruginosa under antibiotic exposure in ICU patients Authors: Thomas E van der Schalk, Matilda Berkell, An Hottebeekx, Leen Timbermont, David E Tabor, Christine Lammens, Claudia Recanatini, Alexey Ruzin, Antonio DiGiandomenico, Hasan Jafri, Jean Chastre, Mark T Esser, Bruno François, Marc J M Bonten, Herman Goossens, Jan A J W Kluytmans, Antonio Oliver, Youri Glupczynski, Samir Kumar-Singh, Surbhi Malhotra-Kumar, the COMBACTE-MAGNET EVADE study group Published: 2026 ABSTRACT AbstractAntibiotic combination therapy is often used to broaden the antimicrobial spectrum, limit resistance and improve treatment efficacy. Several antibiotics show collateral effects where resistance to one antibiotic increases susceptibility to another. In intensive care units (ICUs), antibiotic treatments are frequently adjusted based on patient outcomes, without considering collateral effects. This provides a setting to study these effects in Pseudomonas aeruginosa (PA), a highly adaptable, multidrug-resistant (MDR), nosocomial pathogen. We compared longitudinal PA isolates from twenty-five ventilated ICU patients receiving various antibiotics to laboratory strains undergoing in vitro adaptive evolution under four antipseudomonal monotherapies. Prolonged exposure to certain antibiotics produced resistance with collateral effects. In vitro, increasing antibiotic pressure drove distinct mutational trajectories. In patients, the number of antibiotics administered did not correlate with resistance changes to those antibiotics, suggesting that switching may reduce persistence of resistance. Notably, an inverse correlation between resistance to non-administered antibiotics and the number of different antibiotic classes administered, aligns with the principles of collateral susceptibility driven by multi-class exposure. This study provides s real-world evidence that empirical antibiotic mixing in ICU patients leverages evolutionary trade-offs. Consequently, diversifying antibiotic pressure via multi-class exposure may attenuate the fixation and persistence of MDR phenotypes in critical care.Subject terms: Diseases, Microbiology Introduction Antibiotic resistance is one of the major challenges facing global healthcare1. With the increasing prevalence of antimicrobial-resistant pathogens—driven in part by antibiotic use and misuse—treatment failure has become an increasingly common and severe problem, particularly in hospital settings and in intensive care units (ICU), where infections can be rapidly fatal2. Ventilator-associated pneumonia (VAP), the most common infection among ICU patients, is associated with all-cause mortality rates of 20–50% and remains a major driver of antibiotic use in the ICU2. Treatment of pneumonia accounts for 50–70% of antibiotic usage in ICUs reflecting the medical necessity of prompt and well targeted antimicrobial therapy3,4. Consequently, antibiotic stewardship efforts in ICU have generally focused not on reducing overall antibiotic use, but on optimizing prescribing strategies, such as cycling and mixing different antibiotic classes5, to preserve treatment efficacy while limiting the emergence of resistance. P. aeruginosa is a major cause of VAP and is also a notoriously adaptable pathogen, rapidly developing resistance to antibiotics6. This rapid development of resistance can be attributed to a combination of chromosomal mutations as well as the acquisition of transferable resistance determinants. This has resulted in the emergence and global dissemination of multiple distinct genetic lineages characterized by multidrug-resistant (MDR) and extensively drug-resistant (XDR) phenotypes, commonly referred to as high-risk clones7. However, several studies have also shown that development of antibiotic resistance, or even brief exposure to antibiotics, can lead to collateral changes in antibiotic susceptibility phenotypes of P. aeruginosa8–11. The development of antibiotic resistance mechanisms in P. aeruginosa, along with its associated collateral sensitivities, is deeply intertwined with the organism’s distinctive intrinsic characteristics. These include a broad spectrum of inherent antibiotic resistance mechanisms12, which enable the pathogen to survive exposure to antimicrobial agents and adapt to various environmental pressures. While inherent (intrinsic) antibiotic resistance is generally stable and fitness-compensated, acquired or de novo resistance (resulting from spontaneous mutations or horizontal gene transfer during infection) is often associated with significant fitness costs or metabolic trade-offs. These uncompensated trade-offs can reduce the organism’s overall fitness, rendering it more vulnerable to other antimicrobial classes, a phenomenon known as collateral sensitivity. These changes hold promise for optimizing treatment of severe P. aeruginosa infections by enabling better prediction of effective treatment strategies, identifying synergistic antibiotic combinations13, or treatment strategies based on alternating antibiotics conferring reciprocal collateral susceptibility14,15. P. aeruginosa possesses 12 resistance-nodulation-division (RND) efflux systems, where the cytoplasmic and periplasmic multidrug efflux (Mex) components interact with outer membrane porins to expel a broad range of antibiotics, significantly contributing to multidrug resistance (MDR) in the bacterial cell16. Among these, the role of four systems–MexCD-OprJ, MexAB-OprM, MexEF-OprN, and MexXY-OprM is well described in mediating antibiotic resistance of P. aeruginosa16,17. For instance, MexCD-OprJ specifically exports certain antibiotics such as quinolones, aminoglycosides, and 4th generation cephalosporins, (not the other β-lactams)18. Mutations in nfxB gene, which negatively regulates the expression of MexCD-OprJ8,18, are associated with collateral susceptibility to other antibiotics by other known and unknown mechanism such as potentially altering membrane lipid composition thereby increasing permeability to aminoglycosides19. Altered expression of other efflux pump components, such as OprM (required for MexAB and MexXY function) and impaired β-lactamase activity in the periplasm (due to leakage of AmpC to the extracellular space caused by abnormal outer membrane permeability) are also proposed to cause collateral susceptibility to most β-lactams, including carbapenems and aminoglycosides in nfxB mutants12. It is even shown that similar alterations were detected in vitro and in vivo, the latter in chronic P. aeruginosa populations in patients with cystic fibrosis, with collateral susceptibility changes caused by alterations in nfxB8. Two principal strategies have been implemented to mitigate the emergence of antibiotic resistance in the ICU: antibiotic cycling and antibiotic mixing. Antibiotic cycling involves the scheduled, sequential use of different classes of antibiotics over defined time intervals, with the aim of reducing selective pressure and curbing resistance. In contrast, antibiotic mixing refers to the random allocation of diverse, appropriate antibiotics to individual patients, thereby increasing heterogeneity in selective pressures across the patient population. Despite their theoretical appeal, neither strategy has demonstrated any superiority or consistent efficacy in reducing antibiotic resistance in multiple randomized controlled trials (RCTs)4,5. Combination antibiotic therapy is also frequently employed for broadening the treatment spectrum as well as to exploit any synergistic interaction between agents while trying to decrease the development of antibiotic resistance. While several antibiotic classes display collateral susceptibility in which the evolution of resistance to one antibiotic increases susceptibility to another, in ICU settings, clinicians commonly modify antibiotic regimens dynamically, often switching agents in response to suboptimal clinical outcomes and not considering collateral susceptibilities. The implications of such reactive antibiotic shuffling during the course of ICU treatment remain poorly understood. It is hypothesized that this practice may inadvertently contribute to the selection and propagation of resistant pathogens. In this study, we harness the rapid evolutionary ability of P. aeruginosa to investigate whether and how antipseudomonal therapies influence genomic and phenotypic changes linked to antibiotic resistance and development of collateral sensitivities within relatively short periods of time ( ~ 1 month of P. aeruginosa positivity) both in ICU patients colonized or infected with P. aeruginosa and receiving antibiotic therapy as well as in simulatory in vitro models of adaptive evolution following antibiotic exposure and of cellular infection. Results Of a total of 1958 mechanically ventilated ICU patients enrolled in the ASPIRE-ICU (ClinicalTrial.gov: NCT02413242) and EVADE (ClinicalTrials.gov: NCT02696902) clinical trials, 271 patients tested positive for P. aeruginosa in lower respiratory tract samples (270 by culture, 1 by PCR alone). Patients in ASPIRE-ICU were sampled within 72 h of ICU admission followed by sampling twice weekly until 30 days or the day of ICU discharge with additional sampling moments at the time of pneumonia diagnoses and 7 days post-diagnosis. Patients in EVADE were screened during ICU stay for P. aeruginosa and once tested positive, the patients were sampled the day of receiving treatment (day 1), followed by samplings on days 2, 4, 8, 15 and 22. Additional samples in EVADE were taken in case of suspected pneumonia; when the diagnosis sample was P. aeruginosa positive, additional samples were taken daily until resolution (Fig. 6B). Of the 271 P. aeruginosa positive patients from ASPIRE-ICU and EVADE, 102 patients had only a single positive sample, 1 patient had multiple positive samples on 1 day, and 87 patients lacked antibiotic treatment data or were treated with a single antibiotic class not typically linked associated with collateral sensitivity or cross resistance development. This left 80 patients with longitudinal isolates. Further genomic analysis (including multilocus sequence typing and SNP profiling) revealed that 55 patients either had different P. aeruginosa sequence types or high genomic variability ( > 100 SNPs), indicating distinct strains. No hypermutator strain, as defined by Oliver et al.20, was detected as expected due to the acute setting of the study. Ultimately, 25 patients harboring genetically consistent P. aeruginosa strains (defined as same ST and <100 SNPs between longitudinally collected isolates) were selected, providing a total of 83 longitudinal isolates for in-depth analysis of in situ adaptive evolution under antibiotic pressure. Patient and sample overview. A Patient inclusion flow from ASPIRE-ICU and EVADE. Pseudomonas aeruginosa (PA)-positive patients (N) within ASPIRE-ICU and EVADE were taken forward and based upon availability of the isolates (n), lack of longitudinal isolates, or patients only receiving antibiotics that were out of scope were excluded. Patients that were eligible were analyzed and based upon the longitudinal similarities a patient that had the same ST longitudinally nor had multiple sub-types/clones detected were included in this study. N = number of patients, n = number of isolates. B Patient sampling overview. At the top, ASPIRE-ICU, where the patients were screened for S. aureus carriage status followed by regular sampling for lower respiratory tract samples (ETA or sputum) and peri-anal swabs. Below, EVADE, where all samples utilized for this study were lower respiratory tract samples (ETA or sputum). The screening samples were analyzed by clinical site for P. aeruginosa positivity both culture and qPCR, prior to study inclusion. When the patient was included in the study, a sample was taken pre-dose (D1) and samples were collected up to three weeks post dose. In case the patient developed pneumonia or suspected serious bacterial infection (SSBI) additional samples were collected. In ASPIRE-ICU, a sample at the day of pneumonia diagnosis was collected and seven days after diagnosis. In EVADE, a sample was collected at the time of diagnosis, and sampling was again collected on a daily basis during PAIP until resolution of the infection in case the initial sample was positive for P. aeruginosa. C Experimental study design. The green blocks indicate isolates, orange blocks indicate phenotypic assays and blue blocks indicate sequence-based assays. Exposure to combination of antipseudomonal antibiotics determines resistance suppression and/or collateral susceptibility in surviving P. aeruginosa colonizing or infecting ICU patients Patterns of antibiotic exposure and their differential impact on MIC evolution could be analysed in 25 critically ill ICU patients, either colonized only with P. aeruginosa, or with a documented VAP infection, each of whom provided at least two P. aeruginosa isolated from at least two longitudinally collected LRT samples with matching sequence types (ST) or lineages. These were further included for the genomic and statistical analyses (Fig. 2). On average, patients received 2.24 different antibiotic classes (range, 1—3) over a treatment duration of 27.5 days (range, 2—46). The relationship between administered antibiotics and MIC measurements to these antibiotics as well as to non-administered antibiotics, evaluated as two-fold dilution or changes (increases or decreases), were assessed based on class-level relevance and categorized into two groups: (i) MIC changes for antibiotics that the patient received (“administered antibiotics”) and (ii) MIC changes for antibiotics that the patient did not receive (“non- administered antibiotics”). For instance, when a patient received a carbapenem, the same class MICs included both imipenem and meropenem. Similarly, administration of β-lactam/β-lactamase inhibitor combinations (e.g., piperacillin-tazobactam, ceftolozane-tazobactam, and ceftazidime) was associated with MICs for piperacillin-tazobactam, ceftolozane-tazobactam, ceftazidime, and cefepime. Additional antibiotic groups are listed in Table 1. Overview of mechanically ventilated ICU patients colonized with P. aeruginosa isolates, received antibiotic therapies during ICU stay, antibiotic susceptibility profile of first and last isolated/available isolate, MIC fold changes and genotypic alterations over time between the first and last available isolates. A indicates P. aeruginosa isolated from respiratory samples (endotracheal aspirates) as circles. Star denotes P. aeruginosa isolated during ventilator-associated pneumonia. Treatment with antibiotics is indicated by colored bars if these belonged to one of four classes (Quinolone, Carbapenem, Aminoglycoside and Colistin) against P. aeruginosa and the most commonly used ones were further utilized for experimental evolution in this study (patients 1-25). Many of the patients received other antipseudomonal antibiotics but these were found to be less associated with evolutionary resistance including the two additional highlighted groups, cephalosporins (ceftazidime, and ceftolozane/tazobactam) and β-lactams (piperacillin/tazobactam). B shows the susceptibility profile of initial isolate (left) and PAIP & final isolates (center) obtained from each patient according to EUCAST breakpoints. The legend (left) denotes different colors for susceptibility (green), intermediate (orange) and resistance (red). C shows the MIC fold change per antibiotic for each patient between the initial and last available isolates (right) shows a decrease (green) or increase (red) in MIC (legend right). D indicates emergent genetic alterations in P. aeruginosa during a pneumonia episode or the second isolate in comparison to the initial isolate. Table 1The table summarizes the relationship between administered antibiotics and their corresponding same class MICs, as they belong to the same antibiotic class and were therefore likely subject to direct similar selective pressureAntibiotics administeredSame class MICs testedCarbapenemImipenem, MeropenemPiperacillin/tazobactamPiperacillin/tazobactam, Ceftolozane/tazobactam, Ceftazidime, CefepimeCeftolozane/tazobactamCeftazidimeAminoglycosideAmikacin, TobramycinQuinoloneCiprofloxacinColistinColistin The table summarizes the relationship between administered antibiotics and their corresponding same class MICs, as they belong to the same antibiotic class and were therefore likely subject to direct similar selective pressure Our findings demonstrate that MICs for administered antibiotic classes significantly increased with the cumulative number of days of antibiotic exposure (Pearson’s r = 0.29; Spearman’s R² = 0.613; P < 0.001, Fig. 1), consistent with the concept that prolonged antibiotic use promotes the emergence of resistance to that antibiotic or its class. Interestingly, the number of different antibiotic classes administered did not correlate with MIC changes to administered antibiotic classes (Pearson’s r = 0.01; Spearman’s R² = 0.019; P > 0.05), suggesting that frequent switching between different antibiotic classes may reduce the persistence of selective pressure and mitigate resistance development. Antibiotic Exposure and MIC Shifts in P. aeruginosa. Pearson’s association between antibiotic days or total number of antibiotics administered and MIC changes to administered and non-administered antibiotics in P. aeruginosa isolated from ICU patients. Similarly, MIC changes for non-administered antibiotic classes were not significantly correlated with the duration of exposure to administered antibiotic classes (Pearson’s r = −0.095; Spearman’s R² = 0.035; P > 0.5). However, there was a significant inverse correlation between MICs to non-administered antibiotics and the number of different antibiotic classes administered (Pearson’s r = −0.48; Spearman’s R² = 0.035; P = 0.016 and 0.046, respectively), indicating the presence of collateral susceptibility mechanisms driven by exposure to multiple antibiotic classes. These results were corroborated by ANOVA. Grouping patients by days of antibiotic exposure revealed a significant increase in MICs to administered antibiotics with longer treatment durations (F = 6.395, df = 16, P = 0.006), but no significant change in MICs to non-administered antibiotics (F = 0.638, df = 16, P > 0.5). Conversely, grouping by the number of antibiotic classes used did not significantly alter MICs for administered antibiotics (F = 1.018, df = 2, P = 0.378), but did for non-administered antibiotics (F = 4.381, df = 2, P = 0.025). Pairwise comparisons indicated that the increase in MICs for non- administered antibiotics was statistically significant when comparing patients who received one vs. three antibiotic classes (mean difference: 9.83 doublings; P = 0.028), with a trend towards significance for one vs. two classes (mean difference: 9.10 doublings; P = 0.082), but no significant difference between two and three classes (mean difference: 0.74 doublings; P = 1.00). In vivo evolution of P. aeruginosa in mechanically ventilated ICU patients receiving antibiotics P. aeruginosa isolated from these 25 patients exhibited multiple alterations over time (Fig. 2, Table S2). Several mutations that emerged were in the outer membrane porin OprD (patients PA1, 6, 7, 10, 12, 14, 21, 22, 23, n = 9) that are often associated with carbapenem resistance21 (Fig. 2), as well as alterations in RND efflux systems (PA2 MexB, 4 MexR & MexZ, 8 MexB, 14 MexR, 18 NfxB, 19 MexF, 24 MexZ, n = 8) that can lead to resistance to multiple classes of antibiotics16, like quinolones, β-lactams (penicillin) and carbapenems. When linking genetic alterations to shift in antimicrobial susceptibility phenotypes, there were nine patients that showed category changes from resistance towards susceptible or from susceptible towards resistance (Fig. 2B, C, Table S2). In one patient (PA20), multiple phenotypic MIC changes were observed over time without identifying corresponding genetic alterations in the sequenced isolates (Fig. 2). This suggests the involvement of alternative adaptive mechanisms, such as transient transcriptomic remodeling, biofilm-associated tolerance, or the presence of persister cell subpopulations. Furthermore, we cannot exclude the possibility of complex genomic events—such as large insertions/deletions (INDELs), chromosomal rearrangements, or undetected horizontal gene transfer—that may fall below the detection threshold of current bioinformatic pipelines. On the other hand, strains from multiple patients that had limited MIC changes yet showed changes in resistance-associated genes (PA8, 13, 14, 23). These patient isolates had mostly acquired frameshift mutations in resistance-associated genes like oprD, ampD, mexR and mexB, as well as some other indirectly linked genes (Table S2). While some of these isolates might have had compensatory alterations, based on the available data and in situ analyses, we predicted a loss of function that might indicate (antibiotic) resistance gene silencing22. In patients where a clear phenotype-genotype correlation could be established, a shift towards resistant phenotypes was observed in four cases (PA6 (carbapenems), 18 (CIP), 22 (carbapenems), 24 (CIP and MER)). Decrease in susceptibility that could be linked to longitudinal genetic changes was detected in four cases (PA1 (carbapenems), 7 (carbapenems), 10 (CAZ), 12 (TAZ)). Adaptive evolution under antibiotic pressure results in highly resistant PAO1 strains To dissect the mechanisms underlying adaptive evolution in response to antibiotic pressure and the treatment effects observed in ICU patients, we utilized an in vitro treatment model. Two differently sourced PAO1 strains (PAO1DTU and PAO1MI) were subjected to increasing concentrations of four commonly used anti-pseudomonal antibiotics (CIP, IPM, TOB, or CST, based upon the antibiotics classes used frequently in clinical practice and known for selecting adaptive resistance) at 22-hour intervals for 20 exposures. The two PAO1 strains were selected to investigate potential evolutionary differences based on known PAO1 diversification23 and these showed no genomic differences in the variant analysis prior to the experiment. Both strains evolved similarly in terms of rapidly developing high levels of antibiotic resistance in this study (Fig. 3A, B). After 20 CIP exposures, both strains grew in a CIP concentration 5120 times higher than their starting concentration. IPM, TOB, and CST exposure resulted in growth in 32-, 128-, and 512-fold higher antibiotic concentrations than in their first exposure, respectively. IMP- and CST-exposed strains grew rapidly in Clinical & Laboratory Standards Institute (CLSI) resistant concentrations (R) after only two and three exposures skipping over inhibiting growth of CLSI intermediate (I) concentrations, respectively. CIP exposure resulted in growth inhibited at CLSI I and R concentrations in the same exposures (seven and eight, respectively) for both strains, whereas TOB exposure also showed CLSI I and R growth inhibitions but for both strains at very different exposures (exposure five and nine for PAO1MI, and six and seven for PAO1DTU). Adaptive evolution under antibiotic pressure results in highly resistant P. aeruginosa phenotypes and collateral susceptibility effects. A, B Two P. aeruginosa PAO1 strains, PAO1 MI and PAO1 DTU were subjected to adaptive evolution while exposed to increasing concentrations of CIP, IPM, TOB, and CST during a total of 20 exposures. Bacterial growth at concentrations classified as intermediate by the CLSI guidelines are marked by a cross surrounded by a circle, whereas first growth at a concentration classified as resistant are denoted with an empty circle. For both CIP and TOB, the ECOFF were not skipped during growth and are denoted with an empty diamond. In case no optimal growth was reached, concentrations from the previous day are considered in the graphs. C, D Correlations in the evolution of antibiotic resistant phenotypes for PAO1 MI and PAO1 DTU was assessed by comparing the MIC for isolates growing at intermediate and resistant concentrations as well as the final exposure. E, F Exposure to CIP, IPM, TOB, and CST resulted in development of collateral susceptibility changes for both the PAO1 MI and PAO1 DTU strains. The heatmaps illustrate differences in MIC after antibiotic-exposure compared to parental PAO1 MI and PAO1 DTU reference strains. Rows indicate isolates from different exposures after exposure of individual antibiotics whereas columns indicate result of antibiotic susceptibility testing. CIP ciprofloxacin, IPM imipenem, TOB tobramycin, CST colistin, ECOFF epidemiological cut-off values, CAZ ceftazidime. AZM: azithromycin, I intermediate strain, R first resistant strain, F strain obtained after the 20th final exposure. MIC minimum inhibitory concentration. Despite minor differences in patterns of developing resistance between the two PAO1 strains, the correlation between observed resistance phenotypes was, as expected, high between the isolates that grew in I, R, and final (F) exposures (Pearson correlation coefficient of 0.85 and 0.87 for PAO1MI and PAO1DTU, respectively; Fig. 3C, D). Examining individual antibiotics, PAO1MI and PAO1DTU showed a strong positive correlation in development of a TOB resistance phenotype (Pearson correlation coefficient = 0.99), whereas the CIP-resistant phenotype was less consistent (Pearson correlation coefficient = 0.51). A stronger correlation might have been observed for all antibiotics with use of ECOFF values rather than CLSI breakpoints in the exposures to be closer towards the emergence of resistance. Due to only two data points, available of IPM and CST there was no option to calculate the correlation. Collateral changes in antibiotic susceptibility patterns show consistent associations with emergent resistance phenotypes To determine potential collateral antibiotic susceptibility effects in the two exposed PAO1 strains, the fittest isolates based upon growth in agar with the exposure antibiotic were subjected to (multiple) repeat MIC tests. Following exposure to controlled concentrations of several common anti-pseudomonal drugs (CIP, IMP, TOB, CST) several changes in susceptibility ( ≥ 2-fold higher/lower) were observed in isolates from I, R, and F exposures compared to the parental strains, although slight differences were observed between PAO1DTU and PAO1MI (Fig. 3E, F, Table S3). AZM was included in this study for proof of concept based upon Imamovic et al.8 also taking into account the known effects of different medium for AZM MIC testing of P. aeruginosa24. Maximum collateral effects on the P. aeruginosa PAO1 phenotype occurred following CIP exposure most likely due to upregulated efflux capabilities, resulting in an increased susceptibility to IPM (4x) and TOB (5x), and increased AZM ( > 2x) MICs. Additionally, a slight increase in CAZ (3.9x) susceptibility was observed in PAO1MI in the R exposure. By contrast, IPM exposure did not result in collateral susceptibility to any of the investigated antibiotics, with co-resistance to CAZ being observed in the PAO1DTU strain but only in the F exposure (6x). Following TOB exposure, PAO1MI and PAO1DTU both developed increased susceptibility to CAZ (PAO1MI F: 2x, and PAO1DTU R: 2.6x), and PAO1DTU additionally developed resistance against CIP (4x). PAO1MI from the I exposure grew in a TOB concentration >10x higher than the MIC of the parental PAO1MI strain, yet did not show any changes in susceptibility category to TOB. Exposure to CST did not result in development of any co-resistance (to antibiotics other than CST) in PAO1MI nor in PAO1DTU; however, both collateral susceptibility to CAZ ( > 6x) and AZM ( > 8x) was observed for both isolates in the final exposure. Additionally, an evolution towards increased CIP susceptibility (3.9x) was found with PAO1MI. Overall, the phenotypic patterns of exposure effects observed after CIP, IPM and CST exposure are similar to those previously reported8,11, whereas results obtained after TOB exposure were less consistent. From the comparison of the repeat isolates obtained from longitudinally followed patients with the in vitro developed resistant PAO1s, there seem to be only partially comparable results of genotypic and phenotypic correlation based upon the genomic data with the in vitro derived schematic (Table 2). Table 2Collateral MIC changes after antibiotic exposure of at least two-fold increase or decreaseIn green, decreases in MIC and in orange, increases in MIC that were observed in both PAO1s. Per row for each antibiotic, the MIC changes can be seen. Changes with * were in at least 1 of the 2 PAO1s less than a two-fold. Collateral MIC changes after antibiotic exposure of at least two-fold increase or decrease In green, decreases in MIC and in orange, increases in MIC that were observed in both PAO1s. Per row for each antibiotic, the MIC changes can be seen. Changes with * were in at least 1 of the 2 PAO1s less than a two-fold. Population-level dynamics of mutational succession under antibiotic pressure show distinct evolutionary trajectories of the two PAO1 strains By pooling the three wells per strain to find the most well adapted resistant strain, it was possible to investigate the populations, with cut-offs slightly below 1/3rd (alterations below 30% frequency were not considered), at the I, R and F exposures. We utilized population sequencing to holistically study the dynamics of emergent non-synonymous mutations during the I, R, and F exposures under antibiotic pressure compared to the parent PAO1 strains. Both PAO1MI and PAO1DTU populations harbored similar but also uniquely different non-synonymous SNPs at varying frequencies at all investigated exposures in this study, with some mutations disappearing, or at least falling below the threshold of detection, over time and with sustained antibiotic pressure (Fig. 4). Importantly, some genes were commonly/similarly mutated, the two PAO1 strains showed diversity in mutational trajectories under pressure of the same antibiotic (Fig. 4) yet still showed similar MIC changes (Table S3). A full list of observed SNPs and their frequencies in each P. aeruginosa population can be found in Table S4 as well as the genomic alterations on strain level. No alterations on strain level were detected that were not present in the populations. The data shown here is focused on the nonsynonymous alterations, but there were instances of synonymous alterations and alterations in non-coding sequences without known relations towards antibiotic susceptibility changes. Genetic alterations underlying collateral susceptibility phenotypes in PAO1 strains after adaptive evolution in vitro. Intermediate, resistant, and final exposure isolates, and broths were sequenced to study genetic changes after in vitro evolution under CIP (A, B), IPM (C, D), TOB (E, F), and CST pressure (G, H) compared to the parental reference genomes of PAO1 MI and PAO1 DTU. Mutation frequencies observed at each sequenced exposure are shown vertically as the percentage of reads in the broth of that exposure. A frequency of 1.0 indicates that the whole PA population contains the observed mutation, whereas frequencies lower than 1.0 indicate mutation-harboring subpopulations in the PA community. CIP ciprofloxacin, IPM imipenem,TOB tobramycin, CST colistin. Overall, multiple known resistance-associated alterations or genes were detected starting from the I or R exposures. All CIP-exposed P. aeruginosa populations were found to harbor known fluoroquinolone resistance-mediating and other well-characterized mutations25,26 in DNA gyrase-encoding gyrA as well as in DNA topoisomerase-encoding parC and parE (Fig. 4A, B). Additional CIP resistance associated alterations in efflux systems were detected by alterations of mexR27, regulating mexAB, and nfxB8, regulating mexCD, or the nfxB binding site (in PAO1DTU altered binding site not shown in Fig. 4B). In case of IPM exposure both PAO1s showed alterations in the outer membrane porin OprD, a commonly altered porin in carbapenem-resistant strains28–30. In TOB exposed PAO1 strains, alterations were found in the fusA1 gene that have been associated with aminoglycoside resistance31,32. Other resistance-associated altered genes found after TOB exposure were ptsP33 and mexY34 both only detected in the F exposure (Fig. 4E, F). CST exposure alterations were detected in pmrA, pmrB, speE2, phoP and phoQ (Fig. 4G, H), all have been described previously and are associated with CST resistance35–37. As also observed in vivo, multiple alterations without any causal relation towards antibiotic resistance were also detected. With CIP exposure, PAO1MI and PAO1DTU both showed alterations in genes more often associated with virulence in the F exposure (Fig. 4A, B) like pilB (PAO1MI), potentially leading to a type IV pilus assembly deficient phenotype38 and wbpX (PAO1DTU) potential loss of the common polysaccharide antigen of lipopolysaccharide (LPS)39. Interestingly with PAO1DTU, in the F exposure to IPM additional alterations were found in LPS and exopolysaccharide biosynthesis genes, ssg (Q11L) and wbpL (A77fs, which was also found in patient 11), both genes have been associated with resistance towards multiple antibiotics39. These findings, combined with other studies8,11, indicate that several mutational mechanisms underlie collateral effects but that they are strain-dependent, making direct comparisons between studies and identification of genetic markers predicting these effects most difficult. In addition, when comparing the data from Fig. 3 with Fig. 4 there seems to be a potential relation between the rapid development of resistance and emergence of diverse mutational subpopulations. For both IPM and CST where resistance did develop in only a few exposures both PAO1 strains showed resistance-associated alterations with the R exposure only for a subpopulation, indicating the presence of a susceptible population. On the other hand, when comparing with the slower development of resistance in CIP and TOB, at the I and R exposures, resistance associated genetic alterations were present in almost 100% of the population indicating the importance of looking into ECOFF values for early detection since with these antibiotics where the entire population is enriched with diverse mutations, with PAO1MI showing multiple resistance-associated alterations at the I and R exposures. Reduced growth and motility, but increased cytotoxicity of PAO1 wbpL mutant Several of the P. aeruginosa isolates obtained after in vitro and in vivo evolution under antibiotic pressure were found to develop non-synonymous mutations in LPS biosynthesis genes of the wbp operon. Mutations in the wbp operon were identified in vitro (PAO1DTU (wbpL)) as well as in vivo ((PA11 (wbpL) and PA16 (wbpM) after carbapenem exposure. In the case of PA11 and PAO1DTU there were also other genomic alterations (see tables S2 and S4). These isolates and their parental isolates were selected to study the effect on bacterial fitness, motility, biofilm production, and cytotoxicity (Fig. 5). Comparing the parental PAO1DTU to the IPM-treated mutant from the F exposure (PAO1DTUwbpL), multiple significant differences were detected. As shown in Fig. 5A-F, a decreased growth and motility but an increase in biofilm production as well as an increase in cytotoxicity (only after a time lag of 20 h), were observed with PAO1DTUwbpL (Fig. 5G). This delay could have been due to the reduced growth rate; however, bacterial loads of both strains were similar over time in cytotoxicity co-cultures (Fig. 5H), indicating an enhanced virulence of PAO1DTUwbpL compared to PAO1DTU. Increased cytotoxicity in our in vitro generated wbp mutants might seem paradoxical; however, LPS-deficient mutants or those with reduced LPS chain length show increased cytotoxicity toward host cells, particularly respiratory epithelial cells40. One mechanism might be due to a hyper-induction of the T3SS. Without the protective “shield” of the long O-antigen chains, the T3SS “needle” may have better physical access to the host cell membrane, leading to more efficient delivery of toxins like ExoU or ExoS40. Of the clinical in vivo isolates, both the PA11wbpL and PA16wbpM the mutants had a higher growth rate compared to their parental isolates (Fig. 5B). Only for PA16, there was an increased biofilm production in the wbpM mutant compared to the parental strain (Fig. 5C). For motility however, only PA11 showed significant differences between the parental and the mutant isolate. PA11wbpL was found to have (p = 5.0e-9) increased swimming (Fig. 5D) and reduced swarming (p = 0.020) and twitching (p = 0.002) compared to PA11 (Fig. 5E and F). Bacterial fitness, motility, biofilm formation capacity, and toxicity for three parental (PAO1 DTU, PA 11, PA 16) and mutated strains. Cytotoxicity (A), expressed as cytotoxicity/log (CFU/mL), and bacterial loads (B) in cytotoxicity co-cultures were measured over a 36-h period and evaluated based on the area under the curves. For biofilm forming capacity (C), and swimming, swarming, and twitching motilities (D) the measurements were compared towards the parental isolates. Bacterial was measured by 24-h growth curves measured every 15 min (E) and intrinsic growth rates (F). For graphs (A–F) the means are displayed with 95% confidence intervals (whiskers), for (A, B and E) a line is drawn between the longitudinal means. *: p < 0.05. Discussion This is one of the first studies investigating the collateral effects of antibiotic exposure on P. aeruginosa with a focus on critically ill ICU patients. These data provide one of the first in vivo evidence that empirical multi-class antibiotic exposure in individual ICU patients, while aiming to limit prolonged use of any single class and mitigate resistance amplification, might induce collateral susceptibility to unrelated, non-administered antibiotics. While combination therapies are not considered standard of care for most common bacterial infections, in suspected VAP due to difficult-to-treat pathogens like P. aeruginosa that also have a high propensity for resistance development, combination therapy was until recently recommended41. However, most studies comparing PA-VAP patients on combination or monotherapy have reported similar outcomes in terms of clinical cure rates and mortality42. In addition, a recent multicentric randomized controlled trial (iDIAPSON) did not show difference in patient outcome between combination opposed to monotherapy when comparing the duration of treatment (8 vs 15 days) of therapy for PA-VAP43. Despite compelling experimental data from in vitro studies demonstrating the advantages of combination therapy, there is currently a lack of evidence to recommend or guide clinical use of combination therapy as an antibiotic stewardship strategy to suppress or deescalate emergence of resistance and/or obtain collateral susceptibility (CS) benefits. In our study, mutation driven changes in P. aeruginosa isolated over time from ICU patients either colonized only or developing VAP were numerous and diverse, and in most isolates not deterministic of the phenotype especially where changes in known resistance associated genes did not result in relevant MIC increases. Nonetheless, phenotypic MIC changes pointed to an overall limited or de-escalated emergence of resistance in patients where two or more antibiotic classes were utilized. Conversely, while MICs for administered antibiotics remained stable, those for non-administered antibiotics showed a significant decrease upon administration of two or more antibiotic classes. While these effects might be the (cumulative) result of dilution of selective pressure through antibiotic switching, accumulation of fitness costs, or population bottlenecks in critically ill patients, our results provide evidence consistent with the potential (clinical) benefit of combining antipseudomonal antibiotic classes in mechanically ventilated ICU patients, who are at increased risk of harboring high inoculum (high load) of nosocomial opportunistic pathogens such as P. aeruginosa either as colonizers or as causative agents of severe infections. In high inoculum infections (bacterial burden scenarios) like P. aeruginosa VAP in ICU patients, the mutational resistance threshold is easily surpassed44, making it likely that less susceptible organisms are present at treatment onset and that these subpopulations can then expand under antibiotic pressure. A large US-based analysis found that none of the conventional anti-pseudomonal dual combination regimens achieved a ≥ 95% Pseudomonas coverage, implying that resistant subpopulations are frequently present from the onset in ventilated patients45. This is arguably one of the first studies to our knowledge where the potential for resistance suppression (RS) and patterns compatible with CS emergence under combination therapy have been evidenced in a real-world clinical setting. However, given the inherent constraints of patient sampling and data collection in an acute care setting, our findings are correlational rather than causal. Furthermore, we do not identify specific antibiotic pairs exhibiting directional or reciprocal CS, typically observed in controlled in vitro environments. To further understand whether the same antibiotic might lead to different mutational trajectories as observed in patients, we chose to study two PAO1 strains that differed by < 1% genomic differences, simulating micro evolutionary changes (e.g., SNPs, indels, regulatory mutations) that might arise in vivo in settings of high bacterial loads of a single ST type leading to emergence of diverse subpopulations in patients. Although the PAO1 reference strain has its origins in the clinics, continuous microevolution during culture and storage and lab transfers have resulted in subtle alterations in geno- and phenotype46. With the two ‘microevolved’ PAO1 reference strains, we generated highly resistant P. aeruginosa mutants with distinct genomic alterations, though the overall susceptibility changes observed were relatively similar. While core resistance mechanisms were found in both PAO1 strains, distinct pathways were also found to be altered. These data support our in vivo findings of high genotypic divergence, but importantly, robust phenotypic convergence of resistance and potential CS mechanisms that have the potential to be exploited in the clinics for patient care. Another important feature of note in our study was that by using population sequencing, we could visualize the rise and fall of different genotypic resistance trajectories under pressure. The approach of pooling replicate wells helped to identify robust consensus mutational trajectories that are most likely to survive in the high-pressure environment of an ICU-simulated in vitro model. While this may limit the detection of rare minority sub-lineages, our approach ensures that the reported resistance patterns represent the most stable and reproductively successful phenotypes likely to emerge under selective pressure. As a summatory of all mutations present in the population, our analysis showed that for antibiotics that demonstrated a slow development of resistance i.e., those that underwent an intermediate MIC phase before exhibiting resistance, the entire growing population harbored genomic alterations (as for CIP and TOB exposures) while for antibiotics where the resistance developed rapidly, only a subpopulation harbored resistance-associated genomic alterations (IPM and CST exposures). These two distinct phenomena are reminiscent of the selective sweeps47 (rapid spread of beneficial mutations such as resistance associated mutations) and boom-bust48 (cyclical patterns of bacterial population collapse and regrowth) concepts of microbial evolution and adaptation. For the antibiotics used in this study, the ECOFF values that differed from the I concentration are more important potentially to detect an early rise of resistance than the I concentration. This should thus be monitored besides the MIC as a potential driver for antibiotic resistance development and taken into consideration49. In addition, a focus upon more than or at least equal to two-fold MIC changes facilitated mapping collateral susceptibility changes in a more robust manner compared to other studies6,8,9. A longitudinal investigation of P. aeruginosa–positive patients exposed to quinolones, carbapenems, aminoglycosides, and colistin revealed a high frequency of resistance-associated genetic alterations predominantly linked to antibiotic uptake and efflux mechanisms, but also involving several indirectly associated pathways—most notably, those related to lipopolysaccharide (LPS) biosynthesis. There was a high reproducibility of the genomic alterations as shown by the detected in vivo and in vitro alterations, even in longitudinally collected isolates from patients within a small timeframe in an acute infection setting where similar non-traditional resistance pathways were affected. Overall alterations and potential loss of function in both the type 4 pili as well as the LPS synthesis as a response to antibiotic pressure could result into a survival mechanism for P. aeruginosa, as has been reported in case of clinically relevant CIP exposure at 1 µg/mL, that both the motility and LPS synthesis is down regulated50. The findings of this study show both potential novel antibiotic resistance associated genomic alterations (Table S4) in P. aeruginosa as well as potential novel collateral effects of antibiotic exposure including an (in vitro) increase in virulence. More importantly, we observed the potential and the reliability for future clinical utility of RS and CS guided combination therapy or treatment strategies based on alternating/ sequential antibiotics potentially conferring reciprocal CS, where treatment with one would sweep the resistant mutants as well as create sensitivity for the other and vice versa, that could be exploited to treat severe multi-drug resistant bacterial infections. This study identifies a clinical trend that warrants further investigation via prospective, controlled stewardship protocols. Methods Study design In this study we investigated the effects of antibiotics on P. aeruginosa evolution, in combination with collateral susceptibility effects, both from longitudinally sampled mechanically ventilated patients and through adaptive evolution experiments under antibiotic exposure in vitro. Patients were selected based on longitudinal P. aeruginosa positivity and antibiotic treatments that are often associated with emergence of mutational resistance at varying frequencies, e.g. fluoroquinolones, carbapenems, aminoglycosides and colistin. Patients fulfilling the criteria had the isolates investigated for longitudinal genomic and antimicrobial susceptibility changes. Based on molecular epidemiological data, PAO1 reference strain, rather than PA14, was chosen for the in vitro study since most of the clinical isolates belonged to a similar clade based on type three secretion system effector carriage (data not shown). Further, two ‘genetically similar’ PAO1 reference strains ( < 1% genetic dissimilarity), obtained from two different sources were utilized to simulate microevolutionary genomic changes typically observed under high bacterial loads of a single sequence type (ST), facilitating the emergence of genetically diverse subpopulations within a host that, we hypothesized, might exhibit differential responses to the same antibiotic exposure. In vitro modeling was performed to assess the impact of single antibiotic exposure on strain survival and genomic variation, as the bacterial strains would not have time to adapt under direct exposure to combined agents. This limitation precluded further analysis of adaptive responses under combination treatment conditions and highlights the potential lethality of such regimens against certain subpopulations. The two PAO1s were longitudinally exposed to increasing concentrations of the same antibiotic to ensure that, besides the known traditional resistance mechanisms, other non-traditional resistance mechanisms were likely to be detected. In case of alterations in these potential non-traditional resistance mechanisms that are detected both in vivo and in vitro, additional investigation was conducted for impact on biofilm formation, motility, fitness, and cytotoxicity effects (Fig. 6C). Clinical samples and strains P. aeruginosa isolated from respiratory tract samples (primarily endotracheal aspirates but also sputum) or perianal swabs collected from mechanically ventilated patients admitted to the ICU during March 2015—February 2019, and screened on inclusion criteria as part of two clinical trials, were included in this study (Fig. 6A). Of the two clinical trials, one was observational. ASPIRE-ICU (Advanced understanding of Staphylococcus aureus and Pseudomonas aeruginosa Infections in Europe—Intensive care units, ClinicalTrial.gov: NCT02413242)51 included 30 ICUs and 2035 patients. EVADE (Effort to Prevent Nosocomial Pneumonia caused by Pseudomonas aeruginosa in mechanically ventilated Subjects, ClinicalTrials.gov: NCT02696902)52 a Phase II, randomized, controlled safety and efficacy trial of MEDI3902 included 48 ICUs and 184 randomized patients. MEDI3902 or gremubamab is a bispecific monoclonal antibody against two P. aeruginosa proteins, for the prevention of VAP in adult ICU-patients. Both trials (in ASPIRE-ICU the P. aeruginosa arm) were conducted as part of the IMI-FP7 COMBACTE-MAGNET project. Both ASPIRE-ICU and EVADE followed up patients longitudinally throughout their ICU stay (Fig. 6B). Human Ethics and Consent to participate declarations were approved and collected in ASPIRE-ICU and EVADE and not additionally applicable for this study. Among 1958 patients from both the ASPIRE-ICU and EVADE studies, 271 tested positive for P. aeruginosa in the lower respiratory tract (270 through positive cultures and one through PCR only). A large group of patients (n = 102) were only found P. aeruginosa positive at one single time-point and most of these had only 1 sample available (n = 52). Based on (availability of) treatment data, patients receiving antibiotics associated with mutational resistance, a selection could be made of 80 patients with multiple longitudinal isolates were further analyzed, and those whose isolates shared the same sequence types (STs), and/or showed fewer than 100 SNPs without evidence of subpopulations with altered lineages were included in subsequent analysis. This selection process resulted in the selection of 25 patients with a total of 83 longitudinal isolates (Fig. 6B, Table S1). Adaptive in vitro evolution Strains In vitro antibiotic selection experiments were conducted using two P. aeruginosa PAO1 reference strains: PAO1MI obtained from MedImmune/AstraZeneca and PAO1DTU from the Danish Technological University. Antibiotics Four antipseudomonal antibiotics were utilized in the adaptive in vitro evolution experiments: ciprofloxacin (CIP; 17850, Sigma, Merck, Germany), tobramycin (TOB; T4014, Sigma, Merck), imipenem (IPM; 45532001, ACROS, Belgium), and colistin (CST; C4461, Sigma, Merck). Stock solutions were prepared according to Clinical & Laboratory Standards Institute (CLSI) guidelines53 and stored at 4 °C until use. A fresh stock solution was prepared after 10 treatments for CIP, TOB, and CST, whereas a new stock solution was prepared every 24 h for IPM, as recommended by the manufacturer. In vitro evolution experimental design Adaptive evolution experiments for assessment of collateral susceptibility development under antibiotic pressure were performed based on the method described by Imamovic et al.8, with minor modifications. Briefly, a single P. aeruginosa PAO1 colony (either PAO1MI or PAO1DTU) was grown for 12 h in cation-adjusted Mueller Hinton broth (CAMHB, Becton Dickinson (BD), France). An 8 µL culture aliquot, as described by Imamovic et al.8, was inoculated into 1 mL of CAMHB containing one of four antibiotics (CIP, IPM, TOB, or CST) at ½ minimal inhibitory concentration (MIC), MIC, and 2x MIC. Cultures were incubated at 37 °C with shaking at 180 rpm in 24 well plates (NuncTM non-treated, 144530, ThermoFisher Scientific, Denmark). After 22 h (~7–8 generations based on growth and carrying capacity (data not shown), with no decline in the plateau phase), cultures showing growth (OD600 ≥ 0.4) at the highest antibiotic concentration were transferred into fresh CAMHB with the same, two-fold higher, and a four-fold higher antibiotic concentration compared to the previous exposure. The process was repeated for 20 exposures, with triplicates maintained for both strains in each exposure. The triplicates showing the highest growth were pooled per strain, and utilized for the subsequent inoculation, as well as for quantitative culture by spiral-plating (Eddy Jet, program 6; 50 µL logarithmic spreading, IUL micro) on Mueller Hinton agar (MHA, BD, USA) plates. Intermediate and resistant status was defined according to the CLSI guidelines53. Isolates were cultured overnight at 37 °C on MHA plates containing ¼ of the antibiotic concentration from the corresponding exposure and morphology-based selection (in case of multiple morphologies) performed to be taken for further analyses. For each passage and for each of the two PAO1 strains, wells not exposed to any antibiotic were pooled and taken further for a new plate inoculation as a control to monitor natural strain evolution without antibiotics. Antibiotic susceptibility testing Minimal inhibitory concentrations (MICs) were determined at the first intermediate (I), first resistant (R) and final (F) exposures based upon morphological characteristics using a quantitative gradient diffusion method (E-tests, bioMérieux, France) for CIP (reference# 412311), TOB (reference# 412479), IPM (reference# 412374), and ceftazidime (CAZ, reference# 412293). For azithromycin (AZM, Sigma, Merck) and CST (MIC-Strips Colistin, Merlin, Germany), MIC values were determined using broth microdilution (microbroth) according to EUCAST54,55 guidelines. This approach differed from in vitro evolution experiment, which used CLSI breakpoints to include intermediate concentrations specific to CLSI guidelines and closer towards some clinical cut-offs. MIC determination for clinical isolates obtained from both the ASPIRE-ICU and EVADE studies is described in detail by Torrens et al.56. Whole-genome sequencing Short-read Illumina sequencing Whole-genome sequencing was performed on cultured pooled populations and on isolates from the first intermediate (I), first resistant (R), and final (F) exposures, as well as on clinical isolates. Approximately 2 × 109 colony forming units (CFU) per mL of population cultures, as determined by quantitative culture, were used for DNA extraction. DNA was extracted from first intermediate (CLSI), first resistant (CLSI), and final exposures for both strains, along with the fittest isolate, using the MagAttract HMW DNA Kit (Qiagen, Germany) according to the manufacturer’s instructions. Multiplexed DNA libraries were prepared using the Nextera XT Library Preparation Kit followed by v2 2 × 150-bp paired-end sequencing on a MiSeq instrument (Illumina Inc., USA). Long-read PacBio sequencing Parental PAO1MI and PAO1DTU strains were sequenced using long-read sequencing. Multiplexed microbial libraries were prepared using the Template Prep Kit v.1.0 and the associated Barcoded Adapter 8 A or 8B Kit (Pacific Biosciences, USA) as described by the manufacturer. Sequencing was performed on a Sequel instrument utilizing SMRT Cells 1 M v3 (Pacific Biosciences, USA). Assembly of reference genomes Raw sequence quality was assessed using FastQC (https://www.bioinformatics.babraham.ac.uk/projects/fastqc/) followed by trimming using the BacPipe v1.2.657 implementation of TrimGalore (version 0.6.2, Phred score quality threshold = 25) (https://www.bioinformatics.babraham.ac.uk/projects/trim_galore/). PacBio reads were assessed and de novo assembled into genomes using the SMRTlink v.7.01 implementation of HGAP v.458. Assembled genomes were polished by aligning trimmed short reads to the long-read assembly with BWA-MEM59,60 v0.7.17, converting alignments with Samtools61 v1.7, and polishing with Pilon62 v. 1.23. Circularization was assessed by mapping long reads to the overlap of contigs using the CLC Genomics Workbench (v9.5.1 CLC Bio, Qiagen, Germany). Assembly annotation and further analysis of resistance and virulence elements were performed using the in-house developed pipeline BacPipe v1.2.657, from which MLST v2.0, VirulenceFinder v2.0, ResFinder v3.2, and PROKKA v0.8 were used. Assembly quality was assessed using CheckM63 v.1.1.3, and assemblies with <98.0% completion and >5.0% contamination were excluded from further analyses. Variant calling Raw reads were trimmed as described above. The isolates from the first exposure with growth in intermediate and resistance concentrations, as well as the final exposure, were mapped against the long-read sequenced and annotated PAO1 reference genomes using the CLC Genomics Workbench with a length and similarity fraction of 0.7 and 0.99, respectively. For variant detection, a minimum coverage and count of 10 reads and a minimum frequency of 75% were applied. To detect larger structural variations, variant calling of the populations was performed with different parameters, allowing for varying degrees of stringency. A strict mapping was performed with previously mentioned parameters. Additionally, mappings with a length fraction of 0.5 and a similarity fraction of either 0.6 or 0.8 were carried out. For each population, the three different mappings were used with a minimum coverage and count of three reads and minimum frequency of 35% for variant detection. Variant tables obtained for each strain/population were compared and assessed based on their quality matrices (quality score, ratio, mapped reads, and coverage) before proceeding with further analyses. Genomic alterations that caused amino acid changes or were predicted to be deleterious were identified using PROVEAN v1.164 and SIFT with UniProt-SwissProt 2010_0965. Alterations occurring in transcriptional regulator binding regions were identified using the Pseudomonas genome database (version 20.2)66. Additional analysis of effects of non-synonymous mutations was carried out with PredictProtein v1.2.067. Alterations acquired through exposures were visualized in Muller plots using lollipop (https://pypi.org/project/lolipop/) in Python v3.7.4. For clinical isolates, variant detection was performed using the CLC Genomics Workbench as described above, using the sequence from the first isolate from each patient as a reference for mapping. Longitudinal subsequent isolates from the same patient were mapped to the reference isolate using a length and similarity fraction of 0.7 and 0.99, respectively. Variant detection was carried out with a minimum coverage and count of 10 reads and a minimum frequency of 75%, followed by functionality assessment of identified variants using PROVEAN64 and SIFT65. Bacterial fitness, motility, and biofilm production The impact on bacterial phenotypes, beyond antibiotic resistance, was assessed in terms of fitness, biofilm formation, and motility. These phenotypes were evaluated for isolates selected based on detected mutations. Bacterial fitness was determined by calculating growth curves and rates. A 20 µL aliquot of a fresh 0.5 McFarland bacterial culture added to 180 µL of CAMHB and incubated for 24 h at 37 °C in a 96-well plate reader (MultiSkanGo, ThermoFisher Scientific, Finland). OD600 measurements were recorded every 15 min for 24 h, and the resulting growth curve was analyzed with Growthcurver v0.3.168 to determine growth rates. Motility (swimming, swarming, and twitching) was examined as previously described69. Briefly, isolates were grown overnight on lysogeny broth (LB; Sigma, USA) agar of 1.5% agar (BD, France). Swimming agar plates consisted of 0.3% agar (BD), 0.5% NaCl (Merck), and 1% Bacto Tryptone (BD, France) (wt/vol). Swarming agar plates consisted of 0.5% agar, 0.5% glucose (Merck), and 2.5% nutrient broth (Oxoid, United Kingdom). Twitching motility agar plates were made of 1% agar and 2% LB. Plates were stab-inoculated, with swimming and swarming agar plates incubated for 12-16 h at 30 °C, while twitching agar plates were incubated for 24 h at 37 °C. Motility was quantified by measuring the diameter of the observed growth zones. Biofilm production was measured using a static model. 96-well plates with pin lids (Nunc, ThermoFisher Scientific) and bacterial cultures were prepared and inoculated as described for growth curves. After 24 h, plate lids were washed 3 times for 5 s with 0.9% NaCl solution before being fixed in methanol for 20 minutes. After air drying the lids, biofilms were stained for 15 min with crystal violet solution (2% crystal violet (Merck), 20% ethanol, and 0.8% ammonium oxalate (Ensure, Merck)). Excess crystal violet solution was removed by washing with tap water and the pin lids were air dried before being placed in 33% glacial acetic acid solution for 30 minutes. Plates containing the stain were used for measuring OD492 using a MultiSkanGo plate reader. Respiratory epithelial infection model Respiratory epithelial cells (A549) were grown in Dulbecco’s minimal essential medium (DMEM, Gibco, ThermoFisher Scientific) supplemented with 10% fetal bovine serum (FBS, Gibco). Prior to infection, A549 cells were seeded at a density of 2 × 104 cells/well in a tissue culture plate (Nunc, ThermoFisher Scientific) to achieve seeding density of ~4 × 104 cells/well on the following day for infection. For the infection process, bacterial isolates were grown overnight in CAMHB at 37 °C and 180 rpm. The bacterial cultures were then centrifuged (10 min at 4000 × g) and resuspended in PBS (Invitrogen, ThermoFisher Scientific, Lithuania). After a second centrifugation step, the bacteria were suspended in DMEM. The bacteria were added with a multiplicity of infection (MOI) of 1 cell:100 bacteria and wells were adjusted to a working volume using DMEM without FBS. Plates containing infected cells were placed at 37 °C with 5% CO2, and cytotoxicity and bacterial loads were determined every 4 h over a 36-hour period. Bacterial loads were quantified by serially diluting each well from 10−2 to 10−8 and spiral plating of dilutions 10−4 to 10−8 on MHA plates (Eddy Jet, program 6; 50 µL logarithmic spreading, IUL micro, Spain). Cytotoxicity assay Cytotoxicity for selected isolates was determined by measuring release of lactate dehydrogenase (LDH) using the LDH cytotoxicity assay kit (Abcam, United Kingdom). In vitro co-culture infection plates were spun down at 600× g for 10 min and the supernatant was collected. 10 µL aliquot of the supernatant was mixed with 100 µL of LDH reaction mix and incubated for 30 min. Absorbance was then measured at 450 nm and 650 nm. The percentage of cytotoxicity was calculated according to manufacturer’s instructions. Data were corrected for CFU counts. Statistical analysis All statistical analysis was performed using GraphPad v8.0.2 (Dotmatics, USA). Continuous variables were tested for normality using the Shapiro-Wilk test, followed by statistical comparisons using either parametric ANOVA or t-tests, or non-parametric Kruskal-Wallis or Mann-Whitney tests as appropriate. Correlation was tested by calculating the Pearson correlation coefficients. P-values for all applicable tests were corrected according to the recommended settings of the software. Additional analyses into patient statistics, treatment effects and patient outcomes were performed in SPSS v. 29.0.1.0 (IBM, USA). Supplementary information Supplementary Information Acknowledgements We would like to thank all the local staff and investigators with the sample and data collections in all the participating hospitals. We would like all the technical and other personnel helping with experiments, sample analyses and sequencing: Jelle Vlaeminck, Britto Basil Xavier, Gert Leten, Anouk Vanderstraeten, Kitty Gys, Liesbet Bryssinck, Sofie Van Mieghem, and Sabine Chapelle. We would like to thank the ASPIRE-ICU and EVADE research teams for their contribution and input. This research is supported by the Innovative Medicines Initiative Joint Undertaking under COMBACTE-MAGNET (Combatting Bacterial Resistance in Europe-Molecules against Gram-negative Infections, grant agreement no. 115737) and COMBACTE-NET (Combatting Bacterial Resistance in Europe-Networks, grant agreement no. 115523), resources of which are composed of financial contribution from the European Union’s Seventh Framework Program (FP7/2007-2013) and EFPIA companies’ in-kind contribution. Author contributions Conceptualization was done by S.M.K., Y.G. and S.K.S., Experiments and data analyses were performed by T.S., M.B., A.H., D.T., A.O., S.M.K., S.K.S. Writing the first draft was written by T.S. and M.B., the final draft was written by T.S., Y.G., S.M.K. and S.K.S. ASPIRE-ICU and EVADE trials management was performed by C.R., J.K., M.B., L.T., C.L., H.G., B.F., J.C., H.J., A.R., A.D., and M.E. All authors reviewed and approved the final manuscript. Data availability Data supporting the findings of this study are available within the paper and its supplementary information document. Sequence data generated and analyzed in this study have been deposited in the NCBI Sequence Read Archive with the accession code PRJNA685914 for ASPIRE-ICU, TBD1 for EVADE, and TBD2 for PAO1 long read sequences and evolutionary broths/isolates. All other data generated in this study are available from the corresponding author upon reasonable request. Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. A list of authors and their affiliations appears at the end of the paper. These authors contributed equally: Antonio Oliver, Youri Glupczynski, Samir Kumar-Singh. Contributor Information Surbhi Malhotra-Kumar, Email: surbhi.malhotra@uantwerpen.be. the COMBACTE-MAGNET EVADE study group: Jean Chastre, Bruno François, Marc Bourgeois, Apostolos Komnos, Epaminondas Zakynthinos, Galia Rahav, Nicolas De Schryver, Alain Lepape, Iftihar Koksal, Miguel Sánchez-García, Michael Joannidis, Walter Klimscha, Elisabeth De Waele, Jacques Devriendt, Vincent Huberlant, Pieter Depuydt, Sam Van Boxstael, Mladen Peric, Jasminka Kopic, Michal Hanauer, Tomas Hruby, Vladimir Sramek, Petr Svoboda, Tomas Vymazal, Martin Novacek, Djillali Annane, Jean-Paul Mira, Bertrand Souweine, Pierre-François Dequin, Ferhat Meziani, François Stephan, Saadalla Nseir, Sebastien Gibot, Carole Schwebel, Gaetan Plantefeve, Jean-Luc Diehl, Christian Richard, Christian Lamer, Kada Klouche, Samir Jaber, Georgios Filntisis, Spyros Zakynthinos, Antonia Koutsoukou, Georgios Saroglou, Charikleia Nikolaou, Glykeria Vlachogianni, Ioannis Pnevmatikos, Konstantinos Mandragos, Ildiko Kremer, Zsolt Dezso Rozgonyi, Zsuzsa Marjanek, Ignacio Martin-Loeches, Pierre Singer, Vernon Van Heerden, Yehuda Carmeli, Pedro Povoa, Antonio Alvarez Seoane, Pedro Moura, Filipe Gonzalez, Paula Ramirez, Antonio Torres Marti, Ricard Ferrer Roca, Lorena Oteiza, Dolores Escudero, Enrique Piacentini, Paula Vera, Luis Tamayo, Miguel Angel Gonzalez Gallego, Borja Suberviola Canas, Iglesias Figueira, Rafael Leon, Volkan Korten, Murat Akova, Duncan Wyncoll, Tony Whitehouse, Phil Hopkins, Malcolm Sim, Yoav Golan, Marcus Zervos, Jose Vazquez, Kartikeya Cherabuddi, George Smulian, Nadine Rouphael, James Welker, Mathew Sims, David Van Duin, Todd McCarthy, and Christopher Polk Supplementary information The online version contains supplementary material available at 10.1038/s44259-026-00199-3. References Associated Data Supplementary Materials Supplementary Information Data Availability Statement Data supporting the findings of this study are available within the paper and its supplementary information document. Sequence data generated and analyzed in this study have been deposited in the NCBI Sequence Read Archive with the accession code PRJNA685914 for ASPIRE-ICU, TBD1 for EVADE, and TBD2 for PAO1 long read sequences and evolutionary broths/isolates. All other data generated in this study are available from the corresponding author upon reasonable request.