Constraint-based modeling identifies new putative targets to fight colistin-resistant A. baumannii infections

Acinetobacter baumannii is a clinical threat to human health, causing major infection outbreaks worldwide. As new drugs against Gram-negative bacteria do not seem to be forthcoming, and due to the microbial capability of acquiring multi-resistance, there is an urgent need for novel therapeutic targets. Here we have derived a list of new potential targets by means of metabolic reconstruction and modelling of A. baumannii ATCC 19606. By integrating constraint-based modelling with gene expression data, we simulated microbial growth in normal and stressful conditions (i.e. following antibiotic exposure). This allowed us to describe the metabolic reprogramming that occurs in this bacterium when treated with colistin (the currently adopted last-line treatment) and identify a set of genes that are primary targets for developing new drugs against A. baumannii, including colistin-resistant strains. It can be anticipated that the metabolic model presented herein will represent a solid and reliable resource for the future treatment of A. baumannii infections.

for important pathogens such as Helicobacter pylori 7 , Pseudomonas aeruginosa 8 , Mycobacterium tuberculosis 9 , and Staphylococcus aureus 10 . Usually, such predictions are performed simulating growth in an arbitrarily defined medium, accounting for the main nutrients used by the microbe and without imposing any additional constraint to the model. Indeed, the search space of essential genes predicted can be narrowed by imposing additional constraints on the model. One possibility consists of modulating the flux admissible across each reaction on the basis of the expression values of the corresponding genes. By doing so, it is possible to generate context-specific models that reflect the actual set of reactions employed 11 . This approach promises to reduce i) the gap between the predicted and real cellular metabolic landscapes, and ii) the number of false positives/negatives in EGs predictions. Additionally, it might reveal hints for the synergistic use of antibiotics and, in particular, to the possible additional targets that might arise from the adaptation/response of a microbe's metabolism to a single antibiotic. Indeed, changes in gene expression might redirect the cellular metabolic fluxes in such a way that novel and untapped essential reactions may emerge, representing good candidates for a synergic antibiotic. Despite that the use of antibiotics in combination is sometimes questionable, this approach can be considered in cases of severe infections and it has been shown to be effective in the case of Pseudomonas and Acinetobacter spp. 12,13 .
Here, we explored the system-level metabolic consequences of A. baumannii exposure to colistin. We integrated gene expression data during exposure to colistin 14 with a newly reconstructed genome scale metabolic model, allowing for constraint-based modelling of the type strain ATCC 19606. Our data revealed the metabolic reprogramming that occurred in this strain following the establishment of a stressful condition such as the presence of an antibiotic. Furthermore, the metabolic reconstruction provided here represents an important resource for the future understanding of A. baumannii metabolism and for the detection and identification of novel drug targets.

Results and Discussion
Genome-scale A. baumannii ATCC 19606 model is consistent with large scale phenotypic data. A preliminary draft reconstruction of the A. baumannii ATCC 19606 metabolic model was obtained through the Kbase server (http://kbase.us). This was manually curated as described in Methods. Afterwards, we used previously published large-scale phenotypic data 15 to validate our reconstruction over a large set of experimental tests. Manual curation was performed by comparing FBA outcomes with such auxotrophies data (determined through Phenotype Microarray (PM) technology). During this process, the capability of our model to represent the observed phenotypes was tested.
Growth rates were firstly estimated in silico in simulated Simmons minimal medium (a standard bacteriological medium that contains only essential inorganic salts) under aerobic conditions by iteratively probing each C-source used in PM plates. During these simulations, biomass optimization was selected as the model objective function (O.F.). Results of the simulations (either "growth" or "no growth", i.e. the estimated flux value across biomass assembly reaction) were compared with the activity directly measured during an experimental phenotype microarray experiment, and discrepancies identified between the in silico and experimental data were manually adjusted as possible (such as by filling in missing transport reactions or metabolic gaps).
Following this procedure, we reached an overall agreement of about 84% between the in silico and experimental data: out of the 67 in silico screened metabolites, 24 were correctly found to be carbon and energy sources for A. baumannii ATCC 19606 (true positives) and 32 not (true negatives) while only 10 disagreements remained, 5 false negatives and 5 false positives. All data are briefly summarized in Fig. 1, and a detailed description of the outcomes of the comparison is reported in Supplementary Material S1, Supplementary

Constraint-based modelling identifies metabolic EGs. As already mentioned, the identification of
EGs is one of the key-steps in a drug discovery pipeline. Indeed, both general and condition-specific EGs can be distinguished 20 . The formers are required to sustain life under virtually all growth conditions; conversely, under specific constraints, changes of central metabolism may occur leading not only to a change in flux distribution throughout the network, but also to changes in gene essentiality and to the appearance of condition-specific essential genes. Hence, we systematically evaluated relevant switches in both unconstrained and constrained models (different scenarios), by imposing the necessary constraints to the metabolic reconstruction. Accordingly, screens for EGs were performed for multiple specific conditions: starting by simply changing the set of available nutrients (i.e. simulating different environmental niches) and then by simulating stressful situations such as antibiotic exposure and mutations (using available expression data in such conditions).
Nutrient availability influences identified metabolic Egs. As we were interested in modelling the system in a gradually constrained manner, we initially simulated an arbitrary rich medium, allowing our model to have virtually all the nutrients needed -as likely happens inside of a host 19 . To do so, we set the lower bound of exchange reactions as described in methods. Then we performed in silico gene deletions for each gene in the model. Accordingly, each gene was defined as essential if its elimination destroyed the network's ability to synthesize at least one key biomass molecules (i.e. the model predicts no-growth); otherwise, the gene was considered to be dispensable. Gene essentiality analysis was performed through both FBA and MOMA approaches (see Methods), which both lead to the identification of 67 EGs. Fig. 2A shows the values of the ratio between the predicted growth rate of the gene knock-out mutant and the wild type strain (GR ratio , see Methods). The complete list of EGs and their functions is reported in Supplementary Material S4. Next, we repeated EGs prediction by simulating growth in a minimal medium (Simmons medium, as defined in Supplementary Material S1, Supplementary Table 2). As shown in Fig. 2B, this approach labelled a total of 138 genes as indispensable for growth in this condition (see Supplementary Material S4 for the complete list). Differences emerging from these two simulations highlight how nutrient availability affects cell metabolism and, interestingly, how different environmental pressures influence gene essentiality. Particularly, 67 genes were predicted to be essential under both the tested growth conditions, while 71 are likely to become essential only when limited nutrient availability force the cell to reprogram its metabolic behaviour (i.e. in Simmons medium, Fig. 3A). Figure 2. GR ratio value for each gene deletion in rich (A) and minimal (B) media. Blue and red lines represent MOMA and FBA predictions, respectively. Please note that, in order to make the analysis more comprehensive, also gap-filling genes (i.e. those virtually coding for gap-filling reactions) were included, leading to a total of 1043 simulated knock-outs. Predicted EGs are consistent with available experimental datasets. A large body of data exists concerning A. baumannii gene essentiality. Here we used such information both to validate our EGs prediction and to understand whether the identified EGs sets are particular to A. baumannii ATCC 19606.
First we compared the EGs dataset obtained in the arbitrary rich medium to that obtained with an in vivo experiment on A. baumannii ATCC 17978 pathogenesis 21 . By using LB medium (a well-known bacteriological rich medium), Wang et al. 21 labelled 481 genes as essentials for that strain. However, not all of them were comparable with our predictions since a large fraction was neither metabolic or possessed an orthologous gene in A. buamannii ATCC 19606. In both cases these genes are absent in iLP844. For the same reason, not all the 67 EGs found through our simulation were comparable with the reported experiment. After performing all these necessary restrictions, we reached the result shown in Fig. 3B, i.e. 42 genes have been predicted to be essential by both approaches (in silico and wet-lab) for the two A. baumannii strains considered (Fig. 3B). A complete description of these EGs is provided in Supplementary Material S4 and represents an experimentally validated dataset in the context of A. baumannii drug target identification. Nevertheless, the two experiments show large discrepancies. The most likely reason for such inconsistency is strains genomic diversity, as previously reported for E. coli strains 22 .
Furthermore, our predictions in arbitrary rich medium were compared to those achieved performing the same analysis on the AbyMBEL891 model, an existing model of A. baumannii AYE 19 . In order to implement the simulation, it was necessary to perform a preliminary editing step on the AbyMBEL891 model, as the entire set of gene-reaction-rules was missing from the main reconstruction file. This difficulty in running the analysis highlights the need for a common protocol to be adopted during metabolic reconstruction and a standard to be reached in order to facilitate model re-use and data sharing among research groups. Nevertheless, after including the genes in the Abymbel891 model, we carried out single gene deletion analysis on both models, as described in the methods. As shown in Fig. 3C, 48 genes were predicted to be essential in both models, whereas 18 and 37 EGs were specific for A. baumannii ATCC 19606 and AYE, respectively. Information about the gene function are reported in Supplementary Material S4.
Comparisons were also carried out between our in silico predictions and wet-lab results in minimal (Simmons) medium. Specifically, we compared our EGs set to that obtained by Dorsey et al. through insertional mutagenesis experiments with A. baumannii ATCC 19606 23 , where the metabolic deficiency of insertion derivatives was subsequently confirmed, identifying essentiality of 10 disrupted genes. Repeating the assay in silico, our model correctly represented the phenotypes of the A. baumannii mutants, with 8 out of the 10 genes predicted as essential by Dorsey and colleagues also shown to be essential in iLP844 (Fig. 3D, Supplementary Material S4). Additionally, in 6 out of the 7 cases, A. baumannii ATCC 19606 model growth was correctly restored (as done in the corresponding wet-lab experiments) by adding to the minimal medium the metabolite(s) whose production was affected by the mutation.
Antibiotic treatment defines condition-specific models. Although a large fraction of the predictions was supported by previous experimental data, a possible source of error, using the methodology described above, stems from the observation that not all the reactions of the model will be active during growth in a given physiological condition. In particular, changes in gene expression are likely to influence the activity rate of the corresponding cellular metabolic reactions, leading to the observation that a given reaction can be considered 'turned on' or 'off ' on the basis of the expression levels of the encoding gene(s). Using available computational methodologies, it is possible to modulate the flux across each reaction on the basis of the expression values of the corresponding genes. This permits to take a picture of the current metabolic state and tightening up the predictive capabilities of the model itself. Accordingly, as the dynamic changes of metabolic reprogramming are likely mirrored by changes in gene essentialities, a possible solution for avoiding or reducing false positives is to merge transcriptomics data of the tested scenario into the genome-scale model.
Arguably, one of the most interesting physiological conditions of A. baumannii strains is the exposure to antibiotics and to colistin in particular 14 . Importantly, both the (metabolic) consequences and the occurrence of targets to be used in a synergic treatment are, currently, almost untapped. In order to study the dynamic changes of the metabolic network following antibiotic exposure and to derive a more realistic picture of gene essentiality patterns in a real scenario (antibiotic treatment), we used available transcriptomic data for A. baumannii ATCC 19606 in response to colistin treatment 14 . Up-regulation and down-regulation ratios (and corresponding P-values) of genes were combined with the iLP844 by using MADE (Metabolic Adjustment by Differential Expression) 24 . Briefly, MADE uses statistically significant changes in gene expression measurements to determine binary expression states (highly and lowly expressed reactions) i.e. reactions are turned on and off depending on the changes in mRNA transcript levels. Thus, by mapping gene expression data into the model, the in silico metabolic predictions are more consistent with the actual physiological state of the cell.
In the experiment by Henry et al. 14 , A. baumannii was grown in two different media, i.e. with and without 2 mg/L of colistin, and then sampled at 15 and 60 minutes after exposure. Following the described approach, we integrated the available transcriptomic data regarding all the metabolic genes embedded in our in-silico reconstruction (i.e. about 80 genes). Accordingly, we obtained four distinct models, each representing the predicted functional metabolic state of the cell at both 15 and 60 minutes, treated and untreated with colistin. These models differ in that some of their reactions are (completely) 'turned on' or 'off ' according to the measured levels of their corresponding genes. Afterwards, optimization of the four models was performed, allowing the analysis of flux distribution in the network and the occurring metabolic reshape.
Colistin exposure changes predicted metabolic fluxes in central A. baumannii pathways. In order to highlight changes in the overall metabolic behaviour and to identify changes on the metabolic rewiring occurring after antibiotic exposure, we compared flux distributions at the two time-points by calculating the flux ratio (RF ratio , see Methods) of treated vs. untreated models, for all the reactions. However, FBA only provides one of the possible optimal solutions out of many alternative (and feasible) cellular flux distributions. Hence, in order to correctly predict metabolic changes following antibiotic exposure we restricted the feasible solution space by performing Flux Variability Analysis (FVA) 25 . This approach allows estimating the minimum and maximum flux admissible across each reaction (under the same constraints as in FBA) and hence it can be used to estimate the correctness and accuracy of FBA predictions (see Methods).
As shown in Supplementary Material S5, according to FVA, the range of admissible flux is sometimes very large (spanning from the minimum −1000 to maximum 1000 mmol/g*h −1 in some cases), revealing the lack of accuracy in some of FBA-derived predictions. We here used FVA outcomes (as described in Methods) to filter out those reactions whose fluxes display little variation. In other words, each reaction was considered for downstream analyses only if both the maximum and minimum FVA predicted fluxes did not differ from the FBA predicted flux by more than 20%. Consequently, we were left with 901 reactions at 15 minutes and 970 reactions at 60 min. It is worth noting that several intervals of admissible flux ranges were tested and we report in Supplementary Materials S1, Supplementary Figure 1 the number of reactions filtered for each set of intervals. After carrying out this preliminary step, we observed the effects of the treatment at the metabolic level (for each reaction) by comparing the flux values in the untreated vs. treated condition.
Both qualitative and quantitative flux changes were analysed by dividing the reactions into three categories ('steady' , 'increasing' , and 'decreasing' , see Fig. 4) according to their trends in the examined experimental conditions. Also, we report a survey of the pathways in which they are involved in and their relative abundance for each category. Reactions' fluxes were considered 'steady' if their values did not change in the two conditions, otherwise they were defined to be 'increasing' or 'decreasing' according to the corresponding trend.
As shown in Fig. 4, at both 15 and 60 min time-points there is an increase in flux in most of the reactions. Interestingly, such change in flux mainly occurs in three biosynthetic pathways: fatty acid, peptidoglycan, and lysine biosynthesis. On the other hand, under the given constraints, there is a change in flux in some catabolic pathways (mainly involved in sugars and nucleotide metabolism). In our opinion, such a finding could be related to the rearrangement of the external membrane layer, a well-known effect of colistin treatment. If this is true, it is possible that the cell reacts to the antibiotic treatment by trying to repair the damage established by colistin while at the same time redirecting a certain amount of LPS components to catabolic processes.
Also, we would like to point-out that, although a down-regulation of certain genes involved in fatty acids biosynthesis was detected by Henry et al., here our data suggest that it does not necessarily imply a turning-off of the pathway. On the contrary, in our simulation fatty acid biosynthesis registers an increase in flux, probably as a side-effect of LPS disassembly as stated above. Colistin exposure changes gene essentiality patterns. According to the new constraints taken into account, gene essentiality was re-evaluated by calculating growth ratios (FBA and MOMA) at both 15 and 60 minutes after exposure to colistin. As for the analysis involving nutrient availability, shifts in gene essentiality emerged following antibiotic stress. The complete sets of the predicted EGs for each condition have been reported in Supplementary Material S4. As with the previous case, we can easily recognize genes likely to be essential in both conditions (treated and not) and, more interestingly, genes that emerged as essential only after the treatment. Specifically, following 15 minutes of colistin exposure, a total of 87 EGs were predicted: 66 were required both in presence and absence of colistin, but an additional 21 EGs were marked as condition-specific: 12 related to the non-treated model and 9 related to the treated one, reported in Table 2 and in Fig. 5A.
The same outline has been depicted in the second time-point condition (60 minutes): we identified 12 and 5 condition-specific EGs in the absence and in the presence of the antibiotic, respectively, see Table 2. Moreover, we found the same set of 66 EGs mentioned above (see Fig. 5B), suggesting that this represents a functionally relevant set of genes for sustaining growth in A. baumannii ATCC 19606. Interestingly, however, some genes switch from the 'essential' condition to the 'non-essential' one, following the exposure to colistin.  The two new sets of EGs show how changes in gene expression induced by the presence of the antibiotic might influence gene essentiality patterns in the strain ATCC 19606 and provide additional, nontrivial targets for drug design in such an organism. We also performed additional robustness analyses in order to test whether nutrients depletion occurring in treated and untreated samples during the in vivo experiments could impact the set(s) of predicted EGs. Specifically, the robustness of the number of predicted EGs in each of these conditions (i.e. treated vs. untreated samples) in respect to possible variations in the medium composition was assessed via random permutation. We tested up to 1,000 different nutritional compositions as described in detail in Supplementary Material S1, Supplementary Figure 2). The results showed that possible changes in the nutritional environment had only minor implications for the set of predicted EGs.

Locus Tag
Further, with the aim of discriminating whether the products of all the identified hypothetical EGs are A. baumannii specific or have orthologs in Homo sapiens, meaning that they would not represent good candidates for antibiotic treatment development, the sequences of the identified potential EGs were used to probe the human genome. Based on this BLAST 26 search (see Methods), we excluded from further studies those genes presenting more than 30% sequence identity with their human counterparts. Targeting of such genes is non-ideal since they may cause potential side-effects by perturbing critical components in the human body. All BLAST results are reported in Supplementary Material S6.
Among all the queries, we identified 46 (out of 73) general EGs and 4 (out of 12) condition-specific EGs that do not have any human orthologous. Thus, the 46 EGs represent valuable targets for further development of brand new drugs against A. baumannii ATCC 19606 infections. However, it is relevant to remark that, while these 46 general EGs could have been detected in several conditions, the other 4 condition-specific EGs are the result of specific constraints integrated in the model (gene expressions data). Hence, as already mentioned, they are nontrivial detections and they could represent a suitable horizon in the field of colistin-coupled treatment. The four genes, named HMPREF0010_00949, HMPREF0010_02972 and HMPREF0010_03445, HMPREF0010_01216 respectively encode a malonate decarboxylase (epsilon subunit), an arginine succinate synthase, a 3-4-dihydroxy-2-butanone-4-phopsphate synthase and a quinolate synthase. Interestingly, malonate decarboxylase epsilon subunit has already been characterized in the closely related organism Pseudomonas putida and labelled as an indispensable component of the enzyme for the cyclic decarboxylation ofmalonate 27 . However, to the best of our knowledge, no therapies targeting this protein have been developed to date. The product of 3-4-dihydroxy-2-butanone-4-phopsphate synthase is an intermediate in the biosynthesis of riboflavin. The enzyme requires a divalent cation, preferably Mg 2 + , to be active. The step becomes essential after colistin treatment as the antibiotic is predicted to cause an increase in flux through this pathway, probably following the shutdown of other parts of the network due to the down regulation of the corresponding genes. The arginine succinate synthase, is an enzyme catalysing the penultimate step in arginine biosynthesis (urea-cycle): the ATP-dependent ligation of citrulline to aspartate in order to form arginino-succinate, AMP, and pyrophosphate. The last EGs encodes for a bacterial specific enzyme that catalyses the second step in the de novo biosynthesis of NAD+ from aspartate.

EGs in colistin resistant A. baumannii.
Up to now, we have presented how metabolic reconstruction and mathematical modelling can be used to explore the strain's metabolic response during colistin treatment and how it can lead to the identification of novel potential drug targets. Our last attempt is now to illustrate how, starting from the same available experimental data, the model can be employed as a ready-to use blueprint in order to test new hypothesis.
As it was reported by Moffatt et al. 28 , the mechanism responsible for colistin resistance is linked to LPS. Specifically, mutations in the lpxA, lpxC, and lpxD genes have been reported as the main cause of LPS loss, thus abolishing the initial charge-based interaction with the antibiotic. Hence, to simulate an A. baumannii LPS (LPS − ) deficient and colistin-resistant strain, we removed this component from the biomass formulation in our genome-scale model. Then, to determine which genes are central for the cell's survival in such a condition, we used the transcriptomic data of the mutant strain in the presence/absence of colistin at 60 minutes 14 and mapped the data onto the new LPS − model. After this, we repeated the EGs prediction pipeline described above.
The analysis yielded a total of 55 and 72 EGs in the untreated and treated condition, respectively. Even in this case, the two sets share several elements (54 EGs) that remain mandatory for the cell in the two conditions (listed in Supplementary Material S4). Additionally, it is possible to observe that 18 genes become essential (reported in Table 2) only after antibiotic exposure: 9 of them were already found to be EGs in the wild type strain while 9 represent specific EGs of the mutant. Since the latter are non-trivial EGs (obtained through gene expression integration and introduction of an antibiotic resistance promoting mutation into the model) they have been re-used as seed for an additional BLAST search against the human genome (see Supplementary Material S6). The search led to the identification of 5 genes that do not have orthologs in humans: HMPREF0010_01215 encoding for glutamate-N-acetyltransferase (member of the ornithine acetyltransferase, OAT, family), HMPREF0010_00419, encoding for N-acetyl-gamma-glutamyl-phosphate reductase, HMPREF0010_01382 encoding for N-acetyl-L-glutamate-kinase, all of which are involved in the arginine biosynthesis pathway, HMPREF0010_0275979 that encodes for a glutamate-racemase and HMPREF0010_02344 that encodes for an adenosylhomocysteinase. This group of 4 EGs represents a potential achievement obtained from this work as it suggests specific targets to be taken into consideration when developing therapies in combination with colistin in an LPS-strain.
Predicted EGs are common in A. baumannii. Finally, we checked the distribution of EGs predicted for the strain ATCC 19606 within the entire A. baumannii species. The sets of predicted EGs were searched in all of the 1099 A. baumannii genomes sequenced to date, as described in the Methods. The overall result is shown in Supplementary Material S1, Supplementary Figure 3. The general trend observed was that more than 90% of the genomes analysed possessed the searched queries (identity >30%). Also, our analysis shows that this tendency is kept almost unchanged even when imposing an identity threshold greater than 50%, 70% and 90%. Accordingly, it can be stated that the possible target genes are broadly distributed and their sequence is conserved at the A. baumannii species level. Although we do not have any information about the EGs at such a wide level, this preliminary result is encouraging, since it expresses the possibility that the target genes we indicated for A. baumannii ATCC 19606 are probably common targets in most of A. baumannii type infections.

Conclusions
In this work, we have reconstructed and validated a genome-scale metabolic model of A. baumannii ATCC 19606. The model is comprehensive and accurate, as it covers ~23% of all CDSs in the genome of this microorganism and it was shown to have 84% agreement with Phenotype Microarray growth experiments. Based on the model's reliability, we applied constraint-based modelling to derive a global understanding of the behaviour of this metabolic system. By integrating gene expression data with constraint-based modelling we described the metabolic reprogramming occurring after colistin-exposure in A. baumannii and the changes in the pattern of gene essentiality during this stress condition. All the sets of condition-specific putative target genes that we propose have been compared (and partially validated) with the results obtained from experiments found in the literature. Some of these genes, although not yet experimentally validated, might represent primary targets for future research on the treatment of both the wild type and LPS-mutant (i.e. colistin resistant) strains. Our results have practical implications for the identification of new therapeutics as the identified essential genes can be used in drug-design pipelines. Moreover, we showed that the sequences of predicted EGs for the type strain ATCC 19606 are shared by most of the members of A. baumannii species, encouraging further research to check whether they are valuable drug targets for a larger number of strains than currently known. Finally, it can be anticipated that the iLP844 model illustrated herein represents a reliable and solid platform for further developments and the system-level understanding of the physiology of A. baumannii representatives and for the treatment of their infections.

Methods
Draft model reconstruction. We obtained a draft metabolic model of A. baumannii ATCC 19606 based on the genome annotation using Kbase automated reconstruction method (https://kbase.us/) 29 . This reconstruction was then thoroughly inspected and refined by integrating data from additional functional databases (MetaNetX 30 , Bigg 31 , Seed 32 , KEGG 33 ). Further integration was performed by searching for orthologous genes (genes likely having an identical biological function in a different organism) in closely related organisms (Acinetobacter baumanii AYE, Acinetobacter baylyi ADP1, and Escherichia coli) through a BBH (Bidirectional Best Hit) approach (inParanoid 34 ). Information regarding transport proteins was obtained probing the Transporter Classification Data Base (TCDB 35 ) and transportDB 36 .
In order to predict proper phenotypes, the general biomass producing reaction of Gram negative bacteria automatically generated by Kbase was substituted with a more accurate one that takes into account strain's specific components, which was recovered from the previously reported model of the related strain A. baumannii AYE, AbyMBEL891 19 . Two growth media were considered during the in silico simulations: Rich medium. Lower bounds of salts uptake reactions were set to −1000 mmol/g*h −1 in order to mimic non-limiting conditions. Carbon sources uptake reactions were set to −100 mmol/g*h −1 .
Simmons medium. Lower bounds of all uptake reactions accounting for the nutrients present in Simmons medium 39 (see Supplementary Material S1, Supplementary Table 2), were set to −1000 mmol/g*h −1 , to mimic non limiting conditions, only the C-source (citrate) was set to −5.
FVA. FVA analysis allows the determination of the span of possible flux variability (i.e. the maximum and minimum values of all the fluxes that satisfy the given constraints) while keeping the same optimal objective value. This approach has been used in this work in order to impose bounds to FBA flux predictions (f FBA ), which are notably non-unique. In fact, for any optimal solution found through FBA there may exist alternate flux distribution patterns yielding the same growth rate. Hence, the space of reliable FBA-flux predictions has been restricted by selecting only those that occur in the interval defined as follows: Gene essentiality and flux ratios calculation. Gene essentiality testing was performed by simulating deletion of each gene within the metabolic network and hence setting the associated reactions to carrying no flux (according to the corresponding Gene-Protein-Reaction (GPR) rule). To predict the growth of the mutant strain and determine the set of EGs, we used two different approaches, FBA and MOMA 40 . The main difference between them is that while the first predicts growth yield and metabolic fluxes based on the biological assumption of optimal growth, the second does not assume optimality of growth but approximates metabolic phenotype by performing distance minimization in flux space. The second approach has been shown to be more accurate in predicting lethal phenotypes 40 . The knocked-out gene was defined as 'essential' according to the results obtained computing the ratio (GR ratio ) between the simulated knocked out strain growth rate (µ KO ) and the one predicted for the wild type strain (µ WT ). Formulated as: Following this approach, if GR ratio = 0, then the knocked out gene is labelled as essential. Conversely, in case GR ratio = 1, the removal of the gene has no effect on the growth phenotype. Finally, when 0 < GR ratio < 1, the deleted gene was labelled as fitness-contributing gene, i.e. its removal partially affects the capability of the cell to produce biomass.
As MOMA and FBA predictions may lead to different essential gene sets 41,42,40 , we used both approaches to compute essential genes in all the conditions tested in this work. Although no major differences were observed, results obtained with both methods are presented throughout the manuscript.
In order to evaluate the range of the change in the carried flux of each reaction in the model following colistin exposure, we compute the ratio between the predicted flux in the treated vs. the untreated conditions as follows:

ratio T reated Untreated
Values of RF ratio equal to one indicate that no changes in the activity of the corresponding reactions were observed when simulating growth in the treated vs. the untreated conditions. Conversely, values of RF ratio between 0 and 1 or values greater than 1 will indicate a reduced or increased activity of the corresponding reactions in the treated condition, respectively. Finally, negative values of RF ratio will indicate those reactions whose directionality is predicted to change after the treatment.
Transcriptomics data integration and data visualization. In order to add transcriptional regulatory rules to the metabolic model, we imported the model from COBRA Toolbox into TIGER-1.2.0 0 (Toolbox for Integrating Genome-scale metabolism, Expression, and Regulation) framework (12). Then, the up-and down-regulation ratios of gene expression were mapped into the A. baumannii ATCC 19606 metabolic model by using MADE (Metabolic Adjustment by Differential Expression) 24 . The program uses significant statistical changes in gene or gene expression to create functional metabolic models. By adopting an optimization approach that applies Boolean rules, MADE connects reactions to the binary expression states of associated genes. The four arrays of genes to be switched-off yielded by MADE have been reported in Supplementary Material S8. H. sapiens and A. baumannii species. Protein sequences of the corresponding EGs found in A. baumannii ATCC 19606 were probed against the human proteome to test their validity as potential drug target in infections with this pathogen, i.e. to exclude any cross-interactions between the drug used for the treatment and human proteome elements.

EGs BLAST searches in
Queries were aligned to the protein sequences of H. sapiens using the default search parameters of the NCBI BLASTP online tool (BLOSUM62 matrix and gap costs equal to Existence 11, Extension 1). Results were considered positive (orthologous sequences found) if their sequence identity score value was equal to/greater than 30. In addition, the global distribution of EGs was evaluated at the A. baumannii species level by probing them against all of sequenced genomes retrieved at NCBI ftp site, i.e. 1099 genomes. Particularly, the focus was centred on EGs found in Simmons medium and in rich medium, as well as for those found after 15 and 60 minutes of colistin exposure. BLAST search parameters and analysis of the results were performed as described above.