Pseudomonas aeruginosa isolates from burn wound patients

We recovered and characterized P. aeruginosa strains from the exudates of three burn wound patients. The isolates were tested for antibiotic resistance (Table S1) and for clonal variability using binary arrays [10]. The isolates of the three exudates for which sufficient material was obtained for subsequent determination of in vivo gene expression were named P. aeruginosa PBCLOp10, PBCLOp11 and PBCLOp17. All three isolates were clonally distinct. The three strains belong to clonal complexes of variable abundance in the global population. The genotype of isolate CLOPBp10 is yet unique in a collection of currently 1,600 genotyped P. aeruginosa strains of independent origin, whereas the isolate CLOPBp17 belongs to a clone of intermediate occurrence (0.6% of the population). Isolate CLOPBp11 is a representative of the frequent clone D that worldwide is highly prevalent among keratitis isolates [11]. The antibiotic resistance profiles also differ between the clinical isolates and the control laboratory strains PAO1 and PA14 used as controls for susceptibility test. Strain PBCLOp10 had higher levels of resistance to piperacilin, as well as to antibiotics from the group of meropenems and aminoglicosydes, strain PBCLOp11 show resistance to ciprofloxacin. Strain PBCLOp17 emerged as being multi-resistant, with resistance to all antibiotics in our analysis except carbapenem – meropenem (full results Table S1). The antibiotics resistance pattern suggests that these strains were already specialized as nosocomial pathogens. Taking into account both the difference in the antibiotic resistance phenotype and the difference in the multimarker genotype, we conclude that the isolates are genetically distinct from each other, even though they were isolated from the same habitat and the same geographical site (albeit different patients at different times).

Genomic plasticity or niche specificity?

We compared the transcriptional profiles in vivo of the three different P. aeruginosa isolates from burn wound patients, with those determined in two alternative infection systems, namely a plant (lettuce) and a solid murine tumor, and two in vitro culture conditions (planktonic and sessile). The transcriptional profiles for each of the three independent strains under each of the five conditions were performed in duplicate. By treating the three P. aeruginosa strains as one data set for the purpose of the analysis (See Materials and Methods), a picture of general P. aeruginosa features emerged rather than those specific for the single strain. Genes that showed a significant expression difference (percentage of false-positives pfp<0.05) in at least one pairwise comparison (1734 genes) between each of the five conditions were grouped using hierarchical clustering with a 1-correlation algorithm as the distance metric. As a result, 213 groups of genes were resolved (at a cutoff of 0.15) (Fig. 1). These groups of genes subsequently assembled into five broader groups (at a cutoff at 0.9), where each of the five groups represented genes that were unique to each of the five studied conditions (the entire list of genes is presented in the tables S2, S3, S4, S5, S6 of the SI; pairwise comparisons between the various conditions are shown in table S9). The level of expression is depicted in the hierarchical cluster dendrogram as a heat map (Fig. 1) to facilitate the visualization and comparison of the degree of gene expression under the various conditions. These data clearly indicate that these three independent P. aeruginosa isolates regulate their genes in a similar manner when subjected to the same conditions. Altogether, these analyses indicate that while the genomic contents and plasticity of nosocomial P. aeruginosa strains are important in conferring specific capabilities upon infection, the niche where they thrive largely determines the prevalent transcriptional programs and, most likely, the resulting phenotypes.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 1. Hierarchical clustering of expressed genes, where gene expression is presented as a heat map.

Five broad groups of genes at a cutoff of 0.9 (indicated by the colored boxes) on the hierarchical cluster dendrogram represent signature genes in each of the five conditions. These genes are listed in Tables S2, S3, S4, S5, S6. Distance was measured by the 1-correlation algorithm.

https://doi.org/10.1371/journal.pone.0024235.g001

Assessing models of infection

Permutational multivariate analysis of variance (PERMANOVA) showed that the condition, under which each independent strain was subjected, significantly influenced its transcriptional profile (Pseudo F = 3.863, p = 0.0001). The five studied conditions were all significantly different to each other, with the exception of the burn wound and mouse tumor conditions, which were not significantly different (PERMANOVA Pseudo F = 1.5182, p = 0.1108) (Table 1). This is also clearly seen in the principal coordinate analysis (PCO) plot (Fig. 2), which shows that the transcriptional patterns of both the burn wound and murine tumor infections partially overlap. This underscores the view that the murine tumors represent a more appropriate model for a human infection with P. aeruginosa. By contrast, the in vivo infection of lettuce had a largely different expression pattern as compared to that of the burn wound or the murine tumor (p = 0.0316 and p = 0.0275), respectively. This multi-condition analysis shows that non-mammalian infection models still only partly reflect the real human infection condition and highlights the importance of in vivo transcription profiling of a pathogen directly from the infection site for the better understanding of the infection process and the underlying mechanisms.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 2. Global expression patterns of P. aeruginosa strains in the tumor infection (open circles) and those of the burn wound infection (closed circles) co-localized at the centre left of the PCO plot, while the global expression of the strains from the planktonic growth (asterisks) ordinated to the lower right of the plot, biofilm (open diamonds) ordinated upper of the plot and lettuce infection model (closed diamonds) ordinated at the right of the plot.

This result was confirmed by PERMANOVA, where all conditions were significantly different from each other (p<0.05) except the burn wound and tumor infection (p = 0.1108).

https://doi.org/10.1371/journal.pone.0024235.g002

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Pairwise comparison of the global expression patterns of P. aeruginosa strains between each condition.

https://doi.org/10.1371/journal.pone.0024235.t001

Niche-specific transcriptional signatures

The analysis of the in vivo gene expression profiles of the groups of genes mentioned above pinpoints sets of genes and circuits (signatures) specifically involved in the virulence and survival of P. aeruginosa in different settings (Tables S2, S3, S4, S5, S6, S9). Below we will focus on the in vivo infection settings and address some of the most prominent features for each of these.

Testing specific signature genes.

Among the genes expressed in all conditions, we filtered those that were expressed in one condition only and absent in the others. We obtained a highly specific list of genes expressed only in one condition in our experiments limited to only three up to eleven in number. Those genes are given in SI, figures S2, S3, S4, S5, S6. We tested representative unique genes from two amenable conditions, the plant infection and the biofilm. We used the transposon mutants of P. aeruginosa PAO1 and PA14 strains [31], [32] to verify if the filtered genes were indeed important for one condition and did not influence growth in other conditions. Figure 3 shows the biofilm formation assay results with mutants of the genes predicted to be part of the signature of biofilm formation (Figure S4), namely: PA1656 (being PA14_43050 the PA14 strain homologue of this mutant), PA1660 (PA14_43000), encoding part of a putative type VI secretion system and the hypothetical gene PA1123 (PA14_49850), as well as a gene specifically expressed in the plant infection PA5176 (PA14_68380). The mutants encoding the putative type VI secretion system were defective in the biofilm formation of up to 42% in the wild type strains (p-value<0.05, t values between 6.8–14.9). The mutant PA1123 was significantly impaired in its biofilm formation capability (p<0.05, t = 8.9), whereas PA14_49850 showed no significant change. Similarly, the mutant PA5176, which was shown to be specifically expressed in the plant infection, was slightly affected in its biofilm formation ability (p = 0.045, t = 2.65), whereas its counterpart PA14_68380 in the PA14 showed essentially no change. We also performed plant infection assays with representative mutants of genes predicted to be part of the: (i) plant infection signature, the conserved hypothetical PA5176 (PA14_68380) and (ii) biofilm formation signature, the gene from putative type VI secretion encoding operon PA1660 (PA14_43000) (Figure S6). The results show that the mutant strain PA5176 was twice more effective in plant infection (p-value<0.05, t = 5.1) than the wild type PAO1 (Fig. 4). The PA14 mutant PA14_68380 (PA5176 homologue) was almost three times more effective in proliferation in lettuce leaf (p<0.05, t = 8.6). When testing the biofilm signature mutant PA1660 (PA14_43000) in plant infection there was no significant difference in comparison to wild type strains (Figure 4). These results support the hypothesis formulated above on the genetic signatures for different infection conditions, which are largely independent of the strain variability and more strongly determined by the specific environments.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 3. Relative biofilm formation of P. aeruginosa PAO1 and PA14 strains and their transposon mutants evaluated by crystal violet assay.

An error bars were calculated from eight replicates and two independent experiments.

https://doi.org/10.1371/journal.pone.0024235.g003

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 4. Virulence of the P. aeruginosa PAO1 and PA14 wild type strains compared to their transposon mutants, PA5176 (PA14_68380) and PA1660 (PA14_43000) in the lettuce leaf assay.

The number of bacterial cells (as colony forming units, cfu) present in 10 g of lettuce midrib 3 days post injection is shown. Error bars were calculated from three independent experiment.

https://doi.org/10.1371/journal.pone.0024235.g004

We had a greater number of relevant mutants available in the PAO1 transposon library than in that of PA14 and we thus extended our testing with them. The selected biofilm signature genes, namely PA0263 – hcpC coding for a hemolysin co-regulated protein; PA1661 – another member of the putative type VI secretion system; and the hypothetical gene PA3906 were all impaired biofilm formation (p<0.05, t values between 10.2–15.4), but had no significant change in the plant infection (Figures S7 and S8). The plant infection specific gene PA0734 did not show significant difference in either plant or biofilm infection (Figures S7 and S8).

We thus showed that the predicted condition-specific genes are indeed important under these conditions for both PAO1 and PA14. The putative type VI secretion system here tested is yet unknown and will require further elucidation but it is clear from our results that it is required for biofilm formation on the abiotic surface. For the plant infection per se, this system is not needed. Quite interestingly, however, the mutant of the gene PA5176 (PA14_68380), which belongs to the plant infection signature, was more virulent to plants than the wild type counterparts. It is uncertain at this stage why would it have increased virulence, but it is clearly involved in plant infection. Sequence analysis of the protein encoded by this gene indicates that it has a conserved NUDIX hydrolase domain, typical of ADP-ribose hydrolases, which hydrolyze the ADP-ribose to AMP and ribose- 5-phosphate and is believed to be involved in maintaining the level of ADP-ribose in the cell. Some enzymes from this family have the ability to degrade potentially mutagenic, oxidized nucleotides whereas other control the levels of metabolic intermediates and signaling compounds [33]. Further research will have to be done to unravel the function and role of this gene in virulence.

Regardless of the minor differences between the strains, the joint expression profile analysis of these and the “control” conditions underscores the proposition that each “niche signature” consists of a “core” gene set common to all conditions (e.g. those responsible for DNA replication, ribosomal structure, see house keeping genes in SI Table S8), and a set of niche specific genes whose wiring support proliferation of P. aeruginosa in those various environments (See Tables S2, S3, S4, S5, S6). This observation hints at a “modular” structure of the transcriptional programs in this bacterium, which underscores its ability to thrive ubiquitously in the environment and opportunistic potential.

Conclusions

We report here the global analysis of P. aeruginosa gene expression under in vitro and in vivo conditions. We identified signature genes encoding proteins expressed specifically under each setting tested. Our study is also the first report of in vivo measurement of global gene expression by P. aeruginosa upon burn wound infection and sheds light on the genetic programs used by the pathogen under these conditions. One of the most valuable insights herein is that the main stress P. aeruginosa has to overcome is iron limitation. Although this is common to other microbial infections, it had not been previously shown for humans infected by P. aeruginosa, thereby opening possible new avenues for combating this threatening, ubiquitous pathogen. We also compared, for the first time, the in vivo behavior of genetically distinct P. aeruginosa strains under different infection conditions, namely burn wounds, solid tumors and lettuce, as well as under in vitro conditions. We showed for our set of strains that the P. aeruginosa core genome encodes a pool of genetic traits that are activated similarly under particular conditions regardless of the strain-to-strain genetic variability. The astonishingly high number of hypothetical genes that were differentially expressed only under in vivo conditions underscores the importance of in vivo studies as compared to in vitro studies for the understanding of the infection processes. These genes are obvious candidates for subsequent functional characterization and for research aiming at determining their function within the infection processes. Altogether, our data suggest that P. aeruginosa has a wide variety of conserved virulence mechanisms at its disposal, enabling this bacterium to effectively adapt to and survive in vastly divergent environments.

While focusing on P. aeruginosa, the insights obtained in this work point to general principles – the combination of general and niche-specific signatures, and the importance of in vivo expression profiling - that are likely to be relevant for the global understanding of pathogenicity in opportunistic infections.

Ethics Statement

Samples from burn wound patients were collected during regular wound debridement. The sampling was performed as a part of standard care procedure in the Clinic. The sample provided for this research was subtracted from the samples collected for routine microbiological tests, which are made on the regular basis, therefore no additional procedures were carried on the patients. Three samples were taken for preliminary studies in order to evaluate the possibility to extract bacterial RNA. Verbal informed consent for the isolation of bacterial RNA from those routine patient samples was considered sufficient and was obtained from patients in presence of two members of the clinic. The verbal informed consent was documented by the clinic member. Ultimately, for further research and prior to sample analysis the sampling procedure was duly approved by the Ethical Board of Silesian Medical Academy, which also suggested and provided the form for written informed consent in future.

Bacterial strains

Three independent P. aeruginosa strains were isolated from 3 different burn wound patients with a monoclonal infection of P. aeruginosa being treated at the Centre for Burn Treatment (CLO) in Siemianowice Śląskie, Poland. Only patients with a clinically confirmed P. aeruginosa burn wound infection were chosen for sample collection. Liquid exudates were taken from the burn wound surface using swabs and kept at room temperature in transport medium. The 3 P. aeruginosa isolates herein named PBCLOp10, PBCLOp11 and PBCLOp17 were verified to be unrelated clonal complexes by binary array genotyping [10]. The wild type P. aeruginosa strain PAO1 and its transposon mutants (PW2332, PW3030, PW3947, PW3955, PW3957, PW7598, PW9706) were obtained from Department of Genome Sciences at the University of Washington [31]. The wild type P. aeruginosa PA14 and its transposon mutants (IDs – 31367, 23266, 33144, 41430, 36721) were obtained from the mutant library from the Harvard Medical School (http://ausubellab.mgh.harvard.edu/cgi-bin/pa14/home.cgi) [32].

In vitro bacterial growth

The 3 P. aeruginosa strains were grown separately in 100 ml LB broth at 37°C and at 160 rpm until a 1∶10 dilution measures an optical density of 0.6 at 600 nm (equating to early stationary phase). Afterwards, cells were harvested for RNA isolation as described below. For biofilm growth, overnight pre-inoculum of each P. aeruginosa strain was diluted to an optical density of 0.05 at 600 nm in 10% LB medium, where plastic Permanox® slides (Nunc, Rochester, NY, USA) were suspended into 100 ml of the diluted culture. Flasks were kept at 37°C for 24 h without shaking. Once biofilm formation was observed on the plastic slides, slides were then washed 2 times in fresh LB medium then immersed into RNAprotect Bacteria Reagent (Qiagen, Hilden, Germany) and biofilms removed by a cell scraper. Cells were then harvested for RNA as described below.

Biofilm formation assay

A modified crystal violet microtiter plate test, as described before [34], was used to asses the biofilm formation of wild type P. aeruginosa PAO1/PA14 and its transposon mutants. Briefly, overnight LB cultures were diluted to an A₆₀₀ = 0.05 and 100 µl was inoculated in 96-well microtiter plate (PVC, BD Biosciences). Each strain was inoculated in 8 wells and the experiment was repeated with exchanging the position of strains on the plate. After 24 h incubation the cultures were withdrawn and wells washed with H₂O. 150 µl of crystal violet staining solution (0.1% m/v H₂O) was added to the wells and the plate was incubated for 30 min at room temperature. After incubation the staining solution was removed, wells washed with H₂O, and left to dry. The crystal violet was resolubilized with 200 µl Ethanol (95%) for 30 min at RT. The 150 µl of Ethanol solution was transferred to fresh plate and the absorbance was measured at the 550 nm.

Burn wound sample collection

From the same 3 burn wound patients (described above), liquid exudates were also taken from each burn wound surface prior to the wound cleansing, which is a regular treatment, using sterile forceps and immediately transferred to vials containing RNAprotect Bacteria Reagent. Each sample containing the buffer was vortexed, incubated for 10–15 min at room temperature and centrifuged for 15 min at 4000 x g. The supernatant was discarded and the pellet immediately frozen at −20°C. Frozen samples were transported on dry ice and further stored at −70°C until RNA extraction.

Murine tumor infection

Six-week-old female BALB/c mice were purchased from Harlan (Germany). All animal experiments were approved by the appropriate ethical board (approval ID 33.9.42502-04-050/09 obtained from Niedersächsisches Landesamt für Verbraucherschutz und Lebensmittelsicherheit, Oldenburg, Germany). Cells of the colon adenocarcinoma cell line CT26 (ATCC CRL-2638) were grown in IMDM medium (Gibco, Karslruhe, Germany) supplemented with 10% (v/v) fetal calf serum, 2 mM L-glutamine and 10 mM HEPES. Prior to their injection into mice, CT26 cells were trypsinized, washed and finally resuspended in phosphate-buffered saline (PBS). Cells at a concentration of 10⁶ were injected subcutaneously in the abdomen. Mice bearing tumors of approximately 4–6 mm diameter in size were intravenously injected with 5×10⁶ CFU of P. aeruginosa suspended in PBS. Three days post-infection (p.i.) mice were necrotized and infected tumors prepared for stabilization of bacterial RNA in the following way: the tumors were removed and excised into 2–4 pieces, put onto nylon filters for cell culture (70 µm pore size) and suspended in 2 ml of RNAproctect Bacteria Reagent in 2 cm petri dishes. The tumor was spent through the membrane with a sterile spatula. The resulting mixture in the RNAprotect reagent was collected and centrifuged for 5 min at maximum speed 16,000 x g. The pellet was immediately frozen at −70°C for further RNA extraction (described below).

Lettuce infection model

The protocol for using romaine lettuce leaves as a model of P. aeruginosa infection was performed as previously described [35], [36]. P. aeruginosa strains were grown aerobically overnight at 37°C in LB broth, washed twice with 10 mM MgSO₄, and diluted in sterile MgSO₄ to a cell concentration of 1×10⁸ CFU/ml. Lettuce leaves (Mini-Roma lettuce purchased commercially) were prepared by washing with distilled H₂O containing 0.1% bleach (sodium hypochlorite) followed by additional wash pure distilled H₂O. Lettuce mid-ribs were inoculated with 10 µl of bacterial suspension at a concentration of 1×10⁸ CFU/ml (corresponding to ∼1×10⁶ bacteria in total) by injecting the end of the plastic pipette tip into the rib and then all leaves were placed in plastic containers containing Whatman paper moistened with 10 mM MgSO₄. Lettuce was incubated at 37°C or 30°C, and symptoms were monitored daily over the course of 5 days. As a negative control, lettuce leaves were inoculated with 10 mM MgSO₄. A separate lettuce leaf was used for each strain. The experiments were repeated three times. Five days p.i. a 2 cm² piece from the original place of injection was excised from the leaf and immersed into 3 ml RNAprotect Bacteria Reagent. The sample was vortexed for 30 sec and the solid parts of the plant tissue discarded. The resulting mixture was incubated for 5 min at room temperature and centrifuged for 15 min at 4000 x g. The supernatant was discarded and the pellet frozen at −70°C for further RNA extraction performed as described below. For the PAO1 mutant infection assay the number of bacterial cells in the midrib were determined by CFU after 3 days of the incubation period.

Total RNA isolation

Analysis of gene expression of the pathogen directly at the infection site is limited by various factors associated with RNA extraction. The RNA extracted from clinical samples will most likely be a mixture of bacterial RNA with host RNA. Prior to RNA extraction samples were thawed on ice. RNA isolation was performed using RNAeasy Mini Kit (Qiagen) according to manufacturer's instructions with some minor modifications: samples were treated with 600 µl of TE buffer containing 1 mg/ml Lysozyme, incubated for 10 min with periodic vortexing every 2 min for 15 sec. Then, 1050 µl of RLT buffer containing 1% β-mercaptoethanol was added. The sample was vortexed and centrifuged for 2 min at maximum speed (16000 x g). The supernatant was transferred to a fresh 15 ml tube and 750 µl of absolute ethanol was added. The sample was loaded onto a spin column where the DNA was digested using RNase-free DNase I. The RNA was then eluted twice from each column with 50 µl and then 30 µl of RNase-free water. Eluted RNA was treated a second time with DNase I to ensure that all traces of genomic DNA were removed. The isolated RNA was stored at −70°C. The yield of the isolated RNA was measured by light absorption at 260 nm in an Eppendorf photometer. Integrity and purity was checked by formaldehyde gel electrophoresis or by 2100 BioAnalyzer (Agilent Technologies, Santa Clara, CA, USA).

Bacterial RNA enrichment

Since the extracted RNA from the in vivo samples contained both bacterial and eukaryotic RNA (i.e. human or mouse), the samples had to be enriched for bacterial RNA (Fig. S1). This was achieved with the MicrobEnrich Kit® (Ambion, Austin, TX, USA). The basis of the kit is to hybridize the eukaryotic ribosomal RNA and the messenger RNA to the magnetic beads. Hybridization occurs between specific 18S and 28S rRNA regions and polyA tails of eukaryotic mRNA. Since there are reports that bacterial mRNA including that of P. aeruginosa possesses polyadenylated (polyA) tails to some extent [37] all of the samples including the controls were treated for enrichment.

RNA amplification

One of the main bottlenecks in the transcriptomic analysis of host-pathogen interactions is the amount of bacterial RNA that can be isolated from the sample taken from the site of infection. The well established P. aeruginosa GeneChip® from Affymetrix (Santa Clara, CA, USA) requires 10 µg of RNA per microarray. Even with burn wound infections, where the amount of bacterial cells is relatively large this amount was never obtained. A promising way to overcome this obstacle was with the use of bacterial RNA amplification. This was performed using the MessageAmp Bacteria Kit (Ambion). The procedure consists of the following steps: i) total enriched bacterial RNA is treated with an enzyme polyadenylation polymerase to produce polyA tails, ii) single stranded cDNA is produced using reverse transcriptase and oligo dT primers, iii) second strain cDNA is produced, and iv) double stranded cDNA serves as a template for in vitro transcription using T7 RNA polymerase and T7 oligonucleotides. During the in vitro transcription reaction modified nucleotides were used: biotin-11-CTP (PerkinElmer Life Sciences, Waltham, MA, USA) and biotin-16-UTP (Roche Applied Science, Basel, Switzerland). As a result we obtained antisense biotinylated RNA, ready to use for GeneChip® hybridization.

Microarray hybridization

The use of amplified RNA is common with eukaryotic microarrays, but the original procedure for P. aeruginosa GeneChip® (Affymetrix) was for single stranded terminally labeled cDNA. Comparing to the standard procedure (without amplification), amplifying RNA for microarray analysis does not change the final outcome as previously reported [38]. The hybridization and washing steps were performed in the Affymetrix Array facility at Helmoholtz Centre for Infection Research (HZI) Braunschweig (Dr. Robert Geffers). Since the RNA had been amplified, some changes were introduced to the original Affymetrix protocol, such as that the RNA was fragmented using 5x fragmentation buffer instead of DNaseI treatment. The total amount of amplified and fragmented RNA used per chip was 7.5 µg. The process of sample preparation for each chip was as follows: the RNA from each single condition was pooled together after the enrichment; amplification step was performed on each condition; amplified RNA was hybridized in duplicate onto microarray chips in order to have technical replicates. Initial steps of data analysis were done at the Array facility using Affymetrix Microarray Suite Software 5.0 with default parameters. Once raw data files of scanned pictures were obtained further bioinformatic analysis was made as described below.

Microarray data normalisation

Data normalization and calculation of differential expression was performed with software from Bioconductor microarray analysis suite [39]. The quality of all chips was assessed by fitting a linear model to the probe level data using the function “fitPLM” from the “affyPLM” package. Subsequently, the distribution (boxplots) of RLE (Relative Log Expression) and NUSE (Normalised Unscaled Standard Errors) was manually analyzed. Expression values were computed using the “Robust Multichip Average” algorithm [40].

Differential Expression

As the number of replicates was low, the “Rank Products” algorithm was used to identify differentially expressed genes [41]. It has been shown that this algorithm performs well when the number of replicates is low [42]. The algorithm addresses the multiple testing problems by calculating for every gene an estimate of percentage of false-positives (pfp), if this gene and all genes with lower pfp would be considered as significantly differentially expressed. Thus, it is an estimate of False Discovery Rate (FDR). The value of 0.05 was accepted as a cut-off for pfp. The computations were performed using function Rpadvance from the “RankProd” package. A list of significantly differentially expressed genes was created for every pair of growth conditions (burn wound, lettuce, mouse tumor, planktonic and biofilm). Microarray data discussed here is MIAME compliant and have been deposited to the National Center for Biotechnology Information's Gene Expression Omnibus (GEO, http://www.ncbi.nlm.nih.gov/geo/) and are accessible through GEO series accession number GSE23007.

Hierarchical clustering of differentially expressed genes

Of all genes that showed a significant change of expression in at least one pairwise comparison between the experiments (1734 genes), hierarchical clustering was performed using the 1-correlation as the distance measure and clustering performed using the “hclust” function of the R software suite, http://www.r-project.org/. Subsequently, the outcome of the clustering was used to identify groups of genes that followed a similar pattern in changes to expression, where 213 groups were detected using a cutoff of 0.15. The expression values were averaged within each of the groups to form for each of them a group-wide expression pattern. Finally, the hierarchical clustering was performed on the groups, following the same procedure as previously described. The dendrogram, together with the heatmap composed of the group-wide expression values were plotted.

Specific gene signatures

To identify signature genes a set of five linear models was fit for every gene. Single model assumed that the expression change a gene occurs only in single condition (therefore five models for five conditions). As signature genes for a particular condition those were selected whose absolute value of linear coefficient for this condition exceeded 1.8 (the values of all five coefficients of a particular gene summed to 0). This value was chosen empirically, in order to achieve a reasonable number of signature genes. Afterwards, the expression patterns of chosen genes were inspected visually.

Ordination of the global gene expression profiles of P. aeruginosa strains within each condition

Each of the three independent P. aeruginosa strains at each of the five studied conditions were ordinated using principal coordinate analysis (PCO). A permutational multivariate analysis of variance (PERMANOVA) was performed to determine the statistical significance of the effect of condition on the global transcriptional profile of independent strains as observed in the PCO plot [43], [44]. These routines were performed using PRIMER (v.6.1.6, PRIMER-E, Plymouth Marine Laboratory, UK) [45], [46]. From the gene expression data matrix comprising those 1734 differentially expressed genes, a resemblance-matrix was generated using Euclidean distance by comparing the expression of each gene in regards to every pairwise combination of strains at each condition (burn wound infection, mouse tumor infection, lettuce infection, biofilm and planktonic growth). Euclidean distance is a widely used and accepted measure of distance in multivariate data such as gene expression data [47]. PCO ordination was used to reduce high dimensional data into low dimensional space and is equivalent to principal component analysis (PCA) when Euclidean distance is the distance algorithm chosen. PERMANOVA was performed according to a one-way experimental design. For one-way PERMANOVA, an exact P-value was generated using unrestricted permutation of raw data, where a Pseudo-F and corresponding p-value first report whether there are any overall differences in the gene expression between conditions. If there are, then the Pseudo-F statistic and generated p-values are reported for each pair of conditions. Pair-wise tests from PERMANOVA were also performed on unrestricted permutation of raw data. However, in the pairwise PERMANOVA, the total number of unique permutations when comparing the 3 strains across a pair of conditions was 10 and since the number of unique values under permutations was very low, the resulting p-values for each corresponding t-value can not be any smaller than 0.1. Thus, when only low unique values in the permutation distribution were available, asymptotical Monte Carlo p-values were generated instead of permutational p-values (9999 Monte Carlo samples). The ability to generate significance tests for small sample sizes made this approach powerful. That is, PERMANOVA seemed to be more powerful than other classical multivariate analyses allowing partitioning of variance components [48]. The conditions were considered significantly different if the p-value falls <0.05.