We investigated the phenotypic variability among clinical isolates in a diverse panel of 31 P. aeruginosa strains, including 28 isolates from infected patients (Yan et al., 2017) and the 3 types of strains PAO1, PA14, and PA7, all of which have their genomes sequenced. We grew each strain in synthetic media using glycerol as the sole carbon source and ammonium sulfate as the sole nitrogen source, producing a carbon-to-nitrogen molar ratio of 7.0. Then we used the drop collapse assay (Figure 1A) to classify each strain as surfactant absent, moderate, or high. Previous work had shown that glycerol enables the biosynthesis and secretion of surfactants in the strain PA14 (Boyle et al., 2015), a finding that we confirmed here with the wild-type PA14 and its isogenic mutants in the rhamnolipid synthesis. As expected, the wild type caused drop collapse indicating surfactant secretion, the ΔrhlA mutant did not, indicating loss of surfactants, and the complementation of the ΔrhlA with the L-arabinose inducible P_BADrhlAB restored the drop collapse but only when L-arabinose was added, as expected (Figure 1B). Among the clinical isolates, however, there was a wide variability of surfactant phenotypes. The strongest producers (n=17) showed phenotypes similar to the wild-type PA14, while non-producers (n=8) showed no surfactant secretion similar to the ΔrhlA mutant; moderate producers (n=6) showed a phenotype in between. There was no obvious relation between the body site infected in the patient and the surfactant phenotype (Figure 1C). There was also no coherent distribution with the isolates’ phylogeny, which we built using the sequence variation in core genes, also supported by a poor correlation in Moran’s I test of p=0.14. Ancestral reconstruction of the surfactant secretion trait supports that the ability to make surfactants from glycerol was present in the common ancestor of all strains (Figure 1—figure supplement 2). Importantly, the genes rhlA, rhlB, and rhlC, the genes encoding the secondary biosynthesis pathway, were present in all 31 isolates, and their sequence was 100% conserved. This indicates that the absence of surfactant secretion was not due to a loss of the biosynthetic pathway. Instead, we hypothesized that the phenotypic variation was due to the life histories of the strains: some lineages may have lost their ability to rewire their regulatory network to overcome the specific selective pressures that they had faced during their evolution.

We further investigated whether variation elsewhere in the genome—particularly in the accessory genome (the set of genes missing in at least one strain)—could explain the phenotype differences. Among the 8290 accessory genes, 354 are only absent in those non-producers but present in all mild and strong producers (Supplementary file 1). In a principal component analysis (PCA) of the presence-absence profiles, there was no visible grouping of isolates by their surfactant phenotype (Figure 1—figure supplement 2), which is consistent with the lack of association with the phylogenetic tree built from the core genome.

To search for a general explanation, we conducted analyses that showed producers have faster primary metabolisms in the same media of drop collapse assay. This finding was unexpected because, if anything, secondary metabolism should divert resources from primary metabolism (Yan et al., 2018): strains that produce surfactants would grow slower, not faster. We started by noting differences among the growth curves of the 31 strains, as assessed by the dynamics of the population optical density at 600 nm (OD₆₀₀). The maxima that the population densities reached showed no obvious difference in producers vs non-producers, but the cultures did vary widely in the shape of their lag phase and the exponential phase (Figure 2A). The shape of the growth curve, as a whole, was not predictive of surfactant secretion because hierarchical clustering using the entire time series failed to separate producers and non-producers (Figure 2—figure supplement 1). Instead, the slope of the exponential growth phase was steeper in producers. This was also shown using non-negative matrix factorization (Lee and Seung, 1999), which could decompose each growth curve as a weighted sum of three basis functions (i.e. features). This analysis showed that—although the growth curves of the producers (orange lines; including mild and strong producers) and non-producers (blue lines) largely overlapped (Figure 2A)—producers had higher weights for base 1, which agrees with a faster growth rate in the exponential phase (Figure 2B). This was also confirmed by dividing each growth curve into three phases (Figure 2C and Figure 2—figure supplement 2) and defining seven quantitative features to characterize each growth phase (Figure 2D and Supplementary file 2). Random forest classification revealed that the top two features associated with surfactant production were the maximum and the average specific growth rates in phase I, the exponential growth phase (Figure 2E). Features of phase III (the decay phase), such as the starting value of OD₆₀₀ in phase III which represents the final biomass produced by different strains in phase II, were among the least significant.

These analyses indicate that the ability to make surfactants, which typically peak when growth slows down from exponential (Latifi et al., 1996; Boyle et al., 2015; Mellbye and Schuster, 2014), is associated with the speed of the exponential phases, and hence, an efficient primary metabolism on glycerol. This suggests that the non-producer lineages had particular life histories that are selected for specific metabolic adaptations, and that those adaptations had the collateral damage of impacting their ability to grow on glycerol as a sole carbon source.

We then searched the composition of the intracellular metabolome for differences that distinguish producer and non-producer lineages. We extracted the metabolites from all our strains except for two non-producers (M55212 and F23197), which grew too slowly during the transition between phase I and phase II when rhamnolipid production begins (Materials and methods). Liquid chromatography-mass spectrometry (LC-MS)revealed 67 metabolites whose identities are known, and their abundances varied significantly across the tested strains (Figure 3—figure supplement 1A–B). Hierarchical clustering and PCA (Figure 3—figure supplement 2A–B) showed a reasonable separation between producers and non-producers. Fitting the data using the orthogonal projections to latent structures-discriminant analysis (OPLS-DA) (Cloarec et al., 2005) confirmed that there is a significant association between the broad metabolome and surfactant secretion (Figure 3A, R²=0.82, Q²=0.66, p-value = 5e−4). A Mann-Whitney U test revealed the 15 metabolites whose abundances differed most significantly between producers and non-producers. Then, we used those 15 metabolites to identify pathways perturbed in non-producers using the FELLA algorithm for pathway enrichment analysis (Picart-Armada et al., 2018; Figure 3—figure supplement 3, Supplementary file 3). The most perturbed pathways were the TCA cycle and amino acid metabolism (Figure 3B). Three out of six metabolites in the TCA cycle—fumarate, succinate, and citrate—were significantly changed: fumarate and malate had positive loadings, which indicated higher levels in producers. Citrate, succinate, and to a lesser extent cis-aconitate and alpha-ketoglutarate had negative loadings, which indicated higher levels in non-producers (Figure 3B). Pyruvate remained relatively constant across all the strains (Figure 3—figure supplement 1A), implying that the differential responses in the TCA cycle were independent from the changes in its upstream central carbon metabolism.

Besides the TCA cycle metabolites, most of the annotated compounds in the metabolism of branched chain amino acids (leucine/isoleucine, valine) and sulfur-containing amino acids (cysteine/methionine) were more abundant in non-producers than producers (Figure 3—figure supplement 3). A notable exception was formylmethionine (fMet), which was lower in non-producers. To place this observation in context, we reanalyzed previous metabolomics data (Boyle et al., 2017) from an experiment where we had compared PA14 and its ΔrhlA mutant. Interestingly, the ΔrhlA mutant also had lower levels of fMet compared with the wild type (Figure 3—figure supplement 4). The ΔrhlA mutant grows just as fast in glycerol as the wild type (van Ditmarsch and Xavier, 2011), which means that the link between lower fMet and lack of surfactant secretion is not simply due to a growth defect.

Several theoretical arguments support that non-producers might have slower primary metabolisms because they suffer from higher oxidative stress than the producers. First, the major differences include TCA intermediates (Figure 3B). The TCA cycle harbors five enzymes with Fe-S clusters, aconitase A, aconitase B, succinate dehydrogenase (SDH), fumarase A, and fumarase B (Py and Barras, 2010), which make them especially vulnerable to oxidative stress (Figure 3C). If non-producers are less successful at removing the oxidative stress from growth in glycerol, then the stress accumulated could reduce flux through the TCA cycle, which would slow down growth. The accumulation of succinate and depletion of fumarate in non-producers could be due to a reduced activity of SDH under oxidative stress: SDH is a membrane-bound dehydrogenase linked to the respiratory chain—a major site of ROS production in the cell—and also a member of the TCA cycle that catalyzes the oxidation of succinate into fumarate (Hederstedt and Rutberg, 1981). Since SDH contains [2Fe-2S], [3Fe-4S], and [4Fe-4S] clusters (Ayala-Castro et al., 2008), ROS that damages Fe-S clusters could decrease SDH activity in vivo.

Second, we saw that non-producers also included intermediate metabolites in amino acid biosynthetic pathways which are also substrates of Fe-S containing enzymes (Figure 3B). For example, the glutamate synthetase has a large chain (encoded by gltB) and a small chain (encoded by gltD), and both subunits contain Fe-S clusters that catalyze the production of glutamate from oxoglutarate and glutamine (Curti et al., 1996). Non-producers had slightly higher oxoglutarate and glutamine as well as lower glutamate than producers.

Finally, we used a whole-genome reconstruction to model the intracellular metabolic fluxes during growth in the glycerol medium (Nogales et al., 2017; Nogales et al., 2020). The rhamnolipids are produced when carbon is in excess (Boyle et al., 2015; Mellbye and Schuster, 2014) and, in the simulations, rhamnolipids secretion occurred when C:N flux ratio >6.3 (Figure 4—figure supplement 1). We then set C:N to 10.0, which exceeds the minimum threshold, and we simulated various levels of redox stress level by changing the flux through three redox molecules—NADH (reduced nicotinamide adenine dinucleotide), NADPH (reduced nicotinamide adenine dinucleotide phosphate), and GSH (reduced glutathione)—responsible for the bulk of cellular electron transfer and the main sources of reactive oxygen species (ROS) (Xiao and Loscalzo, 2020). Although redox stress is ultimately caused by high ROS levels, not fluxes, and ROS was not an explicit variable in our model, we assumed that ROS could perturb intracellular metabolic fluxes, particularly the fluxes of redox molecules, and we then asked how cell growth and surfactant synthesis responded to these ROS-mediated indirect perturbations.

For all three redox molecules, we found that the maximum growth rate was maintained at an intermediate flux range (redox homeostasis; Figure 4A–C, upper panels, white area). The growth rate decreased gradually as the flux of redox molecules changed in either direction away from this homeostatic interval (gray shading). The reason for the gradual decrease in growth rate is that deficiency or excess of these redox molecules which participate in the redox control of a great variety of biological processes disturbs the primary metabolism for growth. Importantly, before the growth starts slowing down due to an excess in redox species there was an abrupt shutdown in the overflow on carbon into many central carbon metabolites and—importantly— in the secretion of the HAAs, mono- and di-rhamnolipids that makeup the surfactants (Figure 4A–C, lower panels). The shutdown of overflow reactions under excess of redox molecules mimics the metabolic state in the non-producer strains who lose metabolic plasticity to overflow metabolites without compromising growth (Vemuri et al., 2006). When we forced these secretion fluxes to zero—to simulate a ΔrhlA mutant—the simulation did not predict a growth benefit. This result agrees with data showing that the ΔrhlA mutant has no growth benefit in glycerol media compared with wild type (van Ditmarsch and Xavier, 2011).

This also suggests that regaining surfactant synthesis should not restore fast growth in glycerol to non-producers, which agrees with the idea that loss of surfactant secretion is a consequence—not a cause—of the slower growth in non-producers. Notably, most of the overflowed metabolites are carbon-rich molecules, including HAA, mono-, and di-rhamnolipid (Xavier, 2011). Collectively, these theoretical arguments support a link between the ability to make surfactants and oxidative stress: fast growth and surfactant secretion are both impacted by the accumulation of ROS produced by primary metabolism.

Our model suggested that non-producers suffer from higher oxidative stress. Pseudomonas have a variety of antioxidant enzymes including catalases, glutathione reductases, NADH peroxidases, and cysteine-based peroxidases (Mishra and Imlay, 2012), which in PA14 and PAO1 were only made when the growth rate maintained its maximum value under excess carbon (Xavier et al., 2011; Boyle et al., 2015; Mellbye and Schuster, 2014). Isogenic mutants lacking some of those could make surfactants (Vinckx et al., 2010). If the slower growth in the clinical non-producer strains is due to impairments in some of those mechanisms to reduce oxidative stress, then this could affect surfactant biosynthesis.

To test this hypothesis, we compared the dynamics of hydrogen peroxide (H₂O₂) during growth in glycerol (Figure 4D and E). H₂O₂ is a representative ROS that can diffuse freely between cells and the environment; we can access its dynamics by measuring the fluorescence by the Amplex assay, where the difference between cell cultures and baseline reveals each strain’s net ability to reduce H₂O₂ (Zhao et al., 2012). All strains except for M1608, one of the non-producers, quenched the signal over the first 18 hr (Figure 4E). Consistent with our hypothesis, the non-producers (blue lines) removed less H₂O₂ than moderate producers (green lines) and strong producers (red lines). To factor out the possibility that the non-producers performed worse due to lower cell density, we calculated the H₂O₂ removal rate per cell. We observed a similar relative trend of H₂O₂ degradation between non-producers and producers (Figure 4F): a linear mixed-effect model quantified the difference that mild and strong producers have in per-cell H₂O₂ removal rates, which was in both cases significantly higher than in non-producers (p<0.001; Figure 4G). This is consistent with the notion that non-producers suffer more from oxidative species, which can be caused by high production or slow degradation.

To identify the genes that would explain the varying tolerance to oxidative stresses, we performed a Spearman correlation between gene presence or absence and the time averaged H₂O₂ removal rate per cell. No gene passed the significance threshold (p<0.05) after Benjamini-Hochberg correction for multiple hypothesis testing. It is possible that oxidative stress tolerance is a system-level phenotype that depends on the interactions among many genes in multiple pathways. Alternatively, the phenotype can be determined by a small number of, but different genes for different strains. For example, the worst strain in this assay, M1608, lacks katE (Supplementary file 1), a catalase that degrades H₂O₂ (Mulvey et al., 1988), and another non-producer, F5677, lacks the redox sensor soxR (Hidalgo et al., 1998).

To investigate transcriptomic differences associated with the surfactant phenotype, we selected 11 representative strains (6 strong producers, 3 mild producers, and 2 non-producers), in proportion to the number in each category. We then assessed their transcriptomics in Phase II, which is typically when producers start their surfactant biosynthesis (van Ditmarsch and Xavier, 2011). A PCA of the gene expression profiles could not cluster the strains according to their phenotypes (Figure 5—figure supplement 1). We then turned to RLQ analysis (R, L, and Q are explained in Figure 5A) to find both genes (Figure 5B) and pathways (Figure 5C) associated with rhamnolipid production.

As expected, the genes in the biosynthesis of rhamnose (rmlA, rmlB, rmlC, rmlD), rhamnolipid biosynthesis (rhlA, rhlB, rhlC, rhlR, rhlI), and quorum sensing (lasA, lasB, lasI, pqsABCDEH) were significantly less expressed in the non-producers (Figure 5B, Figure 5—figure supplement 2). Interestingly, in vitro data shows that the expression of genes from the above list, including lasI, lasR, rhlI, rhlR, pqsA, pqsR, in P. aeruginosa PAO1, can be significantly reduced by quorum sensing quencher and oxidative stress by exogenous H₂O₂ (Mohamed et al., 2020), supporting that the loss of the phenotype is linked to oxidative stress. More importantly, the zwf gene encoding glucose 6-phosphate dehydrogenase (G6PDH) was expressed higher in non-producers (Figure 5—figure supplement 2). G6PDH, a major source of NADPH that contributes to antioxidant defenses by reducing oxidized glutathione, was expressed higher, supporting our hypothesis that non-producers experience higher oxidative stress. Other genes, including the peroxide-sensing transcriptional regulator (oxyR), catalases (katA, katB), alkyl hydroperoxide reductases (ahpC, ahpF), glutaredoxin (grx), glutathione reductase (gor), and thioredoxins (trxA, trxB1, trxB2) were not highly expressed in the producers (Figure 5—figure supplement 2). We speculate that the stress levels may occur locally within the cells (e.g. within SDH) so that the oxidative stress response is also local. Another possibility is that the antioxidant defense may be regulated at the metabolic level. A notable example is biosynthesis of phenazine, whose genes (e.g. phzA1/B1/B2) are highly expressed in the producers but not in the non-producers. It was previously shown that the phenazine pyocyanin plays a role in the reduction of oxidative stress as it can restore growth of the ΔoxyR mutant in rich medium (LB) and undergoes H₂O₂ oxidation (Vinckx et al., 2010).

The RLQ analysis also revealed the top five pathways enriched specifically in each phenotype group: strong producers were enriched in acarbose and validamycin biosynthesis (pau00525), phenazine biosynthesis (pau00405), polyketide sugar unit biosynthesis (pau00523), starch and sucrose metabolism (pau00500), and pentose and glucuronate interconversions (pau00040) (Supplementary file 4). By contrast, non-producers were enriched in the biosynthesis of siderophore group nonribosomal peptides (pau01053), ascorbate and aldarate metabolism (pau00053), monobactam biosynthesis (pau00261), D-glutamine and D-glutamate metabolism (pau00471), and sulfur metabolism (pau00920).

So far, we have shown data supporting that surfactant producers can reduce ROS better than non-producers. This explains why strains that grow slower in glycerol also tend to produce less surfactants from that carbon source, likely due to the impact of ROS on growth. It is even possible that the ability to direct metabolic fluxes to surfactant secretion is part of a multifarious response to oxidative stress. Previous work in an engineered strain of P. putida showed that inducing surfactant synthesis reduces production of the final electron acceptor CO₂, suggesting that directing metabolic flux to surfactant secretion could indeed serve as an alternative electron sink (Tiso et al., 2016).

To investigate the link with oxidative stress further, we measured the expression of the rhamnolipid synthesis operon rhlAB to extrinsic oxidative stress using PA14 genetically engineered with a reporter fusion (P_rhlABgfp). This strain allows tracking the expression of rhlAB through a dynamic readout of green fluorescence (Figure 6). A PA14 carrying gfp under a constitutive promoter, P_A1/04/03gfp (Lambertsen et al., 2004), was used in parallel to account for growth effects on the gfp dynamics. We then added H₂O₂ at different doses always at the time when the culture started entering the exponential phase of growth in glycerol. In both PA14::P_rhlABgfp and PA14::P_A1/04/03gfp, the start of the exponential phase was increasingly delayed as the concentration of H₂O₂ was larger (Figure 6A and C), but once the exponential growth started there was no difference in the growth rate under different H₂O₂ concentrations (Figure 6E). This indicates that PA14 can restore normal operation of its primary metabolism after removal of the extrinsic stress.

The apparent removal of H₂O₂ during the lag phase may be achieved by metabolic rewiring that affects rhamnolipid production in the late exponential phase (phase II). Similar to growth dynamics, the promoter activity of both mutants was also delayed, and the time delay is longer at higher H₂O₂ concentration (Figure 6B and D). We determined that the ratio of the median P_rhlAB expression relative to the median P_A1/04/03 increased linearly with the H₂O₂ concentration (Figure 6F, orange). Since P_A1/04/03 measures the growth dependence of promoter activity (e.g. changes in number of RNA polymerases and ribosomes), the increased ratio showed that extrinsic redox stress can induce rhlAB expression via a growth-independent mechanism that likely involves transcriptional activators.

We next compared the above results obtained using PA14 wild-type strain and the ΔrhlA mutant. The gfp dynamics showed that in the absence of H₂O₂ the ΔrhlA strain still attempts to express rhlAB at a similar level as the wild type (p=0.3). We observed no decrease in the growth rate with increasing H₂O₂ concentration (Figure 6E), which indicates that the lack of surfactant biosynthesis had no detrimental impacts on the primary metabolism. Similar to the PA14 wild-type strain, H₂O₂ increased the expression of the rhlAB promoter relative to the constitutive promoter in the ΔrhlA mutant (Figure 6F, blue), but the baseline promoter activity increased significantly compared to the wild type (1.14×, p<0.01). Taken together, these results indicate that the genes for surfactant biosynthesis can be induced by extrinsic redox stress, and that the induction is stronger if surfactant synthesis fails to start.

All analyses so far were done in glycerol, a carbon source relevant for P. aeruginosa rhamnolipid production in bioremediation (Zhao et al., 2021) and clinical settings (Scoffield and Silo-Suh, 2016; Abdel-Mawgoud et al., 2014). But could different environments lead strains that are non-producers in glycerol to produce surfactants, and vice-versa? To address this question, we developed a predictive model from data and then sought out to test its predictions experimentally.

The model is a function f that takes as input a given carbon source mi and predicts whether a strain produces surfactants from that source. To parameterize the model we obtained a dataset using the BIOLOG phenotype microarrays PM1 and PM2a; these two arrays together contain 190 distinct carbon sources in separate wells of two 96-well microtiter plates. We used the same composition as in the glycerol media (minus the glycerol). Then, we grew each of the 31 strains in each of the 190 nutrients and negative controls (wells without carbon source) for 24 hr (shorter than 48 hr used previously to facilitate high throughput). We measured the final absorbance (OD₆₀₀) to quantify the ability to grow in that nutrient (Figure 7A; Figure 7—figure supplement 1A). The data showed that D-glucose (also known as dextrose) was the best carbon source for growth. Also, we saw 40 carbon sources (including the non-metabolizable isomer of D-glucose, L-glucose) in which none of the strains could grow on (OD₆₀₀ <0.05).

Then we assessed the ability of each strain to secrete surfactants from each carbon source and gave a score of 0 (absent), 1 (moderate), or 2 (high) using the drop collapse assay. To avoid experimentalist bias, the drop collapse was done by an experimentalist blind to the surfactant secretion score that we had determined previously in glycerol. The final dataset totaled N=6014 data points (31 strains grown on 194 conditions each, including the 190 carbon sources and 4 negative controls)—a sizeable dataset that we used to parameterize the model.

As with any high throughput approach, the BIOLOG data came with a loss of precision. We could assess that imprecision first by looking at three positive controls—tween20, tween40, and tween80—which come in the PM1 BIOLOG plate and are themselves surfactants, which means that they should produce a score of 3 regardless of surfactant secretion. The error of failing to identify collapse in these compounds was 4.2%. Then, we collected additional BIOLOG data on the PA14 ΔrhlA, which cannot produce surfactants. The rate of false identification of collapse in this strain was 1.3%. Finally, we used the fact that the PM1 plates contain glycerol in one of the wells to compare those results with our previous surfactant secretion scores (Figure 1). The assay correctly determined 8/8 non-producers, which amount to a 100% specificity rate. It failed to detect 10/23 producers, which amount to a 43% sensitivity rate. Taken together, these results show that BIOLOG assay was mildly sensitive but highly specific: while it may miss some surfactant secretion cases, it is unlikely to overestimate the ability of a specific sugar to induce secretion.

In the model, we encoded each carbon source as a set of p hot-encoded binary variables Xi . We included another covariate important to determine the surfactant production outcome: the OD₆₀₀ after at 24 hr (covariate Oi), which quantifies whether that strain could grow on that nutrient. The model therefore had p+1 parameters, each represented by the Greek letter β with an index (i=1,…,p+1) where p is the number of nutrient conditions tested (β1 represents the parameter, or effect size, for the OD₆₀₀, and each β2 to βp+1 represents the parameters or effect sizes for each nutrient). We used the LASSO (Tibshirani, 1996) to impose an L1-penalization for the parameter λ (see equation in Figure 7). Then we used threefold cross-validation, which splits the dataset into training and testing in repeated iterations and uses the test loss (the error rate of the model) to tune the penalty parameter. As common practice with LASSO, we choose the penalty value that was one standard error above the one that produced the minimum loss in the test: λ=9 × 10^–3 (Figure 7—figure supplement 1B). This penalty shrank all but 11 parameters to 0 (Figure 7—figure supplement 1C). The covariates with non-zero effect sizes included OD₆₀₀ and 10 carbon sources (Figure 7D). The top three compounds were—as expected—the positive controls tween80, tween40, and tween20. This provided a sanity check that the model worked as intended. D-glucose came in fourth place: the only carbon source that the model deemed better than glycerol at favoring P. aeruginosa surfactant production.

To validate the glucose result, we grew all the 31 strains in minimal medium using D-glucose as the sole carbon source, now using media that we made ourselves from scratch as opposed to the high throughput—but less precise—BIOLOG plates and we did each test in triplicate for added rigor. Consistent with our model, glucose improved surfactant secretion overall, showing better secretion seven strains relatively to the secretion we had seen in glycerol (Figure 7C); this included three strains that were non-producers in glycerol but turned into producers in glucose. Interestingly, the switch also went in the opposite direction: four strains had worse surfactant scores in D-glucose than in glycerol. This suggests a complex strain-to-strain variability in the metabolic features of P. aeruginosa. Nonetheless, the data showed that—at least for some strains—glucose, a carbon source which should impose less burden on primary metabolism (Guerra-Santos et al., 1984), could free up resources for secondary metabolism.

To test this idea further, we tested another carbon source, succinate, which P. aeruginosa can grow on (Figure 7A) but which—according to our model—should not improve secretion relatively to glycerol. Succinate enters the carbon metabolism directly through the TCA cycle, and compared to glycerol and glucose, the initial TCA cycle flux would be faster and generate more ROS level that would burden primary metabolism and impair surfactant synthesis. Indeed, none of the strains secreted surfactants in succinate (Figure 7C).

To investigate how oxidative stress might impact expression rhlAB in glucose we added H₂O₂ and monitored P_rhlABgfp fluorescence in the PA14 wild type and ΔrhlA strains. As described previously (Figure 6), we used the same constitutive P_A1/04/03gfp as the baseline for comparison. Phase I growth rate was unaffected, indicating that both strains could still remove oxidative stress effectively (Figure 7D). As in glycerol, the rhlAB expression levels showed the same higher baseline level in the ΔrhlA compared to the wild type. But, contrasting with the results from glycerol, D-glucose did not increase rhlAB expression with increasing H₂O₂ in either strain (Figure 7E). Taken together, these results support that carbon sources that place a lower burden on primary metabolism leave resources available to both deal with extrinsic oxidative stress and sustain surfactant production.

Glycerol synthetic media were prepared with 800 ml of Milipore water, 200 ml of 5× minimal salts buffer, 1 ml of 1 M magnesium sulphate, 0.1 ml of calcium sulphate with 0.5 gN/l nitrogen (ammonium sulphate), iron at 5 μM (iron III sulphate), and 3.0 gC/l glycerol, respectively. 5× stock minimal salts buffer was prepared with 12 g of Na₂HPO4 (anhydrous), 15 g of KH₂PO4 (anhydrous), and 2.5 g of NaCl into 1 l water.

The clinical isolates were inoculated in 3 mL of lysogen broth (LB) and incubated at 37°C overnight with shaking at 250 rpm. 500 μL of cell culture was pelleted and washed three times with PBS. 0.0025 units were inoculated into 3.0 g of carbon/liter of medium in glycerol minimal medium (7.7 g/l of glycerol) in BD Falcon (BD Biosciences, San Jose, CA) 96-well flat-bottom plates, with 150 μL of suspension per well. The plate was incubated during 48 hr at 37°C in a Tecan Infinite M1000 or Tecan Infinite M1000 Pro plate reader (Männedorf, Switzerland), with an orbital shaking of 4 mm of amplitude. OD₆₀₀ was measured in 10 min intervals.

The production of rhamnolipids was assessed by drop collapse assay (Jain et al., 1991; Chen et al., 2007). We placed 50 μL of the culture’s supernatant on a polystyrene surface (the lid of a 96-well plate). The presence of rhamnolipids decreases the surface tension of the liquid, making the drop collapse. We considered a strain rhamnolipid non-producer if the drop does not spread, a mild-producer if spreads, but not enough to pass over the lid’s circle that corresponds to the wells, and producer if spreads over it.

The pangenomes of our clinical P. aeruginosa isolates were identified using the roary tool (Page et al., 2015) using the output of genome annotation from prokka (Seemann, 2014). The core genes need to have at least 95% identity using blastp and to be present in at least 99% of all genomes, and then were aligned with the MAFFT method (Katoh et al., 2002).

The phylogenetic tree of clinical P. aeruginosa strains was constructed from core genomes as previously described (Yan et al., 2017). Moran’s I test was carried out using the ape package in R (Paradis and Schliep, 2019). The ancestral state of swarming and rhamnolipid production was reconstructed using corHMM package (Beaulieu et al., 2013).

We first normalized the growth curves conducted in different experimental batches (different microtiter plates) using the PA14 inoculated in the same plate as the reference for normalization. The plate-specific coefficients were obtained by solving a linear regression model ‘log(OD of PA14)~Plate + Time’ (the Wilkinson notation) using Matlab function fitlm, where Plate is plate ids and Time is time points (both are categorical variables). Then the growth curves in each plate were normalized by multiplying the adjustment coefficient of the plate. These normalized growth curves were then smoothed by Savitzky-Golay filter (sgolayfilt function in Matlab). The growth phases (phase I, II, III) were determined by the following rules: (phase I) first-order derivative >0 and second-order derivative >0; (phase II) first-order derivative >0 and second-order derivative <0; (phase III) first-order derivative <0. The mean and maximum specific growth rate for each growth phase were estimated from the best-fit model among three models of bacterial growth curves: (Morens et al., 2004) exponential model yt=y0+y1eμt ; (Rowland et al., 2018) Zwietering-Logistic model yt=y0+A-y01+e4μAλ-t+2-1 ; (Henderson, 1999) Zwietering-Gompertz model yt=y0+(A-y0)e-eμeAλ-t+1 , where y0, y1, μ, λ, A are the parameters to fit. The goodness of fit for each model was evaluated by the value of the coefficient of determination, R2=1−∑t=1n(o(t)−y(t))2∑t=1n(o(t)−o¯)2 , where ot is the growth curve data observed at time point t and o¯=1n∑t=1no(t). Both non-negative factorization and random forest regression were performed in python using sklearn.decomposition.NMF and sklearn.ensemble.RandomForestClassifier, respectively.

All P. aeruginosa strains were grown until the end of exponential phase of growth in glycerol minimal medium. Bacteria was then loaded into 0.25-μm nylon membranes (Millipore) using vacuum, transferred to pre-warmed hard agar plates with the same medium composition and incubated at 37°C during 2.5 hr. The filters were then passed to 35-mm polystyrene dishes (Falcon) with 1 mL of 2:2:1 methanol:acetonitrile:H₂O quenching buffer and incubated there during 15 min on dry ice. Cells were removed by scraping, and the lysate containing quenching buffer was transferred to 1.5 mL tubes and centrifuged at 16,000 rpm for 10 min at 4°C. Supernatant transferred to fresh tubes and stored at –80°C.

The extracts were profiled using LC-MS, identifying a total of 92 compounds (Supplementary file 5). Some compounds contained missing values. These missing values in metabolite abundance can be (Morens et al., 2004) truly missing; (Rowland et al., 2018) present in a sample but its level is below detection limit; (Henderson, 1999) present in a sample at a level above the detection limit but missing due to failure of algorithms in data processing. Here we assume that a metabolite with missing values in all three replicates is truly missing in the sample and removed from our analysis (Supplementary file 5). However, if the missing values were only found in one or two replicates, the missing values were imputed by the average of the non-missing values. After that imputation, all compounds remaining with missing values were removed (Supplementary file 5).

The peak areas were normalized using cross-contribution compensating multiple standard normalization (Redestig et al., 2009) with NormalizeMets R package (De Livera et al., 2018). This method relies on the use of multiple internal standards. Since LC-MS lacks such internal standards, we used instead a set of metabolites assumed to be constant across all the strains. They were selected with a Kuskal-Wallis test, adjusting the p-value with Benjamini-Hochberg method. The ones with a p-value above 0.05 were considered constant (pyruvate, methylglyoxal, (S)–2-acetolactate, tyramine, D-glucose, (S)-lactate, N-acetyl-L-glutamate 5-semialdehyde, 4-aminobutyraldehyde, and glycine), therefore after the normalization step they were removed (indicated in red, Supplementary file 5A). The processed area peaks for all metabolites are included in Supplementary file 5.

The hierarchical clustering of the normalized metabolomic data was performed using gplots R package (Warnes et al., 2015), with Euclidean distance and Ward’s aggregation method. The clustering was performed with all metabolites, including 16 unidentified metabolites that we could not identify (indicated in black, Figure 2—figure supplement 2). We included these 16 unidentified compounds in the clustering analysis because they were detected in all strains and therefore were not artifacts of the LC-MS. Nonetheless, because they were unidentified, we removed them from Figure 3—figure supplement 2 and from the downstream analyses. Two other compounds, fumarate and guanosine, which were only putative initially, were bioinformatically confirmed as all other compounds with the same molecular weight are produced by enzymatic reactions missing in our clinical isolates. The PCA was done with R’s base method, using also the normalized data and previously removing the 16 metabolites with unknown identity. The plot was obtained using ggplot2 R package (Wickham, 2016).

The differential metabolites between rhamnolipid producers and non-producers were determined by a Mann-Whitney test, with p-values adjusted with Benjamini-Hochberg method and a significance level of 0.05. These compounds were fed to FELLA algorithm (Picart-Armada et al., 2018; Picart-Armada et al., 2017). FELLA retrieves a graph describing the relationships among pathway, module, enzyme, reaction, and compounds of P. aeruginosa strain PA14 from the KEGG database. The graph was then used as an input network of differential compounds for its diffusion algorithms (Vandin et al., 2011). The output of FELLA consists of all subnetworks of the entries predicted to have a high connectivity with the differential compounds. We filtered the subnetwork entries to keep only metabolic pathways. The entries shown in Supplementary file 3 are the ones with a significant probability of receiving part of the simulated flux.

OPLS-DA model of metabolomics data was built using ropls R package (Thévenot et al., 2015), fixing the number of orthogonal components to 3. R² and Q², key parameters for assessing the validity of the model, were assessed with sevenfold cross-validation. The significance of the model was determined by permutation test (n=2000). The p-value corresponds to the proportion of Q²_perm above Q². With a p-value below 0.05 we considered the model significant. The loadings of the predictive component of the model were extracted to determine how each metabolite contributes to the separation according to the phenotype.

Custom Python codes were developed with the COBRApy package (Ebrahim et al., 2013) to carry out all metabolic flux modeling and simulations in the paper. Since iJN1411 model was developed for P. putida, we removed genes and associated reactions that are missing in all our strains but present in the iJN1411 model. The futile cycles involving NADH, NADPH, and GSH were also removed. The modified iJN1411 model was further expanded by adding rhamnolipid biosynthesis pathway involving 9 new metabolites and 12 new reactions.

The boundary fluxes of the model were set to mimic the composition of the glycerol minimum medium. For C:N=10, the lower bounds of glycerol and ammonium fluxes were set to –10 and –3, respectively. For C:N=3, the lower bounds were set to –3 and –3, respectively. The flux unit is mmol/gDW/hr throughout the paper. To constrain the total producing flux of NADH (the same for NADPH and GSH) at a certain value C, we first defined a binary variable Xk for each NADH-involving reaction k to indicate whether NADH is produced by this reaction. Given the stoichiometric coefficient of NADH in this reaction (sk) and its flux value (fk), the mathematical constraints for Xk was set by Xk=1 for skfk>0 and Xk=0 otherwise. Therefore, the constraint that equalizes the total NADH producing flux and a constant C is simply ∑kskXkfk=C . However, both Xk and fk are variables and such quadratic constraint has not yet been supported by COBRApy. We overcame this difficulty by defining Ak=Xkfk and linearized the product with the following two inequalities: Xklb≤Ak≤Xkub and fk-(1-Xk)ub≤Ak≤fk-(1-Xk)lb , where lb and ub are the lower and upper bounds of fk . The two constraints ensure that Ak=0 when Xk=0 and Ak=fk when Xk=1. The minimum/maximum flux values of byproduct secretion were simulated by flux variability analysis at maximum growth rate.

The H₂O₂ level in the extracellular medium was quantified with Amplex Red Hydrogen Peroxide/Peroxidase Assay Kit (Invitrogen, Carlsbad, USA, Catalog no. A22188). 500 μL of overnight growth in LB medium cell suspension were spun down, washed twice in PBS, and resuspended into glycerol medium at OD 1. Each reaction was done in a final volume of 100 μL, with a final concentration of 50 μM of Amplex Red reagent, 0.1 U/mL of HRP (Horseradish Peroxidase), and 0.2 cell OD, in glycerol synthetic medium, in BD Falcon (BD Biosciences, San Jose, CA) 96-well flat-bottom plates. The first column of the plate corresponded to the reaction without cells (the volume was substituted by PBS), and the last column instead of cells contained H₂O₂ (final concentration 10 μM). OD₆₀₀ and fluorescence 530/590 nm were measured in 10 min intervals (48 hr 37°C).

The P. aeruginosa strains were inoculated into LB and grown with agitation for 16 hr. The overnight culture was diluted in glycerol synthetic medium the next day and grew to late exponential phase to harvest RNA. Cells were harvested by centrifuge and then immediately resuspended in RNAprotect (QIAGEN). The RNA was extracted using the RNeasy Mini Kit from QIAGEN. The standard RNA sequencing was done by Genewiz (Genewiz, Inc; South Plainfield, NJ) using pair-end indices of 150 bp read length on an Illumina HiSeq platform. The raw fastq files first went through quality check by fastQC (Andrews, 2010), and aligned with STAR (Dobin et al., 2013). The final output was generated using DEseq2 (Love et al., 2014) package in R. The read counts for each gene are included in Supplementary file 6. The SRA accession number for the RNAseq data is listed in Supplementary file 7.

To test for the functional associations between genes and rhamnolipid production, we adopted the RLQ analysis (ade4 package in R) that was developed for identifying the associations between traits and environmental variables. Here we applied it to RNAseq analysis to understand the main co-structures between gene functions and phenotypes mediated gene expression. The theory of RLQ analysis is described in detail elsewhere (Thioulouse et al., 2018). Briefly, it builds on the simultaneous ordination of three tables: an R table (categorical) describing rhamnolipid production (strong-, mild-, and non-producers) for 11 selected isolates (each replicate was treated as an independent observation), an L table (continuous) describing the RNAseq counts of 1474 core genes shared among the 11 isolates, and a Q table (binary) describing 120 KEGG pathways for the 1474 genes. We first applied corresponding analysis (CA) to table L and Q and PCA to table R. In the CA of R and Q, each strain and gene were weighted by their total RNA expression levels, respectively. The three separate ordinations are then used as inputs to the rlq function. The output of the function includes coefficients for the gene functions and phenotypes (loadings) as well as the scores of strains (out of phenotypes) and scores of genes (out of gene functions). RLQ analysis maximizes the squared cross-covariances between the two sets of scores weighted by gene expressions.

H₂O₂ was removed from environment by cells which degrade H₂O₂ intracellularly. The cumulative H₂O₂ removal curve was calculated by subtracting the values of emission of the wells containing each strain from the values of emission of the wells without cells or H₂O₂. Then they were normalized to the relative scale by dividing their values by the corresponding values (at the same time points) of the control strain (PA14) in the same experiment (plate). These normalized curves were then smoothed by the Savitzky-Golay filter (sgolayfilt function in Matlab). The H₂O₂ removal rate per cell was calculated by dividing the first-order derivative of the cumulative removal curves by OD values. To estimate the effect of rhamnolipid production on per-cell H₂O₂ removal rate, we developed a linear mixed-effect model that extends the simple linear regression model by considering both fixed and random effects. The model was specified by the Wilkinson notation ‘RemovalRate ~Time + RHL + (1|CurveID)’, where RemovalRate is per-cell H₂O₂ removal rate, Time is time point (categorical variable), RHL is categorical rhamnolipid production (non-producer, mild-producer, and strong-producer), and CurveID is the unique ID for each strain and replicate. The expression ‘(1|CurveID)’ indicates that CurveID was modeled as random effects. The model was solved by fitlme in Matlab.

Two P. aeruginosa strains carrying chromosomal insertions of the reporter fusion P_rhlABgfp (Lequette and Greenberg, 2005) and P_A1/04/03gfp (a constitutive promoter, to be used as internal standard) (Lambertsen et al., 2004) in a PA14 background (Xavier et al., 2011), respectively, were inoculated into LB, and overnight cultures were harvested and washed with PBS three times. Cells were diluted to OD₆₀₀=0.1 in synthetic media and monitored using TECAN plate reader for both OD₆₀₀ and GFP fluorscence. Upon cells entering the exponential phase, H₂O₂ was added to the cells at a concentration in the assay of 0, 10, 20, 50, and 100 mM, continuing growth monitoring over 48 hr. After identifying growth phase II, the promoter activity at each timepoint within this interval was calculated as: P=dGFPdt∙1OD . The ratio of promoter activity in each experiment was extracted as R_p = median(PrhlAB)median(PA1/04/03) . Similarly, the growth rate was calculated as G=dODdt⋅1OD , and the ratio of strains with different promoters was calculated as RG=median(GPrhlAB)median(GPA1/04/03) . The effect of H₂O₂ was estimated by fitglme in Matlab as: log(R)~a₀+β⋅ [H₂O₂]+ϵ, where R is the measured ratio of either promoter activity or growth rate, and ϵ is denoted as the random effect of experiments. This procedure was used too for performing the promoter activity and growth rate analyses in PA14 ∆rhlA mutant.

We grew each of the P. aeruginosa stains in the BIOLOG phenotype microarrays PM1 and PM2a from Biolog Inc (Hayward, California, USA). We prepared minimal media without any carbon source added (800 ml of Milipore water, 200 ml of 5× minimal salts buffer, 1 ml of 1 M magnesium sulphate, 0.1 ml of calcium sulphate with 0.5 gN/l nitrogen [ammonium sulphate], iron at 5 μM [iron III sulphate]; 5× stock minimal salts buffer was prepared with 12 g of Na₂HPO₄ [anhydrous], 15 g of KH₂PO₄ [anhydrous], and 2.5 g of NaCl into 1 l water). We grow 5 mL bacterial cultures for 12–16 hr at 250 rpm and 37 ℃, then aliquoted washed 1.5 mL in media without carbon source and adjusted the initial OD by further dilution. We added 150 ul of inoculated media onto each well of a BIOLOG plate and incubated for 24 hr overnight shaking at 250 rpm at 37℃. After that we measured the endpoint OD and assessed surfactant secretion using the drop collapse assay.