Metabolic Modeling of Chronic Wound Microbiota Predicts Mutualistic Interactions that Drive Community Composition

Abstract

Aims

To identify putative mutualistic interactions driving community composition in polymicrobial chronic wound infections using metabolic modeling.

Methods and Results

We developed a 12 species metabolic model that covered 74% of 16S rDNA pyrosequencing reads of dominant genera from 2,963 chronic wound patients. The community model was used to predict species abundances averaged across this large patient population. We found that substantially improved predictions were obtained when the model was constrained with genera prevalence data and predicted abundances were averaged over 5,000 ensemble simulations with community participants randomly determined according to the experimentally determined prevalences. Staphylococcus and Pseudomonas were predicted to exhibit a strong mutualistic relationship that resulted in community growth rate and diversity simultaneously increasing, suggesting that these two common chronic wound pathogens establish dominance by cooperating with less harmful commensal species. In communities lacking one or both dominant pathogens, other mutualistic relationship including Staphylococcus/Acinetobacter, Pseudomonas/Serratia and Streptococcus/ Enterococcus were predicted consistent with published experimental data.

Conclusions

Mutualistic interactions were predicted to be driven by crossfeeding of organic acids, alcohols and amino acids that could potentially be disrupted to slow chronic wound disease progression.

Introduction

Chronic wounds are defined as host-pathogen conditions which have not healed in a timely manner (Lazarus et al., 1994). Approximately 2% of the US population (6 million people) have non-healing chronic wounds incurring annual treatment costs of $25 billion (Sen et al., 2009; Frykberg and Banks, 2015). The occurrence of chronic wounds has been surging due to the increasing prevalence of associated health problems such as diabetes and obesity (Valensi et al., 2005; Guariguata et al., 2014). The four most common types of chronic wounds are diabetic foot ulcers, venous leg ulcers, pressure ulcers and non-healing surgical wounds (Kirsner, 2016). Venous leg ulcers occur in lower sections of the legs, pressure ulcers are injuries localized to skin located over bones and diabetic foot ulcers occur at the bottom of feet in individuals having diabetes mellitus. These ulcers are caused by restricted blood flow and increased pressure due to impaired motion.

Chronic wounds are usually colonized by microbial communities rather than single bacterial species (Gjødsbøl et al., 2006; Dowd et al., 2008a, b; James et al., 2008; Wolcott et al., 2009). Polymicrobial infections often require about 12+ months to clear, have recurrence frequencies of 60 to 70% (Richmond et al., 2013; Frykberg and Banks, 2015) and have elevated mortality rates as compared to single-species infections (Pulimood et al., 2002). Furthermore, colonizing bacteria form multispecies biofilms on the wound surface (Hall-Stoodley et al., 2004; James et al., 2008), providing an additional level of protection against antibiotics due to diffusional limitations (Burmølle et al., 2010; Bjarnsholt, 2013) and the presence of metabolically inactive cells (Peterson, 2005; Clinton and Carter, 2015). In vivo rabbit models have demonstrated that polymicrobial infections slow wound healing compared to their respective monoculture infections (Seth et al., 2012; Pastar et al., 2013).

Culture- and molecular-based methods have been used to analyze chronic wound communities (Bowler et al., 2001; Melendez et al., 2010; Rhoads et al., 2012). The most common genera represented in chronic wound infections are Staphylococcus, Corynebacterium, Pseudomonas, Streptococcus, Enterococcus, Enterobacter, Finegoldia and Serratia (Gjødsbøl et al., 2006; Dowd et al., 2008a, b; Frank et al., 2009). Staphylococcus aureus and Pseudomonas aeruginosa are the two most common bacterial pathogens observed in chronic wound infections. These two pathogens have been shown to establish mutualistic interactions including metabolite crossfeeding that allows them to resist antibiotic treatment in multiple types of infection environments including chronic wounds and the cystic fibrosis lung (DeLeon et al., 2014; Hotterbeekx et al., 2017; Marques et al., 2018; Alves et al., 2018). Mutualistic relationships between pathogens reduce competition for available nutrients and result in robust communities associated with prolonged infections and poor clinical outcomes (Loesche et al., 2017). Chronic wound pathogens also form mutualistic relationships with skin commensal species that impact their virulence (Ramsey and Whiteley, 2009; Ramsey et al., 2011). These interactions allow pathogens to survive at infection sites, enhance antibiotic resistance and increase disease severity (Dalton et al., 2011; Peters et al., 2012; Korgaonkar et al., 2013). More detailed knowledge about the mechanisms underlying these interspecies relationships offers the potential for developing novel treatment strategies based on disrupting specific mutualistic interactions rather than just targeting specific pathogens.

In silico metabolic models are powerful tools for the analysis of how host environment impacts species interactions and community stability, composition and robustness (Shoaie et al., 2013; Hanemaaijer et al., 2015; Perez-Garcia et al., 2016, Bosi et al., 2017, Bauer and Thiele, 2018). We have developed a suite of in silico computational methods for efficiently simulating complex and realistic host-associated bacterial communities, connecting host environment and community metabolism, exploring growth-diversity tradeoffs, quantifying metabolite crossfeeding relationships, relating metabolism and disease states, and rationalizing patient-to-patient variability (Henson and Phalak, 2018; Henson et al., 2019). Our modeling framework exploits the availability of 16S rDNA sequencing data to identify the dominant genera present in the host-associated infection, the AGORA database (Magnúsdóttir et al., 2017) to select a genome-scale metabolic reconstruction for a representative species from each genera, and the SteadyCom method (Chan et al., 2017) for efficiently simulating large community models to predict the growth rate, species abundances and metabolite crossfeeding rates between species. We have applied the in silico methods to a 20 species model of commensal species in the human gut (Henson and Phalak, 2018) and to a 17 species model including dominant pathogens of the adult cystic fibrosis lung (Henson et al., 2019).

In this study, we have used our in silico computational methods and 16S rDNA pyrosequencing data collected from 2,963 chronic wound patients (Wolcott et al., 2016) to develop a bacterial community model for investigation of pathogen-pathogen and pathogen-commensal interactions. The dataset contained abundances (i.e. the relative amount of the genera averaged across samples) of the 20 most abundant genera for each type of chronic wound, diabetic foot ulcers, venous leg ulcers, pressure ulcers and non-healing surgical wounds, as well as prevalences (i.e. the fraction of samples containing the genus). Because the original study (Wolcott et al., 2016) concluded that the average bacterial community present at each wound location were not significantly different, the average abundance data for each wound type was assimilated into a combined dataset and used to construct a single 12 species community model representing the most abundant pathogenic and commensal genera. Simulations were performed and analyzed to identify putative mutualistic interactions that could drive community composition and negatively impact the effectiveness of antibiotic treatments.

Materials and Methods

Community metabolic model

16S rDNA pyrosequencing data was obtained from a published study which analyzed chronic wound samples from 2,963 patients treated for decubitus ulcers (767 samples), diabetic foot ulcers (910 samples), venous leg ulcers (916 samples) and non-healing surgical wounds (370 samples) (Wolcott et al., 2016). The publication provided relative abundances of the top 20 bacterial genera for each wound type. Community composition was shown to be independent of the wound type and patient demographics such as age, gender and race. Therefore, we assimilated the average abundance data for each wound type into a single dataset and determined the most abundant genera across all samples. To limit model complexity and focus on the most dominant genera, the community model accounted for the 12 genera with highest average abundances (Figure 1A). These 12 genera accounted for approximately 74% of the 16S read data averaged across all 2,963 samples. To allow direct comparison with community model predictions, the 16S data was normalized such that the abundances of these 12 genera summed to unity. A representative species for each genus was selected from the AGORA database (www.vmh.life) (Magnúsdóttir et al., 2017) according to species occurrence and abundance data contained in the original experimental reference (Wolcott et al., 2016; Table 1). Due to the absence of well-curated metabolic reconstructions, the AGORA models based on literature-derived experimental data and comparative genomics offered the most faithful representation of species-dependent metabolism. The community metabolic model was formulated within SteadyCom (Chan et al., 2017) by combining the species models into a large-scale community model. The 12-species community model accounted for 16,133 reactions, 13,666 metabolites and 9,713 genes.

Figure 1:

Overview of the community modeling framework. (A) Flow chart showing steps in model development, simulation and analysis. (B) Average species abundances obtained from the model ensemble. (C) r and p values obtained from correlation analysis of the model ensemble abundance data. (D) Significant crossfeeding relationships between Staphylococcus (purple bars) and Pseudomonas (green bars) predicted by model ensemble simulations.

Table 1:

The 12 species included in the chronic wound community model along with the prevalences and normalized average abundances of the associated genera from (Wolcott et al., 2016).

Number	Strain	References	Prevalence (%)	Relative abundance
1	Staphylococcus aureus subsp aureus USA300 FPR3757	(Melendez et al., 2010; Rhoads et al., 2012; Wolcott et al., 2016)	63	0.42
2	Pseudomonas aeruginosa NCGM2 S1	(Melendez et al., 2010; Rhoads et al., 2012; Wolcott et al., 2016)	25	0.13
3	Corynebacterium striatum ATCC 6940	(Dowd et al., 2008a; Rhoads et al., 2012; Wolcott et al., 2016)	36	0.11
4	Streptococcus agalactiae A909	(Rhoads et al., 2012; Wolcott et al., 2016)	23	0.07
5	Enterococcus faecalis V583	(Melendez et al., 2010; Tzaneva et al., 2016)	17	0.05
6	Finegoldia magna ATCC 29328	(Dowd et al., 2008a; Wolcott et al., 2016)	25	0.05
7	Anaerococcus vaginalis ATCC 51170	(Rhoads et al., 2012; Jneid et al., 2017)	24	0.05
8	Stenotrophomonas maltophilia D457	(Rhoads et al., 2012; Wolcott et al., 2016)	19	0.04
9	Prevotella bivia DSM 20514	(Wolcott et al., 2016; Jneid et al., 2017)	12	0.03
10	Acinetobacter baumannii AB0057	(Rhoads et al., 2012; Jneid et al., 2017)	9	0.02
11	Serratia liquefaciens ATCC 27592	(Alper et al., 1983; Bowler et al., 2001)	5	0.02
12	Bacteroides fragilis 3 1 12	(Stephens et al., 2003; Jneid et al., 2017)	8	0.02

Model tuning and simulation

The nutrient environment in chronic wound is complex and expected to vary between patients and according to disease progression. A metabolomics study conducted for four chronic pressure ulcer samples detected 122 metabolites with the quantified metabolite concentrations spanning several orders of magnitude (Ammons et al., 2015). Several studies have identified upregulated or downregulated metabolites in chronic wounds compared to healing wounds, but absolute metabolite concentrations were not reported (Sood et al., 2015; Ammons et al., 2015; Junka et al., 2017). The in vitro Lubbock chronic wound biofilm model based on chopped meat-based media (Bolton broth, MRS broth and BHI broth) has been shown to contain nutrients present in chronic wound beds (Sun et al., 2008). While these studies provided important guidance on nutrient selection for the community model, they were not sufficient to completely define a nutrient environment in which all 12 species were capable of growth and the predicted species abundances were in approximate agreement with the 16S data (Wolcott et al., 2016) used in this study.

Therefore, our approach previously used to specify nutrients for a community model of the adult cystic fibrosis microbiota (Henson et al., 2019) was followed. Due to lack of data on species-specific uptake rates, we assumed that all species had the same maximum uptake rate of a given nutrient and used nutrients based on their metabolic requirements. By contrast, the uptake rates were varied across the available nutrients as described below. First, all 21 amino acids and 6 carbon sources (glucose, L-lactate, ribose, galactose, L-arabinose, fructose) known to be available in chronic wound beds (Ammons et al., 2015) were added (Table S1). Then 15 common metals and ions and 30 metabolites that were required for each species to grow in simulated monoculture were included. The metabolites guanosine, inosine, uracil and uridine (Sood et al., 2015; Ammons et al., 2015) and the terminal electron acceptors O₂ and NO₃ were added because they are known to be present in the chronic wound environment. Finally, three putative metabolites were added to increase the growth rates of particular species such that predicted species abundances were in approximate agreement with the 16S data: starch 1 for Corynebacterium; kestose for Enterococcus and; glycerol-3-phosphate for Prevotella.

The 81 metabolites contained in the simulated chronic wound environment were partitioned into 19 groups for the purpose of model tuning (Table S1): (1) 15 metals and ions; (2) 30 essential growth metabolites; (3)-(6) each of the 4 chronic wound metabolites guanosine, inosine, uracil and uridine: (7) 17 amino acids; (8) 4 amino acids isoleucine, leucine, lysine and valine reported to be elevated in chronic wounds compared to the other 17 amino acids (Junka et al., 2017); (9)-(14) each of the 6 carbon sources; (15) O₂; (16) NO₃; (17) starch 1; (18) kestose; and (19) glycerol-3-phosphate. The community uptake rates of metabolites in these 19 groups were tuned by trial-and-error to achieve species abundances in approximate agreement with the 16S data (see Figure 2C).

Figure 2:

Model predictions for monoculture simulations and a 12 species community simulation without prevalence constraints. (A) Single-species growth rates (h⁻¹) where the species are listed by their genera. (B) Single species secretion rates predicted from monoculture simulations for the byproducts acetate (Ac), CO₂, ethanol (Eth), formate (For), H₂S, D-lactate (D-Lac), L-lactate (L-lac), NH₄ and succinate (Succ). (C) Comparison of species abundances predicted from the 12 species community model without prevalence constraints (blue bars) and obtained from normalized 16S patient data (red bars). (D) Relationship between single-species growth rates and species abundances predicted from the community simulation (r = 0.72, p = 0.009).

The SteadyCom method (Chan et al., 2017) was used to formulate and solve the chronic wound community model as described in our previous studies on the human gut microbiota (Henson and Phalak, 2018) and cystic fibrosis (Henson et al., 2019). The SteadyCom method uses a form of community flux balance analysis to calculate the relative abundance of each species with an objective of maximizing the community growth rate. The non-growth associated ATP maintenance for each species was chosen to be 5 mmol/gDw/h, which is in reported ranges for curated bacterial reconstructions (Baumler et al., 2011). Individual species simulations were performed to ensure that each species was able to grow in monoculture on the in silico media. In addition to providing the community growth rate and species abundances, SteadyCom calculated species-dependent uptake and secretion rates of all supplied and secreted metabolites.

The community model was further constrained with genus prevalence data (i.e. the fraction of samples containing the genus) available in the original experimental study (Wolcott et al., 2016). To implement these constraints, the participating species of the community were randomly chosen according to the prevalences using uniform random numbers and then the model was solved. A large number of models were solved to adequately sample the species participation space. A total of 5,250 model ensemble simulations were performed with 250 cases discarded because the community growth rate was zero or the SteadyCom tolerance on the sum of the species abundances was not satisfied. The remaining 5,000 cases were treated as simulated patient samples and their abundances were averaged (Figure 3B). The community uptake rates of the 81 supplied metabolites were further tuned to achieve quantitative agreement with the 16S data (see Figure 3B).

Figure 3:

Prevalence constrained model ensemble predictions. (A) Comparison of genera prevalence data (red bars) (Wolcott et al., 2016) and in silico prevalences of the corresponding modeled species (blue bars). (B) Comparison of species abundances predicted from the model ensemble with species prevalence constraints (blue bars) and normalized 16S patient data (red bars).

Analysis of simulation results

The difference (e) between the normalized 16S abundances (p_id) and the model predicted abundances (p_im) was calculated as the angle between the two abundance vectors (Li et al., 2004),

$$e=sin\left[co{s}^{-1}\left(\frac{p_{im}^T{p}_{id}}{‖p_{im}‖‖p_{id}‖}\right)\right]$$

where k · k denotes the Euclidean norm and e ∈ [0,1]. The two abundance vectors were identical (i.e. parallel) if e = 0 and orthogonal if e = 1. The inverse Simpson equitability index (D_com) was used as a measure of community diversity (Henson and Phalak, 2018),

$$D_{com}=\frac{1}{N}\frac{1}{{\sum }_{i=1}^N{p}_i^2}$$

where N = 12 is the total number of species and p_i is the 16S determined or model predicted abundance of the species i. Significant crossfeeding relationships were identified based on the magnitudes of the secretion and uptake fluxes as detailed in our previous study (Henson and Phalak, 2018). We typically reported the top six crossfeeding relationships between the participating species (Figure 1D).

Staphylococcus and Pseudomonas are the most common genera in chronic wound infections (Gjødsbøl et al., 2006; Fazli et al., 2009; Pastar et al., 2013; Bessa et al., 2015) and are known to exhibit strong interactions (Nguyen and Oglesby-Sherrouse, 2016; Hotterbeekx et al., 2017). Correspondingly, we were interested in community behavior with the presence and absence of these two dominant pathogens. Therefore, the ensemble of 5,000 community simulations was partitioned into our groups: both Staphylococcus and Pseudomonas present (SaPa); Pseudomonas not present (Sa∆Pa); Staphylococcus not present (∆SaPa); and neither Staphylococcus nor Pseudomonas present (∆Sa∆Pa). For the entire ensemble and each of the four partitioned subsets, mutualistic interactions between species were identified by performing correlation analysis on the predicted abundance data (Figure 1C). An interaction was deemed significant if the r-value was greater than 0 and p-value was less than 0.05.

Results

Community composition is shaped by single-species metabolism

Monoculture simulations were performed with the in silico nutrients (TableS1) to access the metabolic capabilities of the 12 species. Staphylococcus was predicted to have the highest single-species growth rate (Figure 2A), consistent with its role as a dominant chronic wound pathogen. Serratia, Stenotrophomonas and Pseudomonas had growth rates greater than 0.3 h⁻¹, suggesting that these species would be competitive in community simulations. Of 9 metabolites commonly found in the chronic wound environment (Sood et al., 2015), acetate, CO₂ and formate were the primary secreted byproducts predicted from monoculture simulations (Figure 2B). Interestingly neither L-lactate nor D-lactate secretion was predicted even though species from genera such as Streptococcus are well known to secrete L-lactate as a primary byproduct (Keevil et al., 1984) via homolactic fermentation. This model behavior was attributable to alternative optima with respect to byproduct secretion patterns (Henson et al., 2019). We chose not to perform manual curation of the metabolic model since in silico lactate synthesis was induced by the presence of other community members.

When the community model was simulated without prevalence constraints, 7 of the 12 species were predicted to coexist (Figure 2C). Consistent with the normalized 16S rDNA gene amplicon data, Staphylococcus was the dominant species and both Pseudomonas and Corynebacterium) were present at abundances greater than 10%. However, the model overpredicted the abundances of Acinetobacter, Serratia and Bacteroides and incorrectly predicted zero abundances for 5 other species. The difference e between the predicted and 16S abundance vectors was 0.38, denoting a moderate prediction error. This difference was not unexpected, as the 16S data was provided as the average over a large number of patient samples while the simulation represented a single predicted sample (Henson et al., 2019). As a result, the normalized 16S data exhibited a greater species diversity (D_com = 0.38) than the simulated sample (D_com = 0.36). The predicted species abundances were strongly correlated to the single-species growth rates (r = 0.72, p = 0.009; Figure 2D), as expected for a community modeling method such as SteadyCom based on growth rate maximization. A sublinear correlation was predicted because some species were more/less efficient at metabolite crossfeeding. For example, Pseudomonas had the fourth highest monoculture growth rate but the second highest abundance.

Incorporation of genera prevalence data improves prediction of community composition

A single model simulation with all 12 species allowed to participate in the community only provided qualitative agreement with normalized 16S data with respect to the abundances of the dominant genera (Figure 2). We hypothesized that incorporation of genera prevalence data as additional constraints would improve model predictions with respect to the coexisting species and their abundances. Given that the 16S abundances were obtained by averaging over 2,963 patient samples, we used prevalence data reported in the original study (Wolcott et al., 2016) to generate an ensemble of 5,000 in silico communities by randomly generating the species allowed to participate (see Materials and Methods). Then the predicted abundances were averaged over the ensemble for comparison to 16S values. This approach was consistent with the constraints-based philosophy of metabolic modeling based on the refinement of model predictions through the imposition of additional data-based constraints (Imam et al., 2015; Conde et al., 2016). Even if allowed to participate in a community, a particular species could be predicted to have a zero abundance. Therefore, participation was a necessary but not sufficient condition for a species to coexist in a simulated community. Below we used the terms “prevalence” and “participation” interchangeably with respect to the model ensemble simulations for simplicity.

A comparison of the genera prevalence data and the in silico prevalences of the corresponding modeled species showed a slight bias even through the model ensemble was generated with the intent of these prevalences being identical (Figure 3A). This disparity was caused by the need to discard 250 of 5,250 total simulation cases because the community growth rate was zero or the SteadyCom tolerance on the sum of the species abundances was not satisfied. The removal of these 250 cases introduced a small systematic bias into the in silico prevalences as most species were advantaged by being allowed to participate more frequently than indicated by data. Despite this small bias, the model ensemble generated substantially improved predictions of the 16S-derived abundances (e = 0.12; Figure 3B) compared to the prevalence unconstrained model (e = 0.38; Figure 2). As expected, the model prevalences and abundances were strongly correlated (r = 0.91, p = 5×10⁻⁵). While the abundances of some minor species were substantially underpredicted (e.g. Enterococcus) or overpredicted (e.g. Bacteroides), we deemed these predictions to be sufficiently accurate to utilize the model ensemble for further analysis of the chronic wound community.

Analysis of community structure and composition

We defined the richness of the community as the number of species with abundances exceeding 1%. To further investigate the effect of imposing prevalence constraints in the ensemble of 5,000 models, we calculated the number of species allowed to participate in each community (Figure 4A) and the actual richness of each community (Figure 4B). Over 90% of the simulations allowed no more than 4 species, with the most likely cases being 2 or 3 species. Because some species allowed to participate were predicted to have zero abundances, the predicted richness was generally less than the number of participating species. Over 90% of the simulated communities had richnesses of no more than 3 species. Therefore, the model ensemble predicted that most individual patient samples would have low diversity. The original study (Wolcott et al., 2016) did not provide data on individual samples that would allow comparison with these modeling results.

Figure 4:

Analysis of chronic wound community structure and growth rates. (A) The number of species allowed to participate in each community simulation. (B) The richness (number of species with calculated abundances exceeding 1%) of each simulated community. (C) The percentages and numbers of the 5,000 simulated models in which both Staphylococcus and Pseudomonas were allowed to participate (SaPa), Pseudomonas was not allowed to participate (Sa∆Pa); Staphylococcus was not allowed to participate (∆SaPa); and neither Staphylococcus or Pseudomonas were allowed to participate (∆Sa∆Pa). (D) A box and whisker plot showing the community growth rates for each of the four partitioned cases, where the red line corresponds to the median, the black dotted lines (whiskers) indicate the variability outside the lower and upper quartiles, and the red circles represent outliers.

Because Staphylococcus and Pseudomonas are the two dominant pathogens in chronic wound infections (Gjødsbøl et al., 2006; Fazli et al., 2009; Pastar et al., 2013; Bessa et al., 2015), we partitioned the ensemble of 5,000 community models into four groups based on the allowed participation of these two species (see Materials and Methods): both Staphylococcus and Pseudomonas present (SaPa); Pseudomonas not present (Sa∆Pa); Staphylococcus not present (∆SaPa); and neither Staphylococcus or Pseudomonas present (∆Sa∆Pa). Each group was populated by a sufficient number of models to allow statistical analysis of the impact of each dominant pathogen on community structure and species interactions (Figure 4C). The SaPa group was predicted to have the highest average growth rate with little variability except for some outliers (Figure 4D). The growth rate decreased, and variability increased as Pseudomonas (Sa∆Pa), Staphylococcus (∆SaPa) or both species (∆Sa∆Pa) were removed from the communities. These predictions suggest mutualistic interactions between the two pathogens and possibly with some commensal species that enhance community fitness.

Interestingly, all outlier communities were characterized by high growth rates except for the ∆Sa∆Pa group. We analyzed the compositions of these outliers for each group by comparing predicted abundances of the outlier communities to those of communities with outliers removed (Figure S1). This analysis revealed several putative mutualistic relationships including: Serratia and/or Bacteroides with Staphylococcus and/or Pseudomonas; and Streptococcus with Finegoldia. The outlier-free cases for the ∆Sa∆Pa group were predicted to have much higher diversity than the other groups, suggesting that the presence of Staphylococcus and/or Pseudomonas increased community growth at the expense of diversity. These predictions were consistent with our previously posited hypothesis that infectious disease progression correlates to high growth and low diversity of the evolving community (Henson and Phalak, 2018).

Staphylococcus and Pseudomonas form a mutualistic relationship

We used the ensemble of 820 community models in which both dominant pathogens Staphylococcus and Pseudomonas could participate (SaPa) to identify putative mutualistic interactions between the 12 modeled species. The only significant mutualistic relationship predicted was between the two pathogens themselves (Table 2, Figure S2), suggesting that this interaction drove community growth and composition. We performed additional analysis to test this hypothesis. Compared to the average species abundances calculated from the entire 5,000 model ensemble, the 820 SaPa cases produced a much larger Pseudomonas abundance such that the abundances of the two dominant pathogens averaged almost 90% (Figure 5A). The Pseudomonas abundance was greater than the Staphylococcus abundance in 780 communities (Figure 5B), indicating that Pseudomonas was the primary beneficiary of the mutualistic interaction.

Table 2:

Species abundance correlation analysis for cases in which Staphylococcus and Pseudomonas could participate (SaPa), Pseudomonas was not allowed to participate (Sa∆Pa); Staphylococcus was not allowed to participate (∆SaPa); and neither Staphylococcus nor Pseudomonas could participate (∆Sa∆Pa).

Group	Positively correlated species	r value	p value	Number of cases	Analysis
SaPa	Staphylococcus and Pseudomonas	0.69	<10e-6	820	Figure 5
Sa∆Pa	Staphylococcus and Acinetobacter	0.57	<10e-6	235	Figure 6
Sa∆Pa	Staphylococcus and Corynebacterium	0.22	<10e-6	895	Not shown
Pa∆Sa	Pseudomonas and Serratia	0.88	<10e-6	18	Figure 7
Pa∆Sa	Pseudomonas and Streptococcus	0.72	<10e-6	125	Figure S4
Pa∆Sa	Pseudomonas and Acinetobacter	0.41	<10e-2	42	Not shown
Pa∆Sa	Pseudomonas and Bacteroides	0.4	<1e-2	39	Not shown
∆Sa∆Pa	Corynebacterium and Stenotrophomonas	0.6	<10e-6	107	Figure S5
∆Sa∆Pa	Corynebacterium and Serratia	0.5	<1e-2	24	Not shown
∆Sa∆Pa	Corynebacterium and Bacteroides	0.6	<10e-6	41	Not shown
∆Sa∆Pa	Streptococcus and Enterococcus	0.53	<10e-6	51	Figure 8
∆Sa∆Pa	Streptococcus and Acinetobacter	0.46	<10e-3	30	Not shown
∆Sa∆Pa	Streptococcus and Serratia	0.61	<1e-1	12	Not shown
∆Sa∆Pa	Streptococcus and Bacteroides	0.92	<10e-6	30	Not shown
∆Sa∆Pa	Enterococcus and Finegoldia	0.78	<10e-6	66	Not shown
∆Sa∆Pa	Enterococcus and Stenotrophomonas	0.3	<1e-2	63	Not shown
∆Sa∆Pa	Enterococcus and Acinetobacter	0.91	<10e-6	25	Not shown
∆Sa∆Pa	Enterococcus and Bacteroides	0.98	<10e-6	18	Not shown
∆Sa∆Pa	Finegoldia and Anaerococcus	0.75	<10e-6	85	Not shown
∆Sa∆Pa	Finegoldia and Acinetobacter	0.63	<10e-6	40	Not shown
∆Sa∆Pa	Finegoldia and Serratia	0.86	<10e-6	17	Not shown
∆Sa∆Pa	Finegoldia and Bacteroides	0.51	<1e-2	20	Not shown
∆Sa∆Pa	Anaerococcus and Stenotrophomonas	0.49	<10e-6	73	Not shown
∆Sa∆Pa	Anaerococcus and Acinetobacter	0.8	<10e-6	35	Not shown
∆Sa∆Pa	Anaerococcus and Bacteroides	0.54	<10e-2	28	Not shown
∆Sa∆Pa	Stenotrophomonas and Acinetobacter	0.5	<10e-2	29	Not shown
∆Sa∆Pa	Stenotrophomonas and Bacteroides	1	<10e-6	20	Not shown
∆Sa∆Pa	Prevotella and Bacteroides	0.7	<1e-2	12	Not shown
∆Sa∆Pa	Serratia and Bacteroides	0.83	<1e-2	8	Not shown

Figure 5:

Model ensemble predictions for SaPa simulations showing a mutualistic relationship between Staphylococcus and Pseudomonas. (A) Average species abundances for all 5,000 ensemble simulations (blue bars) and 820 SaPa simulations (red bars). (B) Staphylococcus and Pseudomonas abundances for 820 simulated communities containing both species where the colorbar indicates the number of simulations represented by each circle. The two species show a mutualistic interaction (r = 0.69, p < 10⁻⁶). (C) Community growth rates and equitability for 820 simulated communities containing both species. (D) The six most significant crossfeeding relationships between Staphylococcus (purple bars) and Pseudomonas (green bars).

The SaPa model ensemble predicted a significant positive correlation (r= 0.53, p < 10⁻⁶) between community equitability and growth rate (Figure 5C). These results suggest that the incorporation of less abundant commensal species such as Corynebacterium enhanced community growth. When combined with predictions that the SaPa ensemble produced the highest growth rates with the lowest variability (Figure 4D), these predictions indicate that the mutualistic interaction produced resilient communities not negatively affected by the addition of commensal species. The mutualistic relationship between Pseudomonas and Staphylococcus was supported by bi-directional metabolite crossfeeding, with ethanol, L-lactate and succinate being the primary crossfed metabolites (Figure 5D). Interestingly, L-lactate and D-lactate were not secreted by either species in monoculture (Figure 2B) due to alternative optima with respect to byproducts. These predictions suggest that the exchange of these two byproducts was important for maintaining the interaction. In fact, Staphylococcus is known to consume lactate in vivo to enhance its competitiveness (Ferreira et al., 2013).

Staphylococcus and Acinetobacter form a mutualistic relationship in the absence of Pseudomonas

Next, we used the ensemble of 2,410 community models in which Pseudomonas was absent (Sa∆Pa) to predict mutualistic interactions between Staphylococcus and the 10 remaining species. Compared to the entire 5,000 model ensemble (Figure 6A), the 2,410 Sa∆Pa simulations produced a substantially higher average Staphylococcus abundance and richer communities in which only Enterococcus and Finegoldia failed to coexist (Figure 6A).Two mutualisms involving Staphylococcus were identified (Table 2, Figure S2); we focused on the Staphylococcus and Acinetobacter relationship (Figure 6B) because the correlation was most positive (i.e. mutualistic) and experimental literature characterizing the interaction was available. Compared to the SaPa cases (Figure 5A), the absence of Pseudomonas resulted in increased average Acinetobacter abundance over the 235 cases in which Acinetobacter could participate.

Figure 6:

Model ensemble predictions for Sa∆Pa simulations showing a mutualistic relationship between Staphylococcus and Acinetobacter. (A) Average species abundances for all 5,000 ensemble simulations (blue bars) and 2,410 Sa∆Pa simulations (red bars). (B) Staphylococcus and Acinetobacter abundances for 235 simulated communities in which both species could participate where the colorbar indicates the number of simulations represented by each circle. The two species showed a mutualistic interaction (r = 0.57, p < 10⁻⁶). (C) Community growth rates and equitability for 235 simulated communities with both species. (D) The five most significant crossfeeding relationships between Staphylococcus (purple bars) and Acinetobacter (green bars).

A significant positive correlation (r = 0.53, p < 10⁻⁶) between community equitability and growth rate was predicted for the Sa∆Pa model ensemble (Figure 6C), suggesting that commensals such as Streptococcus enhanced community growth. However, the Sa∆Pa ensemble produced lower growth rates than the SaPa ensemble (Figure 4D) due to the absence of Pseudomonas. Therefore, increased richness of the Sa∆Pa ensemble was accompanied by decreased growth. These predictions are consistent with our hypothesis that low abundance of dominant pathogens such as Pseudomonas corresponds to an earlier disease stage with relatively low growth and high diversity (Henson and Phalak, 2018). Compared to the mutualistic interaction between Staphylococcus and Pseudomonas in the SaPa ensemble (Figure 6D), Staphylococcus and Acinetobacter mutualism was supported by lower crossfeeding rates of amino acids rather than organic acids and alcohols. Acinetobacter was predicted to be the primary beneficiary of crossfeeding, explaining its ability to coexist with Pseudomonas absent. This predicted mutualistic relationship has experimental support, as Staphylococcus and Acinetobacter are major nosocomial pathogens involved in burn infections (Furuno et al., 2008; Barbut et al., 2013) and both genera are known to develop antibiotic resistance (Boucher et al., 2009).

Pseudomonas and Serratia form a mutualistic relationship in the absence of Staphylococcus

The ensemble of 506 community models in which Staphylococcus was absent (∆SaPa) was analyzed to predict mutualistic relationships between Pseudomonas and the 10 remaining species. The ∆SaPa simulations produced a high average Pseudomonas abundance and a large increase in the average abundance of Anaerococcus (Figure 7A); only Finegoldia failed to appear in any community. Of the four significant mutualistic relationships predicted (Table 2, Figure S3), we focused on Pseudomonas and Serratia even though the two species were both allowed to participate in only 18 ∆SaPa communities (Figure 7B). A similar analysis of the Pseudomonas and Streptococcus interaction (125 cases) is provided in the Supplementary Materials (Figure S4). Interestingly, the average Serratia abundance was almost unchanged from the full model ensemble despite Serratia having a positive correlation with Pseudomonas. Furthermore, Anaerococcus did not have a significant correlation with Pseudomonas despite Anaerococcus having a larger average abundance compared to the full model ensemble. These results demonstrate that significant species interactions cannot easily be discerned from abundance data averaged over heterogeneous samples.

Figure 7:

Model ensemble predictions for ∆SaPa simulations showing a mutualistic relationship between Pseudomonas and Serratia. (A) Average species abundances for all 5,000 ensemble simulations (blue bars) and 506 ∆SaPa simulations (red bars). (B) Pseudomonas and Serratia abundances for 18 simulated communities in which both species could participate where the colorbar indicates the number of simulations represented by each circle. The two species showed a mutualistic interaction (r = 0.88, p < 10⁻⁶). (C) Community growth rates and equitability for 18 simulated communities with both species. (D) The six most significant crossfeeding relationships between Pseudomonas (purple bars) and Serratia (green bars).

A significant positive correlation (r = 0.95, p < 10⁻⁶) between community equitability and growth rate was predicted for the ∆SaPa model ensemble despite the small number of samples (Figure 6C). The mutualistic relationship was primarily supported by lactate crossfeeding, with Serratia having a large uptake rate of L-lactate and Pseudomonas consuming D-lactate (Figure 6D). The ability of Serratia to enhance its competitiveness through L-lactate crossfeeding explains why Serratia was predicted to be dominant in the 18 ∆SaPa communities in which it appeared (Figure 6B). Pseudomonas and Serratia are often found to coexist in infections associated with chronic wounds and corneal ulcers (Mayo et al., 1987; Dowd et al., 2008a). Furthermore, the two genera are known to secrete the quorum sensing molecule N-butanoyl l-homoserinelactone (C4 HSL) which might be used in interspecies communication (Tashiro et al., 2013).

Streptococcus and Enterococcus form a mutualistic relationship in the absence of Staphylococcus and Pseudomonas

To identify putative mutualistic interactions between less abundant species, we analyzed the ensemble of 1,264 community models in which both Staphylococcus and Pseudomonas were absent (∆Sa∆Pa). This ensemble produced lower and more variable growth rates than the other ensembles due to lack of the two dominant, growth-promoting pathogens (Figure 4). This growth reduction was accompanied by an increase in community richness as all 10 remaining species were able to coexist in some communities and no species was predicted to have an average abundance less than 3% (Figure 8A). This enhanced richness translated into higher equitabilities than predicted for the other ensembles (Figure 8C), again supporting the hypothesis that pathogen emergence results in resilient communities characterized by increased growth and reduced diversity.

Figure 8:

Model ensemble predictions for ∆Sa∆Pa simulations showing a mutualistic relationship between Streptococcus and Enterococcus. (A) Average species abundances for all 5,000 ensemble simulations (blue bars) and 1,264 ∆Sa∆Pa simulations (red bars). (B) Streptococcus and Enterococcus abundances for 51 simulated communities in which both species could participate where the colorbar indicates the number of simulations represented by each circle. The two species showed a mutualistic interaction (r = 0.53, p < 10⁻⁶). (C) Community growth rates and equitability for 51 simulated communities with both species. (D) The four most significant crossfeeding relationships between Streptococcus (purple bars) and Enterococcus (green bars).

Rather than focusing on mutualisms with respect to a single species, we used the ∆Sa∆Pa simulation results to identify mutualistic relationships between any pair of the 10 remaining species (90 possible cases). The analysis produced 22 significant interactions (Table 2, Figure S3), suggesting that mutualistic benefits could be spread across more species in the absence of dominant pathogens. For example, the commensal Corynebacterium positively interacted with less abundant pathogen Stenotrophomonas (Figure S5). We focused on the mutualistic relationship between Streptococcus and Enterococcus (Figure 8B) because these two genera are known to coexist in infections (Chávez de Paz et al., 2015; Gao et al., 2016). As before, the community growth rate and equitability were positively correlated (r = 0.49, p = 10⁻⁵, Figure 8C) in the 51 communities in which both species could participate. The two species interacted through the crossfeeding of multiple metabolites (Figure 8D), with L-lactate consumption important for Enterococcus coexistence. Streptococcus and Enterococcus are known to form thick and dense biofilms on root canal dentin and glass slides (Gao et al., 2016). Furthermore, Enterococcus has been shown to be more resistant to starvation in coexistence with Streptococcus (Gao et al., 2016), an interaction our model attributed to L-lactate crossfeeding.

Discussion

Polymicrobial infections in chronic wounds are responsible for poor clinical outcomes and cause elevated mortality rates as compared to single-species infections (Tzaneva et al., 2016). Colonizing species establish mutualistic relationship through multiple mechanisms including metabolite crossfeeding to promote community stability and resilience (West et al., 2006; Ramsey et al., 2011). Robust community structures mitigate the effectiveness of antibiotic treatments and promote the evolution of antibiotic resistance through mechanisms such as horizontal gene transfer (Davies and Davies, 2010; Hall and Mah, 2017). The communities place an increasing bioburden on the host and play a critical role in impaired/delayed wound healing (Gardner and Frantz, 2008; Tuttle, 2015). While recent studies based on 16S rRNA (Price et al., 2009) and rDNA (Rhoads et al., 2012; Wolcott et al., 2016) sequencing have revealed key bacterial taxa involved in chronic wound infections, knowledge about the interspecies mechanisms that drive community structure and function have remained elusive.

We developed a 12 species community metabolic model to identify putative interactions that drive the composition of chronic wound communities. The 12 modeled bacterial species covered 74% of 16S rDNA pyrosequencing reads of genera from 2,963 chronic wound patients (Wolcott et al., 2016). The metabolism of each species was described with a genome-scale metabolic reconstruction obtained from the AGORA database (Magnúsdóttir et al., 2017). Therefore, our predictions of community metabolism were dependent on the accuracy of these metabolic reconstructions. We used the limited data available from chronic wound metabolomics studies (Sood et al., 2015; Ammons et al., 2015; Junka et al., 2017) as a starting point to define community uptake rates as required in the SteadyCom modeling framework (Chan et al., 2017). Model tuning was used to define 81 host-derived nutrients and their uptake rates such that each species was capable of monoculture growth and predicted species abundances were in rough agreement with normalized 16S values from the original study (Wolcott et al., 2016). The tuning process required the introduction of 30 metabolites to achieve monoculture growth and three putative chronic wound metabolites to enhance the growth rates of particular species: starch 1 (Corynebacterium), kestose (Enterococcus) and glycerol-3-phosphate (Prevotella). As discussed in our previous modeling study on cystic fibrosis communities (Henson et al., 2019), the 30 essential metabolites suggest limitations for the AGORA genome-scale metabolic models (Magnúsdóttir et al., 2017) with respect to biosynthetic pathways leading to biomass formation. The presence of the three growth enhancing metabolites in chronic wound beds would need to be tested through metabolomics.

The tuned single-species models offered a wide range of predicted metabolic capabilities with respect to their growth rates and metabolite secretion patterns (Figure 2). As found in our previous modeling studies on gut (Henson and Phalak, 2018) and cystic fibrosis (Henson et al., 2019) communities, pathogens such as Staphylococcus, Pseudomonas and Stenotrophomonas generally had higher growth rates than commensal species, suggesting that they are more metabolically capable of dominating the community. When compared to normalized 16S-derived abundances averaged across the 2,963 patients, the tuned community model predicted relatively high abundances for the most highly represented genera (Staphylococcus, Pseudomonas, Corynebacterium) but underpredicted or overpredicted abundances of the remaining genera and generated a relatively low diversity community.

We sought to improve the prediction of community composition by imposing genera prevalence data available from the original study (Wolcott et al., 2016) as additional in silico constraints. The prevalence data was used to generate an ensemble of 5,000 communities in which the participating species of each community were randomly determined. While the in silico prevalences averaged over the 5,000 simulations deviated slightly from the 16S-derived values (see Materials and Methods), the average species abundances predicted by the prevalence- constrained model ensemble showed substantially improved agreement (Figure 3). These results demonstrate the difficulties in predicting 16S-derived abundances averaged over large numbers of patient samples with a single community model that is best thought of as simulating a single patient sample.

Because Staphylococcus and Pseudomonas are the dominant pathogens observed in most chronic wound infections (Gjødsbøl et al., 2006; Fazli et al., 2009; Pastar et al., 2013; Bessa et al., 2015), we were interested in community behavior in the presence and absence of these two pathogens. To overcome the lack of individual patient sample data in the original study (Wolcott et al., 2016), the ensemble of 5,000 community simulations was partitioned into four groups: both Staphylococcus and Pseudomonas allowed to participate (SaPa, 820 cases); Pseudomonas not allowed to participate (Sa∆Pa, 2,410 cases); Staphylococcus not allowed to participate (∆SaPa, 506 cases); and neither Staphylococcus or Pseudomonas allowed to participate (∆Sa∆Pa, 1,264 cases). We sought to computationally identify mutualistic relationships between species for each of the four scenarios since mutualisms reduce competition for available nutrients and result in robust communities associated with prolonged infections and poor clinical outcomes (Bowler et al., 2001). These putative mutualistic interactions were viewed as future targets for experimental testing and possible therapeutic disruption to enhance treatment efficacy.

When the pathogens Staphylococcus and Pseudomonas could participate in communities, the only significant mutualistic relationship predicted was between the two pathogens themselves. These SaPa communities were characterized by pathogen dominance, low diversity and high growth rates with little variability (Figure 5), characteristics we previously attributed to resilient communities well progressed towards a fully developed disease state (Henson and Phalak, 2018; Henson et al., 2019). Mutualism was supported by bi-directional crossfeeding of organic acids, amino acids and ethanol between the two species, making the identification of a single crossfeeding relationship for disruption a challenge. These predictions are supported by studies showing that the presence of Pseudomonas along with Staphylococcus generates larger chronic wounds and delays/prevents the healing process (Madsen et al., 1996; Gjødsbøl et al., 2006; Kirketerp Møller et al., 2008).

In the absence of Pseudomonas, Staphylococcus was predicted to form mutualistic relationships with the less abundant pathogen Acinetobacter and the commensal Corynebacterium. By spreading mutualism across three pairs of species, the Sa∆Pa ensemble produced slightly more diverse communities at the expense of slower and more variable growth (Figure 6). These results suggest that infections lacking Pseudomonas should be more easily treated, a hypothesis supported by the aforementioned studies (Madsen et al., 1996; Gjødsbøl et al., 2006; Kirketerp Møller et al., 2008). The Staphylococcus–Acinetobacter interaction was driven by lower metabolite crossfeeding rates than those predicted for Staphylococcus and Pseudomonas, another indication that Pseudomonas-free infections should be more easily cleared. These predictions could yield new insights into the treatment of the so-called ESKAPE pathogens (Enterococcus faecalis, Staphylococcus aureus, Klebsiella pneumoniae, Acinetobacter baumannii, Pseudomonas aeruginosa, Enterobacter species) which are the leading cause of nosocomial infections (Boucher et al., 2009; Santajit and Indrawattana, 2016).

When Staphylococcus was omitted from the simulated communities, Pseudomonas was predicted to have mutualistic relationships with four other species: Serratia, Streptococcus, Acinetobacter and Bacteroides. Consistent with the trends mentioned above, this increase in the number of mutualistic interactions resulted in the ∆SaPa ensemble producing more diverse communities which exhibited slower and more variable growth (Figure 6). The effect of removing Staphylococcus in the Sa∆Pa communities was greater than the effect of removing Pseudomonas in the Sa∆Pa communities, consistent with the role of Staphylococcus as the single dominant pathogen in chronic wound infections whose absence correlates to better clinical outcomes (Demling and Waterhouse, 2007; Pastar et al., 2013). The Pseudomonas-Serratia interaction was primarily driven by L-lactate and D-lactate exchange between the two species, a prediction that could be tested through in vitro experiments.

Of the 90 pairwise interactions possible when both Staphylococcus and Pseudomonas were removed, 22 interactions were predicted to be significantly mutualistic. The ∆Sa∆Pa ensemble exhibited substantially higher diversity and lower and more variable growth than the other three ensembles, consistent with earlier stage infections that lack dominant pathogens. One particularly interesting mutualistic relationship involved the commensal Streptococcus and ESKAPE pathogen Enterococcus, which has some experimental support (Gao et al., 2016). Our model predicted that this interaction was driven largely by L-lactate consumption by Enterococcus, demonstrating how pathogens may take advantage of metabolic byproducts secreted by commensals to increase their abundance when more dominant pathogens are absent.

Supplementary Material

Supp TableS1

Supplementary Table 1: In silico media components and their community uptake rates for chronic wound community simulations without and with species prevalence constraints.

Supp figS1-5

Supplementary Figure 1: Comparison of average and outlier community compositions for the four groups of community simulations.

Supplementary Figure 2: Correlation analysis of predicted species abundances for SaPa and Sa∆Pa community simulations.

Supplementary Figure 3: Correlation analysis of predicted species abundances for ∆SaPa and ∆Sa∆Pa community simulations.

Supplementary Figure 4: Model ensemble predictions for ∆SaPa simulations showing a mutualistic relationship between Pseudomonas and Streptococcus.

Supplementary Figure 5: Model ensemble predictions for ∆Sa∆Pa simulations showing a mutualistic relationship between Corynebacterium and Stenotrophomonas.

Significance and Impact of the Study.

Approximately 2% of the US population suffers from non-healing chronic wounds infected by a combination of commensal and pathogenic bacteria. These polymicrobial infections are often resilient to antibiotic treatment due to the nutrient-rich wound environment and species interactions that promote community stability and robustness. The simulation results from this study were used to identify putative mutualistic interactions between bacteria that could be targeted to enhance treatment efficacy.