Growth-altering microbial interactions are responsive to chemical context

Microbial interactions are ubiquitous in nature, and are equally as relevant to human wellbeing as the identities of the interacting microbes. However, microbial interactions are difficult to measure and characterize. Furthermore, there is growing evidence that they are not fixed, but dependent on environmental context. We present a novel workflow for inferring microbial interactions that integrates semi-automated image analysis with a colony stamping mechanism, with the overall effect of improving throughput and reproducibility of colony interaction assays. We apply our approach to infer interactions among bacterial species associated with the normal lung microbiome, and how those interactions are altered by the presence of benzo[a]pyrene, a carcinogenic compound found in cigarettes. We found that the presence of this single compound changed the interaction network, demonstrating that microbial interactions are indeed dynamic and responsive to local chemical context.


Introduction
Microbes are rarely alone. Microbial communities are nature's workhorses, from degrading tree litter in the forest to degrading food in the colon [1,2]. Complex communities of bacteria and fungi surround the roots of plants and colonize the surfaces of our teeth [3,4]. Interactions between microorganisms within these communities can entirely determine the overall interaction of the community with the environment. A potent example is the influence of Clostridium scindens on Clostridium difficile [5]. C. difficile is an intestinal pathogen that can reside indefinitely in the intestinal tract of healthy humans alongside hundreds of other species. However, after antibiotic treatment, C. difficile often outcompetes its neighbors and produces toxins, causing the host to experience intense diarrhea, fever and weight loss. Interestingly, the addition of a single species-C. scindens-to the intestinal community can prevent C. difficile overgrowth and the associated negative symptoms [5]. A single interaction can make the difference between health and disease. Similarly, disease severity can be influenced by microbial interactions. For example, Burkholderia cenocepacia increases mortality among cystic fibrosis patients with Pseudomonas aeruginosa-associated lung infections, illustrating that interactions among microbes are just as relevant to human health as the identities of the species with which humans interact [6]. However, it is difficult to measure and characterize microbial interactions [7]. Interactions have many underlying causes including direct interactions through signaling or toxin molecules, competition and cross-feeding, or the result of one species actively changing environmental factors such as pH [8]. Interactions are particularly difficult to measure in complex, multi-species communities. Recent efforts to elucidate the network of microbial interactions relied on time-series measurements of metagenomic sequence data, inferring interaction information from microbial abundance dynamics [9,10]. While useful, these approaches often assume that the nature of microbial interactions is fixed, an assumption which can mask important biological insights. Returning to the interaction between C. scindens and C. difficile, it was shown that C. scindens prevents C. difficile overgrowth by producing secondary bile acids, which is only possible in the presence of primary bile acids [5]. Clearly, in at least some cases, the chemical and nutritional context of the environment determines the interactions which are possible. The question of "how fixed are microbial interactions?" is still open, and answering this question requires innovation in the ways we measure microbial interactions, and requires many more observations of microbial communities in many different contexts.
We present a novel screening approach to quantify microbial interactions in vitro, enabling us to measure how those interactions change as a function of the chemical environment. As a test case, we chose to examine the influence of benzo[a]pyrene (BaP) on the interactions between a subset of "core" lung bacterial species [11,12]. BaP is a polycyclic aromatic hydrocarbon that is found in cigarette smoke which can interfere with DNA replication [11]. We quantified the pairwise interactions between bacterial species in control and BaP media conditions using a novel inoculation protocol and semi-automated image analysis. We found that at least one interaction was completely altered from growth-promoting to growth-inhibiting, while many other interactions changed more subtly in sign or magnitude. This proof-of-principle demonstration highlights the utility of our new framework and the fact that microbial interactions are highly dynamic and responsive to the chemical environment. Improved tools will lead to greater awareness and understanding of microbial interactions, and to an increased ability to target them therapeutically.

Species selection
Previous research suggests the existence of a core microbiome in the human lung [12]. We selected bacterial species which are commonly associated with the respiratory tract [12]: P. aeruginosa PA01, P. aeruginosa PA14, Haemophilus influenza type B ATCC 10211, Haemophilus parainfluenzae ATCC 7901, and Staphylococcus aureus ATCC 29213.

Media and culture protocol
All five species were cultured in brain-heart infusion (BHI) medium (BD) supplemented with L-histidine (0.01 g/L) (Sigma), hemin (0.01 g/L) (Sigma) and β-NAD (0.01 g/L) (Sigma) [13]. For the agar plates, we added granulated agar (BD) at 1.2% by weight. We prepared a stock solution of BaP (Sigma) dissolved in DMSO at 10 mg/mL and filter sterilized this solution (0.2 μm pore size). We added 250 μL of this solution into 1L of supplemented BHI for BaP conditions for a final concentration of 2.5 μg/ml. For reference, a single cigarette contains 3.4-28.4 ng of BaP [14].
On day 0, we made the agar plates and the liquid medium. We allowed liquid cultures to grow for 24 hours in a shaking incubator at 37˚C and 5% CO 2 . On day 1, we collected OD600 measurements for each of the liquid cultures, diluted the liquid cultures to equal OD600 with fresh medium, and evenly spread 7 mL on agar plates to create a lawn of each species which grew for 24 hours. On day 2, each species was stamped onto fresh 6-well agar plates using a custom stamping mechanism which ensured equal initial spacing and colony size (Fig 1). Each species was grown alone and in pairwise combinations with the other four species. The stamped 6-well plates were placed in the incubator for 24 hours before imaging (Fig 1).

Stamping mechanism
In measuring bacterial areas as one of our final metrics, it was necessary to ensure that the initial bacterial colonies were stamped at a consistent size and spacing. We selected a starting spot size of 0.5 mm diameter, which were placed 3.5 mm apart (from center to center). To achieve these specifications, we used metal nails to pick up the bacteria from the lawn and a 3D-printed mechanism to stamp the bacteria onto the plates (Fig A in S1 Supplemental Material). The metal nails had uniform tip size and were non-porous, which improved initial inoculation consistency (1 ½ inch steel box nails, size 4D). The stamping mechanism had two fixed slots, each sized to allow the nails to slide in. These slots ensured that the two nails were the same distance apart for each stamping. The nails were autoclaved and the stamping mechanism was routinely disinfected with CaviCide (Metrex). The STL file for 3D printing the stamping mechanism has been included as S1 File and is also available in the online repository for this project. All five species were grown in overnight liquid cultures. Cultures were diluted with fresh media, spread evenly over agar plates and incubated for 24 hours to grow into lawns. We devised a stamping mechanism to maintain consistent spacing and sizes of initial colonies. Nails were used to lift cells from the lawns. The nails were placed into slots in a 3D-printed stamping mechanism and cells were carefully "stamped" onto fresh BHI agar poured in 6-well plates. In the control condition, all five species were grown alone, and all pairwise combinations were grown together (six replicates were performed of each condition). For the BaP condition, the procedure was identical except that BaP was added to the growth media. The stamped colonies were incubated for 24 hours and the resulting colonies were imaged at 2x magnification. The images were automatically segmented and colony areas calculated using image analysis software (see Materials and methods). https://doi.org/10.1371/journal.pone.0164919.g001

Imaging bacteria
Our images were captured by an EVOS microscope (ThermoFisher Scientific) at 2x magnification and using the brightfield mode (Fig 1). We collected images for six biological replicates of each condition (Fig 2). In order to convert colony areas from pixels to mm 2 , additional images were taken of size standards.

Data analysis
Images were analyzed in CellProfiler, an open-source software package [15]. While bacterial colonies were traced automatically in CellProfiler, we manually checked each image to ensure that edge traces were accurate. We collected measurements of colony area (in pixels), shape and location. Colony areas were converted from pixels to mm 2 based on the images of paper standards (Fig 3). In order to infer microbial interactions, colony areas in paired growth conditions were compared to colony growth alone using a two-sided Wilcoxon signed-rank test in R [16]. The p-values produced were adjusted for multiple testing using the false discovery rate (FDR) method. Similarly, we compared the interactions from the control condition to the condition with BaP exposure in order to identify cases where the nature of an interaction changed. To accomplish this, we converted colony areas to area fold changes by dividing the observations by the mean area when the species was grown alone. The observed fold changes in control versus BaP conditions were compared using a two-sided Wilcoxon signed-rank test and pvalues were adjusted for multiple testing using the FDR method.
All of our images and analysis scripts are available in our online repository: https://github. com/mbi2gs/BaP_microbialInteractions Final colony areas across all conditions. Colony areas were obtained from the images using semi-automated image analysis in CellProfiler. Areas were converted from pixels to mm 2 based on images of paper standards. Control conditions are shown in blue while BaP conditions are shown in red. Individual data points are shown in gray (note that some outliers are outside the axis range of these plots). The colony areas correspond to the species labels on the left, while the labels along the bottom indicate the paired species. Species interactions are inferred by comparing pairwise growth with growth alone. For example, notice that when grown next to P. aeruginosa PA14, all the other species reach smaller colony areas than when grown alone, suggesting that P. aeruginosa PA14 negatively interacts with all four other species. https://doi.org/10.1371/journal.pone.0164919.g003

Interspecies interactions observed in the control condition
Final colony areas of four of the five species were significantly altered by at least one interaction in the control media condition (p-value < 0.05 by two-sided Wilcoxon signed-rank test corrected by FDR). We observed that both of the P. aeruginosa species grew more when paired with H. parainfluenzae or H. influenzae in control media, and the increase of growth was statistically significant in the case of P. aeruginosa PA01 (Fig 4A). S. aureus grew less when it was paired with both of the P. aeruginosa species. P. aeruginosa PA01, S. aureus, H. parainfluenzae and H. influenzae all grew less when paired with P. aeruginosa PA14.
We also grew each species alone and paired with itself in order to quantify self-interactions. In this instance, none of the colonies grown alone were significantly different from those paired with self (p-value > 0.4 by two-sided Wilcoxon signed-rank test corrected by FDR) (Fig  B in S1 Supplemental Material).

BaP alters interspecies interactions
We did not observe statistically significant changes in bacterial colony growth when grown alone in the presence of BaP after correcting for multiple testing, although the increase in H. influenzae colony size was significant before FDR correction (Fig 3). In terms of interactions, we observed that S. aureus grew significantly less when paired with P. aeruginosa PA14 in the BaP media condition (Fig 4B; p-value < 0.05 by two-sided Wilcoxon signed-rank test corrected by FDR). This same interaction (P. aeruginosa PA14 inhibition of S. aureus) was also observed to be statistically significant in the control condition (Fig 4A). Some interactions from the control condition, while no longer statistically significant in the BaP condition, did retain the same trends from the control condition. P. aeruginosa PA01 still grew less when paired with P. aeruginosa PA14 and S. aureus growth was inhibited when paired with P. aeruginosa PA01 (Fig 4B). The trend of other interactions from the control condition seemed to change in the BaP condition (Fig 4C). The only statistically significant change was the interaction between H. influenzae and P. aeruginosa PA01. In the control condition, P. aeruginosa PA01 grew more when paired with H. influenzae, while in the BaP condition, P. aeruginosa PA01 grew less when paired with H. influenzae (Fig 4C). A similar trend can be seen in the interaction between H. parainfluenzae and P. aeruginosa PA01, although it is not statistically significant after multiple testing correction.
In some instances, we observed colonies coming into physical contact after 24 hours of growth. We repeated the experiment for P. aeruginosa PA01 and H. parainfluenzae and included a semipermeable membrane (0.1 μm pore size) separating the two species (Fig C in  S1 Supplemental Material). There was no significant difference between colony areas in conditions with and without the membrane (p-value > 0.08 by two-sided Wilcoxon signed-rank test).

Discussion
We devised an in vitro experimental framework that allows for the semi-automated assessment of pairwise interactions between microbes grown in vitro. We demonstrate our framework on a set of five representative lung bacterial species. We identified several growth-altering interactions in both control and BaP conditions, and significantly, we saw that the nature of at least one observed interaction was significantly different between the two conditions, indicating that interspecies interactions are altered by the presence of BaP.
Our pipeline includes a novel stamping mechanism for standardizing colony sizes, combined with quantitative image analysis (Fig 1 and Fig A in S1 Supplemental Material). We incorporated the open-source image analysis software CellProfiler into our workflow to increase throughput and quantify interactions [15]. Our methodology allows for automated, quantitative measurements of colony attributes whereas previous research using similar techniques has relied on manual observations [17]. For this work, we chose to focus on colony area Heat maps indicate the fold change in colony area during pairwise growth with respect to growth alone. Fold changes in colony area are shown for the species in the columns. The species along the rows are those which influenced the growth of the species in the columns. Statistically significant changes in colony areas are indicated by yellow stars (p-value < 0.05 by two-sided Wilcoxon signed-rank test corrected by FDR). A. Fold changes in colony area in control media. S. aureus grows significantly less when grown next to P. aeruginosa PA01, and P. aeruginosa PA01 grows significantly less when paired with P. aeruginosa PA14. B. Fold changes in colony area in BaP media. Only one interaction is statistically significant in this context, and it is an interaction that was also observed in the control condition. C. Panel A subtracted from panel B. The difference in fold change is an indication of how different the interactions are in the two conditions. Statistical significance was determined by two-sided Wilcoxon signed-rank test on the fold change data for each interaction, and then FDR corrected. Only one interaction was significantly changed between the two conditions. P. aeruginosa PA01 grew more when paired with H. influenzae in control media, but grew less in BaP media. Some interactions were essentially identical between the two conditions (e.g. P. aeruginosa PA01 grows less when paired with P. aeruginosa PA14, regardless of the presence of BaP), and others were changed, if not in a statistically significant way (e.g. P. aeruginosa PA01 grows more when paired with H. parainfluenzae in control media, but less in BaP media). D. An illustration of the BaP-dependent interactions between H. influenzae and P. aeruginosa PA01. We put question marks beside the interactions which are observed but not statistically significant. Note the representation of the hypothesis that BaP causes H. influenzae to switch from a positive to a negative influence on P. aeruginosa PA01. as an indicator of interaction strength. Colony area is a function of both growth rate and duration. We assume that a bacterial colony will grow more slowly and/or stop growing earlier when the cells are under stress (due to a negative interaction with a neighboring species) [18]. Similarly, we assume that a colony will grow faster and/or longer if it is participating in a beneficial interaction. It should be noted that there is some variation in the initial inoculation density, which can influence the final colony area. For this reason, it is important to include sufficient experimental replicates. We chose to use single colony growth as a control representing an absence of any interaction. We compared single colony growth to double colony growth in which a species was paired with itself ( Fig B in S1 Supplemental Material). This paired-withself growth is a control representing perfect metabolic and spatial competition. Under these specific conditions, we found that single and paired-with-self growth were statistically the same. However, in future experiments, both controls can be used to benchmark interactions with other species.
While the workflow does not give much insight into the mechanism of the interaction, the use of a semi-permeable membrane allows the experiment to discriminate between contactdependent and-independent interactions. Contact-dependent interactions would require cells to come into physical contact, as when toxins are injected into neighboring cells via protein secretion systems [19]. Contact-independent interactions act over a distance, and are responsible for the majority of the interactions observed in this study since generally the colonies were not physically interacting. Potential diffusion-based causes of interactions include (but are not limited to) metabolic competition, cross-feeding, signaling, or toxin-mediated interactions [20,21].
In this study we observed many interspecies interactions, some of which have never been reported in the literature. BaP is a carcinogen present in cigarettes, which we added to the media at concentrations higher than that likely found in the lung after smoking. However, the specific concentration matters less than the overall outcome that changes in chemical context influence the nature and strength of interspecies microbial interactions. In both control and BaP conditions neighboring species grew less when paired with P. aeruginosa PA14 (and to a lesser extent, PA01; Fig 3). One contributing factor is that P. aeruginosa PA14 and PA01 colonies were enormous (Fig 2), suggesting that these two species utilized nutritional resources more rapidly than neighboring species could access them. It is also known that many species of P. aeruginosa excrete molecules that broadly inhibit growth among neighboring species, including S. aureus and Haemophilus species [17,22,23]. We also observed that both H. influenzae and H. parainfluenzae promote growth of P. aeruginosa PA01 (and to a lesser extend PA14), an interaction which to our knowledge has never been reported before (Fig 4A), highlighting the utility of our approach for identifying relevant and novel interspecies interactions.
An interesting outcome was the sensitivity of the observed interactions to the presence of BaP. In the control condition, P. aeruginosa PA01 growth was enhanced when paired with H. influenzae, while in the presence of BaP, P. aeruginosa PA01 growth was inhibited when paired with H. influenzae, completely reversing the nature of the interaction (Fig 4D). While this reversed interaction was the most dramatic observation, many interactions changed in nature or intensity to some extent, if not statistically significantly (Fig 4C). It is remarkable that the addition of a single chemical could so drastically alter elements of the interaction network, which has important implications in the context of human health where the effect of pharmaceuticals on the human microbiota are rarely understood [24]. There is growing evidence that microbial interactions are not fixed, but are context-dependent and in continuous feedback with the local chemical backdrop [17,25,26]. Computationally, it has been shown that pairs of bacterial species can be made to compete or cooperate based solely on the nutritional context in which they are grown [27]. The context-dependence of bacterial interactions becomes relevant during attempts to predict the dynamics of important microbial communities such as the gut microbiome. Previously, mathematical frameworks for inferring microbial interaction networks have been based on the assumption that interactions are fixed [10,28]. Advances in predicting and engineering community dynamics will come as we increase our understanding of the effect of context on microbial interactions.
Future applications of our workflow can easily utilize metrics other than colony area. Readily-included metrics include colony shape and color. While we did not observe visible phenotypic differences in this particular study, previous research suggests that shape and color can reveal important interactions. For example, pseudomonads can exhibit a "swarming" phenotype which drastically alters colony shape, and Streptomyces coelicolor can produce a large variety of pigments which differ depending on the identity of neighboring species [29,30]. Furthermore, this method is extensible to higher order interactions between three (or more) species, similar to previous work exploring how spatial arrangement of colonies influences interactions [31].
The novel pipeline applied here enables faster screening of microbial interactions in vitro. Our results derived from representative lung bacteria highlight the fact that microbial interactions are dynamic and responsive to environmental perturbations. Improved tools such as those presented here will lead to greater understanding of such interactions, and to an increased ability to modulate them therapeutically.
Supporting information S1 File. STL file suitable for 3D printing the stamping mechanism described in this work. (STL) S1 Supplemental Material. Additional resources describing the novel stamping mechanism, control growth conditions and membrane-separated growth. (DOCX)