Metabolic modeling of the International Space Station microbiome reveals key microbial interactions

Background Recent studies have provided insights into the persistence and succession of microbes aboard the International Space Station (ISS), notably the dominance of Klebsiella pneumoniae. However, the interactions between the various microbes aboard the ISS and how they shape the microbiome remain to be clearly understood. In this study, we apply a computational approach to predict possible metabolic interactions in the ISS microbiome and shed further light on its organization. Results Through a combination of a systems-based graph-theoretical approach, and a constraint-based community metabolic modeling approach, we demonstrated several key interactions in the ISS microbiome. These complementary approaches provided insights into the metabolic interactions and dependencies present amongst various microbes in a community, highlighting key interactions and keystone species. Our results showed that the presence of K. pneumoniae is beneficial to many other microorganisms it coexists with, notably those from the Pantoea genus. Species belonging to the Enterobacteriaceae family were often found to be the most beneficial for the survival of other microorganisms in the ISS microbiome. However, K. pneumoniae was found to exhibit parasitic and amensalistic interactions with Aspergillus and Penicillium species, respectively. To prove this metabolic prediction, K. pneumoniae and Aspergillus fumigatus were co-cultured under normal and simulated microgravity, where K. pneumoniae cells showed parasitic characteristics to the fungus. The electron micrography revealed that the presence of K. pneumoniae compromised the morphology of fungal conidia and degenerated its biofilm-forming structures. Conclusion Our study underscores the importance of K. pneumoniae in the ISS, and its potential positive and negative interactions with other microbes, including potential pathogens. This integrated modeling approach, combined with experiments, demonstrates the potential for understanding the organization of other such microbiomes, unravelling key organisms and their interdependencies. Video Abstract Supplementary Information The online version contains supplementary material available at 10.1186/s40168-022-01279-y.

International Space Station (ISS) is a unique and controlled system to study the interplay between the human microbiome and the microbiome of their habitats. The ISS is a hermetically sealed closed system, yet it harbors many microorganisms that survive extreme environmental conditions such as microgravity, radiation and elevated CO 2 levels [5][6][7][8]. Recent studies have demonstrated the unique link between crew member microbiomes and surface microbiomes on the ISS [9]. These microorganisms might have been brought into the ISS via routine payloads and astronauts [9].
Several studies have focused on microbial isolation [10] and molecular microbial community analyses of the ISS microbiome, through experiments performed on vacuum filter debris [11], high-efficiency particle arrestance (HEPA) filters [12,13], the ISS environmental surfaces [8,14] and astronauts' microbiome [9]. A recent report using a shotgun metagenomic sequencing approach on intact cells of the ISS environmental microbiome revealed the succession and persistence of certain microbial populations on the ISS environmental surfaces [15]. The study posited a dominant, viable presence of Biosafety Level -2 (BSL-2) pathogens such as Klebsiella pneumoniae, Staphylococcus aureus, Enterococcus faecalis and Salmonella enterica ( Figure S1).
K. pneumoniae is well-known for its ability to cause pneumonia and other nosocomial infections and is largely studied for its known resistance to a wide spectrum of antibiotics, such as carbapenems, and its hypervirulence [16][17][18][19][20]. The present study is motivated by the evidence presented about the dominance of K. pneumoniae at multiple locations of the ISS, its succession over time [15], and the potential clinical implications it could have on the health of the astronauts inhabiting the ISS.
Metabolic interactions are a key driver in shaping microbial communities [21]. Studying these metabolic interactions can be instrumental in understanding the interplay between various microorganisms in diverse communities [22][23][24]. Genome-scale metabolic modeling is a powerful tool to study and understand microbial metabolism [25]. These modeling approaches capture in detail, the known metabolic reactions happening in a microorganism, along with the enzymes that catalyze them, representing them in a mathematical form amenable to simulations [26]. Beyond single microorganisms, metabolic modeling can also be extended to study microbial communities; many paradigms have been developed, including those based on graph theory, and constraint-based modeling, as reviewed elsewhere [27,28].
In this study, we leverage the metagenome datasets that have captured the microbial composition of the ISS surfaces [15] to specifically predict how various species in these communities can influence one another's metabolism, leading to mutually beneficial interactions and stable microbiomes. In particular, we focus our study on K. pneumoniae and its coexisting species, to elucidate the possible underlying metabolic interactions that drive the structure of these communities. Through a combination of a systems-based graph-theoretical approach, Met-Quest [29,30], and a constraint-based community modeling approach, SteadyCom [31], we illustrate the central role of metabolic interactions and dependencies in shaping the ISS microbiome.

Methodology
In this study, a two-pronged computational approach was adopted to investigate and decipher the prevailing microbial interactions in the ISS. In the first graph-based approach [29], bipartite graphs were created to represent the metabolic networks of microorganisms taken individually, as well as when in pairs or larger communities. The metabolic networks underlying the organisms were represented as bipartite graphs, where two groups of nodes represent metabolites and reactions, and directed edges connect substrates to reactions and reactions to products. These networks were used to identify possible metabolic dependencies between different microorganisms. In the second constraint-based approach [31], microbial growth rates were predicted, individually and in communities, to gauge the nature of their interactions, as we describe in the following sections. Figure 1 shows a broad overview of the approach of this study, outlining the steps beginning with the identification of microorganisms coexisting with K. pneumoniae to various metabolic network analyses.

Data
Microbial abundances associated with samples taken across three flights at eight ISS locations were previously measured using metagenomic sequencing [15]. Among eight locations, only seven locations yielded measurable metagenomics data. In the present study, the data were further pruned such that only shotgun metagenome sequences associated with propidium monoazide (PMA)treated samples that recorded a presence of K. pneumoniae were retained. The PMA-treatment removed naked DNA and dead cells, thus retaining sequences associated with viable and intact cells only [32]. Further, only microorganisms that were found to coexist with K. pneumoniae with >1% relative abundance at the location were used in this study.
For this study, the available whole-genome sequences (WGS) of microorganisms from the ISS, which were compatible with KBase [33] along with reference genomes for the relevant microorganisms taken from the Fig. 1 Overview of the analysis pipeline. Beginning with identifying a list of microorganisms that coexist with K. pneumoniae, the genome-scale metabolic models were built using KBase and the extent of benefit derived by each individual microorganism in different scenarios was computed using MetQuest and the Metabolic Support Index (MSI) as a metric. SteadyCom, a constraint-based approach, was also employed here to determine the effect of a microorganism on the growth of another and to further classify the nature of their interactions into their respective types based on the observed change in growth rates NCBI database, were collected. The RefSeq and GenBank accession identification numbers, along with the underlying reasoning for the choice of these sequences, are provided in Supplementary Table S1. Although the relative abundance of Pantoea sp. PSNIH2 was >1% in some of the locations considered in this study, as its reference genome sequence was removed from the NCBI database, at the time of this study, this microorganism therefore had to be excluded. Metabolic models were constructed for the 52 microorganisms using ModelSEED [34]. The detailed process is elaborated in Supplementary Methods M1. The details of the reconstructions, templates and information about reaction counts are specified in Supplementary Table S2.

Predicting metabolic dependencies in the community
To investigate the microbial interactions underlying the chosen microbial communities of the ISS, MetQuest [29], a graph-theoretic algorithm was first employed. Through a novel dynamic-programming-based enumeration, MetQuest assembles reactions into pathways of a specified size producing a given target from a specified set of source molecules ('seed'). Through a guided breadth-first search, MetQuest can identify all the metabolites that can be traversed from a set of seed metabolites. This functionality was exploited in this study as previously published [30]. Using this algorithm, the present study presents a top-down view of metabolic dependencies in these communities, beginning with predicting the most supportive families, then the most dependent genera, and finally the beneficial and dependent microorganisms by looking at pairs of potentially interacting microorganisms. In each scenario, the extent of benefit derived was calculated using a measure known as the Metabolic Support Index (MSI) [30]. This was also extended to measure the possible support an organism derives from (or provides to) a community where it is present, denoted Community Support Index (CSI).
The composition of the seed metabolites was chosen so as to incorporate the co-factors and coenzymes that would be present in the environment and required by a living cell, and capture the minimal nutrient content offered by the ISS. The minimal medium of each metabolic model was identified by minimizing the components required to achieve a minimum growth rate of 0.1 h -1 , using built-in functions in COBRApy [35], namely minimal_medium from the cobra.medium module. This approach was adopted on the premise that each microorganism is capable of individual growth. For each location at each flight, the minimal medium sets corresponding to each constituent microorganism were pooled together to form a consolidated medium. Additionally, a base set of co-factors and coenzymes were added to each of these sets. The same medium was used for both the graphbased and constraint-based analyses. Supplementary Table S3 contains the list of all such metabolites.

Pairwise interactions
For a pair of microorganisms, A and B, MSI is calculated as follows: Here, n stuck, A | A ∪ B refers to the number of stuck reactions in A, when present in the community alongside B, while n stuck, A | A refers to the number of stuck reactions in A, when it is present in isolation. Individual bipartite graphs were constructed (n = 52), and the stuck reactions were first determined for every individual microorganism in the dataset, using the MetQuest algorithm as described above, which provided insight into the intrinsic metabolic capabilities of each individual microorganism. Following this, for every pair of coexisting microorganisms that may potentially interact, bipartite graphs were constructed for the two-member community (n = 426), and the stuck reactions were determined for each constituent member of each community, in a flight-locationspecific approach (n = 761), to derive the effect of its partner by virtue of being in this community, through the calculation of the MSI.
Building further on these interactions, as captured by the MSIs between various microbial pairs, microbial association networks were constructed and visualized using Cytoscape [36]. These networks capture the extent to which a microorganism is able to relieve the stuck reactions of another through metabolic exchanges.

Higher-order interactions
The enhancement of metabolic capabilities rendered by a microorganism(s) by virtue of it thriving in a community was derived from the MSI, here referred to as the community support index (CSI). The CSI, denoted CSI A→Ã throughout this manuscript, was calculated as the fraction of reactions relieved in the rest of the community over the reactions 'stuck' in the same in the presence of the microorganism. For a community X and an inhabitant microorganism, or family of microorganisms, A, the effect of A on the rest of the community Ã = (X − A), where Ã denotes the rest of the community X devoid of microorganism/family A, is computed as follows: (1) MSI(A|A ∪ B) = 1 − n stuck,A|A∪B n stuck,A|A (2) CSIÃ|X = 1 − n stuck,Ã|X n stuck,Ã|Ã Here, n stuck,Ã|Ã refers to the number of stuck reactions in the community X without microorganism A for the given set of seed metabolites. A reaction is considered stuck when it does not have the necessary precursors for the reaction to happen. In other words, these are reactions that cannot occur under the present conditions, without extraneous help in terms of other metabolites from other microorganism(s), or the environment itself. Similarly, n stuck,Ã|X refers to the number of reactions in Ã that remain stuck, even in the presence of A. We consider only internal reactions (as against transport or extracellular reactions) while computing these numbers.
For each of the sites considered in this study, bipartite graphs were first constructed (n = 129) for communities devoid of one constituent microorganism, and their respective stuck reactions were determined in the appropriate environment (i.e., the seed metabolites described above and listed in Supplementary Table  S3). This approach provided insight into the intrinsic metabolic capabilities of each community devoid of that microorganism. Following this, a total 11 bipartite graphs were constructed, taking into consideration all inhabitant microorganisms of a particular site. The stuck reactions were determined for each community to gather the effect of the interactions by calculating the CSI A→Ã . A CSI A→Ã of unity indicates that A completely supports Ã and relieves all its stuck reactions, whereas a value of zero indicates that A has no effect on Ã .
Microorganisms could potentially be metabolically similar, and therefore redundant to the community, offering little or no metabolic support individually. Therefore, microbial interactions were also studied by grouping them by their respective families (n = 7), at each location. The metabolic support provided by groups of microorganisms (e.g., a family) can be readily computed by replacing A in X − A with the set of microorganisms comprising the family in Eq. (1).
It is also possible to estimate the community benefit, in terms of enhancement of metabolic capabilities, an individual microorganism might receive by virtue of it thriving in a community. Here, the CSI, denoted CSIÃ →A , was calculated as the fraction of reactions relieved in the microorganism when in a community assemblage over the reactions stuck in the individual. For a community X and an inhabitant microorganism A, the effect of the rest of the community Ã on A is computed as follows: The CSIÃ →A thus captures the 'benefit' a microorganism or a group of microorganisms receives from another microorganism or group, by virtue of its stuck reactions being relieved via metabolic exchanges with its coexisting (3) CSI(A|X) = 1 − n stuck,A|X n stuck,A|A microorganisms. For clarity, all CSIs and MSIs have been represented as percentages.

Determining the nature of interactions
Microbial interactions have major relevance in building and shaping the community. Based on the effects a microorganism has on another, the interactions can be categorized into six types, namely amensalism, commensalism, competition, neutralism, mutualism and parasitism [37]. The present study reports a constraint-based community modeling study using the algorithm, Steady-Com [31]. SteadyCom [31] was employed to determine the biomass production rates of all microorganisms that are a part of a community under steady state. The seed metabolites described earlier, in the network-based approach were used here as the medium (see Supplementary  Table S3). The lower bounds of the exchange reactions for the uptake of these metabolites were constrained to −10 mmol/gDW-h [38]. Joint models were then created for all pairs of microorganisms that can potentially interact by virtue of being at the location at that point in time. These joint models were optimized using SteadyCom [31] to find the individual biomass rates when existing as a community.
The growth rate observed for each microorganism in community was compared with that observed for the individual microorganism. A 10% or higher increase or decrease in growth rate, when in a community, was taken to be a significant effect [39]. The interactions were accordingly classified into the types listed above. The equation defined below was used to determine the effect of microorganism B on A, in a two-membered community constituting A and B.
where v bio, AB is the community biomass production rate of AB and v bio, A is the biomass production rate of A.

K. pneumoniae and A. fumigatus interactions
Overnight grown cultures of K. pneumoniae (10 6 cells per 100 μL) and purified conidia of A. fumigatus (10 6 per 100 μL) were either mixed or grown alone in a 10-mL sterile potato extract (0.4%; w/v) dextrose (2%; w/v) liquid medium and grown under normal and simulated microgravity conditions using high aspect ratio vessel (HARV; Synthecon inc., Houston, TX) culturing units [40]. The autoclavable HARVs provided oxygenation to the culture media via their large-diameter gas permeable membrane.

Data acquisition and genome-scale metabolic network reconstruction
According to the data published by Singh et al. 2018 [15], among the PMA-treated samples, reads of K. pneumoniae were detected in a total of 11 sites-locations #1, #2, and #5 in Flight 1; location #5 in Flight 2; and locations #1, #2, #3, #4, #5, #7, and #8 in Flight 3 (Fig. 2). The relative abundances of microorganisms were calculated at the respective sites, and upon further pruning of this data to include solely those with >1% abundance, a total of 50 different strains of microorganisms that coexist with . The octagonal location nodes have been color coded according to the ISS Node they belong to. Locations #1, #2 and #3 belong to Node 3; locations #4 and #5 to Node 1; location #6 (not shown here) to PMM and location #7 to the US Laboratory. The microbial nodes have been color-coded such that those existing at more than one location with K. pneumoniae are shown in grey, implying that a microorganism found in a large number of locations (such as K. pneumoniae in F3) will be filled with a darker shade of grey K. pneumoniae at varied locations and time points were chosen for further study. Analysis of the diversity of these microorganisms at each of these locations over time (Fig. 2) reveals the following trends. In Flight 1, among the three locations that were colonized by K. pneumoniae, there was an observed dominance of fungi in location #1, Enterobacteriaceae in location #3, and Erwiniaceae in location #5. In Flight 2, K. pneumoniae was documented to be present only in location #5, coexisting with extant and dominant Erwiniaceae, a member from the Enterobacteriaceae family and another from the Paenibacilleae family. In Flight 3 however, K. pneumoniae spread across the ISS, inhabiting nearly all sampled locations. Additionally, many more members of Erwiniaceae, specifically the Pantoea genus, emerged into view at a majority of these locations, as shown in Fig. 2. Further, with the exception of A. niger and Rhodoturula sp. strain JG-1b, all fungi considered in the dataset appear to be concentrated in location #3. Location #3 is also dominated by a large number of Staphylococcaceae. The list of microorganisms, along with their relative abundances at each location in each flight, is provided in Supplementary Table S4.

Enterobacteriaceae are pivotal contributors to community metabolism
Through a 'leave-one-out' approach, microorganisms were knocked out one at a time from the community to estimate the extent of benefit provided by that single microorganism to the remaining coexisting microorganisms. From the analysis, it was evident that some microorganisms do have an influence on the metabolic capabilities on the rest of the community, albeit low in most cases. Notable among these, were P. rubens at location #3 during Flight 3 ( CSI A→Ã = 1.15%), and at locations #2 ( CSI A→Ã = 0.40%) and #5 ( CSI A→Ã = 0.36%) during Flight 1, and K. oxytoca at location #2 in Flight 1 ( CSI A→Ã = 0.49%) (Supplementary Table S5 and Supplementary Figure S2).
These low values of CSI A→Ã are expected, given the possible high metabolic overlap between members within a community, thereby rendering individual organisms less important to a community. To examine this, we grouped organisms based on their phylogenetic affiliation, and removed these groups, one at a time, from the community, to study the importance of a given family to the community.
Overall, with respect to the locations considered in this study, the Enterobacteriaceae family comprising species of the Klebsiella, Escherichia, Salmonella, Shigella, and Enterobacter genera, was often found to be the most beneficial of all clusters, with the highest CSI A→Ã s in six out of a total eleven communities under study (Fig. 3 and Supplementary Table S6), five of which were in Flight 3. With regards to the remaining five communities, Trichocomaceae have the highest CSI A→Ã s at locations #2 and #5 during Flight 1, Erwiniaceae have the highest CSI A→Ã s at location #1 in Flight 3 and at location #5 during Flight 2, and Paenibacillaceae has the highest CSI A→Ã at location #2 in Flight 3. Here, A refers to the family of microorganisms that provides metabolic support.

Pantoea species thrive in the support of the microbiome
The CSIÃ →A of an individual microorganism in the presence of its coexisting microorganisms was calculated to gain insight into the combined effect of the community on an individual. The CSIÃ →A values of Pantoea were found to be among the highest (Supplementary Table S7 and Fig. 4), indicating that these species are benefitted to a comparatively greater extent from the remaining members of the community. Across all locations and flights in the dataset, the CSIÃ →A values of these species ranged between 0 and 6.52%, wherein the values at the lower end were predominantly observed in those Pantoea species found at location #1 in Flight 3. P. vagans was found to exhibit the highest dependency amongst all the Pantoea, at location #5 in Flight 2 (Fig. 4).
With the exception of E. coli and S. enterica, the CSIÃ →A values of other Enterobacteriaceae fell in the range of 0-0.91%, suggesting an overall lower dependence of members of the Enterobacteriaceae on their coexisting microorganisms (Fig. 4). K. pneumoniae showed no dependency ( CSIÃ →A = 0%) in six out of the 11 locations that they were found. In addition, S. enterica exhibited the highest community dependence amongst all microorganisms in consideration, with a CSIÃ →A = 8.79% at location #3 during Flight 3, and along with E. coli, showed high CSIÃ →A values (Fig. 4).
Analysis across the three flights at location #5 reveals that four out of five organisms present at all instances are benefitted to a lesser extent in Flight 1 compared to Flights 2 and 3 ( Fig. 4 and Supplementary Table S7).

K. pneumoniae are beneficial to its coexisting microorganisms
In every community, microorganisms exchange metabolites, influencing one another's growth and survival. We performed a pairwise analysis of a total of 761 flight-location-specific communities to identify how every pair of microorganisms interacts with one another, at a given location, during a given flight, and through various metabolic exchanges support each other.
In communities with K. pneumoniae and a member outside of the Enterobacteriaceae, the MSI values in the presence of K. pneumoniae were found to fall in the range of 0-6.46%. Most Pantoea species were found to be highly dependent on K. pneumoniae, with MSIs at the higher end (4.41-6.46%) of the spectrum (Supplementary Table  S8). With the exception of E. coli and S. enterica, the MSI values of all other Enterobacteriaceae members were found to be 0% in the presence of K. pneumoniae (Supplementary Figures S3 and S4). Out of 118 pairs, K. pneumoniae was found to have a non-zero MSI value in only nine such pairs. In these nine pairs, the other interacting member belonged to a species of the Penicillium genus.
Since K. pneumoniae was present across the three flights only at location #5, the corresponding microbial association networks were constructed to gain insights into microbial interaction patterns over time (Fig. 5). In Flight 1, fungi such as P. rubens and Rhodotorula sp. JG-1b are prevalent and shown to decline in abundance over time (Flight 2 and Flight 3). On the contrary, members of the Panteoa genus were observed to dominate this location over time, with many more Pantoea species surfacing. In agreement with earlier results discussed, many species of the Enterobacteriaceae family shown here such as K. pneumoniae and E. cloacae, were beneficial to members of the Erwiniacea family to a great extent (Fig. 5).

Constraint-based modeling suggests the domination of amensalistic and parasitic interactions
The graph-theoretic approaches described earlier do not explicitly account for the growth of microorganisms in the presence of one another or their rates, although a higher MSI does point towards a possible higher biomass in the community [30]. Complementary to the above approaches, constraint-based modeling approaches such as SteadyCom [31] can shed light on the ability of microorganisms to grow in a steady-state. Based on these predicted growth rates, it is also possible to classify the nature of the interaction between various pairs of microorganisms [37].
A total of 761 flight-location-specific pairs of microorganisms were studied. Among these, 458 pairs exhibited amensalism and 166 were predicted to be parasitism. Of the remaining, 48 were commensal, 36 were competitive, 34 were mutualistic and 19 were neutral.
Looking closely into the nature of interactions of K. pneumoniae and its coexisting microorganisms, in 118 such pairs, 72 were predicted to be amensal, 14 commensal, 13 involved in competition, 10 were parasitic interactions, seven were mutual and two were neutral. In communities with the Enterobacter species, K. pneumoniae was involved in amensalistic interactions as observed across all such flight-location pairs. While in the interactions with E. cancerogenus and Enterobacter sp. strain NFIX59, the growth rate of K. pneumoniae was predicted to decrease, in the interactions with all other Enterobacter species, there was no significant effect on the predicted growth rate of K. pneumoniae.
In eight out of nine flight-location-specific communities of K. pneumoniae with E. coli, E. coli was predicted to have no significant change in growth rate while that of K. pneumoniae was predicted to decrease, suggesting an amensalistic interaction. Further, in Flight 3, at location #3, E. coli was predicted to be parasitic towards K. pneumoniae. In the interactions with S. enterica, in four out of the six locations, the growth rate of K. pneumoniae was decreased in its presence, whereas that of S. enterica was increased, resulting in a parasitic interaction. This is in accordance with the previously observed high MSI values at locations #3, #5, #7, and #8 during Flight 3. In the other cases during Flight 1 at location #2 and Flight 3 at location #1, the interaction was predicted to be amensalistic, wherein the growth rate of K. pneumoniae was reduced.
With other Klebsiella species, the interactions were predicted to be predominantly amensalistic, with the growth rate of K. pneumoniae reduced, except in the interaction with Klebsiella sp. strain MS 92-3, where the reverse was observed. K. pneumoniae was predicted to be parasitic towards K. aerogenes strain IIIF7SW-P1.
The interactions with Pantoea species, such as P. agglomerans, P. conspicua, Pantoea sp. strain 3.5.1, Pantoea sp. strain A4, Pantoea sp. strain At-9b, Pantoea sp. strain FF5, Pantoea sp. strain NGS-ED-1003, Pantoea sp. strain OXWO6B1 and P. vagans, were amensalistic with a decreased growth rate of the Pantoea species, in all cases except in Flight 3 at location #1, where the interactions were predicted to be competitive. Interactions with P. ananatis however were found to be commensal, with an observed increased growth rate. P. dispersa and Pantoea sp. strain IMH were found to compete with K. pneumoniae at their coexisting locations ( Fig. 6 and Supplementary Table S9).
Staphylococcus sp. in Flight 3 at location #2 were found to be mutualistic with K. pneumoniae strain F3-2P(2*), with the exception of S. saprophyticus that was observed to be involved in commensalism. During Flight 3, location #3 however, this interaction of S. saprophyticus with K. pneumoniae was predicted as mutualistic  Table S9). This observed change in the nature of interactions may perhaps be attributed to the differences in the simulated medium compositions at the different locations.
The interactions of K. pneumoniae with A. niger were observed to be parasitic, with an observed decreased growth rate of A. niger and an increased growth rate of K. pneumoniae. With Penicillium species, K. pneumoniae had an amensalistic effect on them, resulting in their decreased growth rates. The interactions with Rhodotorula sp. strain JG-1b were found to depend on the flight and location in which they coexisted. During Flight 1, at location #1, an increased growth rate was observed in Rhodotorula sp. strain JG-1b, whereas there was no significant effect on that of K. pneumoniae, thereby suggesting a commensal interaction. In locations #2 and #5, however, a neutral interaction was predicted, with no significant change in growth rates of either species. In Flight 3, K. pneumoniae was observed to be amensalistic towards Rhodotorula sp. strain JG-1b at location #1 and commensal at location #2. K. pneumoniae increased the growth rate of R. toruloides in Flight 1 at location #1 ( Fig. 6 and Supplementary Table  S9). Details regarding the nature of interactions of other microorganisms in the community have been delineated in Supplementary Table S9. Microbial association networks for pairs of microorganisms inhabiting location 5, across the three flights. Cytoscape was used to construct and visualize these networks across three flights: Flight 1 (F1), Flight 2 (F2) and Flight 3 (F3). The nodes are labelled with microorganisms that inhabit the site at that respective time point. The directed edges are directed from the metabolically supportive microorganism to the metabolically dependent microorganism. The color of these directed edges are mapped to the Metabolic Support Indices (MSI), represented as percentages, such that the gradient from light blue to dark blue represents an increasing MSI

Detrimental effect of K. pneumoniae on Aspergillus fumigatus
The parasitic interactions of K. pneumoniae on A. niger were observed using various microscopic techniques (data not shown). Since A. niger is not a pathogenic bacterium, we extensively characterized an ISS isolated A. fumigatus, a BSL-2 bacterium, to understand the detrimental effect of K. pneumoniae. The scanning electron micrographs confirmed the parasitic effect of K. pneumoniae over A. fumigatus (Fig. 7). Under normal gravity conditions, A. fumigatus was found to be healthy where conidiophore holding vesicle with healthy metulae and phialides harboring conidia were noticed (Fig. 7A), whereas simulated microgravity affected the phialides structure (Fig. 7C) and subsequently conidia-forming cells were not clearly visible. When K. pneumoniae was co-cultured with the fungus, under normal gravity conditions, the bacterial cells (artificially colored in red) were shown to destroy the conidia-forming cells and also partially degraded metulae and phialides (Fig. 7B).
Furthermore, when K. pneumoniae cells were grown with the fungus and incubated in simulated microgravity conditions, bacterial cells destroyed the fungal structure where only vesicle was seen along with the conidia and phialides, additionally the metulae were disformed and the hyphae (Fig. 7D) appeared to be disfigured compared to those seen in Fig. 7A or B. When co-cultured, the hyphae of A. fumigatus appeared to be healthy when grown under normal gravity (Fig. 7B) and mostly dead when exposed to the simulated microgravity (Fig. 7D). The number of conidia were also less and on top of the conidia-forming cells the fungal morphological architecture changed. These changes were more obvious in microcosm where both K. pneumoniae and A. fumigatus were co-cultured and exposed to the simulated microgravity (Fig. 7D) than in the setup that were not exposed to simulated microgravity (Fig. 7B). Interpretation of image data was based on careful analysis of multiple images (n = 100), not only on a single micrograph of each test condition.
Additional SEM images (n = 4 to 6) per condition are provided in Supplementary Figure S5A to S5D. The mixed nature of the microorganisms (bacteria/fungi), the high magnification necessary to resolve the smallersized bacterium (creating narrow field of view) and large morphology (mycelium, and conidiophore/conidia) of the fungus combined to prevent adequate SEM observations that would be necessary to reasonably quantify the cells. Thus, statistical analysis was not possible using the SEM approach. This visualization approach demonstrates that K. pneumoniae in combination with simulated microgravity negatively influenced the growth of the fungi. Both these bacterial and fungal strains were isolated from ISS and potential detrimental effect of K. pneumoniae as predicted by the metabolic model during this study was experimentally demonstrated, confirming the degeneration of A. fumigatus morphological architecture. In addition to these morphological modifications, K. pneumoniae also reduced the biofilm forming capability of the fungus when grown together (Supplementary Figure S6).

Discussion
In this study, we used a metabolic perspective to analyze the ISS microbiome with a focus on K. pneumoniae, a BSL-2 human pathogen, and its coexisting microorganisms. In our two-pronged approach, we first used a graph-theoretical approach to unravel the metabolic interdependencies in the microbiome, and we further complemented this with constraint-based analyses, to identify the consequences of microbial interactions on growth. We begin by looking at higher-order interactions in the community and further delve deeper and look at pairwise interactions that have been previously shown to be major drivers of community dynamics [41]. We traverse three taxonomic levels-family, genus and species and provide insights into the keystone species of the community through metabolic network analyses.
The key results of this study are four-fold. First, among the families considered, species belonging to the Enterobacteriaceae family were often found to be the most beneficial, among the ISS microbiome. Secondly, Pantoea species were predicted to be extensively dependent on their coexisting microorganisms. Third, through metabolic network analysis of microorganisms taken pairwise, K. pneumoniae was found to be beneficial to many of its coexisting microorganisms, especially to those species belonging to the Pantoea. Fourth, metabolic interactions in the community broadly fell under the categories of amensalism and parasitism. The parasitic interaction under normal and simulated microgravity between K. pneumoniae and A. fumigatus was experimentally checked.
Even though <7% of the metagenomic reads of ISS environmental surfaces constituted BSL-2 microorganisms, the dominant and persistent human microbial pathogens were Pantoea, Klebsiella, Staphylococcus, Erwinia and Penicillium [15]. Among the ISS surfaces, Zero-G Stowage Rack (Location #5) had more Pantoea reads compared to other locations of all flights. When metagenomic reads of all BSL-2 microorganisms were compiled, K. pneumoniae was found to be persistent and dominant in Zero-G Stowage Rack of all three flights (Fig. 2). In general, K. pneumoniae reads were more and dominant in all seven locations sampled in Flight 3. Unlike in Flight 3, K. pneumoniae reads were retrieved sparingly in Flight 1 (Cupola and WHC) and Flight 2 (Zero-G Stowage Rack). The cleaning reagent used in ISS consists of benzalkonium chloride, which might be able to eradicate the fungal population; thus fungal succession was not observed, unlike bacterial members which were persistent in the ISS locations examined [42]. It is also documented that K. pneumoniae cells were resistant to the quaternary ammonium compound concentration used as cleaning agents in ISS [43]. On applying graph-theoretical algorithms and surveying the microbiome at multiple levels, we predicted that among the families in the dataset, members of the Enterobacteriaceae family were perhaps the most beneficial to other microorganisms in the community, whereas in most cases, members of the families such as Erwiniaceae, Staphylococcaceae and Sporidiobolaceae (yeast) offered little or no additional metabolic benefit to the community. It has been documented that some Rhodotorula species produced silver nanoparticles containing antimicrobial properties against a wide variety of Gram-positive and negative microorganisms [44]. The members of the family Sporidiobolaceae (Rhodotorula, pink yeast) were isolated from the ISS locations [8], their antagonistic and parasitic behavior predicted during this study should be tested against opportunistic microbial pathogens to aid in the development of appropriate countermeasures. Similarly, the amensalism predicted towards Erwiniaceae and mutualism of Staphylococcaceae members with K. pneumoniae need further study utilizing the ISS strains.
At the genus level, we observed that Pantoea genus typically derives the most benefit in the community. On looking at the pairwise interactions, we observed that amongst other Enterobacteriaceae, K. pneumoniae provides metabolic support to many of its coexisting species, notably those from the Pantoea genus. The members of the genus Pantoea were primarily considered as plant pathogens, but subsequently, they have been isolated from many aquatic and terrestrial environments, including ISS [8] as well as in association with insects, animals and humans [45][46][47][48][49]. During this study [8], the metagenomic sequences of the opportunistic human pathogens associated with Pantoea genus were P. agglomerans, P. conspicua, P. brenneri, P. ananatis and P. dispersa, whereas the plant pathogenic Pantoea species were not observed. The competitive metabolomic properties predicted during this study by K. pneumoniae with Pantoea members (Fig. 6A) might be due to the assimilation of similar compounds for sustaining their growth. Detailed phenotypic metabolic profiles of these microbes are needed to confirm the competition and amensalism predicted in this study.
Finally, from the constraint-based analyses, many amensal and parasitic interactions were noticed. This is corroborated by earlier findings that K. pneumoniae can inhibit fungal conidia germination and hyphal growth, as well as biofilm formation of Aspergillus species [50]. Recent reports documented that K. pneumoniae shifted to a pathogenic state, potentially leading to septic infections when cooccurred with A. fumigatus in immunocompromised individuals [50]. In addition, this study shows that K. pneumoniae cells prevented germination of conidia and hyphal development of the fungi. It has been further shown that K. pneumoniae bacterial cells induced the fungal cell wall stress response mechanisms and suppressed the filamentous growth of fungi [50]. The simple in silico metabolic model of this study predicted the antagonistic (parasitic) metabolic interaction between K. pneumoniae and A. niger (Fig. 6B), which further enabled to validate parasitism in vivo, using the strains isolated from the ISS. However, this experimental evidence might not be enough to fully prove the metabolic prediction. Additional positive and negative controls are necessary, which would further improve the evidence of antagonistic metabolic interaction predicted during this study. Furthermore, quantitative measurements of K. pneumoniae and Aspergillus growth are needed to validate the computational prediction of this study and warrants further study.
Our study does have some limitations. First, the metabolic models used in the study are automated reconstructions, which despite gap-filling, could potentially contain gaps and blocked reactions and need to be curated [51]. Nevertheless, such automated reconstructions are being widely used in many studies as they serve as useful predictive tools and representations of the metabolism of microorganisms [52]. Secondly, wherever it is applicable, genomes generated from the cultured ISS microorganisms were utilized. Conversely, when strains were not isolated from the ISS environment, and thus, genomes were not available, the reference genomes of the corresponding type strains were used. However, reference genomes provide a useful perspective of an organism's metabolic capabilities and have also been used in many other studies that incorporate metabolic models [53,54]. In the absence of exact estimates of environmental metabolites, we posited that the environment had at least the minimal set of metabolites necessary for a given organism to survive, and also included additional cofactors and coenzymes. We assumed that each microorganism is capable of minimal independent growth and thus calculated the minimal medium required to support survival. It is possible that some of the interactions remain conserved, in richer media, although others may disappear, depending on the inherent metabolic capabilities of the organisms.
The graph-based approach provides a somewhat static snapshot of the metabolic interactions happening between the organisms. Indeed, the graph-theoretic approach does not account for differential importance of reactions, or for example, the increase in growth rate, that may result from potential interactions. This is captured by the complementary constraint-based modeling approaches that were employed, and together, we believe, they provide a more complete picture of the possible interactions between the microbes. As previously highlighted in the 'Methods' section as well, there have been previous reports wherein a higher MSI does point towards a possible higher biomass in the community. As reported for Candida albicans, in addition to the genome-scale metabolic model, refining the model using phenotypic microarray and other wet-lab confirmation are needed [55]. Yet, it is highly useful and offers valuable insights into the potential microbial interactions in the community [30,56] and well complements the constraint-based methods. Of course, it is important to emphasize that our computational methodology studies microbes through a metabolic lens only, and in reality, other interactions may be plausible. Nevertheless, metabolic interactions have been previously shown [21] to be important drivers of community structure and microbial interactions. Importantly, this study offers a first glimpse into the metabolic interactions of the ISS microbiome, upon which several hypotheses can be formulated for future experimental design.

Conclusion
In summary, our analyses show the key role played by Klebsiella and other Enterobacteriaceae in mediating the metabolic interactions taking place between microorganisms in the ISS. Metabolic modeling, through a combination of graph-theoretic approaches and steadystate constraint-based modeling, paints a more comprehensive picture of possible microbial interactions, which are as yet inscrutable to this extent by experimental approaches. Our results point towards key dependencies of microorganisms in various locations on the ISS and can pave the way for possible interventions that may rely on targeted disinfection of surfaces aboard the ISS. Our approach also underscores the importance of complementary modeling approaches in dissecting a fairly complex microbiome and understanding various possible interactions. Our methodology is fairly generic and can be readily extended to predict microbial interactions in other interesting milieu and generate testable hypotheses for wet lab experiments, as we have demonstrated in this study with Klebsiella and Aspergillus species.