Costless metabolic secretions as drivers of interspecies interactions in microbial ecosystems

Metabolic exchange mediates interactions among microbes, helping explain diversity in microbial communities. As these interactions often involve a fitness cost, it is unclear how stable cooperation can emerge. Here we use genome-scale metabolic models to investigate whether the release of “costless” metabolites (i.e. those that cause no fitness cost to the producer), can be a prominent driver of intermicrobial interactions. By performing over 2 million pairwise growth simulations of 24 species in a combinatorial assortment of environments, we identify a large space of metabolites that can be secreted without cost, thus generating ample cross-feeding opportunities. In addition to providing an atlas of putative interactions, we show that anoxic conditions can promote mutualisms by providing more opportunities for exchange of costless metabolites, resulting in an overrepresentation of stable ecological network motifs. These results may help identify interaction patterns in natural communities and inform the design of synthetic microbial consortia.


INTRODUCTION 23
The astonishing number of microbial species observed in nature 1-3 seems to contradict 24 classical ecological theory, which predicts far less biodiversity in many nutrient-poor 25 environments 4,5 . A variety of different explanations have been proposed as possible 26 solutions to this inconsistency, including resource partitioning 6 , differential nutrient use 7 , 27 spatial segregation 8 , and metabolic cross-feeding 9-11 . In environments poor in 28 resources, cross-feeding has been shown to enhance the capacity of microbes to survive, 29 either through the secretion of valuable compounds 12-14 , or by maintaining 30 thermodynamic gradients necessary for continued metabolism 15 . Despite their 31 prevalence, it is not clear how these cooperative phenotypes emerge, as they often 32 involve the exchange of metabolites that are costly for the producer. This apparent 33 altruism introduces the potential for the rise of cheating organisms that do not contribute 34 common goods but still benefit metabolically from others, challenging community stability 35 community benefits as a product of otherwise selfish acts by individual microbial species. 48 This phenomenon has been explored in a macroecological context 21-23 and can be 49 illustrated by the example of a vulture consuming the remains of a lion kill. Here, the lion 50 gains nutritional benefit from its hunt and leaves behind scraps of food that are in turn 51 eaten by the vulture. In this way, though the lion did not expend energy to facilitate access 52 to food explicitly for the vulture, it did unintentionally contribute to the vulture's success 53 through its own selfish action 24 . It is known that, in the microbial world, metabolic waste 54 products secreted at no cost to the producing organism (e.g. E. coli secreting acetate 55 under limited oxygen) can serve to support other species 13 . However, it is not obvious 56 whether such behavior extends beyond a few fermentation byproducts. Moreover, little 57 information exists on how costless secretions vary across microbial species and growth 58 media composition. Most importantly, even if the metabolites secreted by an organism 59 under a given condition were to be known, it still would be difficult to ascertain whether 60 such byproducts would be likely to enable or enhance growth of other species. 61

62
In this study, we use computational metabolic modeling to quantify the magnitude of 63 environmental modification brought about by costless metabolite secretion, as well as the 64 interspecies interactions that can arise from this type of exchange. In a microbial analog 65 to the lion-vulture interaction, we seek to understand how metabolites released as a 66 product of selfish action by individual species yield unintended benefits to partner 67 organisms, resulting in emergent interspecies cooperation. Based on this framework, we 68 present a computational pipeline based on flux balance analysis (FBA) 25 that predicts 69 the growth phenotypes and cooperative interactions mediated by costless metabolites for 70 14 microbial species under a large combinatorial set of environmental conditions. In this 71 way, we obtain a global view of cross-feeding opportunities that can mediate the 72 emergence of cooperation and the maintenance of biodiversity in natural communities. In 73 addition, we complement our metabolic modeling with a dynamical modeling framework 74 to understand whether costless secretions on their own can promote long-term stability 75 in model synthetic microbial communities. While the present work focuses entirely on 76 putative secretions and interactions predicted computationally, we wish to highlight that 77 we restricted our analysis to microbes associated with high quality, manually curated (and 78 therefore in most cases individually tested) in silico models and that in many cases, 79 specific predictions can be shown to be consistent with previously established empirical 80 knowledge. For the most part, however, the current analysis should be viewed as the 81 exploration of a large space of stoichiometrically possible costless interactions 82 (inscrutable to such an extent at the experimental level), whose global patterns could 83 motivate and inform future experimental and theoretical endeavors. 84

RESULTS 86
Metabolite secretions can be costly or costless, depending on environmental 87 context. Understanding whether or not the secretion of a specific metabolite by a given 88 organism is associated with a decrease in fitness (interpreted here as growth rate) is 89 difficult to assess experimentally, but can be readily addressed using genome-scale 90 models of metabolism (see Methods). For example, it is possible to impose the secretion 91 of a given compound at a given rate, and then ask whether this constraint is expected to 92 cause a reduction in growth. A small set of simulations of this kind for a single organism 93 ( Fig. S1) exemplifies the broad spectrum of possible outcomes: based on the specific 94 carbon sources, different metabolites can be produced, sometimes at the expense of 95 growth capacity, other times with no apparent effect (neutral), or even to its benefit. Due 96 to the basic assumptions of the genome-scale models we employed (especially the 97 maximization of growth as the objective function) we know that these last two kinds of 98 secretions are compatible (or even necessary) for metabolism to operate at maximal 99 Simplified schematic of an in silico experiment: A growth medium ( + ) containing two carbon sources ( , ) with or without oxygen (Ω) is provided to genome-scale metabolic models of two microbial organisms ( , ). If at least one organism grows, any costlessly-secreted metabolites ( + ) are added to the medium, which is fed back to the organisms. This process is repeated for a series of iterations , and terminates at iteration ) , defined as the last iteration in which any new metabolites were secreted into the medium. explained by the alternative hypothesis that organisms do secrete multiple byproducts in 125 the first iteration, but these byproducts contribute weakly to additional secretions in 126 subsequent iterations. 127

128
In aggregate, our simulations showed a rightward shift in the diversity of metabolites 129 secreted under anoxic conditions when compared to the number secreted when oxygen 130 was present. To understand this effect, we looked at the distribution of the number of 131 metabolites secreted after the first iteration, which is equivalent to growing each organism 132 on its own in the provided medium. This distribution for = 1 was unimodal for both 133 conditions, centered between two and three metabolites with oxygen and around five 134 metabolites without oxygen ( Figure 2b). After this first iteration, the maximum number of 135 secreted metabolites was 11 with oxygen and 16 without oxygen. In the anoxic 136 simulations, the central carbon metabolites most commonly secreted after the first 137 iteration were fermentation byproducts such as acetate, formate, succinate, and ethanol. 138 These metabolites were secreted in 87.5%, 74.5%, 25.7%, and 20.2% of growth-yielding 139 simulations respectively. With oxygen, the most commonly secreted central carbon 140 metabolites after the first iteration were formate and acetate, secreted in 46.8% and 141 18.3% of growth-yielding simulations respectively. We may therefore chiefly attribute the 142 shift between the oxic and anoxic secretion curves to the anoxic export of incompletely-143 reduced core metabolism intermediates. 144 . The last iteration is defined as the iteration in which no additional metabolites were secreted into the medium. Despite the large variability in number of expansions and number of secreted metabolites, we observe a poor correlation between these distributions, indicating that a simulation resulting in a high number of expansions does not necessarily result in a high number of metabolites being secreted ( Figure S3).
In addition to a positive shift observed between anoxic and oxic conditions, our results 145 also show a shift in the quantity of metabolites secreted between the first and last iteration 146 of each computational experiment (Figure 2c). This effect reflects organisms taking up 147 metabolites secreted by themselves or their partner, and secreting different metabolites 148 as a response. After the last medium expansion iteration for all simulations, the total 149 number of secreted metabolites followed similar distributions with a maximum at 18 and 150 21 metabolites for oxic and anoxic conditions, respectively. This positive shift suggests a 151 response from one or both organisms to a medium iteratively enriched by costless 152 byproducts, which hints at their potential metabolic utility. Principal component analysis 153 (PCA) shows that neither the environment nor the species alone can explain the variability 154 in secretion profiles ( Figure S4), suggesting that a combination of both variables accounts 155 for the range in costlessly-secreted products. 156 secreted compounds across all simulations, nitrogen-containing compounds such as 162 nitrite, ammonium, urea, and trimethylglycine were secreted in 73.5% of the analyzed 163 cases, suggesting maintenance of an appropriate carbon-to-nitrogen ratio in the cell. We 164 note specifically that nitrite is secreted in fewer than 100 simulations with oxygen, but 165 almost universally in anoxic simulations -a phenomenon previously observed in 166 anaerobic enteric bacteria 26 . Organic acids make up the second most abundant category 167 of costlessly-secreted byproducts, constituting 23% and 36% of unique metabolites with 168 and without oxygen respectively. Notably, we also observe secretion of nucleotides, 169 peptides, and carbohydrates in a combined 9% and 13% of simulations with and without 170 oxygen respectively. Altogether, this space of secreted metabolites points to a large 171 variety of molecules that can be freely produced, suggesting that costless metabolic 172 secretion may provide substantial degrees of environmental enrichment. This effect 173 becomes magnified considering the relative scarcity of resources provided in our minimal 174 medium, which suggests that costless secretions play a fundamental role in promoting 175 metabolic diversity in natural environments. 176

177
Given the abundance and complexity of secretions from different organisms, as well as 178 the possible ecological connections they may promote, we asked whether specific 179 metabolite secretions were highly correlated. As patterns in environmental modification 180 through secretion have an impact on the species composition of a microbial community 181 27 , it becomes important to understand which metabolites co-occur within our set of 182 simulations. To address this question, we performed a Spearman correlation analysis to 183 determine common secretion patterns ( Figure S6). In the presence of oxygen, we observe 184 a strong co-occurrence of glycerol, lactate, succinate, malate, and acetate, which 185 correlates with the high frequency of secretion of these carbon-containing compounds 186 ( Figure S5a). We also observe positive, but weaker correlations between these 187 metabolites and other central carbon compounds such as fumarate, citrate, and 2-188 oxoglutarate. Our analysis also points to the simultaneous release of multiple nitrogen-189 containing compounds, chiefly urea, ammonium, and nitrate. Without oxygen, we observe 190 We note that while a potentially useful metabolite can be secreted into the environment 202 by one species, it does not necessarily mean that it will be consumed by a second 203 organism. We place particular importance on this distinction, as any interspecies 204 interaction must also take into account the decision to import a novel metabolite found in 205 the environment. To map this distinction, we examine the space of costless metabolites 206 that are exchanged by each organism across all in silico experiments ( Figure S5b). Here, 207 the most commonly exchanged organic metabolites were central carbon intermediates, 208 secreted mostly in anoxic conditions. These secretion patterns mirror those of anoxic gut 209 bacteria, which divide the task of digesting complex polysaccharides by exchanging 210 intermediate organic acids 9,30 . Importantly, we observed that amino acids, secreted 211 chiefly by S. cerevisiae, but also in a substantial number of simulations by S. enterica, K. 212 pneumoniae, and E. coli, were among the most highly-exchanged costless metabolites. 213 This phenomenon has been previously documented in relation to overflow metabolism in 214 S. cerevisiae 31 and E. coli 32,33 , as well as in yeast-bacteria symbioses 34,35 , and account 215 for exchange in over 10 4 simulations with and without oxygen in our study. This high 216 prevalence of exchange underscores the metabolic utility of these secreted byproducts, 217 particularly when contrasted with patterns of secretion in which the most commonly 218 released metabolites were of low or no metabolic utility to a partner organism (e.g. water).  necessitating dependence on a rich set of metabolic products produced by the host or 242 other commensal microbes. In our simulations, however, these organisms failed to grow 243 in all environments and with all species pairs even after any costless metabolites were 244 secreted. This failure to sustain growth of highly dependent organisms suggests that there 245 is an upper limit to the degree to which costless metabolite production can enable species 246 growth, especially in the minimal environments that were tested. Aside from these 247 extreme cases, our analysis sheds light on the performance of generalist organisms, such 248 as E. coli, K. pneumoniae, S. cerevisiae, and S. enterica. These organisms grew in at 249 least half of all tested environmental conditions, in contrast with organisms such as M. 250 extorquens or Z. mobilis, which exhibited much more limited pairwise growth capabilities. 251 These patterns suggest a greater dependence of these organisms on the metabolic 252 byproducts of their partners, particularly in anoxic conditions. These patterns underscore 253 the importance of not only the number of metabolites secreted, but also of the specific 254 metabolic needs of the receiving organism in determining the contribution of costless 255 metabolites to the growth of a partner. This metric therefore aims to reflect the cooperative potential of each carbon source pair 295 relative to that of each carbon source in isolation. In this way, when averaging a single 296 carbon source over its cooperativity index, we obtain a relative degree to which a carbon 297 source "depends" on another to sustain growth. By framing cooperativity in this context, 298 we observed that simple sugars such as glucose and sucrose had relatively low 299 cooperativity indices, that is, they were able to sustain growth efficiently on their own. In 300 contrast, more complex molecules and dipeptides had higher average cooperativity 301 indices, indicating they performed better in the presence of another carbon source. We 302 grouped these average cooperativity indices through hierarchical clustering ( Figure S7) 303 and observed general clustering by carbon source type -especially with sugars and 304 amino acids appearing in distinct groups. This analysis illustrates the nonlinear effects of 305 adding additional nutrients to a minimal medium, underscoring the observed complex 306 metabolite usage patterns in organism pairs. 307 308 Organisms competing for the same carbon source can simultaneously benefit each 309 other through costless secretions. Our analysis so far has examined the contexts in 310 which a metabolite can be secreted costlessly, as well as the potential for these 311 metabolites to promote growth. Based on these insights, we wished to more 312 fundamentally understand these interspecies interactions and how they compare to 313 ecological expectations of cooperation and competition. To do this, we defined six types 314 of possible interactions: non-interaction, commensalism (unidirectional exchange), and 315 mutualism (bidirectional exchange), each with or without competition for a primary carbon 316 source. We chose to decouple competition for nutrients from cooperation via secreted 317 metabolites in order to more fully understand the degree to which the latter can promote 318 organism coexistence despite resource scarcity (Figure 5a). When analyzing our dataset 319 under this framework, we found that competition for one or both carbon sources 320 showed greater amounts of mutualistic interactions without oxygen. When E. coli was 342 grown anaerobically but its partner was grown with oxygen, the vast majority of 343  In order to simulate how these interactions could contribute to stable symbioses, we 369 created a dynamical chemostat model of two arbitrary species consuming carbon sources 370 and exchanging costless metabolites according to each motif type (see Methods). By 371 varying the specific growth rates of each species from 0 to 1 hr -1 , we simulated the growth 372 of the pair under each motif type for 500 hours. If both species were still present at the 373 end of the simulation, we determined the motif type to enable stability at that combination 374 of specific growth rates. We mapped the space of stable species pairs under each motif 375 type, observing that competitive interactions generally have a reduced parameter space 376 for enabling stability (Figure 7c). Notably, though motif N1b was highly prevalent in the 377 costless FBA simulation set, this motif represents classic competitive exclusion and 378 cannot result in long-term stability. In contrast, though complete nutrient-organism 379 orthogonality can yield stability over the whole space of parameters (N2a), this motif was 380 not predicted to occur in the mechanistic simulations. An intermediate case between 381 these two extremes (N2b) is the one in which there is a balance between competition and 382 independence with respect to external carbon source utilization: in this case, which 383 frequently occurs in our dataset, stability is achievable only for a narrow set of parameters. 384 385 A marked increase in stability is predicted when costless metabolite exchange is enabled 386 (commensalism and mutualism). For motif C2b, for example, both organisms are 387 competing for a carbon source and organism 1 is providing one or more costless 388 metabolites to organism 2. Our dynamical modeling showed that the growth rate of 389 organism 1 must be greater than that of organism 2 in order for both species to be stable. 390 When feedback was allowed to occur (mutualism), the potential for stability vastly 391 increases across our parameter space. M2a and M2b even allowed for very low specific 392 growth rates for both organisms, indicating a strong dependence on costless metabolites 393 for long-term coexistence. 394

DISCUSSION 396
We have investigated the pairwise growth phenotypes and interactions of 14 diverse 397 microbial species in over 10 6 computational experiments. We found that resource-poor 398 environments provide the basis for release of a wide variety of useful metabolic products 399 secreted without cost by their producing organism; these costless metabolic products 400 provide, in an oxygen-dependent manner, valuable environmental enrichment, nearly 401 doubling the potential of minimal environments to sustain growth. We further found that 402 exchange of costless metabolites establishes beneficial uni-and bidirectional 403 interspecies interactions, associated with different chance of stability of the ensuing 404 consortia. Overall, both the metabolic capabilities of the organisms and the environmental 405 contexts in which they are grown (particularly oxygen availability) determine which 406 metabolites will be secreted without cost and how these secretions will contribute to 407 interspecies interactions. We nonetheless emphasize that even in the simple, minimal environments we studied, 424 our modeling framework, based on fundamental stoichiometric constraints and metabolic 425 efficiency assumptions, predicts the widespread prevalence of molecular products that 426 are secreted without a metabolic burden and that can benefit other organisms. An 427 important implication of this prediction is that costless metabolites may significantly 428 contribute to enriching environments and sustaining biodiversity, even when organisms 429 are competing for the same primary nutrients. By using costless secretions to cooperate 430 while simultaneously competing for primary nutrients, organisms may escape some of the 431 limitations of pure competition, which has been predicted to limit biodiversity 45  conditions. The process of generating a genome-scale metabolic model has been 488 outlined conceptually 66-69 and described procedurally 70 by various groups, and generally 489 comprises an automatic generation of a model based on pathway and genome data 490 followed by manual curation by integrating phenotyping, metabolomic, or transcriptomic 491 data 71 . We note that although an automatically-generated draft metabolic model can be 492 constructed for virtually any organism for which a genome annotation exists, the space of 493 high-quality, experimentally-verified metabolic models that have undergone the manual 494 curation process summarized above is comparatively very small 72 . This is due to the time 495 and resources needed to complete the curation process, which can span from six months 496 70 to more than ten years for the iteratively-refined model of E. coli K-12 64 . We 497 nonetheless consider this process to be essential in producing models that can generate 498 the mechanistic cross-feeding predictions detailed here, which rely on verified metabolic 499 capabilities in monoculture. 500

501
The models used in this analysis span four taxonomic kingdoms, including 502 representatives from eight bacterial taxa, as well as a variety of primary metabolic 503 strategies ( Supplementary Information 1). In addition, these models describe several 504 organisms that are commonly used for in vivo studies (E. coli K-12, S. enterica LT2, etc.), 505 making the resulting costless cross-feeding predictions particularly useful for synthetic 506 ecology experiments and microbial community assembly. 507 508 Each model was imported into MATLAB (The MathWorks, Inc., Natick, Massachusetts) 509 using the COnstraint-Based Reconstruction and Analysis (COBRA) Toolbox 73 , a 510 software platform for constraint-based modeling of metabolic networks. In order to enable 511 in silico cross-feeding to be correctly classified, the namespace of all of the metabolic 512 compounds in each of the models was standardized to be internally consistent. This was 513 performed via a computational pipeline with additional manual curation for irregularly-514 annotated metabolites. 515 516 Computational methodology description and inputs. Our computational method 517 comprises a set of programs written in MATLAB that use Flux Balance Analysis (FBA) to 518 mechanistically define the growth status and metabolic exchange of microbes through 519 costlessly-secreted byproducts. Briefly, FBA is a mathematical method that determines 520 an optimal distribution of metabolic flux through a biochemical network that will maximize 521 a given objective, usually biomass 74 . An FBA problem is framed in the context of several 522 constraints, namely: (i) , the stoichiometric matrix of dimensions × where is the 523 number of metabolites and is the number of reactions in the model; (ii) , the vector of 524 all reaction fluxes; and (iii) ABC and ADE , flux constraints placed on , defined by 525 enzymatic capacity and experimentally measured uptake rates. 526

527
We employ FBA to determine if an organism is able to grow on the in silico growth media 528 conditions we define, in addition to which metabolites are taken up and costlessly 529 secreted. We first apply FBA by maximizing for growth and obtaining an optimal growth 530 rate for an organism, FGHIJK (ADE) . To determine which metabolites are secreted costlessly, 531 we set this growth rate as a minimum for the biomass flux and apply FBA again, recording 532 any metabolites that were secreted. We also apply the additional constraint of minimizing 533 all reaction fluxes across the network to more closely simulate efficient use of the optimizations separate for each model without accounting for spatial or temporal 551 community structure. It is also for this reason that we establish the biomass fluxes of each 552 in silico organism as the objective functions to be optimized, as we are concerned with 553 secretion of potentially useful metabolic byproducts that arise out of "selfish" optimal 554 growth. This assumption of maximum growth with proteome optimality is also key for 555 translating these organisms and predictions to in vivo synthetic ecologies, where biomass 556 optimization more closely describes the behavior of organisms in batch or continuous 557 culture 76 . 558 559 Our algorithm requires six inputs: 1: a data structure containing the genome-scale 560 metabolic models to be used, 2: a list of carbon sources, 3: the number V of in silico 561 organisms to be simulated together (for pairwise simulations V = 2), 4: the number W/ 562 of carbon sources to be provided to each simulation, 5: a Boolean variable Ω = {1,0} that 563 specifies if oxygen will be made available to the in silico organisms, and 6: a list of 564 metabolites that makes up a simulated base growth medium, ABC . This base medium 565 contains various nitrogen, sulfur, and phosphorus sources, as well as vitamins, ions, and 566 metals needed for growth of the organisms ( Supplementary Information 3). 567

568
We focused on pairwise species growth with two carbon sources ( V , W/ = 2). Although 569 each genome-scale metabolic model we used has been manually curated to reflect in 570 vivo metabolic capabilities, very few experiments have been performed to verify FBA-571 generated predictions for more than a single species 77,78 . We therefore limit the number 572 of in silico species to two, in order to interpret the growth and cross-feeding predictions 573 with greater confidence. This limit also constrains the combinatorial space of the 574 simulations, which grows exponentially and becomes numerically intractable with more 575 models and carbon sources. In addition, limiting simulations to V = 2 allows for greater 576 experimental accessibility for assembling synthetic ecologies based on costless 577 metabolite exchange. Our algorithm can nonetheless be applied to any { V , W/ > 0}. 578

579
The list of all possible carbon sources was defined primarily from the carbon sources 580 contained in the BIOLOG Phenotyping MicroArray 1 (PM1) plate, which is used for 581 phenotyping and curation of genome-scale metabolic models 79-81 . The carbon sources 582 we selected are common mono-di-and polysaccharides, all 20 amino acids, dipeptides, 583 and organic acids contained in the PM1 plate. We also supplemented the list with 584 additional carbon sources known to be consumed by the in silico organisms, for a total of 585 108 ( Supplementary Information 2). 586

587
To permit uptake of the metabolites in the medium, the constraint on the uptake flux bound 588 ADE for each exchange reaction pertaining to a medium metabolite was removed in each 589 of the models and . This bound was fully removed ( ADE = 1000 * ℎ ⁄ ) for 590 non-limiting medium components, and was set to ADE = 10 * ℎ ⁄ for the 591 growth-limiting carbon sources and . This latter value is drawn from experimentally-592 estimated uptake rates of sugars by E. coli in exponential growth conditions 64 , and is 593 applied equally to all other species to simulate general availability of the carbon sources 594 in the environment. All other exchange reaction ADE values are set to zero to block 595 uptake of metabolites not in the medium. 596 597 Computing growth, secretion, and cross-feeding. We describe the FBA operations at 598 the core of our algorithm as a function that, given a medium condition and organisms 599 and , outputs the binary growth status of the organisms, as well as the set of to simulate the long-term stability of each pairwise interaction type observed in our in 639 silico experiments. We first established a graph theory framework to map each simulation 640 to a specific interaction motif, each of which accounted for the general interaction type 641 (non-interacting, commensal, or mutualistic), the number of carbon sources consumed by 642 the pair, and the competition status for the carbon sources ("a" denotes no competition, 643 "b" denotes competition) (Figure 5a). We next applied a differential equation-based 644 growth model to each specific motif. Since motifs with two carbon sources can be 645 represented by more than one motif topology, we selected one representative topology 646 from these motifs to simplify the space of dynamical modeling simulations. These the population e to reach half of its maximum growth rate in g/L. 675

676
We then combine equations E2-4 to fit the particular motif being modeled ( Figure S8). 677 The values of the parameter values are described in Supplementary Information 4 and  678 are based on values reported by Smith 83 , Balagaddé et al. 84 , and those based on 679 reasonable estimates for resource consumption. For each motif, we vary the specific 680 growth rate of both organisms from 0 to 1 hr -1 and run the simulation for 500 hours. If 681 both organism abundances are above 0.05 g/L at the end of the simulation, we 682 determine the motif to be stable at the prescribed growth rates. Increasing the secretion flux of a 'costly' product, such as succinate, imposes a reduction in growth rate when glucose and glycerol are supplied as carbon sources. When the carbon sources are replaced with citrate and trehalose, succinate is secreted without a cost to growth rate. (b) With glucose and glycerol as carbon sources, E. coli is predicted to have a wide range of fluxes at which formate can be secreted without a cost to its growth rate. Formate would, according to our definition, be secreted 'costlessly' by E. coli under the applied environmental conditions. (c) Some costlessly-secreted metabolites must be secreted at a given rate in order to maximize growth. If an upper bound is placed on acetate secretion, E. coli must allocate resources away from biomass in order to cope with its limited ability to secrete fermentation byproducts. Acetate would therefore also be considered a costlessly-secreted metabolite by our definition. Figure S2. Detailed example of single in silico experiment, illustrating three phases. Initialization: A minimal medium ABC common to all simulated conditions (composed of salts, metals, vitamins, as well as nitrogen, phosphorous, and sulphur sources) is defined prior to execution of the pipeline. This medium is supplemented with two carbon sources, and .The Boolean variable Ω = {0,1} defines whether or not oxygen is present in the environment. Here, Ω = 1. These together define the initial medium set, d . Expansion: The function is applied to genome-scale metabolic models of two organisms ( , ) in a series of iterations, . In each iteration, simulates the growth of both organisms in the current medium condition and returns the Boolean growth statuses + = { B , e } of both organisms and the set of any costlessly-secreted metabolites, + . Here, in the first iteration, h = {1,0} since organism grew but organism did not. Since at least one organism in the pair grew, the medium is updated ( +gh = + + + ) and is applied again until no new metabolites are secreted. Completion: When no new metabolites are added to the medium, the experiment is complete. The last iteration with any new secreted metabolites is defined as / . Figure S3. Correlation between total number of metabolites secreted costlessly and the number of expansions in each in silico experiment for (a) oxic and (b) anoxic conditions. We observe a poor correlation between number of secreted metabolites and number of expansions in both oxic and anoxic simulations. This lack of correlation suggests a lower rate of metabolite exchange with increasing iterations, with most organisms quickly stabilizing their environment within one or two expansions. With oxygen, for example, only the K. pneumoniae and Synechocystis pair exhibited more than three medium expansions, with acetate, formate, citrate, and L-malate being the only metabolites secreted at these iterations. These scenarios accounted for only 40 simulations. Without oxygen, there were 697 experiments that reached more than three medium expansions, with 10 organisms being represented. However, this anaerobic set was dominated by the S. cerevisiae-P. aeruginosa pair, with fermentation byproducts being secreted at late iterations.