Microbes use transporters to regulate the release of metabolites based on value

Microbes release a wide range of metabolites into their environment, yet the reasons for this release and the factors that influence release rates are not well understood. Here, using a combinaUon of computaUonal and experimental approaches on seven different microbes we show that release rates are negaUvely correlated with the value of metabolites, and that this relaUonship explains more variability in release rates than other frequently associated factors. These findings are in accordance with the idea that metabolite release generally consUtutes a fitness cost that microbes have evolved to reduce. This new conceptual explanaUon for why microbes release metabolites has important implicaUons for how we think about the emergence of cross-feeding interacUons. Main Introduc)on Metabolites are frequently exchanged between microbes in natural communiUes (1). Such cross-feeding plays a key role in maintaining community diversity and funcUon (2–6). While certain occurrences of cross-feeding arise from extracellular metabolism such as polymer degradaUon by extracellular enzymes (6, 7), many observaUons of cross-feeding stem from the .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted August 19, 2025. ; https://doi.org/10.1101/2025.08.19.671024doi: bioRxiv preprint release of intracellular metabolites into the environment by microbial community members (8– 10). Recent in vitro quanUficaUons of microbial cultures have demonstrated that this release includes high-energy glycolyUc intermediates, amino acids and organic acids (11–14). The ubiquity of metabolite release may help explain the high frequency of auxotrophs isolated from natural environments that may be supported by co-culUvated partners (15–18). However, why so many diverse intracellular metabolites are being released and what explains differences in release rates remains unanswered. Previous efforts to answer why microbes release intracellular metabolites have revealed fundamental principles of metabolism, including rate-yield trade-offs and constraints on proteome allocaUon (19–23), mechanisms of pathway regulaUon and maintenance of homeostasis (24), opUmal catabolic pathway lengths in different environments (25, 26) and redox balancing (27, 28). These explanaUons are context-, metabolite- or species-specific. However, to fully understand metabolite release, we need to consider broader principles that apply across diverse microbial taxa and environments. Cell lysis and passive diffusion across the cell membrane are two such commonly proposed broad mechanisms (29–32). Here we move beyond these mechanisUc explanaUons and ask whether metabolite release generally confers a benefit or imposes a cost on the source organism. While recent work has explored how metabolite release can be beneficial, e.g. by alleviaUng costs associated with molecular noise (32, 33), we here propose and test an alternaUve hypothesis that metabolite release comes with a cost that microbes have evolved to reduce. Metabolite release rates are nega)vely correlated with metabolite value If metabolite release generally imposes a fitness cost, and natural selecUon acts to minimize such costs, then the release rates of specific metabolites should be negaUvely correlated with their relaUve importance, or value, to the microbe (Fig. 1A). Release of more valuable metabolites – that require more energy to produce – should have a more severe negaUve impact on fitness and be more strongly selected against. To quanUfy metabolite release rates across different species, we first analysed an exisUng Ume- series exometabolome dataset with high temporal resoluUon and absolute quanUficaUon of 19 to 37 metabolites collected from batch cultures of E. coli, B. licheniformis, S. cerevisiae and C. .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted August 19, 2025. ; https://doi.org/10.1101/2025.08.19.671024doi: bioRxiv preprint glutamicum in high glucose concentraUons (12). We esUmated net uptake and release rates for all metabolites using simple linear regression, modelling extracellular metabolite concentraUon as a linear funcUon of the area under the growth curve (AUC), unUl the point where the extracellular concentraUon saturated or the exponenUal growth phase ended (Methods, Fig. 1B, and Figs. S1-S4). The iniUal dynamics of 44-95% of metabolites were well described by this simple linear model (two-sided Wald test, FDR 0.5, Fig. S1-S4). To esUmate metabolite value, we used previously reconstructed enzyme-constrained genome- scale metabolic models (ecGEMs) (34–36). Genome-scale metabolic models (GEMs) are mathemaUcal representaUons of a species’ metabolic capacity that can for example be used to predict opUmal metabolic phenotypes across various environments (37, 38). The addiUonal enzyme constraints account for the total enzyme pool limitaUon in cells (20), enabling predicUon of relevant trade-offs and temporal dynamics, e.g. acetate overflow in E. coli (35). Various measures of metabolite value have been proposed in the past (39–41). Here we define metabolite value as the negaUve change in growth rate on metabolite release close to opUmal growth as predicted by enzyme-constrained GEMs (Fig. 1C). This measure explicitly reflects the fitness penalty on growth rate associated with the marginal release of a metabolite, and it can easily be used for different species across different environmental contexts. With these two measures we could now test our hypothesis that release rate should be negaUvely correlated with metabolite value. Indeed, this holds for E. coli, S. cerevisiae, B. licheniformis in glucose medium (Pearson ρ, P < 0.001, Fig. 1D). This did not hold for C. glutamicum (Fig. 1D), possibly because the engineering of its central carbon metabolism for enhanced lysine producUon has also affected the release of other metabolites (42). To test whether our hypothesis held beyond simple glucose environments, we conducted new well-controlled batch culUvaUons of E. coli in minimal media with galactose, L-malate or L- alanine as the single carbon source. These carbon sources were chosen to span the variability in E. coli metabolism to avoid biasing the results towards a parUcular class of compounds or metabolic flux pakerns (Methods, Fig. S5), since this has been shown to produce similar exometabolome pakerns (43). Triplicate batch bioreactor experiments were performed under strictly controlled aerobic condiUons with acUve regulaUon of pH and dissolved oxygen (Fig. S6). RelaUve extracellular concentraUons of 126 metabolites were then quanUfied at three exponenUal-phase samples and one sample from staUonary phase. Of these metabolites, 41- .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted August 19, 2025. ; https://doi.org/10.1101/2025.08.19.671024doi: bioRxiv preprint 56% were released at significant rates during exponenUal phase (two-sided Wald test, FDR < 0.05, Figs. S7-S10), consistent with previous findings showing that microbes generate rich exometabolomes (12, 13, 44). From this set, we were able to esUmate absolute release rates for 34 metabolites that we compared with esUmated metabolite values. Again, we found significant negaUve correlaUons for all three condiUons (Pearson ρ, P < 0.001, Fig. 1E). Finally, we analysed a third dataset that in addiUon to E. coli and Pseudomonas puGda contained environmental isolates from the genera Enterobacter and Pseudomonas (43). Unlike previous datasets, these experiments were conducted without pH and oxygen control, offering an opportunity to test the hypothesis under more variable condiUons. Despite these differences, we again found consistent negaUve correlaUons (Fig. 1F), suggesUng that the relaUonship between metabolite value and release rate is robust across species, nutrient condiUons, and experimental planorms. Importantly, the ecGEMs used here were reconstructed by different research groups using different pipelines, suggesUng that our results are not dependent on the choices made during model reconstrucUon. This conclusion was also supported by consistent

when metabolite values were esUmated using four other E. coli GEMs (Fig. S11). .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted August 19, 2025. ; https://doi.org/10.1101/2025.08.19.671024doi: bioRxiv preprint .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted August 19, 2025. ; https://doi.org/10.1101/2025.08.19.671024doi: bioRxiv preprint Figure 1: Correla'on between microbial metabolite release rates and metabolite values. A) Hypothesis: if metabolite release generally cons:tutes a fitness cost that microbes have evolved to reduce, metabolite release rates should be nega:vely correlated with metabolite values. Three metabolites are used to conceptually illustrate their rela:ve posi:on based on their es:mated release rate and metabolite value, see panels B and C. B) Metabolite release rates were es:mated from :me-series exometabolomics data from batch cultures by linear regression of extracellular concentra:ons vs area under the curve (AUC) of biomass. This panel shows extracellular concentra:ons of 37 metabolites measured during exponen:al phase of E. coli cul:vated in glucose medium by Paczia et al. (2012), and the linear regressions used to es:mate net release rates. Linear regressions with an R2 0.05 (two-sided Wald test) are drawn in red. Full :me-series, including sta:onary phase and linear regression for E. coli, B. sub:lis, C. glutamicum and S. cerevisiae are shown in Figs. S1-S4. Metabolite abbrevia:ons: α-KG: alpha- ketoglutarate, DHAP: dihydroxyacetone phosphate, FBP: fructose 1,6-bisphosphate, F6P: fructose 6-phosphate, G6P: glucose 6- phosphate, 2/3PG: 2/3-phosphoglycerate, GA3P: glyceraldehyde 3-phosphate, R5P: ribose 5-phosphate, PEP: phosphoenolpyruvate, RU5P/X5P: ribulose 5-phosphate/xylulose 5-phosphate, E4P: erythrose-4-phosphate. C) We used enzyme- constrained genome-scale metabolic models to es:mate metabolite values, quan:fied as the nega:ve change in growth rate (Δy) on metabolite release (Δx) close to op:mal growth. D-F) Three different datasets were used to test our hypothesis: D) Paczia et al. (2012), including S. cerevisiae, B. licheniformis, C. glutamicum and E. coli on glucose (12), E) E. coli on three other carbon sources (this work), and F) Vila et al. (2023) covering a smaller number of metabolites on E. coli, P . pu:da and two environmental strains of the genera Enterobacter and Pseudomonas on up to 8 different carbon sources (43). The scaeer plots show mean es:mated rates +- standard errors vs es:mated metabolite values. Black-filled circles are datapoints where the lower error bar is omieed because mean minus standard error is less than 0 and cannot be included on a log-scale. Axis labels and colour legend are shared for panel D-F. Annota:ons in panel F show the Pearson correla:on and associated P value in parentheses. Metabolite value is the most important factor explaining variability in metabolite release rates The consistent negaUve correlaUon between metabolite release rates and metabolite values encouraged us to ask how well this factor explains release variability compared to other metabolic or physiochemical properUes previously associated to metabolite release (33, 45). We focused on the data from bioreactor batch cultures of E. coli with glucose, galactose, L-malate or L-alanine as the carbon source, where condiUons were well-defined, the measured metabolites were diverse and reference values for intracellular metabolite concentraUons were available (46–49). We first built univariate linear models to quanUfy the importance of these factors: metabolite value, intracellular concentraUon, compound class, solubility in oil vs water (log P), molecular weight, topological polar surface area, charge, hydrogen bond acceptor and donor count, rotatable bond count, turnover and carbon source (Fig. 2A). Metabolite value explained 43% of the variability, more than any other factor (Fig. 2B). The structure-based classificaUon of compound class also had good predicUve power (29% explained), primarily because carboxylic and keto/hydroxy acids were released at significantly higher rates than amino acids and other compounds (Fig. 2A). However, these two factors are clearly confounded as class-level differences mirrored significant differences in metabolite values (Fig. 2C). The number of rotatable bonds, solubility in oil vs water (log P) and the number of hydrogen bonds, each related to membrane permeability (50, 51), were also predicUve of metabolite release, suggesUng passive diffusion as a relevant mechanism. Metabolite turnover is predicted to .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted August 19, 2025. ; https://doi.org/10.1101/2025.08.19.671024doi: bioRxiv preprint correlate negaUvely with release rates, according to the noise-averaging cooperaUon hypothesis (low turnover metabolites are more suscepUble to noise and therefore more useful to share within the populaUon) (33), but our dataset shows likle support for this. When we asked how well the best linear model could predict metabolite release rates, we found that 63% of the variability could be explained by metabolite value, compound class, intracellular concentraUon and charge (Fig. 2B). However, predicUve accuracy declined when applied to out-of-sample condiUons (45%, P = 0.04, 52.9 ± 3.8% expected from random model, Fig. S12) or out-of-sample metabolites (50%, P < 0.001, 55.2 ± 0.7% expected from random model, Fig. S13), likely due to correlaUons among the factors, which warrants cauUon when interpreUng factors’ relaUve importance (Fig. S14). Nevertheless, excluding metabolite value as a factor reduced model performance on out-of-sample condiUons (10% decline in median R2, P = 0.003, one-sided Mann-Whitney U, N = 20, Fig. S15), and model quality was largely dependent on including metabolite value or compound class (Fig. S16). Furthermore, we found differences in metabolite value between condiUons to be negaUvely correlated with differences in release rate (Pearson ρ, P = 2e-3, Fig. S17). This not only supports the relevance of this factor in predicUng out-of-sample rates but also raises the quesUon of whether microbes adapt their release rates to context-dependent costs. Finally, when we esUmated each factor’s importance separately for each microbe in each dataset (except C. glutamicum), metabolite value sUll emerged as the most robust explanatory factor (Fig. 2D). Cell lysis is commonly considered a key mechanism for metabolite release (31, 32). To measure its importance in our experiments, we used E. coli batch cultures with paired sampling of intra- and extracellular metabolites, as well as flow cytometry with live/dead staining to quanUfy the fracUon of lysed cells. In general, less than 10% of extracellular concentraUons could be explained by the release of intracellular metabolites following cell lysis, and osen much less (Fig. 3A). We then extended this analysis by comparing our collecUon of E. coli extracellular metabolite concentraUons to published intracellular concentraUons (46–49). To avoid asserUng equal lysis rates across condiUons, we asked how large a fracUon of the cell populaUon would need to be lysed to account for the extracellular levels, considering only the contribuUon from intracellular metabolite pools (Fig. 3B). Only 10.4% (157/1508) of the data points fall within the range of measured lysis fracUon (0.2-2.3%, mean ± std across condiUons = 0.6 ± 0.6%) or below, and more than 56.3% (849/1508) are above the limit where more than the whole cell populaUon is required. Glutamate, glutamine and NAD are the only metabolites where most .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted August 19, 2025. ; https://doi.org/10.1101/2025.08.19.671024doi: bioRxiv preprint data points can be accounted for by cell lysis, due to high intracellular concentraUons and moderate to low extracellular levels (Fig. S18). DegradaUon of cell debris may also contribute to extracellular metabolite pools (52). In E. coli, this degradaUon is dependent on the presence of Lon protease in the cell lysate (53). However, the benefit of cell debris degradaUon by Lon was only significant well beyond exponenUal phase, and only 19 ± 6% of oligopepUdes (6-10 amino acids) in the lysate were catabolised by Lon over 20 hours (53). Thus, in our dataset where 90% of the samples were collected before 26 hours (67% in exponenUal phase), we expect only a fracUon of the proteome from lysed cells to contribute to the extracellular concentraUons of amino acids. When we included 10% proteome degradaUon in our analysis, only the extracellular levels of tryptophan and threonine addiUonally became fully explained by cell lysis (Fig. S19). However, apart from amino acids, the metabolites in Fig. 3B are not typical degradaUon products. Overall, these results suggest that cell lysis and proteome degradaUon are insufficient to explain extracellular levels of most metabolites – consistent with previous work (12, 31). .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted August 19, 2025. ; https://doi.org/10.1101/2025.08.19.671024doi: bioRxiv preprint Figure 2: The importance of different factors in explaining metabolite release rates. A) Among the tested chemical and metabolic factors, log-scaled metabolite value explains the most variability in log-scaled E. coli release rates. The tested factors are shown in order of how much variability (R2) in log-scaled metabolite value they explain as a univariate linear model, shown in panel B. Compound class abbrevia:ons: CA: Carboxylic acids, AA: Amino acids, KA: Keto/hydroxy acids, OC: Organooxygen compounds, O: Other. Sta:s:cal test for compound classes and carbon source: Kruskal-Wallis H-test followed by Conover’s test with BH correc:on. For the other factors the reported P values are from Pearson correla:ons. B) The best linear model, ranked by BIC score, is a four-factor model that explains 63.4% of the variability in log-scaled release rates. C) Differences in metabolite values between compound classes. Abbrevia:ons and sta:s:cal test as in Fig. 2A. D) Log-scaled metabolite value consistently ranks among the most important factors when evaluated separately for each species in each dataset. Factor abbrevia:ons: RBC: Rotatable Bond Count, HBDC: Hydrogen Bond Donor Count, TPSA: Topological Polar Surface Area, HBAC: Hydrogen Bond Acceptor Count. .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted August 19, 2025. ; https://doi.org/10.1101/2025.08.19.671024doi: bioRxiv preprint Figure 3: Contribu'on from cell lysis and associated release of intracellular metabolites to extracellular concentra'ons. A) Using paired intra- and extracellular metabolite samples from the exponen:al phase of E. coli batch cultures we quan:fy what frac:on of extracellular metabolite concentra:ons can be explained by cell lysis. Point size reflects order-of-magnitude increase in extracellular concentra:on from the 2-hour sample to the late-exponen:al phase sample (Fig. S30), and all values except leucine in the L-malate condi:on were in the μΜ range. B) We expanded this analysis to all metabolite measurements in our E. coli datasets by using literature values for intracellular concentra:ons (Methods). For each metabolite, we es:mated the frac:on of cell popula:on that would need to be lysed for the intracellular concentra:ons released by cell lysis to explain the measured extracellular concentra:ons. The green region covers the range of measured frac:ons of lysed cells from A), and the red region is where more than the whole cell popula:on is required to explain extracellular concentra:ons. Abbrevia:ons: 2PG: 2- phosphoglycerate, 3PG: 3-phosphoglycerate, α-KG: Alpha-ketoglutarate, DHAP: Dihydroxyacetone phosphate, E4P: Erythrose 4- phosphate, FBP: Fructose 1,6-bisphosphate, F6P: Fructose 6-phosphate, G6P: Glucose 6-phosphate, PEP: Phosphoenolpyruvate, R5P: Ribose 5-phosphate, RU5P: Ribulose 5-phosphate, XU5P: Xylulose 5-phosphate. Metabolic disrup)on and evolu)on change metabolite release rates Given the effect of gene deleUons on intracellular metabolite levels (54), we assumed that the lack of negaUve correlaUon between metabolite value and release rate in C. glutamicum (Fig. 1D) could be a result of its engineered metabolism (42). To test the sensiUvity of metabolite release pakerns to modificaUons in core metabolism, we measured the exometabolome of eight E. coli knockout (KO) strains in late exponenUal phase in galactose medium (Fig. 4A). These gene knockouts target different parts of core metabolism, were previously shown to affect intracellular metabolite pools (54), and were predicted to change intracellular flux pakerns (Fig. S20). In line with our expectaUons, 14-38% (16-42 of 111) of the extracellular metabolite concentraUons differed significantly from the wild type across the KO strains (Welch’s t-test, FDR < 0.05, Fig. 4B). In contrast, only 2.7% of metabolite concentraUons were different in the negaUve control strain ΔlacA. Each KO strain displayed disUnct exometabolome pakerns (Fig. 4C), and we observed the most pronounced differences for metabolites in proximity of the deleted reacUon (Figs. S21-S23), consistent with intracellular effects of enzyme-deleUons (54). .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted August 19, 2025. ; https://doi.org/10.1101/2025.08.19.671024doi: bioRxiv preprint The WT reference for these KO strains from the KEIO collecUon (55) - E. coli BW25113 - had largely comparable rates as E. coli MG1655 used in the experiments presented in Fig. 1 (Fig. S24). Our central hypothesis for the negaUve correlaUon between metabolite value and release rate is that bacteria face a trade-off: the cost of prevenUng metabolite release versus the energy lost through that release. If this trade-off is under strong selecUon, then disrupUng the balance (e.g. via a gene knockout) should trigger rapid evoluUonary compensaUon to restore it. To test this, we evolved the ΔaceE and ΔsucB mutants – chosen for their severe growth defects and disUnct exometabolome characterisUcs – for ~100 (replicates M2-M4) and ~200 generaUons (M5-M7) in galactose medium, respecUvely (Fig. S25). The strains evolved in chemostats, where byproducts are conUnuously removed to avoid accumulaUon and later uptake of metabolites. From each chemostat endpoint, we isolated three strains and selected from each replicate the fastest growing clone (Fig. 4, D and E). These clones showed 94-146% (ΔaceE) and 50-82% (ΔsucB) growth rate increases relaUve to their ancestors (Fig. 4, F and G). To assess changes in metabolite release, we measured extracellular concentraUons of 5 metabolites at the end of exponenUal growth (Fig. 4, D and E). These metabolites were selected because their release was strongly disturbed in the ancestors (Figs. S21-S23), and either near the deleted reacUon (pyruvate and lactate for the ΔaceE; citrate, isocitrate and cis-aconitate for ΔsucB) where we expected the strongest effects, or located elsewhere in the metabolic network and of different chemical classes to probe more widespread effects. Lactate levels were significantly reduced in 2/3 ΔaceE isolates (Fig. 4F), which also exhibited the highest growth rates. The reducUons were, however, not sufficient to reach WT concentraUons. In contrast, all three ΔaceE isolates showed increased pyruvate release, contrary to expectaUons. These results were qualitaUvely invariant to growth rate normalizaUon (Fig. S26). For the ΔsucB isolates, we found in general small reducUons in extracellular concentraUons (Fig. 4G), but some of these differences change qualitaUvely upon growth rate normalizaUon (Fig. S27). For the metabolites elsewhere in the metabolic network, the outcomes were variable (Figs. S28 and S29). To esUmate the fitness contribuUon of reducing metabolite release, we used ecGEMs to predict relaUve growth rate penalUes associated with the observed release rates, using the ancestors’ mean growth rates as references. In the ΔaceE ancestor, lactate release is predicted to reduce fitness by 23% whereas pyruvate release has negligible impact (0.02%). The reduced release of lactate in the isolates from chemostat M2 and M4 is predicted to reduce the negaUve fitness .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted August 19, 2025. ; https://doi.org/10.1101/2025.08.19.671024doi: bioRxiv preprint effect to only 16% and 7%, respecUvely. These contribuUons are small compared to the observed increase (94-146%). The pakern is similar for the ΔsucB isolates: the esUmated net fitness reducUon of citrate, isocitrate and cis-aconitate in the ancestor is only 0.37%, far from sufficient to explain the improved growth rates. To beker understand the reasons for increased growth and changed release rates, we idenUfied sequence variants in the selected fast-growing strains and mutaUons that had fixed in the chemostat populaUons (Table S1-S4). Aside from the ΔaceE-M4 isolate, which likely hypermutated due to a mutT mutaUon, the evolved ΔaceE isolates have few sequence variants, and either related to galactose uptake (e.g. galS in M2, M4 and M7, regulaUng galactose uptake) or modulaUng fluxes near the deleted reacUon (e.g. poxB in M2, ldhA in M4, and icd, aceK and sdhA in M5, M6 & M7 and M7, respecUvely). Fixed non-synonymous variants in the chemostat populaUons support these observaUons, e.g. ldhA and galS were fixed in M2 and ilvG was fixed in M4. The ΔsucB populaUons have many more fixed mutaUons, but among the ones shared across replicates we find again galS and aceK. Together, these results show that metabolite release rates change as microbes evolve, but in our experiments, these changes were likely secondary to other, more dominant drivers of fitness – leading to inconsistent changes in metabolite release across replicates. .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted August 19, 2025. ; https://doi.org/10.1101/2025.08.19.671024doi: bioRxiv preprint Figure 4: Effect of gene knockouts on metabolite release rates and the effect of subsequent evolu'on to higher fitness. A) Growth curves of selected KO mutants, with sampling :me-points indicated by open circles. B) Comparison of extracellular metabolite concentra:ons in late exponen:al (OD600~1.2-1.3) between KO mutants and WT. Significant differences are marked by solid black outlines (Welch’s t-test, FDR < 0.05). C) A PCA plot of exometabolome profiles, showing that each strain has a dis:nct paeern. D & E) Batch culture growth curves and sampling :mepoints for isolates from the endpoint of the evolu:on experiment for ΔaceE and ΔsucB, respec:vely. F & G) Growth rates and extracellular metabolite concentra:ons for metabolites near the deleted reac:on, for ΔaceE and ΔsucB, respec:vely. Sta:s:cal significance was assessed using Welch’s t-test.

To beker understand metabolic interacUons between microbes, it is necessary to develop and test hypotheses for the fundamental principles that govern them. Here we propose that the broad release of intracellular metabolites is in general a loss of energy and that for each metabolite there is a trade-off between the cost of its release and the cost of prevenUng it. How microbes act to prevent release remains poorly understood, but one possibility is expressing .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted August 19, 2025. ; https://doi.org/10.1101/2025.08.19.671024doi: bioRxiv preprint transporters that immediately re-consume released metabolites (56, 57). This hypothesis connects to previous work linking fitness costs to microbial traits, including amino acid uptake rates (14), proteome composiUon (10, 39), properUes of extracellular enzymes (58) and metabolic strategies (59). We find consistent support for our hypothesis: (i) Metabolite values are negaUvely correlated with metabolite release rates across different species and nutrient environments, (ii) Metabolite value explains the most variability in release rates among all tested factors, (iii) Release rates cannot be explained by cell lysis alone. We also find that release rates are sensiUve to changes in intracellular metabolism, both caused by gene knockouts and subsequent evoluUon. But on the Umescale of our evoluUonary experiments, changes in release rates were inconsistent across replicates and did not always match our expectaUons. This may be because alternaUve paths towards increased fitness were favoured that were independent of metabolite release, or that increased growth rate was selected for more than increased yield. Beyond enabling order-of-magnitude predicUons of release rates using metabolic models, these findings impact how we think about metabolite release and the evoluUon of microbial cross- feeding. While it is commonly assumed that cross-feeding either emerges from extracellular enzymes, from the release of waste products or is an act that benefits other species at the expense of the releaser (2, 60–63), our results suggest another alternaUve: Microbes release metabolites because using, retaining or recapturing them is not worth the investment. This increases the scope for commensal cross-feeding, which in this scenario can arise without the need to invoke the evoluUon of cooperaUon or being vulnerable to cheaters (64). .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted August 19, 2025. ; https://doi.org/10.1101/2025.08.19.671024doi: bioRxiv preprint

1. C. Kost, K. R. PaUl, J. Friedman, S. L. Garcia, M. Ralser, Metabolic exchanges are ubiquitous in natural microbial communiUes. Nat. Microbiol., 1–9 (2023). 2. E. J. Culp, A. L. Goodman, Cross-feeding in the gut microbiome: Ecology and mechanisms. Cell Host Microbe 31, 485–499 (2023). 3. A. E. Douglas, The microbial exometabolome: ecological resource and architect of microbial communiUes. Philos. Trans. R. Soc. B Biol. Sci. 375, 20190250 (2020). 4. G. D’Souza, S. Shitut, D. Preussger, G. Yousif, S. Waschina, C. Kost, Ecology and evoluUon of metabolic cross-feeding interacUons in bacteria. Nat. Prod. Rep. 35, 455–488 (2018). 5. L. M. Solden, A. E. Naas, S. Roux, R. A. Daly, W. B. Collins, C. D. Nicora, S. O. Purvine, D. W. Hoyt, J. Schückel, B. Jørgensen, W. Willats, D. E. Spalinger, J. L. Firkins, M. S. Lipton, M. B. Sullivan, P . B. Pope, K. C. Wrighton, Interspecies cross-feeding orchestrates carbon degradaUon in the rumen ecosystem. Nat. Microbiol. 3, 1274–1284 (2018). 6. S. Pontrelli, R. Szabo, S. Pollak, J. Schwartzman, D. Ledezma-Tejeida, O. X. Cordero, U. Sauer, Metabolic cross-feeding structures the assembly of polysaccharide degrading communiUes. Sci. Adv. 8, eabk3076 (2022). 7. M. Gralka, R. Szabo, R. Stocker, O. X. Cordero, Trophic InteracUons and the Drivers of Microbial Community Assembly. Curr. Biol. 30, R1176–R1188 (2020). 8. M. Daniels, S. van Vliet, M. Ackermann, Changes in interacUons over ecological Ume scales influence single-cell growth dynamics in a metabolically coupled marine microbial community. ISME J. 17, 406–416 (2023). 9. Y.-C. Chuang, N. W. Haas, R. Pepin, M. G. Behringer, Y . Oda, B. LaSarre, C. S. Harwood, J. B. McKinlay, Bacterial adenine cross-feeding stems from a purine salvage bokleneck. ISME J. 18, wrae034 (2024). 10. M. T. Mee, J. J. Collins, G. M. Church, H. H. Wang, Syntrophic exchange in syntheUc microbial communiUes. Proc. Natl. Acad. Sci. 111, E2149–E2156 (2014). .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted August 19, 2025. ; https://doi.org/10.1101/2025.08.19.671024doi: bioRxiv preprint 11. K. Campbell, J. Vowinckel, M. Mülleder, S. Malmsheimer, N. Lawrence, E. Calvani, L. Miller- Fleming, M. T. Alam, S. Christen, M. A. Keller, M. Ralser, Self-establishing communiUes enable cooperaUve metabolite exchange in a eukaryote. eLife 4, e09943 (2015). 12. N. Paczia, A. Nilgen, T. Lehmann, J. Gätgens, W. Wiechert, S. Noack, Extensive exometabolome analysis reveals extended overflow metabolism in various microorganisms. Microb. Cell Factories 11, 122 (2012). 13. J. S. L. Yu, C. Correia-Melo, F. Zorrilla, L. Herrera-Dominguez, M. Y . Wu, J. Hartl, K. Campbell, S. Blasche, M. Kreidl, A.-S. Egger, C. B. Messner, V. Demichev, A. Freiwald, M. Mülleder, M. Howell, J. Berman, K. R. PaUl, M. T. Alam, M. Ralser, Microbial communiUes form rich extracellular metabolomes that foster metabolic interacUons and promote drug tolerance. Nat. Microbiol. 7, 542–555 (2022). 14. M. Zampieri, M. Hörl, F. Hotz, N. F. Müller, U. Sauer, Regulatory mechanisms underlying coordinaUon of amino acid and glucose catabolism in Escherichia coli. Nat. Commun. 10, 3354 (2019). 15. G. D’Souza, S. Waschina, S. Pande, K. Bohl, C. Kaleta, C. Kost, Less Is More: SelecUve Advantages Can Explain the Prevalent Loss of BiosyntheUc Genes in Bacteria. EvoluGon 68, 2559–2570 (2014). 16. R. Gregor, G. T. Vercelli, R. E. Szabo, M. Gralka, R. C. Reynolds, E. B. Qu, N. M. Levine, O. X. Cordero, Vitamin auxotrophies shape microbial community assembly in the ocean. bioRxiv [Preprint] (2024). hkps://doi.org/10.1101/2023.10.16.562604. 17. B. Ryback, M. Borneld-Miller, J. A. Vorholt, Metabolic adaptaUon to vitamin auxotrophy by leaf-associated bacteria. ISME J. 16, 2712–2724 (2022). 18. G. Yousif, F. Zorrilla, S. Dash, L. Oña, A. Shekhar, S. Giri, R. Guan, S. Harshe, M. Itermann, D. Welter, V. Benes, K. R. PaUl, C. Kost, Obligate cross-feeding of metabolites is common in soil microbial communiUes. bioRxiv [Preprint] (2025). hkps://doi.org/10.1101/2025.01.29.635426. .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted August 19, 2025. ; https://doi.org/10.1101/2025.08.19.671024doi: bioRxiv preprint 19. V. Baldazzi, D. Ropers, J.-L. Gouzé, T. Gedeon, H. de Jong, Resource allocaUon accounts for the large variability of rate-yield phenotypes across bacterial strains. eLife 12, e79815 (2023). 20. M. Basan, S. Hui, H. Okano, Z. Zhang, Y . Shen, J. R. Williamson, T. Hwa, Overflow metabolism in Escherichia coli results from efficient proteome allocaUon. Nature 528, 99– 104 (2015). 21. M. Huelsmann, O. T. Schubert, M. Ackermann, A framework for understanding collecUve microbiome metabolism. Nat. Microbiol. 9, 3097–3109 (2024). 22. C. Malina, R. Yu, J. Björkeroth, E. J. Kerkhoven, J. Nielsen, AdaptaUons in metabolism and protein translaUon give rise to the Crabtree effect in yeast. Proc. Natl. Acad. Sci. 118, e2112836118 (2021). 23. T. Pfeiffer, S. Schuster, S. Bonhoeffer, CooperaUon and CompeUUon in the EvoluUon of ATP- Producing Pathways. Science 292, 504–507 (2001). 24. M. L. Reaves, B. D. Young, A. M. Hosios, Y .-F. Xu, J. D. Rabinowitz, Pyrimidine homeostasis is accomplished by directed overflow metabolism. Nature 500, 237–241 (2013). 25. R. Heinrich, S. Schuster, H.-G. Holzhüker, MathemaUcal analysis of enzymic reacUon systems using opUmizaUon principles. Eur. J. Biochem. 201, 1–21 (1991). 26. E. Costa, J. Pérez, J.-U. Kres, Why is metabolic labour divided in nitrificaUon? Trends Microbiol. 14, 213–219 (2006). 27. M. J. van Hoek, R. M. Merks, Redox balance is key to explaining full vs. parUal switching to low-yield metabolism. BMC Syst. Biol. 6, 22 (2012). 28. G. N. Vemuri, M. A. Eiteman, J. E. McEwen, L. Olsson, J. Nielsen, Increasing NADH oxidaUon reduces overflow metabolism in Saccharomyces cerevisiae. Proc. Natl. Acad. Sci. 104, 2402–2407 (2007). 29. A. Bar-Even, E. Noor, A. Flamholz, J. M. Buescher, R. Milo, Hydrophobicity and Charge Shape Cellular Metabolite ConcentraUons. PLoS Comput. Biol. 7 (2011). .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted August 19, 2025. ; https://doi.org/10.1101/2025.08.19.671024doi: bioRxiv preprint 30. K. Campbell, L. Herrera-Dominguez, C. Correia-Melo, A. Zelezniak, M. Ralser, Biochemical principles enabling metabolic cooperaUvity and phenotypic heterogeneity at the single cell level. Curr. Opin. Syst. Biol. 8, 97–108 (2018). 31. J. B. McKinlay, Are Bacteria Leaky? Mechanisms of Metabolite ExternalizaUon in Bacterial Cross-Feeding. Annu. Rev. Microbiol. 77, 277–297 (2023). 32. S. Gude, G. J. Pherribo, M. E. Taga, Emergence of Metabolite Provisioning as a By-Product of Evolved Biological FuncUons. mSystems 5, 10.1128/msystems.00259-20 (2020). 33. J. G. Lopez, N. S. Wingreen, Noisy metabolism can promote microbial cross-feeding, eLife (2022). hkps://doi.org/10.7554/eLife.70694. 34. B. J. Sánchez, C. Zhang, A. Nilsson, P .-J. Lahtvee, E. J. Kerkhoven, J. Nielsen, Improving the phenotype predicUons of a yeast genome-scale metabolic model by incorporaUng enzymaUc constraints. Mol. Syst. Biol. 13, 935 (2017). 35. P . S. Bekiaris, S. Klamt, AutomaUc construcUon of metabolic models with enzyme constraints. BMC BioinformaGcs 21, 19 (2020). 36. Z. Mao, J. Niu, J. Zhao, Y . Huang, K. Wu, L. Yun, J. Guan, Q. Yuan, X. Liao, Z. Wang, H. Ma, ECMpy 2.0: A Python package for automated construcUon and analysis of enzyme- constrained models. Synth. Syst. Biotechnol. 9, 494–502 (2024). 37. J. D. Orth, I. Thiele, B. Ø. O. Palsson, What is flux balance analysis? Nat Biotech 28, 245– 248 (2010). 38. C. Gu, G. B. Kim, W. J. Kim, H. U. Kim, S. Y . Lee, Current status and applicaUons of genome- scale metabolic models. Genome Biol. 20, 121 (2019). 39. H. Akashi, T. Gojobori, Metabolic efficiency and amino acid composiUon in the proteomes of Escherichia coli and Bacillus subUlis. Proc. Natl. Acad. Sci. 99, 3695–3700 (2002). 40. M. Mori, C. Cheng, B. R. Taylor, H. Okano, T. Hwa, FuncUonal decomposiUon of metabolism allows a system-level quanUficaUon of fluxes and protein allocaUon towards specific metabolic funcUons. Nat. Commun. 14, 4161 (2023). .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted August 19, 2025. ; https://doi.org/10.1101/2025.08.19.671024doi: bioRxiv preprint 41. A. R. Pacheco, M. Moel, D. Segrè, Costless metabolic secreUons as drivers of interspecies interacUons in microbial ecosystems. Nat. Commun. 10, 103 (2019). 42. T. Georgi, D. Rikmann, V. F. Wendisch, Lysine and glutamate producUon by Corynebacterium glutamicum on glucose, fructose and sucrose: Roles of malic enzyme and fructose-1,6-bisphosphatase. Metab. Eng. 7, 291–301 (2005). 43. J. C. C. Vila, J. Goldford, S. Estrela, D. Bajic, A. Sanchez-GorosUaga, A. Damian-Serrano, N. Lu, R. Marsland, M. Rebolleda-Gomez, P . Mehta, A. Sanchez, Metabolic similarity and the predictability of microbial community assembly. bioRxiv [Preprint] (2023). hkps://doi.org/10.1101/2023.10.25.564019. 44. R. Baran, E. L. Brodie, J. Mayberry-Lewis, E. Hummel, U. N. Da Rocha, R. Chakraborty, B. P . Bowen, U. Karaoz, H. Cadillo-Quiroz, F. Garcia-Pichel, T. R. Northen, Exometabolite niche parUUoning among sympatric soil bacteria. Nat. Commun. 6, 8289 (2015). 45. P . Wendering, Z. Nikoloski, COMMIT: ConsideraUon of metabolite leakage and community composiUon improves microbial community reconstrucUons. PLOS Comput. Biol. 18, e1009906 (2022). 46. B. D. Bennek, J. Yuan, E. H. Kimball, J. D. Rabinowitz, Absolute quanUtaUon of intracellular metabolite concentraUons by an isotope raUo-based approach. Nat. Protoc. 3, 1299–1311 (2008). 47. J. O. Park, S. A. Rubin, Y .-F. Xu, D. Amador-Noguez, J. Fan, T. Shlomi, J. D. Rabinowitz, Metabolite concentraUons, fluxes and free energies imply efficient enzyme usage. Nat. Chem. Biol. 12, 482–489 (2016). 48. T. Sajed, A. Marcu, M. Ramirez, A. Pon, A. C. Guo, C. Knox, M. Wilson, J. R. Grant, Y . Djoumbou, D. S. Wishart, ECMDB 2.0: A richer resource for understanding the biochemistry of E. coli. Nucleic Acids Res. 44, D495–D501 (2016). 49. L. B. Thorfinnsdo‡r, L. García-Calvo, G. H. Bø, P . Bruheim, L. M. Røst, OpUmized Fast FiltraUon-Based Sampling and ExtracUon Enables Precise and Absolute QuanUficaUon of the Escherichia coli Central Carbon Metabolome. Metabolites 13, 150 (2023). .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted August 19, 2025. ; https://doi.org/10.1101/2025.08.19.671024doi: bioRxiv preprint 50. G. Falcón-Cano, C. Molina, M. Á. Cabrera-Pérez, Reliable PredicUon of Caco-2 Permeability by Supervised Recursive Machine Learning Approaches. PharmaceuGcs 14, 1998 (2022). 51. X. Liu, B. Testa, A. Fahr, Lipophilicity and Its RelaUonship with Passive Drug PermeaUon. Pharm. Res. 28, 962–977 (2011). 52. S. Pontrelli, K. B. Villi, A. Sichert, J. Trouillon, A. Rutz, Z. C. Landry, S. H. Rüdisser, R. Stocker, U. Sauer, DegradaUon of extracellular polymeric substances shapes microbial community diversity. PLOS Biol. 23, e3003287 (2025). 53. S. E. R. Gibson, I. Frost, S. J. Hierons, T. Moses, W. C. K. Poon, S. A. West, M. J. Cann, Bacteria encode post-mortem protein catabolism that enables altruisUc nutrient recycling. Nat. Commun. 16, 1400 (2025). 54. T. Fuhrer, M. Zampieri, D. C. Sévin, U. Sauer, N. Zamboni, Genomewide landscape of gene– metabolome associaUons in Escherichia coli. Mol. Syst. Biol. 13, 907 (2017). 55. T. Baba, T. Ara, M. Hasegawa, Y . Takai, Y . Okumura, M. Baba, K. A. Datsenko, M. Tomita, B. L. Wanner, H. Mori, ConstrucUon of Escherichia coli K-12 in-frame, single-gene knockout mutants: the Keio collecUon. Mol. Syst. Biol. 2, 2006.0008 (2006). 56. D. E. Rancourt, J. T. Stephenson, G. A. Vickell, J. M. Wood, Proline excreUon by Escherichia coli K12. Biotechnol. Bioeng. 26, 74–80 (1984). 57. B. Enjalbert, P . Millard, M. Dinclaux, J.-C. Portais, F. LéUsse, Acetate fluxes in Escherichia coli are determined by the thermodynamic control of the Pta-AckA pathway. Sci. Rep. 7, 42135 (2017). 58. D. R. Smith, M. R. Chapman, Economical EvoluUon: Microbes Reduce the SyntheUc Cost of Extracellular Proteins. mBio 1, e00131-10 (2010). 59. A. Flamholz, E. Noor, A. Bar-Even, W. Liebermeister, R. Milo, GlycolyUc strategy as a tradeoff between energy yield and protein cost. Proc. Natl. Acad. Sci. 110, 10039–10044 (2013). .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted August 19, 2025. ; https://doi.org/10.1101/2025.08.19.671024doi: bioRxiv preprint 60. R. K. Friks, A. L. McCully, J. B. McKinlay, Extracellular Metabolism Sets the Table for Microbial Cross-Feeding. Microbiol. Mol. Biol. Rev. 85, 10.1128/mmbr.00135-20 (2021). 61. J. A. Lopez, B. Liu, Z. Li, M. S. Donia, N. S. Wingreen, When is microbial cross-feeding evoluUonarily stable? bioRxiv [Preprint] (2025). hkps://doi.org/10.1101/2025.05.16.654511. 62. S. Pande, H. Merker, K. Bohl, M. Reichelt, S. Schuster, L. F. de Figueiredo, C. Kaleta, C. Kost, Fitness and stability of obligate cross-feeding interacUons that emerge upon gene loss in bacteria. ISME J. 8, 953–962 (2014). 63. D. Preussger, S. Giri, L. K. Muhsal, L. Oña, C. Kost, Reciprocal Fitness Feedbacks Promote the EvoluUon of MutualisUc CooperaUon. Curr. Biol. 30, 3580-3590.e7 (2020). 64. A. R. Zomorrodi, D. Segrè, Genome-driven evoluUonary game theory helps understand the rise of metabolic interdependencies in microbial communiUes. Nat. Commun. 8, 1563 (2017). 65. K. A. Datsenko, B. L. Wanner, One-step inacUvaUon of chromosomal genes in Escherichia coli K-12 using PCR products. Proc. Natl. Acad. Sci. 97, 6640–6645 (2000). 66. N. E. Lewis, K. K. Hixson, T. M. Conrad, J. A. Lerman, P . CharusanU, A. D. PolpiUya, J. N. Adkins, G. Schramm, S. O. Purvine, D. Lopez-Ferrer, K. K. Weitz, R. Eils, R. König, R. D. Smith, B. Ø. Palsson, Omic data from evolved E. coli are consistent with computed opUmal growth from genome-scale models. Mol. Syst. Biol. 6, 390 (2010). 67. R. J. G. B. Campello, D. Moulavi, J. Sander, “Density-Based Clustering Based on Hierarchical Density EsUmates” in Advances in Knowledge Discovery and Data Mining, J. Pei, V. S. Tseng, L. Cao, H. Motoda, G. Xu, Eds. (Springer Berlin Heidelberg, Berlin, Heidelberg, 2013), pp. 160–172. 68. H. Gallart-Ayala, I. Konz, F. Mehl, T. Teav, A. Oikonomidi, G. Peyratout, V. van der Velpen, J. Popp, J. Ivanisevic, A global HILIC-MS approach to measure polar human cerebrospinal fluid metabolome: Exploring gender-associated variaUon in a cohort of elderly cogniUvely healthy subjects. Anal. Chim. Acta 1037, 327–337 (2018). .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted August 19, 2025. ; https://doi.org/10.1101/2025.08.19.671024doi: bioRxiv preprint 69. V. van der Velpen, T. Teav, H. Gallart-Ayala, F. Mehl, I. Konz, C. Clark, A. Oikonomidi, G. Peyratout, H. Henry, M. Delorenzi, J. Ivanisevic, J. Popp, Systemic and central nervous system metabolic alteraUons in Alzheimer’s disease. Alzheimers Res. Ther. 11, 93 (2019). 70. J. A. Myers, B. S. CurUs, W. R. CurUs, Improving accuracy of cell and chromophore concentraUon measurements using opUcal density. BMC Biophys. 6, 4 (2013). 71. Q. Ren, B. Henes, M. Fairhead, L. Thöny-Meyer, High level producUon of tyrosinase in recombinant Escherichia coli. BMC Biotechnol. 13, 18 (2013). 72. A. Goelzer, J. Muntel, V. Chubukov, M. Jules, E. Prestel, R. Nölker, M. Mariadassou, S. Aymerich, M. Hecker, P . Noirot, D. Becher, V. Fromion, QuanUtaUve predicUon of genome- wide resource allocaUon in bacteria. Metab. Eng. 32, 232–243 (2015). 73. Y . Zhang, J. Cai, X. Shang, B. Wang, S. Liu, X. Chai, T. Tan, Y . Zhang, T. Wen, A new genome- scale metabolic model of Corynebacterium glutamicum and its applicaUon. Biotechnol. Biofuels 10, 169 (2017). 74. S. L. McGill, Y . Yung, K. A. Hunt, M. A. Henson, L. Hanley, R. P . Carlson, Pseudomonas aeruginosa reverse diauxie is a mulUdimensional, opUmized, resource uUlizaUon strategy. Sci. Rep. 11, 1457 (2021). 75. J. Kuepper, J. Dickler, M. Biggel, S. Behnken, G. Jäger, N. Wierckx, L. M. Blank, Metabolic Engineering of Pseudomonas puUda KT2440 to Produce Anthranilate from Glucose. Front. Microbiol. 6, 1310 (2015). 76. J. Niu, Z. Mao, Y . Mao, K. Wu, Z. Shi, Q. Yuan, J. Cai, H. Ma, ConstrucUon and Analysis of an Enzyme-Constrained Metabolic Model of Corynebacterium glutamicum. Biomolecules 12, 1499 (2022). 77. K. Wu, Z. Mao, Y . Mao, J. Niu, J. Cai, Q. Yuan, L. Yun, X. Liao, Z. Wang, H. Ma, ecBSU1: A Genome-Scale Enzyme-Constrained Model of Bacillus subUlis Based on the ECMpy Workflow. Microorganisms 11, 178 (2023). .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted August 19, 2025. ; https://doi.org/10.1101/2025.08.19.671024doi: bioRxiv preprint 78. J. Guo, H. Zhang, C. Wang, J.-W. Chang, L.-L. Chen, ConstrucUon and analysis of a genome- scale metabolic network for Bacillus licheniformis WX-02. Res. Microbiol. 167, 282–289 (2016). 79. J. Nogales, J. Mueller, S. Gudmundsson, F. J. Canalejo, E. Duque, J. Monk, A. M. Feist, J. L. Ramos, W. Niu, B. O. Palsson, High-quality genome-scale metabolic modelling of Pseudomonas puUda highlights its broad metabolic capabiliUes. Environ. Microbiol. 22, 255–269 (2020). 80. Z. Mao, J. Niu, J. Zhao, Y . Huang, K. Wu, L. Yun, J. Guan, Q. Yuan, X. Liao, Z. Wang, H. Ma, ECMpy 2.0: A Python package for automated construcUon and analysis of enzyme- constrained models. Synth. Syst. Biotechnol. 9, 494–502 (2024). 81. N. E. Lewis, K. K. Hixson, T. M. Conrad, J. A. Lerman, P . CharusanU, A. D. PolpiUya, J. N. Adkins, G. Schramm, S. O. Purvine, D. Lopez-Ferrer, K. K. Weitz, R. Eils, R. König, R. D. Smith, B. Ø. Palsson, Omic data from evolved E. coli are consistent with computed opUmal growth from genome-scale models. Mol. Syst. Biol. 6, 390 (2010). 82. A. Ebrahim, J. A. Lerman, B. O. Palsson, D. R. Hyduke, COBRApy: COnstraints-Based ReconstrucUon and Analysis for Python. BMC Syst. Biol. 7, 74 (2013). 83. A. Varma, B. W. Boesch, B. O. Palsson, Stoichiometric interpretaUon of Escherichia coli glucose catabolism under various oxygenaUon rates. Appl. Environ. Microbiol. 59, 2465– 2473 (1993). 84. E. Reznik, P . Mehta, D. Segrè, Flux Imbalance Analysis and the SensiUvity of Cellular Growth to Changes in Metabolite Pools. PLOS Comput. Biol. 9, e1003195 (2013). 85. Y . Ram, E. Dellus-Gur, M. Bibi, K. Karkare, U. Obolski, M. W. Feldman, T. F. Cooper, J. Berman, L. Hadany, PredicUng microbial growth in a mixed culture from growth curve data. Proc. Natl. Acad. Sci. 116, 14698–14707 (2019). 86. J. L. Reed, T. D. Vo, C. H. Schilling, B. O. Palsson, An expanded genome-scale model of Escherichia coli K-12 (iJR904 GSM/GPR). Genome Biol. 4, R54 (2003). .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted August 19, 2025. ; https://doi.org/10.1101/2025.08.19.671024doi: bioRxiv preprint 87. A. M. Feist, C. S. Henry, J. L. Reed, M. Krummenacker, A. R. Joyce, P . D. Karp, L. J. Broadbelt, V. HatzimanikaUs, B. Ø. Palsson, A genome-scale metabolic reconstrucUon for Escherichia coli K-12 MG1655 that accounts for 1260 ORFs and thermodynamic informaUon. Mol. Syst. Biol. 3, 121 (2007). 88. J. D. Orth, T. M. Conrad, J. Na, J. A. Lerman, H. Nam, A. M. Feist, B. Ø. Palsson, A comprehensive genome-scale reconstrucUon of Escherichia coli metabolism—2011. Mol. Syst. Biol. 7, 535 (2011). 89. J. M. Monk, C. J. Lloyd, E. Brunk, N. Mih, A. Sastry, Z. King, R. Takeuchi, W. Nomura, Z. Zhang, H. Mori, A. M. Feist, B. O. Palsson, iML1515, a knowledgebase that computes Escherichia coli traits. Nat. Biotechnol. 35, 904–908 (2017). 90. Y . Djoumbou Feunang, R. Eisner, C. Knox, L. Chepelev, J. HasUngs, G. Owen, E. Fahy, C. Steinbeck, S. Subramanian, E. Bolton, R. Greiner, D. S. Wishart, ClassyFire: automated chemical classificaUon with a comprehensive, computable taxonomy. J. CheminformaGcs 8, 61 (2016). 91. S. Seabold, J. Perktold, “Statsmodels: Econometric and StaUsUcal Modeling with Python” (AusUn, Texas, 2010; hkps://doi.curvenote.com/10.25080/Majora-92bf1922-011), pp. 92– 96. 92. M. Lukačišinová, B. Fernando, T. Bollenbach, Highly parallel lab evoluUon reveals that epistasis can curb the evoluUon of anUbioUc resistance. Nat. Commun. 11, 3105 (2020). 93. S. Chen, Ultrafast one-pass FASTQ data preprocessing, quality control, and deduplicaUon using fastp. iMeta 2, e107 (2023). 94. Simon Andrews, FastQC: A Quality Control Tool for High Throughput Sequence Data, Babraham InsUtute (2010). 95. H. Li, Minimap2: pairwise alignment for nucleoUde sequences. BioinformaGcs 34, 3094– 3100 (2018). .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted August 19, 2025. ; https://doi.org/10.1101/2025.08.19.671024doi: bioRxiv preprint 96. P . Danecek, J. K. Bonfield, J. Liddle, J. Marshall, V. Ohan, M. O. Pollard, A. Whitwham, T. Keane, S. A. McCarthy, R. M. Davies, H. Li, Twelve years of SAMtools and BCFtools. GigaScience 10, giab008 (2021). 97. P . Cingolani, A. Plaks, L. L. Wang, M. Coon, T. Nguyen, L. Wang, S. J. Land, X. Lu, D. M. Ruden, A program for annotaUng and predicUng the effects of single nucleoUde polymorphisms, SnpEff: SNPs in the genome of Drosophila melanogaster strain w1118; iso- 2; iso-3. Fly (AusGn) 6, 80–92 (2012). 98. E. Garrison, G. Marth, Haplotype-based variant detecUon from short-read sequencing. arXiv arXiv:1207.3907 [Preprint] (2012). hkps://doi.org/10.48550/arXiv.1207.3907. 99. S. D. Burz, A.-L. Abraham, F. Fonseca, O. David, A. Chapron, F. Béguet-Crespel, S. Cénard, K. Le Roux, O. Patrascu, F. Levenez, C. Schwintner, H. M. Blo‡ère, C. Béra-Maillet, P . Lepage, J. Doré, C. Juste, A Guide for Ex Vivo Handling and Storage of Stool Samples Intended for Fecal Microbiota TransplantaUon. Sci. Rep. 9, 8897 (2019). 100. B. D. Bennek, E. H. Kimball, M. Gao, R. Osterhout, S. J. Van Dien, J. D. Rabinowitz, Absolute metabolite concentraUons and implied enzyme acUve site occupancy in Escherichia coli. Nat. Chem. Biol. 5, 593–599 (2009). 101. M. Mori, S. Schink, D. W. Erickson, U. Gerland, T. Hwa, QuanUfying the benefit of a proteome reserve in fluctuaUng environments. Nat. Commun. 8, 1–8 (2017). 102. B. Volkmer, M. Heinemann, CondiUon-Dependent Cell Volume and ConcentraUon of Escherichia coli to Facilitate Data Conversion for Systems Biology Modeling. PLoS ONE 6, e23126 (2011). 103. V. Simensen, C. Schulz, E. Karlsen, S. Bråtelund, I. Burgos, L. B. Thorfinnsdo‡r, L. García- Calvo, P . Bruheim, E. Almaas, Experimental determinaUon of Escherichia coli biomass composiUon for constraint-based metabolic modeling. PLOS ONE 17, e0262450 (2022). .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted August 19, 2025. ; https://doi.org/10.1101/2025.08.19.671024doi: bioRxiv preprint

We thank all members of the Mitri lab at the University of Lausanne for discussions and feedback. We also thank people at the Department of Biotechnology and Nanomedicine at SINTEF for technical support and advice on the experimental design, and Vassily HatzimanikaUs and the Laboratory of ComputaUonal Systems Biotechnology at EPFL for feedback on the computaUonal analyses. Furthermore, we would like to thank Simon van Vliet, Olga Schubert, Guilhem Panneau, Jean C. C. Vila and Uwe Sauer for good discussions, Olga Schubert and Kim Schlegel for providing the KEIO strains, SebasUan Burz, You Zheng Teo and Clément Vulin for support with the live/dead staining protocol, and Vladimir Sentchilo for support with the flow cytometry and sampling for metabolomics analyses. We also thank the Metabolomics unit at UNIL, and specifically Hector G. Ayala and Julijana Ivanisevic, not only for conducUng the LC-MS analyses, but also for providing guidance and recommendaUons on related makers. Special thanks to Florent Mazel and Daniel Segrè for feedback on the dras manuscript. Finally, we thank Lilja B. Thorfinnsdo‡r, Jean C. C. Vila and Stephan Noack for sharing of experimental data. Funding Swiss NaUonal Science FoundaUon Swiss Postdoctoral Fellowship TMPFP3_217172 (SS), NaUonal Center of Competence in Research Microbiomes grant SNF 51NF40_180575 (SM, SS, AG, MAV, JL) Faculty of Biology and Medicine, University of Lausanne (EU) Swiss NaUonal Science FoundaUon Eccellenza grant PCEGP3_181272 (SM, MAV) SINTEF Industry (GB) Research Council of Norway, SFI Industrial Biotechnology, grant no 309558 (GB) Author contribu)ons ConceptualizaUon: SS, SM; Methodology: SS, AQ, JSL, GB, MAV, EU; InvesUgaUon: SS, GB, AG, EU; VisualizaUon: SS; Funding acquisiUon: SS, SM, PE; Project administraUon: SS, SM; Supervision: SS, SM; WriUng – original dras: SS; WriUng – review & ediUng: SS, AQ, GB, JSL, PE, EU, MAV, SM, AG Compe)ng interests Authors declare that they have no compeUng interests. .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted August 19, 2025. ; https://doi.org/10.1101/2025.08.19.671024doi: bioRxiv preprint Data and materials availability Strains and isolates used in this study are available on request. Formaked data and code for data analyses and visualizaUon are available on GitHub at hkps://github.com/Mitri- lab/metabolite-release. A permanent archive of this repository will be deposited to Zenodo upon publicaUon. The raw sequencing data and metadata have been deposited to the NCBI SRA database under the BioProject ID PRJNA1270783.

Strains All experiments were conducted with either E. coli K-12 MG1655, E. coli K12 BW25113 (wild-type ancestor for KEIO KO strains) or the following KO mutants from the KEIO collecUon (55): ΔaceE (JW0110), ΔcyoD (JW0419), ΔlacA (JW0333), ΔnuoA (JW2283), Δpgi (JW3985), Δrpe (JW3349), ΔsucB (JW0716), Δzwf (JW1841). The KEIO KO strains were verified by PCR with one primer binding inside the KanR casseke inserted during gene knock-out (Keio_K1_Rev for genes on the (- ) strand; Keio_KT_RvComp_Fw for genes on the (+) strand) and one gene-specific primer binding downstream of the knocked-out gene (primers are listed in Table S5) (55, 65). For all experiments, the KEIO KO strains were used with the kanamycin resistance casseke in place. Precultures and growth media For all experiments detailed below the strains were precultured as described here unless otherwise stated: From glycerol stocks stored at -70 °C each strain was streaked out onto LB agar plates and incubated overnight (37 °C). Then, one colony of each strain was picked and used to inoculate liquid precultures (10 mL LB in 50 mL Erlenmeyer flasks) that were incubated overnight (37 °C, 200 rpm). The kanamycin-resistant KEIO KO strains and strains evolved from these were always precultured in liquid LB and streaked out onto LB agar plates with 25 µg/mL kanamycin, while E. coli K-12 MG1655 and E. coli K12 BW25113 were always precultured in LB and streaked onto LB agar plates without selecUon. Unless otherwise noted, precultures were washed three Umes by centrifugaUon, with the supernatant removed aser each step, and the cells resuspended in M9 medium with no carbon source to eliminate residual LB before inoculaUon. Unless otherwise stated, M9 medium was always prepared by mixing (per litre): 200 mL 5X M9 salts (M6030, Sigma-Aldrich, St. Louis, MO, USA), 2 mL 1M MgSO4, 0.1 mL 1M CaCl2, 1 mL 0.5 g/L FeSO4·7H2O, MQ water and appropriate amounts of a carbon source stock .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted August 19, 2025. ; https://doi.org/10.1101/2025.08.19.671024doi: bioRxiv preprint soluUon. The final M9 medium was then pH-adjusted to pH = 7.0, filter sterilized (0.22 μm) and stored at 4°C unUl use. Selec)ng carbon sources for E. coli K-12 MG1655 batch cultures To complement the exisUng exometabolome data from batch cultures of E. coli on glucose (12), we aimed to select three carbon sources that differed in chemical class and caused variaUon in metabolic flux pakerns. We used the ecGEM eciJO1366 of E. coli K-12 MG1655 without any modificaUons and parsimonious FBA to simulate opUmal flux pakerns across 104 different growth supporUng metabolites (35, 66). We then performed a PCA analysis of the predicted fluxes (as Boolean values) to compare flux pakerns between the carbon sources (Fig. S5). We used HDBSCAN (67) to idenUfy clusters and selected from the different clusters in total 10 compounds that also differed in chemical class (sugars, amino acids and organic acids) and their entry point into central carbon metabolism (Fig. S5). We then screened the growth of the selected carbon sources in triplicates over 48 hours in 96-well plates (Fig. S5). For this experiment, we prepared M9 medium with each of the 10 different carbon sources to a concentraUon scaled to 120 mM carbon atoms. From washed LB precultures, E. coli K-12 MG155 was inoculated to OD600 = 0.05 in 200 μL of media and culUvated with conUnuous shaking (double orbital, 425 cpm) in a BioTek Synergy H1 (Agilent Technologies, Winooski, VT, USA) plate reader at 37 °C for 48 hours. We ulUmately chose galactose, L-alanine and L-malate, as they supported rapid growth, differed in chemical class, entered central metabolism at different locaUons and were grouped into different clusters based on flux pakerns (Fig. S5). E. coli K-12 MG1655 batch cultures in bioreactors with different carbon sources E. coli K-12 MG1655 was culUvated in lab-scale bioreactors using the three chosen carbon sources. Their concentraUons were normalized by the number of carbon atoms to a carbon atom concentraUon of 120 mM, corresponding to 20 mM galactose (G0625, Sigma-Aldrich), 30 mM L-malic acid (02290, Sigma-Aldrich) and 40 mM L-alanine (A7627, Sigma-Aldrich). M9 medium with these three carbon sources were prepared as described above using 25X carbon source stock soluUons. The strain was streaked from a frozen glycerol vial onto an LB agar plate and incubated at 37 °C overnight. One colony was transferred to a 250 mL baffled shake flask containing 40 mL LB medium. Aser 7 hours, a culture volume corresponding to a start OD600 = 0.05 was inoculated .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted August 19, 2025. ; https://doi.org/10.1101/2025.08.19.671024doi: bioRxiv preprint to a 500 mL baffled shake flask containing 100 mL M9 + 40 mM L-alanine. This flask was incubated for 36 h (200 rpm, 37 °C). For the two other carbon sources, the strain was incubated in LB for 17 h before inoculaUng a culture volume corresponding to OD600 = 0.05 in a 500 mL baffled shake flask containing 100 mL M9 + 20 mM L-galactose or M9 + 30 mM L-malic acid, respecUvely. The flasks were incubated for 24 h (200 rpm, 37 °C). Before inoculaUon to bioreactors, M9 precultures were centrifuged (3220 rcf, 10 min, 4 °C), supernatant removed, and pellet resuspended in M9 medium without carbon. Batch culUvaUons were performed in 1 L DASGIP bioreactors (Eppendorf DASGIP, Jülich, Germany) with an iniUal volume of 600 mL M9 medium supplemented with either 20 mM galactose, 30 mM L-malic acid or 40 mM L-alanine, three replicates per condiUon. The reactors were equipped with two Rushton impellers and probes for measurements of dissolved oxygen (DO) and pH. Submerged aeraUon was maintained at 0.5 vvm, and agitaUon was cascaded between 400 and 800, controlled by a DO set point of 30 %. pH set point was 7.0 and controlled using 2 M NaOH and 2 M HCl. Start OD 600 was 0.05. Sampling was performed regularly during the exponenUal growth phase with addiUonal samples collected during staUonary phase. OD600 was measured at all Ume points and supernatant stored: 2 mL culture was centrifuged (3220 rcf, 10 min, 4 °C), and supernatant frozen in -20 °C. At the end of exponenUal phase and at the end of the experiment, samples were collected for cell dry weight (CDW). 50 mL culture was centrifuged (3220 rcf, 10 min, 4 °C), supernatant removed, cell pellet resuspended in 50 mL MQ water, centrifuged, supernatant removed, cell pellet resuspended in 5-10 mL MQ water and transferred to a pre-weighed aluminium beaker and dried at 105 °C for 24 hours. Beakers were then weighed again and CDW calculated. Exometabolome analyses of )me-series samples from E. coli K-12 MG1655 batch cultures with different carbon sources Four samples from each of the nine batch cultures (three replicates of three different media) plus media samples were analysed with LC-MS and GC-MS to measure the extracellular concentraUon of metabolites at different Ume points throughout the experiment. The Ume points were selected based on Ume-series culUvaUon data to cover the exponenUal growth phase (three samples) plus a staUonary phase sample (Fig. S6). Samples from the different .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted August 19, 2025. ; https://doi.org/10.1101/2025.08.19.671024doi: bioRxiv preprint replicates were aligned based on the Uming of the end of the exponenUal growth phase (Fig. S6). LC-MS of selected bioreactor samples was conducted by the Metabolomics facility at UNIL using their method “h igh-coverage targeted analysis of polar metabolites”, following this procedure: Media samples were thawed on ice. An aliquot (20 µL) was extracted with the addiUon of ice-cold methanol (80 µL), vortexed (30 sec) and centrifuged (15000 rpm, 15 min, 4°C). The supernatants were transferred to vials for liquid chromatography – tandem mass spectrometry (LC-MS/MS) analyses. The extracts were analyzed by hydrophilic interacUon chromatography coupled to tandem mass spectrometer (6496 iFunnel , Agilent Te chnologies) using mulUple reacUon monitoring - MRM approach in both, posiUve and negaUve ionizaUon modes, to maximize the polar metabolome coverage. The analyUcal condiUons have been described in detail elsewhere (68, 69). Raw LC-MS/MS metabolome data were processed using the Mass Hunter QuanUtaUve analysis sosware (Agilent Technologies). The peak areas (or extracted ion chromatograms (EICs) for the monitored MRM transiUons) were used for relaUve comparison of metabolit e levels between different condiUons or groups of samples. Data quality assessment, including signal dris correcUon, was performed using pooled quality control (QC) samples analyzed periodically throughout the enUre batch. Metabolite quanUficaUon was performed using stable isotope labeled internal standards and calibraUon curves following the signal dris correcUon. Data processing was done using MassHunter QuanUtaUve analysis. Peak area integraUon was manually curate d, and concentraUons were reported by selecUng a 6-point porUon of the linear standard curves relevant for the observed concentraUons. The relaUve LC-MS data was standardized before subsequent analyses and data processing. Both the relaUve and absolute LC-MS data were cleaned of obvious outliers. For the absolute data the outliers were caused by carry-over from previous samples leading to too high malate (six samples) and succinate (one sample) values. For the relaUve LC-MS data there were addiUonal outliers of unknown reasons (39 values in total) related to in total 14 metabolites. For several metabolites there were more missing values (not detected/not quanUfied) in the absolute than in the relaUve data. For a few metabolites (phenylalanine, proline, creaUne) we leveraged good linear relaUonships between relaUve and absolute data (R2>0.8) to esUmate the missing absolute values using a simple linear transformaUon. .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted August 19, 2025. ; https://doi.org/10.1101/2025.08.19.671024doi: bioRxiv preprint GC-MS analysis of the same samples was conducted in the following way: AnalyUcal standards were prepared as M9 medium with increasing concentraUons of acetate, lactate, pyruvate, formate and propionate and stored at -70°C unUl use. Samples from the E. coli bioreactor batch cultures and analyUcal standards were thawed on iced. Then, 100 μL was transferred to an Eppendorf tube and kept on a cold block (-20 °C) while adding 5 μL 11% HCl and 500 μL diethyl ether. Tubes were capped, vortexed on a thermoblock (2000 rpm, 10 min, 1°C) and centrifuged (13 000 rcf, 5 min, 4°C). The top organic phase layer was pipeked to a glass vial and derivaUzed with 20 μL N-(t-butyldimethylsilyl)-N-methyltrifluoroacetamide (MTBSTFA) and vortexed briefly. Samples were then placed on a heat block (90 min, 35°C) and kept at 16°C unUl analysis. The samples were injected (1 μL) by a Pal3 autosampler onto an Agilent 8890-5977B GC-MSD (Agilent Technologies) with a VF-5MS (30 m x 0.25 mm x 0.25 mm) column. The samples were injected with a split raUo of 15:1, helium flow rate of 1 mL/min and inlet temperature of 230 °C. The temperature was held for 2 min at 50 °C, raised at 25 °C/min to 175 °C, 30 °C/min to 280 °C and held for 3.5 min. The MSD was run in scan mode from 40-500 Da. Analyte abundances were calculated using the MassHunter QuanUtaUve Analysis sosware (Agilent Technologies). Absolute concentraUons were calculated using standard curves. Galactose concentraUons were measured using the L-Arabinose/D-Galactose assay kit (K-ARGA, Megazyme, Bray, Ireland), following their rapid protocol for analyses in 96-well plates. Es)ma)on of metabolite release rates of E. coli, B. licheniformis, C. glutamicum and S. cerevisiae in glucose medium Time-series data on biomass (in OD600) and absolute extracellular metabolite concentraUons from batch culUvaUons of E. coli, B. licheniformis, C. glutamicum and S. cerevisiae in high glucose concentraUons (10-20 g/L) (12) was kindly provided by the authors (mean and standard deviaUons). For metabolites reported as a sum of two metabolites we akributed equal amounts to those metabolites to enable direct comparison with metabolite values esUmated for individual metabolites (described below). This includes 2-phosphoglycerate/3-phosphoglycerate and ribulose 5-phosphate/xylulose 5-phosphate. To esUmate the metabolite release and uptake rates of each species we first converted the OD600 values to gDW/L. For E. coli we used the conversion factor we esUmated from our own cell dry weight measurements in M9 galactose medium (0.346 gDW/L/OD600), which is comparable, but on the lower end of values found in the literature (0.36-0.515 gDW/L/OD600) (70, 71). For the three other species we used literature .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted August 19, 2025. ; https://doi.org/10.1101/2025.08.19.671024doi: bioRxiv preprint values: B. licheniformis (data from B. subUlis) 0.48 gDW/L/OD600 (72); C. glutamicum 0.27 gDW/L/OD600 (73); S. cerevisiae (mean of literature values) 0.746 gDW/L/OD600 (70). We then integrated the translated biomass Ume-series data to obtain biomass AUC data that we leveraged together with the mean absolute exometabolome data to esUmate the uptake or release rates for each measured metabolite for each species using linear regression (Figs. S1- S4). Rates were only esUmated for the exponenUal growth phase, and only unUl each metabolite either saturated or was clearly being re-consumed. Theore&cal jus&fica&on of es&ma&ng uptake or release rates from extracellular concentra&ons We start out from a simple assumpUon that the change in extracellular concentraUon 𝐶! of metabolite i can be explained by the net release or uptake 𝑎!(𝑡) (by any mechanism) by the microbes with biomass density 𝑋(𝑡) in the batch culture: 𝑑𝐶! 𝑑𝑡 = 𝑋(𝑡) ⋅ 𝑎!(𝑡) The concentraUon 𝐶!(𝑡) at a given Umepoint t is thus: 𝐶!(𝑡) = + 𝑋(𝑡) ⋅ 𝑎! (𝑡)𝑑𝑡 " "# + 𝐶!(𝑡#) If we assume that the uptake or release rate 𝑎!(𝑡) is constant, we see that the concentraUon 𝐶! (𝑡) is described by a linear funcUon of the constant release rate and the integrated biomass (area under the curve, AUC): 𝐶!(𝑡) = 𝑎! + 𝑋(𝑡)𝑑𝑡 " "# + 𝐶! (𝑡#) = 𝑎! ⋅ 𝑋$%& (𝑡) + 𝐶# Hence, when metabolites are released at constant rates the dynamics should be well explained by a linear regression of extracellular metabolite concentraUons on biomass AUC, and the specific net rate (uptake or release) is the slope of the curve. Es)ma)on of metabolite release rates of E. coli in galactose, L-malate and L-alanine For our bioreactor batch culUvaUons of E. coli in galactose, L-malate and L-alanine, we used paired cell dry weight measurements and OD600 readings to calculate conversion factors used to translate OD600 values for all Ume points (galactose: 0.346 ± 0.016; L-malate: 0.279 ± 0.017; L- alanine: 0.296 ± 0.011; mean ± std in gDW/L/OD600). The translated Ume-series data was used to calculate the biomass area under the curve used to esUmate metabolite release and uptake rates further described below. .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted August 19, 2025. ; https://doi.org/10.1101/2025.08.19.671024doi: bioRxiv preprint We esUmated the metabolite uptake or release rate by performing for each metabolite and each carbon source a linear regression of measured concentraUon (relaUve or absolute) vs biomass AUC using the three samples (Umepoints T1-T3) acquired from the three replicate batch cultures during exponenUal phase plus the iniUal M9 medium (T0; Figs. S7-S9). To avoid underesUmaUng release rates for metabolites that potenUally were being re-consumed at the end of the exponenUal phase we discarded the last Umepoint from the exponenUal phase in cases where the rate was posiUve (indicaUng metabolite release) but there was a significant reducUon in extracellular concentraUon from T2-T3 (P < 0.05, one-sided t-test). For 17 (of 126) metabolites we chose to not include T0 for at least one of the three condiUons in the linear regression, either because of very large variability of T0 or because the trend from T0 to T1 was clearly different from the general trend from T1-T3. We only esUmated the rate for metabolites where we had at least three data points from at least two different Ume points. For the metabolites where we had both relaUve and absolute data, we used the absolute data if there were enough absolute data points to get adequate slope esUmates. For seven metabolites (glutamine, alpha-aminoadiapate, serine, lysine, lactate, NAD and isocitrate) where both absolute and relaUve quanUficaUon was performed, we could esUmate the rate more accurately on the relaUve data set (because of more missing data points in the absolute data set). However, there were sufficient absolute measurements to esUmate the absolute spread (standard deviaUon) of these data points, enabling translaUon of the slope value from Z-score to μmol/gDW/h. Es)ma)on of metabolite release rates of E. coli, P. p u 9 d a, an Enterobacter sp. and a Pseudomonas sp. in various carbon source environments The dataset from Vila et al. (2023) was kindly provided by the authors. The dataset contains exometabolome data from three Umepoints (16, 28, and 48 hours) of E. coli MG1655, P. p u G d a KT2440, and two environmental isolates of the genera Enterobacter and Pseudomonas, respecUvely (43). All species were culUvated in M9 medium with one of five carbon sources (D- glucose, D-fructose, glycerol, pyruvate or L-malate). AddiUonally, the dataset includes exometabolome data from two Umepoints (28 and 48 hours) of E. coli and the Enterobacter sp. on D-ribose, L-arabinose and D-galactose, and of P. p u G d a and the Pseudomonas sp. on acetate, fumarate and succinate. To esUmate metabolite release rates, we first used the plate reader growth curves to calculate biomass AUC aser 16, 28 and 48 hours and to idenUfy the end of exponenUal growth of each organism in each condiUon. The end of exponenUal phase was .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted August 19, 2025. ; https://doi.org/10.1101/2025.08.19.671024doi: bioRxiv preprint automaUcally idenUfied as the first peak in the second derivaUve of the smoothed growth curve with an OD600 above 60% of maximum. We converted the AUC from OD600 to gDW/L using the same conversion factor for all the different carbon source environments: 0.36 gDW/L/ OD600 for E. coli and the Enterobacter sp., mean of measured values and literature values (70, 71); 0.476 gDW/L/ OD600 for P. p u G d a and the Pseudomonas sp, mean of literature values (74, 75). Finally, we performed a linear regression between extracellular metabolite concentraUons and AUC on both duplicates of each strain/carbon source combinaUon including all datapoints before the end of exponenUal growth + 2 hours (a buffer to account for some inaccuracy in the automaUc detecUon). We only included rate esUmates for metabolites that were quanUfied at least twice for each strain. As the dataset did not include an early Umepoint or media measurement we assumed the iniUal concentraUon of each metabolite to be 0. If a metabolite was detected more at more than one Umepoint we discarded the esUmated rate if there was a qualitaUve discrepancy in slope value when the slope was esUmated with or without the T0 concentraUon assumed to be 0. If the carbon source was among the measured metabolites, this metabolite was excluded from rate esUmaUon in those specific condiUons. Genome-scale metabolic models For E. coli, S. cerevisiae and C. glutamicum we used the previously developed enzyme- constrained genome-scaled metabolic models (GEMs) eciJO1366 (35), ecCGL1 (76) and ecYeastGEM 8.3.4 (34), respecUvely. For B. licheniformis there was no available enzyme- constrained GEM, and we therefore opted to use the enzyme-constrained GEM ecBSU1 of the closely related species B. subGlis (77), which was available at github.com/Ubbdc/ecBSU1. A normal GEM for B. licheniformis is available (78), but with this model we could not achieve any feasible flux balance soluUon. To simulate the environmental isolate of the genus Enterobacter we used the E. coli GEM eciJO1366. For the environmental isolate of the genus Pseudomonas and for P. p u G d a we used the enzyme-constrained version of the P. puGda GEM iJN1463 (79) automaUcally generated by the ECMpy2 method (80), which we here refer to as eciJN1463. We made a few minor changes to the published GEMs before predicUng metabolite values. In the E. coli GEM eciJO1366 the proteome weight of the forward reacUon GALKr (galactokinase) was extremely high (0.016), limiUng the growth rate on galactose to much less than the growth rate we measured in M9 galactose medium. To enable eciJO1366 to reach the measured growth rate with galactose as the only carbon source we replaced the proteome weight of the forward .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted August 19, 2025. ; https://doi.org/10.1101/2025.08.19.671024doi: bioRxiv preprint GALKr reacUon with the proteome weight of the reverse GALKr reacUon which was much smaller (0.00037). For the ecYeastGEM we only merged exchange reacUons that were previously split to simplify its use. Both ecBSU1 and eciJN1463 were translated from json to sMOMENT format in xml (35), modified by merging split exchange reacUons and fixing metabolite compartment names and annotaUons. For ecBSU1 we also removed the enzyme constraint on the reacUon “PPAm_num1” (inorganic diphosphatase) as it caused unrealisUc uptake and release rates of phosphate and diphosphate, respecUvely. The C. glutamicum GEM ecCGL1 was also converted from the ECMpy format (36) to the sMOMENT format (35) but otherwise kept unchanged. All GEMs were used with their default biomass equaUons as objecUve when maximizing growth in all FBA or pFBA analyses (81). Reframed 1.5.3 (github.com/cdanielmachado/reframed) or COBRApy 0.29.1 (82) were used to load, manipulate and run constraint-based analyses with Gurobi 10 (Gurobi OpUmizaUon, LLC) as the solver. Calcula)on of metabolite values and metabolite turnover using genome-scale metabolic models In constraint-based modelling a shadow price is an esUmate of how sensiUve the objecUve funcUon in a flux balance analysis (FBA) is to increased influx or efflux of a metabolite (83, 84). Hence, one can interpret the shadow price as metabolite value as it quantifies the impact on growth caused by the release of a metabolite (to have positive values we here define metabolite value as the shadow price multiplied by -1). A similar approach to measure the cost of metabolite release has been used to identify costless secretions (41). The metabolite value (or the shadow price) can be predicted for a specific organism in a specific context by using a GEM of the corresponding organism (or closely related) and then constraining this GEM to the specific context (i.e. primarily defining metabolite uptake rates). While shadow prices are usually estimated directly by the solver used to run FBA (37), the robustness of these values is often questioned because of numerical issues. To get more robust shadow price estimates we solved for each metabolite in each context a separate FBA, where we constrained the GEM to have an (additional) release of 0.01 mmol/gDW/h of that metabolite and quantified the change in the objective function, i.e. the change Δy in growth rate caused by the forced metabolite release. The metabolite value was then quantified as Δy/0.01. If the metabolite was predicted to be released by FBA, we added 0.01 mmol/gDW/h to the FBA-predicted value to estimate the rate. This was only relevant for acetate for E. coli in the L-malate condition, for ethanol, acetate, formate and pyruvate for S. cerevisiae, and for pyruvate for P. putida in the L-malate condition. .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted August 19, 2025. ; https://doi.org/10.1101/2025.08.19.671024doi: bioRxiv preprint Using this approach, we esUmated metabolite values for E. coli in batch cultures with glucose, galactose, L-malate or L-alanine as the carbon source by constraining the uptake of the corresponding exchange reacUon in eciJO1366 to the uptake rate of the carbon source as esUmated from the experimental data. The same approach was used for C. glutamicum, B. licheniformis and S. cerevisiae using the corresponding models. For E. coli, P. p u G d a and the two environmental isolates of the genera Enterobacter and Pseudomonas we first extracted maximum growth rates for each organism in each environment from growth curves (using curveball (85)). Then, we used the respecUve GEMs to calculate the uptake rates of the corresponding carbon sources required to sustain the maximum growth rates. These uptake rates were then used to constrain the GEMs when esUmaUng the metabolite values as described above. The turnover (i.e. the sum of posiUve fluxes producing an intracellular metabolite) was esUmated by running parsimonious FBA (66), using the same GEMs constrained in the same way as described for above for metabolite values. To validate that the negaUve correlaUon between log-transformed metabolite values and release rates we compared esUmates from eciJO1366 with esUmates from the 4 benchmark E. coli K-12 MG1655 GEMs iJR904 (86), iAF1260 (87), iJO1366 (88) and iML1515 (89) that are of increasing complexity. We then compared esUmated metabolite values and esUmated release rates using data from E. coli in bioreactors with glucose, galactose, L-malate or L-alanine as the sole carbon source, the same data used in Fig. 2A (Fig. S11). In the L-malate condiUon for the iJR904 model we discarded the metabolite value for formate as an outlier because it was extremely small (<10-12), below the solver tolerance. Note that no metabolite values were negaUve and hence no other values were discarded upon log-transformaUon. Metabolite classifica)on, chemical proper)es and intracellular concentra)ons The charge and molecular weight of metabolites were obtained from eciJO1366 (35). All other chemical properUes and InChIKeys were obtained from PubChem using the python interface PubChemPy (github.com/mcs07/PubChemPy). The InChIKeys were then used to classify metabolites into a chemical taxonomy using ClassyFire (90). To get an appropriate level of detail of metabolic classificaUon we used the “Class” level category. However, we separated the class “Carboxylic acids and derivaUves” into “Amino acids” and “Carboxylic acids” based on the “Subclass” level. We also merged the classes “Hydroxy acids and derivaUves” and “Keto acids .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted August 19, 2025. ; https://doi.org/10.1101/2025.08.19.671024doi: bioRxiv preprint and derivaUves” into the joint class “Keto / hydroxy acids”. Finally, metabolites that did not fall into either of the abovemenUoned categories nor the class “Organooxygen compounds” were joined into the category “Other”. Linear models for evalua)ng the importance of different factors for predic)ng metabolite release rates Linear models were created and analysed using statsmodels (91). Compound class and carbon source was incorporated as categorical variables. Metabolites predicted to have zero turnover was given a value of 10-4, less than the minimum predicted turnover. Release rates where corresponding values were missing in any of the compared factors were discarded before model fi‡ng. To evaluate the staUsUcal significance of the out-of-sample predicUons we performed 104 permutaUons of model fi‡ng and out-of-sample esUmaUons where the labels (either metabolite or condiUon) was shuffled on each permutaUon. We then calculated the fracUon of out-of-sample R2 values from these permutaUons below the observed R2 value to esUmate the P value. Cul)va)on and exometabolome analyses of KEIO knockout strains One objecUve was to study how disrupUon of key metabolic fluxes would affect extracellular metabolite concentraUons and how this would change if these strains were allowed to evolve in a simple nutrient environment. The KEIO collecUon is a collecUon of non-essenUal E. coli KO strains that we could use for this purpose (55). To idenUfy relevant KEIO KO strains we used eciJO1366 and parsimonious FBA (81) to predict the effect of a gene knockout on opUmal metabolic fluxes, and previous metabolomics data (54) of these strains to see the effect on intracellular concentraUons (Fig. S20A). We eventually chose seven strains (ΔaceE, ΔcyoD, ΔnuoA, Δpgi, Δrpe, ΔsdhB, ΔsucB) that were predicted to have different opUmal flux distribuUons, had many significantly changed intracellular metabolite concentraUons, and targeted different part of key metabolic pathways (Fig. S20). We included ΔlacA as a negaUve control expected to behave like the WT as done previously (92). The selected strains from the KEIO collecUon were precultured on LB agar and liquid LB as previously described, and glycerol stocks used for experiments described below were prepared by adding glycerol to a final concentraUon of 25 %. .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted August 19, 2025. ; https://doi.org/10.1101/2025.08.19.671024doi: bioRxiv preprint For exometabolome sampling, the KEIO strains and E. coli BW25113 were precultured as detailed above. They were then re-inoculated to OD600 = 0.05 in 10 mL M9 + 40 mM galactose (+ 25 µg/mL kanamycin for KEIO strains) in 50 mL Erlenmeyer flasks (200 rpm, 37 °C, 31 h). 1 mL culture was sampled into Eppendorf tubes and washed twice with M9 without carbon source by centrifugaUon (6000 rcf, 6 min, 4°C). UlUmately, samples were resuspended in 0.5 mL M9 without carbon source and used to inoculate 200 μL M9 + 40 mM galactose (+25 µg/mL kanamycin for KEIO strains) in a flat bokom 96-well plate to OD600 = 0.05 and 6 mL M9 + 40 mM galactose (+25 µg/mL kanamycin for KEIO strains) to straight glass tubes (15.8 cm tall, 1.5 cm diameter) to OD600 = 0.005. The 96-well plate was incubated in a Synergy H1 plate reader with conUnuous shaking (double orbital, 425 cpm, 37 °C), and the glass tubes in a shaking incubator (200 rpm, 37 °C, 45° angle), both for 108 hours. The experiment in the well plate was used to monitor more closely the growth (as OD600) of these strains and to help with the Uming of the exometabolome sampling. All strains were culUvated in three replicates. Two samples were collected during mid-exponenUal growth phase for all cultures, mostly to ensure that at least one sample from each replicate was obtained before the end of exponenUal phase. 1 mL culture volume was centrifuged (14 000 rpm, 10 min, 4 °C), and supernatant stored in -70 °C for exometabolome analysis. The second sample for each replicate (Fig. 4A) was analysed using LC- MS at the UNIL Metabolomics facility, following their protocol described above. AddiUonally, we also analysed three samples of pooled inoculums to verify that any carry-over with the inoculum was much lower (or below detecUon limit) of than the sample metabolite concentraUons. Three replicate pooled samples were obtained by inoculaUng 6 mL M9 + 20 mM galactose medium with all the strains, each strain to an OD600 = 0.005, and subsequently handled as described above for the culture samples. Chemostat evolu)on experiment with ΔaceE and ΔsucB in bioreactors Two of the KEIO KO strains, ΔaceE and ΔsucB, were selected for the evoluUon experiment with conUnuous culUvaUon (chemostats) in bioreactors based on their reduced growth compared to their ancestor and difference in exometabolome pakerns (Figs. 4A and S21-S23). The strains were streaked from glycerol vials onto LB + 25 µg/mL kanamycin agar plates and incubated at 37 °C overnight. 5-6 colonies were transferred to 500 mL baffled shake flasks containing 75 mL LB + 25 µg/mL kanamycin and incubated for 22 h (200 rpm, 37 °C). .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted August 19, 2025. ; https://doi.org/10.1101/2025.08.19.671024doi: bioRxiv preprint Chemostat culUvaUons were performed in 1 L DASGIP bioreactors with a constant volume of 300 mL M9 + 20 mM galactose medium + 25 µg/mL kanamycin (prepared as previously described, but without pH adjustment because of pracUcal challenges with the large volumes required, so pH ≈ 7.1). CulUvaUon was performed at 37 °C with a submerged airflow maintained at 0.5 vvm, and a cascaded agitaUon of 400-700 rpm, controlled by a DO set point of 30 %. pH control was started at day 6 (ΔaceE) or day 20 (ΔsucB) with a set point of 7.1, controlled using 2 M NaOH (Fig. S25A). The reactors (three replicates per strain) were inoculated to OD600 = 0.1. During the late stage of exponenUal growth, pumps were started to feed fresh culture medium and remove used medium. Medium was fed using a pre-calibrated peristalUc pump and used medium was removed by a steel pipe placed right above the culture surface. The steel pipe was coupled to a peristalUc pump operaUng at a flow rate approximately 1.5x of the medium in flow, to avoid volume accumulaUon. The set diluUon rate was adjusted several Umes during the experiment to compensate for (someUmes surprising) changes in growth dynamics (from 0-0.1 h-1 for ΔaceE, 0- 0.3 for ΔsucB, Fig. S25B). The actual diluUon rate was monitored by keeping the flasks with fresh culture medium on scales. OD600 was measured daily, and weekly, larger samples were collected. Culture was centrifuged (3220 rcf, 10 min, 4 °C), and supernatant and pellet stored at -70 °C. In addiUon, cultures were preserved at -70 °C as glycerol stocks. The number of generaUons were esUmated from the measured diluUon rates across the experiment plus the generaUons during the iniUal batch culture phase before the culture diluUon was started. Cul)va)on of single colonies from evolu)on experiment and sampling of exometabolome Culture samples from different bioreactors and Umepoints collected in the chemostat experiment were streaked out onto LB + 25 µg/mL kanamycin agar plates and incubated overnight at 37 °C. Three colonies were picked and individually inoculated in 20 mL LB + 25 µg/mL kanamycin in 100 mL Erlenmeyer flasks and incubated (37 °C, 200 rpm, 16 h). 12 mL of the culture was then centrifuged (3220 rcf, 6 min, RT), and the pellets stored at –20 °C for sequencing. The isolates were stored as glycerol stocks at -70 °C made from these cultures. 2 mL samples from the precultures were washed as previously described, resuspended in M9 without carbon source, and used to inoculate a flat bokom 96-well plate to OD600 = 0.05 with 200 µL M9 + 25 mM galactose + 25 µg/mL kanamycin. The M9 medium was prepared as .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted August 19, 2025. ; https://doi.org/10.1101/2025.08.19.671024doi: bioRxiv preprint previously described, but to a pH of 7.1 as in the chemostats. The well plate was incubated in a Synergy H1 plate reader with conUnuous shaking (double orbital, 425 cpm, 37 °C, 120 h) and the OD600 was read every 20 minutes. Based on these growth measurements we selected the fastest growing isolates from the last Umepoint of each of the six bioreactors (three bioreactors per strain). These isolates, the iniUal KEIO KO strains ΔaceE and ΔsucB, and the KEIO ancestor E. coli BW25113 were precultured and subsequently washed as described above. The washed preculture were used to inoculate 10 mL M9 + 20 mM galactose to OD600 = 0.01 in straight glass tubes (15.8 cm tall, 1.5 cm diameter) in triplicates and incubated for 20-84 hours (200 rpm, 37 °C, 45° angle). 1 mL samples for exometabolome analysis were collected at OD600 ≈ 1 (Fig. 4, D and E), filtered using a 0.22 μm Millex-GV PVDF syringe filter (SLGVR33RS, MilliporeSigma, Darmstadt, Germany) into a 1.5 mL Eppendorf tube stacked in a cold block (-20 °C), and then stored at -70 °C unUl it was analysed. The exometabolome LC-MS analyses were conducted by the UNIL Metabolomics facility, following the procedure described above. DNA extrac)on of samples from selected isolates and chemostat culture samples The DNA extracUon of 14 different samples from the chemostat cultures (#37A D44 M2 Pellet, #38A D44 M3 Pellet, #39A D44 M4 Pellet, #28B D30 M5 Pellet, #29B D30 M6 Pellet, #30B D30 M7 Pellet) and the selected isolates (sucB-M5-D30-4, sucB-M6-D30-6, sucB-M7-D30-4, sucB- Ancestor, aceE-M2-D44-2, aceE-M3-D44-3, aceE-M4-D44-1, aceE-Ancestor) was performed by adapUng the protocol from the FastPure Bacteria DNA IsolaUon Mini Kit (DC103-01, Vazyme, Nanjing, China). 1 mL GA buffer was added to the thawed pellet samples, and 230 µL of these resuspensions were kept in Eppendorf tubes for DNA extracUon. The rest was centrifuged (4000 rpm, 6 min, RT) and the pellet stored at –20 °C. Genomic DNA (gDNA) was extracted from the 230 µL sample following the manufacturer’s protocol. In the final step, eluUon of the DNA was performed twice with fresh eluUon buffer, eluUng a total of 100 µL DNA. Extracted gDNA was stored at –70 °C unUl sequencing. Genomic sequencing processing Genomic DNA from the selected isolates and chemostat samples was sequenced by Novogene (Planegg, Germany) using the Illumina NovaSeqX PE 150. The isolates and the bioreactor samples were sequenced to 1 Gb and 10 Gb sequencing depth, respecUvely. Good quality of .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted August 19, 2025. ; https://doi.org/10.1101/2025.08.19.671024doi: bioRxiv preprint both datasets was ensured using fastqc v. 0.12.1 before trimming the reads with fastp v. 0.24.0 (--qualified_quality_phred 20 --cut_front --cut_tail --average_qual 20 --length_required 50) (93, 94). Both read sets were mapped against the E. coli BW25113 reference genome (NCBI RefSeq assembly GCF_000750555.1) with minimap2’s v. 2.28 default “sr” parameters (95). The resulUng alignments were filtered with samtools view v. 1.20 (-b -f 3 -q 60) to filter out reads not properly mapped in pairs or mapped to the reference with a MAPQ lower than 60 (96). Average coverage was assessed for both datasets showing a minimum average coverage for the chemostat culture samples of 2017 and 190 for the sequenced isolates. For both datasets, variants were idenUfied with freebayes v. 1.3.6 (--min-alternate-count 3 -p 1 - -min-alternate-fracUon 0.05 --pooled-conUnuous --haplotype-length 0) and annotated with snpEff v. 5.0 using the BW25113 reference genome (97, 98). A custom python script was used to filter the annotated variants. For the isolates, variants with a frequency lower than 0.95 and a coverage lower than 100 were filtered out. For the chemostat culture samples, variants with a frequency lower than 0.05 and coverage lower than 100 were filtered out. A variant was considered fixed if its alle frequency was at least 0.9. Paired intracellular and extracellular metabolomics and live/dead staining M9 medium with 20 mM galactose, 30 mM malate or 40 mM L-alanine was prepared as detailed above. E. coli K-12 MG1655 was precultured on LB agar plates and in liquid LB as detailed above. 1 mL preculture was washed as described above and used to start a second preculture in 20 mL M9 + 40 mM L-alanine, inoculated to OD600 = 0.05. Another LB preculture of E. coli K-12 MG1655 was used to inoculate a second M9 + 20 mM galactose and a M9 + 30 mM malate preculture in the same way 12 hours later. Aser 36/24 hours, 1 mL was collected from each of the three precultures and washed twice as detailed above. These washed precultures were used to inoculate 10 mL of the same media (M9 + either 20 mM galactose, 30 mM malate or 40 mM L-alanine) in straight glass tubes (15.8 cm tall, 1.5 cm diameter). These glass tubes were incubated for two days (200 rpm, 37 °C, 45° angle). Growth was monitored by measuring OD600 directly in the glass tubes (NANOCOLOR VIS II, MACHEREY-NAGEL, Düren, Germany) at 21 Umepoints. The glass tubes were vortexed briefly before any OD reading or sampling. A reference exometabolome 0.5 mL sample was collected aser 2 hours following the procedure detailed below. Then, 2 mL samples for paired intra- and extracellular metabolome and live/dead analyses were collected in the late exponenUal phase (at OD600 ≈ 1.1, Fig. S30). Exometabolome samples were obtained by filtering 0.5 mL culture using a 0.22 μm Millex-GV PVDF syringe filter (SLGVR33RS, MilliporeSigma) into a 1.5 mL Eppendorf tube stacked in a cold .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted August 19, 2025. ; https://doi.org/10.1101/2025.08.19.671024doi: bioRxiv preprint block (-20 °C). The samples for intracellular metabolomics, i.e. cells, were obtained immediately aser following this procedure: A 25 mm 0.22 μm Durapore PVDF filter (GVWP02500, MilliporeSigma) was placed in a new filter holder. A 5 mL syringe was connected to the inlet of the filter holder and a vacuum pump (250 mbar) was connected to the outlet of the filter holder. Then, 2 mL of 37 °C MQ water was added to pre-wet the filter. Immediately aser, 1 mL of culture sample was added to the syringe and allowed to pass through the filter. If the culture did not pass through solely based on the vacuum pump, light addiUonal pressure was added by using the syringe itself to push the culture through the filter. This process took 10-30 seconds. SUll using the same set-up, the filter was finally rinsed with 5 mL of 37 °C MQ water as previously recommended (49). Two tweezers were quickly rinsed in MQ water and used to carefully transfer the filter from the filter holder to 2 mL lysis tubes filled with 1.4 mL of 80:20 methanol:water (v/v), pre-chilled to -20 °C and kept on a cold block also pre-chilled to -20 °C. All samples were then quickly transferred to -70 °C for storage. Live/dead staining was performed using a similar approach as previous work (99): Briefly, the live sample was made by diluUng 10 μL culture in 990 μL PBS and the dead sample was made by mixing 100 μL culture with 1 mL 70% isopropanol. Both samples were vortexed and incubated at room temperature for 1 h. Then, the isopropanol was removed from the dead sample by centrifugaUon (8000 rcf, 4 min, RT), pipe‡ng off the supernatant, resuspending the pellet in 1 mL PBS and vortexing. This procedure was repeated once. Then, both the live and the dead sample were stained with propidium iodide (PI; P4170, Sigma-Aldrich) and SYBR green (Invitrogen S7563, Thermo Fisher ScienUfic, Carlsbad, CA, USA) with the following procedure: 48 μL sample (live or dead) was mixed in a 96-well plate with 49 μL PBS, 2 μL 0.5 mg/mL PI and 1 μL 100X SYBR green and incubated for 15 min in the dark. Then, a 0.1X samples of both the stained live and the stained dead samples were created by diluUng 10 μL of each sample in 90 μL of PBS. The stained live/dead samples were then analyzed immediately on a CytoFLEX S flow cytometer (Beckman Coulter, Brea, CA, USA) and the CytExpert sosware (v2.4.0.28). The dead samples were used to make the gaUngs required to quanUfy the fracUon of dead cells in the live samples (Table S6). The extra- and intracellular metabolomics samples were analysed by the metabolomics facility at UNIL to quanUfy the five metabolites leucine, glutamate, aspartate, fructose 6-phosphate and cis-aconitate. These metabolites were chosen because they were chemically and metabolically diverse and likely to be quanUfiable both intracellularly and extracellularly based on previous experience with similar samples. Also, as glutamate is the most abundant intracellular metabolite (49, 100), it serves as a best-case scenario in our akempt to test how much of the extracellular metabolite levels that can be explained by cell lysis and the associated release of .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted August 19, 2025. ; https://doi.org/10.1101/2025.08.19.671024doi: bioRxiv preprint intracellular metabolites. Intracellular metabolite samples were first normalized by the amount of protein detected in the same sample and subsequently converted to absolute intracellular concentraUons using the esUmated protein density in E. coli (13.5 ⋅ 10'( µg µm)⁄ ) (101). The fructose 6-phosphate value for sample 1B, Umepoint 1, was discarded as an outlier as it was one order of magnitude larger than any other fructose 6-phosphate measurements. Calcula)on of contribu)on from intracellular metabolites and cell lysis to extracellular concentra)ons Approximate cell volume was esUmated from measured OD600 and a previous esUmate of OD to cell volume for E. coli of 3.6 μL/ΟD600/mL (102). From approximate cell volume, measured absolute intracellular concentraUons and cell lysis fracUons we then calculated the corresponding change in extracellular concentraUons. This potenUal contribuUon was then compared to the net change in extracellular concentraUons between the T0 reference sample (aser 2-hours) and the late exponenUal phase sample. If all T0 values were below the detecUon limit, we assumed the iniUal concentraUon of this metabolite to be 0. If only some values at T0 were below the detecUon limit we imputed these values to the mean of the detected values, giving conservaUve esUmates for the net change in extracellular concentraUons. To esUmate a potenUal contribuUon from protein depolymerizaUon we esUmated cell density in gDW/L from OD600 (0.346 gDW/L/OD600) and from this specific amino acid density (in g/L) from recent esUmates of E. coli amino acid mass fracUons (103). Where mass fracUons were akributed to pairs of amino acids (glutamate/glutamine and aspartate/asparagine) we assumed equal contribuUons. Amino acid density was converted from g/L to μΜ by dividing by the amino acids’ molecular weight subtracted the weight of a water molecule to account for the condensaUon reacUon in protein polymerizaUon. A similar procedure as described above was followed when we extended this analysis by using literature values for intracellular concentraUons. When extracellular concentraUons were reported as two indisUnguishable molecules, we akributed equal amounts to the two metabolites. This includes 2-phosphoglycerate/3-phosphoglycerate and ribulose 5-phosphate/xylulose 5-phosphate. To idenUfy typical biomass degradaUon products we used the metabolites consumed by the biomass equaUon of the E. coli model iML1515 (89), although excluding the soluble pool (as categorized in (87)) since this should correspond to the intracellular concentraUons already accounted for. Of the metabolites in our dataset (Fig. 3B), only the amino acids were among the consumed biomass components and therefore considered to be typical biomass degradaUon products. .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted August 19, 2025. ; https://doi.org/10.1101/2025.08.19.671024doi: bioRxiv preprint .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted August 19, 2025. ; https://doi.org/10.1101/2025.08.19.671024doi: bioRxiv preprint