Metabolite correlation networks reveal complex phenotypes of adaptation-driving mutations Authors: Sharon Samuel, Weronika Jasinska, Ezra Sternlicht, Matvey Nikelshparg, Daniel Kleiner, Shai Pilosof, Shimon Bershtein Published: 2026 ABSTRACT AbstractMetabolic adaptation is often driven by mutations in pleiotropic genes that affect metabolism globally. However, identifying which phenotypes produced by such mutations drive selection remains difficult. Here, we consider metabolite correlation networks as a systems-level phenotype whose emergent properties reflect the combined effects of biochemical reactions and regulation. We hypothesize that mutation-induced changes in the covariance structure of these networks constitute complex phenotypes linking genetic change to bacterial fitness, and that node-level shifts can identify metabolites and pathways mechanistically associated with adaptation. We test this hypothesis by applying adaptive laboratory evolution to a metabolically suboptimal Escherichia coli strain in which metK, encoding methionine adenosyltransferase (MAT), is replaced by an ortholog from Ureaplasma urealyticum. Untargeted metabolite correlation networks reveal that evolution shifts S-adenosylmethionine (SAM), the MAT product, from a peripheral node to a key connector between network clusters. None of the accumulated pleiotropic mutations causing the shift directly affect MAT activity or SAM metabolism, indicating that this transition emerges from system-level metabolic reorganization. Targeted metabolomics of nodes showing similar shifts identifies additional metabolites involved in SAM-related pathways. These findings establish metabolite correlation networks as a framework for mapping genotype–phenotype–fitness relationships and prioritizing metabolic nodes linked to adaptive mechanisms.Subject terms: Molecular evolution, Metabolomics, Biochemical networks Introduction Mapping genotype-phenotype-fitness relationships in evolutionary adaptations is a central problem in evolutionary biology1. The primary obstacle lies in the hierarchical organization of biological systems, which spans multiple levels of increasing complexity from individual macromolecules to cell systems, cells, organisms, and populations2,3. Each level exhibits unique emergent properties that cannot be predicted from lower-level components4–6, such that mutational effects originating in the genome are reshaped as they propagate across the organizational levels7. Identifying which of the phenotypes arising at different levels influence fitness and determine the outcome of selection remains a major challenge1,8–11. Among the cellular systems through which mutational effects propagate, metabolism represents a particularly important intermediate layer connecting genetic variation to fitness8. Its central role in organismal survival, demonstrated across numerous laboratory evolution studies in bacteria10,12–22, makes metabolism a focal point for studying adaptive evolution. Metabolism is carried out by a dense and tightly regulated network of biochemical reactions that catalyze the synthesis and interconversion of metabolites. Under constant conditions, repeated measurements of metabolite abundances reveal highly significant correlations across biological replicates23–27. The interdependency of metabolite concentration is a unique phenomenon not observed among constituents of other cellular systems, such as mRNA transcripts and proteins, and is rooted in the direct conversion of metabolites into one another via biosynthetic pathways23. Early metabolomics studies revealed that only a small subset of metabolite pairs exhibit strong and reproducible correlations across biological replicates, often independent of direct pathway adjacency23. This observation challenged the intuitive expectation that metabolite correlation would simply recapitulate metabolic maps. Subsequent theoretical work demonstrated that correlation patterns arise as emergent properties of the dynamical biochemical system28. Complementary analyses grounded in metabolic control theory further clarified that strong correlations could originate from shared control distribution, mass conservation, chemical equilibrium constraints, or coordinated parameter perturbations, rather than direct substrate–product relationships23. Two main sources drive metabolite correlations: intrinsic metabolic fluctuations caused primarily by stochastic internal changes in cellular components and pathways29–33 and environmental sensitivity, where even micro-perturbations in growth conditions, such as temperature or aeration gradients, affect metabolite levels25,34. Intrinsic noise is largely averaged out in population-level measurements, whereas environmental fluctuations affect entire populations and therefore dominate covariance structure28. Because fluctuations in metabolite concentrations tend to propagate beyond immediate pathway neighbors to distant, co-regulated metabolites, metabolite correlation patterns capture system-wide metabolic coordination and can be leveraged to provide insights into global signatures of physiological state under specific genetic and environmental conditions23,35. Correlated metabolites can be represented as networks, where nodes correspond to individual metabolites, and edges represent pairwise correlations25. The edges are undirected, indicating mutual correlation without implying causation or directionality, and weighted, reflecting the correlation strength. A key feature of these networks is their dynamic nature; they readily adapt in response to genetic and environmental changes, mirroring the system’s current metabolic state23,25. Metabolite correlation networks constitute, therefore, systems-level phenotype arising from the cumulative effects of biochemical reactions and regulatory mechanisms. Consequently, mutation-induced changes in network structure and function can be viewed as complex phenotypic effects of genetic perturbations. Here, we propose that the phenotypic consequences of mutations driving metabolic adaptation can be assessed at a cellular systems-level that integrate their widespread impacts on metabolic structure, flux organization, and regulation. We suggest that metabolite correlation networks represent such a level. This is because correlation patterns between metabolite abundances reflect the integrated influence of direct enzymatic conversions shaping flux magnitude and partitioning together with indirect regulatory processes affecting gene expression and biochemical control. Our central hypothesis is that comparative analysis of non-random and reproducible network covariance reorganization, particularly at the node-level, across distinct evolutionary states enables post hoc association of mutation-driven metabolic reprogramming with underlying molecular mechanisms and fitness changes (Fig. 1). Conceptual framework linking genetic perturbations to systems-level metabolic phenotypes. Bottom panels: Genetic perturbation (right) alters enzyme activity and metabolic reactions relative to the wildtype state (left), reshaping metabolic fluxes and regulatory interactions. Upper panels: These distributed biochemical changes give rise to coordinated shifts in metabolite abundances. Quantifying changes in metabolite correlation networks captures this reorganization as a measurable, complex systems-level phenotype. Network alterations can be assessed at whole-network (global), meso-scale (e.g., node neighborhoods), and individual node levels. Nodes whose roles and connectivity change across evolutionary states generate candidate metabolites for targeted exploration of the molecular mechanisms of adaptation. To test this hypothesis, we analyzed laboratory metabolic adaptation of E. coli populations driven by an inefficient enzyme. In natural settings, inefficient enzymatic activities arise in bacterial populations through diverse genetic and ecological processes, including inactivating mutations19,36–39, environmental changes favoring promiscuous enzymatic activities40,41, or horizontal gene transfer events introducing poorly optimized enzymes22,42,43. Adaptive mutations under such selection regimes rarely occur in the inefficient enzyme itself; instead, they typically affect other metabolic or regulatory genes, often as loss-of-function (LOF) mutations19,36,38,44–46. These off-target mutations frequently involve pleiotropic genes (genes affecting multiple traits), such as those encoding RNA polymerase subunits, translation termination release factors, or metabolic enzymes controlling cellular redox potential36,47,48. Therefore, adaptation to an inefficient enzyme offers a powerful model for testing whether reproducible reorganization of metabolite correlation networks can be causally associated with genetic changes and fitness improvements. To examine how metabolic adaptation reshapes correlation networks, we engineered a metabolically suboptimal E. coli strain by replacing the endogenous metK gene encoding methionine adenosyltransferase (MAT) with a poorly adapted ortholog from U. urealyticum49, and subjected the strain to laboratory evolution. Analysis of metabolite correlation networks with 2,118 metabolites constructed from untargeted metabolomics revealed extensive strain-specific global rewiring of network structure across evolutionary states, linking mutation-driven changes to growth-rate differences used here as a proxy for fitness. Notably, evolution shifted the node representing S-adenosylmethionine (SAM), the product of the MAT reaction, from a peripheral position to a central connector between network clusters. Using targeted metabolomics, we predicted the metabolic identity of selected individual nodes whose roles and properties in the untargeted metabolite correlation networks shifted during adaptive evolution in a manner similar to the SAM node. We found that a substantial fraction of the predicted metabolites participates in SAM-related metabolic pathways and global metabolic regulation, thereby prioritizing testable mechanistic hypotheses for metabolic adaptation driven by complex mutational effects. Results Adaptation to an inefficient enzyme involves off-target pleiotropic mutations that reshape metabolism and enhance fitness To model inefficient enzyme evolution in laboratory settings, we replaced the chromosomal metK gene in E. coli with an ortholog from U. urealyticum, while retaining the endogenous E. coli promoter (Fig. 2A). The metK gene encodes methionine adenosyltransferase (MAT), an essential enzyme responsible for synthesizing S-adenosylmethionine (SAM) from ATP and methionine50. SAM is involved in numerous biochemical processes, including the methylation of DNA, proteins, RNA, and small molecules, the synthesis of enzyme cofactors, and the production of polyamines51. The experimental system. A Gene editing and adaptive laboratory evolution of E. coli strains. The panel was created in BioRender. Samuel, S. (2026) https://BioRender.com/e91kkktB Growth curves and growth rates (in 1/hour) of WT, ‘engineered’, and ‘mut’ strains as well as ‘evolved’ and ‘evolved-WT’ populations. The shaded area represents the standard deviation across four independent measurements. C– D Data collection and construction of metabolite correlation networks. C A pipeline for processing untargeted metabolomics data for constructing metabolite correlation networks. Raw chromatograms from LC-MS analysis of 19 replicates per batch (across two batches) were processed using Compound Discoverer software (Thermo Scientific), generating data files containing the relative abundances of 2,118 individual intracellular metabolic features (Met_1, Met_2, …, Met_n) in the different samples (S_1, S_2,…, S_m*2, when m represents the number of samples in each experimental batch) along with their corresponding mass-to-charge (m/z) ratios and retention times. For each strain, two pairwise correlation matrices (one per batch) were created using Pearson correlations between relative abundances of metabolite pairs across 19 replicates and converted into correlation networks. These networks were then filtered using the PCLRC algorithm to reduce non-specific (transitive) interactions. Finally, a single network was constructed, containing only the links shared by both batch networks. D Principal component analysis (PCA) of the relative abundances of untargeted metabolic features extracted from both batches. Each data point represents a biological replicate within a batch. The plot was generated using MetaboAnalyst 5.0. Source data are provided as a Source Data file. While both E. coli MAT (EcMAT) and U. urealyticum MAT (UuMAT) display similar enzymatic constants, UuMAT is substantially more sensitive to SAM inhibition, and its catalytic efficiency drops by more than half in the presence of the intracellular SAM levels of E. coli49. The E. coli strain with the orthologous metK replacement (hereafter termed the ‘engineered’ strain) showed a severe growth rate reduction from 1.3 h⁻¹ to 0.64 h⁻¹ (One-way ANOVA across strains: F4,5 = 1133, p = 1. 41 ⋅ 10−7, followed by Tukey’s HSD, p < 0.001) (Fig. 2B, Supplementary Table 1), a metabolic imbalance characterized by a significantly decreased SAM and increased methionine levels (One-way ANOVA across strains: SAM, F4,185 = 168.2, p < 2 ⋅ 10−16, followed by Tukey’s HSD, p < 2 ⋅ 10−7; methionine, F4,185 = 434.6, p < 2 ⋅ 10−16, followed by Tukey’s HSD, p < 2 ⋅ 10−7) (Supplementary Fig. 1A), and an 80% decrease in the intracellular MAT abundance (Supplementary Fig. 1B). We subjected both the ‘engineered’ and wild-type (WT) E. coli strains to laboratory evolution by serial passaging for 320 generations, resulting in the ‘evolved’ and ‘evolved-WT’ populations, respectively (Fig. 2A). The growth rate of the ‘evolved’ population increased significantly from 0.64 h⁻¹ to 1.196 h⁻¹ (Tukey’s HSD, p < 0.001), though still below the WT rate (1.3 h⁻¹). Similarly, the ‘evolved-WT’ population showed a growth rate improvement from 1.3 h⁻¹ to 1.6 h⁻¹ (Tukey’s HSD, p < 0.001) (Fig. 2B, Supplementary Table 1). To identify the accumulated mutations, we isolated three random variants from both the ‘evolved’ and ‘evolved-WT’ populations from the last serial passage (generation 320), validated that their growth rates closely match the average growth rates of their respective populations (Supplementary Fig. 2), and performed whole genome sequencing (Table 1). There were no shared mutations between the two sets, indicating that the metK replacement profoundly reshaped the mutational fitness landscape of the ‘engineered’ strain. All ‘evolved-WT’ clones carried a known 82 bp deletion in the rph-pyrE operon, restoring pyrimidine synthesis17,52,53, whereas all ‘evolved’ clones shared two fixed mutations: a metK promoter mutation and a nonsense mutation in the pleiotropic pntA, inactivating transhydrogenase and shifting NADH/NADPH balance36,54. Other mutations, such as in pleiotropic genes rpoB (encoding RNA polymerase subunit beta47) and prfC (encoding peptide chain release factor 348), appeared sporadically and have been seen in other lab-evolution studies (referenced in Table 1), suggesting potential adaptive roles even though they did not reach fixation. Table 1Mutations found in isolated variants‘evolved’ population‘evolved-WT’ populationisolated variantsisolated variantsGeneMutationv1v2v3v1v2v3Ref.pntANonsense CAA → TAA (Q322STP)✓✓✓36,54,87–89metKIntergenic (−8) T → A✓✓✓This workprfCMissense GAT → AAT (D222N)✓53,90fesIntergenic (−89) G → T✓89,91yagEIntergenic (−70) ΔA✓rpoBMissense CAC → TAC (H526Y)✓17,47,53,87flgBIn frame deletion ΔTAACAGCCT (1131349)✓92sufSFrameshift deletion ΔT (1759368)✓88,93rph-pyrEΔ82 bp (3815848)✓✓✓17,52,53flhDFrameshift insertion C (1978164) → CGGTG✓✓94,95insAB1Gene duplication (1978500-1979250)✓✓paoCMissense GAT → GTT (D13V)✓cadASynonymous TCC → TCT (S331)✓ Mutations found in isolated variants For further analysis, we have arbitrarily chosen ‘evolved v1’ and ‘evolved-WT v1’ variants (hereafter, ‘evolved’ and ‘evolved-WT’ strains, respectively). We found that evolution significantly increased the intracellular UuMAT abundance in the ‘evolved’ strain by ~1.4-fold (two-sided pairwise t-test with Bonferroni adjustment, p < 0.001) (Supplementary Fig. 1B) and partially alleviated metabolic imbalance by significantly raising SAM (Tukey’s HSD, p < 0.001) and reducing methionine levels (Tukey’s HSD, p < 0.001) (Supplementary Fig. 1A). In principle, a recurrent mutation in the metK promoter might solely account for these changes by boosting UuMAT expression, potentially rendering other accumulated mutations inconsequential to the observed phenotypic and growth improvements. To isolate the effects of the metK promoter mutation, we reintroduced that single mutation into the ancestral background (‘mut’ strain). Although ‘mut’ matched the ‘evolved’ strain in UuMAT abundance (Supplementary Fig. 1B), its growth rate remained slightly lower (1.173 h⁻¹ vs. 1.196 h⁻¹), whereas its lag phase grew significantly longer (62 min vs. 53 min; One-way ANOVA across strains: F4,5 = 199.3, p = 1.07 ⋅ 10−5, followed by Tukey’s HSD, p < 0.05) (Fig. 2B, Supplementary Table 1). Moreover, ‘mut’ accumulated significantly more SAM than the evolved strain (Tukey’s HSD, p < 0.001) despite indistinguishable methionine levels, yielding a significantly higher SAM/methionine ratio (One-way ANOVA across strains: F4,5 = 463.4, p < 2 ⋅ 10−16, followed by Tukey’s HSD, p < 0.001) (Supplementary Fig. 1A). Thus, although the mutation in the metK promoter clearly demonstrates a strong adaptive value and has likely reached fixation through a selective sweep during the early stages of evolution, additional mutation(s) that have accumulated in the ‘evolved’ variant (Table 1) further modified the metabolic phenotype and enhanced the fitness of the ‘evolved’ strain. In contrast to the ‘evolved” strain, in the ‘evolved-WT’ strain, EcMAT abundance and SAM levels both rose significantly (EcMAT: two-sided pairwise t-test with Bonferroni correction for multiple comparisons, p = 0.0194; SAM: Tukey’s HSD, p = 0.003), but methionine levels also increased (Tukey’s HSD, p = 0.03), keeping the SAM/methionine ratio at wild-type levels (Supplementary Fig. 1). Because none of the ‘evolved-WT’ mutations overlap with those in the ‘evolved’ strain, the shifts in MAT and SAM likely reflect broader metabolic and regulatory changes driven by its distinct mutational trajectory (Table 1). Collectively, these results underscore that the mechanisms and outcomes of metabolic adaptation depend strongly on a strain’s initial metabolic state. Strains show distinct metabolic states, separating evolution and genome editing Bacterial populations propagated under seemingly identical laboratory conditions often display significant variability in fitness across replicates, even when fitness is measured with high precision34. This variability is further intensified when measurements are taken on different days, a phenomenon known as the ‘batch effect’34. Fitness variability is primarily attributed to physiological sensitivity to subtle environmental perturbations that inevitably exist between replicates in laboratory settings, due to minor temperature gradients in the incubator, slight differences in vessel shape, and variations in shaking speed. Given the highly dynamic nature of the metabolome and its hyper-responsiveness to environmental conditions, minuscule environmental changes are likely to induce even greater variability in metabolite levels and, potentially, reveal metabolite correlation patterns23. Crucially, adaptive mutations introduce strain-specific changes in enzyme activities, regulatory interactions, redox balance, and resource allocation, thereby reshaping how each genetic background responds to the same micro-environmental fluctuations. As a result, identical environmental variability can produce distinct metabolite correlation patterns across strains, reflecting mutation-induced rewiring of metabolic control and flux coupling. In this sense, environmental variability acts as a probe that exposes underlying metabolic dependencies. At the same time, biological variability across replicates and batches can introduce correlations that are not reproducible and do not reflect stable system-level relationships. To account for the metabolic variability and identify reproducible metabolite correlation patterns, we implemented the following strategy. First, we increased the number of replicates in a single batch experiment from the standard ten replicates24 to nineteen. Second, we conducted two such batch experiments on different dates, resulting in a total of 38 replicates per strain. We consider correlation reproducible across replicates and independent batches as stable features of metabolic organization, whereas non-reproducible correlations likely reflect hypersensitivity to minor environmental fluctuations. We applied this strategy to all strains. Each replicate was grown to a mid-exponential phase and subjected to polar intracellular metabolite extraction and LC-MS metabolic profiling under positive ionization mode (“Methods”). The processing of raw chromatograms produced data files containing the relative abundances of all individual untargeted metabolic features. In each replicate, we identified 2,118 identical metabolic features (Fig. 2C, Supplementary Fig. 3, Supplementary Data 1). Next, we performed Principal Component Analysis (PCA) on the metabolite abundance data separately for each batch and found a clear strain-specific clustering, although differences between the batches were also apparent. For example, the ‘evolved-WT’ and ‘evolved’ strains were more distinctly separated along both PC1 and PC2 in batch 1 compared to batch 2 (Supplementary Fig. 4). However, when we combined the data from both batches into a single dataset, the PCA showed a sharp, strain-specific group separation, with strong clustering of replicate data points from both batches (Fig. 2D). The largest shift along PC1 was caused by the orthologous metK replacement (from the WT to the ‘engineered’ strains). Evolutionary adaptation (‘evolved’ strain), or the introduction of a single mutation in the metK promoter (‘mut’ strain) shifted the data points back towards the WT strain. Notably, the better growing ‘evolved’ strain moved closer to the WT than the ‘mut’ strain, suggesting that PC1 is likely to report on resource consumption or growth rate. Conversely, the evolutionary adaptation of the ‘evolved-WT’ strain, which accumulated a completely different set of mutations compared to the ‘evolved’ strain, pushed the data points away from the WT along PC2, indicating that PC2 captures an additional adaptive property that is at least partially independent of PC1 (Fig. 2D). Taken together, the PCA analysis demonstrates that each strain represents a distinct metabolic state, where differences in relative metabolite abundance patterns reflect the evolutionary relationship between strains. Combining the metabolite abundance data from two batch experiments further strengthens these distinctions. Metabolite correlation networks reveal strain-specific rewiring while preserving a core of robust correlations Having established that each strain is characterized by a unique metabolite abundance pattern, we constructed metabolite correlation networks by converting metabolite abundances into correlation matrices using Pearson’s correlations. Separate matrices were generated for each experimental batch and filtered using the Probabilistic Context Likelihood of Relatedness (PCLRC) algorithm55 to reduce non-specific transitive correlations25. To identify correlations robust to subtle environmental perturbations, we retained only those observed in both batches, resulting in a single overlap network per strain (Fig. 2C). Across all strains, the Jaccard similarity of the correlations between both batches was low (0.126–0.156, Fig. 3A), underscoring the strong influence of environmental conditions on metabolite correlation patterns. Comparison of the resulting overlap networks between strains also revealed low levels of shared links, ranging from 0.098 (between the WT and ‘engineered’ networks) to 0.19 (between the ‘evolved’ and ‘mut’ networks) (Fig. 3A), indicating extensive strain-specific rewiring. Jaccard similarity analysis and hierarchical clustering of the metabolic networks. A Yellow cells indicate similarity between the two batch networks within each strain, while white cells represent similarity between the overlap networks of different strains. The red dendrogram shows unsupervised hierarchical clustering based on the Jaccard similarity values. B Unsupervised hierarchical clustering based on the Jaccard similarity values of the individual strain-specific batches. Values on the left indicate Jaccard dissimilarity. Unsupervised hierarchical clustering analysis of the Jaccard similarity values always grouped both strain-specific batches together (Fig. 3B). It also grouped the WT and ‘evolved-WT’ overlap networks and the ‘evolved’ and ‘mut’ overlap networks, while positioning the ‘engineered’ overlap network separately (Fig. 3A,B). This clustering recapitulated the hierarchy of strain separation observed in the PCA (Fig. 2D), indicating that network-level differences reflect the underlying metabolic state differences rather than analytical artifacts. Because metabolite correlation networks are constructed from replicate-to-replicate covariance patterns, we explicitly evaluated whether differences in network structure could arise from technical variability or unequal variance across strains. Metabolite abundance profiles revealed strong separation by strain identity (PERMANOVA; F = 642.95, R² = 0.97, p = 0.001), with substantially smaller effect sizes for within-strain between-batch variability than for between-strain comparisons (Supplementary Data 2). Similarly, Jaccard distance analysis indicated that strain identity (F = 1.99, R² = 0.54, p = 0.002) explained more variation in network dissimilarity than batch (F = 2.9, R² = 0.19, p = 0.002). Together, these results indicate that strain-specific biological differences, rather than technical noise, dominate both abundance and network-level variation. To assess whether the observed inter-strain similarities exceeded random expectation, we generated 1,000 degree-preserving randomizations of each overlap network. The observed Jaccard similarities were significantly greater than those of randomized networks (one-sided upper-tail z-test, p « 10−4) (Supplementary Fig. 5), indicating that shared correlations between the empirical networks are not attributable to chance. Therefore, despite the massive rewiring of correlation networks caused by genetic changes, there remains a common robust core of correlations that withstood genetic perturbations. Specifically, the ‘evolved-WT’ strain retained 36% of the correlations found in the ancestral WT strain, and the ‘evolved’ strain preserved 25% of the correlations detected in the ancestral ‘engineered’ strain (Fig. 4). The ‘mut’ strain, carrying the single mutation in the metK promoter, and the ‘evolved’ strain carrying, in addition to the metK promoter mutation, four off-target mutations (Table 1), shared only 34.3% and 29.9% of the correlations, respectively, indicating the strong effect of the off-target mutation(s) on network structure. Shared positive and negative correlations between strains. Notably, 2772 correlations shared between the ‘evolved’ (9% of correlations) and ‘evolved-WT’ (8.5% of correlations) were absent from the WT network, suggesting partially convergent metabolic adaptation under identical environmental conditions. Across all five strains, 2477 correlations (4.7% of WT correlations) formed a shared core (Fig. 4). Taken together, these results show that metabolite correlation networks undergo extensive strain-specific rewiring during genetic perturbation and evolution, yet retain a conserved core of robust covariance structure shared across strains. Adaptation to inefficient MAT restructured the covariance relationships of SAM As described above, the ‘engineered’ → ‘evolved’ transition led to a significant increase in intracellular SAM levels (Supplementary Fig. 1A). Although this increase can be partly explained by the promoter mutation in metK (Table 1), the significant difference in SAM abundance and SAM/methionine ratios observed between the ‘evolved’ and ‘mut’ strains (Supplementary Fig. 1A) suggests an important role of the off-target mutation(s) in the metabolic adaptation of the ‘evolved’ strain (Table 1). We therefore asked (i) whether the altered SAM abundance was accompanied by specific changes in the covariance relationships of the SAM node within the metabolite correlation network; and, if so, (ii) whether these changes reflected focal perturbations centered on SAM or, alternatively, were embedded within broader network reorganization. To address these questions, we considered separately positive and negative correlations23–26,28 and found that positive correlations dominate our networks, accounting for 78.7% to 88.6% of all correlations (Supplementary Table S2). The predominance of positive correlations is an expected emergent property of metabolic systems, supported by previous studies (see Supplementary Note I), which also reported similar proportions of correlation signs56–58. To quantify the overall difference between covariance structures of the metabolite correlation networks, we performed a global-level analysis using a suite of network metrics (see Supplementary Note I, for the detailed analysis). We found substantial differences between the ‘evolved’ and ‘mut’ networks. Relative to the ‘mut’ network, the ‘evolved’ network contained fewer positive and negative correlations. The ‘evolved’ network also showed significantly lower positive correlation weights (two-sided Wilcoxon rank-sum test with Benjamini–Hochberg false discovery rate (BH-FDR) correction, p < 3 × 10⁻⁸), reduced hub connectivity (t-test, p = 0.017), and lower mean neighborhood connectivity (two-sided t-test, p < 10⁻⁸) (Supplementary Note II). In addition, ‘small-world’ characteristics, which reflect the tendency of a network to combine locally cohesive groups with sparse long-range connections, were disrupted by the WT → ‘engineered’ transition but were more effectively restored in the ‘evolved’ network than in the ‘mut’ network (see Supplementary Note III). Together, these results indicate that the off-target mutation(s) present in the ‘evolved’ strain reshaped the broader covariance structure in which the SAM node is embedded. Next, we conducted the meso-scale network analysis (Fig. 1). For each strain, we constructed bilayer networks separating positive and negative correlations into distinct layers and partitioned these networks into clusters of nodes called ‘modules’ using the Infomap algorithm (see “Methods”). Although overall distributions of module sizes were broadly similar across networks (Fig. 5A,B), we detected substantial differences in the sizes of large and medium modules, as well as in their correlation-sign composition and node-level properties (Fig. 5A,B and Supplementary Data 3). During the ‘engineered’ → ‘evolved’ transition, the largest module decreased from 521 to 344 nodes (Fig. 5A). This reduction in module size was accompanied by significant increases in the average clustering coefficient, which reflects the local cohesiveness of module nodes, from 0.50 to 0.53 (two-sided Wilcoxon rank-sum test with continuity correction, p = 0.002); in average betweenness centrality, which estimates the extent to which nodes lie on shortest paths through the network, from 0.0017 to 0.0019 (two-sided Wilcoxon rank-sum test with continuity correction, p = 0.048); and in the average number of correlations per node, or node degree, from 82.2 to 86.6 (two-sided Wilcoxon rank-sum test with continuity correction, p < 2.2 × 10⁻¹⁶). Notably, the largest modules in both the ‘engineered’ and ‘evolved’ networks consisted entirely of positive correlations. Furthermore, the four largest modules in the ‘evolved’ network were approximately 13–28% smaller than their counterparts in the ‘mut’ network, underscoring the contribution of off-target mutation(s) to the observed meso-scale restructuring (Fig. 5A,B and Supplementary Data 3). Together, these results suggest that ‘engineered’ → ‘evolved’ adaptation reorganized the network into smaller, yet more internally cohesive modules, while preserving or enhancing intramodular connectivity. Change in node roles upon genetic engineering and adaptive evolution. A Distribution of network module sizes (number of nodes). Only modules with 10 nodes or more are shown. Arrows point to modules that include SAM (only SAM nodes involved in positive correlations are shown). B Cumulative distribution of module sizes. C Normalized Mutual Information (NMI) analysis of the module composition among networks. D Node roles. Role definitions are accompanied by a toy example. Each dot represents a node with an assigned role according to its connectivity within (Z) and between (C) modules. Color-coded numbers indicate the relative proportion of nodes with the assigned role. Nodes representing SAM involved in positive correlations are depicted as yellow diamonds. Source data are provided as a Source Data file. To further quantify structural differences, we compared module composition across strains using Normalized Mutual Information (NMI; see “Methods”). Pairwise NMI values ranged from 0.50 to 0.56 (Fig. 5C), indicating only moderate overlap in module structure and suggesting substantial reorganization of node assignments. Despite this broad reorganization, SAM remained assigned to the largest module in both the WT → ‘engineered’ and WT → ‘evolved-WT’ transitions (Fig. 5A). In contrast, during the ‘engineered’ → ‘evolved’ transformation, SAM relocated to a substantially smaller module (Fig. 5A). SAM, however, remained in a larger module in the ‘mut’ network, further underscoring the distinct structural impact of the off-target mutation(s) present in the ‘evolved’ strain (Fig. 5A). The relocation of SAM suggests that adaptive mutations may alter the structural roles of individual metabolites. To systematically identify such changes, we classified all nodes based on within-module degree (Z), reflecting connectivity within their module, and participation coefficient (C), reflecting the distribution of links across modules59 (see “Methods”). Nodes were categorized as peripheral (low Z, low C; weakly connected within their module), connectors (low Z, high C; linking multiple modules), module hubs (high Z, low C; highly connected within a module), or network hubs (high Z, high C; highly connected both within and across modules). Across all networks, the proportions of peripheral nodes dominated (79.4–83.1%), followed by connectors (15.7–21%), while module and network hubs were rare (9–16% collectively) (Fig. 5D) We determined that SAM shifted from a peripheral to a connector role during the ‘engineered’ → ‘evolved’ transition (positive correlations), assuming the connector role exclusively in the ‘evolved’ network, again underscoring the substantial impact of the off-target mutation(s). This transition was accompanied by a seven-fold decrease in node degree, an approximately two-fold decrease in clustering coefficient, and a five-fold reduction in neighborhood connectivity, which quantifies the average degree of a node’s neighbors (Fig. 6A–C). In contrast, SAM’s betweenness centrality, which is determined by the number of shortest paths passing through the node, increased by at least ten-fold in the ‘evolved’ network compared to all other strains (Fig. 6D). Conversely, the clustering coefficient and betweenness centrality of SAM in the ‘mut’ strain more closely resembled those in the ‘engineered’ network than in the ‘evolved’ one (Fig. 6C, D), highlighting the contribution of the off-target mutation(s) to the metabolic role of SAM. SAM node properties in the correlation networks (for positive correlations). A Node degree. B Neighborhood connectivity. C Clustering coefficient. D Betweenness centrality. Source data are provided as a Source Data file. Collectively, these findings suggest that the ‘engineered’ → ‘evolved’ transition not only relocated SAM from the largest to a smaller module but also reduced its local centrality by severing interactions with tightly connected neighborhoods. Concurrently, SAM’s role in linking distinct modules became more pronounced in the ‘evolved’ network, potentially suggesting altered coordination of SAM metabolism across the broader metabolic system. Thus, adaptation driven by off-target mutation(s) in the ‘evolved’ strain affected not only intracellular SAM abundance, but also the covariance relationships of SAM within the metabolite correlation network. Moreover, this focal restructuring of SAM covariance occurred within a globally reorganized network background. To determine whether the properties of SAM node in the ‘evolved’ network are a general adaptive property of the evolving population and not merely an adaptation of an arbitrarily chosen single variant, we repeated the analysis at the whole population level using the same evolutionary passages from which the individual clones were isolated (see “Methods“, Supplementary Data 4, Supplementary Table S3 and Supplementary Note IV) We found that meso- and node-scale properties closely recapitulated the single-clone results, including module size distributions (Supplementary Fig. 6A), module composition similarity (Supplementary Fig. 6B), SAM relocation to a smaller module (Supplementary Fig. 6A), SAM role shift (Supplementary Fig. 7A), and a decrease in the SAM clustering coefficient and an increase in betweenness centrality (Supplementary Fig. 7B), highlighting a shared mechanism of adaptation between distinct genetic clones in the population. These results support the mechanistic generality of our conclusions and demonstrate the robustness of the proposed experimental–analytical framework. Shifts in metabolite roles enable identifying candidate metabolites for targeted mechanistic analysis The pronounced shift of SAM from a peripheral to a connector node indicates that changes in node properties can serve as informative signals for identifying metabolites for targeted mechanistic investigation. Structural reassignments within the network potentially reflect redistribution of metabolic coordination, suggesting that metabolites undergoing similar shifts may represent key adaptive targets (Fig. 1) We, therefore, systematically screened the ‘evolved’ network for metabolites that exhibited changes similar to those observed for SAM. Specifically, we identified metabolites that met one or more of the following criteria: (1) a unique shift from peripheral to connector in the ‘evolved’ network; (2) a ≥ 5-fold increase in betweenness centrality in the ‘evolved’ relative to the ‘WT’ network (comparable to the change observed for SAM); and (3) a ≥ 40% reduction in clustering coefficient in the ‘evolved’ relative to the ‘WT’ network (similar in magnitude to that of SAM). In total, 365 metabolites met at least one of those criteria, 67 met two, and 11 fulfilled all three (Supplementary Data 5). Using deep neural network–based predictions of high-resolution fragmentation spectra from LC-MS/MS data60–63, we successfully assigned putative identities to 18 of these nodes (Supplementary Data 5; see “Methods”). Among these were SAM and 5-methylthioadenosine (MTA), whose identities were validated using authentic standards (“Methods”). Out of 18 predicted metabolites, five participate in purine biosynthesis (Supplementary Table S4). Specifically, L-glutamine and SAICAR serve as a substrate and an intermediate, respectively, in the purine biosynthetic pathway64. Xanthosine functions as a precursor in the purine salvage pathway65,66. 5,10-Methenyltetrahydrofolate is a direct precursor of 10-formyltetrahydrofolate, which donates one-carbon units required for purine synthesis67. S-hydroxymethyl glutathione acts as a precursor of formate68, which contributes to an alternative route for the production of 10-formyltetrahydrofolate69. SAM is part of a tightly interconnected network of metabolic pathways involved in the synthesis of purines, methionine, and polyamines51,67,70. We found that five of the predicted metabolites (L-glutamine, SAICAR, 5,10-methenyltetrahydrofolate, xanthosine, and S-hydroxymethyl glutathione) participate in de novo or salvage purine biosynthesis, two (L-threonine and L-homoserine phosphate) are involved in regulation of de novo methionine synthesis, and three (palmitoylputrescine, N-acetylputrescine, and methylthioadenosine) are linked to polyamine biosynthesis. Three additional metabolites, enterobactin, cyclic AMP (cAMP), and γ-glutamylcysteine, play global roles in E. coli metabolism, physiology, and growth (Supplementary Table 4 and Supplementary Note V). Overall, these results demonstrate that metabolite correlation-network analysis can identify nodes whose properties shift under adaptive pressure and, when combined with computational prediction of their metabolite identities, can prioritize candidate metabolites and pathways for subsequent mechanistic investigation of metabolic adaptation. Discussion Understanding how mutations translate into fitness-relevant phenotypes remains a central challenge in evolutionary biology because mutational effects propagate through multiple organizational layers while influencing organismal performance. Here, we addressed this challenge by examining metabolic adaptation through the lens of metabolite correlation networks, which capture systems-level coordination across metabolic pathways. Using laboratory evolution of a metabolically suboptimal E. coli strain combined with untargeted metabolomics, we reconstructed correlation networks across evolutionary states and analyzed how their structure reorganizes during adaptation. Adaptive mutations drive extensive rewiring of metabolite correlation networks while selectively restoring the balance between local clustering and global connectivity disrupted by genetic perturbation. In parallel, shifts in the network roles of specific metabolites, most notably S-adenosylmethionine (SAM), identify metabolic nodes whose altered connectivity provides mechanistic clues to the underlying adaptive processes. Importantly, the correlation-based framework we suggest provides advantages over mechanistic metabolic models in evolutionary contexts driven by pleiotropic mutations. Mechanistic approaches, including Metabolic Flux Analysis (MFA)36,71, are highly effective when the affected pathways are known and the underlying reaction topology remains stable. However, adaptive mutations, particularly in global regulators or redox-related genes can simultaneously perturb hundreds of reactions, modify effective pathway connectivity, and activate latent metabolic routes in ways that are difficult to anticipate. Under such conditions, assumptions of fixed topology and predefined metabolite coverage may no longer hold. In our system, mutations in pntA, prfC, and rpoB (Table 1) likely produced widespread perturbations in the metabolic network, triggering shifts in global metabolic states, activating latent pathways, and redistributing fluxes across pathways. For example, LOF mutations in pntA are known to alter the NADPH/NADP⁺ ratio, influencing numerous cellular reactions36. This shift could thermodynamically favor the reverse direction of certain NADP(H)-dependent reactions72 and modulate (inhibit or activate) enzymes in key branch points of the metabolic pathways36,73. Such systemic effects would require extensive remodeling of mechanistic models and expansion of targeted metabolite sets71. In contrast, metabolite correlation networks constructed from untargeted metabolomics provide topology-agnostic, system-wide coverage and enable detection of mutation-induced metabolic reorganization without prior knowledge of affected pathways. Although individual correlations do not imply causality, comparative network analysis across evolutionary states reveals reproducible structural signatures and highlights metabolites whose shifting network roles generate testable mechanistic hypotheses. An important added value of the correlation network analysis lies in providing an unbiased prioritization of metabolites and pathways. Metabolites are prioritized not simply because they belong to a known pathway or because their abundance changes, but because their position and role within the metabolite covariance structure are reorganized in a manner associated with the adaptive transition. We demonstrate that this strategy reduced thousands of untargeted metabolic features to a smaller manageable set of candidate nodes for targeted identification. The recovery of metabolites involved in SAM-related pathways supports the biological relevance of this network-based prioritization, whereas the identification of metabolites such as enterobactin, cAMP, and γ-glutamylcysteine nominates additional physiological processes, such as iron acquisition, global carbon regulation, and redox/glutathione metabolism, that may be coupled to adaptation but would not necessarily have been selected in a narrowly targeted SAM-centered analysis. Moreover, in many adaptive scenarios, such as adaptation to a novel environment, pleiotropic mutations may play major roles without a clearly defined metabolic bottleneck or focal metabolite. In such cases, correlation networks may be particularly useful for prioritizing metabolites, pathways, and physiological axes for mechanistic follow-up. The extensive hub loss, correlation rewiring, module fragmentation, and changes in path length observed during genome editing and adaptive evolution indicate that modular organization is a central network property undergoing evolutionary change in our system. In network theory, modularity has often been proposed as a key contributor to robustness in biological systems74,75, as compartmentalization of interactions can limit the propagation of stochastic perturbations and stabilize metabolite concentrations. However, model-based analyses of biochemical networks demonstrate that modularity does not universally enhance metabolic robustness, but instead introduces a trade-off between buffering stochastic concentration noise and buffering functional perturbations such as LOF mutations or regulatory rewiring76. While increased modularity can isolate local disturbances, it may simultaneously reduce compensatory flux redistribution across the system, thereby increasing vulnerability to functional damage. Directly interpreting evolutionary changes in metabolite correlation network properties in terms of metabolic robustness is nontrivial. For example, the link between modularity changes and functional resilience of metabolic fluxes remains indirect. Furthermore, adaptive evolution in our system involves pleiotropic mutations that simultaneously perturb regulation, redox balance, and pathway activity, complicating the separation of functional adaptation from stochastic variability. Thus, although metrics such as modularity, clustering, and path length describe global metabolic reorganization, their translation into specific robustness principles requires further mechanistic investigation. The emergence of S-adenosylmethionine (SAM) as a connector node with elevated betweenness centrality in the ‘evolved’ network illustrates how correlation-network reorganization can be linked to fitness and point to mechanistic hypotheses. In the ‘evolved’ strain, SAM relocated to a smaller module and shifted from peripheral to connector, while exhibiting reduced clustering and increased betweenness. This pattern suggests reduced local embedding but enhanced cross-neighborhood integration. Because both the ‘mut’ and ‘evolved’ strains share the metK promoter mutation and exhibit comparable methionine levels, differences in SAM network role likely reflect altered utilization rather than production. Lower SAM abundance in the ‘evolved’ strain despite similar methionine levels is consistent with improved SAM consumption efficiency. This hypothesis can be tested experimentally through controlled metabolic perturbations or flux analysis. Although the initial network analyses were performed on single representative isolates from evolved populations, whole-population metabolomics from the same evolutionary passages reproduced the key meso-scale and node-level findings, indicating that the observed network reorganization reflects shared adaptive phenotypes rather than clone-specific idiosyncrasies. To conclude, we present an experimental and analytical framework that leverages metabolite correlation networks as systems-level phenotypes linking genotype to fitness. This framework integrates untargeted metabolomics with network reconstruction, structural comparison across evolutionary states, and identification of metabolites whose altered roles nominate them for targeted mechanistic interrogation (Fig. 1). By capturing mutation-induced reorganization of metabolic coordination without assuming fixed biochemical topology, our approach provides a practical strategy for connecting adaptive mutations to functional metabolic hypotheses. Beyond the specific case of inefficient enzyme adaptation, our findings point to a broader principle: adaptive evolution can reorganize systems-level metabolic structure to restore or maintain functionality following perturbation. Testing the generality of this structural restoration principle across diverse evolutionary contexts, together with complementary measurements and mechanistic approaches such as metabolic flux analysis, perturbation experiments, and quantitative metabolic modeling, should help establish how specific mutational effects give rise to the observed covariance structure and clarify how network-level organization contributes to robustness and adaptation. Methods Culture medium preparation Bacterial strains were cultivated in a minimal M9 medium supplemented with casamino acids. The medium contained M9 salts (15 g/L KH₂PO₄, 2.5 g/L NaCl, 33.9 g/L Na₂HPO₄, 5 g/L NH₄Cl), casamino acids (0.1%), thiamine (0.5 μg/L), magnesium sulfate (1 mM), and D-glucose (2 g/L). The final volume was adjusted with sterilized water. M9 salts were obtained from Formedium (Swaffham, UK). Thiamine (99%) and Bacto Casamino Acids were purchased from Thermo Fisher Scientific (St. Louis, MO), while D-glucose ( ≥ 99.5%) and magnesium sulfate ( ≥ 99.5%) were sourced from Sigma-Aldrich (St. Louis, MO). The medium was sterilized by autoclaving. Genome editing and experimental evolution In this study, wild-type (WT) E. coli K-12-MG1655 strain was employed as the model organism. A replacement of the essential metK gene encoding methionine adenosyltransferase (MAT) with an ortholog from Ureaplasma urealyticum was conducted in two steps. First, a Tn7 transposon-based system was used to introduce the coding sequence of orthologous metK under the control of the endogenous E. coli metK promoter into the genome of E.coli77. Once the orthologous metK gene was successfully integrated, a λ Red recombination system78 was used to delete the endogenous E. coli metK. The whole genome sequencing validated that the resulting ‘engineered’ strain was genetically identical to the WT except for the altered gene. Subsequently, the WT and unevolved bacterial populations were subjected to adaptive laboratory evolution by serial passaging. The populations were grown in 96-well microtiter plates and propagated by 1/100 dilution into fresh media every 24 h. A total of 48 passages produced ‘evolved-WT’ and ‘evolved’ populations. An additional ‘mut’ strain was produced by incorporating a single (−8) T to A metK promoter mutation using a λ Red recombination system on the ‘engineered’ strain genetic background. Each of the listed bacterial strains was stored in a glycerol stock at −80 °C for further analysis. Growth rate evaluation To measure the growth rates of bacterial strains, a 5 mL starter culture of each strain was grown overnight at 37 °C in the culture medium. The overnight cultures were subsequently diluted at 1:100, and 50 mL of each diluted culture was incubated under identical conditions until reaching an optical density (OD) at 600 nm of 0.5. The cultures were then further diluted 1:50 to the same final volume and incubated for an additional 10 h at 37 °C. Optical density was measured and recorded every 30 min using an Infinite 200 PRO plate reader (Tecan, Switzerland) at 600 nm. To compare the growth rates of the bacterial strains, we fitted the growth data using a logistic model79: In this re-parameterized logistic model, m1 represents the asymptotic maximum (A), m2 corresponds to the maximum specific growth rate (μ max), and m3 indicates the lag phase duration (λ). To statistically compare the different parameters between strains, we performed a one-way ANOVA, followed by pairwise comparisons using Tukey’s HSD. Quantitative western blot analysis of MAT levels A quantitative Western Blot was performed to assess the relative expression levels of methionine adenosyltransferase (MAT). Cells were lysed using a lysis buffer (25 mM Tris, pH 8.0; 150 mM KCl; 1 mM BTT; 1% Triton X-100 (Sigma); 0.5% Benzonase (Sigma)) after a 15 min pre-incubation. The samples were diluted in Tris buffer (pH 8.0) to ensure that 10 μg of protein was loaded per well onto a 12.5% SDS-PAGE gel. Electrophoresis was conducted at 120 V for 80 min, followed by electrotransfer onto a Trans-Blot Pure nitrocellulose membrane (Bio-Rad, Hercules, CA). The membrane was blocked with AdvanBlock-PF blocking solution (Advansta, San Jose, MO) and washed with AdvanWash washing solution (Advansta). Custom-raised rabbit anti-Ureaplasma urealyticum MAT polyclonal antibodies (Genemed Synthesis, San Antonio, TX) were added at 1/2000 dilution, and the membrane was incubated for 1 h at room temperature with shaking. After additional washes with AdvanWash, the membrane was incubated with goat anti-rabbit horseradish peroxidase (HRP)-conjugated secondary antibodies diluted 1/2000 (Catalog number 402-SH026; Lot: AP9840; Abm) for one hour at room temperature. The immunoblot signal was detected using enhanced chemiluminescence (ECL; Advansta). The relative intracellular abundances of methionine MAT were compared using two-sided pairwise t-tests with Bonferroni correction. Whole genome sequencing A comprehensive whole genome sequencing (WGS) analysis was conducted to obtain genomic data for the ‘engineered’, ‘evolved’, ‘evolved-WT’, and ‘mut’ strains, which were then compared to the genome sequence of the WT strain. DNA extraction was performed using the DNeasy UltraClean Microbial Kit (QIAGEN, Hilden, Germany) according to the manufacturer’s protocol. Sequencing was carried out by the Mantoux Bioinformatics team at the Ilana and Pascal Mantoux Institute for Bioinformatics, Weizmann Institute, Israel. The MiSeq Micro V2 (300 cycles) kit (Illumina, San Diego, CA) was used for sequencing, and library preparation included the incorporation of nine-base Unique Molecular Identifiers (UMIs). LC-MS analysis For individual variants, two experimental batches, each consisting of 19 biological replicates per WT, ‘engineered’, ‘evolved’, ‘evolved-WT’ and ‘mut’ strains, were prepared on two different days, resulting in a total of 38 replicates per strain. For whole populations, two experimental batches, each consisting of 12–13 replicates per WT, ‘evolved’, and ‘evolved-WT’ were prepared on two different days, each resulting in a total of 24–26 replicates per population. In all instances, cultures were grown overnight at 37 °C in the culture medium, followed by a 1:100 dilution and further incubation until reaching an OD of 0.5 at 600 nm. A secondary 1:50 dilution was then performed, with additional growth until the OD reached 0.5 again. Cells were then harvested and centrifuged at 4 °C for 15 min at 4000 × g. The supernatant was discarded, and the pellet was resuspended in 1 ml of ice-cold phosphate-buffered saline and centrifuged again under the same conditions. Afterwards, ice-cold 80% ultrapure methanol was added to the pellet, mixed well by vortexing, put for 10 s in liquid nitrogen, and incubated for 30 min at −20 °C. To remove cell debris, samples were centrifuged at a maximum speed at 4 °C for 30 min. The supernatant was collected into a fresh Eppendorf tube and by using a SpeedVac vacuum concentrator (Thermo Fisher Scientific), the liquid was evaporated. Pure standards of S-adenosylmethionine (SAM) and methionine were prepared to facilitate their identification in the chromatograms. Metabolite analysis was performed using Acquity UPLC I-Class System (Waters Corporation, Milford, MA) coupled with a Q Exactive hybrid quadrupole-orbitrap mass spectrometer (Thermo Fisher Scientific) in positive ionization mode. Screening for unknown features using compound discoverer Raw data files were processed using Compound Discoverer version 3.3 SP2 (Thermo Fisher Scientific) to screen for unknown compounds. The analysis workflow included retention time alignment, unknown peak detection and ion association, peak filtration (intensity ≤ 1e6), gap filling, and removal of background components unrelated to experimental samples (detected in blanks with a ratio ≤ 2.5). Additional steps involved elemental composition prediction, database searching via ChemSpider, spectral matching with mzCloud, and mass list comparisons. To avoid redundant identification of multiple adduct ion compositions of the same compound, metabolic features were consolidated based on similarity in retention time (RT difference ≤ 0.1), high correlation (R ≥ 0.9), and m/z differences matching known adducts. The program annotations of SAM and methionine were validated using pure standards, and their levels across strains were compared using one-way ANOVA, followed by pairwise comparisons using Tukey’s HSD test. Principal components analysis (PCA) of the metabolites A two-dimensional principal components analysis (PCA) was performed using MetaboAnalyst 5.080, a freely accessible web-based platform for metabolomics data analysis, statistical evaluation, and data visualization. Prior to analysis, the data were log-normalized. PERMANOVA for metabolite abundances and network similarity To assess differences in metabolite abundance composition among groups, permutational multivariate analysis of variance (PERMANOVA) was performed in MetaboAnalyst 5.0 using Euclidean distance matrices calculated from log-transformed data. A global PERMANOVA was conducted to test for overall differences across all groups. In addition, pairwise PERMANOVA analyses with BH-FDR correction for multiple comparisons were performed to evaluate differences between individual group pairs. For network comparisons, differences in community structure were evaluated using PERMANOVA based on the Jaccard distance matrices. This analysis was conducted in R using the ‘adonis2’ function from the ‘vegan’ package. Construction of correlation networks Correlation analysis was used to evaluate associations between variables in the untargeted LC-MS metabolomics results. Pearson’s pairwise correlations were computed based on the peak intensities of metabolic features, generating an adjacency matrix of correlation coefficients. To remove transitive interactions, we employed the Probabilistic Context Likelihood of Relatedness (PCLRC) algorithm, an extension of the Context Likelihood of Relatedness (CLR) approach55,81. PCLRC integrates correlation analysis with mutual information and resampling techniques to filter out non-significant associations. In each iteration, 75% of the data is randomly sampled, correlations are computed, and CLR is applied. Only the top Q% of interactions are retained, and after k iterations, edges present in at least 95% of iterations form the final network. Following established recommendations24,55, we set Q% to 0.3 and k to 1000. For network construction, metabolic features were grouped by bacterial strain and experimental batch. An undirected, weighted network, in which edge weights represented the magnitude of correlations, was generated for each group, retaining only consistent correlations across technical replicates. This approach improved network reliability by minimizing noise and random variation. Network analysis Average Shortest Path and Network Diameter were calculated with ‘mean_distance’ and ‘diameter’ functions, respectively, using ‘igraph’ package in R82. The weights parameter was set to 1/(correlation coefficient of the edges) to account for edge strength, treating higher weights as closer connections in the network. Node degree was calculated with the ‘degree’ function from the ‘igraph’ package in R83. Modularity was calculated by applying the ‘Infomap’ algorithm from the ‘infomapecology’ package in R83 on a weighted bilayer network composed of layers of positive and negative correlations. In this approach, each metabolic feature (“physical node”) is represented by two “state nodes” in the network—one in the positive layer and one in the negative layer—enabling the analysis of both types of correlations within the same framework. These two state nodes may not belong to the same module, and they may not share similar characteristics. Normalized Mutual Information (NMI) was calculated using the ‘NMI’ function in the ‘infomapecology’ package in R83. This measure allows the comparison of two partitions even when they have a different number of modules. Given two network partitions, the NMI is calculated as84: Where A and B are the two network partitions; CA is the number of modules in partition A, CB is the number of communities in partition B, N is the number of nodes, Nij is the overlap between A’s module i and B’s module j, i.e., the number of nodes that the two modules have in common. Ni is the total number of nodes in A’s module i, and Nj is the total number of nodes in B’s module j. The NMI ranges from 0 to 1, where 0 signifies that the community structures are independent and 1 that they are identical. Neighborhood Connectivity was calculated using the ‘neighborhood.connectivity’ function from the ‘influential’ package in R85: Where Ni is the set of neighbors of vertex i, and k is one of these neighbors. Clustering Coefficient was calculated using the ‘transitivity’ function from the ‘igraph’ package in R82. It is defined as the number of triangles connected to a node i divided by the number of triples centered around it: where Ti is the count of triangles connected to node i, and di is the degree value of node i. Global Clustering Coefficient was calculated as the ratio of the count of triangles and connected triples in the graph. Node Role was determined by its within-module connectivity and its participation across different modules, both of which reflect the node’s function and position in the network59. The classification of node roles was based on two key parameters—the within-module degree (Zi) and the participation coefficient (Ci). Zi measures how well-connected a node is within its own module. Nodes with high Zi values are considered “module hubs” because they have strong connections within their module. In contrast, nodes with low Zi values have weaker connections within their module and are categorized as “non-hubs.” Ci evaluates how a node’s connections are distributed across different modules. A Ci value close to 1 suggests that the node connects to many modules, while a value close to 0 implies the node is confined to its own module. This is used to further distinguish between roles: Peripheral nodes are non-hubs with a low Ci (Ci ≤ 0.62), indicating that their connections are mainly within their own module; Connectors are non-hubs with a high Ci (Ci > 0.62), meaning their connections span multiple modules; Module hubs are hubs with a low Ci (Ci ≤ 0.62), as they are highly connected within their module but don’t bridge different modules; Network hubs are hubs with a high Ci (Ci > 0.62), indicating that they connect different modules and play a central role in linking various parts of the network. Targeted metabolomics To advance the identification of unknown compounds based on fragmentation spectral information, we used SIRIUS 6.1.0 software60. The isotope patterns of the molecular ions from the MS1 and the fragmentation spectra from the MS2 of the selected compounds were manually exported from CompoundDiscoverer and imported into SIRIUS. The following parameters were used for the computation: instrument (Orbitrap), MS2 mass accuracy (5 ppm), possible adducts ([M + H]+, [M+Na]+, [M + K]+), molecular formula generation (de novo + bottom up). In addition, CSI:FingerID fingerprint prediction62, CANOPUS compound class identification61, CSI:FingerID structure database search, and MSNovelist63 were activated; all other parameters were maintained at their default settings. Reporting summary Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article. Supplementary information Supplementary Information Transparent Peer Review file Reporting Summary Description of Additional Supplementary Files Supplementary Data 1 Supplementary Data 2 Supplementary Data 3 Supplementary Data 4 Supplementary Data 5 Source data Source Data Acknowledgements We thank Yariv Brotman for the theoretical and technical guidance and Gabriel Frank and Adrian Serohijos for the detailed reading of the manuscript. Author contributions Conceptualization: S.B., S.P. Investigation: S.S., W.J., E.S., M.N., D.K. Supervision: S.B., S.P. Visualization: S.S. Writing—original draft: S.B. Writing—review and editing: S.B., S.P., S.S. Peer review Peer review information Nature Communications thanks the anonymous reviewers for their contribution to the peer review of this work. A peer review file is available. Funding S.B. was supported by the Israel Science Foundation personal research grant 593/21. S.P. was supported by grants from the Israel Science Foundation (grant 1281/20) and Human Frontiers Science Program Organization (grant number RGY0064/2022). Data availability All data are available in the article Supplementary Info and Supplementary Data files. The raw metabolomics data generated in this study have been deposited in the MetaboLights database under accession code MTBLS14569. Sequencing data have been deposited to BioProject under accession code PRJNA1503581. Source data are provided as a Source Data file. Source data are provided with this paper. Code availability The R code encompassing all steps of data analysis is available on GitHub (https://github.com/Ecological-Complexity-Lab/metabolic_corrs_networks_analysis.git)86. A detailed README.md file provides execution instructions. Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Contributor Information Shai Pilosof, Email: pilos@bgu.ac.il. Shimon Bershtein, Email: shimonb@bgu.ac.il. Supplementary information The online version contains supplementary material available at https://doi.org/10.1038/s41467-026-77850-0. References Associated Data Supplementary Materials Supplementary Information Transparent Peer Review file Reporting Summary Description of Additional Supplementary Files Supplementary Data 1 Supplementary Data 2 Supplementary Data 3 Supplementary Data 4 Supplementary Data 5 Source Data Data Availability Statement All data are available in the article Supplementary Info and Supplementary Data files. The raw metabolomics data generated in this study have been deposited in the MetaboLights database under accession code MTBLS14569. Sequencing data have been deposited to BioProject under accession code PRJNA1503581. Source data are provided as a Source Data file. Source data are provided with this paper. The R code encompassing all steps of data analysis is available on GitHub (https://github.com/Ecological-Complexity-Lab/metabolic_corrs_networks_analysis.git)86. A detailed README.md file provides execution instructions.