Dynamic remodeling of Escherichia coli interactome in response to environmental perturbations

Proteins play an essential role in the vital biological processes governing cellular functions. Most proteins function as members of macromolecular machines, with the network of interacting proteins revealing the molecular mechanisms driving the formation of these complexes. Profiling the physiology‐driven remodeling of these interactions within different contexts constitutes a crucial component to achieving a comprehensive systems‐level understanding of interactome dynamics. Here, we apply co‐fractionation mass spectrometry and computational modeling to quantify and profile the interactions of ∼2000 proteins in the bacterium Escherichia coli cultured under 10 distinct culture conditions. The resulting quantitative co‐elution patterns revealed large‐scale condition‐dependent interaction remodeling among protein complexes involved in diverse biochemical pathways in response to the unique environmental challenges. The network‐level analysis highlighted interactome‐wide biophysical properties and structural patterns governing interaction remodeling. Our results provide evidence of the local and global plasticity of the E. coli interactome along with a rigorous generalizable framework to define protein interaction specificity. We provide an accompanying interactive web application to facilitate the exploration of these rewired networks.

growth" (NZG) [1,2].Accurate descriptions of the molecular mechanisms supporting these vastly different physiological states in different environments is crucial to elucidating the fundamental relationship between genotype and phenotype.
While bacteria are known to regulate biochemical responses through transcriptional control of gene operons, post-translational regulation is also thought to mediate their adaptation to changing physiological demands [3][4][5].Proteomics research therefore plays a particularly important role in elucidating the physiological state of bacteria due to the crucial role of proteins in executing essential cellular functions.Comparative proteome studies have found that the distribution of protein resources in bacteria is related to their growth rate [3].Peebo et al. used a chemostat to isolate E. coli within a range of µ = 0.2-0.9h, and found that more proteins in slow-growing cells were used for energy generation, carbohydrate transport, and metabolism, whereas most proteins in fast-growing cells functioned in biological processes closely related to protein synthesis pathways [4].More recently, Schmidt et al. measured the relative abundance of more than 2300 proteins in E. coli under 22 culture conditions and found that growth rate was positively correlated with amino acid transport and ribosomal biogenesis, and negatively correlated with energy generation pathways [5].
Proteins do not function in isolation in the cell, rather they selectively interact to form large multi-subunit complexes that are collectively known as the "interactome."Elucidating the composition and overall network properties of the interactome is key for revealing the molecular mechanism of cell growth and environmental adaptation on a proteome-wide level.Butland et al. and Hu et al. used affinity purification mass spectrometry (AP-MS) to define a dense network of protein-protein interactions (PPIs) among the soluble protein complexes of E. coli cultured in LB medium [6,7], while Babu et al. [8] reported PPIs among membrane-associated proteins under a single static growth condition.
Multiplex co-fractionation mass spectrometry (mCF-MS) is a flexible approach for detecting and comparing protein complexes and PPI networks under different cellular contexts [9].In mCF-MS, cellular lysates are biochemically fractionated prior to mass spectrometrybased shotgun sequencing and relative protein quantification.Ionexchange chromatography (IEX) is a particularly effective method to resolve complex protein mixtures [9][10][11], while sophisticated computational analysis tools are then used to assign proteins to a given complex based on the similarity of their co-elution profiles [12][13][14].Due to its quantitative high-throughput nature, mCF-MS technology allows for direct comparison of interactome differences between distinct samples after controlling for spurious variance using biological replicates.
The protein interactome is a dynamic system that changes in response to different stimuli and environments.Studies on the dynamic response of S. cerevisiae PPIs to environmental disturbances showed that more than half of the PPIs only existed under specific culture conditions [15].Changes in a single protein or a small amount of protein would lead to changes in PPIs, and some protein modifications would also lead to changes in interactions [15].Other studies have also revealed that the location and abundance of proteins in S. cerevisiae

Significance Statement
The protein interactome contains the network of physical interactions that enable the functions of most proteins.Protein interactions can be disrupted by many triggers, such as pathogen infection or mutations in protein-coding genes, yet most studies in the field focused on characterizing the interactome in a static manner, with few devoted to investigating the dynamic nature of these interactions.In this study, we profiled the dynamics of the Escherichia coli interactome in response to changes in its growth environment.
Our results shed light on the mechanisms governing protein interaction remodeling, while also providing a rigorous analytical framework for quantifying interaction dynamics on an interactome-wide scale, representing an important step toward accurate modeling of dynamic biological systems.cells influence PPI formation [16,17].The proteome profile of bacteria similarly varies under different conditions, and so it follows that the bacterial interactome is expected to exhibit dynamic assembly patterns.However, there has been no systematic study yet on the bacterial interactome under different conditions, with the existing E. coli interactome only constructed for the LB medium.[18] While protein interac- In this study, we used mCF/MS along with a customized data analysis pipeline to compare differences in the PPI networks of E. coli cells grown under 10 different culture conditions.We defined and investigated extensive condition-dependent remodeling predicted by this dataset.Projecting this PPI remodeling against evolutionary and biochemical traits allowed us to pinpoint key biological factors driving protein interaction dynamics, while statistical assessment of the dynamic networks revealed fundamental mechanistic principles underlying interactome plasticity.Along with a robust computational framework to support future studies of this nature, we developed an interactive web application to facilitate exploration of these results by the broader research community.

Experimental design
We grew E. coli BW25113 under 10 alternate growth conditions representing different types of environmental variations (

Generation of condition-dependent CF/MS proteome profiles
We used multiplex co-fractionation mass spectrometry (mCF/MS; see Section 4) to generate proteomic profiles encompassing 1937 E. coli proteins, in replicate, across 96 IEX-HPLC fractions for each of the 10 growth conditions (Figure 1A).We used isobaric Tandem Mass Tag (TMT) stable isotope chemical labeling to quantify the conditionspecific protein elution profiles for just under half of all curated protein-coding genes (Figure 1B).These mCF/MS profiles consisted of quantitative measurements of each protein relative levels in each fraction, using summed precursor (MS2) ion intensity as a proxy for protein abundance (Figure 1C).The proteins profiled in this dataset represent strong coverage (>70%) of the five largest functional annotation pathways in the KEGG [19] database (Figure 1D), with all pathways well-represented with >53% coverage on average.Additionally, chromatograms showed excellent reproducibility between HPLC runs (triplicate technical injections are shown in Figure 1E).Given a key premise of mCF/MS is that physically-interacting proteins display similar elution profiles, we found that pairs of proteins with high-confidence experimental evidence of interacting in the STRING database [20] had significantly higher correlations in our dataset than random protein pairs (p < 2.2 e-16; Figure 1F), demonstrating strong recovery of canonical protein interactions.

Computational strategy for profiling interactome remodeling from dynamic CF/MS data
A key challenge of the mCF/MS routine is that the multiplexed measurements are made at the peptide-level, after in vitro-digestion of the protein fractions, while quantitative inferences must be assigned at the higher protein-level.Moreover, while our dataset was expected to be dynamic, existing tools such as EPIC [12], PrInCe [13], and CCprofiler [14] were devised for analyzing static PPI networks using standard non-multiplexed CF/MS data, with each proposing different pre-processing strategies.Thus, we developed a pre-processing workflow starting from the peptide-level with the optimal set of steps and parameters for our data that led to the best recapture of literature-reported PPIs (see Section 4).
Figure 2 outlines the computational analysis pipeline we developed to profile dynamic protein interaction remodeling based on conditional mCF/MS data.The pipeline consists of two modules, with the first dedicated to pre-processing the mCF/MS data (Figure 2A), and the second responsible for quantitatively scoring protein interaction remodeling (Figure 2B).
The second module is concerned with quantitatively profiling protein interaction remodeling from the pre-processed mCF/MS data.Our strategy begins by computing conditional similarity scores for each pair of interacting proteins within a reference interactome predicted from the mCF/MS data using established algorithms (see Section 4).We then leveraged the quantitative nature of the 10 mCF/MS reporter channel measurements to quantify the nature and extent of remodeling exhibited by each putative binding partner in response to each growth environment based on changes in the similarity of the interactor mCF/MS profiles.This allowed us to profile changes in the interaction patterns at different levels of biological organization, starting from individual pairs of proteins and multi-protein complexes, through to whole interactome, in response to the different growth conditions.
Our pipeline enabled the identification of key pathways and molecular mechanism driving interactome remodeling, as described in later sections, while providing a rigorous generalizable framework for interactome remodeling using mCF/MS data.A detailed breakdown of each step in the analysis pipeline can be found in Section 4.
Prior to applying this pipelinee, we first predicted a reference interactome spanning 6,152 high-confidence pairwise interactions among the quantified E. coli proteins in the mCF/MS data by combining and scoring all 10 datasets using the established EPIC software [12] (see Section 4 and Table S1).This unified interactome was shown to be highly modular (Louvain modularity = 0.89), encompassing 267 putative multi-component complexes ranging in size from just three subunits to a large ribosomal assembly consisting of 35 polypeptides (Table S2).While most proteins (68.3%) were assigned to a single complex, several "moonlighting" proteins were predicted to function as members of multiple distinct complexes, with nine proteins operating in as many as four complexes, including molybdopterin cofactors (moaE, mobA, modE, moeA) and rRNA methylases (rlmA, rlmB, rlmG).

Extent of global remodeling of E. coli interactome
We used the standard LB growth medium as a baseline condition to quantify the extent of interactome remodeling occurring in response to the other nine different environmental perturbations (as summarized in Table 1).By scoring each constituent interaction's conditional remodeling relative to the LB reference, we placed the predicted protein complexes on a spectrum of stable-to-disrupted complexes detected within each growth medium, with complexes whose underlying mCF/MS profiles showed the largest increase in dissimilarity being assigned higher remodeling scores (see Section 4).We found that while individual complexes reacted differently depending on the specific culture conditions, the overall extent of interactome remodeling was remarkably similar across all conditions, with most complexes (57.8% on average) remaining quantitatively unchanged rel-ative to the LB baseline, whereas only a small <5% fraction exhibited high (>0.5)remodeling scores suggestive of disrupted interactions (Table S2).
Figure 3 shows the interactome-wide patterns of protein complex remodeling seen among select representatives of the four major test conditions evaluated in the comparative experiment.We highlight an example complex from each condition based on both biological relevance and being impacted by relatively high levels of PPI remodeling.The quantitative nature of the mCF/MS measurements enabled deciphering the nature of the intra-complex remodeling occurring (e.g., subunit loss versus changes in overall macromolecular F I G U R E 2 Computational pipeline for quantifying interactome remodeling from dynamic CF/MS data.(A) CF/MS data processing.Pipeline for processing peptide-level CF/MS data leading to generation of protein-level profiles.(B) Profiling protein interaction remodeling.Analysis workflow for quantifying interactome remodeling under different conditions.abundance relative to LB), and we show the underlying mCF/MS data for a representative remodeled interaction per complex.
For example, in the condition where galactose was the primary carbon source, we found evidence of extensive formation of a protein complex composed mainly of galactose metabolism enzymes, where the assembly extends from one large unit into two distinct subunits connected by the galM epimerase protein.Additionally, the dynamic association of garD, a galactarate dehydrase, suggests a peripheral role for this protein with the core complex.Together, these remodeling patterns suggest the presence of condition-dependent macromolecules driven by the availability of galactose and simultaneous trimming of less essential interactions.
In response to high temperature (42 • C), we found that one of the top remodeled complexes showed evidence of decomposition into two separate subcomplexes, with one consisting of the heat shock response proteins from the "hsl" family, and the other consisting of hydrogenase proteins from the "hya" family of genes.Meanwhile, the dimeric histone-like master transcriptional regulators hupA and hupB and chaperone protein htpG dissociated entirely from a prominent assembly seen in LB.A potential explanation for this behavior could be induction of a heat shock transcriptional response for the organism to cope with the high temperature.
A striking example of the influence of highly-connected hub proteins on complex remodeling was observed when E. coli was cultured in a chemostat at near-zero growth rate.The complex with the highest remodeling score in this condition involved several transporter proteins, with the respiratory enzyme glpB playing a central role by being the only member that physically interacts with all complex members.Both the downregulation and elution shifting of glpB under the near-zero growth condition coincided with the destruction of all its intra-complex interactions and the consequent loss of this complex, demonstrating the dynamic nature of complexes with reliance on highly-connected subunits for their structural integrity.
Finally, another complex that had a high remodeling score was detected preferentially among starved cells, showing extensive disruption primarily due to the loss of interactions involving the two highlyconnected subunits ade and add.Proteins in this conditional assembly have been associated with multiple types of post-translational protein modifications, including phosphorylation and acetylation [21,22], implying altered activity of an upstream signaling mechanism as a key response to the starvation condition, presumably triggering energy conservation mechanisms.

Biological pathways driving interactome remodeling
Using annotated pathway membership information from the EcoCyc curation database [23] combined with our reference EPIC-derived interactome, we quantified the extent of conditional remodeling detected in each condition among interaction partners mapping to 38 major biochemical pathway families, classifying them according F I G U R E 3 Global patterns of interactome remodeling.(A) Leftmost network plots visualize protein complexes predicted from the dynamic CF/MS data.Nodes represent protein complexes color-coded by quantitative extent of remodeling in given condition compared to base growth media (LB) and sized according to number of member proteins.Middle network plots visualize example complexes selected due to high level of remodeling and biological significance.Rightmost plots visualize CF/MS profiles of example intra-complex pairwise protein interactions with high remodeling scores.
to the spread of remodeling scores around the baseline LB condition (Figure 4A).While most (61.5%)biosynthesis pathway families showed relatively stable behavior across conditions compared to LB, notable exceptions included strengthening of genetic machinery involved in cell replication in conjunction with consistent weakening of interactions related to the synthesis of polyprenyl and tetrapyrrole, which are known to be growth inhibitory.Interestingly, metabolic regulator biosynthesis displayed a wide range of dynamic remodeling, being strengthened in some conditions and weakened in others.
Dissecting the interaction behavior showed that these changes are primarily driven by changes in the elution profiles of the proteins rather than changes in relative abundance, suggesting possible biochemical alterations to the structures of these complexes.
In contrast to the biosynthesis pathways, only one-third of pathways associated with biomolecule degradation were relatively stable across conditions.Reactive oxygen species degradation, a hallmark stress response, was consistently impacted across all changes in the media, alongside changes in the machinery linked to the degradation of amino acids and aldehydes.On the other hand, we observed a relative weakening in assemblies linked to degradation of fatty acids and lipids, aromatic compounds and alcohol, with alterations primarily driven by abundance changes.Additionally, interactions involving signal transduction pathways were elevated concomitant with increased activity of signaling cascades in response to the environmental perturbations (Figure 4A).A similar pattern was observed for enzyme assemblies linked to metabolic detoxification and glycan formation.
Conversely, persistent global weakening of interactions among components of the protein modification and energy generation pathways was seen under non-conventional environments, reflecting a shift to increased conservation.
We also identified pathways most influential in driving the remodeling within each growth condition.We consequently examined assemblies linked to the top four strengthened and weakened pathways detected within each culture (Figure 4B).Consistent with the global trends (Figure 4A), the protein modification machinery was among the most consistently and severely impacted systems across all conditions relative to LB, indicative of the key role played by dynamic post-translational modifications.Interestingly, acid resistance (ability to withstand pH < 2.5) was among the most-strengthened pathways detected under near-zero growth yet one of the most weakened in starved cells, despite being generally associated with the stationary phase in the literature.
Hypergeometric enrichment tests revealed that amino acid degradation, cofactor biosynthesis, and precursor metabolite generation pathways are significantly enriched (FDR < 0.01) among both the top 5% remodeled and top 5% stable proteins on average across conditions (Figure 5C,D).Amine degradation and aromatic compound biosynthesis were exclusively strongly represented among the most stable proteins (Figure 5D), while a larger set of 13 pathways spanning diverse biological mechanisms were enriched among the most highly remodeled proteins, indicative of the multi-faceted nature of interactome disruption across the diverse growth conditions.

Structural and functional properties of interactome remodeling
Our mCF/MS data enabled the capture of two distinct types of data patterns suggestive of interaction remodeling: (1) qualitative: elution (HPLC retention time) shifts, that is, impacting the overlap of the proteins' cross-fraction co-elution patterns, and (2) quantitative: intensity fold-changes, that is, the ratio between protein relative expression (co-abundance) between conditions (Figure 2B).While we opted for a single integrative analysis strategy designed to generate a single score combining the changes seen in both patterns (see Section 4), we also compared the patterns of each reference interaction and found that the magnitude of changes in interactors' co-abundance tended to be significantly greater than that exhibited by coelution changes, suggesting that expression-level regulation plays a prominent role as the dominant mechanism influencing downstream interaction remodeling (Figure 5A).
To identify biological properties that distinguish proteins based on their levels of interaction remodeling, we computed associations between diverse protein traits and their corresponding averaged integrated remodeling score within each growth condition separately.A negative correlation between remodeling scores and summed protein intensities indicated that low-abundance proteins were more prone to remodeling (Figure 5B), consistent with past findings from mammalian interactome remodeling studies [24].Classifying the proteins based on their evolutionary age also revealed that certain conditions, including near-zero growth and stationary phase, favored ancient protein interaction stability, implying that adaptation to certain types of environments, such as growth on xylose-rich media, is a modern adaptation (Figure 5C).Strikingly, we found that membrane proteins were relatively more stable than their cytosolic and periplasmic counterparts (Figure 5D), while proteins subject to phosphorylation were likewise generally more stable (Figure 5E).Counter-intuitively, proteins annotated as containing intrinsically disordered structures had a lower median remodeling score (Figure 5F), suggesting they tend to form constitutive assemblies.Similarly, highly-connected hub proteins that participated in many (six or more) interactions were more stable, likely due to their persistent and essential role in maintaining interactome structure (Figure 5G), while unexpectedly the number of complexes a protein participates in was less influential (Figure 5H).
We explored the overarching structural patterns of interactome remodeling by examining the relationship between the integrated remodeling scores and the mathematical properties at both the pairwise interaction level and the protein complex level.We found no correlation between each interaction's betweenness score within the interactome and its corresponding remodeling propensity (Figure 5I).This indicates that unlike hub proteins, individual central interactions have the same tendency to be remodeled as more peripheral ones.
Meanwhile, tightly-connected protein complexes had relatively lower remodeling scores on average than those with sparser intra-complex connections, suggesting higher structural resilience to changes in the surrounding environment (Figure 5J).

DISCUSSION
Here, we present the results of a multi-factorial perturbation experiment investigating the global robustness and localized dynamics of the E. coli interactome in response to different types of environmental perturbations.We cultured a K-12 laboratory strain under 10 different growth media and generated high-throughput mCF/MS data to enable comparative protein interaction prediction and quantitative profiling of interactome dynamics.We found rewired protein complexes that were altered preferentially, or even exclusively, in certain physiological contexts that highlight key players in environmental adaptation responses.These included a galactose metabolism complex forming in response to the availability of galactose, remodeling of a heat shock complex under high-temperature, and destruction of a complex associated with post-translational modifications in starved cells.
We also pinpointed macromolecules and their associated biological pathways driving this remodeling in comparison to crucial (housekeeping) assemblies that remain universally unaltered (stable core).Notable pathway remodeling patterns included strengthening of cell replication pathways alongside weakening of growth inhibitory pathways, consistent weakening of protein modification and energy generation pathways across the different environments, and elevated levels of signal transaction pathways.Our integrative scoring approach also revealed interactome-wide biological, biophysical, and structural patterns governing the tendency of bacterial interactions to become disrupted or strengthened.Notably, we found that high-abundance proteins, phosphorylated proteins, and highly-connected proteins tended to form more stable assemblies that were resistant to environmental perturbations.
Our experimental workflow included freeze-thawing and sonication of the E. coli cells prior to co-fractionation (see Section 4), and while we acknowledge that this aspect of the experiment could potentially affect the integrity of the protein interactions, this method has demonstrated strong results in our previous CF/MS-based protein-protein interaction studies (e.g., Havugimana et al., 2012 [10], Wan et al., 2015 [11], Pourhaghighi et al., 2020 [25], Havugimana et al., 2022 [9]).We also note that while each of the experimental perturbations explored in this study presents rich material for biological exploration within the larger context of E. coli studies, we focused our analysis primarily on the large-scale interactome remodeling patterns, and present our data in an accessible manner to enable specialized in-depth follow up studies on the different conditional responses captured here.To facilitate exploration of the results, an interactive web application visualizing the dynamic E.coli mCF/MS profiles is available at https://bnfweb.bu.
edu/EcoliDynamicInteractome/.Finally, we note that the experimental and computational pipelines reported here provide a generalizable workflow for future studies of interactome dynamics in other settings.

Strain
The wild-type E. coli K-12 strain BW25113 [8] was stored in the lab and used to generate the data for all 10 culture conditions.

Media
Chemical reagents for media preparation were purchased from Sigma-Aldrich unless specified otherwise.The LB broth miller was purchased from Fisher BioReagents™.Twenty-five grams of LB broth power, with tryptone 10 g, yeast extract 5 g, and NaCl 10 g, was suspended in 1 L of Mini-Q water and sterilized by autoclaving.The LB plates were produced by adding 1.5 g agar to 100 mL LB medium mixture before autoclaving.
M9 minimal medium without carbon source was prepared in the fol- favored the stability of ancient proteins while others the modern proteins.(D) Confidence intervals of averaged protein remodeling scores for proteins from different cellular compartments.Membrane proteins tended to be the most stable across conditions.(E) Confidence intervals of averaged protein remodeling scores for proteins with and without evidence of phosphorylation in the dbPSP database.Phosporylated proteins were marginally more stable across conditions.(F) Confidence intervals of averaged protein remodeling scores for proteins with and without evidence of intrinsically disordered structures in the DisProt database.No dominant pattern was observed among intrinsically disordered proteins.(G) Confidence intervals of averaged protein remodeling scores based on number of interactions.Hub proteins with higher number of interactions were more stable across conditions.(H) Confidence intervals of averaged protein remodeling scores for proteins based on number of complexes.Despite the higher stability of hub proteins, there was no strong association between protein complex membership and remodeling.(I) Relationship between edge betweenness score of interactions relative to the interactome structure and remodeling scores.Consistently near-zero correlation suggests that central interactions exhibit the same amount of remodeling as periphery ones.(J) Relationship between protein complex density and averaged protein remodeling scores.The downward trend suggests tightly-connected complexes are less prone to remodeling.
The Amino acid medium (AA) was made by supplementing the medium with glucose with a final concentration of 4 g/L, and the amino acids solution (50×), which was purchased from Sigma (R7131), was used to replace NH 4 Cl in the M9 medium as the nitrogen.The amino acids solution consisted the following individual amino acids: Arginine 10.0 g/L, Asparagine 2.84 g/L, Aspartic Acid 1.0 g/L, Cystine 2.5 g/L, Glutamic Acid 1.0 g/L, Glycine 0.5 g/L, Histidine 0.75 g/L, Hydroxy-L-Proline 1.0 g/L, Isoleucine 2.5 g/L, Leucine 2.5 g/L, Lysine 2.0 g/L, Methionine 0.75 g/L, Phenylalanine 0.75 g/L, Proline 1.0 g/L, Serine 1.5 g/L, Threonine 1.0 g/L, Tryptophan 0.25 g/L, Tyrosine 1.16 g/L, Valine 1.0 g/L.Other supplement salt mixture in AA was the same as M9 minimal medium.For chemostat growth, 3 g/L glucose in M9 minimal medium was used.

Media list Growth conditions and why we choose them
Complex medium 1 LB The most commonly media used for E. coli cultivation.For comparison with published protein complex to illustrate BP method used in this project is good and credible for E. coli protein complex studies.
2 Glucose+AA Amino acids replaced ammonium as nitrogen source. Minimal

Cultivation
For the preculture, a single colony was picked from the LB plate and grown overnight in 50 mL LB medium in a 250-mL Erlenmeyer flask at 37 • C, 200 rpm.For the batch cultures, the cells from a preculture were washed twice with sterilized ice-cold phosphate-buffered saline (PBS) and re-inoculated into 100 mL of the appropriate medium in a 500-mL Erlenmeyer flask and grown at 37 • C, 200 rpm.The cells were first grown to exponential phase and then transferred into a second shake-flask containing fresh medium under the respective condition and growing to early exponential phase.The cells undergoing temperature stress were grown at 42 For cell cultivation, frozen glycerol stocks were inoculated into 5 mL LB medium and grown overnight.The precultures were 1:10 diluted with fresh glucose M9 medium, and allowed to grow in the batch mode to a specific OD before continuous operation initiated.
Then cultures were stabilized in chemostat mode at a dilution rate (D) (µ min ) = 0.00097 h (near-zero growth) until achieving the steady-state.
After sample collection, a continuous increase of D started until cells could not keep up with the rising D (resulting in culture washout) and achieved the maximal specific growth rate (µ max ) at ≈0.8 h.
Starved cells were continuously grown after reaching stationary phase for 1 day.

Fractionated proteins digestion and peptides desalting
Take a small aliquot (25 µL) of proteins from each fraction of the 96deep well collection plate, to a new 96-well plate to measure protein concentration using BCA assay in a Bio-tek microplate reader.The rest of proteins of each fraction were dried by a Savant SpeedVac SC210A (Thermo Fisher Scientific, USA), resuspended in 100 µL buffer

TMT labeling
We The desalted peptides were eluted into a PlateOne™ 96-well 0.5 mL polypropylene plate and divided to three new low-profile 96-Well PCR Plate and dried by SpeedVac.These three plates were stored in −80 • C, one waiting for mass spec analysis and the others stored as backups.

LC-MS/MS
The dried peptides were resuspended in 100 µL of solvent A and 15 µL of each sample was taken and loaded on an EASY nLC 1200 system coupled to a Q Exactive HF mass spectrometer equipped with an EASY-Spray ion source (all from Thermo Fisher Scientific, USA Fisher Scientific, USA).Each fractionated sample was eluted from the column with a 120-min gradient.

Protein identification with MaxQuant
The MS/MS raw files were searched using MaxQuant Version 1.6.0.16 against the E. coli database (Uniprot, download data:2018/06/25).The database consists of four, 313 E. coli proteins as well as known contaminants.The extracted MS/MS spectra were searched against a both forward (native) and reversed (decoy) sequences, with protein identification allowing for two missed trypsin cleavage sites, methionine oxidation and N-terminal acetylation as variable modifications, and carbamidomethylation of cysteine residues as a fixed modification.
Reporter ions (MS2 spectra) were used for quantification with 10plex TMT using a matched reporter mass tolerance of 0.003 Da.Peptide search tolerance was set to 4.5 ppm for MS1, while MS2 fragment tolerance was set to 10 ppm.Candidate peptide and protein identifications were filtered based on a stringent 1% false discovery rate (FDR) threshold.Match between runs was active with an alignment window of 20 min and a match retention time window of 0.7 min.Other MaxQuant parameters were performed using default settings.

CF/MS data preprocessing
We sought to develop a pre-processing workflow with the optimal set of steps and parameters for our data that would lead to the best recapture of literature-reported protein interactions.Starting from the peptide-level data, the following four steps represent the pre-processing steps in our final pipeline: (i) Normalization.To correct for the different number of peptides identified in each of the 96 fractions (850-5940 peptides per fraction), MS intensity values were normalized within each fraction, converting the raw intensity of each peptide to its proportion of all peptides within that fraction, followed by log2-transformation.
(ii) Smoothing signals.Each peptide's 96-fraction conditional profile is smoothed by taking a moving average of four fractions to smooth out the short-term fluctuations across neighboring fractions.
(iii) Filter outlier sibling peptides.On average, each protein in this dataset has 14 peptides mapped to it (IQR 5-19).It is expected that the peptides that map to the same protein should have similar profiles to each other, and as such any peptide that deviates significantly from its group is likely a faulty measurement or incorrect mapping.To filter out these outlier peptides, we performed average-linkage hierarchical clustering on the sibling peptides based on their similarity to each other, split the resulting dendrogram into two clusters, and retained the peptides belonging to the larger cluster as being representative of the protein.
(iv) Construct protein profiles.Finally, we collapsed the sibling peptide profiles into their corresponding protein profiles by averaging their per-fraction intensities, leading to a final set of 1937 proteins for downstream analysis.
In developing this pipeline, we benchmarked different strategies for each step as follows: 1. Generate protein profiles from peptide profiles using different sets of processing parameters.
3. Compare distances to literature-curated PPIs using a ROC analysis.
For step 2, the distances between protein profiles were computed using three metrics: Pearson distance, Euclidean distance, and Wasserstein distance, with Pearson distances eventually leading to the best performance.This dataset has two replicates, and as such four distances were computed per protein pair, with the average of these four distances taken as the representative distance between the two proteins.This distance was then converted into a signal-to-noise ratio by dividing the distance by the average cross-replicate distance, that is, noise, of the given two proteins.
For step 3, the following datasets were used as reference protein interactions to benchmark against: 1.A set of one million pairwise interactions from the STRING database v11 [20].STRING is a repository of interactions compiled from seven different sources.The dataset was filtered for high-confidence interactions only (STRING score > 0.7).Only pairs for which elution data is present in our data for both members were retained for further analysis, leading to a dataset of 24,912 interactions among 1821 unique proteins.
BioGRID is a literature-curated database of genetic and protein interaction data.Only pairs for which elution data is present for both members in our data were retained for further analysis, leading to a dataset of 36,204 interactions among 1690 unique proteins.
tions are rewired in response to different environment-specific factors such as changing gene expression and protein abundance patterns, we sought to explore the global patterns and mechanisms underlying E. coli interactome rewiring as opposed to specific environmental responses.

F I G U R E 1
Profiling E. coli interactome dynamics.(A) Experimental design.Quantitative CF/MS data was generated from E. coli cultured under 10 different media.(B) E. coli proteome coverage.Fraction of E. coli protein-coding genes quantified in our dataset.(C) Coverage of largest E. coli pathways in the KEGG database.(D) Dynamic CF/MS profiles.Heatmaps visualizing the CF/MS data generated for each of the growth media.(E) Reproducibility of data.Overlap of replicate HPLC data.(F) Recovery of known E. coli interactions.Comparison of Pearson correlation between pairs of proteins known to interact in the STRING database and all possible pairs of proteins in our dataset.

F I G U R E 4
Biological pathways driving interactome remodeling.(A) Average remodeling scores of main E. coli pathway families.Boxplots represent distribution of each pathway family's remodeling scores in each of the growth media.Pathway families colored according to spread of remodeling scores around base growth medium (LB).(B) Top remodeled and stable pathways across conditions.Top four strengthened and weakened pathways for selected growth conditions based on remodeling scores.(C) Pathways enriched among top remodeled proteins.Results of hypergeometric enrichment test for top 5% most disrupted proteins based on averaged remodeling scores.(D) Pathways enriched among top stable proteins.Results of hypergeometric enrichment test for top 5% most stable proteins based on averaged remodeling scores.F I G U R E 5 Structural and functional properties of interaction remodeling.(A) Boxplots comparing distribution of co-elution and co-abundance remodeling scores across all interactions.(B) Relationship between summed protein MS2 intensities and averaged protein remodeling scores.The downward trend suggests higher remodeling scores for lower abundance proteins.(C) Confidence intervals of averaged protein remodeling scores for two categories of proteins based on evolutionary age using information from the GenOrigin database.Some types of environmental changes
3, cells under the respective condition were collected by centrifugation at 3197 g at 4 • C for 20 min, washed twice with ice-cold PBS buffer, harvested by centrifugation at 10,000 g and the cell pellet was stored at −80 • C until further processing.For all batch cultures (totally six culture conditions), each culture condition generated three independent culture cells that were subjected to three independent protein samples extraction.The cultured cells under each chemostat culture condition (totally four) had only one biological replicate and each was used to generate protein samples extraction three times.1 BP (biochemical purification) with ion exchange chromatography (HPLC-IEX) Per 100 µg frozen cells were resuspended in 0.5 mL fresh protein extraction buffer (20 mM Tris-HCl, pH 8.0, 150 mM NaCl, 0.2 mM EDTA, 10% glycerol, protease inhibitor (PI, Roche, Cat.No. 04693159001), 0.25 mM TCEP), and lightly disrupted by sonication on ice.The soluble proteins were obtained by Centrifugation at 15,000 × g for 10 min at 4 • C, a small aliquot of the supernatant was taken to determine proteins concentration using a BCA assay.An Agilent 1260 Infinity II was equipped with columns PolyWAX LP 204WX0510 (200 × 4.6 mm i.d., 5 µm, 1000-Å) and PolyCAT A 204CT0510 (200 × 4.6 mm i.d., 5 µm, 1000-Å) (PolyLC INC, MD,USA).Before sample injection, the columns were balanced with 10 bed volumes of buffer A (0.75 mM AmAc).About ∼2.5 mg of fresh soluble protein sample was injected to the IEC column and separated by a linear gradient of 4%-30% buffer B (2.5 M AmAc) for 90 min and 30%-60% buffer B for 30 min, with a flow rate of 0.4 mL/min.Totally 96 fractions were collected using a 96-deep well plate (Thermo Scientific™ Abgene™ AB0564) with 1 min intervals.

( 8 . 5 M
urea-100 mM Tris-HCl, pH 8.0, 5 mM TCEP) reduced at 37 • C for 60 min, alkylated with 15 mM IAA for 60 min, diluted with 50 mM Tris-HCl with urea <1 M, digested by incubation with sequencing-grade modified trypsin (1/20∼1/100, w/w) overnight at 37 • C.After digestion, the reaction was stopped by adding FA with a final concentration of 1%.The peptides were desalted by a Sep-Pak tC18 96-well µElution plate (Waters, USA, Product Number 186002318).Before loading samples, the desalting plate was wetted with 0.5 mL methanol twice and washed with 0.5 mL 0.1% FA twice.Peptide fractions were loaded on the plate by centrifugation for 1 min at 100 g.Based on the number of proteins from each fraction measured with a BCA assay previously, the total peptides for each fraction loaded to the desalting plate should not exceed 1% of the sorbent weight (10 mg).The desalting plate was washed twice with 0.5 mL of 0.1% FA to remove the unbinding materials and eluted with 150 µL 0.1% FA/60% ACN twice, the elution samples were collected with PlateOne™ 96-well 0.5 mL polypropylene plate (USA Scientific, USA, Product number 1896-5110).The fractioned desalted samples in the plate were divided into three lowprofile 96-Well PCR Plates (Bio-rad, USA, Product number HSP9601) and dried by Speedvac.Dried peptides in one plate were used for further TMT-labeling, and the other two plates were stored at −80 • C as backups.

Table
• C high temperature stress.High temperature makes E. coli fragile.When put E. coli culture from 42 • C to refrigerator and then put back to 42 • C, E. coli almost cannot grow (death cell precipitation appear).It did not happen in 37 • C (37 • C-4 • C-37 • C, grow normal).When E. coli grown in 42 • C, some flasks grow faster and some flasks grow slower, not uniform.Maybe growth heterogeneity happened in high temperature stress.We also have "E. coli grown in chemostat aerobic condition with the same growth rate" sample, can do proteome comparison if necessary.
used 20 µL of 50 mM HEPES buffer to resuspend each fraction of µL 5% hydroxylamine incubating at room temperature for 15 min to quench the reaction.The labeled peptides in the 11 plates were pooled into a 96-well plate and desalted with a Sep-Pak tC18 96-well µElution plate.
from the same fraction of each plate and put them into a new plate.The 11th plate was a mixture of the 10 plates and used for normalization.Added 10 µL of 2 µg/µL TMT reagent to each fraction of peptides to make the TMT/sample = 5:1 (v/v) for labeling.The mixtures were incubated for 1 h at room temperature.After the reaction, added 2