Report

Global Network Reorganization During Dynamic Adaptations of Bacillus subtilis Metabolism

See allHide authors and affiliations

Science  02 Mar 2012:
Vol. 335, Issue 6072, pp. 1099-1103
DOI: 10.1126/science.1206871

Abstract

Adaptation of cells to environmental changes requires dynamic interactions between metabolic and regulatory networks, but studies typically address only one or a few layers of regulation. For nutritional shifts between two preferred carbon sources of Bacillus subtilis, we combined statistical and model-based data analyses of dynamic transcript, protein, and metabolite abundances and promoter activities. Adaptation to malate was rapid and primarily controlled posttranscriptionally compared with the slow, mainly transcriptionally controlled adaptation to glucose that entailed nearly half of the known transcription regulation network. Interactions across multiple levels of regulation were involved in adaptive changes that could also be achieved by controlling single genes. Our analysis suggests that global trade-offs and evolutionary constraints provide incentives to favor complex control programs.

A major challenge in biology is to understand the organization and interactions of the various functional and regulatory networks in cells. The underlying complexity arises from the intertwined nonlinear and dynamic interactions among a large number of cellular components. To better understand these interacting molecular networks, the acquisition of appropriate, preferably time-resolved quantitative data is a prerequisite (13). Because the acquisition of such data is technically demanding, few studies have reported transcript, protein, and metabolite abundances, and most studies have been restricted to steady-state conditions (48). Consequently, only subsets of components have been monitored dynamically for very short- (9, 10) or long-term responses (11, 12) to environmental perturbations. These studies typically revealed coordinated abundance changes (11, 12), major transcriptional reconfigurations in response to environmental change (9, 11), and an unanticipated complexity of unicellular organisms (7). Data interpretation, however, has generally been restricted to multivariate statistical analysis methods that indicate general but not mechanistic relationships between different molecular entities. Focusing on single data types with sophisticated computational analysis has been informative (3) but increases the risk of missing the functionally relevant multilevel control mechanisms (2), limiting the description of the underlying molecular mechanisms and, hence, the depth of biological insight.

To elucidate the dynamic interplay between metabolic and regulatory networks systematically, we investigated dynamic shifts in availability of the preferred carbon sources, glucose and malate, of the bacterium Bacillus subtilis (13). Because these nutrients repress the use of other substrates but are themselves used together, they represent a tractable model for analyzing dynamic decision-making by cells when faced with compounds of similar nutritional value. We induced dynamic shifts by adding glucose or malate to cultures of B. subtilis growing exponentially on the other substrate. To elucidate the cellular adaptation mechanisms, we determined transcript, protein, and absolute metabolite abundances, as well as promoter activities (Fig. 1 and table S1) (14), providing dynamic data with up to 24 time points for two shift experiments that cover short-term metabolic to longer-term protein-level adaptation (15).

Fig. 1

Overview of our experimental design, computational analysis, and information flow. The dynamic shift experiments yielded measured data from which nonmeasurable quantities were inferred by statistical and model-based analysis. Arrows show the flow of information, including the use of inferred data as parameters for subsequent modeling steps. ChIP, chromatin immunoprecipitation; LC-MS, liquid chromatography–mass spectrometry; GFP, green fluorescent protein; HPLC, high-performance liquid chromatography; OD, optical density.

To enable data integration, we developed problem-driven, yet generic solutions in three areas (table S1). To generate consistent data, we minimized biological variability (Fig. 1) by withdrawing samples from the same bioreactor culture in triplicate experiments (fig. S1) and, in a few cases, from standardized small-scale cultivations (fig. S2). We used naming and formatting conventions with unique identifiers for all considered constituents to permit the efficient exchange of data and knowledge. Furthermore, to maximize reliability and coverage for subsequent data analysis and modeling, we combined overlapping dynamic data acquired from different analytical platforms and on different time scales (Fig. 1). Data from the different transcript-array platforms did not require further consolidation (figs. S3 and S4), but we generated consensus proteomics data by calculating confidence-weighted averages of protein abundances obtained from two-dimensional gels and liquid chromatography–mass spectrometry (SOM 1). Replicate time series of transcript and protein data were then smoothed and interpolated by Bayesian multicurve regression analysis (SOM 1). Because metabolite data showed changes on various time scales, we developed a special algorithm that aligns time courses on the basis of Kalman smoothing (Fig. 2A and SOM 1) (16).

Fig. 2

(A) Processing of time-series data from parallel experiments. The example shown here is fructose 1,6-bisphosphate during the malate–to–glucose-plus-malate shift. Measurements from three independent cultures (black data points) yield an averaged and smooth time course (blue line). (B and C) Initial dynamic uptake rates of malate (green) and glucose (red) after addition to wild-type B. subtilis growing on the respective other substrate were obtained by fitting of splines. Black lines and shadings represent the specific growth rate and 95% confidence intervals, respectively (SOM 3). (D) The global importance of transcriptional regulation after glucose addition is confirmed by an increasing correlation (red) between flux sensitivity (measuring the impact of a change in environmental conditions on a metabolic flux) and protein abundance over time. This increase does not occur after malate addition (blue). White areas denote statistically significant correlation values with respect to an approximate 95% confidence interval; gray shaded areas denote the absence of statistically significant correlation.

We exploited the profusion of data (table S1) to extend the genome annotation of B. subtilis by statistical data analysis and integration with prior knowledge (table S2). Analysis of mRNA abundances identified 4393 transcriptionally active genomic regions, 109 of which were not previously described, including 21 putative protein-coding sequences and 23 antisense RNAs (table S3). Additionally, 2422 genomic regions were differentially transcribed after the nutrient shifts (table S1). Clustering of mRNA profiles and functional classification of differentially transcribed genes enabled detailed functional annotation of 46 genes (table S4) and probable function assignment of 853 previously not annotated genes (table S1). For the metabolically important transcription factors CcpA, CcpC, CcpN, and CggR (1719), we identified DNA binding sites by chromatin immunoprecipitation–on-chip analysis in malate plus glucose (table S1). We found 184 CcpA binding sites, most of which were located near to promoters of genes differentially expressed immediately after glucose addition (fig. S5). By filtering this group for high-scoring, previously undetected CcpA sites, we generated an improved regulatory network topology (SOM 2) for subsequent analysis of regulatory events.

Metabolomics data revealed instantaneous malate uptake into cells grown on glucose but substantially delayed uptake of glucose into cells grown on malate (Fig. 2, B and C). Sensitivity analysis of steady-state fluxes in the stoichiometric network model in combination with proteomic data predicted transcriptional regulation to be more important for the dynamics induced by glucose than by malate (Fig. 2D). Network component analysis (20) quantified the strengths of all 2900 transcription factor interactions with target genes, and 1488 of the interactions were implicated in at least one shift (689 for malate addition; 1244 for glucose addition) (SOM 2). Furthermore, we identified 110 posttranscriptional regulation events in protein synthesis (77 and 23 were specific to malate and glucose addition, respectively) through a dynamic model that correlates time profiles of promoter activity, mRNA abundance, and protein abundance for 300 genes for which all three data types were available (tables S5 to S7). Overall, our integrated analysis suggests that apparently similar adaptation processes are mediated by fundamentally different control mechanisms, namely a predominantly posttranscriptional regulation after malate addition and a greater reliance on transcriptional regulation after glucose addition.

Next, we inferred the dynamic activity profiles of 154 transcription factors by network-component analysis (20) from the transcript abundances of their 1754 known target genes (Fig. 3A and SOM 2). One hundred twenty-seven transcription factors changed their activity profile significantly in at least one shift (61 for malate addition and 91 for glucose addition). A rapid change of activity (<5 min), most prominent for the transporter regulators MalR (malate) and GlcT (glucose), and the deviations between the inferred transcription factor activities and the measured transcription factor mRNA abundance provided evidence for posttranslational regulation (SOM 2). Network component analysis predicted that the activity of 51 transcription factors was modulated posttranscriptionally, and 39 of these (for example, NadR, KipR, CcpN, FruR) are regulated by yet unknown effectors. Overall, both nutrient shifts induced substantial, global network reconfiguration at all levels of regulation.

Fig. 3

(A) Activities of the main transcriptional regulators controlling central metabolism during the shifts of glucose (GM) or malate (MG) to glucose plus malate. Blue curves represent log2 expression profiles of target genes; purple curves denote the inferred transcription factor activities. Straight lines in the middle sections indicate the relative contribution (proportional to color intensity) of different transcription factors to target gene expression changes (blue, activation; red, repression). (B and C) Averaged time profiles of transcripts (light blue), proteins (dark blue), metabolic fluxes (green), and metabolites (red) during the dynamic shift experiments of glucose (B) or malate (C) to malate plus glucose.

To identify the relevant primary changes in central metabolism, where essentially all genes were differentially expressed during one or both shifts, we correlated time courses of metabolic fluxes with those of the abundances of the corresponding enzymes. We estimated dynamic extracellular rates by interpolation (SOM 3) and network-wide pseudodynamic metabolic fluxes with a stoichiometric network model and the intracellular metabolite concentrations (Fig. 3, B and C, and SOM 4). Positive correlations between fluxes and enzyme abundances indicated genetic rather than metabolic regulation of the reaction rates (21). Upon the addition of malate, the abundances of only the nicotinamide adenine dinucleotide phosphate–dependent malic enzyme (YtsJ) and two acetate production enzymes (Pta and AckA) correlated with their respective reaction rates (R > 0.8) and thus putatively control flux (Fig. 4A). After the addition of glucose, enzymes for 11 central reactions putatively controlled flux, and in six cases increasing enzyme abundance potentially overcame the bottleneck in glucose uptake. The corresponding genes are organized in two operons: ptsGHI for glucose transport and phosphorylation and cggR-gapA-pgk-tpiA-pgm-eno for lower glycolysis. This focused analysis of multiple “omics” data sets suggested the pts and cggR operons as primary control targets for adaptation, where the half-maximal responses of the pts and cggR operon-encoded proteins within 30 to 60 min of glucose addition could account for the ~60-min delay in glucose uptake (Fig. 2, B and C). We evaluated the relative contributions of both operons to the delay with a cggR deletion mutant and a mutant with constitutive ptsG expression (Fig. 4B and SOM 5). When grown on malate, both mutants used glucose immediately without the delay of the wild-type strain. As the cggR or ptsG expression alone is sufficient to facilitate immediate glucose uptake, it is not obvious why B. subtilis does not express these genes constitutively.

Fig. 4

(A) Regulatory mechanisms relevant for control of metabolic fluxes during the two shifts. Key controlling enzymes are indicated by a positive correlation between reaction rate and enzyme abundance after addition of malate (left) or glucose (right). Reactions were considered only if both enzyme abundance and flux changed by more than one standard deviation and if the correlation coefficient was positive. The size of the purple circles is proportional to the correlation coefficient (see inset). Transcription factors (middle) were identified by network component analysis to be responsible for changes in enzyme expression, as indicated by color coding (gray denotes unknown factors). (B) Dynamics of glucose uptake rate upon addition of glucose to a culture growing exponentially on malate in mutants with constitutive ptsG expression (green), cggR deletion (blue), and the respective parent strains (solid red, vertical shading and dashed red, horizontal shading). Shaded areas denote 95% pointwise confidence intervals. gDCW, grams of dry cell weight. (C) Predicted growth effects of different control strategies using a simplified dynamic model. In silico competition experiments were carried out between two strains that differ in how they control malate and glucose uptake; both systems can be expressed constitutively or only upon substrate availability (SOM 6). Cocultures in a chemostat setting were pulsed every 3 hours with substrate (similar results were obtained with pulse frequencies of 1 and 9 hours), where the probability of selecting malate versus glucose as substrate was varied. The cost of metabolic system maintenance (ρ) normalized by the estimated cost for B. subtilis0; the solid line indicates the estimated value of ρ; the dashed line to the right is the upper bound of the 95% confidence interval) constituted a free simulation parameter. Control strategies with higher growth rates were considered advantageous, and the best control strategies are identified by color coding.

B. subtilis needs to optimize its use of two qualitatively distinct substrates: Glucose results in high growth yields at low metabolic rates, whereas malate results in low growth yields at high metabolic rates (13). Similar trade-offs exist between respiratory and fermentative strategies in adenosine triphosphate generation, both of which can confer condition-specific evolutionary advantages (22). To investigate optimal control strategies for substrate usage in Bacillus, we developed a simplified dynamic model that describes the substrate and biomass dynamics of the shift experiments quantitatively (SOM 6). The maintenance cost for enzyme expression (ρ) is a key parameter in the model; the estimated value of ρ0 ≈ 0.06 implies that full expression of a metabolic (substrate) system would reduce the maximal specific growth rate by ≈5%. In silico competition of strains that either constitutively express the metabolic systems or induce their expression only upon substrate availability allowed us to quantify the dynamic control effects. In simulations of a dynamically changing environment, we varied the substrate probabilities and the maintenance cost ρ because of the uncertainties in the estimated ρ0 (Fig. 4C). Surprisingly, there almost always existed a unique, evolutionarily dominating control strategy. The winning strategy depended on quantitative parameters. Hence, active regulation or constitutive subsystem expression are not advantageous per se. Even with the uncertainties in ρ0, the experimentally observed strategy of constitutive malate usage capacity (Fig. 2B) but induced (and delayed) glucose-specific metabolism confers an evolutionary advantage if B. subtilis predominantly encounters malate (Fig. 4C). This is biologically plausible given the organism’s habitat in the vicinity of plant root systems that often secrete carboxylic acids such as malate (13, 23).

Our systems approach helps reveal how previously known regulatory mechanisms are combined to effect nutritional transitions. Despite more than half of the B. subtilis gene complement being involved in the adaptive response to glucose, our methodology could discern the key regulatory events. The overall control strategy of B. subtilis can be rationalized in terms of its evolutionary advantages; however, these advantages, and therefore the overall control design, depend on quantitative system characteristics—regulation is not beneficial per se. The dynamic data presented here may be used for further computational analyses such as multivariate statistics and large-scale structural or kinetic network models (24). We hope that our publicly available tools for mathematical analyses and modeling will facilitate future large-scale and dynamic systems biology studies in B. subtilis and other species.

Supporting Online Material

www.sciencemag.org/cgi/content/full/335/6072/1099/DC1

Materials and Methods

SOM Text (SOM 1 to 6)

Figs. S1 to S20

Tables S1 to S8

References (2528)

Database S1

References and Notes

  1. Materials and methods are available as supporting material on Science Online.
  2. Acknowledgments: This work was funded through the European Commission 7th Framework project BaSysBio (LSHG-CT-2006-037469), coordinated by P. Noirot, the Department of Education, Science and Training (CG110055), and the National Health and Medical Research Council (455646). Data, algorithms, and mathematical models have been deposited in a publicly accessible relational database (www.basysbio.eu/nutrientshift/ and table S1).
View Abstract

Navigate This Article