To understand how protein complexes are organized and how metabolism is connected to these, one has to look first at suitable large-scale studies on protein complexes to have sufficient data. In general, proteins which are well connected in protein complexes are called “hubs” and one can distinguish between permanently well-connected “party” hubs and proteins that have many connections, but specific interactions are reserved for small time slots, known as “date” hubs (Figure 1, [11]). Furthermore, for all protein complexes, there are central core complexes and accessory components outside. The accessory components may either be adaptor proteins for interactions, occurring as shared components in various complexes, or are only accessory components for one specific complex [10].

Structural considerations for the network topology include consideration of network centrality. Thus, general metabolic pathways are connected by short pathways and currency metabolites are well buffered to achieve optimal balancing of the network. Whether such networks are truly “small world”—like [12] or not [13] is still a matter of debate. Small world-like behavior often reflects agglomeration, and evolutionary forces drive such processes (e.g., pathway duplication, pathway recruitment etc., [14]), thereby enhancing exactly this growth type, including metabolic enzyme complexes. There is also the concept of a large central component with several smaller bystander networks and a comparatively high number of singletons [15]. This is again a typical finding from interactomics [16], but partly reflects true effects of evolutionary forces at work. However, partly also natural limitations of knowledge (most data instances can be connected, so we get a large central component, and similar reasoning for other subnetworks) becomes a problem.

Selection optimizes metabolic networks in bacteria further. For instance, metabolic pathways in bacteria are organized to be optimally switched by central transcription factors and, in this respect, there is certainly a selection for optimal control. Controllability in different types of networks is currently a hot topic of research [17].

Regarding large-scale studies on prokaryotic complexes, focusing on one of the smallest bacteria known and profiting from its compact genome, Kühner et al. [4] used tandem affinity purification-mass spectrometry (TAP-MS) on the small Gram-positive bacterium M. pneumoniae. This revealed 62 homomultimeric and 116 heteromultimeric soluble protein complexes which are partly conserved across bacteria, including many novel ones. Interestingly, a third of the complexes are involved in higher levels of proteome organization. These larger multi-protein complex entities link successive steps in biological processes like a conveyor belt involving shared multifunctional components. This interesting finding of a factory-like arrangement of bacterial protein complexes churning out a maximum of proteins and processed metabolites was supported by structural analysis on 484 proteins (single particle electron microscopy, cellular electron tomograms and bioinformatical models). Thus, Kühner et al. [4] show details of the factory and interlinked protein complexes, including detailed structure prediction. Regarding time-dependent nuclear complexes, they found multiple regulators and regulatory interactions per prokaryotic gene, such as new noncoding transcripts. For instance, there are 89 of them in antisense configuration to known genes in M. pneumoniae [3]. With similar techniques, Butland et al. [18] analyzed E. coli complexes using affinity tagged proteins of 1,000 open reading frames (nearly a quarter of the genome). 648 were homogeneously purified and analyzed by mass spectrometry. The direct experimental approach revealed new interactions, as well as interactions predicted previously based on bioinformatic approaches from genome sequence or genetic data. Furthermore, looking in detail at both data sets ([3,18]) shows that many important interactions are conserved in both bacteria.

The question of conservation of prokaryotic protein complexes and their interactions was also analyzed by Parrish et al. [8] in the food-borne pathogen Campylobacter jejuni (NCTC11168). Yeast two-hybrid screens identified 11,687 interactions with 80% of all bioinformatically predicted proteins participating. Furthermore, this study places a large number of poorly characterized proteins into networks with hints about their functions. Interestingly, a number of their subnetworks are not only conserved compared to E. coli, but also to S. cerevisiae. Furthermore, biochemical pathways can be mapped on protein interaction networks. This has been shown in this study for the chemotaxis pathway of C. jejuni. As an application aspect, a large subnetwork of putative essential genes suggests new antimicrobial drug targets for C. jejuni and related organisms. In summary, this landmark study [8] nearly doubled the binary interactions described for prokaryotes at that time, and showed that many of the identified complexes are conserved in their central components between various prokaryotic organisms.

Figure 2 shows a number of complexes available from these studies and found either in E. coli or Staphylococcus aureus or both as well as their connection to metabolism.

The annotation and validation of all the implied prokaryotic interaction data and protein complexes is nontrivial. One important way to achieve this is to make them accessible by a Wiki, for instance, the “SubtiWiki” or the “WikiPathways”. These Wikis provide a knowledgebase for the Gram-positive model bacterium Bacillus subtilis [19], or pathways in general [20]. The SubtiWiki includes the companion databases SubtiPathways [21] and SubtInteract with graphical presentations of metabolism, its regulation, as well as protein–protein interactions and complexes, and is highly recommended as an exemplary resource to study systems biology of protein complexes in bacteria.

Moreover, protein modifications have to be charted. They influence protein complex formation, and removing or adding a protein modification allows corresponding protein complexes to change with time. An important mechanism is phosphorylation. Interacting proteins may specifically bind to these protein modifications or only bind if these modifications are absent. Thus, Van Noort et al. [22] compared this mechanism with other posttranscriptional regulatory mechanisms in M. pneumoniae. This organism is particularly suited for such studies because it encodes only two protein kinases and one protein phosphatase. This fact allows an elegant identification of the protein-specific effects on the phosphorylation network using specific deletion mutants. Van Noort and co-workers also considered changes in protein abundance and lysine acetylation [22]. Introduction of the mutations did not alter the transcriptional response, but deletion of the two putative N-acetyltransferases affected protein phosphorylation, which demonstrates the cross-talk between the two posttranslational modification systems. Phosphoproteome studies were also reported for B. subtilis. Jers et al. [23] identified nine previously unknown tyrosine-phosphorylated proteins in B. subtilis, and the majority of these were at least in vitro PtkA substrates.

For growth, intracellular model pathogens rely on different metabolic resources and exploit suitable protein complexes for their utilization and thus regulate metabolism accordingly. Therefore, comparing the wild-type S. aureus strain 8325 and the isogenic deletion mutants (either lacking the eukaryotic-like protein serine/threonine kinase PknB or the phosphatase Stp, [24]) remarkable differences were found. Those differences were in nucleotide metabolism and cell wall precursor metabolites, such as peptidoglycan and cell wall teichoic acid biosynthesis in S. aureus. This phosphatase and the kinase are also antagonistic players in central metabolism, affecting enolase, triose phosphate isomerase, fructose biphosphate aldolase, pyruvate dehydrogenase, phosphate acetyl transferase, and others.

Similarly, S. aureus pathogenicity potential depends on the iron status of the host [25]. Combining difference in-gel electrophoresis and mass spectrometry with multivariate statistical analyses, Friedman et al. [25] revealed clusters of cellular proteins responding to distinct iron-exposure conditions (iron chelation, hemin treatment), as well as genetic changes (∆fur). 120 proteins representing several coordinated biochemical pathways and regulons were affected by changes in iron-exposure status, for instance the heme-regulated transport system (hrtAB), a novel transport system. During iron starvation, pH decreased and acidic end-products accumulated so that iron was released from the host iron-carrier protein transferrin.

Complexes may thus rapidly assemble and disassemble according to the metabolic situation. To achieve this efficiently, “moonlighting” enzymes have a hidden second function only apparent in the “moonlight”, i.e., an alternative metabolic condition revealing its nonstandard function. Aconitase is a good example; with sufficient iron content, its iron-sulfur cluster is present and the enzyme catalyzes isomerization of citrate to isocitrate. However, under low iron, a hidden second activity is apparent: without an iron-sulfur cluster the enzyme binds iron-responsive elements in RNA to block translation. Such enzymes are thus found in two different complexes (e.g., metabolic complex or RNA-binding complex) and change their life from metabolism to control of gene expression in response to the availability of their substrates (“trigger enzymes”; [9]). Other enzymes have acquired a DNA-binding domain. They act as direct transcription repressors by binding DNA in the absence of substrate. Furthermore, sugar permeases of the phosphotransferase system control transcription activity by phosphorylating regulators in the absence of a specific substrate [26]. Finally, regulatory enzymes may control transcription factors by inhibitory protein–protein interactions. Duplication and subsequent functional specialization, a general motor of enzyme evolution, is also a major evolutionary pattern found here.

On a biophysical level, dynamics of metabolic protein complexes involve also molecular channeling of metabolites, as well as molecular crowding effects (Figure 3).

Substrate channeling directly transfers a product to an adjacent cascade enzyme without mixing with the bulk phase, which is again most easily achieved in a static or transient multienzyme complex (Figure 3a). Besides enhanced reaction rates, unstable substrates are protected and metabolic fluxes regulated. Furthermore, this avoids unfavorable equilibria, toxic metabolite inhibition, substrate competition or kinetics [43]. Substrate channeling has also biotechnological potential for metabolic engineering, and cell-free synthetic pathway biotransformation.

Substrate channeling is an old field, started by the Cori’s in the 1950s [44]. Paul Srere coined the concept of “metabolon” to describe improved channeling of substrates in the citric acid cycle [45], encapsulating the concept of what is described here. To study channeling became quite popular in the ‘80s’ (see Tombes and Shapiro [46] on phosphorylcreatinine shuttling; Yang et al. [47] on β-oxidation) and ‘90s’ (see Kholodenko et al. [48]; Miziorko et al. [49] for cholesterol synthesis; Welch and Easterby [50] review a number of different metabolic examples). There is also previous modeling work that shows dynamic channeling is capable of decreasing the metabolite pool sizes (but also able to increase them) [51,52].

Hence, channeling speeds up prokaryote metabolism and involved enzymes. It has already been probed and even changes have been indirectly monitored for carbohydrate metabolism combining a variety of methods: Bauler et al. [53] modeled in this way a two-step reaction, using a simple spherical approximation for the enzymes and substrate particles. These authors applied Brownian dynamics to show that spatial proximity and channeling is helpful. Closely aligned active sites are the most effective reaction pathway in their results, but they must not be too close so that the ability of the substrate to react with the first enzyme is not hindered. Protein interaction data comparing E. coli, yeast and humans support [54] that indirect protein interactions between related enzymes achieve metabolic channeling. Interestingly, protein complexes include nonenzymatic mediator proteins, sometimes related to signal transduction, to form channeling modules. In E. coli reactions, possessing such interactions show higher flux. Channeling could lead to more cross-talk. However, Pérez-Bercoff et al. [54] find that scaffolding proteins limit this, keeping protein complexes in separate places. Furthermore, there are interesting differences in the channeling of glucose towards gluconate and other catabolic end-products like pyruvate and acetate, with respect to phosphate status for different Pseudomonas strains (Pseudomonas aeruginosa versus P. fluorescens) [55]. Enzyme activities including glucose dehydrogenase, glucose-6-phosphate dehydrogenase and pyruvate carboxylase change in a coordinated fashion in response to changes in growth, glucose utilization or gluconic acid secretion. This includes a shift of glucose towards a direct oxidative pathway under phosphate deficiency which may perhaps also be implied in the different abilities of the two strains to produce gluconic acid. Comparison of enzyme–enzyme interactions in metabolic networks of E. coli and S. cerevisiae shows evidence for direct metabolic channeling [56]. Enzyme–enzyme interactions occur more often for pathway neighbors with at least one shared metabolite. Non-neighbouring interactions are often regulatory.

Molecular crowding: Crowding effects do change prokaryotic enzymes, metabolism and promote protein complexes in prokaryotes. Where metabolic channeling is a specific effect between metabolic proteins (enzymes and protein mediators) in a complex, molecular crowding is instead a more general, unspecific effect by the combined variety of biomolecules (Figure 3b), including nucleic acids, proteins, polysaccharides, as well as other soluble and insoluble components and metabolites (total concentration 400 g/L). The reason for the crowding effect is thus that together these biomolecules occupy a significant proportion (20–40%) of the total cellular volume in cytoplasm and nucleus, respectively [57]. Biophysical effects from crowding differ thus in different compartments of cells. Many nuclear processes such as gene transcription, hnRNA splicing and DNA replication, assemble large protein–nucleic acid complexes. Macromolecular crowding provides a cooperative momentum for these [58], boosting functionally important nuclear activities. In cell membranes, membrane proteins occupy approximately 30% of the total surface area leading to crowding effects on the surface as well as unique effects for the even more movement restricted integral membrane [58]. Thus Wang et al. [59] directly monitored the effect of strong crowding on pressure-induced reduction of unfolding of a protein (staphylococcal nuclease) by tryptophan fluorescence.

Besides such unspecific crowding effects, there are also complex-promoting activities from proteins from the milieu. Thus, in NMR studies, it was observed that high molecular weight glycoproteins are efficient molecular seeds for protein aggregation [60]. Such additional effects were also invoked by McGuffee and Elcock [61], using a simulation model which successfully describes the relative thermodynamic stabilities of proteins measured in E. coli, modeling 50 highly abundant macromolecule types at experimentally measured concentrations. Morelli et al. [62] show a simple way to model the effects of macromolecular crowding on biochemical networks. To succeed, they had to scale bimolecular association and dissociation rates correctly. They used kinetic Monte Carlo simulations and looked at crowding effects, comparing a constitutively expressed gene, a repressed gene, and a model for the bacteriophage λ genetic switch. Each molecular assembly was modeled both with and without nonspecific binding of transcription factors to genomic DNA. Furthermore, crowding effects shifted association–dissociation equilibria rather than slowing down protein diffusion, which sometimes had unexpected effects on biochemical network performance. Norris and Malys [63] show even changes of Michaelis-Menten kinetic constant K_m, and rate constant k_cat for the enzyme glucose-6-phosphate dehydrogenase under crowding. k_cat increased at very low concentrations of crowding agent or at high crowded concentrations during heating (45 °C), adding PEG. Simulations applying the Arrhenius equation agree with these observations.

More subtle effects of how enzymes are influenced by crowding are apparent in simulations and only partly supported by experimental data: Adenylate kinase was coarse grain modeled by Echeverria and Kapral [64], showing large-scale hinge motions during enzymatic cycles. Multiparticle collision dynamics included effects due to hydrodynamic interactions. A stationary random array of hard spherical objects provided crowding in the simulation. Adenylate kinase prefers a closed conformation for high volume fractions (smaller obstacle radius and tighter packing). Average enzymatic cycle time and characteristic times of internal conformational motions of the protein change, as do the transport properties. Under crowding, diffusive motion becomes up to ten times slower with longer orientational relaxation time. In general, and according to simulations on seven different proteins, those experiencing the strongest crowding effects have larger conformational changes between open and closed states [65]. In Brownian dynamic simulations, Ando and Skolnick [66] modeled a simplified E. coli cytoplasm with 15 different macromolecule types at physiological concentrations and sphere representations using a soft repulsive potential. These authors compare their data with the experiment; at cellular concentrations, the calculated diffusion constant of GFP was shape independent and much larger than in the experiments. However, including hydrodynamic interactions using the equivalent sphere system reproduced the in vivo experimental GFP diffusion constant without further parameter adjustment. Nonspecific attractive interactions reduced strongly diffusivity of the largest macromolecules [66]. The authors observed attractive clusters around these, but not if hydrodynamic interactions dominated. The latter led also to size-independent intermolecular dynamic correlations. Both models are interesting, and the noted differences between both models should now be directly compared to further experimental data. Even the change in the binding free energy due to crowding could be quantitatively described by the scaled particle theory model without any fitting parameters [67]. Crowders of different sizes were predicted by the same model with an additive setup. Crowding increased the fraction of specific complexes and nonspecific transient encounter complexes were reduced in a crowded environment as the nonspecific complexes had greater excluded volume [67]. However, more experimental data are needed to confirm these detailed predictions.