Dynamic interactions of proteins in complex networks: a more structured view

Abstract

Virtually every process in a cell is carried out by macromolecular complexes whose actions need to be perfectly orchestrated. The synchronization and regulation of these biological functions is indeed critical and is usually carried out by complex networks of transient protein interactions. Here, we review some of the many strategies that proteins in regulatory networks use to achieve the dynamic plasticity necessary to rapidly respond to diverse cellular needs. More specifically, we present recent work on the molecular bases of transient peptide-mediated interactions and the role that post-translational modifications and disordered regions might play. Finally, in light of some recent findings, we speculate on the possibility of a new regulatory code for intrinsically disordered proteins and the potential biophysical and functional advantages that disorder might provide.

Abbreviations

CD2BP2

CD2 binding protein 2

ELM

Eukaryotic Linear Motif database

Gsk3

glycogen synthase kinase 3

IUP

intrinsically unstructured protein

MAPK

mitogen-activated protein kinase

PDB

Protein Data Bank

Plk1

polo-like kinase 1

PTM

post-translational modification

pTyr

phosphorylated tyrosine

SH2

Src homology 2

SH3

Src homology 3

Introduction

Proteins are the main perpetrators of most biological processes. However, they seldom act alone and most cellular functions are, in fact, carried out by large ensembles of proteins that need to be perfectly orchestrated. It has been estimated that there are ∼ 800 macromolecular complexes involved in the functioning of a yeast cell, of which 500 have already been described. Furthermore, it has been suggested that this number is likely to rise to ∼ 3000 molecular machines in human [1]. The individual functions of these complexes are very diverse and cover virtually the whole functional spectrum of a cell, from DNA replication and protein synthesis (i.e. housekeeping functions) to sensory perception or the acquisition of nutrients from a symbiotic organism. In addition to the large molecular machines, many cellular processes (e.g. energy production or amino acid synthesis) occur through the coordinated action of a number of discrete proteins and small complexes in what we know as a metabolic route. To date, some 100 of these functional pathways have been annotated, classified and stored [2,3]. The synchronization and regulation of these different processes and cell responses through signalling events is indeed critical and usually requires complex (i.e. nonlinear) responses produced through intricate networks of very dynamic and versatile protein–protein interactions.

Although both cellular processes and their regulation are based on the interaction of multiple proteins, the biophysics of these interactions is quite different. Large molecular machines (e.g. the ribosome or the RNA polymerases) are often built around a stable core of proteins that are rarely found in isolation and that define the basic function of the complex, which can be modulated through the attachment of peripheral protein components [1]. The direct relationships that constitute the different metabolic routes are frequently created by the interaction of two globular domains in different proteins and characterized by a large contact interface that ensures strong binding [4,5]. These interactions can be permanent or transient, meaning that the proteins involved are either always together or exist in isolation and meet only to fulfil a concrete function, but the speed at which the association/dissociation of such interactions occurs is usually not a critical issue. In this case, as for large molecular machines, it seems more important to achieve strong binding with a stable and relatively fixed position of the protein components, which serves to provide an optimal orientation of the several active sites, thus increasing the global efficiency of the process [6,7].

The situation is radically different in signalling and regulatory networks. Interactions embedded in such networks need to be extremely dynamic and versatile to be able to respond quickly to certain stimuli and to adapt this response over time [8]. Thus, instead of using large domain–domain binding interfaces, the interactions in regulatory networks are often characterized by small interfaces, with only a few molecular contacts involved, in which a short peptide in one protein is bound by a recognition domain in another. Around 50 of these protein recognition modules have already been described, and for many of them a high-resolution 3D structure is available. However, the mode of binding to other proteins has only been structurally characterized for about two-thirds of them [9,10]. The use of specialized modular protein–protein interaction domains, which are based on the recognition of short amino acid motifs, is indeed a very powerful strategy for nature to quickly explore a significant portion of the vast interaction space with a limited cost, in the sense that only one or a few residues in a protein need to change over the course of evolution to create a new interaction or to disable an existing one [11]. However, employing such a limited number of domains to mediate several thousands of interactions has brought into question the level of specificity of such interactions. On the one hand, cell regulatory networks would require very specific interactions in which one domain only recognizes and binds one other protein to avoid cross-reactions and undesired cross-talk between pathways. On the other hand, the small number of contacts in peptide-mediated interaction interfaces, needed to maintain their dynamic aspect, makes it difficult to reach this level of specificity [10].

Here, we review some of the tricks and strategies that proteins in cell regulatory networks use to achieve the desired dynamic plasticity without compromising other important aspects such as the binding strength or high specificity. In particular, we discuss recent work on the modular bases of peptide-mediated interactions, and how post-translational modifications (PTMs) and unstructured regions in proteins add an extra layer of regulatory flexibility to fine-tune the system’s response to the many and varied cellular needs.

High specificity in peptide-mediated interactions

Transient protein interactions, such as those found in signalling pathways, are often mediated by a globular domain of typically 50–150 residues that binds a linear, extended peptide. The peptides bound by such domains are characterized by a common consensus motif of 3–10 residues which is recognized in the interaction. Although they show common patterns, usually only a few positions in the motif are fixed, whereas other positions may allow certain flexibility in the amino acid composition. Consensus motifs are found in otherwise unrelated proteins, often in loops or unstructured regions, which are accessible for interaction with the recognition domain; however, motifs do adopt a well-defined structure upon binding to the domain. It has been repeatedly shown that isolated motifs are able to bind their domain with sufficient strength to establish a functional interaction [12], yet the interface formed between peptide and domain is usually much smaller (200–500 Å²) than those found in domain–domain interactions or complexes (∼ 2000 Å²) [4,5], which makes them better suited for transient interactions.

Because of their shortness, linear motifs are difficult to identify computationally from sequence data alone, and established experimental methods for protein interaction detection often miss these transient interactions [9]. In recent years, specific methods for the detection of peptide-mediated interactions have been developed based on the availability of high-throughput interaction data and the assumption that sequences containing a similar consensus motif should bind proteins via a common feature such as a recognition domain or a motif [13,14]. Several methods explicitly focus on the discovery of motifs that have arisen through convergent evolution, by removing homologous regions or proteins from a given set of binding partners for a particular protein or domain [14,15]. Once a recognition domain has been identified, high-throughput peptide display methods allow the detailed characterization of the consensus motif [16,17]. Most of what is known about domain–motif interactions is stored and classified in the Eukaryotic Linear Motif database (ELM) [18], which provides a literature-curated collection of experimentally validated consensus motifs involved in peptide-mediated interactions. It currently contains 81 motifs that bind to 51 different domains.

A representative example of a peptide-recognition domain is the Src homology 3 (SH3) domain, which binds proline-rich peptides that form a polyproline type II helix [19] (Fig. 1A). The classic SH3 consensus motif is PxxP (x stands for arbitrary amino acid residues), yet several variants have been observed, including [KRY]xxPxxP (class I; [KRY] means that Lys, Arg or Tyr are allowed in position 1) and PxxPx[KR] (class II), which bind the domain in opposite orientations. SH3 domains are abundant in many genomes, from 27 in yeast to 91 in fly and 304 in human [20]. Although interactions between domains and peptides involve only a few residues, they are known to be highly specific in vivo. Lim and colleagues [21] showed that only the SH3 domain in Sho1, but none of the other 26 SH3 domains in yeast, can bind the proline-rich peptide from polymyxin B sensitivity 2 (Pbs2). Furthermore, even single point mutations in arbitrary positions of the motif (x) affect the binding specificity [21]. These results indicate that peptide-mediated interactions have evolved to avoid cross-reactivity or, in other words, to maximize specificity. However, note that few examples have been studied to this level of detail, so it is not clear how widespread the complete encoding of specificity in sequence alone is.

Figure 1.  Examples of peptide-binding domains. (A) SH3 domain of the tyrosine kinase Fyn binding a typical ligand, a proline-rich peptide (PDB 1fyn) [110]. (B) SH2 domain of growth factor receptor-bound protein 2 (Gbr2) binding a phospho-Tyr-containing peptide (PDB 1jyr) [111]. (C) The Src kinase in the inactive state: the N-terminal SH3 domain (blue) binds to the linker between SH2 domain and kinase domain (green, residues emphasized). Note that the linker does not contain the typical PxxP motif, but nevertheless forms a polyproline type II helix and interacts with the SH3 domain. The SH2 domain (green/cyan) binds the phosphorylated C-terminus (red). Together the two intramolecular interactions cause the kinase domain to be inactive (PDB 1fmk) [38].

Another typical peptide-recognition domain, PDZ (postsynaptic density 95; discs large; zonula occludens 1), binds motifs through β-strand addition [22] and specifically recognizes the C-termini of its interaction partners, although exceptions have been observed in which the bound structure mimics a C-terminus [23]. PDZ domains are frequently found in fly and human proteomes, with 81 and 214 instances, respectively. However, in yeast, there are only two occurrences, with low sequence identity to canonical PDZ domains so that it is not clear if these putative domains are functional [24]. For many years, PDZ domains were thought to fall into distinct classes of preferentially recognized peptides based on position −2, with 0 being the C-terminal residue [25]. However, MacBeath and colleagues recently performed a genome-wide scan for PDZ–motif interactions in mouse, and found that the 157 mouse PDZ domains analysed do not fall into distinct binding classes; rather, their specificity is evenly distributed throughout selectivity space [26]. The group then created multidimensional profiles to represent the binding preferences of each of those PDZ domains, considering sequences from both peptide and domain [27]. The profiles were built around information extracted from high-resolution 3D structures of PDZ domains binding peptide ligands, from which the authors were able to derive 38 position pairs representing molecular contacts between domain and peptide. Training of such a detailed model for each PDZ domain was only possible because of the large amounts of data available from the previous study, containing confirmed interaction pairs but also peptides that do not bind a given PDZ domain. The models are capable of successfully predicting PDZ–motif interactions in mouse given sequence information only. They have also been tested on fly and worm PDZ domains and found – perhaps unsurprisingly – to work with lower accuracy in those organisms [27]. In another recent specificity study, Tonikian et al. [28] performed large-scale random phage library displays to analyse and classify approximately half of the PDZ domains in human and worm. They report at least 16 specificity classes, which significantly extend the classifications based on position −2, yet the variety is not as broad as in the mouse study by MacBeath and colleagues described above. Tonikian et al. [28] also performed mutational studies on PDZ domains and showed that, although the recognition of C-termini is robust, some mutations can alter the specificity of the domain even when the changed position is not in direct contact with the peptide.

Contextual specificity

Extremely detailed profiles for individual recognition domains, as described for PDZ, help explain the high specificity observed in vivo for peptide-mediated interactions. However, there are still domains with overlapping specificity profiles, and the relatively small number of contacts between domain and peptide cannot fully explain how they differentiate between interaction partners. It is assumed that orthogonal specificity information is encoded in the cellular context, i.e. the environment in which the interaction takes place. Factors such as subcellular localization and expression patterns shape this context. For example, although the interaction is possible in vitro, the SH3 domain of Fyn in T cells does not compete with the glycine–tyrosine–phenylalanine domain of CD2 binding protein 2 (CD2BP2) for the proline-rich motif in the cytoplasmic tail of CD2 because Fyn is located in the lipid rafts, whereas CD2BP2 occurs in the detergent-soluble membrane fraction [29]. Nevertheless, even in a given subcellular site and time, several interaction domains and corresponding binding peptides are found. In a recent study, we used high-resolution 3D structures extracted from the Protein Data Bank (PDB) [30] to study the molecular details of interaction specificity encoded in domain-peptide binding interfaces [10]. We found structures covering 30 different recognition domains and 47 peptides; thus, for more than half of the interactions in ELM, atom-level details are available. Using in silico alanine scanning, we showed that the motif contributes 80% of the binding energy on average, whereas the region surrounding the motif, termed context, provides the remaining 20%. By exchanging peptides that bind the same domain, we created non-native interactions and analysed these again using alanine scanning, showing that the binding of the motif itself is virtually always favourable, but that there are many unfavourable residue contacts in the context, in native but particularly in constructed (i.e. non-native) interactions. These results suggest that the motif may have evolved to ensure binding, although the context uses suboptimal interactions to achieve high specificity. Moreover, the suboptimal residue contacts observed in the context might also contribute to weaken the strength of the native interactions, promoting their transient nature. However, such subtle regulation is unlikely to work in domain–domain contacts, because large interfaces with strong affinities might ‘ignore’ unfavourable interactions in the periphery. In addition, our study confirmed that motifs assume a well-formed structure upon binding, whereas the structure of the context is more variable. However, the structural deviations do not correlate with energetic ones, so differences in the structure cannot account for most of the unfavourable interactions observed in the context [10].

Modularity of recognition domains and domain–peptide interactions

Like the SH3 and PDZ domains, many peptide-recognition domains are modular in the sense that they fold independently and their N- and C-termini are close in space, which facilitates their integration into exposed regions of an existing protein without disrupting its structure (see Fig. 1A,B). Besides their modular structure, there is also a functional layer of modularity in protein interactions, shown by heterologous chimeric proteins with the same function as native ones [31]. In general, because the peptides are so short, and because of the modular nature of the recognition domains binding them, a new connection in a cellular pathway can arise by the insertion of such a domain and mutations in a few residues, thus facilitating the evolution of new pathways or interconnections.

Phosphotyrosine-binding Src homology 2 (SH2) domains were the first type of peptide-recognition domains discovered [32] (Fig. 1B). In addition to binding a phosphorylated tyrosine (pTyr), different SH2 instances preferentially recognize three to five residues C-terminal to the pTyr. Currently, ELM lists six SH2 ligand classes, but considering the results from the large-scale PDZ specificity studies described above, it may well be that further subdivision will be needed. However, in the case of domains that bind post-translationally modified residues, a major part of the binding energy is contributed by the interaction with the modified residue, such that the consensus motifs may be shorter than those for domains recognizing unmodified peptides. Nevertheless, some instances of domains that usually bind phosphorylated peptides (e.g. SH2 or the phosphotyrosine-binding domain) appear to have evolved larger peptide-binding surfaces, allowing them to recognize unmodified ligands, albeit with weaker affinity [23,32–35].

The modularity observed in peptide-recognition domains also applies to the possibility of combining several of them to trigger specific responses. Indeed, complex and highly specific signalling behaviour with nonlinear responses in higher eukaryotes is often achieved by combining multiple peptide-binding domains in the same molecule or complex. Adaptor proteins or scaffolds enable high specificity by binding several proteins and thus tethering them together for interactions [36]. Many regulatory proteins contain peptide-recognition domains which serve substrate localization as well as activity regulation [37]. A well-studied example is the Src kinase, in which an SH2 domain binds the phosphorylated C-terminus and an N-terminal SH3 domain binds the linker between SH2 and the kinase domain [38,39] (Fig. 1C). Individually, neither of these interactions is sufficient to repress kinase activity [40]. Activation of the kinase requires either dephosphorylation of the C-terminus or the presence of extramolecular ligands for the SH2 or SH3 domains [37,41]. The most potent activators are proteins with correctly spaced ligand peptides for both domains [42], which makes them highly specific interaction partners for Src. Comparison of Src kinases in different species, including Monosiga brevicollis, a unicellular organism closely related to metazoans, provides insight into whether the insertion of such modular interaction domains first served substrate localization or catalytic regulation. Li et al. [43] showed that phosphorylation of the C-terminus of Monosiga’s Src family kinase and subsequent binding to the SH2 domain does not inhibit kinase activity. Thus, it is assumed that incorporation of modular binding domains initially served substrate localization, and that subsequent evolution of the intricate intramolecular interactions has allowed the peptide-binding domains to regulate kinase activity depending on whether they are bound to an external ligand or to intramolecular ligands [37].

Peptide-mediated interactions and, in particular, complex constructs with multiple peptide-binding domains that exploit synergistic effects provide fine-tuneable tools for the construction of proteins and circuits in synthetic biology [40]. For example, Dueber et al. [44] showed that cooperative input-dependent behaviour of a catalytic domain can be generated by creating a fusion protein with one or more peptide-binding domains on one side of the catalytic domain and corresponding peptide ligands on the other. In their experiment, a single SH3 domain and proline-rich peptide allowed for linear activation of the catalytic domain (with respect to the amount of external peptides added), whereas three to five domains and peptides provided a range of effects, from cooperativity to switch-like behaviour. The study also points out that the affinity between peptide and domain is important, because a single weak intramolecular peptide was not sufficient to repress the catalytic domain, although multiple strongly binding peptides created a protein that could not be activated [44].

The role of post-translational modifications in peptide-mediated interactions

Like the SH2 domains described above, peptide-mediated interactions often require specific PTMs to form binding sites for the recognition domains. Those PTMs are created by dedicated enzymes and are very often reversible, which gives them the ability to act as molecular switches and provides a way to dynamically regulate complex cellular processes. Many different types of amino acid PTMs are known and range from the addition of simple chemical groups, such as the phosphorylation of Ser, Thr and Tyr residues, the methylation of Arg residues and the hydroxylation of Pro residues, to the attachment of small proteins, including the sumoylation and ubiquitylation of Lys residues [45,46]. In principle, several PTMs of the same type can be added to a single residue, and a polypeptide chain can be customized by different PTMs. Together, this creates a large number of regulatory protein variants that can produce different cellular phenotypes. For example, specific Lys residues on the flexible N- and C-terminal tails of histones can be mono-, di- or trimethylated, which is important for chromatin organization and the epigenetic regulation of gene expression [46,47].

The combination of different types of PTMs or of several instances of the same type on one protein (multisite PTMs) results in different variations of PTM-induced peptide-mediated interactions. In cooperative interactions involving multisite PTMs, a signal is only generated after a given number of sites on the same protein have been modified. For example, the cell-division cycle 4 (Cdc4) protein of the budding yeast Saccharomyces cerevisiae binds to its target, substrate inhibitor of cyclin-dependent protein kinase 1 (Sic1), only when the target has been phosphorylated on at least six Ser/Thr residues. However, an example for a sequential PTM-dependent interaction is the human version of Cdc4, which binds to phosphothreonine (pThr)-containing sites on cyclin E. This in turn leads to polyubiquitylation of cyclin E, which is then degraded by the proteasome [46]. In the epigenetic regulation of transcription, polymethylation plays an important role, because certain Lys residues of histone tails have to be trimethylated in order to activate transcription [47]. Multiple PTMs on a protein can even be antagonistic if a PTM attached to one residue hinders the interaction with another modified residue [45,46]. Thus, multisite PTMs are important for the dynamic coordination of signalling events, as well as the qualitative and quantitative control of protein function in vivo. The dynamic nature of PTMs manifests itself both in the form of their reversibility (e.g. phosphate groups attached by protein kinases can be removed by protein phosphatases), and in their fast attachment and removal kinetics [45].

PTMs are known to affect protein function either by inducing conformational changes in the structure of a protein or through interaction with specific modular recognition domains, such as the ubiquitin-binding domain (UBD), which mediates the interaction between the proteasome and polyubiquitylated cyclin E, or the SH2 domain and phosphotyrosine-binding domain (PTB), which recognize Tyr-phosphorylated peptides. As explained above, the modular nature of those binding domains provides an efficient way to couple target proteins to peptides carrying a particular PTM or even a combination of different PTMs. Such domains usually bind the modified residue in a conserved binding pocket, whereas the surrounding amino acids provide the required specificity to distinguish between different peptide motifs with the same PTM. In combination with multisite modification, the selective recruitment of effector proteins through PTM-dependent interaction domains plays an important role in the regulation of many cellular processes, including intracellular signalling, chromatin organization and transcription [45–47].

Phosphorylation as a paradigm of PTM in cell regulatory processes

Arguably, the best studied PTM is the phosphorylation of Ser, Thr and Tyr residues, which plays a crucial role in signal transduction pathways and controls many important biological processes such as cell growth, differentiation and DNA repair [46–50]. For example, it has been shown that the just-in-time assembly of protein complexes during the different phases of the cell cycle is phosphorylation dependent [51], and that transcriptional regulation and phosphorylation of cell-cycle components have co-evolved to achieve a joint control of system dynamics [52]. Although only a single phosphate group can be added to a residue, the fact that those phosphorylations can occur on multiple sites in a protein increases the number of theoretically possible protein phospho-isoforms in a combinatorial way [46]. Phosphorylations are generated by protein kinases, which constitute one of the largest families of genes in eukaryotes, representing ∼ 2% of the protein-coding genome [48,49,53]. In human, > 510 protein kinases are known [53,54], and it has been estimated that about one third of all human proteins may be phosphorylated [48,50].

Protein kinases can be roughly classified into two main groups: Ser/Thr-specific kinases, which constitute ∼ 80% of all kinases; and Tyr-specific kinases [48,53]. All eukaryotic kinases are structurally similar and share a common fold for the catalytic domain, which is ∼ 250 amino acids long and consists of a small N-terminal lobe of β sheets and a larger C-terminal lobe of α helices (Fig. 1C). ATP binds in a cleft between the two lobes, and the substrate binds along that cleft. A set of conserved residues within the kinase catalytic domain then catalyses the transfer of the γ-phosphate group of ATP to the hydroxyl oxygen of the Ser, Thr or Tyr residue of the substrate [48,49]. In order to ensure that each kinase phosphorylates the right targets at the correct sites, many specificity-determining factors work in concert. The main factor involves the specific contacts between the residues of the kinase active site and the substrate residues surrounding the phosphorylation site. In those cases where several different substrates of a given kinase are known, comparison of those substrates at the sequence level has shown that it is often possible to derive a consensus sequence motif characteristic for a specific kinase or kinase family. However, as for the motifs mediating domain–peptide interactions, these consensus motifs are in most cases not sufficient to achieve the required level of kinase specificity in vivo, because they are usually too short and degenerated to be able to confidently distinguish real substrates from spurious motif hits occurring by chance. Consequently, other specificity-determining factors have evolved to ensure the desired specificity, such as the need for the phosphorylation site to be accessible on the protein surface, subcellular co-localization of the kinase and substrate, and modular-domain-mediated, docking or scaffolding interactions distal to the phosphorylation site [48,49,55].

In kinase-substrate selection, both specificity-determining interactions mediated by a modular interaction domain and docking interactions based on recognition of a short peptide motif by the kinase itself, but separate from the active site, can also be PTM dependent [55]. For example, polo-like kinase 1 (Plk1) contains a polo-box domain that binds to the kinase domain of Plk1, inhibiting catalytic activity. Only upon binding of the polo-box domain to a phosphorylated substrate with a particular consensus motif does the active site become free, allowing Plk1 to phosphorylate the substrate at another site [48,56,57]. This effect is also known as ‘priming phosphorylation’, because the substrate first needs to be ‘primed’ by phosphorylation on a specific site before it can be phosphorylated by Plk1 on another site [48]. As an example, in phosphorylation-dependent docking interactions mediated by a short peptide motif the substrates of glycogen synthase kinase 3 (Gsk3) first have to be phosphorylated on a residue C-terminal to the one recognized by the active site of Gsk3. This primed motif can then bind to a docking groove next to the catalytic cleft of Gsk3, activating the kinase [48,55,58]. Comparing the two main groups of kinases, docking interactions are more prevalent in Ser/Thr kinases, whereas Tyr kinases, like the previously mentioned Src kinase, mostly utilize modular interaction domains to increase their substrate specificity [48,55].

As mentioned above, subcellular co-localization can also help increase specificity by limiting the number of substrates accessible to the given kinase. Scaffolding interactions represent a somewhat extreme variant of co-localization in which scaffold proteins recruit the kinase and substrate to the same complex. Protein kinases can thus achieve different substrate specificity depending on the exact composition of the scaffolding complex [48]. A well-studied example is the yeast mitogen-activated protein kinase (MAPK) kinase kinase (MAPKKK) Ste11 (sterile 11) which is part of a protein interaction network that can be divided into three distinct signalling cascades: the mating, filamentous growth and high-osmolarity responses (Fig. 2). To ensure that activation of Ste11 subsequently activates only one of those pathways, in addition to mutual inhibition [59], different scaffold proteins are used to prevent unwanted pathway cross-talk [31,48,60,61]. The importance of scaffolding interactions in directing information flow has been demonstrated by Park et al., [31] who created chimeric scaffolds using components of the yeast mating and high-osmolarity MAPK pathways to convert a mating signal into an osmolarity response. Recently, Bhattacharyya et al. and Bashor et al. have shown that scaffolds not only have a passive role in signalling cascades, but are also able to allosterically activate binding partners [60] and fine-tune pathway dose-response and dynamics [61], respectively. Moreover, Good et al. demonstrated that the scaffold protein Ste5 (sterile 5) catalytically unlocks the MAPK Fus3 for phosphorylation by Ste7 (sterile 7), which explains why Fus3 is not activated in the filamentous growth pathway [62] (Fig. 2). Finally, kinase specificity in vivo also depends on so-called systems-level effects, such as the competition with other substrates and error correction through dephosphorylation by protein phosphatases. Together with a requirement for multisite phosphorylation, phosphatases can effectively minimize the effects of random off-target phosphorylation [48].

Figure 2.  Scaffold proteins prevent MAPK pathway cross-talk in the yeast interactome. In the yeast protein interaction network, the mitogen-activated protein kinase (MAPK) kinase kinase (MAPKKK) Ste11 (sterile 11) participates in three signalling cascades: the mating pheromone response (magenta), filamentous growth (cyan) and high osmolarity response (grey) pathways. To acquire signalling specificity and eliminate pathway cross-talk, the scaffold proteins Ste5 (sterile 5) and Pbs2 (polymyxin B sensitivity 2) recruit Ste11 and the respective pathway-specific MAPK, Fus3 (cell fusion 3) or Hog1 (high osmolarity glycerol response 1), to the same complex (indicated by dashed lines). Pbs2 also acts as a MAPK kinase (MAPKK) and phosphorylates Hog1 in the high osmolarity response pathway, whereas in the other two pathways this function is carried out by Ste7 (sterile 7), shown in orange. The MAPKs Fus3, Kss1 (kinase suppressor of Sst2 mutations 1) and Hog1 (in green) then either directly or indirectly activate specific transcription factors to respond to the given environmental stimulus [48,112]. Phosphorylations (P) are shown in red.

In the past, experimental identification of phosphorylation sites through mutational analysis and Edman degradation of phosphopeptides was both time-consuming and laborious and often required large amounts of purified protein [50,63]. With the advent of high-throughput MS-based methods for large-scale screening of phosphorylated peptides, the number of known phosphorylation sites in different species grows at an ever increasing pace. However, the fact that phosphorylation often occurs at low stoichiometry or on low-abundance proteins has hampered the detection of many in vivo phosphorylation sites [64]. To address this issue, several affinity-based strategies have been developed which try to enrich samples in modified peptides by combining quantitative proteomics techniques with specific enrichment methods. Such strategies allow us to increase the number of identified modification sites by almost an order of magnitude, as has recently been shown by Dephoure et al. [65] in constructing a quantitative atlas of human mitotic phosphorylation, detecting > 14 000 phosphorylation sites in ∼ 3700 different proteins. Current studies estimate the relative levels of phosphorylation in a cell to be in the range of 86–90% phospho-Ser (pSer), 10–12% pThr and 0.1–2% pTyr [48,66]. The low relative level of Tyr phosphorylation has led to the development of immunoprecipitation methods using specific pTyr antibodies to increase the detection of Tyr phosphorylation sites [64].

Although some known PTMs are annotated in the large protein databases UniProt [67] and HPRD [68], the wealth of available phosphorylation data has led to the development of dedicated databases of phosphorylation sites such as Phospho.ELM [50], PHOSIDA [69] and PhosphoSite [54]. The manually curated Phospho.ELM database, for example, now contains ∼ 12 100 pSer, 2400 pThr and 2100 pTyr sites from the literature, covering almost 4100 different eukaryotic proteins [50]. Moreover, databases such as Phospho3D [70] and mtcPTM [71] combine knowledge about the position of phosphorylation sites with the 3D structure of the respective protein or a close homologue to show phosphorylation sites in their structural environment, along with annotations of solvent accessibility, secondary structure and residue conservation. We anticipate that, with the wealth of data coming from high-throughput proteomics initiatives, these resources will play a crucial role in the near future.

Unveiling connectivity in phosphorylation-dependent signalling networks

Because of the increasing number of known phosphorylation sites, the aim of computational tools has shifted from predicting novel sites to trying to link known ones to specific protein kinases and to decipher the role of phosphorylations in the context of cellular signalling networks [50,72,73]. Those networks are important for detecting and processing different environmental stimuli, as well as for making complex response decisions [60]. With the current proteomics data, it is possible not only to discover the phosphorylated residues in a protein, but also to study which protein kinase is responsible for such phosphorylations. When trying to predict kinase substrates, to accurately address in vivo kinase specificity, it is crucial to consider not only the consensus sequence motif, but also sources of contextual information, such as subcellular co-localization, multisite phosphorylation or evolutionary conservation because it often implies functional importance [49,69,74–76]. The fact that phosphorylation sites are predominantly found in fast-evolving loop and hinge regions, however, leads to difficulties in correctly aligning them in orthologous proteins [69]. In a recent study, Linding et al. [77] integrated motif-based predictions with the network context of kinases and phosphoproteins to link protein kinases to specific phosphorylation sites. In the first step, they used predictions based on the consensus motif to assign known phosphorylation sites to one or more kinase families. Then, in the second step, they employed a probabilistic association network, covering physical interactions, genome and literature co-occurrence, as well as co-expression data, to identify for each phosphorylation site the most likely member of a kinase family responsible for the phosphorylation. The authors demonstrated that contextual information provided 60–80% of in vivo substrate specificity, and constructed a human phosphorylation network with 7143 site-specific kinase–substrate interactions between 1759 substrates and 68 kinases [77]. The network has recently been updated with current data and stored in the NetworKIN database with > 20 000 interactions spanning almost 4000 substrates and ∼ 70 kinases from 20 different kinase families [73]. Note that the context appears to be much more important here than in studies on individual domain–peptide interactions mentioned above. This may be because of the dynamic nature of phosphorylation events, which, among other regulatory factors, strongly depend on co-localization of substrates with kinases and phosphatases, and binding by the recognition domain is only possible in the modified state, whereas other recognition domains like SH3 or PDZ can bind their partner independent of such molecular changes. Furthermore, differences in organism complexity might play a role, because NetworKIN is based on human data, whereas most more-detailed studies have been performed in model organisms like yeast or worm.

To achieve a tight regulation and dynamic control of intracellular signalling, most kinases act in a complex network of kinases, phosphatases and other effector proteins, which are recruited by docking interactions or dedicated modular interaction domains [46,53,76]. For example, receptor tyrosine kinases, such as the epidermal growth factor receptor, dimerize upon binding of an external ligand and phosphorylate each other at multiple Tyr sites. Downstream effectors then specifically recognize the pTyr residues through SH2 domains [46]. To systematically link those modular domains to the specific phosphorylation sites that they recognize, Miller et al. [72] created NetPhorest, an atlas of consensus sequence motifs covering 179 kinases and 104 phosphorylation-dependent binding domains. NetPhorest utilizes phylogenetic trees to capture the similarity of binding domains as an indicator of their substrate specificity, as well as machine learning approaches to assess the relative affinities between domains and peptides [72]. Olsen et al. [66] tried to uncover the in vivo phosphorylation dynamics of signalling networks by using quantitative MS for the identification of phosphorylation sites as a function of stimulus, time and subcellular localization. They detected 6600 phosphorylation sites on 2244 proteins and quantified the temporal changes in phosphorylation after stimulating HeLa cells with epidermal growth factor. About half of the proteins were phosphorylated at multiple sites, showing different modification kinetics, which demonstrated that phosphorylation is generally regulated in a site-specific way and that different phosphorylation sites on the same protein usually serve different functions. In addition to the identification of many novel components of the epidermal growth factor signalling pathway and their temporal dynamics, the study revealed that phosphorylation is intimately tied to other regulatory systems, such as ubiquitylation, RNA processing and transcriptional regulation [66].

Current studies have shown that a combination of phosphoproteomics, phenotypic response profiles from biochemical perturbations and computational modelling is necessary to fully unveil the architecture and regulation of cellular signalling networks. For example, Bakal et al. [78], combined phosphoproteomics data with the results of a high-throughput RNA interference screen, testing > 17 000 double knockouts, and computational models of kinase specificity to identify regulators of the Jun N-terminal kinase in fly. Construction of a Jun N-terminal kinase phosphorylation network, integrating both genetic and kinase–substrate interactions, then provided structural and mechanistic insights into the systems architecture of that signalling pathway and revealed a potential link between Jun N-terminal kinase activity and remodelling of the cytoskeleton [78]. The utility of genetic interactions in elucidating the structure of signalling networks has also been shown by Fiedler et al. [79], who generated an epistatic miniarray profile consisting of ∼ 100 000 pairwise quantitative genetic interactions among kinases, phosphatases and regulatory proteins to create a systems view of signalling in yeast. Comparison with a literature-curated phosphorylation network revealed significant enrichment of positive genetic interactions (i.e. the double gene knockouts had better fitness than the single knockouts) between proteins involved in the same pathway, as well as between kinase–phosphatase pairs that share a common substrate. Negative genetic interactions, characterized by a double knockout with worse fitness than either single knockout, were found to be indicative of pairs of kinases and pairs of phosphatases that share a common substrate but are part of compensatory or alternative pathways. Genetic interaction maps thus represent a valuable complementary tool for resolving the architectures of signalling pathways [79].

Unstructured protein regions provide an additional level of regulation

It has been observed that most linear motifs in peptide-mediated interactions and residues targeted for PTMs are not embedded in globular protein domains but are rather found in linker regions [10,77]. Thus, in this respect, there seems to be a direct relationship between locally unstructured regions of proteins and the above described strategies for network regulation. However, the function of disordered regions in proteins goes beyond that of exposing peptide segments to recognition partners or to modifying enzymes that would create PTMs. Intrinsically unfolded proteins (IUPs) provide an additional regulatory strategy that has proven evolutionarily very successful in higher organisms. More than half of eukaryotic proteins have long segments (> 30 residues) that are not structured in their native isolated forms [80], and > 20% of eukaryotic proteins are expected to be fully unfolded [81]. IUPs are a recent acquisition in evolutionary terms and are much less common in Eubacteria and Archaea [82,83], which correlates well with the increase in regulatory complexity. Indeed, many IUPs are functionally associated with regulatory processes and signal transduction and are themselves strongly regulated by effects on the levels of synthesis and degradation as well as via PTMs, which is consistent with their regulatory role [84]. Thus, their direct connection with important pathologies in complex organisms, such as cancer [85] or cardiovascular [86] and neurodegenerative diseases [87,88], is not surprising.

Biophysical and evolutionary advantages of being unfolded

The essential role of IUPs in regulation arises from the unique thermodynamic advantage offered by existing as an ensemble of very different conformations in fast exchange [89,90] (Fig. 3). As mentioned above, most binding peptides are found in disordered regions of proteins, but the advantages that these offer are still the subject of debate. In terms of binding energy, a preformed binding interface offers the advantage of avoiding the entropic penalty associated with the restriction of the conformational space that necessarily occurs in an unfolded protein upon binding [91]. However, when a preformed binding stretch is in a free, unbound state, it lives in a frustrated state, and only binding of the ligand can relieve this frustration. A preformed binding site is therefore optimized to maximize binding but is not expected to be the best choice to maintain the balance between bound and free forms, which is characteristic of regulatory systems [92].

Figure 3.  Conformational selection and induced fit models. Functional proteins have the information content needed to perform their function. In folded proteins the information is stored in the 3D coordinates (and possibly the dynamics) of a major structure. In functional IUPs, however, this information is stored in the conformational ensembles and the possible interactions they enable. Often the information is read by a binding partner (shown in grey), either through conformational selection (increasing the population of a pre-existing conformation) or by induced fit (generating a new conformation to optimize the interaction with the binding partner).

The relative advantages of folded versus intrinsically unfolded recognition domains in protein–protein interaction networks can also be analysed for their capability to adapt to new demands [93,94]. Protein interaction networks are probably one of the motors of the fast evolution of higher organisms. On the one hand, fast evolution in response to new environmental changes is a key property of complex systems and interaction networks. On the other hand, networks are also involved in long-term evolution, because acquisition of a novel function requires evolution of the network itself by incorporating new nodes or acquiring new interactions. The substrate of evolution is variability, but the accumulation of mutations often conflicts with optimization under fixed conditions. In this sense, IUPs may offer a way to promote the evolution of protein interaction networks by providing an evolutionarily neutral environment in which substantial variability can be accumulated, from which in turn new interactions may appear. The evolutionary dynamics of IUPs can also be functionally associated with the coexistence of multiple similar conformations (redundancy), which parallels the role of gene duplication as a driving force for the evolution of new functions for old folds [94].

Mechanistically, IUP evolvability probably arises from the minimization of the number of internal contacts, which is a characteristic feature of this class of proteins. Residue-type bias in IUPs towards small nonstructuring residues and against large hydrophobic residues leads to a minimization of inter-residue interactions [95], which leads to low compactness values in so-called meta-structures [96]. Variability in a given sequence position is expected to minimally affect the overall properties of the complete protein. This is in contrast to the situation in folded domains, where most mutations lead to loss of function through denaturation, even when the affected residues are not directly responsible for the domain’s function. This association between protein disorder and mutational tolerance is supported by the correlation between sequence diversity and flexibility in different protein folds [97].

A new regulatory code for IUPs

As discussed above, the role of disordered protein regions in interaction networks seems to go beyond their ability to expose certain residues, and thus the multiple structures sampled by IUPs should not be boldly equated to a disordered state. Living organisms are characterized by a flow of information, and protein–protein interactions form the core of an information-processing system [98,99]. Disorder is often implicitly associated with lack of information, implying the inability to undertake complex functions. The specification of a set of coordinates to the atoms of a single 3D protein structure, out of a large number of a priori possibilities based only on the primary structure, implies a large information content on top of that included in the sequence itself [100]. The current paradigm equates static structure, order and information to a specific function, and the loss of structure to a loss of the necessary information to perform this function. However, the emerging paradigm, unveiled by the discovery of the widespread occurrence of IUPs performing key functions, is that the information content needed to perform a function is compatible with the coexistence of multiple structures. If we equate high information content with structure, the determination of the structure of an IUP would not be a paradox.

The question then arises as to how the information is stored in IUPs. It is known that many, although not all, IUPs adopt a folded conformation in the presence of a suitable partner [101]. In this case, the final information encoded in the complex is made explicit only after the encounter of the binding partners, which means that one needs to consider folding and binding together to decipher the structure–function relationship in IUPs [91]. Two main mechanistic models have been suggested (Fig. 3). The first, known as the conformational selection model, assumes that the active conformation is present in the ensemble that describes the isolated species and is stabilized by binding, shifting the populations [102,103]. Alternatively, the induced folding model postulates that the interactions occurring during the process of forming the complex drive the molecule towards the bound conformation [91,104]. The Levinthal principles also apply to IUPs, and these two models mainly differ in the unimolecular or bimolecular character of the folding funnel upon association. A synergistic model combining both previously described strategies has also been suggested [105]. The conformational selection model is illustrated by the interaction of the unstructured C-terminal domains of Sendai virus nucleoprotein and phosphoprotein. It has been demonstrated that the binding region of Sendai virus nucleoprotein forms a highly populated (∼ 70%) α helix in its unbound state that is further stabilized when forming a four-helix bundle protein assembly with phosphoprotein [106]. Therefore, the already stable helical stretch is selected for the interaction. In the binding and induced folding model, the most notorious example is the complex of the phosphorylated kinase-inducible domain with the KIX domain of the CREB-binding protein [107]. In this case, phosphorylated kinase-inducible domain folds on forming a complex with the KIX domain, and adopts a helical secondary structure even though it is completely unfolded in its free state.

The plasticity and bimolecular nature of the folding/binding event may lead to promiscuity of IUPs that are able to recognize different partners and even adopt diverse structures upon binding to different targets. A good example is the unstructured nuclear activation binding domain of the CREB-binding protein, which folds into α helices of different topologies upon binding to a p160 nuclear receptor co-activator [108] or the interferon regulatory factor IRF-3 [108]. Other cases of binding promiscuity are the C-terminal regulatory region of p53 [109], and the hypoxia-inducible factor 1α protein [101]. Different binding modes and promiscuity can have important consequences in the thermodynamics of protein–protein interactions. There is an entropic penalty associated with the disorder-to-order transition upon binding. This cost can be compensated for by a favourable enthalpy. The presence of preformed structured regions in the free-state precisely modulates the enthalpy–entropy balance and regulates the affinity of the interaction which is normally low. These particular properties are very well suited for signal transduction and regulatory processes which require binding to initiate the signaling, but also dissociation when the signal is complete.

Concluding remarks

Complex organisms, such as human, need a fairly high number of cellular processes to function, and these processes must be tightly regulated to ensure their coordinated action. This regulation is often carried out through intricate networks of transient protein–protein interactions able to produce complex responses to certain cellular states or environmental signals. However, the level of organism complexity reaches a point where any process would require a much higher number of proteins and interactions for its regulation and control than for the specific function itself. To overcome this paradox, nature has used one of its favourite and most effective tricks: modularity. The use of protein domains whose specific function is to facilitate interactions between proteins by recognizing specific sequence motifs, as well as the possibility of combining several of them, makes it possible to produce several thousand interactions with a few dozen of these domains. To extend the combinatorial possibilities even further, interactions occurring through the participation of such domains often depend on variable PTMs of certain residues in the sequence stretch that is being recognized. The combination of recognition domains and PTMs produces a sufficiently high number of interactions to ensure the required level of specificity and avoid cross-reactions. However, recognition alone is not enough to achieve the desired outcome, and these interactions have to display certain biophysical properties, such as a weak and reversible binding, that will permit the efficient transmission of a given signal. There is increasing evidence that protein structural disorder might play an important role in establishing these properties, which would represent yet a new indication of the strong relationship between protein structure and function. We believe that a more structured view of transient protein interactions will ultimately lead to a better understanding of the molecular bases of cell regulatory networks.

Acknowledgements

PA and MP acknowledge the financial support received from the Spanish Ministerio de Educación y Ciencia (BIO2007-62426 and BIO2007-63458) and the European Commission (LSHG-CT-2005-512028). RAP is a Spanish FPU fellowship recipient. PB has a Ramon y Cajal contract sponsored by the Spanish Ministerio de Ciencia y Innovación and co-sponsored by the IRB by funds provided by the Generalitat de Catalunya.