Signalling complexes and clusters: functional advantages and methodological hurdles

Cells have developed sophisticated molecular systems to sense information from the environment and transmit it into the intracellular space. These molecular systems have evolved to integrate, codify and transport information across the cell and process it into appropriate biological responses. Signalling pathways are high-fidelity, robust information-processing machineries that can distinguish sometimes weak signals from a very noisy environmental background with high precision and selectivity. It is now generally accepted that this is achieved by a multifaceted molecular network that is controlled by numerous negative and positive feedback and feed-forward loops (Brandman and Meyer, 2008; Jordan et al., 2000).

Although signalling pathways have been the subject of intensive studies for over 50 years, we are only beginning to understand the interconnections of this vast molecular network (Kholodenko, 2006). A major limitation to our understanding has been the experimental approaches that are used to investigate cellular events. The size of individual signalling molecules is well below the resolution of conventional light microscopy, and electron microscopy is incompatible with living cells, so biochemical techniques have long been the tools of choice to study signalling. This approach has successfully identified the armoury of signalling molecules such as receptors, enzymes, second messengers and simple ions. Biochemical approaches led to the discovery of post-translational modifications of signalling intermediates but, because they do not work at the level of single molecules in living cells and are instead in vitro approaches that average information from a large number of cells, these data lack spatiotemporal details. Only recently have new developments in optical and biophysical methods enabled significant improvements in the spatial and temporal resolution of live-cell imaging. Such advances have led to observations of often unexpected morphological and structural details and dynamics of cellular processes.

This Commentary summarises current views on how signalling molecules assemble in multimolecular complexes and clusters (Fig. 1), and how they use their physical properties to transduce environmental information into a large variety of cellular processes. Innovations in live-cell imaging techniques and the challenges of data processing, which can help us to understand the formation and lifetime of signalling assemblies and generation of characteristic biological outputs, are also discussed. The growing amount of literature describing small-molecule secondary messengers and the importance of their compartmentalisation for signalling has been summarised in numerous recent reviews (e.g. Baillie, 2009; Cahalan, 2009). Here, we focus on macromolecular interactions and larger multimolecular assemblies that are involved in the complex systems of cell signalling.

Signalling pathways need to have certain elementary properties to reliably maintain the organisation and function of cells and organisms. Sensitivity, reversibility, a capacity to be regulated and robustness are among the most crucial requirements. In addition, the signalling network not only broadcasts information within the cell but must also adapt the response to the cell's current state. Two key elements that are thought to drive signal transduction in cells are the localisation of proteins and their post-translational modifications.

Whereas the number of known post-translational modifications involved in cellular signalling is growing with the introduction of novel techniques, the role of free diffusion in the localisation of signalling proteins has been placed under scrutiny as more detailed information from live-cell studies becomes available (Bray, 1998; Bray et al., 1998; Doyle and Mamula, 2001; Eisenhaber and Eisenhaber, 2007). Diffusion is essential for spreading information across an ‘open space’ but, at the same time, it is too inefficient and of too low fidelity to be the main ‘driving force’ behind most macromolecular interactions in cells.

Interaction of two or more macromolecules undergoing three-dimensional ‘random walk’ diffusion in open space (Fig. 2A) depends on their concentration and ability to move rapidly over long distances (Phillips, 2009). Sub-nanomolar concentrations of growth factors (e.g. Garcia-Maya et al., 2006) and a low number of membrane-bound ligands (Irvine et al., 2002) have been reported to stimulate cellular responses. Therefore, it is likely that proximal signalling molecules interact at low concentrations, at least during early stages after cell stimulation. Understanding these events is further complicated by the multidomain nature of signalling macromolecules that form multimolecular complexes (e.g. signalosomes) in a highly oriented manner and in a crowded environment. The interactions that occur between freely diffusing small-molecule substrates (present in micromolar concentrations in cells) and enzymes (e.g. metabolic) are less influenced by intracellular crowding owing to a large difference in the size of solute and ‘crowders’ (Zhou et al., 2008).

The cell interior (i.e. the membranes, cytoplasm and nucleoplasm) is an environment with a high level of molecular crowding, which interferes with free diffusion, as has been demonstrated with some exemplar molecules (Dix and Verkman, 2008; Goulian and Simon, 2000; Rich et al., 2001; Weiss et al., 2004; Zhou et al., 2008). The presence of numerous structural barriers, such as membranes and filaments that undergo continuous changes (such as actin polymerisation), also drives intracellular components out of thermal Brownian (free) motion (Brangwynne et al., 2008; Luby-Phelps, 2000; Novak et al., 2009). It is therefore favourable to reduce ‘reaction space’ by compartmentalisation of signalling pathways. Intracellular organelles such as the Golgi complex, endoplasmic reticulum or mitochondria (Bivona et al., 2003; Chiu et al., 2002; Mironov et al., 2005) have been reported to function as transient confinement zones for molecules delivering specific signals. Similarly, cytoskeletal filamentous structures are involved in diverse multimolecular processes such as centriole formation and axonal signalling (Kuriyama, 2009; Perlson et al., 2005), thereby also acting to compartmentalise signalling molecules.

Membranes and cytoskeletal structures reduce the reaction space by one and two dimensions, respectively, and hence increase the probability that molecular interactions will occur. This is thought to be one reason that explains the large number of signalling molecules temporarily or permanently bound to membranes and filaments (Eigen, 1974; Phillips et al., 2009). However, without active intervention, the reduced dimensionality of membranes and filaments still only marginally increases the likelihood of interactions between individual molecules. Hence, additional factors such as the orientation of binding sites towards solute, cooperative binding of components and an increase in the half-life of interactions drive signalling molecules to assemble into complexes and clusters (Bray, 1998; McCloskey and Poo, 1986) (see below and Fig. 2).

One way of overcoming the low efficiency of random macromolecular interactions in the cellular environment is to form complexes of three or more elements in which at least one component interacts with more than one other molecule (Fig. 1, middle panel). This leads to the transient but favourable orientation of individual molecules within the complex, thereby allowing rapid interaction with additional components. On the other hand, such a system needs to define its stability (and therefore its lifetime) by balancing the strengths of intermolecular interactions within a complex in order to fulfil the requirement of reversibility (i.e. the ability to rapidly disassemble). Transient interactions between signalling intermediates and anchored complexes (e.g. on a membrane) are predicted to be more efficient than transient interactions with unanchored complexes (Fig. 2B) (Bray, 1998; Bray et al., 1998; Cho and Stahelin, 2005). The interactions between ‘free’ components and a complex are often responsible for switching the complex to an activated state, thus enabling the existence of pre-formed inactive complexes without unintentionally stimulating downstream events (e.g. Bray et al., 1998; Weiss and Littman, 1994). Such constitutive ‘silent’ complexes in resting cells have been observed experimentally (e.g. MEK-ERK complex) (Fukuda et al., 1997), and these are thought to accelerate the transmission of signals by rapid localisation of pathway components in one assembly, thereby reducing reaction space. One of the most effective features of signal transduction is thought to involve the coexistence of signalling complexes and freely diffusing intermediates dynamically interacting in a tightly controlled network (Fig. 2B) (Bray, 1995). Complexes with multiple freely diffusing components tend to generate various outputs required for diversification of cellular responses (such as lymphocyte differentiation) (Dustin, 2009a).

The existence of multimolecular complexes that drive essential cellular functions has been known for decades. The machineries of DNA replication, RNA transcription and protein translation, as well as Krebs cycle multi-enzyme complexes, have been well characterised and function as excellent models for studying how cells use these macromolecular assemblies (Stryer, 1995; Velot et al., 1997). The number of known macromolecular complexes that are involved in signal transduction is rapidly growing. Current experimental and theoretical work takes the non-stochastic molecular-crowding model of signal transduction further still, proposing the formation of larger clusters that are probably composed of several complexes (Fig. 1, right-hand panel) (Cho, 2006; Das et al., 2009b; Greenfield et al., 2009; Harding and Hancock, 2008).

Clustering of macromolecules and complexes provides efficient and rapid spatial confinement, thereby avoiding long-distance diffusion in the crowded cellular environment (Fig. 2C). The half-life of molecular interactions in clusters is prolonged by the high probability of re-binding owing to an increased local concentration of components and favourable orientation of binding sites (Fig. 2D) — this phenomenon might influence the quantitative and qualitative characteristics of signals broadcasted from these clusters. Clustering of signalling intermediates can also reduce the effects of negatively-regulating enzymes (e.g. phosphatases and proteases) by sterically reducing access to activated molecules, and therefore further stabilise the signalling pathway.

Molecular clustering was first proposed as a simple means by which to transmit signals, such as the transfer of an activating signal by transmembrane receptors, across the plasma membrane. This general principle has been best described for dimerisation of growth-factor receptors in metazoans (Plotnikov et al., 1999; Sako et al., 2000) and oligomerisation of chemotactic receptors in bacteria (Maddock and Shapiro, 1993). Dimerisation and oligomerisation of receptors have also been proposed to be involved in the activation of lymphocytes by immune receptors (Reth and Wienands, 1997; Schamel et al., 2005; Weiss and Littman, 1994). Clustering of plasma-membrane receptors is mainly driven by multivalent ligands (Fig. 2C) (e.g. Andrews et al., 2009; Fung et al., 2009; Tolar et al., 2005), but the existence of small pre-formed receptor oligomers has also been reported (Andrews et al., 2009; Livnah et al., 1999; Schamel et al., 2005).

The discovery of microscale assemblies of intracellular signalling intermediates has been facilitated by improvements in imaging techniques (Bunnell et al., 2002; Prior et al., 2003; Tian et al., 2007; Yokosuka et al., 2005). Clustering of membrane-associated proteins in often non-overlapping membrane domains has been observed as a general phenomenon in isolated plasma-membrane sheets of different cell types (Lillemeier et al., 2006; Wilson et al., 2001). However, the driving force behind the formation of intracellular signalling microclusters is not yet understood. Microclusters containing proximal signalling molecules often derive from receptor microclusters (Bunnell et al., 2002; Campi et al., 2005). Whereas membrane lipid microdomains (also known as lipid rafts) might also play a role in their formation and stabilisation (Cebecauer et al., 2009), transmembrane and cytosolic adapters that are involved in early signalling events can stabilise protein-protein interactions and thereby function as nucleation factors for signalling microclusters (Douglass and Vale, 2005). Indeed, most non-receptor microclusters have been defined in experiments in which adapter signalling molecules, such as LAT and SLP-76 in T cells, were imaged (Bunnell et al., 2002; Yokosuka et al., 2005).

Electron microscopy and single-particle video tracking studies of fibroblasts have demonstrated the existence of inner-plasma-membrane Ras nanoclusters with an average size of 5-20 nm across and a half-life of 0.1-1.0 seconds (Murakoshi et al., 2004; Prior et al., 2003). Total internal reflection fluorescence (TIRF) microscopy studies of lymphocytes on activating surfaces have allowed visualisation of T- and B-cell receptors (TCRs and BCRs) and associated proximal signalling molecules in microclusters that have a lifetime of up to several minutes and are ~1 μm across on average (Bunnell et al., 2002; Campi et al., 2005; Depoil et al., 2008; Yokosuka et al., 2005). The two observed types of clusters probably have distinct functions in the cell; short-lived Ras nanoclusters seem to be involved in rapid propagation of receptor signals to downstream intermediates, whereas microclusters observed in lymphocytes are hypothesised to have a role in the sustained signalling that is required for lymphocyte differentiation. Both types of cluster have the capacity to temporarily accumulate specific signalling intermediates and to segregate these from negative regulators that are present in the cell. However, the distinction between nano- and microclusters is a somewhat grey area; for example, neuronal receptors in stimulated synapses have been described to assemble in diverse macromolecular patterns, the relationship of which to nano- and microclusters is unclear (Newpher and Ehlers, 2008; Serge et al., 2002).

What is the distinction between simple signalling complexes and signalling nano- and microclusters? How many individual macromolecules are required to form a complex or cluster? Whereas only one or two copies of a specific component are usually present in a complex, seven Ras molecules were found, on average, in a nanocluster, independent of their activation state (see below) (Prior et al., 2003; Henis et al., 2009). It was estimated that approximately ten or more receptors compose a single microcluster on the surface of T and B cells (Depoil et al., 2008; Varma et al., 2006). Colocalisation studies indicate that these signalling clusters are formed from pathway components, as predicted by biochemical studies (Bunnell et al., 2002; Campi et al., 2005; Hibino et al., 2003; Tian et al., 2007; Yokosuka et al., 2005). In the following subsections, we discuss specific signalling clusters in more detail, with emphasis on Ras nanoclusters and microclusters that have been described in leukocytes.

Spatial segregation and clustering of signalling intermediates, together with positive feedback loops, provide an essential tool to transform generally analogue signals (such as a concentration gradient of growth factor) into digital or switch-like outputs. Digitalisation of analogue information received by surface receptors is mainly required for decisive cellular responses such as survival or death; it also reduces the sensitivity of a system to noise. Hancock and co-workers have focussed their attention on how high-fidelity Ras nanoclusters can function as an analogue-digital(-analogue) converter and why this cannot be achieved by individual signalling molecules (Harding and Hancock, 2008).

Ras-family proteins are small GTPases (there are three known Ras isoforms) and are probably the most extensively studied proteins for their ability to form signalling clusters. Ras proteins are essential components of signalling in various cell types, including growth-factor-receptor pathways in fibroblasts and antigen-receptor pathways in lymphocytes. Almost half of the plasma-membrane-bound Ras molecules have been found to be associated with distinct membrane nanoclusters under specific experimental conditions (Henis et al., 2009; Prior et al., 2003; Tian et al., 2007). Clustered, but not individually distributed, Ras proteins recruit and activate their downstream effector, Raf (Hibino et al., 2003; Tian et al., 2007). Interference with the clustering of Ras proteins was shown to completely abolish downstream signalling, demonstrating the importance of higher ordering of signalling molecules at the membrane (Cho, 2006; Plowman et al., 2005; Tian et al., 2007). A plausible explanation for this requirement is that clustered, but not individually distributed, Ras molecules significantly reduce the time required for downstream intermediates to interact with and be activated by Ras in the molecularly crowded cytoplasm (Fig. 2C). This brings us back to the aforementioned optimal signalling module, which is composed of a membrane-bound upstream complex (in this case a cluster) and freely diffusing downstream substrates or regulatory molecules (Bray, 1998). Positive feedback keeps clustered Ras molecules in the active state (see below). The interaction of downstream intermediates with such clusters leads to the activation of these clusters with high probability, independently of the stimulus strength, e.g. ligand concentration (Tian et al., 2007). In this manner, an analogue external stimulus is converted into a digital signal by increasing the number of activated Ras clusters, functioning here as ‘nano-switches’ (Harding and Hancock, 2008).

Nanoclustering has also been proposed to increase the robustness of signalling systems (Gurry et al., 2009; Harding and Hancock, 2008). In other words, the existence of a large number of nanoclusters containing several signalling units, together with the conversion of noisy analogue information into a digital switch-like message, means that the system is more resistant to the external and internal fluctuations that are generally present in a highly complex system such as a cell (Brandman and Meyer, 2008). Overall, the existence of nanoclusters enables high fidelity and robustness of the Ras pathway, traits that are essential for the crucial cellular decisions that this pathway determines.

Ras activation is also essential for lymphocyte development and lymphocyte effector functions (Genot and Cantrell, 2000), although the presence of Ras nanoclusters has not yet been demonstrated in these cells. In silico modelling combined with biochemical analysis suggests that switch-like signalling by the Ras pathway occurs in T cells (Das et al., 2009a). The Ras guanine-nucleotide exchange factors (RasGEFs) Sos and RasGRP have been shown to function in concert to enable a bistable (digital) response to a TCR stimulus. In addition, hysteresis (short-term signal memory) is generated by the Sos-Ras·GTP positive feedback loop, allowing sustained signalling in the presence of discontinuous stimulation (Das et al., 2009a). Computational analyses show that this positive feedback, which is responsible for the bistable response of the Ras pathway, favours an increase in cluster number or size (Das et al., 2009b). Segregation of Ras and RasGEFs into small plasma-membrane domains would also increase the efficiency and fidelity of the system, because new activated Ras molecules are rapidly generated at the boundary between the cluster and the inactive substrates that are homogenously distributed throughout the membrane. Clustering and increased fidelity of the system only occurs in the presence of slow reactant diffusion, emphasising the importance of membrane anchorage of Ras molecules and their effectors (Das et al., 2009b; Lommerse et al., 2006).

T cells recognise their ligands on the surface of antigen-presenting cells (APCs). Early contact between a T cell and an APC leads to large macromolecular rearrangements, which localise to a specialised lymphocyte-APC junction called the immunological synapse (IS) (Dustin and Cooper, 2000). Signalling events that take place in the IS, which were recently summarised and schematised by Morgan Huse, have illustrated the high level of molecular and structural complexity that underlies T-cell activation (Huse, 2009). Formation of an IS has also been observed for other immune cells, including stimulated B cells and natural killer cells (Batista et al., 2001; Davis et al., 1999). For technical reasons, the IS is typically oriented orthogonally to the focal plane (Fig. 3A) when imaged by conventional microscopy, which limits visualisation of some details. Mimicking the APC surface by immobilising activating antibodies on optical glass, and the use of supported planar bilayers containing lipid-anchored activating ligands, has enabled reorientation of the IS to the focal plane and an improvement in IS resolution (Fig. 3B,C) (Bunnell et al., 2001; Dustin et al., 1997). Tens of discrete receptor-ligand pairs modulate the lymphocyte response in vivo; however, only two or three types of adhesive and antigenic ligands are responsible for the activation of lymphocytes on APC-mimicking surfaces (Dustin, 2009b; Nguyen et al., 2008). The lipid-anchored ligands of supported planar bilayers have free lateral mobility, and supported bilayers therefore better represent the natural IS (Dustin, 2009b). A more physiological setting was developed by Oddos and co-workers, who used laser tweezers to reorient their cell conjugates to align the IS with the optical plane (Fig. 3A), improving the spatiotemporal resolution of microcluster visualisation (Oddos et al., 2008).

TIRF microscopy, a method for imaging fluorescence near the optical glass surface, has improved the level of detail that can be observed at the IS (see below) by reducing out-of-focus noise, and has led to the demonstration of microclusters containing TCRs or BCRs and downstream signalling intermediates in activated lymphocytes (Bunnell et al., 2002; Campi et al., 2005; Depoil et al., 2008; Tolar et al., 2009; Yokosuka et al., 2005). Such microclusters are assembled on the surface of lymphocytes within seconds of forming contact with an activating surface. New receptor-containing microclusters are generated in the periphery of the IS for more than 30 minutes in an actin-dependent process that is mostly independent of downstream signalling (Bunnell et al., 2002; Campi et al., 2005). Lymphocyte activation is then driven by the transient interaction of downstream signalling intermediates with receptor microclusters, which is dependent on intact upstream signalling (Bunnell et al., 2002; Varma et al., 2006; Weber et al., 2008; Yokosuka et al., 2005). In fact, any interference with microcluster formation or stability rapidly attenuates signal transduction and blocks lymphocyte activation (Ilani et al., 2009; Tolar et al., 2009; Varma et al., 2006; Weber et al., 2008). Together with the requirement for nanoclusters in growth-factor signalling, these data underscore the importance of macromolecule clustering for proper cell function.

Microclusters are dynamic, and the association and dissociation of their components occurs with varying kinetics for diverse signalling molecules. Neither association nor dissociation interferes with the overall stability of microclusters that have a lifetime of greater than 2 minutes. In addition to the association and dissociation of individual cluster components by diffusion into the cytosol (e.g. Grb2 and Gads of signalling microclusters) (Bunnell et al., 2002), the composition of microclusters can be changed by segregation and internalisation of new clusters containing specific molecules, which can trap downstream substrates and form distinct intracellular signalling assemblies (e.g. LAT-SLP76 clusters formed from receptor microclusters) (Barr et al., 2006; Bunnell et al., 2006). The network of receptor and signalling microclusters in lymphocytes is further extended by the segregation or independent formation of multimolecular assemblies that contain diverse surface receptors such as CD2 or integrins (Baker et al., 2009; Kaizuka et al., 2009). Such co-stimulatory microclusters often share some types of signalling molecules with the primary antigen-receptor microclusters. Whereas individual antigen-receptor microclusters might be insufficient to generate a variety of signals, assemblies that also contain co-stimulatory and/or cytokine receptors might have the capacity to modulate the final output by regulating the quality, level or localisation of individual signalling and effector events. The transient association and segregation of signalling microcluster components, together with the use of diverse adapters and scaffolding proteins, gives microclusters the capacity to regulate diverse processes, such as specific immune responses to a wide variety of pathogens (Baker et al., 2009; Janeway et al., 2001; Nguyen et al., 2008).

‘Spatiotemporal patterning’ has been suggested by Wülfing and co-workers as an alternative model to the formation of signalling assemblies to explain the highly dynamic localisation of molecules in zones of active signalling in lymphocytes (Singleton et al., 2009). The authors claim that, “with its diversity in time and space, spatiotemporal patterning will regulate the probability of interactions between signalling intermediates”. It cannot be inferred from their work what mechanism underpins local accumulation of the numerous studied signalling molecules in certain areas of the IS, because the model (effector T cells interacting with professional APCs) used in the work, although highly physiological, was not compatible with high-resolution microscopy (Singleton et al., 2009). The low spatial resolution of data for the IS oriented orthogonally to the focal plane leaves open the possibility that the assembly of signalling molecules in complexes and clusters is, at least in part, responsible for the observed transient local increase in concentration of signalling molecules. Therefore, these two models — the formation of signalling assemblies, and spatiotemporal patterning — might not be mutually exclusive.

Clustering of signalling molecules has also been observed in many other cell types, such as Escherichia coli, Drosophila melanogaster neuronal cells, and human neurons and dendritic cells (Featherstone et al., 2002; Fulcher et al., 2009; Greenfield et al., 2009; Serge et al., 2002). In bacteria, for example, clustering of chemotactic receptors has been shown to contribute to their ability to sense a wide range of often subtle changes in nutrient concentrations (Ames et al., 2002). In contrast to what was observed for Ras nanoclusters, no anchoring sites have been found for such clusters and a stochastic self-assembly process is thought to be responsible for their nucleation (Greenfield et al., 2009; Wang et al., 2008). Transient accumulation and clustering of receptors in synaptic membranes of human neurons, assisted by a meshwork of adapter and scaffolding proteins and the plasticity of such macromolecular rearrangements, facilitate the sensing of rapid changes in neurotransmitter concentrations between the presynaptic and postsynaptic sides. Indeed, the size of these clusters regulates the strength of the neural synapse. Cooperativity of synaptic receptors and ion channels in clusters then enhances the efficiency of excitatory and inhibitory neuronal transmission, and promotes the regulation of both processes (Levi et al., 2008; Newpher and Ehlers, 2008; Renner et al., 2008; Schutz et al., 2000). These data support the importance of macromolecular clustering for accurate information processing in living cells.

The rapid development of new imaging techniques has permitted us to address in greater detail the higher-order organisation of signalling molecules in living cells. Ever since Ernst Abbe identified the wavelength of light as a major limiting factor of image resolution in 1873 (Abbe, 1873) (Fig. 4A), microscopists have been inventive in ways to push this border. However, although the use of shorter wavelengths (in electron microscopy) has been incredibly successful for discovering structures at the nanoscale level, this technique is unfortunately incompatible with living cells. Recently, efforts have been concentrated on making improvements in visible-light microscopy. General developments in light microscopy related to live-cell imaging have been extensively reviewed in this journal recently (Frigault et al., 2009), so here we focus on fluorescence microscopy techniques suited to studying sub-micrometer signalling assemblies.

Fluorescence microscopy, in contrast to transmission light imaging, requires labelling of observed molecules or structures with fluorescent probes. Immunostaining of molecules in cells under study using antibodies with a covalently attached small organic fluorescent dyes or expression of a genetically modified fusion to a fluorescent protein allows specific detection of cellular structures. The first method is mainly employed for fixed cells, whereas the latter allows for live-cell microscopy applications.

Although Abbe's law makes it impossible to directly resolve individual molecular clusters by conventional light microscopy, fluorescence intensity fluctuations and single-molecule approaches have long been used to estimate the properties of individual molecules (Schutz et al., 2000; Walter et al., 2008). In fluorescence-correlation microscopy (FCS), these fluctuations are measured in the resolution-limited focal volume, and concentration and motility (or accumulation) of molecules are determined through statistical analysis. Another powerful indirect method to deduce interactions between molecules in complexes is Förster (fluorescence) resonance energy transfer (FRET). If the emission spectrum of one fluorophore overlaps with the excitation spectrum of another, the former can transfer its excitation energy to the latter (Ciruela, 2008). Because this energy transfer only occurs within a radius of ~3-6 nm, a detected FRET signal indicates that the two fluorescently labelled molecules are within close proximity, giving a resolution that is ~40× better than the diffraction limit. Both FCS and FRET have been successfully used to study signalling complexes and clusters in living cells (e.g. Lasserre et al., 2008; Sako et al., 2000; Tian et al., 2007; Tolar et al., 2009).

The most limiting dimension for resolution in optical microscopy is the z axis, which has a diffraction limit that is three- or four-times higher than in the optical xy plane (600-800 nm versus 200-300 nm, respectively; see Fig. 4). In TIRF microscopy, the limit of the z axis is pushed down to ~100 nm by excitation of only a thin layer above the glass coverslip (Axelrod, 1981). Light that is directed onto the boundary between the optically dense glass and the medium at an angle beyond the critical angle is reflected, except for an ‘evanescent wave’ that penetrates into the sample with exponentially decreasing energy (Fig. 5A). This dramatically reduces the observed imaging depth to ~100 nm. Although the resolution in the optical plane (xy) is not directly affected, the complete lack of out-of-focus light dramatically improves the signal-to-noise ratio (SNR). Visualisation of a sample using TIRF microscopy is restricted to a very narrow space above the optical glass, but this method is particularly useful for the observation of highly dynamic signalling microclusters on the surface of cells (Axelrod, 1981; Bunnell et al., 2002). When TIRF is used for imaging of cells on planar lipid bilayers enriched with stimulating molecules, the behaviour of specific proteins in the IS can be studied in great detail (Campi et al., 2005; Depoil et al., 2008; Yokosuka et al., 2005).

Recently, several methods based on limited fluorescence excitation have been developed for direct visualisation of sub-resolution structures, circumventing Abbe's diffraction law. Two main routes were taken to reduce the size of the detected volume: first, non-linear physics has been applied in the imaging techniques of stimulated emission depletion (STED) microscopy and non-linear saturated structured illumination microscopy (SSIM); second, sequential, statistical excitation of only a few single fluorophores spaced at distances larger than the diffraction limit has been applied in the techniques of photoactivation localisation microscopy (PALM) and stochastic optical reconstruction microscopy (STORM).

A non-linear physical effect is used to limit the excitation volume in STED microscopy (Hell and Wichman, 1994). Here, the sample is scanned with an excitation laser and a concentrically overlaid doughnut-shaped depletion laser at the emission wavelength of the fluorophore (Donnert et al., 2006; Hell, 2007). The high intensity of the depletion laser is sufficient to force fluorophores at the edges of the excited volume to the ground state by inducing stimulated emission of a coherent photon. This restricts the detected volume to the centre of the laser beam, and its size is only limited by the ratio of excitation:depletion laser intensities. The high laser energies that are needed for STED microscopy raise problems of phototoxicity, heat production and sometimes unexpected physical behaviour of fluorophores. This has so far prevented STED from becoming a mainstream technique and poses challenges particularly for its use in live-cell microscopy. By contrast, STED has successfully pushed the xy resolution into the macromolecular scale and been used to visualise cellular nanostructures (Fig. 5) (Sieber et al., 2007) or the dynamic accumulation of lipids and membrane proteins in nanometric domains (Eggeling et al., 2009).

Non-linear SSIM illuminates the sample with patterned excitation light at high intensity, leading to partial fluorophore saturation. The resulting non-linear fluorescence emission allows the decoding of underlying subresolution spatial frequencies from a series of diffraction-limited images (Gustafsson, 2005). SSIM requires several acquisitions for each image and high excitation intensity, which might impact on acquisition speed, fluorophore bleaching and photodamage. Application of SSIM has been used to provide detailed images of nanoscale cellular structures such as microtubules and nuclear pores (Kner et al., 2009; Schermelleh et al., 2008).

The statistical approach to reduce the excitation volume is based on photoswitchable fluorophores and, unlike STED and SSIM, works with low laser intensities, with the trade-off of low acquisition speeds. In both of the basic methods that apply this principle — PALM (Betzig et al., 2006) and STORM (Rust et al., 2006) — the sample is exposed to a low dose of light at the switching wavelength to activate only a small number of fluorophores that are statistically spaced at distances larger than the diffraction limit. The switched fluorophores are then imaged and the location of each fluorophore is calculated as the centre of the diffraction-limited spot. Subsequently, only the activated fluorophores are bleached, and a new round of low-light photoswitching is started. In this way, with multiple iterative cycles of photoswitching, imaging and bleaching, a high-resolution map of the fluorophores can be determined. Similar to STED microscopy, these techniques do not have a theoretical resolution limit, but the practical limits are the number of detected photons per fluorophore and the number of possible photoswitching cycles (Moerner, 2007). PALM and STORM have been used to achieve visualisation of macromolecular structures as small as ~20 nm (Fig. 5B), including nanoclusters of influenza virus haemagglutinin on the plasma membrane of fibroblasts (Hess et al., 2007) and clustering of chemotactic receptors in bacteria (Greenfield et al., 2009).

Although PALM and STORM are principally based on two-dimensional imaging of a single switchable fluorophore, three-dimensional (Huang et al., 2008; Shtengel et al., 2009; Vaziri et al., 2008) and multicolour (Bates et al., 2007; Shroff et al., 2007) adaptations have also been published. Similar progress has also been reported for STED microscopy (Meyer et al., 2008; Punge et al., 2008). Despite the severe time constraints of current nanoscopy techniques, some attempts have been made to study dynamic events in live cells at video-rate acquisition speed (Conley et al., 2008; Manley et al., 2008; Westphal et al., 2008). The contribution of available fluorescent probes to improving spatial resolution, and the directions of their further development, have been reviewed extensively (Fernandez-Suarez and Ting, 2008; Giepmans et al., 2006).

Whereas individual molecules that are diffusing move independently, components of macromolecular assemblies travel together for at least a certain period of time. Important information can therefore be obtained by tracking individual signalling molecules and clusters. The tracking of objects has been a challenge ever since the first efforts to visualise leukocyte motility and molecular motors (Comandon, 1917; Gelles et al., 1988; Haston and Wilkinson, 1988; Metchnikoff, 1861; Wilkinson, 1982). Open-source and commercial software packages (e.g. ImageJ and MetaMorph, respectively) provide appropriate tools for tracking small numbers of objects with high SNR. High-resolution imaging of signalling events often generates data with a high number of objects and low SNR (e.g. Bunnell et al., 2002; Yokosuka et al., 2005). However, temporary disappearance of objects under study from an image stream can complicate the analysis of their position and motility. Therefore, specific experimental settings require adaptation of existing tools or even the development of new scripts. Many of these are based on standard software platforms (e.g. MATLAB or ImageJ) and are available for further adaptation for specific purposes (Bonneau et al., 2005; Gelles et al., 1988; Jaqaman et al., 2008; Ji and Danuser, 2005; Manley et al., 2008; Sage et al., 2005; Sbalzarini and Koumoutsakos, 2005; Serge et al., 2008; Waterman-Storer et al., 1998).

Object (e.g. cluster) analysis and tracking based on low-SNR images is challenging for two main reasons. First, there is a spatial challenge: clusters should be sufficiently separated to be resolved individually and not cross trajectories with one another often. But the investigator usually needs statistical relevance (ideally exhaustive) for their findings, which requires imaging of large object numbers. Second, there is a temporal challenge: detection and estimation can be performed in an efficient way only for sufficient SNR, which typically requires longer exposure times. Conversely, long delays between images, which allow substantial displacements from frame to frame, obviously hamper tracking both by making the assignment for each track more difficult and because crossing of tracks can occur. For each experiment, the best compromise for object density and time interval needs to be established individually. The dyes, instrument setup and algorithm used might also need to be adapted for each application. Establishing dynamic nanoscopy will involve similar challenges.

An example of a de novo developed analytical tool for specific image analysis — namely, to decipher the non-randomness and non-homogeneity of plasma-membrane organisation — is multiple-target tracing (MTT) (Serge et al., 2008). This technique takes advantage of the high resolution provided by single-molecule microscopy and generates dynamic maps at high densities of tracked objects (Fig. 6; supplementary material Movies 1 and 2). Introducing deflation by subtracting detected peaks allows the detection of peaks of lower intensity. Detection of objects with performance-reaching theoretical limits can be achieved using MTT, which also efficiently reconnects trajectories by integrating the statistical information from the past events. Similar strategies were developed by Jaqaman and colleagues (Jaqaman et al., 2008). Both tools helped to provide essential information about molecular clustering and spatiotemporal dynamics of processes on the plasma membrane (Jaqaman et al., 2008; Serge et al., 2008).

In addition to object tracking, the spatial distribution of signalling proteins can be tested by statistical and clustering analysis software, enabling distinction between randomly distributed (e.g. at the plasma membrane) and clustered (or co-clustered) molecules. The reader is referred to an article summarising current approaches in such software development (Zhang et al., 2006).

Live-cell imaging of structures at the molecular level has been facilitated by recent developments in imaging and computational technologies. The existence and importance of signalling complexes and clusters has stimulated interest in their possible functions in regulating cellular processes. It was suggested that selective macromolecular assemblies, together with a network of feedback and feed-forward signals, are responsible for the high sensitivity, fidelity and robustness of signalling systems, traits that are required for efficient information transfer in a cell's noisy environment. Although the thermodynamic and kinetic advantages of molecular complexes and clusters have only recently begun to attract our attention, some groups have already developed computational tools to model these processes. However, there is still the need to accurately define biological questions and to use the best available technology to verify in silico data.

In addition to their docking and accumulating roles, complexes and clusters might function as depots for regulatory proteins (Ray et al., 2007). Diverse macromolecular assemblies might therefore exchange their components and in this way modulate the output of global signalling processes. Conversely, a majority of signalling molecules freely diffuse in the cytoplasm or nucleoplasm and form complexes and clusters only transiently, a property that imparts the advantage of rapidly adapting to environmental changes in a highly complex signalling network.

The transient character and variable composition of signalling assemblies present major challenges for imaging techniques and data processing. Current technology does not provide a simple solution for the study of signalling complexes and clusters. Hence, further progress — especially the enhancement of temporal resolution of nanoscopy and the availability of more specific and photostable probes — is required for detailed characterisation of content and dynamics of these entities in living cells. In addition, application of high-resolution imaging for the analysis of supramolecular assemblies is currently limited to experiments that use non-physiological stimulation of cells (e.g. activating supported planar bilayers for lymphocytes). Therefore, learning more about events that occur in a physiological setting will require a combination of high-resolution in vitro experiments and less structurally detailed in vivo studies (e.g. using intravital microscopy) as suggested recently (Dustin, 2009a). We expect that future development will lead to the availability of techniques to study cellular processes at the molecular level in vivo.

We thank Mike Ferenczi, Dan Davis, Tomas Brdicka and Christian Liebig for critical reading of the manuscript and helpful discussions. The research of M.C. and A.I.M. is funded by MRC (grant G0700771), M.S. and A.I.M. by BBSRC (CISBIC component grant), and A.S. by INSERM and CNRS.