To gain insight into the dynamics of how Ras transmits signals to downstream effectors under different network configurations or perturbations, we sought a dynamic in vitro reconstitution of Ras signal processing that would allow us to track effector outputs across multiple Ras turnovers. We reasoned that a microsphere surface charged with Ras could serve as a platform for the assembly and disassembly of fluorescent effector molecules from solution in response to inputs, much like the native Ras system (bound to the plasma membrane) functions in cells (Figure 2A). Signaling networks of defined composition could then be prepared from recombinant proteins and robust measurements of the dynamic output behavior could be determined by tracking the amount of effector on the surface over time for many individual beads and averaging their responses. Although such a system would not fully capture all biophysical features of cellular Ras signaling, such as GTPase diffusion in a fluid plasma membrane, partitioning between membrane microdomains, or GTPase exchange from the membrane, it serves as an excellent starting point to understand how these systems behave with fixed Ras molecules in a highly controlled setting (Tian et al., 2007; Silvius et al., 2006). Moreover, although the signaling activity of most Ras effectors is more complex than binding alone (see, for example: Jelinek et al., 1996; Stokoe et al., 1994), the regulated interaction of effectors with GTPase is the foundation on which any other complex signaling mechanisms will unfold, and thus represents a universal and fundamental feature of all Ras signaling systems that demands our understanding.

Figure 2 with 1 supplement see all

Download asset Open asset

We first asked whether we could observe GTP-dependent translocation of an effector molecule to a Ras-coated bead catalyzed by a guanine nucleotide exchange factor (GEF). For our initial studies, we chose to use the catalytic domain from the RasGRF GEF, which is constitutively active and, unlike other GEFs, contains no allosteric feedback sites (Freedman, 2006). Ni-NTA microspheres were charged with a 16x-histidine tagged H-Ras•GDP (OFF state) that could not dissociate from the bead and incubated in the presence of 50 nM (~K_D) of a model effector: a fluorescently tagged Ras-binding domain (RBD) from the C-Raf kinase (Block et al., 1996). Under these basal conditions, the amount of fluorescence on the bead was comparable to the background levels of fluorescence from the effector in solution. We then added as input 2 μM of the catalytic domain of the RasGRF GEF and 5 mM of either GDP or GTP and monitored the output of effector fluorescence on the bead (Figure 2B). This amount of nucleotide in solution is in vast excess of the small amount of bead-bound Ras present in the reactions, providing essentially an infinite supply of nucleotide for these reactions on the timescale we examine (detailed in 'Materials and methods'). Upon GEF and nucleotide addition, there was noticeable accumulation of fluorescent effector on the bead surface of the GTP containing reactions within seconds, and considerable fluorescent signal was observed by 10 min. In contrast, no fluorescent effector accumulated on the surface of reactions containing GDP, indicating that GEF-catalyzed translocation of the effector was dependent on Ras becoming GTP loaded. Having seen GTP-dependent GEF-catalyzed translocation of an effector to a Ras-charged bead surface, we were now in position to prepare signaling networks of arbitrary configuration and assay their output dynamics in multiplex using our microscopy-based assay (Figure 2C).

We first asked whether we could observe quantitative differences in the system’s output behavior when different amounts of GEF activity were used as inputs. When identically loaded beads were stimulated with increasing amounts of RasGRF, we observed both faster rates of effector translocation and higher steady state amplitudes of effector output (Figure 2D, Figure 2—figure supplement 1A–B). Initial rates of effector translocation appeared to show a hyperbolic response to increasing GEF, while steady states effector levels showed a linear response. Together, these data demonstrate that our reconstituted Ras signal processing system produced outputs that responded quantitatively to the amount of GEF present in the system.

In addition to being able to turn on, a dynamic reconstitution of Ras signal processing must be reversible and be able to turn off. To test the reversibility and turn-off of our system, we prepared beads loaded with H-Ras•GTP (ON state), incubated them with a saturating excess amount of C-Raf RBD effector (2.5 μM total, 50 nM fluorescently labeled), and monitored the loss of effector signal from the bead over time (Figure 2E). In the absence of any GAP, effector fluorescence decayed with a rate constant of 0.01 min^-1, which is similar to the expected rate of intrinsic hydrolysis by H-Ras under our assay conditions (Neal et al., 1988; Gibbs et al., 1984). This result is consistent with previous observations that intrinsic hydrolysis by the GTPase continues to occur even in the presence of saturating amounts of effector (Herrmann et al., 1995). When increasing amounts of the catalytic domain from purified Neurofibromin-1 GAP (NF1-GAP [Scheffzek et al., 1998; Xu et al., 1990]) were included in the reactions, effector signal disappeared from the bead at an increased rate in a dose-dependent manner with a hyperbolic dependence on the GAP concentration (Figure 2E, Figure 2—figure supplement 1C). Thus, as with the turn-on of the system, these data indicate that our reconstituted Ras signal processing system displays turn-off that responds quantitatively to the amount of GAP present in the system.

Having found that our system can produce effector outputs that are turned on by GEF and turned off by GAP, we wanted to verify that the system was truly multi-turnover and that the output dynamics would respond to both these activities working in concert. To this end, we incubated Ras•GDP-loaded beads with 50 nM fluorescent effector, initiated signaling with 2 μM RasGRF GEF and 5 mM GTP, and then added 1 μM NF1-GAP at 25 min (Figure 2F). As before, addition of GEF stimulated assembly of effector on the bead as the levels of Ras•GTP increased. When NF1-GAP was added to the reactions, the system responded with a rapid decrease in effector levels before stabilizing at a non-zero plateau corresponding to the non-equilibrium steady state maintained by the balance of effector, GEF and GAP activities present in the reaction.

Taken together, these data imply that our on bead reconstitution of H-Ras signal processing can semi-quantitatively track dynamic effector outputs across multiple cycles of Ras activation and deactivation during signaling. This system now puts us in position to explore how different mutational states, network configurations, protein identities, or feedback mechanisms affect signal processing by Ras GTPase systems.

Mutations of Ras (especially at the G12, G13, or Q61 positions) are frequently associated with cancer or other diseases (Barbacid, 1987). These alleles are primarily thought to impact Ras signaling through three mechanisms: 1) decreasing the intrinsic hydrolysis rate of the GTPase, 2) blocking GAP-mediated hydrolysis of the GTPase, and 3) potentially altering the interaction and preference of the GTPase for downstream effectors (Figure 3A) (Trahey and McCormick, 1987; Barbacid, 1987; Rajalingam et al., 2007; Smith and Ikura, 2014). The same mutant allele of Ras can elicit different phenotypes in different cell types and tissues. Thus, we wanted to use our in vitro systems reconstitution assay to determine which system configurations are most sensitive to these oncogenic perturbations.

Figure 3 with 2 supplements see all

Download asset Open asset

Using our dynamic, multi-turnover reconstitution of Ras signal processing, we examined how signaling networks bearing the G12V allele of the Ras GTPase distorted effector outputs relative to the wild-type Ras GTPase. By labeling wild type and G12V Ras GTPases with different fluorophores, we could distinguish beads loaded with each variant in a common solution of network components to see differences in effector outputs from each system side-by-side.

For these and future experiments, we display the output response data in two ways: (1) we show the absolute response, which conveys information about both the amplitude and the shape of the output response, and (2) we show the responses after normalizing to the maximum value of the response, which conveys information only about the shape or dynamic profile of the output. The latter is particularly useful for seeing how the shape of two responses differs when the amplitudes are substantially different.

With this approach, we first examined the output of 50 nM C-Raf RBD effector from G12V or wild-type Ras networks without GAP activity in response to a step input of 2 μM of the GEF RasGRF (Figure 3B, Video 1). Under this network configuration, wild type and G12V Ras systems produced very similar outputs with almost no difference in the total integrated effector output and only small differences in the overall dynamics of their responses. This suggests that neither the intrinsic hydrolysis nor changes in C-Raf RBD effector interactions of the G12V substitution is particularly perturbative to the output of the signaling system under this GAP-free network configuration.

Video 1

Download asset

We then looked at the system responses of wild type and G12V Ras systems to the exact same step-input (2 μM RasGRF GEF) but in networks that now included 1 μM basal NF1-GAP (Figure 3C, Video 2). Unlike in the GEF-only networks, both the dynamics and amplitude of the effector output were substantially distorted by the Ras-G12V allele. In this network configuration, wild-type Ras produced a transient response that peaked within an hour and declined to a steady state less than 20% its maximum value. In contrast, outputs from G12V were sustained and increased in magnitude for over 6 hr before settling at a steady state more than 40 times higher than wild-type Ras. Thus, the G12V mutation is significantly perturbative in a high-GAP network context.

Video 2

Download asset

Together these data and model imply that the balance of positive and negative regulatory activities in a signaling network impacts the severity by which Ras-G12V distorts signals. Similar results were also observed for G12C and Q61L alleles (Figure 3—figure supplement 1). To determine which particular configurations are most distorted by the G12V allele, we measured the effector output response across four different input strengths ([GEF] activity) and four different NF1-GAP levels. We then calculated a distortion score as the fold-change integrated output from Ras-G12V relative to wild-type Ras and interpolated the responses from these 16 configurations to produce a phase diagram of signal distortion by G12V under different network conditions (Figure 3D). This revealed that G12V alleles were most perturbative with low-GEF inputs and a high-GAP network context, conditions in which the GAP activity would, for wild-type Ras, completely dominate over the small amount of activating GEF input. These observations are consistent with models of oncogenic Ras signaling in which low-level inputs or noise from the environment that would normally be filtered out by basal GAP-activity are misinterpreted by the cell as bona fide activating signals.

Our comparison of wild type and G12V Ras signaling systems and our associated model illustrated the importance of the network composition in shaping signal processing outputs. Each individual network component is, in essence, a separate 'dial' of the Ras signaling system that can be turned by adjusting the concentration of that component (Figure 4A). Because expression levels of signaling components vary across different cell types and are often different in oncogenic states, we wondered how the level of each network component impacted signal processing wild-type Ras signaling networks. To this end, we fixed a particular input (2 μM RasGRF GEF) and, starting from a particular initial system configuration (~2500 Ras molecules × μm², 50 nM C-Raf effector, no GAP activity) with an associated output response, asked how titration of individual system components modulated the effector output.

Figure 4

Download asset Open asset

Diverse dynamic outputs achieved by titration of components

Our data show that the identity and concentration of each component in a Ras signaling network can have a profound impact on the timing, duration, shape, or amplitude of effector outputs. This implies by tuning the abundance and identity of network components and controlling the strength of inputs, a variety of diverse dynamic effector output programs can be realized by Ras signaling system (Figure 5A). To explore this space of output programs more thoroughly, we fixed a particular effector concentration (50 nM, ~K_D) and measured system output from different Ras densities, p120GAP concentrations, and input strengths ([GEF]). In total, we experimentally measured output responses for 96 configurations corresponding to four different input strengths, four different GAP concentrations, and six different Ras densities (Figure 5B, normalized responses in Figure 5—figure supplement 1).

Figure 5 with 3 supplements see all

Download asset Open asset

The diversity of dynamic output responses we observe highlights the versatility and tunability of the Ras signaling system: sustained responses of arbitrary amplitudes can be produced as well as transient responses with different peak times and magnitudes of overshoot, all by simply by co-varying the levels of different system parameters. Because the number of output responses we measured is large, we extracted three features from each output response trace that describe the behavior – integrated signal intensity, initial rate of activation, and a transient score that reflect the amount of overshoot in the response – and interpolated these values for three different Ras density bins to create an experimentally determined signal processing phase diagram for each output feature as a function of the network configuration (Figure 5—figure supplement 2).

These phase-diagrams not only summarize the output responses we measure but also clarify the structure of the space of Ras signaling behaviors. This provides a roadmap for understanding how output responses change as we alter system parameters and helps predict the impact of perturbations that move the system from one region of the space to another. The isoclines in these diagrams also highlight the existence of different network configurations with equivalent signaling behaviors. These correspond to neutral paths in network-space that the signaling system can drift along without immediate consequence to signaling output.

So far, our characterization of Ras signal processing has used the C-Raf RBD as the sole downstream effector, but in living cells these networks typically contain multiple effectors targeting different output responses that are in competition with one another for access to activated Ras, with each of these effectors possessing its own affinity for Ras•GTP and expression level in the cell (Herrmann, 2003; Smith and Ikura, 2014). Indeed, our kinetic modeling implied that competition was another important source of dynamic complexity in these systems (Figure 3—figure supplement 2). Because our reconstituted signal processing is microscopy based, we can track the behavior of multiple distinct competing effectors processing signals on the same bead simultaneously by labeling each effector with a different color fluorophore (Figure 6A). To this end, we purified and labeled RBDs from the A-Raf and B-Raf kinases, which have lower (k_off = 5.52x10^-4 s^-1, k_on = 7.20x10³ M^-1s^-1) and higher (k_off = 1.48x10^-4 s^-1, k_on = 1.32x10⁴ M^-1s^-1) affinities for Ras•GTP than C-Raf (k_off = 2.15x10^-4 s^-1, k_on = 1.02x10⁴ M^-1s^-1), respectively (Fischer et al., 2007), to examine the signal processing behavior of two-effector systems in either GAP-free or high NF1-GAP networks.

Figure 6 with 2 supplements see all

Download asset Open asset

We first considered networks containing equivalent, physiological amounts of C-Raf and B-Raf effectors, which both have high affinity for Ras•GTP (Smith and Ikura, 2014). In response to a 2 μM RasGRF GEF step input in a GAP-free network, C-Raf and B-Raf processed these signals with different amplitudes and completely different dynamics (Figure 6B). Initially, C-Raf and B-Raf outputs assembled at comparable rates, but within 1 hr C-Raf output peaked and began to decrease while B-Raf continued to increase in output monotonically over the entire time course. When these step-responses were re-examined in a high NF1-GAP network context, we continued to see different effector responses between C-Raf and B-Raf: C-Raf output peaked within 30 min before sharply declining to a steady state value 25% of its maximum. In contrast, B-Raf output peaked later at 1 hr, and declined to a 75% its peak maximum, a much higher steady state compared to C-Raf (Figure 6C). Thus, in this case, one effector (B-Raf) produces a transient response while a different effector (C-Raf) produces a more sustained output.

Why do different effector molecules interpret upstream Ras signals with different output dynamics as we observed? Two factors contribute to this phenomenon. First, as we previously saw with one-effector systems, the time it takes for a binding process to equilibrate impacts the extent to which it can track the transient dynamics of its target. Distinct effector molecules have different binding affinities that are determined by different on and off rates from the target, and thus will interpret Ras dynamics differently. Second, these effector molecules are in competition with one another for access to the supply of time-varying activated Ras. Because we are observing the nonequilibrium binding dynamics of effectors to Ras, an effector may be competitive in the short-term (kinetically determined phase) but less so in the long-term (equilibrium-determined phase) (Motulsky and Mahan, 1984). Appropriately, the inclusion of competing effectors with physiological parameters in our model produced dynamics recapitulating such different effector responses (Figure 6—figure supplement 2).

We then performed a similar analysis for two-effector systems containing equivalent, physiological amounts of C-Raf and A-Raf RBDs, where A-Raf has a weaker affinity for Ras•GTP and higher off-rate from the GTPase than C-Raf (Smith and Ikura, 2014). In GAP-free networks, the step-responses for A-Raf and C-Raf differed most significantly in terms of their amplitudes, with A-Raf output 20 times lower than C-Raf output (Figure 6—figure supplement 1). The normalized traces of each effector output show only minor differences in the dynamics of these outputs. In high NF1-GAP networks, we could only detect an output response from the C-Raf effector; the ability of A-Raf to assemble on the bead was either filtered out by the low affinity combined with high levels of GAP activity, or was simply too weak to detect above the background A-Raf in solution (Figure 6D).

Clearly, effector molecules with distinct identities result in differential interpretation of inputs, but how easy is it for a given effector molecule to acquire a new dynamic behavior? Given that A-Raf, B-Raf, and C-Raf have different affinities for the GTPase, we wondered whether mutations in C-Raf that changed its affinity would be sufficient to drive new dynamic behaviors. To this end, we characterized the step-response of two-effector network containing equivalent amounts (50 nM) of the C-Raf RBD and the C-Raf^N64A point mutant that has decreased affinity for Ras•GTP (Block et al., 1996). Different dynamic output behaviors for each effector were observed for both GAP-free (Figure 6—figure supplement 1) and high NF1-gap (Figure 6G) networks. In both contexts, the lower affinity C-Raf N64A mutant peaked earlier and had lower overall amplitude than the wild-type C-Raf RBD, and like before, high GAP-networks accentuated these temporal differences. Consistent with this, we found that parameter changes as small as 2-fold could produce subtle shifts in the timing of effector reponses of our kinetic model (Figure 6—figure supplement 2). This shows that new dynamic behaviors are readily realized by mutation of an effector molecule.

Taken together, this analysis shows that distinct effector molecules can perceive the same input to a Ras signaling system with different dynamics and amplitudes depending on their affinities and biochemical properties. Consequently, a single-step input can be in principle be used encode multiple classes of temporally distinct outputs that peak and decline out of phase with one another, allowing for a sequence of different activities to be organized during signal processing. For example, we were able to produce a three-wave activation response of three distinct effectors in our kinetic model by simply modifying concentrations and off-rates (Figure 6—figure supplement 2). Furthermore, the context of other regulators (e.g. extent of GAP activity in the network) can influence how these different dynamic responses unfold, magnifying temporal distinctions in some cases while restricting the ability of certain effectors to assemble productively at all in other cases. Finally, because even simple point mutations to an effector can dramatically alter its output dynamics, new dynamic patterns are not difficult to produce and can be easily accessed during evolution.

The signaling networks we have examined thus far are solely the product of constitutive enzymatic activities and effector assembly processes unfolding in the simplest possible Ras GTPase signaling circuit. Our analysis found that in high-GAP systems, the 'ground state' output for a step-response will transiently overshoot the final steady state. In some instances, this behavior could be useful for the cell, for example to create an adaptive response or to produce distinct temporal phases in multiple downstream effector outputs; in other instances this overshoot behavior could prove undesirable, for example if the overshoot provoked a proliferative response to non-proliferative level of input.

Many cellular circuits modulate intrinsic behaviors of a signaling system by including additional layers of regulation and feedback control that could alter the signaling properties of the system. To gain insight into how such regulation might alter the ground state signaling behavior of Ras GTPase systems, we examined the effect of introducing two different modes of GTPase→GEF positive feedback (defined as active Ras promoting more activation of Ras) on system signaling behavior.

Recruitment-based feedback mechanism (GTPaseON → GEF localization) amplifies weak inputs

One common mode of generating positive feedback in signaling is through recruitment. For example, in yeast, the GTPase Cdc42, when activated, recruits its own GEF Cdc24 (via the scaffold Bem1), thus leading to further Cdc42 activation (Butty et al., 2002). Similar positive feedback GEF recruitment could occur in Ras signaling networks. To explore the effects of recruitment-based positive feedback, we made a synthetic GEF in which we fused the C-Raf RBD effector domain to the RasGRF GEF to produce 'RasGRF-RBD'. In this case, the catalytic activity of RasGRF-RBD is always constitutive, but activated Ras will assemble the synthetic GEF on the bead surface to provide a higher effective concentration of GEF and potentially increase the apparent GEF activity (Figure 7A). Thus, this feedback mechanism takes an ON GEF molecule and makes it MORE ON as Ras•GTP levels increase. We experimentally measured output responses for this feedback-containing system in 96 system configurations corresponding to four different input strengths, four different GAP concentrations, and six different Ras densities (Figure 7B).

Figure 7 with 1 supplement see all

Download asset Open asset

The inclusion of recruitment-based positive feedback in the system had considerable consequence for both the dynamics and amplitude of the effector output’s we observed. For example, in high GEF / high GAP regimes, multiple local maxima in the output dynamics were observed (Figure 7—figure supplement 1). Interesting behaviors were also observed for the output amplitudes. For example, because the fluorescent output-effector and the feedback-effector-GEF compete with one another, signaling amplitude did not increase monotonically with increasing INPUT strength: at the highest level of GEF INPUT will examined (2000 nM), signal was substantially lower than at 1000 nM GEF INPUT, and more comparable in amplitude to the 200 nM GEF INPUT.

How do these differences alter the signaling properties of the system compared to not having this feedback present? Manual inspection of the curves from the system with or without feedback suggested that recruitment-feedback was particularly effective at increasing output in high GAP network contexts. For example, high GEF / low GAP system configurations produced similar responses with or without feedback (Figure 7C), but feedback was critical to produce an output from a low GEF / high GAP system configuration (Figure 7C). To address this systematically, we calculated a feedback gain for each network configuration we experimentally examined as the ratio of the integrated signal from the feedback system to the non-feedback system. We then interpolated these experimentally determined values for two different Ras density levels to produce phase diagrams showing the impact of feedback on different network configurations (Figure 7D).

These data imply that this type of feedback mechanism can produce strong effects in certain network configurations but have little to no effect in other configurations. In particular, networks that contained high amounts of basal GAP-activity but only small amounts of input GEF showed the strongest differences in signal. These correspond to regimes in which the amount of GAP activity in the system dominates the small amount of intrinsic catalytic activity of the GEF, but does not overcome the small amount of localized GEF activity arising from the RasGRF-RBD feedback GEF (Figure 7C). In contrast, the behaviors of the wild type and feedback systems were most similar under high GEF, low GAP network configurations. In this regime, the GEF activity from the catalytic domain is sufficient to provide strong activation, and any additional boost in activity for localizing the GEF provides only marginal gains (Figure 7D). Thus, this type of feedback mechanism seems most powerful for amplifying signals arising from weak inputs in a high turnover background.

Allosteric positive feedback mechanism dampens system overshoot

For the second feedback mode, we replaced the RasGRF GEF with SOScat, the catalytic domain from the Son of Sevenless GEF. SOScat has an intrinsic feedback mechanism built in, such that the molecule has very low GEF activity in the absence of Ras•GTP, but Ras•GTP binding to a distal site on SOScat allosterically stimulates GEF activity to very high levels (Margarit et al., 2003; Freedman, 2006). Thus, in contrast to the synthetic RasGRF effector fusion (which is always ON), this feedback mechanism takes an OFF GEF molecule and makes it ON in response to Ras•GTP (Figure 8A). We experimentally measured output responses for this feedback-containing system in 96 system configurations corresponding to four different input strengths, four different GAP concentrations, and six different Ras densities (Figure 8B, normalized responses in Figure 8—figure supplement 1).

Figure 8 with 1 supplement see all

Download asset Open asset

Figure 9

Download asset Open asset

Figure 9—source data 1

Relative gene expression level data from a variety of human tissue and cell types that was used to produce the plot in Figure 8A.

This table contains the relative expression-level data that was used to prepare the plot in Figure 9A. These data were obtained from Genevestigator as outlined in the main-text 'Materials and methods'.

https://doi.org/10.7554/eLife.12435.024

Download elife-12435-fig9-data1-v2.xlsx

These data show a very different effect of the allosteric SOScat feedback on signaling outcomes than was observed for the recruitment-based feedback of the RasGRF-RBD fusion. Unlike with recruitment-based feedback, allosteric-feedback did not provide large gains in output amplitudes in any of the regimes we looked like. However, we did observe that transient overshoot behavior was almost completely absent in the output responses of the SOScat system, even in the highest p120GAP condition we inspected. Indeed, the output dynamics from allosteric-feedback networks appeared much more stable and monotonic than those from RasGRF networks that contained no feedback. These effects were most dramatic in high GAP network contexts in which RasGRF produced a large transient overshoot phase, whereas SOScat produced a monotonic approach to the steady state (Figure 8C). However, this increased stability was still noticeable even for low GAP high GEF networks in which RasGRF produces a sustained response (Figure 8C).

These differences in comparison to the RasGRF-RBD positive feedback system can be at least partially attributed to the OFF→ON feedback mechanism of SOScat in the context of a step-input: for a RasGRF GEF step-input, the system experiences a sudden increase in catalytic activity all at once; for SOScat GEF, although, this step-input in protein levels becomes a ramp-input in terms of protein activity because it takes multiple rounds of GEF activity to produce enough Ras•GTP for SOScat to achieve significant activity (Figure 8D). Thus, this type of feedback mechanism appears to play less of a role in amplifying weak signals and more of a role in dampening dynamic and transient features that would arise naturally in the step-response without this feedback. For SOScat, which mediates Ras activation in response to growth signals, this overshoot minimization might be a desirable feature that reduces the risk of inappropriate proliferation. This use of biochemical feedback is somewhat analogous to the use of feedback control used in some electronic systems to avoid overshoot of the steady-state in response to step-inputs.

In this work, we developed a multi-turnover reconstitution of Ras signaling to explore the space of dynamic output behaviors that could be produced by Ras GTPase systems and to characterize how each network component contributes to these behaviors. Using these assays explored how different perturbations such as oncogenic mutation, component levels, inclusion of additional effector molecules, or introducing positive feedback altered the landscape of available outputs.

Our experiments imply that, much in the same way that a single genome can encode multiple cell types that are regulated through differential gene expression, a single signaling system like Ras can encode multiple dynamic signal processing behaviors by regulating the concentration and identity of network components. This regulation can be direct by acting at the level of gene expression. For example, a simple survey of published p120GAP, Ras, and c-Raf mRNA expression levels across a variety of tissue types reveals a staggering amount of diversity in what types of network configurations are present in different cell types and tissues (Figure 9A, Table 1). The true diversity in these configurations is likely even greater given the plethora of additional GEFs, GAPs, Ras variants and effectors that cells can deploy. Regulating the concentration of these activities can also be achieved indirectly by the differential recruitment of these molecules by the receptors that initiate Ras signaling, which changes their effective concentration at the plasma membrane. Indeed, many of the catalytic domains that we looked at in this study show regulated interaction dynamics with the plasma membrane in response to extracellular signals (Gureasko et al., 2010). Thus, different cells can position their signaling systems at different points in the space of available Ras network configurations and modulate these configurations in response to extracellular cues to provide versatile top-level control of the amplitude and duration of proximal signal processing events (Figure 9B).

This versatility is not without trade-offs, however. In particular, we observed many different paths in network-space from one signaling processing behavior to another with much higher or sustained amplitude (Figure 9C). These paths include classic oncogenic substitutions like G12V in Ras, but can also be realized by increased GEF activity, decreased GAP activity, or inclusion of high-affinity effectors that increase, extend and sustain signaling responses. While some of these perturbations have not been definitively recognized as drivers of cancer, many are associated with other RASopathies in humans, like Noonan syndrome or type 1 Neurofibromatosis (Schubbert et al., 2007; Bollag et al., 1996). Thus, the same flexibility that allows Ras systems to realize many different signaling behaviors creates many opportunities for misregulation in response to perturbation.

How any particular perturbation distorted signaling output was highly dependent on network configuration. This was most obvious from comparing the effect of Ras G12V perturbation on GAP-free and high-GAP networks (see Figure 3) but is also readily apparent from inspection of the structure of our experimental maps between network configuration and signaling outputs (see Figure 5—figure supplement 2). Even networks configurations that produced highly similar output behaviors could nonetheless respond divergently to perturbations. These are network configurations along a contour in the phase diagrams and represent neutral paths along which a cell or species can move without immediate consequence to the system. For example, a low signaling output that is maintained by a weak GEF activity alone might also be produced by a higher GEF activity balanced by a high GAP activity. However, these two configurations would respond very differently to substitution with oncogenic alleles of Ras like G12V. This observation demonstrates the limited predictive power of static steady-state measurements of cellular states and highlights the need to obtain dynamic data about the pre-steady state and impulse-response behavior of cellular systems using fine-grained time courses or new methods such as optogenetic pathway activation (Toettcher et al., 2013).

The dependence of a perturbation on network configuration can also afford cells new opportunities that may be positive rather than deleterious. As an example, our analysis of the feedback gain produced by the RasGRF-RBD fusion revealed that some configurations had almost no impact on the system output, while others were highly impacted and produced little to no signal without the presence of feedback (see Figure 7B–D ). Thus, access to certain regions of this space is not feasible without first acquiring permissive modifications to the feedback architecture of the system. Acquisition of this permissive architecture though, can occur in regions in which there is minimal consequence to the system output. Once this feedback mechanism is present, the space of behaviors available to the system changes and previously non-functional regions of the space can now be accessed.

By interrogating the space of available behaviors to a signaling system in an unbiased way as we have in the present work, we learn not only what the behavior of any particular system configuration is, but also how systems respond to change and what paths exist to travel to new configurations with new behaviors. For Ras, this space appears rich with dynamic possibilities and sufficient neutral network structure to provide evolution with ample fodder to facilitate the use of Ras for the wide array of diverse signaling roles at it plays across different cell types and species, but at the risk of harmful perturbation by diseased alleles or expression states.

One striking observation from this work was the importance of effector molecules in determining how a dynamic Ras•GTP signal is interpreted. This is, in fact, a critical aspect of how these particular signaling systems work as activated Ras itself has no enzymatic activity toward other molecules, but instead serves only as a platform for the recruitment of many possible competing effector molecules within the cell. Moreover, activated Ras cannot engage more than one effector simultaneously and thus competition between effectors as well as upstream regulators like GAPs contributes to the system’s output dynamics.

For simple one-effector systems that we studied with the C-Raf RBD effector, the concentration of effector shaped not only the amplitude of the output, but also the dynamics of that output as well (see Figure 4D). This is because higher effector concentrations equilibrate against their target faster than lower effector concentrations, resulting in different abilities to capture transient features of the target dynamics. Effector concentration is thus more than a passive 'volume' knob that reports on Ras•GTP levels and instead is an active system component that interprets Ras activity to sculpt effector-specific output dynamics owing to its own assembly and disassembly kinetics.

The importance of this property of effectors was even more apparent in two-effector Ras signaling systems, in which we found that equivalent amounts of different competing effectors interpreted the same system inputs with markedly different outputs that differed both in amplitudes as well as in duration and dynamics (see Figure 6). These systems showed a variety of interesting multi-effector programs, such as one effector exhibiting a transient response while another was sustained, one effector responding while another did not at all, or two transient responses that peaked and declined with different timing and duration. Moreover, we found that different dynamic behaviors could readily be produced by introducing point mutations in an effector that altered affinity for activated Ras. These observations stress the important roles that both kinetic and thermodynamic aspects of effector assembly and competition play in shaping how an individual effector interprets dynamic Ras•GTP levels in the context of the rest of the network.

An interesting consequence of these different effector behaviors and dynamics is that it can naturally result in the temporal partitioning of distinct activities during a signaling response. This can allow some effector outputs to be restricted to early phases of signaling, only to decline and be displaced by other more dominant effectors at later stages. These observations extend recent observations of hierarchies of binding by different effectors to Ras under equilibrium conditions with non-hydrolyzable analogs (Smith and Ikura, 2014). Thus, the differential perception of Ras•GTP signals by distinct effectors may not be a flaw in the method by which cells make measurements, but a useful feature by which cells can use a single upstream signaling molecule like Ras to dictate a complex temporal program of multiple downstream outputs.

Binding to activated Ras is only the first step in signal propagation for many effectors. For example, binding of Raf kinases we used in this study to Ras primarily serves to deliver the kinase to the plasma membrane where its interaction with lipids (Ghosh and Bell, 1997), other Raf kinases (Freeman et al., 2013), scaffolds (Brennan et al., 2011; Ritt et al., 2006) and other macromolecules alters its kinase activity and thus how it sends out signals downstream (REF). In fact, Raf can even activate downstream signaling in the absence of Ras by artificial membrane recruitment (Stokoe et al., 1994; Leevers et al., 1994). However, this only further emphasizes the importance of effector binding dynamics in the context of cellular signal processing, as binding to Ras is a physiological prerequisite for these other mechanisms to take place. Moreover, different Ras effectors such as PI3 Kinase or Ral-GDS will have their own molecule-specific layers of regulation that take place upon interaction with Ras at the plasma membrane. These processes will be influenced by the underlying effector-binding dynamics in different ways depending on the kinetics of these downstream steps. Our work demonstrates that cells have simple systems for modulating and controlling these fundamental binding dynamics and further indicates that known control mechanisms should be analyzed in this complex dynamic context.

Another layer of dynamic and regulatory complexity is also likely to arise as more classes of Ras GTPases are included in the signaling networks. Indeed, our present study has only investigated effector interaction with H-Ras, but there exist many additional Ras isoforms such as K-Ras and N-Ras, K4A-Ras, and K4B-Ras that may engage these effectors in different ways to produce different dynamics. Additionally, there exist related GTPases such as Rap which can serve as platform for Ras effectors but that do not necessarily promote signal propagation (Wynne et al., 2012; Cook et al., 1993). Understanding how the underlying distribution of GTPase isoforms dictates signal processing behavior is another critical component of cellular signaling and we hope to extend our in vitro system to explore these fundamental questions in the future.

The network level biochemical approach to interrogating signaling systems we employed in this study occupies a relatively underexplored area in our understanding of cellular decision making systems, but is similar to approaches used to understand dynamic mechanical systems in cells like microtubules. Indeed, because additional complexities can emerge when multiple energetically driven processes are coupled together to promote the dynamic assembly and disassembly of competing effectors, exploring how these systems behave in vitro under different configurations sheds new light on the phenomenology of how biochemical signaling devices function and respond to perturbation.

The ability to prepare non-equilibrium steady states in which Ras is actively cycling between ON and OFF states may also prove useful in developing new strategies to ameliorate erroneous signaling associated with diseased states. For example, the fact that we could reconstitute radically different signaling behaviors for wild type and G12V Ras under high GAP conditions is consistent with the notion that wild-type Ras ordinarily cycles quickly but G12V does not. This difference in the lifetime of a Ras•GTP molecule compared to RasG12V•GTP could potentially serve as an additional dynamic selectivity handle for small molecules in which we want to only target the oncogenic form of Ras. The assays we developed are well suited to compare how small molecules differentially impact signaling through these different forms of Ras side by side under active energy-consuming conditions.

More generally, the simplicity of the approach we present here paves the way for further studies on other types of non-equilibrium signaling systems that center around the assembly of molecules from the cytoplasm on a surface such as the other members of the Ras superfamily like Rho, Rac, Cdc42, and Rab GTPases, as well as other completely different multi-turnover signaling systems like receptor tyrosine kinases or lipid kinases. Some aspects of the H-Ras system may be shared with these systems, while other aspects may be different owing to idiosyncratic features of a particular system of molecules. These systems could also be extended in other ways as well to explore how other biophysical constraints impact these signaling processes. Our reconstitutions could, for example, be extended to lipid-coated beads to explore how membrane fluidity or lipid identity shape effector outputs. Multi-currency networks that include multiple Ras isoforms or contain more than one type of GTPase could also be examined to look at higher-order networks and cascades. Only by building these systems, turning them on and watching them run can we begin to understand how they actually perform and operate in different signaling regimes.

A combination of standard and custom ImageJ macros were used to prepare the primary image data for further analysis. First, drift in the stage was corrected using a macro based around the MultiStackReg plugin. Two or three color multi-tiff timecourses were split into separate channels. Matrix transformations to register timecourses were obtained using the constant GTPase fluorescence on every bead from the blue channel. These matrices were then used to register the timecourses of the red or green channels. The three channels were then recombined to produce the properly registered multi-tiff used for analysis of the beads. The ImageJ macro code was tweaked depending on the particular experiment, but a representative example of the code is shown below:

matrix path= "C:\Users\scoyl_000\Google Drive\MICROSCOPY\150403\matrix\";

inputpath = "C:\Users\scoyl_000\Google Drive\MICROSCOPY\150403\raw\";

outputpath = "C:\Users\scoyl_000\Google Drive\MICROSCOPY\150403\processed\";

function scott_register(input,output,filename)

{

open(input+filename);

run("Split Channels");

run("MultiStackReg", "stack_1=[C3-" + filename + "] action_1=Align file_1=["+matrixpath+"matrix.txt] stack_2=None action_2=Ignore file_2=[] transformation=[Rigid Body] save");

run("MultiStackReg", "stack_1=[C1-" + filename + "] action_1=[Load Transformation File] file_1=["+matrixpath+"matrix.txt] stack_2=None action_2=Ignore file_2=[] transformation=[Rigid Body]");

run("MultiStackReg", "stack_1=[C2-" + filename + "] action_1=[Load Transformation File] file_1=["+matrixpath+"matrix.txt] stack_2=None action_2=Ignore file_2=[] transformation=[Rigid Body]");

run("Merge Channels...", "c1=C1-" + filename + " c2=C2-" + filename + " c3=C3-" + filename + " create");

}

list = getFileList(inputpath);

for (i = 0; i < list.length; i++)

scott_register(inputpath, outputpath, list[i]);

We then corrected for uneven sample illumination or background artifacts using a rolling ball background substraction of 50 pixels. These processed images were subsequently used to analyze the signaling behavior of each bead.

To analyze the images, beads were either identified and stored as regions-of-interest (ROIs) automatically from the GTPase fluorescent signal on beads using a custom macro that makes use of ImageJ’s 'find particles' function, or in the case of experiments with beads that deliberately contain differing levels of Ras density, beads were identified and stored as ROIs by hand. The particular parameters of the automated bead finding varied depending on the particular experiment, but a representative ImageJ macro is shown below:

inputpath = "C:\Users\scoyl_000\Google Drive\MICROSCOPY\150403\processed\";

outputpath = "C:\Users\scoyl_000\Google Drive\MICROSCOPY\150403\beads\";

function scott_findbeads(input,output,filename)

{

open(input+filename);

run("Split Channels");

selectWindow("C3-" + filename);

run("Smooth", "stack");

//set threshold and find particles

Stack.setPosition(1, 18,1);

setSlice(20);

setThreshold(3744, 23285);

run("Analyze Particles...", "size=400-15000 circularity=0.60-1.00 display exclude clear include add slice");

//close any and all open windows

close();

close();

close();

//reopen original file

open(input+filename);

//transfer ROIs to overlay and save

run("From ROI Manager");

saveAs("Tiff", output+filename);

//clear manager

roiManager("Delete");

close();

}

list = getFileList(inputpath);

for (i = 0; i < list.length; i++)

scott_findbeads(inputpath, outputpath, list[i]);

Once beads were identified and stored as ROIs, a variety of measurements were made for each bead using a custom macro. We measured the average total Ras fluorescence in the ROI, and the average fluorescent effector signal in the ROI at every timepoint in the experiment. We also measured the area and perimeter of the bead. As with other macros, the exact details of the code varied depending on the number of effectors we were simultaneously examining or other aspects of the setup, but a representative ImageJ macro is shown below:

inputpath = "C:\Users\scoyl_000\Google Drive\MICROSCOPY\150318\one_shot_processed\";

outputpath = "C:\Users\scoyl_000\Google Drive\MICROSCOPY\150318\one_shot_data\";

setBatchMode(true);

list = getFileList(inputpath);

for (i = 0; i < list.length; i++)

crunchimage(inputpath, outputpath, list[i]);

setBatchMode(false);

//The Functions that are used in the macro are below

function crunchimage(input,output,filename)

{

//open the image

open(input+filename);

//clear log file and reulsts table

run("Clear Results");

print("\Clear");

//import ROIs from overlay

run("To ROI Manager");

//count ROis

count=roiManager("count");

//record Ras AMPS and RBD timecourse for each ROI

for(i=0;i<count;i++)

{

// first record the AMP for the Ras field;

recordAMP(i);

// next calculate the time series for the data

recordTimecourse(i);

// print newline marker

print("!");

}

selectWindow("Log");

saveAs("Text", output+filename);

close();

}

function recordAMP(index)

{

//function that reports area, mean intensity, and perimieter

// of an ROI used to get everything EXCEPT the timecourse data

run("Clear Results");

roiManager("Select",index);

// select channel 2 and select the midpoint of hte stack

Stack.setPosition(2, 18,18);

// make measurements

run("Measure");BeadArea=getResult("Area",0);BeadMean=getResult("Mean",0);BeadPerim=getResult("Perim.",0);print(BeadArea+","+BeadMean+","+BeadPerim+",");

}

function recordTimecourse(index)

{

run("Clear Results");

roiManager("Select",index);

sliceCount=nSlices()/2;

// print(sliceCount);

for(k=0;k<sliceCount;k++)

{

Stack.setPosition(1, 18,k+1);

run("Measure");

timepointK=getResult("Mean",k);

print(timepointK+",");

}

}

The output of this macro is a file that contains a list of every single bead trace in the multi-tiff image. Each trace begins with the area measurement of the bead, the mean total Ras intensity of the bead, and the perimeter measurement of the bead, followed by the effector measurement of the bead at every timepoint. Each measurement ends with a ',' and is on a newline. At the end of each bead trace, a stop marker '!' is printed. These data are then transformed by GREP and shell script into a CSV where each line contains all the relevant information about each single bead’s signaling trace. These data can then be loaded into either Matlab or Excel for further analysis.

Once in Excel, data were typically further analyzed as follows: (i) intensity measurements were normalized to perimeter instead of area, (ii) the time-series data for a given bead was normalized such that the time-zero effector measurement was zero, and (iii) single bead traces were binned based on total GTPase levels to obtain statistics on the signaling behavior. Beads were assigned to the nearest of the 6 Ras density beads: 150 molecules / μm², 300 molecules / μm², 600 molecules / μm², 1200 molecules / μm², 2500 molecules / μm², 10,000 molecules / μm². The individual bead traces within a given bin were then averaged together to produce an average response for the associated density bin and network configuration. Each trace was typically the average of 15–80 beads from the combination of two independent experiments. The standard error of the mean for a given trace was typically <15%.

Kintek Student Explorer (Johnson et al., 2009) was used to simulate the dynamic behavior of a variety of models for GTPase activation. Time was modeled in seconds and concentrations in nanomoles. Rate constants for association and dissociation of molecules from the GTPase were based on published Biacore measurements (Fischer et al., 2007), and catalytic rate constants for GEF and GAP activities were based published solution measurements (Bollag and McCormick, 1991; Freedman, 2006). GEF was not modeled explicitly but rather directly incorporated in the rate constant for nucleotide release. Each simulation was allowed to run for 42000 s (~700 min).

Three models were initially explored. The first model was a two-state GTPase model that did not account for competition between GAP and effector. This was modeled in Kintek using the following equations and parameters:

The second model was a two-state GTPase model that explicitly modeled competition between GAP and effector. This was modeled in Kintek using the following equations and (physiological) parameters:

The third model was a three-state GTPase model that included an additional post-hydrolysis GTPase state (GI) which was refractory to GEF activation. This state converts to the GDP form on a slow timescale. This was modeled in Kintek as:

The third model was the best at explaining the transient behaviors of the system that we observed as well as the differences between WT and G12V Ras, and thus was used as the basis of all subsequent modeling and simulations. For any given simulation in the text, the initial conditions and any changes to associated rate-constants are indicated in the figure legend.

Relative log-transformed expression levels for p120GAP, C-Raf, and H-Ras across a variety of cell-types and tissue-types were obtained from data contained within the Genevestigator software package (see data in Table 1). The three-dimensional phenotypes associated with each cell or tissue type was plotted as a 3D scatterplot using Matlab.

The funders had no role in study design, data collection and interpretation, or the decision to submit the work for publication.

SMC was supported by a National Science Foundation Graduate Research Fellowship. This work was supported by NIH grants RO1 GM55040, PN2 EY016546, P50 6M081879 (WAL), and the Howard Hughes Medical Institute (WAL). We thank T Anooki, L Bugaj, R Gordley, RS Isaac, A Mitchell, G O’ Donoghue, K Roybal, M Thomson, J Walter, and A Weeks for helpful discussions.

1. Received: October 20, 2015
2. Accepted: January 13, 2016
3. Accepted Manuscript published: January 14, 2016 (version 1)
4. Version of Record published: February 23, 2016 (version 2)

© 2016, Coyle et al.

This article is distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use and redistribution provided that the original author and source are credited.