Dendritic trafficking faces physiologically critical speedprecision tradeoffs
 Cited 6
 Views 1,751
 Annotations
Abstract
Nervous system function requires intracellular transport of channels, receptors, mRNAs, and other cargo throughout complex neuronal morphologies. Local signals such as synaptic input can regulate cargo trafficking, motivating the leading conceptual model of neuronwide transport, sometimes called the ‘sushibelt model’ (Doyle and Kiebler, 2011). Current theories and experiments are based on this model, yet its predictions are not rigorously understood. We formalized the sushi belt model mathematically, and show that it can achieve arbitrarily complex spatial distributions of cargo in reconstructed morphologies. However, the model also predicts an unavoidable, morphology dependent tradeoff between speed, precision and metabolic efficiency of cargo transport. With experimental estimates of trafficking kinetics, the model predicts delays of many hours or days for modestly accurate and efficient cargo delivery throughout a dendritic tree. These findings challenge current understanding of the efficacy of nucleustosynapse trafficking and may explain the prevalence of local biosynthesis in neurons.
https://doi.org/10.7554/eLife.20556.001eLife digest
Neurons are the workhorses of the nervous system, forming intricate networks to store, process and exchange information. They often connect to many thousands of other cells via intricate branched structures called neurites. This gives neurons their complex treelike shape, which distinguishes them from many other kinds of cell.
However, like all cells, neurons must continually repair and replace their internal components as they become damaged. Neurons also need to be able to produce new components at particular times, for example, when they establish new connections and form memories. But how do neurons ensure that these components are delivered to the right place at the right time? In some cases neurons simply recycle components or make new ones where they are needed, but experiments suggest that they transport other essential components up and down neurites as though on a conveyor belt. Individual parts of a neuron are believed to select certain components they need from those that pass by. But can this system, which is known as the sushibelt model, distribute material to all parts of neurons despite their complex shapes?
Using computational and mathematical modeling, Williams et al. show that this model can indeed account for transport within neurons, but that it also predicts certain tradeoffs. To maintain accurate delivery, neurons must be able to tolerate delays of hours to days for components to be distributed. Neurons can reduce these delays, for example, by manufacturing more components than they need. However, such solutions are costly. Tradeoffs between the speed, accuracy and efficiency of delivery thus limit the ability of neurons to adapt and repair themselves, and may constrain the speed and accuracy with which they can form new connections and memories.
In the future, experimental work should reveal whether the relationships predicted by this model apply in real cells. In particular, studies should examine whether neurons with different shapes and roles finetune the delivery system to suit their particular needs. For example, some neurons may tolerate long delays to ensure components are delivered to the exactly the right locations, while others may prioritize speedy delivery.
https://doi.org/10.7554/eLife.20556.002Introduction
Dendritic and axonal trees of neurons often have many tens or even thousands of branches that can extend across the entire nervous system. Distributing biomolecular cargo within neuronal morphologies is therefore a considerable logistical task, especially for components that are synthesized in locations distant from their site of use. Nonetheless, molecular transport is important for many neurophysiological processes, such as synaptic plasticity, neurite development and metabolism. For example, longlasting forms of synaptic plasticity appear to depend on anterograde transport of mRNAs (Nguyen et al., 1994; Bading, 2000; Kandel, 2001) and specific mRNAs are known to be selectively transported to regions of heightened synaptic activity (Steward et al., 1998; Steward and Worley, 2001; Moga et al., 2004) and to developing synaptic contacts (Lyles et al., 2006).
On the other hand, local biosynthesis and component recycling are known to support dendritic physiology, including some forms of synaptic plasticity (Kang and Schuman, 1996; Aakalu et al., 2001; Vickers et al., 2005; Sutton and Schuman, 2006; Holt and Schuman, 2013) and maintenance of cytoskeletal, membrane and signalling pathways (Park et al., 2004, 2006; Grant and Donaldson, 2009; Zheng et al., 2015). Neurons therefore rely on a mixture of local metabolism and global transport, but the relative contributions of these mechanisms are not understood. Analyzing the performance of global trafficking provides a principled way to understand the division of labor between local and global mechanisms.
In this paper, we examine how well trafficking can perform given what we know about active transport and the typical morphologies of neurites. There are two parts to this question. First, how can active transport achieve specific spatial distributions of cargo using only local signals? Second, how long does it take to distribute cargo to a given degree of accuracy and what factors contribute to delays?
Intracellular trafficking is being characterized in increasing detail (Buxbaum et al., 2014b; Hancock, 2014; Wu et al., 2016). Microscopic cargo movements are stochastic, bidirectional, and inhomogeneous along neurites, leading to to the hypothesis that trafficking is predominantly controlled by local pathways that signal demand for nearby cargo, rather than a centralized addressing system (Welte, 2004; Bressloff and Newby, 2009; Newby and Bressloff, 2010a; Doyle and Kiebler, 2011; Buxbaum et al., 2015). These local signals are not fully characterized, but there is evidence for multiple mechanisms including transient elevations in secondmessengers like $C{a}^{2+}$ and ADP (Mironov, 2007; Wang and Schwarz, 2009), glutamate receptor activation (Kao et al., 2010; Buxbaum et al., 2014b), and changes in microtubuleassociated proteins (Soundararajan and Bullock, 2014).
A leading conceptual model ties together these details by proposing that local signalling and regulation of bidirectional trafficking determines the spatial distribution of cargo in neurons (Welte, 2004; Buxbaum et al., 2015). Doyle and Kiebler (2011) call this the ‘sushi belt model’. In this analogy, molecular cargoes are represented by sushi plates that move along a conveyor belt, as in certain restaurants. Customers sitting alongside the belt correspond to locations along a dendrite that have specific and potentially timecritical demand for the amount and type of sushi they consume, but they can only choose from nearby plates as they pass.
Stated in words, the sushi belt model is an intuitive, plausible account of the molecular basis of cargo distribution. Yet it is unclear whether this model conforms to intuition, and whether it implies unanticipated predictions. Can this trafficking system accurately generate global distributions of cargo using only local signals? Does the model predict crosstalk, or interference between spatially separated regions of the neuron that require the same kind of cargo? How quickly and how accurately can cargo be delivered by this model, given what is known about trafficking kinetics, and do these measures of performance depend on morphology or the spatial pattern of demand?
We address these questions using simple mathematical models that capture experimentally measured features of trafficking. We confirm that the sushibelt model can produce any spatial distribution of cargo in complex morphologies. However, the model also predicts that global trafficking from the soma is severely limited by tradeoffs between the speed, efficiency, robustness, and accuracy of cargo delivery. Versions of the model predict testable interactions between traffickingdependent processes, while the model as a whole suggests that timecritical processes like synaptic plasticity may be less precise, or less dependent on global transport than is currently assumed.
Results
A simple model captures bulk behaviour of actively transported cargo
Transport along microtubules is mediated by kinesin and dynein motors that mediate anterograde and retrograde transport, respectively (Block et al., 1990; Hirokawa et al., 2010; Gagnon and Mowry, 2011). Cargo is often simultaneously bound to both forms of motor protein, resulting in stochastic backandforth movements with a net direction determined by the balance of opposing movements (Welte, 2004; Hancock, 2014; Buxbaum et al., 2014a, Figure 1A). We modelled this process as a biased random walk, which is general enough to accommodate variations in biophysical details (Bressloff, 2006; Bressloff and Earnshaw, 2007; Müller et al., 2008; Bressloff and Newby, 2009; Newby and Bressloff, 2010a; Bressloff and Newby, 2013).
Figure 1 shows this model in a onedimensional cable, corresponding to a section of neurite. In each unit of time the cargo moves a unit distance forwards or backwards, or remains in the same place, each with different probabilities. In the simplest version of the model, the probabilities of forward and backward jumps are constant for each time step (Figure 1B, top panel). Cargo can also undergo extended unidirectional runs (Klumpp and Lipowsky, 2005; Müller et al., 2008; Hancock, 2014). The model can account for these runs with jump probabilities that depend on the previous movement of the particle (Figure 1B, bottom panel, Materials and methods).
While the movement of individual cargoes is stochastic, the spatial distribution of a population (Figure 1C) changes predictably. This is seen in Figure 1D, which shows the distribution of 1000 molecules over time, without (top panel) and with (bottom panel) unidirectional runs. The bulk distribution of cargo can therefore be modelled as a deterministic process that describes how cargo concentration spreads out in time.
A convenient and flexible formulation of this process is a massaction model (Voit et al., 2015) that spatially discretizes the neuron into small compartments. In an unbranched neurite with $N$ compartments, the massaction model is:
where ${u}_{i}$ is the amount of cargo in each compartment, and ${a}_{i}$ and ${b}_{i}$ denote trafficking rate constants of cargo exchange between adjacent compartments. This model maps onto the wellknown driftdiffusion equation when the trafficking rates are spatially homogeneous (Figure 1E; Smith and Simmons, 2001). We used this to constrain trafficking rate constants based on singleparticle tracking experiments (Dynes and Steward, 2007) or estimates of the mean and variance of particle positions from imaging experiments (Roy et al., 2012, see Materials and methods).
With a compartment length of 1 μm, the simulations in Figure 1D gave mean particle velocities of 15 μm per minute, which is within the range of experimental observations for microtubule transport (Rogers and Gelfand, 1998; Dynes and Steward, 2007; Müller et al., 2008). The variances of the particle distributions depended on whether unidirectional runs are assumed, and respectively grew at a rate of ~0.58 and ~1.33 μm^{2} per second for the top and bottom panels of Figure 1D. The mass action model provides a good fit to both cases (Figure 1F). In general, the apparent diffusion coefficient of the model increases as run length increases (Figure 1—figure supplement 1A). The accuracy of the massaction model decreases as the run length increases. However, the model remains a reasonable approximation for many physiological run lengths and particle numbers, even over a relatively short time window of 100 s (Figure 1—figure supplement 1B).
Biophysical formulation of the sushi belt model
The advantage of the mass action model is that it easily extends to complex morphologies with spatially nonuniform trafficking rates, and can accommodate additional processes, including sequestration of cargo. The sushibelt model (Doyle and Kiebler, 2011) proposes that local mechanisms modify local trafficking rates and capture cargo as it passes. For these local signals to encode the demand for cargo, some feedback mechanism must exist between the local concentration of cargo and the signal itself. There are many biologically plausible mechanisms for locally encoding demand (see Materials and methods). For our main results, we did not focus on these details and simply assumed a perfect demand signal. We have thus addressed the performance of the transport mechanism per se, with the most forgiving assumptions about the reliability of the demand signal.
The mass action model of sushibelt transport is:
where $u$ represents the concentration of cargo on the network of microtubules, indexed by the compartment. In each compartment, molecules can irreversibly detach from the microtubules in a reaction $u}_{i}\stackrel{{c}_{i}}{\to}{u}_{i}^{\star$, where $u}_{i}^{\star$ denotes the detached cargo. Biologically, cargo will eventually degrade. However, in this study we are concerned with how cargo can be rapidly distributed so that detached cargo can satisfy demand for at least some time. Therefore, for simplicity we assume degradation rates are effectively zero.
We first asked whether modifying the trafficking rates alone was sufficient to reliably distribute cargo. Thus, we set all detachment rate constants (${c}_{i}$) to zero, and considered a model with trafficking only between compartments, as shown in Figure 2A. Mathematical analysis shows that, for a fixed set of trafficking parameters, the distribution of cargo approaches a unique steadystate distribution over time, regardless of the initial distribution of cargo. The steadystate occurs when the ratio of cargo concentrations between neighboring compartments is balanced by the trafficking rates:
where ${u}_{p}$ is the level in a ‘parent’ compartment (closer to soma), ${u}_{c}$ is the level in the adjacent ‘child’ compartment (closer to periphery) and $b$ and $a$ are the trafficking rate constants between these compartments.
If $\stackrel{~}{u}}_{i$ represents the local demand signal in compartment $i$, then Equation (3) gives the condition for cargo distribution to match demand:
An example demand profile and the corresponding trafficking rate relationships are shown in Figure 2B. This condition ensures that cargo is delivered in proportion to local demand. The absolute concentration at steadystate is determined by the total amount of cargo produced (Figure 2—figure supplement 1); in the case of mRNA, this might be controlled at the somatic compartment by transcriptional regulation. In this paper, we focus on the relative accuracy of cargo distribution when some fixed amount of cargo is produced at the soma.
To illustrate demandmodulated trafficking in a realistic setting, we used a reconstructed model of a CA1 pyramidal neuron (Migliore and Migliore, 2012). To provide a demand signal, we modelled excitatory synaptic input at 120 locations within three dendritic regions (red dots, Figure 2C) and set demand, ($\stackrel{~}{u}}_{i$), equal to the average membrane potential in each electrical compartment (see Materials and methods). As expected, cargo was transported selectively to regions of high synaptic activity (Video 1), matching the demand profile exactly at steady state (Figure 2D). Therefore, local control of trafficking rates (equivalently, motor protein kinetics) can deliver cargo to match arbitrarily complex spatial demand.
Transport bottlenecks occur when trafficking rates are nonuniform
We next investigated the consequences of solely modifying trafficking rates to distribute cargo. A particularly striking prediction of this model is that changes in trafficking (or, equivalently, demand signals) in regions close to the soma can strongly affect cargo delivery times to distal sites. As the demand signal $\stackrel{~}{u}}_{i$ approaches zero in a compartment, the trafficking rates into that compartment also approach zero, cutting off the flow of cargo along the neurite (Figure 3A). The smallest demand signal, $\u03f5$, often determines the ratelimiting time constant for cargo delivery to an entire dendritic tree. We refer to this scenario as a ‘transport bottleneck.’ Figure 3A–C illustrate how decreasing $\u03f5$ to zero causes arbitrarily slow delivery of cargo in a simple threecompartment model.
To illustrate bottlenecks in a more realistic setting, we imposed a bottleneck in the reconstructed CA1 model by setting demand in the middle third of the apical dendrite to a lower level than the rest of the dendritic tree, which was set uniformly high. As expected, the cargo distribution converged much more quickly for uniform demand than with a bottleneck present (Figure 3D).
However, less intuitive effects are seen on the convergence times of cargo in specific compartments. Figure 3E plots convergence time for ${u}_{i}$ to reach a fraction of the steady state value for each compartment. While distal compartments showed prolonged convergence times, (Figure 3E, upper right portion of plot), the bottleneck shortened the transport delay to proximal compartments (Figure 3E, lower left portion of plot). This occurs because the bottleneck decreases the effective size of proximal part the neuron: cargo spreads efficiently throughout the proximal dendrites, but traverses the bottleneck more slowly.
Another counterintuitive effect is seen when demand varies independently at proximal and distal locations, as might occur during selective synaptic stimulation (see e.g., Han and Heinemann, 2013). In Figure 3F we simulated demand at proximal and distal portions of the apical dendrite independently and quantified the total convergence time. Proximal demand alone (Figure 3F ‘proximal’) resulted in the fastest convergence time. Convergence was slowest when the demand was restricted to distal dendrites (Figure 3F, ‘distal’). Interestingly, when both distal and proximal sites signalled demand (Figure 3F ‘both’), convergence was substantially faster than the distalonly case, even though cargo still needed to reach the distal neurites. Uniform demand across the entire tree (Figure 3F ‘entire cell’) resulted in a similarly short convergence time.
Together, these results show that locally modulating trafficking movements will have testable effects on global transport times. The presence and relative contribution of this mechanism can be probed experimentally by characterizing the convergence rate of a cargo that aggregates at recently activated synapses, such as Arc mRNA (Steward et al., 1998). This could be achieved using quantitative optical measurements in combination with synaptic stimulation at specific synaptic inputs.
Local control of trafficking and detachment results in a family of trafficking strategies
We next considered the full sushibelt model (Equation 2) with local demand signals controlling both trafficking and detachment rate constants (Figure 4A). This provides additional flexibility in how cargo can be distributed, since the model can distribute cargo by locally modulating trafficking rates, detachment rates, or both (Figure 4B). If trafficking is much faster than detachment ($a,b\gg c$), then the previous results (Figures 2–3) remain relevant since the distribution of cargo on the microtubules will approach a quasisteady state described by equation (3); cargo may then detach at a slow, nonspecific rate (${c}_{i}=$ constant, with $c\ll a,b$). Figure 4C shows an example of this scenario, which we call demanddependent trafficking (DDT). The spatial distribution of cargo is first achieved along the microtubules (red line, Figure 4C), and maintained as cargo detaches (blue line, Figure 4C).
Alternatively, models can match demand by modulating the detachment process rather than microtuble trafficking. In this case, the trafficking rates are spatially uniform (${a}_{i}={b}_{i}$) so that cargo spreads evenly, and the detachment rates are set proportionally to the local demand, $\stackrel{~}{u}}_{i}^{\star$:
The result of this strategy, which we call demanddependent detachment (DDD), is shown in Figure 4D. Unlike DDT, DDD avoids the transport bottlenecks examined in Figure 3, and can achieve target patterns with $\stackrel{~}{u}}_{i}^{\star$ equal to zero in certain compartments by setting ${c}_{i}=0$.
Mixed strategies that locally modulate both detachment and trafficking are also able to deliver cargo to match demand. Figure 4E shows the behavior of a model whose parameters are a linear interpolation between pure DDT and DDD (see Materials and methods).
Rapid cargo delivery in the sushibelt model is errorprone
Although it is mathematically convenient to separate the timescales of trafficking and detachment in the model, this separation may not exist in biological systems tuned for rapid transport. However, removal of timescale separation in the sushibelt model results in mistargeted delivery of cargo, as we now show.
We returned to the CA1 model of Figures 2–4 and considered a scenario where there is demand for cargo at the distal apical dendrites (Figure 5A). If the detachment rate constants are sufficiently slow, then, as before, delivered cargo matched demand nearly exactly in both the DDT and DDD models (Figure 5A, left). Increasing detachment rates led to faster convergence, but resulted in cargo leaking off the microtubule on the way to its destination (Figure 5A, right). Thus, for a fixed trafficking timescale, there is a tradeoff between the speed and accuracy of cargo delivery. The tradeoff curve shown in Figure 5B shows that both accuracy and convergence time decreased smoothly as the detachment rates were increased. This tradeoff was present regardless of whether the trafficking rates (Figure 5B, red line) or detachment rates (Figure 5B, blue line) were modified to meet demand (compare to Figure 4C and D, respectively). However, DDD outperformed DDT in this scenario, since the latter caused bottlenecks in proximal dendrites.
We considered a second scenario in which there was a uniform distribution of demand throughout the entire apical tree (Figure 5C). As before, fast detachment led to errors for both transport strategies, this time by occluding cargo delivery to distal synaptic sites (Figure 5C, right). A smooth tradeoff between speed and accuracy was again present, but, in contrast to Figure 5A–B, the DDT model outperformed DDD (Figure 5D). Intuitively, DDT is better in this case because DDD results in cargo being needlessly trafficked to the basal dendrites.
Together, these results show that increasing the speed of cargo delivery comes at the cost of accuracy, and that the performance of different trafficking strategies depends on the spatial profile of demand. The balance between demanddependent trafficking and detachment could be probed experimentally. For example, one could perform an experiment in which distal and proximal synaptic pathways are stimulated independently, while optically monitoring the trafficking of proteins and mRNAs that are known to be selectively distributed at recently activated synapses. Interactions of the kind seen in Figure 5A,C and Figure 3F would allow one to infer whether DDT, DDD or a mixture of both strategies are implemented biologically.
Finetuned trafficking rates and cargo recycling introduce new tradeoffs
We next wanted to understand (a) how severe the speedaccuracy tradeoff might be, given experimental estimates of neuron size and trafficking kinetics, and (b) whether simple modifications to the sushibelt model could circumvent this tradeoff. We examined the DDD model in an unbranched cable with a realistic neurite length (800 μm) and an optimistic diffusion coefficient of 10 µm^{2} s^{−1}, which we set by inversely scaling the trafficking rate constants with the squared compartment length (see Materials and methods and Figure 6—figure supplement 1). All cargo began in the leftmost compartment and was delivered to a small number of demand ‘hotspots’ (black arrows, Figure 6A). Similar results were found when the DDT model was examined in this setting (data not shown).
When the detachment timescale was sufficiently slow, the cargo was distributed evenly across the demand hotspots, even when the spatial distribution of the hotspots was changed (Figure 6A1; Video 2). Increasing the detachment rate caused faster convergence, but erroneous delivery of cargo. In all cases, hotspots closer to the soma received disproportionate high levels of cargo (Figure 6A2; Video 3). Importantly, the tradeoff between these extreme cases was severe: it took over a day to deliver 95% of cargo with 10% average error, and over a week to achieve 1% average error (blue line, Figure 6B).
We next attempted to circumvent this tradeoff by two strategies. First, motivated by the observation that too much cargo was delivered to proximal sites in Figure 6A2, we increased the anterograde trafficking rate of cargo near the soma so that more cargo would reach distal sites. By carefully finetuning a linearly decreasing profile of trafficking bias (illustrated in Figure 6A, bottom panel), we obtained a model (Figure 6A3; Video 4) that provided accurate and fast delivery (within 10% error in 200 min) for a distribution of six, evenly placed hotspots.
However, this model’s performance was very sensitive to changes in the spatial pattern of demand (Figure 6A3, middle and bottom; Video 5). Increasing the anterograde trafficking rates produced nonmonotonic speedaccuracy tradeoff curves (green, red, and cyan curves Figure 6B), indicating that the detachment rates needed to be finetuned to produce low error. Randomly altering the spatial profile of demand hotspots resulted in variable tradeoff curves for a finetuned trafficking model (red lines, Figure 6C); an untuned model was able to achieve more reliable cargo delivery albeit at the cost of much slower delivery times (blue lines, Figure 6C).
Next, we considered a variant of the sushibelt model that allowed for the reversible detachment/reattachment of cargo from the microtubules (Figure 7A):
Inspection of this scheme reveals that it is similar in form to the DDT model analyzed in Figure 2 and 3: the reversible detachment step simply adds an additional transient state in each compartment. As we noted in the DDT model, cargo distributions can match demand over time with arbitrarily low error (see Equation 4). However, transport delays still exist. While releasing cargo to the wrong location is not an irreversible error, it slows delivery by temporarily arresting movement – known as a diffusive trap (see e.g. Bressloff and Earnshaw, 2007.
We found that cargo recycling creates a new tradeoff between convergence time and excess cargo left on the microtubules. Models that deliver a high percentage of their cargo (${c}_{i}>{d}_{i}$) converge more slowly since they either release cargo into the diffusive traps (Figure 7A1) or have a slow detachment process (Figure 7A2). Models that deliver a low percentage of their cargo (${d}_{i}>{c}_{i}$) converge quickly since they release little cargo into diffusive traps, allowing cargo to travel along the microtubules and reach all destinations within the neuron (Figure 7A3). Figure 7B shows the convergence of the three examples (A1, A2 and A3) over time. Figure 7C shows that the new tradeoff between cargo utilization and convergence time is similarly severe to the speedaccuracy tradeoff in the sushibelt model without reattachment. Models with reattachment that utilize cargo efficiently (for example, Figure 7A2) converge on similarly slow timescales to models without reattachment that deliver cargo accurately (for example, Figure 6A1). Models with less than 10% excess cargo required more than a day to reach steadystate within a tolerance of 10% mean error. On the other hand, models that converged around 10^{3} minutes (17 hr) required more than 90% of cargo to remain in transit at steadystate (Figure 7C).
Distinct celltype morphologies face order of magnitude differences in speed, precision and efficiency of trafficking
To establish the biological significance of these findings, we examined tradeoffs between speed, precision and excess cargo in reconstructed morphologies of five neuron cell types, spanning size and dendritic complexity (Figure 8A). We simulated trafficking and delivery of cargo to a spatially uniform target distribution in each cell type to reveal morphologydependent differences. In all cases we used optimistic estimates of transport kinetics, corresponding to a diffusion coefficient of 10 µm^{2} s^{−1} (the rate constants were normalized to compartment size as in Figure 6—figure supplement 1).
Figure 8B shows spatial plots of the distribution of cargo on the microtubules (${u}_{i}$, cyantomagenta colormap) and the distribution of delivered cargo (${u}_{i}^{\star}$, blacktoorange colormap) for a model with an irreversible detachment rate of 8 × 10^{−5} s^{−1}. These parameters produce a relatively slow release of cargo: for each morphology, a sizable fraction of the cargo remains on the microtubules at ~3 hr, and it takes ~1–2 days to release all of the cargo. While the speed of delivery is roughly equivalent, the accuracy varied across the neural morphologies. The hippocampal granule cell converged to very low error (~11.7% mean error), while the larger L5 pyramidal cell converged to ~27.7% error. The smaller, but more elaborately branched, Purkinje cell converged to a similarly high average error of ~29.1%.
As before, faster detachment rates produce faster, but less accurate, delivery; while slower detachment rates produce more accurate, but slower, delivery. These tradeoffs across the entire family of regimes are plotted in Figure 8C (left). Adding a reattachment process largely preserved the effect of morphology on transport tradeoffs (Figure 8C, right). We fixed the detachment rate to be fast, since fast detachment produced the most favorable tradeoff in Figure 7C. Tradeoffs between excess cargo and speed of delivery emerged as the reattachment rate was varied (Figure 8C, right) and were more severe for the Purkinje cell and L5 pyramidal cell, and least severe for the Granule cell. Morphology itself therefore influences the relationship between delivery speed and precision, and/or excess cargo required, suggesting that different cell types might benefit from different trafficking strategies.
Discussion
The molecular motors that drive intracellular transport are remarkably efficient, achieving speeds of approximately 15 µm per minute (Rogers and Gelfand, 1998; Dynes and Steward, 2007; Müller et al., 2008). A naïve calculation based on this figure might suggest that subcellular cargo can be delivered precisely within a few hours in most dendritic trees. However, this ignores the stochastic nature of biochemical processes – motors spontaneously change directions and cargo can be randomly delivered to the wrong site. Such chance events are inevitable in molecular systems, and in the case of active transport they lead to diffusion of bulk cargo in addition to directed movement. If this kind of biochemical stochasticity played out in the sushi restaurant analogy, then the waiting time for a dish wouldn’t simply equate to the time taken for the chef to prepare the dish and for the belt to convey it. Instead, the restaurant would be beleaguered by fickle customers who pick up dishes they do not want, either withholding them for an indefinite period, or setting them on another belt destined for the kitchen.
Mathematical models provide a rigorous framework to test the plausibility and the inherent relationships in conceptual models. Our study formalized the foremost conceptual model of dendritic transport (Doyle and Kiebler, 2011) to account for trafficking in realistic dendritic morphologies. Over a wide range of assumptions the model exhibits inherent and surprisingly punishing tradeoffs between the accuracy of cargo delivery and the time taken to transport it over these morphologies. Using conservative estimates based on experimental data, the canonical sushibelt model predicts delays of many hours or even days to match demand within 10%. Producing excess cargo and permitting reversible detachment from the microtubules can mitigate this tradeoff, but at a substantial metabolic cost, since a large amount of excess cargo is required.
These predictions are unsettling, because nucleustosynapse transport appears to play a role in timecritical processes. Elevated synaptic activity can initiate distal metabolic events including transcription (Kandel, 2001; Deisseroth et al., 2003; Greer and Greenberg, 2008; Ch'ng et al., 2011) and this has been shown to be an important mechanism of neuronal plasticity (Nguyen et al., 1994; Frey and Morris, 1997, 1998; Bading, 2000; Kandel, 2001; Redondo and Morris, 2011). Moreover, neuronal activity has been observed to influence trafficking directly through secondmessengers (Mironov, 2007; Wang and Schwarz, 2009; Soundararajan and Bullock, 2014), consistent with the hypothesis that trafficking rates are locally controlled. Genes that are transcribed in response to elevated activity can regulate synaptic strengths (Flavell and Greenberg, 2008; Bloodgood et al., 2013; Spiegel et al., 2014), and it has been suggested that nucleustosynapse trafficking of Arc directly regulates synaptic plasticity (Okuno et al., 2012). None of these findings imply that all kinds of molecular cargo are transported from the soma to distal dendritic locations, since mRNA can be sequestered and locally translated within dendrites (Kang and Schuman, 1996; Cajigas et al., 2012; Holt and Schuman, 2013). However, the speed, precision and efficiency tradeoffs revealed in the sushi belt model provide a principled way to understand why some processes might require local biosynthesis, while others operate globally.
The different ways that local demand signals can influence trafficking and detachment can impact global performance, sometimes nonintuitively. Many of these effects should be experimentally testable. For example, transport bottlenecks can be induced if demand signals target local trafficking rates along microtubules (the DDT model). Transport to distal compartments will be substantially faster when proximal demand is introduced (see Figure 3). On the other hand, uniform trafficking combined with locally controlled detachment (DDD model, Figure 4D) can avoid bottlenecks, and often leads to faster transport. However, this is not always the case, as was shown in Figure 5D, where uniform trafficking is slower/inaccurate because cargo explores the basal dendritic tree even though there is no demand in that region. Spatial tuning of trafficking speed permitted more efficient cargo delivery in the model (see Figure 6). However, this has yet to be observed experimentally and would require extremely stereotyped morphology and physiological needs for it to be effective.
Intuitively, speed/precision tradeoffs arise because there is a conflict between exploring the dendritic tree and capturing cargo in specific locations. For irreversible cargo detachment, the capture rate needs to be roughly an order of magnitude slower than trafficking, otherwise, compartments proximal to the soma receive disproportionately high levels of cargo. This scaling is unfavorable for achieving high accuracy: if it takes roughly 100 min to distribute cargo throughout the dendrites, it will take roughly 1000 min (16–17 hr) before the cargo dissociates and is delivered to the synapses. If, instead, cargo is able to reattach, then fast reattachment favors exploration at the cost of greater excess (i.e. nonutilized) cargo, while slow reattachment hinders transport, since more cargo is detached and thus immobile. Even when the vast majority of cargo is produced as excess, global delivery times of several hours persist. Furthermore, if a neuron needs to rapidly replace a cargo that is already present in high concentrations, the strategy of generating excess cargo will result in large dilution times.
Overall, our results show that there are multiple ways that neurons can distribute cargo, but each differs in its speed, accuracy and metabolic cost. Therefore, optimizing for any one of these properties comes at the expense of the others. For example, in the model without reattachment (Figure 4), the same distribution of cargo can be achieved by: (a) locationdependent trafficking followed by uniform release, (b) uniform trafficking followed by locationdependent release, or (c) a mixture of these two strategies. Experimental findings appear to span these possibilities. (Kim and Martin, 2015) identified three mRNAs that were uniformly distributed in cultured Aplysia sensory neurons, but were targeted to synapses at the level of protein expression by localized translation (supporting option b). In contrast, the expression of Arc mRNA is closely matched to the pattern of Arc protein in granule cells of the dentate gyrus (possibly supporting option a; Steward et al., 1998; Farris et al., 2014; Steward et al., 2014). Trafficking kinetics do not just differ according to cargo identity – the same type of molecular cargo can exhibit diverse movement statistics in singleparticle tracking experiments (Dynes and Steward, 2007). These differences lead us to speculate that different neuron types and different cargoes have adapted trafficking strategies that match performance tradeoffs to biological needs.
It is possible that active transport in biological neurons will be more efficient and flexible than models predict. Real neurons might use unanticipated mechanisms, such as a molecular addressing system, or nonlinear interactions between nearby cargo particles, to circumvent the tradeoffs we observed. For this reason, it is crucial to explore, quantitatively, the behavior of existing conceptual models by replacing words with equations so that we can see where discrepancies with biology might arise. More generally, conceptual models of subcellular processes deserve more quantitative attention because they can reveal nonobvious constraints, relationships and connections to other biological and physical phenomena (Smith and Simmons, 2001; Bressloff, 2006; Fedotov and Méndez, 2008; Newby and Bressloff, 2010b; Bhalla, 2011; Bressloff and Newby, 2013; Bhalla, 2014). Other modelling studies have focused on the effects of stochasticity and local trapping of cargo on a microscopic scale, particularly in the context of low particle numbers (Bressloff, 2006; Bressloff and Earnshaw, 2007; Fedotov and Méndez, 2008; Newby and Bressloff, 2010b; Bressloff and Newby, 2013). We opted for a coarsegrained class of models in order to examine transport and delivery across an entire neuron. The model we used is necessarily an approximation: we assumed that cargo can be described as a concentration and that the multiple steps involved in cellular transport can lumped together in a mass action model.
By constraining trafficking parameters based on prior experimental measurements, we revealed that a leading conceptual model predicts physiologically important tradeoffs across a variety of assumptions. Experimental falsification would prompt revision of the underlying models as well as our conceptual understanding of intracellular transport. On the other hand, experimental confirmation of these tradeoffs would have fundamental consequences for theories of synaptic plasticity and other physiological processes that are thought to require efficient nucleustosynapse trafficking.
Materials and methods
All simulation code is available online: https://github.com/ahwillia/WilliamsetalSynapticTransport
Model of singleparticle transport
Let ${x}_{n}$ denote the position of a particle along a 1dimensional cable at timestep $n$. Let ${v}_{n}$ denote the velocity of the particle at timestep $n$; for simplicity, we assume the velocity can take on three discrete values, ${v}_{n}=\{1,0,1\}$, corresponding to a retrograde movement, pause, or anterograde movement. As a result, ${x}_{n}$ is constrained to take on integer values. In the memoryless transport model (top plots in Figure 1B, D and F), we assume that ${v}_{n}$ is drawn with fixed probabilities on each step. The update rule for position is:
We chose ${p}_{}=0.2$, ${p}_{0}=0.35$ and ${p}_{+}=0.45$ for the illustration shown in Figure 1. For the model with historydependence (bottom plots in Figure 1B, D and F), the movement probabilities at each step depend on the previous movement. For example, if the motor was moving in an anterograde direction on the previous timestep, then it is more likely to continue to moving in that direction in the next time step. In this case the update rule is written in terms of conditional probabilities:
In the limiting (nonstochastic) case of historydependence, the particle always steps in the same direction as the previous time step.
We introduce a parameter $k\in [0,1]$ to linearly interpolate between this extreme case and the memoryless model.
The bottom plots of Figure 1B and D were simulated with $k=0.5$.
To estimate the concentration and spatial distribution of cargo in real units, we used a 1 µm/s particle velocity and a 1 s time step to match experimental estimates of kinesin (Klumpp and Lipowsky, 2005, and references). We assumed a dendritic diameter of 7.2705 µm.
Relationship of singleparticle transport to the massaction model
The massaction model (Equation 1, in the Results) simulates the bulk movement of cargo across discrete compartments. Cargo transfer is modelled as an elementary chemical reaction obeying massaction kinetics (Keener and Sneyd, 1998). For an unbranched cable, the change in cargo in compartment $i$ is given by:
For now, we assume that the anterograde and retrograde trafficking rate constants ($a$ and $b$, respectively) are spatially uniform.
The massaction model can be related to a driftdiffusion partial differential equation (Figure 1E) by discretizing into spatial compartments of size $\mathrm{\Delta}$ and expanding around some position, $x$:
We keep terms to second order in $\mathrm{\Delta}$, as these are of order $dt$ in the limit $\mathrm{\Delta}\to 0$ (Gardiner, 2009). This leads to a driftdiffusion equation:
Measurements of the mean and meansquared positions of particles in tracking experiments, or estimates of the average drift rate and dispersion rate of a pulse of labeled particles can thus provide estimates of parameters $a$ and $b$.
How does this equation relate to the model of singleparticle transport (Figure 1A–B)? For a memoryless biased random walk, the expected position of a particle after $n$ time steps is $E\left[{x}_{n}\right]=n\left({p}_{+}{p}_{}\right)$ and the variance in position after $n$ steps is $n\left({p}_{+}+{p}_{}{\left({p}_{+}{p}_{}\right)}^{2}\right)$. For large numbers of noninteracting particles the mean and variance calculations for a single particle can be directly related to the ensemble statistics outlined above. We find:
This analysis changes slightly when the singleparticle trajectories contain long, unidirectional runs. The expected position for any particle is the same $E\left[{x}_{n}\right]=n\left({p}_{+}{p}_{}\right)$; the variance, in contrast, increases as run lengths increase. However, the massaction model can often provide a good fit in this regime with appropriately refit parameters (see Figure 1F). Introducing run lengths produces a larger effective diffusion coefficient and thus provides faster transport. As long as the singleparticles have stochastic and identically distributed behavior, the ensemble will be welldescribed by a normal distribution by the central limit theorem. This only breaks down in the limit of very long unidirectional runs, as the system is no longer stochastic (Figure 1—figure supplement 1).
Stochastic interpretation of the massaction model
An important assumption of the massaction model is that there are large numbers of transported particles, so that the behavior of the total system is deterministic. Intuitively, when each compartment contains many particles, then small fluctuations in particle number don’t appreciably change concentration. Many types of dendritic cargo are present in high numbers (Cajigas et al., 2012).
When few cargo particles are present, fluctuations in particle number are more functionally significant. Although we did not model this regime directly, the massaction model also provides insight into this stochastic regime. Instead of interpreting ${u}_{i}$ as the amount of cargo in compartment $i$, this variable (when appropriately normalized) can be interpreted as the probability of a particle occupying compartment $i$. Thus, for a small number of transported cargoes, the massaction model describes the average, or expected, distribution of the ensemble.
In this interpretation, the massaction model models a spatial probability distribution. Let ${p}_{i}$ denote the probability of a particle occupying compartment $i$. If a single particle starts in the somatic compartment at $t=0$, and we query this particle’s position after a long period of transport, then the probability ratio between of finding this particle in any parentchild pair of compartments converges to:
which is analogous to Equation (3) in the Results.
In the stochastic model, the number of molecules in each compartment converges to a binomial distribution at steadystate; the coefficient of variation in each compartment is given by:
This suggests two ways of decreasing noise. First, increasing the total number of transported molecules, $n$, proportionally decreases the noise by a factor of $1/\sqrt{n}$. Second, increasing ${p}_{i}$ decreases the noise in compartment $i$. However, this second option necessarily comes at the cost of decreasing occupation probability and thus increasing noise in other compartments.
Estimating parameters of the massaction model using experimental data
The parameters of the massaction model we study can be experimentally fit by estimating the drift and diffusion coefficients of particles over the length of a neurite. A common approach is to plot the mean displacement and mean squared displacement of particles as a function of time. The slopes of the bestfit lines in these cases respectively estimate the drift and diffusion coefficients. Diffusion might not accurately model particle movements over short time scales because unidirectional cargo runs result in superdiffusive motion, evidenced by superlinear increases in mean squareddisplacement with time (Caspi et al., 2000). However, over longer timescales, cargoes that stochastically change direction can be modelled as a diffusive process (Soundararajan and Bullock, 2014).
The massaction model might also be fitted by tracking the positions of a population of particles with photoactivatable GFP (Roy et al., 2012). In this case, the distribution of fluorescence at each point in time could be fit by a Gaussian distribution; the drift and diffusion coefficients are respectively proportional to the rate at which the estimated mean and variance evolves over time.
These experimental measurements can vary substantially across neuron types, experimental conditions, and cargo identities. Therefore, in order to understand fundamental features and constraints of the sushi belt model across systems, it is more useful to explore relationships within the model across ranges of parameters. Unless otherwise stated, the trafficking kinetics were constrained so that ${a}_{i}+{b}_{i}=1$ for each pair of connected compartments. This is equivalent to having a constant diffusion coefficient of one across all compartments. Given a target expression pattern along the microtubules, this is the only free parameter of the trafficking simulations; increasing the diffusion coefficient will always shorten convergence times, but not qualitatively change our results. In Figures 6–8 we fixed the diffusion coefficient to an optimistic value of 10 µm^{2} s^{−1} based on experimental measurements (Caspi et al., 2000; Soundararajan and Bullock, 2014) and the observation that long run lengths can increase the effective diffusion coefficient (Figure 1—figure supplement 1).
Steadystate analysis
The steadystate ratio of trafficked cargo in neighboring compartments equals the ratio of the trafficking rate constants (Equation 2). Consider an unbranched neurite with nonuniform anterograde and retrograde rate constants (Equation 1). It is easy to verify the steadystate relationship in the first two compartments, by setting ${\dot{u}}_{1}=0$ and solving:
Successively applying the same logic down the cable confirms the condition in Equation 2 holds globally. The more general condition for branched morphologies can be proven by a similar procedure (starting at the tips and moving in).
It is helpful to reexpress the massaction trafficking model as a matrix differential equation, $\dot{\mathbf{u}}=A\mathbf{u}$, where $\mathbf{u}={\left[{u}_{1},{u}_{2},...{u}_{N}\right]}^{T}$ is the state vector, and $A$ is the statetransition matrix. For a general branched morphology, $A$ will be nearly tridiagonal, with offdiagonal elements corresponding to branch points; matrices in this form are called Hines matrices (Hines, 1984). For the simpler case of an unbranched cable, $A$ is tridiagonal:
For both branched and unbranched morphologies, each column of $A$ sums to zero, which reflects conservation of mass within the system. Assuming nonzero trafficking rates, the rank of $A$ is exactly $N1$ (this can be seen by taking the sum of the first $N1$ rows, which results in $1$ times the final row). Thus, the nullspace of $A$ is onedimensional. Equation (3) describes this manifold of solutions: the level of cargo can be scaled by a common multiplier across all compartments without disrupting the relation in (2).
The steadystate distribution, $\stackrel{\mathbf{~}}{\mathbf{u}}$, is a vector that spans the nullspace of $A$. It is simple to show that all other eigenvalues $A$ are negative using the Gershgorin circle theorem; thus, the fixed point described by Equation 2 is stable. The convergence rate is determined by the nonzero eigenvalue with the smallest magnitude of $A$. There are no other fixed points or limit cycles in this system due to the linearity of the model.
Biologically plausible model of a local demand signal
There are many biochemical mechanisms that could signal demand. Here we briefly explore cytosolic calcium, ${\left[Ca\right]}_{i}$, as a candidate mechanism since it is modulated by local synaptic activity and ${\left[Ca\right]}_{i}$ transients simultaneously arrest anterograde and retrograde microtubular transport for certain cargoes (Wang and Schwarz, 2009). We represent the effect of the calciumdependent pathway by some function of calcium, $f\left(\left[C{a}_{i}\right]\right)$. This function could, for example, capture the binding affinity of ${\left[Ca\right]}_{i}$ to enzymes that alter the kinetics of motor proteins; the Hill equation would provide a simple functional form. If all outgoing trafficking rates of a compartment are controlled by cytosolic calcium — i.e. for any parentchild pair of compartments we have $a=f\left({\left[Ca\right]}_{p}\right)$ and $b=f\left({\left[Ca\right]}_{c}\right)$ — then condition in Equation 4 is satisfied:
where ${\stackrel{~}{u}}_{i}=1/f([Ca{]}_{i})$. We emphasize that other potential signalling pathways could achieve the same effect, so while there is direct evidence for ${\left[Ca\right]}_{i}$ as an important signal, the model can be interpreted broadly, with ${\left[Ca\right]}_{i}$ serving as a placeholder for any local signal identified experimentally. Further, ${\left[Ca\right]}_{i}$ itself may only serve as a demand signal over short timescales, while other, more permanent, signals such as microtubuleassociated proteins (Soundararajan and Bullock, 2014) are needed to signal demand over longer timescales.
Simulations in realistic morphologies
We used a customwritten Python library to generate movies and figures for all simulations in realistic morphologies (Williams, 2016). We obtained the CA1 pyramidal cell model from the online repository ModelDB (Hines et al., 2004), accession number 144541 (Migliore and Migliore, 2012). We used the default spatial compartments and set the trafficking and dissociation parameters of the massaction transport model without reference to the geometry of the compartments. Model simulations were exact solutions using the matrix exponential function from the SciPy library at logarithmically spaced timepoints (Jones et al., 2001). In Figure 2 we simulated electrical activity of this model with excitatory synaptic input for 5 s using the Python API to NEURON (Hines et al., 2009). We used the average membrane potential over this period to set the target demand level. In Figures 3 and 4, we imposed artificial demand profiles with regions of lowdemand and highdemand (an orderofmagnitude difference) as depicted in the figures. Time units for simulations of the CA1 model were were normalized by setting trafficking rates ${a}_{i}+{b}_{i}=1$ (which corresponds to a unit diffusion coefficient).
In Figure 8, we obtained representative morphologies of five cell types from neuromorpho.org (Ascoli et al., 2007). Specifically, we downloaded a Purkinje cell (PurkinjesliceageP436), a parvalbuminpositive interneuron (AWa80213), a Martinotti cell (C100501A3), a layer5 pyramidal cell (32L5pyr28), and a granule cell from the dentate gyrus (041015vehicle1). In these simulations, we scaled the trafficking parameters inversely proportional to the squared distance between the midpoints of neighboring compartments, which is mathematically appropriate to keep the (approximated) diffusion coefficient constant across the neural morphology. We confirmed that compartment size had minimal effects on the convergence rate and steadystate cargo distribution when the trafficking rates were scaled in this way in the reduced cable model (Figure 6—figure supplement 1).
For simulations with reattachment in Figure 8, we set the detachment rate (${c}_{i}$) equal to the trafficking rates (${a}_{i},{b}_{i}$) for a one micron compartment. We did this based on the observation that a fast detachment rate provided the most favorable tradeoff curve in Figure 7C.
Incorporating detachment and reattachment into the massaction model
For compartment $i$ in a cable, the differential equations with detachment become:
When ${a}_{i},{b}_{i}\gg {c}_{i}$, then the distribution of cargo on the microtubules ($u}_{i$) approaches a quasisteadystate that follows Equation 3. In Figure 4, we present DDT and DDD models as two strategies that distribute cargo to match a demand signal $\stackrel{~}{u}}_{i}^{\star$. As mentioned in the main text, a spectrum of models that interpolate between these extremes are possible. To interpolate between these strategies, let $F$ be a scalar between 0 and 1, and let $\stackrel{~}{u}}^{\star$ be normalized to sum to one. We choose ${a}_{i}$ and ${b}_{i}$ to achieve:
along the microtubular network and choose ${c}_{i}$ to satisfy
Here, $N$ is the number of compartments in the model. Setting $F=1$ results in the DDT model (demand is satisfied purely by demandmodulated trafficking, and nonspecific detachment, Figure 4C). Setting $F=0$ results in the DDD model (demand is satisfied purely by demandmodulated detachment, and uniform/nonspecific trafficking, Figure 4D). An interpolated strategy is shown in Figure 4E ($F=0.3$).
The massaction model with reattachment (Equation 6) produces the following system of differential equations for a linear cable, with ${d}_{i}$ denoting the rate constant of reattachment in compartment $i$
We examined the DDD model with $N=100$ compartments and diffusion coefficient of 10 μm^{2}s^{−1}. The maximal detachment rate constant and the reattachment rates were tunable parameters, while the reattachment rates were spatially uniform. Results were similar when reattachment was modulated according to demand (data not shown, see supplemental simulations at https://github.com/ahwillia/WilliamsetalSynapticTransport).
Globally tuning transport rates to circumvent the speedspecificity tradeoff
In Figure 6, we explored whether finetuning the trafficking rates could provide both fast and precise cargo distribution. We investigated the DDD model with fast detachment rates in an unbranched cable with equally spaced synapses and $N=100$ compartments. Large detachment rates produced a proximal bias in cargo delivery which we empirically found could be corrected by setting the anterograde and retrograde trafficking rates to be:
where $i=\{1,2,\mathrm{\dots}N1\}$ indexes the trafficking rates from the soma ($i=1$) to the other end of the cable ($i=N1$), and $D=10\phantom{\rule{thinmathspace}{0ex}}\mu {m}^{2}/s$ is the diffusion coefficient. Faster detachment rates require larger values for the parameter $\beta $; note that $\beta <D/2$ is a constraint to prevent ${b}_{i}$ from becoming negative. This heuristic qualitatively improved, but did not precisely correct for, fast detachment rates in the DDT model (data not shown).
Intuitively, the profile of the proximal delivery bias is roughly exponential (Figure 6B), and therefore the anterograde rates need to be tuned more aggressively near the soma (where the bias is most pronounced), and more gently tuned as the distance to the soma increases. Importantly, tuning the trafficking rates in this manner does not alter the diffusion coefficient along the length of the cable (since ${a}_{i}+{b}_{i}$ is constant by construction). These manipulations produce a nonzero drift coefficient to the model, which corrects for the proximal bias in cargo delivery.
References
 1

2
NeuroMorpho.Org: a central resource for neuronal morphologiesJournal of Neuroscience 27:9247–9251.https://doi.org/10.1523/JNEUROSCI.205507.2007

3
Transcriptiondependent neuronal plasticityEuropean Journal of Biochemistry 267:5280–5283.https://doi.org/10.1046/j.14321327.2000.01565.x
 4

5
Molecular computation in neurons: a modeling perspectiveCurrent Opinion in Neurobiology 25:31–37.https://doi.org/10.1016/j.conb.2013.11.006
 6
 7
 8

9
Diffusiontrapping model of receptor trafficking in dendritesPhysical Review E 75:041915.https://doi.org/10.1103/PhysRevE.75.041915

10
Directed intermittent search for hidden targetsNew Journal of Physics 11:023033.https://doi.org/10.1088/13672630/11/2/023033

11
Stochastic models of intracellular transportReviews of Modern Physics 85:135–196.https://doi.org/10.1103/RevModPhys.85.135

12
In the right place at the right time: visualizing and understanding mRNA localizationNature Reviews Molecular Cell Biology 16:95–109.https://doi.org/10.1038/nrm3918
 13

14
Singlemolecule insights into mRNA dynamics in neuronsTrends in Cell Biology 25:468–475.https://doi.org/10.1016/j.tcb.2015.05.005
 15

16
Enhanced diffusion in active intracellular transportPhysical Review Letters 85:5655–5658.https://doi.org/10.1103/PhysRevLett.85.5655

17
Synapsetonucleus signalingCurrent Opinion in Neurobiology 21:345–352.https://doi.org/10.1016/j.conb.2011.01.011

18
Signaling from synapse to nucleus: the logic behind the mechanismsCurrent Opinion in Neurobiology 13:354–365.https://doi.org/10.1016/S09594388(03)00076X

19
Mechanisms of dendritic mRNA transport and its role in synaptic taggingThe EMBO Journal 30:3540–3552.https://doi.org/10.1038/emboj.2011.278

20
Dynamics of bidirectional transport of arc mRNA in neuronal dendritesThe Journal of Comparative Neurology 500:433–447.https://doi.org/10.1002/cne.21189
 21

22
NonMarkovian model for transport and reactions of particles in spiny dendritesPhysical Review Letters 101:218102.https://doi.org/10.1103/PhysRevLett.101.218102
 23
 24
 25

26
Molecular motors: directing traffic during RNA localizationCritical Reviews in Biochemistry and Molecular Biology 46:229–239.https://doi.org/10.3109/10409238.2011.572861

27
Stochastic methods: a handbook for the natural and social sciences (Springer series in synergetics) (4th edition)Springer.

28
Pathways and mechanisms of endocytic recyclingNature Reviews Molecular Cell Biology 10:597–608.https://doi.org/10.1038/nrm2755
 29
 30

31
Bidirectional cargo transport: moving beyond tug of warNature Reviews Molecular Cell Biology 15:615–628.https://doi.org/10.1038/nrm3853

32
Efficient computation of branched nerve equationsInternational Journal of BioMedical Computing 15:69–76.https://doi.org/10.1016/00207101(84)900084

33
ModelDB: A Database to Support Computational NeuroscienceJournal of Computational Neuroscience 17:7–11.https://doi.org/10.1023/B:JCNS.0000023869.22017.2e
 34
 35
 36
 37

38
SciPy: Open Source Scientific Tools for PythonSciPy: Open Source Scientific Tools for Python.
 39
 40
 41
 42
 43
 44
 45

46
Dendritic hyperpolarizationactivated currents modify the integrative properties of hippocampal CA1 pyramidal neuronsJournal of Neuroscience 18:7613–7624.
 47

48
ADP regulates movements of mitochondria in neuronsBiophysical Journal 92:2944–2952.https://doi.org/10.1529/biophysj.106.092981
 49
 50
 51

52
Quasisteady state reduction of molecular MotorBased models of directed intermittent searchBulletin of Mathematical Biology 72:1840–1866.https://doi.org/10.1007/s1153801095138
 53
 54
 55
 56

57
Making memories last: the synaptic tagging and capture hypothesisNature Reviews Neuroscience 12:17–30.https://doi.org/10.1038/nrn2963
 58
 59

60
Models of motorassisted transport of intracellular particlesBiophysical Journal 80:45–68.https://doi.org/10.1016/S00063495(01)759942
 61
 62
 63
 64

65
Localization and local translation of arc/Arg3.1 mRNA at synapses: some observations and paradoxesFrontiers in Molecular Neuroscience 7:101.https://doi.org/10.3389/fnmol.2014.00101
 66
 67

68
150 years of the mass action lawPLoS Computational Biology 11:e1004012–1004017.https://doi.org/10.1371/journal.pcbi.1004012
 69

70
Bidirectional transport along microtubulesCurrent Biology 14:R525–R537.https://doi.org/10.1016/j.cub.2004.06.045

71
PyNeuron ToolboxPyNeuron Toolbox.
 72
 73
Decision letter

Upinder S BhallaReviewing Editor; National Centre for Biological Sciences, India
In the interests of transparency, eLife includes the editorial decision letter and accompanying author responses. A lightly edited version of the letter sent to the authors after peer review is shown, indicating the most substantive concerns; minor comments are not usually included.
[Editors’ note: a previous version of this study was rejected after peer review, but the authors submitted for reconsideration. The first decision letter after peer review is shown below.]
Thank you for submitting your work entitled "Dendritic trafficking faces physiologically critical speedprecision tradeoffs" for consideration by eLife. Your article has been favorably evaluated by Naama Barkai (Senior Editor) and three reviewers, one of whom is a member of our Board of Reviewing Editors. Our decision has been reached after consultation between the reviewers. Based on these discussions and the individual reviews below, we regret to inform you that your work will not be considered further for publication in eLife.
Here is a synthesis of the views of the reviewers, highlighting the main points.
1) All three reviewers appreciated the development of a general and simplecoarsegrained model of cargo localization, and the incorporation of severalaspects of experimentally observed trafficking.
2) There is a shared common concern about whether real neurons do experience the bottlenecks and very slow settling of cargo distributions that the model predicts. All three reviewers felt that the real settling time was likely tobe faster than the model predicts.
3) There were also shared concerns about the model assumptions, particularlywith respect to the mechanisms assumed for the system. The reviewers felt thatsome likely, and experimentally supported mechanisms were left out of theanalysis. Some of the mechanisms that the reviewers mentioned included cargo reloading, feedback, and having sufficiently large cargo traffic to feed transient local requirements for cargo. The reviewers felt that these mechanisms might eliminate the slow timecourses predicted by the current study.
4) The reviewers were hoping to see a more complete mapping to experimentswith stronger mechanistic predictions and closer ties to biologicalobservations.
Reviewer #1:
This study puts quantitative flesh on the bones of the sushibelt model for transport in the dendrites and its interaction with local signals resulting in cargo offloading.
At the outset it is important to make the point that the sushibelt model as originally proposed was a wordmodel, and the process of converting it to mathematical form is nontrivial and itself involves many mechanistic assumptions and insights. While I generally appreciate the motivation of the study, I have concerns about some of the assumptions and feel that some aspects of the model behavior and study conclusions may be artifacts of these assumptions. Specifically, I feel that the key conclusions about speedspecificity tradeoffs and the timecourse for attaining desired distributions are overstated. Below I indicate how other model assumptions, or even factoring another term into the same model (point 6), may overcome the stated limitations in achieving target cargo distributions.
1) The first aspect of the model involves the microtubulebased transport.
Here the authors adapt and simplify a biased 1D randomwalk model by
Muller et al. They consider two variants: simple independent probabilities,and historydependent probabilities. This first set of model conversion iswellgrounded in the literature and results in a familiar model form. Thistakes them to an analysis of cargo distribution as a function of transportrates, and they show that a variety of distributions can be achieved throughthis mechanism. If I understand correctly, the main novelty in this figureis the result that even with long runs the massaction model fits thestochastic one reasonably well. It would be nice to have a panel of thisdependency, that is, a graph of fit vs. run length.
2) The next step of the modeling is to consider how the local rates oftransport might be modulated. The authors come up with a somewhat limitingmodel here, assuming only a single signal (Ca for convenience), and achieveforward/backward rate control by taking the signal levels in successive spatialcompartments. Here one could readily imagine that different localsignals might be a more versatile (and spatially more precise) way to achievecontrol of forward and backward rates of transport. Can the authors examinethis?
3) The electrical calculations in Figure 2E are poorly described. I assumethat the authors use the full Migliore 2012 model to obtain an electricalpotential distribution upon synaptic stimulation. It would have been usefulto have seen the electrical potential and its timecourse. Over what timewas the 'average potential' taken? How did the potential map to the rates?
The authors refer to the methods section but there is insufficient informationthere.
4) The authors go through a few more elaborations in the model, before bringing in a 'detachment' scheme that finally takes their model to something more like the full sushibelt model. In Figures 1–3 I am concerned that the analysis talks about density of cargo on the motors rather than free cargo in thedendrites. First, it would be valuable to make this distinction clearer tothe reader. Second, it would be valuable to discuss whether these predictionshave physiological observations to compare with. I do not have a sense forhow much cargo sits on the motors, and how much variability is observedin the distribution of motorattached vs. detached forms.
5) When detachment is incorporated into the model, the authors find that onegets nonspecific cargo delivery, as well as depletion of cargo attached tothe motors. I am concerned that these phenomena are more a reflection ofassumptions than physiology. Specifically, the unloading of cargo is anopenloop, stimulusdriven process in the model. I wonder how many of thesefindings would hold if the unloading rate were driven not just by stimulus,but also by feedback based on amount of desired cargo that was alreadypresent. That is, a term dependent on ${u}_{i}^{*}$. Further, the degradation itselfcould also be driven by feedback. I suspect that the set point might bereached much faster with these elaborations.
6) Looking in more detail at the equations in the subsection “Incorporating detachment and degradation into the massaction model”, I was trying to understand the effect of loading density of the cargo. Specifically,if ${u}_{i}$ is large and the desired ${u}_{i}^{*}$ is small, surely the system shouldgive a very rapid convergence to the target ${u}_{i}^{*}$? In other words, if thereis a huge amount of cargo available and going past, then one can quicklyobtain what one needs in any location to a high degree of accuracy. Itseems to me that the loading term should also play a role in the analysis on
Figures 5 and 6. Thus the 'slow detachment' case could actually be fastin absolute time terms if one were to factor in lots of available cargo.
I do not see this factor in the analysis in the subsection “Conservative experimental estimates of trafficking parameters suggest that the tradeoff between speed and specificity is severe”.
Reviewer #2:
In this manuscript, the authors developed a theoretical model for transport of cargoes on microtubules. Analytical solutions of the model show that such transport can either be fast or precise but not both. (Precision in this case means similarity to target cargo concentration at the destination.) In particular, the authors considered two different transport schemes were considered: (1) specific transport, uniform detachment, and (2) uniform transport, specific detachment.
1) A consequence of the first transport scheme is that bottlenecks will occur with the same probability in the main dendrites and the subsequent branch dendrites, since the rate constants within the whole neuron are modeled by the same function. However, it is reasonable to ask if neurons in reality do exhibit bottlenecks in the main dendrite and the branch/daughter dendrites at the same frequency.
2) Perhaps there could be more detailed studies of transports using a combination of these two transport schemes (Figure 4E). For example, will an intermediate strategy improve speed and precision, i.e., can a scheme involving intermediate transportspecificity and intermediate detachmentspecificity circumvent the problems of bottlenecks and cargo leakage? Perhaps a phase space plot that illustrates the effect of the combinations of schemes on accuracy and transport time may help to conveythe information better.
Reviewer #3:
The manuscript by Williams applies a "sushibelt delivery model" to cargo transport In CA1 pyramidal neurons. The goal is to understand the tradeoff between speed versus precision during cargo transport along microtubules by motor proteins. The manuscript opens with a rather general discussion (massaction model) of convectiondiffusion in a channel that is coarsegrained at the level of adjacent boxes. This idea is extended to model pyramidal neurons where the steady state distribution of cargo is calculated w.r.t. a target profile. The authors then model a bottleneck situation by assuming low cargo transition rates (epsilon) into a compartment, and test how the system converges to steady state as a function of epsilon. The model is taken further by introducing detachment from microtubules (possibly followed by diffusionrecapture – details unclear), and looking at efficiency of transport under two possibilities – cargo is selectively transported to target or is uniformly distributed (combined with detachment).
The goal is laudable because it attempts to present a generalized and simple mathematically solvable coarsegrained description of cargo localization in a complex neuronal geometry. Most of the assumptions of the mathematical model appear valid, their rate constants seem to match the experimental velocities and they seem to have taken into consideration various scenarios during cellular transport. However, we feel that the paper starts off being rather general, and remains moreorless so till the end. For example, Figures 4CD show that the target cargo distributions are always achieved irrespective of the transport/detachment ratio. What is one expected to learn from this, and how might it be useful to plan future experiments? If the message is that many strategies can be employed to achieve target distributions, then this is a rather weak message unless this theme is developed further with specific examples and suggestions. Similarly, the observation (subsection “Convergence rate”, last paragraph) that transport will achieve steady state faster if bottlenecks are removed – why is this surprising? This part is followed up by a few poorly explained lines where the results (Figure 3F) seem interesting, but are obscured by unnecessary usage of complicated Latin words. On the same lines, in the places where it is mentioned, the connection to experiments is rather weak. Whether these assumptions hold true in a biological setting has not been tested for any neuronal cargoes. Live imaging to show that at least a few cargoes follow this model would have helped.
The authors talk about detachment and degradation of cargoes. But how does reloading of cargoes occur in instances when continuous supply is required?
The authors suggest using their model that accurate transport is slower, and faster rates of transport requires much greater complexity and is very sensitive to perturbations. This is hard to visualize in case of most neuronal functions where efficient robust and rapid signaling does occur during processes such as long term memory formation, signaling at the synapse etc. Again, there appears to be a disconnection between real biology and the model.
Taken together, we feel that this work would not have sufficient impact to warrant publication in eLife. This is in contrast to models of microtubule transport (e.g. Lipowsky group PNAS paper) which have made more specific mechanistic predictions that advanced the field and inspired new experiments. We suggest the authors also improve the writing of this manuscript in consultation with some experimental colleagues. Perhaps addition of some experiments as preliminary tests of models would also help.
[Editors’ note: what now follows is the decision letter after the authors submitted for further consideration.]
Thank you for submitting your article "Dendritic trafficking faces physiologically critical speedprecision tradeoffs" for consideration by eLife. Your article has been favorably evaluated by Naama Barkai (Senior Editor) and three reviewers, one of whom is a member of our Board of Reviewing Editors. The following individual involved in review of your submission has agreed to reveal their identity: Ambarish Kunwar (Reviewer #3).
The reviewers have discussed the reviews with one another and the Reviewing Editor has drafted this decision to help you prepare a revised submission.
The reviewers all felt that the paper was much improved from the earlier version. They made some small suggestions for minor improvements and clarifications, not requiring further review.
Summary:
In this study Williams et al. explore simulations of the 'sushibelt' family of models, including local signals to unload cargo. They use deterministic methods after comparing with some preliminary stochastic calculations. They embed the model into a detailed neuronal morphology. The authors exercise the model on some notsoobvious predictions such as speed and accuracy tradeoffs.
Essential revisions:
1) I'm surprised that the feedback and cargo recycling processes do not bring rapid settling to the system without large cargo excess. I think this is one of the key findings of the paper. I would suggest that the authors move some of the panels from Figure 5—figure supplement 2 into the main body of the paper, so as to better present this result.
2) The authors have based their entire simulation on a real life neuron (CA1 Pyramidal Cell) with a fixed number of compartments (742). It would be good if the authors throw some light on (i) how sensitive their simulations are to the number of compartments in the neuron, and (ii) how the compartment size is related to the cargo size.
https://doi.org/10.7554/eLife.20556.019Author response
[Editors’ note: the author responses to the first round of peer review follow.]
Here is a synthesis of the views of the reviewers, highlighting the main points.
1) All three reviewers appreciated the development of a general and simplecoarsegrained model of cargo localization, and the incorporation of severalaspects of experimentally observed trafficking.
We are pleased that the reviewers recognized the need for a formalization of conceptual models in this field and an assessment of the relationships they imply. We further believe that studies like ours are important because researchers in related fields do not always appreciate the relevance of lowlevel biophysical processes to higher level physiological function, and because word models that often inform experimental designs and interpretations of data imply relationships that are not evident until the model is formalized. Not all of the reviewers appreciated these points.
2) There is a shared common concern about whether real neurons do experience the bottlenecks and very slow settling of cargo distributions that the model predicts. All three reviewers felt that the real settling time was likely tobe faster than the model predicts.
Our goal is to challenge intuitions about how existing trafficking models behave with actual numbers. Settling times have never been measured, as such we believe our findings, coupled with the useful suggestions from the reviewers, will advance the field by motivating attempts to actually measure settling times.
A) Either settling times are fast (and accurate) in biology, which calls into question the current working model of the field.
B) Or, nucleustosynapse trafficking is not as fast (or accurate) as expected by the reviewers and other experts in the field.
Either of these possibilities is significant. Determining which is correct will require a series of experimental studies, possibly in multiple different preparations, cell types, and species. Our model can be used to generate specific hypotheses for these experiments, which would be difficult to formulate otherwise. To aid this, we now discuss connection between the model and experiments more explicitly.
On the other hand, our model may have had flawed assumptions. The crucial concern raised in the reviews (that our model did not allow recirculation of cargo) has now been addressed. The other very broad suggestion was some kind of feedback. The model formulation assumes an accurate (perfect, in fact) mechanism that signals demand for cargo according to the actual quantity needed, but does so with fixed rates. We have clarified this in the revision.
Recycling of cargo provides a means to allow more rapid delivery of cargo to potential sites of demand. This does indeed speed up settling times, but with the additional metabolic cost of requiring excess cargo (see next point), and the best case still predicts several hours for typical morphologies.
Beyond such mechanisms, there is no evidence, to our knowledge, of any more sophisticated feedback mechanism, e.g. some kind of mechanism that routes cargo to specific dendrites. We discuss how feedback could optimize delivery times further, but we don't believe that it is profitable to iteratively tweak the model by postulating additional mechanisms that have yet to be observed. Instead, we believe it is far more useful to motivate experimental falsification of the predictions of a parsimonious model.
3) There were also shared concerns about the model assumptions, particularlywith respect to the mechanisms assumed for the system. The reviewers felt thatsome likely, and experimentally supported mechanisms were left out of theanalysis. Some of the mechanisms that the reviewers mentioned included cargo reloading, feedback, and having sufficiently large cargo traffic to feed transient local requirements for cargo. The reviewers felt that these mechanisms might eliminate the slow timecourses predicted by the current study.
As discussed in the previous point, we have now allowed for cargo reloading and discussed feedback. Cargo reloading introduces a new tradeoff between trafficking accuracy, speed and excess cargo. Two new figures are devoted to this (Figure 5—figure supplement 2 and new Figure 6C). The excess cargo required to speed up settling times is surprisingly severe and still results in settling times of several hours.
A further point that we have now discussed is that excess cargo will lead to long dilution times if an existing species of cargo needs to be replaced. We did not model dilution explicitly as this would require a lot of guesswork and in any case could only serve to exacerbate the long settling times predicted by the existing model.
4) The reviewers were hoping to see a more complete mapping to experimentswith stronger mechanistic predictions and closer ties to biologicalobservations.
Our original manuscript contained several experimental predictions, but we recognize that these connections could have been made more clearly. We have substantially tightened the manuscript and made an effort to ensure that our major points and predictions are highlighted more prominently and clearly. We highlight three salient connections to experimental data below:
A) In Figure 1, along with details documented in the Methods and online code repository (https://github.com/ahwillia/WilliamsetalSynapticTransport), provide a mapping of singleparticle movements to the parameters of the model we study.
B) Figure 3, in particular panel F, presents a series of direct experimental predictions for the DDT model related to trafficking bottlenecks. Candidate systems for testing these predictions include activitydependent trafficking of Arc mRNA.
C) Figure 6 is completely new and provides a comprehensive summary of various model assumptions and predictions in realistic neuron morphologies.
Reviewer #3 commented that our results are “general” and that this is somehow a negative attribute. We are slightly puzzled by this assessment. Just as there are many kinds of experimental papers with different goals, there are many kinds of theory and modelling papers. Some, such as the PNAS paper the reviewer cites, focus on detailed molecular interactions. Others, such as ours, attempt to synthesize a complex, messy set of observations and biological questions into something intelligible that reveals high level relationships. It is a mistake to think that a highlevel focus doesn't apply because the object of study is biochemical in nature. We believe that both kinds of studies are important and that it is inappropriate to judge one type of study according to the attributes of a completely different study. Moreover, our study does, in fact, make numerous predictions that are testable in spite of their generality, as we discuss in detail below and in the revised manuscript.
Reviewer #1:
This study puts quantitative flesh on the bones of the sushibelt model for transport in the dendrites and its interaction with local signals resulting in cargo offloading.
At the outset it is important to make the point that the sushibelt model as originally proposed was a wordmodel, and the process of converting it to mathematical form is nontrivial and itself involves many mechanistic assumptions and insights. While I generally appreciate the motivation of the study, I have concerns about some of the assumptions and feel that some aspects of the model behavior and study conclusions may be artifacts of these assumptions. Specifically, I feel that the key conclusions about speedspecificity tradeoffs and the timecourse for attaining desired distributions are overstated. Below I indicate how other model assumptions, or even factoring another term into the same model (point 6), may overcome the stated limitations in achieving target cargo distributions.
1) The first aspect of the model involves the microtubulebased transport.
Here the authors adapt and simplify a biased 1D randomwalk model by
Muller et al. They consider two variants: simple independent probabilities,and historydependent probabilities. This first set of model conversion iswellgrounded in the literature and results in a familiar model form. Thistakes them to an analysis of cargo distribution as a function of transportrates, and they show that a variety of distributions can be achieved throughthis mechanism. If I understand correctly, the main novelty in this figureis the result that even with long runs the massaction model fits thestochastic one reasonably well.
The reviewer understood the purpose of this figure correctly, but we have nonetheless clarified it in the revision.
It would be nice to have a panel of this dependency, that is, a graph of fit vs. run length.
We have added this as a supplemental figure (Figure 1—figure supplement 1).
2) The next step of the modeling is to consider how the local rates oftransport might be modulated. The authors come up with a somewhat limitingmodel here, assuming only a single signal (Ca for convenience), and achieveforward/backward rate control by taking the signal levels in successive spatialcompartments. Here one could readily imagine that different localsignals might be a more versatile (and spatially more precise) way to achievecontrol of forward and backward rates of transport. Can the authors examinethis?
There is likely to be a multitude of signals but the point is that these would somehow converge to have the net effectof controlling trafficking rates or detachment rates. One could assume an arbitrarily complex signaling pathway that achieves the same net result but for clarity we just illustrated how this might work for one signal, simply to show that it was feasible.
Moreover, we showed that even a single signal, such as calcium, can in principle allow arbitrarily precise trafficking. A more complex model involving multiple signals could not be more spatially precise, although it might work just as well.
3) The electrical calculations in Figure 2E are poorly described. I assumethat the authors use the full Migliore 2012 model to obtain an electricalpotential distribution upon synaptic stimulation. It would have been usefulto have seen the electrical potential and its timecourse. Over what timewas the 'average potential' taken? How did the potential map to the rates?
The authors refer to the methods section but there is insufficient informationthere.
We have clarified this by completely rewriting this section (subsection “Biophysical formulation of the sushi belt model”, last paragraph and subsection “Simulations in realistic morphologies”, first paragraph) and simplified Figure 2 by removing extraneous panels. The reviewer nevertheless correctly understood the important technical points on his/her first read. We believe showing the electrical trace is not necessary as this would be a distraction from the main point, which is simply that local demand can vary and can be controlled by signals such as synaptic input. Details of the exact mapping between electrical signals and demand for cargo would not affect our results as these only depend on having a localized demand signal of some kind.
4) The authors go through a few more elaborations in the model, before bringing in a 'detachment' scheme that finally takes their model to something more likethe full sushibelt model. In Figures 1–3 I am concerned that the analysistalks about density of cargo on the motors rather than free cargo in thedendrites. First, it would be valuable to make this distinction clearer tothe reader.
The reviewer arrived at the correct conclusions and interpretations. We have, however, completely rewritten the paper to better motivate variants of the model according to thequestions these address. In particular, we present the full sushibelt model up front in equation 2 and emphasize that Figures 1–3 refer to transport “on the motors.”
Second, it would be valuable to discuss whether these predictionshave physiological observations to compare with. I do not have a sense forhow much cargo sits on the motors, and how much variability is observedin the distribution of motorattached vs. detached forms.
Despite an extensive literature search, we are also unaware of experimental studies that have given precise estimates of these parameters. We therefore examined a broad class of models over wide parameter ranges that should account for the unknown ratio of bound cargo (several fold). We have also added an entirely new analysis of a model with cargo recycling and excess cargo on the motors(subsection “Finetuned trafficking rates and cargo recycling introduce new tradeoffs”; subsection “Distinct celltype morphologies face order of magnitude differences in speed, precision and efficiency of trafficking”, last paragraph; Figure 5—figure supplement 2 and Figure 6). Indeed, we hope our study will draw attention to these parameters and motivate experimentalists to make the appropriate measurements.
5) When detachment is incorporated into the model, the authors find that onegets nonspecific cargo delivery, as well as depletion of cargo attached tothe motors. I am concerned that these phenomena are more a reflection ofassumptions than physiology. Specifically, the unloading of cargo is anopenloop, stimulusdriven process in the model. I wonder how many of thesefindings would hold if the unloading rate were driven not just by stimulus,but also by feedback based on amount of desired cargo that was alreadypresent. That is, a term dependent on ${u}_{i}^{*}$. Further, the degradation itselfcould also be driven by feedback. I suspect that the set point might bereached much faster with these elaborations.
We assumed that some mechanism sets the trafficking rates to get a specific distribution of cargo. By assuming that the steady state is equal to the desired distribution we are implicitly assuming feedback. We have now made this clear by completely rewriting the manuscript (subsection “Biophysical formulation of the sushi belt model”, first paragraph). We have also now considered the effects of cargo reloading, which does indeed reduce the settling time.
All of our analysis assumes a simple massaction model with fixed rates. This is the simplest nontrivial assumption. Arbitrarily allowing for any possible nonlinear feedback system would of course permit faster delivery. In the best case, if some mechanism could signal rapidly between neurites, the cell could shut off transport to dendrites that don't need cargo. We might further suppose that detachment rates approach zero when cargo is at target, and some very large (but finite) value otherwise. The lower bound on delivery time with these two assumptions is then the path length divided by the maximum velocity of the motors. This is the naive bestcase transit time that we now discuss in the opening paragraph of the Discussion.
However, this best case is not particularly illuminating because, while physically possible, it requires us to invoke mechanisms that have never been shown to exist, and are not specified in the current, published versions of the sushibelt model. Enumerating other kinds of mechanisms that can give intermediate performance is unlikely to be useful in the absence of empirical evidence.
On the other hand, the formulation of the sushi belt model we presented makes the simplest assumptions consistent with empirical data and the resulting predictions can be tested in a variety of ways. The consequences of the predictions and possible ways to test them are now discussed at length. As we emphasized in our opening, the value of our study is not confined to whether the predictions hold true, because a violation will tell us how to refine the current working model in the field.
6) Looking in more detail at the equations in the subsection “Incorporating detachment and degradation into the massaction model”, I was trying to understand the effect of loading density of the cargo. Specifically,if ${u}_{i}$ is large and the desired ${u}_{i}^{*}$ is small, surely the system shouldgive a very rapid convergence to the target ${u}_{i}^{*}$? In other words, if thereis a huge amount of cargo available and going past, then one can quicklyobtain what one needs in any location to a high degree of accuracy. Itseems to me that the loading term should also play a role in the analysis on
Figures 5 and 6. Thus the 'slow detachment' case could actually be fastin absolute time terms if one were to factor in lots of available cargo.
I do not see this factor in the analysis in the subsection “Conservative experimental estimates of trafficking parameters suggest that the tradeoff between speed and specificity is severe”.
We agree. In the case of reversible detachment, adding large amounts of cargo can support fast and accurate transport in the model. We have added a new results and analyses addressing this point (subsection “Finetuned trafficking rates and cargo recycling introduce new tradeoffs”, subsection “Distinct celltype morphologies face order of magnitude differences in speed, precision and efficiency of trafficking”, last paragraph; Figure 5—figure supplement 2 and Figure 6). However, to get very fast transport requires large cargo excesses which may be energetically inefficient to produce or may lead to “traffic jams” on the microtubules (Welte, 2004). We do not rule this out as a possibility, but present these results so that readers can judge for themselves.
We measure error in the relative concentration of cargo, that is, normalizing for the absolute amount. We do this because we did not want to focus on whether a neuron produces enough cargo overall, but instead on whether it distributes it appropriately. Moreover, relative abundances of signaling components are often what matters to tune spatial physiological properties (e.g. relative strengths of synaptic inputs).
Reviewer #2:
In this manuscript, the authors developed a theoretical model for transport of cargoes on microtubules. Analytical solutions of the model show that such transport can either be fast or precise but not both. (Precision in this case means similarity to target cargo concentration at the destination.) In particular, the authors considered two different transport schemes were considered: (1) specific transport, uniform detachment, and (2) uniform transport, specific detachment.
1) A consequence of the first transport scheme is that bottlenecks will occur with the same probability in the main dendrites and the subsequent branch dendrites, since the rate constants within the whole neuron are modeled by the same function.
The reviewer is correct that the same equation governs how trafficking rates are modulated by demand. However, the demand signal need not be uniform across the main dendrite and smaller branches. We think there may be a misunderstanding here, so we were careful to make the complete rewrite of this section clearer (subsection “Biophysical formulation of the sushi belt model”).
However, it is reasonable to ask if neurons in reality do exhibit bottlenecks in the main dendrite and the branch/daughter dendrites at the same frequency.
Our results simply state that if trafficking rates are controlled by local demand signals, then bottlenecks can occur. There are two things to draw from this: (1) ongoing physiological processes may actually cause bottlenecks and subsequent nonlocal effects in dendritic physiology; (2) experimental induction of bottlenecks provides a means to test whether trafficking is, in fact, controlled by local demand. We have completely rewritten the section on bottlenecks to clarify these points and their relevance (subsection “Transport bottlenecks occur when trafficking rates are nonuniform”). We agree it would be exciting to test this experimentally and we have offered suggestions as to how this can be done (see last paragraph of the aforementioned subsection).
2) Perhaps there could be more detailed studies of transports using a combination of these two transport schemes (Figure 4E). For example, will an intermediate strategy improve speed and precision, i.e., can a scheme involving intermediate transportspecificity and intermediate detachmentspecificity circumvent the problems of bottlenecks and cargo leakage? Perhaps a phase space plot that illustrates the effect of the combinations of schemes on accuracy and transport time may help to convey the information better.
We did consider this possibility and agree that it is important. In the new manuscript we draw the reader’s attention to Figure 5B and D. In one scenario, the transportspecific model (which we now call DDT) outperforms the detachmentspecific model (now DDD), while the opposite model performs better in the other scenario. Thus, changing the pattern of cargo demand will change which model performs best (including different intermediate models). In addition, Figure 5—figure supplement 1 also shows that strategies that are optimal for one pattern of demand fail to be optimal for others.
Reviewer #3:
The manuscript by Williams applies a "sushibelt delivery model" to cargo transport In CA1 pyramidal neurons. The goal is to understand the tradeoff between speed versus precision during cargo transport along microtubules by motor proteins. The manuscript opens with a rather general discussion (massaction model) of convectiondiffusion in a channel that is coarsegrained at the level of adjacent boxes. This idea is extended to model pyramidal neurons where the steady state distribution of cargo is calculated w.r.t. a target profile. The authors then model a bottleneck situation by assuming low cargo transition rates (epsilon) into a compartment, and test how the system converges to steady state as a function of epsilon. The model is taken further by introducing detachment from microtubules (possibly followed by diffusionrecapture – details unclear), and looking at efficiency of transport under two possibilities – cargo is selectively transported to target or is uniformly distributed (combined with detachment).
The goal is laudable because it attempts to present a generalized and simple mathematically solvable coarsegrained description of cargo localization in a complex neuronal geometry. Most of the assumptions of the mathematical model appear valid, their rate constants seem to match the experimental velocities and they seem to have taken into consideration various scenarios during cellular transport.
We were encouraged by these general statements, particularly the acknowledgement that our assumptions appear valid and the recognition that the goal was to shed light on a complex biological system using a simple and interpretable model.
Given this, we were surprised by some of reviewer 3’s concerns which we address point bypoint below. We believe that many of these concerns were the result of the language and presentation of our original manuscript, which in retrospect was not very clearly presented. We have worked hard to improve the presentation and make the scope, purpose, predictions and caveats clearer.
However, we feel that the paper starts off being rather general, and remains moreorless so till the end.
We do not agree that generality is a shortcoming in a scientific study. Moreover, we make numerous specific, testable predictions about global settling times of cargo in realistic neural morphologies, among other things. We have tried to make the more specific predictions and experimental questions clearer in the revised manuscript.
For example, Figures 4CD show that the target cargo distributions are always achieved irrespective of the transport/detachment ratio. What is one expected to learn from this, and how might it be useful to plan future experiments? If the message is that many strategies can be employed to achieve target distributions, then this is a rather weak message unless this theme is developed further with specific examples and suggestions.
We think the reviewer may have misinterpreted the message of this figure, so we have completely rewritten the paper and introduced the family of trafficking strategies in a much more logical way. We show that there are many possible mechanisms, but crucially we show that they are not equivalent in testable ways. For example, we revealed transport bottlenecks as an experimentally measurable behavior that is particular to the DDT model (and are absent in DDD, as they are now called).
Most importantly, we show that this entire class of models is subject to a speedaccuracy tradeoff and that the optimal strategy depends on the spatial profile of demand (Figure 5). Thus, different neuron types with different morphologies might exhibit trafficking strategies that are tuned to better cope with, say proximal vs distal fluctuations in demand. We think that researchers who are interested in higher level regulation of trafficking will find this very relevant and not at all obvious.
Similarly, the observation (subsection “Convergence rate”, last paragraph) that transport will achieve steady state faster if bottlenecks are removed – why is this surprising?
What is surprising is that elevated global demand results in faster convergence than distal demand alone. One might naively expect the opposite. Furthermore, this is experimentally testable, as we discussed, yet no attempt has been made to observe this effect that we are aware of, which further supports the case that this is not an obvious prediction of the underlying trafficking model. Again, we were concerned that the writing in the original manuscript might have obscured this point. This and related sections have been completely rewritten.
This part is followed up by a few poorly explained lines where the results (Figure 3F) seem interesting, but are obscured by unnecessary usage of complicated Latin words.
We have removed jargon, revised the figure and improved clarity on this result (subsection “Transport bottlenecks occur when trafficking rates are nonuniform”, Figure 3). This is an important point of our paper as it proposes a possible experimental approach for testing and characterizing the DDT model. Specifically, we propose that selectively stimulating different regions of the dendritic tree (as done in, e.g., Han & Heineman’s study) and tracking a cargo with activitydependent trafficking (e.g., Arc mRNA) could expose the existence of bottlenecks (see last paragraph of the aforementioned subsection).
On the same lines, in the places where it is mentioned, the connection to experiments is rather weak. Whether these assumptions hold true in a biological setting has not been tested for any neuronal cargoes. Live imaging to show that at least a few cargoes follow this model would have helped.
We share the reviewer’s desire for detailed experimental studies that address these measurements. However, we believe this is an unreasonable request for this manuscript and in any case we do not have the facilities to perform these experiments. Secondly, all papers need a scope and ours is already substantial. As the reviewer initially appreciated, extracting meaningful, testable predictions from a conceptual model of a very complex process is extremely challenging. It is also crucial for progress because it is hard to identify compelling experiments – such as measurements of total settling times, or the bottleneck effects – without a rational analysis.
We would also like to point out that in this review process our paper has already triggered a passionate discussion of the need to measure quantities that nobody has previously measured, such as global settling times, based on the predictions of a current theoretical model. This is the kind of outcome we would anticipate from a successful theory paper.
The authors talk about detachment and degradation of cargoes. But how does reloading of cargoes occur in instances when continuous supply is required?
This is a very useful suggestion. As discussed elsewhere, we have examined this in the revised manuscript (new section on reloading, subsection “Finetuned trafficking rates and cargo recycling introduce new tradeoffs”, Figure 6 and Figure 5—figure supplement 2).
The authors suggest using their model that accurate transport is slower, and faster rates of transport requires much greater complexity and is very sensitive to perturbations. This is hard to visualize in case of most neuronal functions where efficient robust and rapid signaling does occur during processes such as long term memory formation, signaling at the synapse etc. Again, there appears to be a disconnection between real biology and the model.
We disagree with this conclusion. The reviewer is simply asserting that synaptic plasticity is fast and precise, and can be both fast and precise while relying on nucleus tosynapse communication. We do not know of any published data directly supporting this assertion, and the reviewer does not cite any. Even in extensively studied systems like the Aplysia gill withdrawal circuit or the Shaffer collateralCA1 synapse, the precision of synaptic plasticity is unknown. Similarly, there is little direct evidence that global trafficking is crucial for other forms of synaptic signaling on short timescales. In fact, there is mounting evidence, discussed and referenced at length in our manuscript, that local biosynthesis is important for many processes the reviewer alludes to. To understand why this is the case, we need to analyze the limitations of global transport, which is one of the goals of our paper.
Taken together, we feel that this work would not have sufficient impact to warrant publication in eLife. This is in contrast to models of microtubule transport (e.g. Lipowsky group PNAS paper) which have made more specific mechanistic predictions that advanced the field and inspired new experiments.
We appreciate the work reviewer appears to refer to (Klumpp and Lipowsky, 2005; Muller et al., 2008) and we cite these in our manuscript.
However, detailed models cannot address the highlevel conceptual and physiological questions that are of interest to us and many other neurobiologists. Models that attempt to answer highlevel questions using intricate mechanistic details are far more sensitive to assumptions and measurement error than appropriately coarsegrained models (see, for example, O'Leary, Sutton & Marder 2015, Curr Opinion in Neurobiology).
We feel it is a mistake to think that highlevel, conceptual studies automatically lack 'impact' when the object of study is biochemical in nature. Conceptual questions are what motivate mechanistic work, after all.
Secondly, as the other reviewers appreciated, it is important to put mathematical flesh on the bones of the wordmodels that guide many experimentalists and other researchers who try to intuit the capabilities of global trafficking when formulating theories of how, say, synaptic plasticity might work.
Thirdly, general models can provide distinct and complementary insights to those gathered from detailed models and can make detailed studies more relevant to researchers outside a specific domain. For example, our study connects active transport to important questions in neuroanatomy and plasticity.
Finally, our study does make concrete predictions which, if addressed experimentally, would certainly advance the field. We return once again to the question of global settling times as an example: the reviewers expressed a strong belief that these would be faster than the model predicts, yet these have never been measured experimentally and there is no specific, existing motivation to do so. We believe our study can provide this motivation.
[Editors' note: the author responses to the rereview follow.]
Essential revisions:
1) I'm surprised that the feedback and cargo recycling processes do not bringrapid settling to the system without large cargo excess. I think this is oneof the key findings of the paper. I would suggest that the authors move someof the panels from Figure 5—figure supplement 2 into the main body of the paper, so as to better present this result.
We have moved the figures to the main results, as suggested. They now constitute Figures 6 and 7 of the revised manuscript.
2) The authors have based their entire simulation on a real life neuron (CA1 Pyramidal Cell) with a fixed number of compartments (742). It would be good if the authors throw some light on (i) how sensitive their simulations are to the number of compartments in the neuron, and (ii) how the compartment size is related to the cargo size.
We have performed an additional analysis of the sensitivity of settling time to compartment size (new Figure 6—figure supplement 1) which is seen to asymptote as compartment number increases. Notably, coarsening (reducing number of compartments) decreases estimated settling time. Thus, spatial discretization makes our estimates of the severity of the speedprecision tradeoff conservative.
https://doi.org/10.7554/eLife.20556.020Article and author information
Author details
Funding
Department of Energy Computational Science Graduate Fellowship
 Alex H Williams
Howard Hughes Medical Institute
 Terrence J Sejnowski
National Institutes of Health (P41GM103712)
 Terrence J Sejnowski
National Institutes of Health (1P01NS079419)
 Timothy O'Leary
The funders had no role in study design, data collection and interpretation, or the decision to submit the work for publication.
Acknowledgements
We thank Aoife McMahon, Lasani Wijetunge, Eve Marder, Subhaneil Lahiri, Friedemann Zenke, and Benjamin Regner for useful feedback on the manuscript, and thank Jeff Gelles and Simon Bullock for useful discussion. This research was supported by the Department of Energy Computational Science Graduate Fellowship, NIH Grant 1P01NS079419, NIH Grant P41GM103712, the Howard Hughes Medical Institute.
Reviewing Editor
 Upinder S Bhalla, National Centre for Biological Sciences, India
Publication history
 Received: August 11, 2016
 Accepted: November 29, 2016
 Version of Record published: December 30, 2016 (version 1)
 Version of Record updated: January 6, 2017 (version 2)
Copyright
© 2016, Williams 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.
Metrics

 1,751
 Page views

 323
 Downloads

 6
 Citations
Article citation count generated by polling the highest count across the following sources: Crossref, Scopus, PubMed Central.
Download links
Downloads (link to download the article as PDF)
Download citations (links to download the citations from this article in formats compatible with various reference manager tools)
Open citations (links to open the citations from this article in various online reference manager services)
Further reading

 Computational and Systems Biology
 Neuroscience

 Computational and Systems Biology
 Neuroscience