Coordinated spinal locomotor network dynamics emerge from cell-type-specific connectivity patterns

It has long been known that the spinal cord is capable of generating locomotor movements even in the absence of detailed instructive input from the brain (Brown and Sherrington, 1911). The core dynamical features of spinal locomotor circuitry are left–right alternation and rostro-caudal propagation of cyclical activity over a range of frequencies. These core features are most obvious during swimming in aquatic vertebrates (Grillner and Wallén, 1985), but they are also apparent during limbed locomotion (Bonnot et al., 2002; Yakovenko et al., 2002; Cazalets, 2005; Ivanenko et al., 2006; Cuellar et al., 2009; Saltiel et al., 2016).

Despite decades of work, the mechanisms by which rhythm is generated for coordinated locomotion are not yet fully understood. One line of research has focused on the intrinsic properties of individual excitatory spinal interneurons, whose mutual interactions serve to orchestrate rhythmic oscillations among downstream interneurons and motor neurons (Grillner and Kozlov, 2021; Rancic and Gosgnach, 2021; Grillner and El Manira, 2015). This perspective is supported by the observation of rhythmic bursting of excitatory interneurons in response to tonic input (Song et al., 2020; Agha et al., 2024), but it is potentially challenged by the observation that targeted disruptions of sources of phasic inhibition interfere with rhythmogenesis (Buchanan and McPherson, 1995; Moult et al., 2013; Picton et al., 2022; Agha et al., 2024). Another line of work has suggested that rhythmic oscillations may be generated as a network-level phenomenon via random recurrent connectivity, without requiring specialized single-cell properties (Lindén et al., 2022). While this model is supported by the observation of ‘rotational’ population dynamics rather than purely biphasic dynamics (Lindén et al., 2022), the possible roles of particular cell types and their characteristic connectivity patterns are not addressed. In addition, the observation of recruitment of speed-selective interneuron populations (McLean and Fetcho, 2011; Grillner and El Manira, 2015) and the functional roles of cell-type-specific intersegmental connectivity within spinal cord (Sengupta and Bagnall, 2023) have not yet been fully accounted for in either the cellular- or the network-level perspective.

Here, we construct a hierarchy of increasingly detailed models in which spinal central pattern generation occurs in a distributed manner, and the essential ingredients for producing coordinated locomotion are cell-type- and speed-specific connectivity motifs rather than specialized cellular properties or random connectivity. In all of these models, we represent individual neurons in a highly simplified way, so the dynamical properties of the network are entirely driven by emergent population dynamics due to connectivity patterns among units. First, we show that letting connectivity be determined by intersegmental phase relationships in a single-population, inhibition-dominated network is sufficient to account for left–right alternation and rostro-caudal propagation. Next, we show that dividing our units into fast- and slow-preferring populations is sufficient to account for variable-frequency control with constant phase lag via speed-dependent recruitment. Finally, in our most detailed model, we show that the strength and modularity of recurrent excitation facilitates faster locomotion, but that there is a trade-off between maximum speed and controllability at intermediate speeds.

Together, these results point to an updated model of the spinal locomotor network as a distributed pattern generator, in which rhythm generation and other locomotion features emerge from collective population dynamics via patterned connectivity and speed-dependent recruitment.

To investigate the degree to which population-specific connectivity patterns might account for the phenomenology of the spinal network described above, we developed a family of models at different levels of detail. In order to focus our investigation on connectivity patterns and the emergent population dynamics that result from them, we modeled the dynamics of individual neurons in a highly simplified way:

where $r_i(t)$ is interpreted as the firing rate of unit $i$ at time $t$, ${\tau }_i$ is the membrane time constant, $D_i$ is the tonic drive, $W_{ij}$ are recurrent weights from other units in the circuit, and $[.{]}_{+}$ denotes rectification. The axonal time delays ${\mathrm{\Delta }}_{ij}$ are proportional to the number of segments between units $i$ and $j$.

The three subsections that follow describe a hierarchy of increasingly detailed models, all of which are described by this basic dynamical equation. This approach enables us to focus mainly on basic computational properties within the relatively abstract, high-level model, whereas the lower-level models focus on implementation of distinct cell-type populations for obtaining additional features.

Phase relationships determine connectivity in a distributed pattern generator

Recent work mapping cell-type-specific connectivity patterns in zebrafish (Menelaou et al., 2014; Sengupta et al., 2025; Sengupta and Bagnall, 2023), mouse (Alaynick et al., 2011; Goulding, 2009), and other organisms has revealed a substantial presence of both excitatory and inhibitory connectivity traversing multiple segments. This is somewhat difficult to interpret in light of classical models of the spinal locomotor circuit, in which dynamical single-cell properties such as bursting or synaptic fatigue generate oscillations within each segment, and these oscillators are then coupled together with intersegmental interactions (see Cohen et al., 1992; Sigvardt and Miller, 1998; Grillner et al., 2007; Ausborn et al., 2021 for reviews). We hypothesized that spatially patterned connectivity alone may be sufficient to drive coordinated locomotion even in the absence of dynamical single-cell properties. To test this, we constructed a high-level model in which dynamical single-cell properties are absent, and the basic features of left–right alternation and segment-to-segment propagation are driven by connectivity patterns. We took inspiration from ‘moving bump’ models of brain circuits, for example in models of head-direction circuits, which have shown that inhibition-dominated, asymmetric connectivity can give rise to activation sequences (Zhang, 1996; Samsonovich and McNaughton, 1997; Murray and Escola, 2017). Such models are also motivated by recent work in zebrafish demonstrating the importance of strong phasic inhibition for generating rhythmic locomotor activity (Agha et al., 2024). This high-level model featured a single, homogeneous population of units described by Equation 1, where all units have identical membrane time constant $\tau$ and common excitatory input drive $D$.

The connections $W_{ij}$ between units in this model were all set to be negative or zero, so that units in the model receive excitation only from the tonic drive, not from each other. Previous related work on threshold-linear recurrent networks with binary inhibitory weights and tonic input has shown that, despite their relative simplicity, such networks are capable of supporting rich dynamics (Curto and Morrison, 2023). Due to the (piecewise) linearity of the activation function, the magnitude of the nonzero weights has no effect on the dynamics, so we set $W_{ij}=-1$.

Given these simplifying choices, the nontrivial question in designing this model is which pairs of units should inhibit one another. The key principle that this model led us to is that the connectivity between pairs of units should encode the desired phase relationships between those units. A version of this idea is already present in classical models, where left and right populations of neurons within each segment laterally inhibit one another, ensuring that they fire exactly out of phase. Here, we apply this same principle to model the longitudinal connectivity of the spinal locomotor network.

According to this principle, a mid-body unit should inhibit contralateral units in the same segment and in nearby segments, as well as inhibiting ipsilateral units more distally (shown schematically for a $N=9$-segment model in Figure 1A). To make this precise, we can summarize the desired phase relationships for a pair of units $i$ and $j$ as follows:

(2) ${\displaystyle {\displaystyle {\displaystyle {\varphi }_{ij}=\frac{(s_i-s_j)\text{mod}N}{N}+\frac{(a_i-a_j)\text{mod}2}{2},}}}$

where $N$ is the number of segments. In this equation, $s_i$ denotes the segment number of unit $i$, and $a_i=0$ if unit $i$ is in the left hemi-segment or 1 if it is in the right hemi-segment.

Figure 1

Download asset Open asset

In order to realize these desired phase relationships, we set the connection strength from neuron $j$ to neuron $i$ as $W_{ij}=f({\varphi }_{ij})$, where ${\varphi }_{ij}\in [0,1)$ is calculated from Equation 2, and

(3) ${\displaystyle {\displaystyle {\displaystyle f(\varphi )=\{\begin{array}{l}-1,{\varphi }_l<\varphi <{\varphi }_u,\\ 0,\text{else}.\end{array}}}}$

This equation describes local contralateral inhibition together with more-distal ipsilateral inhibition. Because ${\varphi }_{ij}=0.5$ for two units that are out of phase, it is necessary to choose ${\varphi }_l<0.5$ and ${\varphi }_u>0.5$. Further, in order for propagation to proceed head-to-tail rather than tail-to-head, it is necessary for the connectivity to be asymmetric, with a relatively larger window of disinhibition downstream and a relatively smaller window of disinhibition upstream. Hence, we chose ${\varphi }_l=0.3$ and ${\varphi }_u=0.8$ to define the window of inhibition. Finally, we can note that the most-distal ipsilateral projections described by Equation 3 (of length $\gtrsim N/2$) are not actually necessary, since the units targeted by such inhibition already receive contralateral inhibition from nearby active units on the opposite side. Indeed, such long-range projections are not typically found in the zebrafish (Higashijima et al., 2004; Callahan et al., 2019; Sengupta et al., 2021). Hence, in this model and in the others presented below, all connections between segments more than 13 segments apart were set to 0 (Figure 1B).

Driving this network with tonic input, and setting the axonal delay ${\mathrm{\Delta }}_{ij}=0$ for simplicity, we observed the key features of coordinated locomotion: phasic bumps of activity in each segment, with smooth propagation of the bump from head to tail and strict alternation between left and right within each segment (Figure 1C). While this high-level model is a highly simplified and abstract representation of the spinal locomotor network, it illustrates several key ideas that may be relevant for characterizing locomotor circuitry in biological organisms and will continue to be present in the more elaborate models developed in the following subsections. First, rhythm can be generated by the network as a whole, rather than by dynamical single-neuron properties or single-segment oscillations. Second, the network’s recurrent connectivity is dominated by inhibition, illustrating that inhibition is capable of and perhaps necessary for sculpting the dynamics. Finally, the pattern of inhibitory connectivity is determined by the desired phase relationships between pairs of units. In the following sections, we will build upon this model by introducing multiple interneuron populations and allowing for more-heterogeneous connectivity between units.

Fast and slow speed modules implement frequency control via recruitment

While the single-population model described above was able to implement coordinated locomotion, it did so at a single, fixed frequency. This frequency was set by the only timescale in the model: the membrane time constant. Inspired by observations of speed-module structure in spinal circuitry in zebrafish (McLean et al., 2008) and mouse (Rancic et al., 2020), whereby interneurons are recruited at either fast or slow locomotion speeds but not both, we made a minimal change to our initial model that introduces an additional timescale, replacing each unit by two units: a ‘fast’ unit with a 1-ms membrane time constant, and a ‘slow’ unit with a 10-ms time constant, as observed experimentally in larval zebrafish (Menelaou et al., 2022). Trivially, in the absence of connectivity between the fast and slow modules, this model amounts to two copies of our earlier model and is capable of operating at two speeds: a fast speed if tonic input is provided only to the fast-preferring units, and a slow speed if tonic input is provided only to the slow-preferring units (Pujala and Koyama, 2019). In each case, the period of left–right alternation and head-to-tail propagation would be roughly proportional to the corresponding membrane time constant (though perhaps not perfectly proportional if axonal delays are included).

In order for the two parts of the network to oscillate at a single global frequency and to interpolate continuously between slow and fast speeds, connectivity between the two modules is required. To begin, we adopted the simplest choice and made the connectivity weights the same for between-module projections and within-module projections, with the same spatial connectivity patterns for between- and within-module projections (Figure 2A, B). We found that the circuit exhibited coordinated locomotion characterized by head-to-tail propagation and left–right alternation at a single global frequency that was intermediate between the characteristic frequencies of the fast and slow modules. Further, the locomotion frequency could be controlled in a graded manner by driving the two modules with different inputs. Locomotion was fast when the fast units received most of the tonic drive, and it was slow when the slow units received most of the tonic drive, with smooth interpolation of intermediate frequencies as the ratio of tonic drives was varied (Figure 2C).

Figure 2

Download asset Open asset

Analyzing the model in greater detail revealed that varying the relative levels of tonic drive to the two populations strongly modulated the frequency (Figure 3A, B). This is made possible by selective recruitment of fast and slow speed modules, such that the fast population is active at faster locomotion frequencies, while the slow population is active at slower locomotion frequencies, with a smooth crossover between these two regimes (Figure 3C). Such recruitment has been observed experimentally (McLean et al., 2008) but has not been accounted for by previous models. In Figure 3—figure supplement 1, we further show that the disengagement of the slow module as frequency increases can occur due to inhibition from the increasingly active fast module, without requiring changes in the tonic drive to the slow population, which is also observed experimentally (McLean et al., 2008).

Figure 3 with 2 supplements see all

Download asset Open asset

To establish that the model exhibits coordinated locomotion across all frequencies, we additionally computed the left–right phase difference between the pairs of units on either side of each segment, finding that the phase differed by half of a period across all locomotion frequencies (Figure 3D). We then computed the phase difference between pairs of units in adjacent segments, finding that this phase was approximately equal to $1/N$ across all locomotion frequencies, where $N$ is the number of segments (Figure 3E), so that the length of the spinal cord exhibits approximately one period of oscillation. This constant phase lag relationship is a core feature of locomotion that has been observed during swimming in a variety of animals (Skinner and Mulloney, 1998). While it does not appear by default in models that chain individually oscillating segments together via longitudinal excitation, constant phase lag emerges in our model as a natural consequence of the fact that the desired phase relationships between segments are built into the circuitry via long-range inhibitory projections. Because the phase relationships are determined by the connectivity, which is the same at all speeds, the phase relationships do not depend on locomotion frequency.

These results show that, by including fast and slow subpopulations and coupling these sub-networks together, the model succeeds in producing coordinated locomotion across a broad range of frequencies, with control of frequency via selective recruitment of the two subpopulations. While this model succeeds in producing much of the phenomenology of the spinal locomotor network, it does not yet fully address the diversity of excitatory and inhibitory cell types that are known to play a role in the biological circuitry (Sengupta and Bagnall, 2023). We turn our attention to this question in the following section.

Excitatory and inhibitory cell types

To address the role of cell-type-specific connectivity patterns in greater detail, we next incorporated excitatory and distinct inhibitory populations by replacing each unit in the above model with four units, leading to an eight-population model (since each of the four types has fast and slow subtypes) (Figure 4A). One of these populations (corresponding to V2a interneurons in the zebrafish) consisted of excitatory units with descending, ipsilateral projections. The other three populations consisted of inhibitory units, essentially breaking the inhibitory population from the earlier model into three populations that have identical activity (since they all receive the same inputs) but differ in their projection targets. One population of inhibitory units (corresponding to V1 interneurons in the zebrafish) had ascending, ipsilateral projections; another (corresponding to V2b interneurons in the zebrafish) had descending, ipsilateral projections; and a third (corresponding to dI6 and V0d interneurons in the zebrafish) had contralateral projections. The tonic drive was provided equally to all units.

Figure 4 with 1 supplement see all

Download asset Open asset

The outgoing spatial connectivity of each of these cell types is illustrated in Figure 4B. Each cell-type projects equally to all of the units within each segment that it targets, so that all units within each segment receive the same inputs. As in the two-population model above, we set the spatial connectivity patterns for the inhibitory units according to the desired phase relationships between units, with short-range contralateral inhibition and intermediate-range ipsilateral inhibition. For the excitatory units, we assumed that the projections are descending only in order to facilitate head-to-tail propagation. All units within each speed module had the same membrane time constant and axonal conduction velocity. As in the earlier models, providing distinct tonic drives to the fast- and slow-module units led to coordinated locomotion at a range of frequencies (Figure 4C).

Before analyzing the full model in detail, we decoupled the two speed modules from one another and began by studying the effects of various single-cell and cell-type-specific connectivity properties on the characteristic oscillation frequency of an individual speed module. Unsurprisingly, the locomotion frequency depended strongly on the membrane time constants, with smaller values of these parameters leading to faster frequencies (Figure 5A, B). Because the characteristic frequencies of the two speed modules set the upper and lower limits of locomotion frequency once the modules are coupled together in the full model, it is likely advantageous for an organism to have values of these parameters that differ strongly in fast- and slow-preferring neurons. This agrees with observations from the zebrafish, where the membrane time constants and axonal delays differ for fast- and slow-preferring excitatory and inhibitory interneurons (Menelaou and McLean, 2019; Menelaou et al., 2022). For our subsequent simulations, we fixed these parameters for the fast and slow modules at the experimentally determined values indicated in Figure 5A, B; Menelaou and McLean, 2019; Menelaou et al., 2022.

Figure 5

Download asset Open asset

We next investigated the effect of connectivity properties on locomotion frequency for the decoupled speed module. We found that the frequency was modulated by more than a factor of two as the excitatory projection distances were varied (Figure 5C). This agrees with observations from zebrafish, where the extents of intersegmental projections have been shown to differ for fast- and slow-preferring excitatory interneurons, with fast-preferring V2a interneurons projecting more distally than slow-preferring V2a neurons (Menelaou et al., 2014). For our subsequent simulations, we fixed these parameters for the fast and slow modules at the experimentally determined values illustrated in Figure 5B (Menelaou et al., 2014).

Having shown that the spatial extent of excitatory projections has a strong effect on locomotion frequency, we next asked whether varying connectivity properties would also modulate the range of possible frequencies in the full model with two coupled speed modules. Varying the global strength of excitatory projections had a strong effect on the range of possible frequencies, with stronger excitation facilitating faster locomotion (Figure 6A). In particular, whereas the purely inhibitory model with experimentally determined membrane time constants and axonal conduction velocities realizes a maximum frequency much lower than that observed in larval zebrafish (approximately 20 Hz, black line in Figure 6A), the inclusion of excitatory interneurons facilitates maximum frequencies of over 50 Hz, which is approaching peak swim speeds in larval zebrafish (Agha et al., 2024). Thus, while excitatory interneurons are not necessary for producing coordinated locomotion in our model, they do facilitate faster locomotion, suggesting that this may be a fundamental role for feedforward excitation in the spinal network.

Figure 6

Download asset Open asset

Given that connectivity within and between speed modules has been shown in zebrafish to be modular, with stronger projections within modules than between modules (Song et al., 2020), we asked what would be the effect of varying modularity in the model. We defined modularity as the difference between intra- versus inter-module connection strength divided by the sum of these quantities, such that modularity of 1 corresponds to fully decoupled modules, while modularity of 0 corresponds to identical connection strengths within versus between modules.

Varying the modularity of all four populations together, we found that there was essentially no effect on the maximum or minimum possible frequencies. Further, the model lost the ability to produce locomotion at intermediate frequencies as modularity was increased (Figure 6B). However, when we varied modularity among only the excitatory or only the inhibitory populations, we observed much more significant changes in the maximum frequency (Figure 6C, D). These changes occurred in opposite directions, with excitatory (inhibitory) modularity favoring faster (slower) speeds, suggesting that the lack of an observed change in frequency range when both types of modularity were varied together (Figure 6B) was due to cancellation between these two effects.

Together, these results show that the strength of feedforward excitation and the modularity of excitatory connectivity have a strong effect on the range of possible locomotion frequencies. There is a trade-off, however, in that the model loses the ability to smoothly interpolate between fast and slow frequencies in cases where the excitatory connectivity becomes too strong or too modular (Figure 6A, C). This requirement that excitation not be too strong is in accord with experimental observations from zebrafish (Agha et al., 2024), which have shown that peak excitatory post-synaptic currents are much weaker than peak inhibitory post-synaptic currents in V2a interneurons, consistent with the possibility that excitation may be globally weaker than inhibition in the spinal circuitry. Further, the fact that the model exhibits a frequency range similar to that of the zebrafish for parameters that are close to the critical values where smooth frequency control becomes impossible suggests that the spinal locomotor circuit faces a trade-off between speed and controllability, and that its excitatory connectivity may be configured in a way that optimizes this trade-off.

Having established the roles played by single-cell and connectivity properties of different cell types in the eight-population model, we fixed these parameters and analyzed the behavior of the model over the range of possible tonic drives to fast and slow populations (Figure 4—figure supplement 1). Similar to the two-population model, the eight-population with coupled fast and slow speed modules model exhibited head-to-tail propagation with constant phase lag (Figure 4—figure supplement 1G), left–right alternation (Figure 4—figure supplement 1F), and frequency-dependent recruitment of fast and slow populations (Figure 4—figure supplement 1E). In addition to varying the frequency of oscillations, we also found that the overall amplitude of interneuron activity in the model could be varied by co-varying the drives to the fast and slow populations (Figure 4—figure supplement 1C, D). This provides a potential mechanism to independently control frequency and amplitude of locomotion, although the manner in which the amplitude of interneuron activity relates to the amplitude of locomotion would depend on the assumptions made about how interneuron activity drives the activity of motor neurons, which we have not included in our models.

We next investigated the effects of perturbing the model by partially ablating (i.e. attenuating the outgoing activity of) each interneuron population (Figure 7). At all locomotion speeds, we found that ablating excitatory units decreased locomotion frequency. This is in agreement with experiments in zebrafish, where ablation of excitatory V2a interneurons had the same effect (Eklöf-Ljunggren et al., 2012). Further, we found that ablating inhibitory units with ascending ipsilateral projections decreased locomotor frequency, while ablating inhibitory units with descending projections increased locomotor frequency across all locomotion speeds. This is also in agreement with experiments in zebrafish, where ablation of inhibitory V1 interneurons slowed swimming (Kimura and Higashijima, 2019), while ablation of V2b interneurons led to faster swimming (Callahan et al., 2019; Sengupta et al., 2025). Finally, we found that ablating the contralaterally projecting inhibitory units led to a modest increase in frequency, but that coordinated locomotion was lost when the degree of ablation became too great. The impact on frequency was most obvious at fast speeds, with a more modest impact at slow speeds. This is consistent with recent experiments in zebrafish, which found the impact of attenuating contralateral inhibitory projections from dI6 neurons on coordination was most obvious at fast speeds (Agha et al., 2024). Similar results were found in Xenopus, where silencing contralaterally projecting inhibitory interneurons can eliminate rhythm generation (Moult et al., 2013) or lead to an increase in swim frequency (Dale, 1995). Together, these results show that, where comparisons with experimental data are possible, perturbations to our model lead to effects on locomotion frequency that generally agree with experimental observations. This agreement provides support for the possibility that the basic mechanisms underlying variable-frequency locomotion in our model—namely cell-type-specific connectivity patterns and speed-module recruitment—may also be at play in the spinal locomotor network.

Figure 7

Download asset Open asset

Robustness in a biophysical model

One potential pitfall with the ‘rate models’ considered above is that the oscillatory behavior seen in these models might be overwhelmed by the inherent stochasticity of a biophysical setting. To test the robustness of our connectivity-based mechanisms for rhythmogenesis and frequency control, we built a spiking-neuron version of each model.

The spiking models were generated from the rate models by replacing each unit in the rate model with $n$ leaky-integrate-and-fire (LIF) neurons. A rough estimate suggests that the locomotor circuit is made up of 50–100 neurons of each cell type per segment (Agha et al., 2024; Callahan et al., 2019; Sengupta et al., 2021; Kishore et al., 2020). In accordance with this estimate, most of our simulations use $n=80$, though versions with smaller $n$ also performed well (Figure 8—figure supplement 1). Connectivity between neurons in the spiking model was determined by including a random fraction $p$ of all possible connections between groups of neurons corresponding to connected units in the rate model. The strength of these connections was determined by the strength of the corresponding connection in the rate model (for details, see Methods).

In addition to the stochasticity introduced via the randomized connectivity, we introduced stochasticity in the spiking model by replacing the tonic input drive of the rate model with EI-balanced Poisson input. This input was independently sampled for each neuron, with a uniform rate for each speed population. In the rate model, a given tonic drive represented the firing rate that an isolated unit would maintain without recurrent connections. To similarly control the input to the spiking model, the rates of the EI-balanced Poisson inputs were chosen to produce a target firing rate in an isolated LIF neuron. Since frequency control depended primarily on the ratio of the fast and slow tonic drives, and the target firing needed to be high enough to ensure spikes would occur reliably within the maxima of the oscillations, we chose our target rates to be 100 Hz times the equivalent tonic drive.

The spiking version of the single-population model demonstrated left–right alternation and head-to-tail propagation (Figure 8A). As with the rate model, the frequency of firing rate oscillations in this model did not depend on the input drive but did depend on the membrane time constant (Figure 8—figure supplement 1C). We found that coherent oscillatory behavior survived at $n=80$ even when the connectivity fraction $p$ was reduced down to 0.05 (i.e. only 5% of connections are made), but that oscillations broke down for smaller values of $p$ (Figure 8—figure supplement 1A). At $n=10$, we found that the cutoff for stable oscillations was $p=0.6$ (Figure 8—figure supplement 1B).

Figure 8 with 1 supplement see all

Download asset Open asset

We next simulated the spiking model with fast and slow populations, the counterpart to our two-population rate model (Figure 8B, C). Choosing the membrane time constants to be the same as the rate model (i.e. 1 ms for the fast population and 10 ms for the slow population) led to issues with controllability. In particular, it led to a breakdown of coherent oscillations at intermediate driving. Referring to our results with the single-population model, we hypothesized that this was caused by the large discrepancy between the natural frequencies of the fast and slow populations. Choosing the membrane time constants to be 2 ms for the fast population and 10 ms for the slow population, we recovered frequency control. While the connectivity fraction $p$ did not affect controllability in this model, it did change the maximum frequency of the network (Figure 8—figure supplement 1D).

Extending our spiking model to the full eight populations of our final rate model (i.e. dividing the inhibitory population into 3 separate populations and including an excitatory population), we again observed stable oscillations with left–right alternation, head-to-tail propagation, and controllable frequency with constant phase lag. As with the rate model, we varied the strength and modularity of the excitatory connections (Figure 8D, E). As in the rate model, increasing excitatory connection strength increased the maximum frequency (Figure 8D). Rather than a loss of controllability at intermediate frequencies as in the rate model, however, the range of possible frequencies was instead reduced via an increase in the lowest frequency achieved by the network, suggesting that the fast population dominates the dynamics in the spiking network. Moreover, in contrast to the rate model, there was no change in maximum frequency upon increasing the modularity among excitatory units (Figure 8E), suggesting a possible saturation effect whereby the fast population already dominates the large-frequency dynamics of the network even at low degrees of modularity. Although this spiking network model responded somewhat differently than the rate models to certain parameter changes, the fact that rhythmogenesis and frequency control emerge from this model without fine-tuning supports the conclusion that network-level interactions driven by cell-type-specific connectivity patterns are sufficient for producing locomotor behavior in a more biophysically detailed model.

In this study, we began by postulating that cell-type-specific connectivity alone could be sufficient for producing the main phenomenological features of the spinal locomotor circuit, without requiring dynamical single-cell properties. We found that coordinated locomotion could be achieved in an inhibition-dominated network in which connectivity is determined by desired phase relationships and variable-speed control is implemented by recruitment of frequency-selective populations. Further, while structured excitatory connections were not necessary for producing coordinated locomotion or frequency control, they were useful for increasing peak locomotor frequency, albeit at the cost of losing some control at intermediate frequencies. Together, this family of models shows that network-level interactions are sufficient to generate coordinated, variable-speed locomotion. It further provides new interpretations of intersegmental excitatory and inhibitory connectivity, as well as the basic, recruitment-based mechanism of speed control.

A main conclusion of our models is the importance of intersegmental, inhibition-dominated connections for achieving coordinated locomotion, where patterns of ipsilateral and contralateral inhibition are established by desired phase relationships. Similarly, a very recent network-level model based on the mouse locomotor circuit has proposed that these are key features for obtaining coordinated locomotion in that context as well (Komi et al., 2024). In addition, a recent study found that patterning recurrent neural network on the locomotor circuitry of C. elegans produced a system that required fewer parameters and less training data than general multi-layer perceptron models for controlling a swimming agent (Bhattasali et al., 2022). The congruence of these results—all of which rely on emergent dynamics driven by cell-type-specific connectivity patterns to generate rhythm, without requiring complex intracellular dynamics—suggests that network-level interactions are likely key drivers of locomotor dynamics across species and fit cross-species observations of spinal interneuron diversity. However, while these results suggest that intrinsic cellular mechanisms are not necessary to generate coordinated locomotor rhythms, they are certainly sufficient in certain circumstances, as demonstrated by previous lesion and pharmacological studies (Grillner and Kozlov, 2021).

Another conclusion of our models is that both projection strength and modularity among fast and slow excitatory units increase the maximum possible frequency, but at the cost of losing some control at intermediate frequencies. Recent studies in zebrafish suggest the possibility that the spinal circuit may overcome this limitation by having multiple subtypes of excitatory interneurons with different degrees of modularity. Specifically, V2a neurons with descending-only axons exhibit a greater degree of speed-dependent recruitment (suggesting a higher level of modularity), while those with bifurcating axons exhibit less (Agha et al., 2024). Moreover, V2a neurons with bifurcating axons fire more reliably, compared to more sparsely firing descending-only V2a neurons, and they form stronger connections to motor neurons (Menelaou and McLean, 2019; Agha et al., 2024), consistent with a hierarchical organization distinguishing interneuron rhythmogenesis from motor neuron recruitment. Including motor neurons and these distinct subtypes into a model and testing their effects will be an interesting direction for future work.

Given that the models that we have presented favor simplicity over realism, attempting to capture as much phenomenology as possible with a minimal number of tunable parameters, they are undoubtedly missing features that may be important for describing more-nuanced functional aspects of locomotion in aquatic vertebrates. For instance, we did not include in our model variations in intersegmental projection distances for dI6 and V0d neurons related to speed (Satou et al., 2020), nor did we include excitatory commissural interneuron classes, including V0v (Kawano et al., 2022) and V3 interneurons (Böhm et al., 2022). Moreover, other recent models of the locomotor circuit in zebrafish have highlighted the importance of electrical synapses for rapidly initiating swim bouts and characterizing early stages of development (Roussel et al., 2020; Roussel et al., 2021; Kim and Riecke, 2023). Including these components would be a worthwhile extension of the models presented here. Finally, investigating the ways in which the circuit-level rhythm genesis of our model might work together with and complement mechanisms based on intracellular dynamics, as assumed in classical models, will be an important direction for future work.

A somewhat unique aspect of our approach has been to develop a hierarchy of models to describe the same neural circuit at varying levels of detail. (See also the related approach in Roussel et al., 2021, which developed a series of models describing the same circuit at different stages in development.) Having such a family of models that fit within the same modeling framework enabled us to (1) use the higher-level models to better motivate the choices made in our lower-level models and (2) draw connections between models at different levels to gain additional insight into the functional roles of particular populations. Following this approach, we thus obtained greater insight into the neural mechanisms underlying behavior than would be possible from any one model individually. We expect that this general approach could be useful more broadly for characterizing neural circuits and their relation to behavior.

This is a computational study, so no data have been generated for this manuscript. Modeling code to accompany this work is available at https://github.com/fwandler/ZFswim (copy archived at Wandler, 2025).

1. 1. Agha MA
   2. Kishore S
   3. McLean DL

   (2024) Cell-type-specific origins of locomotor rhythmicity at different speeds in larval zebrafish

   eLife 13:e94349.

   https://doi.org/10.7554/eLife.94349.2

   • PubMed
   • Google Scholar
2. 1. Alaynick WA
   2. Jessell TM
   3. Pfaff SL

   (2011) SnapShot: spinal cord development

   Cell 146:178.

   https://doi.org/10.1016/j.cell.2011.06.038

   • PubMed
   • Google Scholar
3. 1. Ausborn J
   2. Shevtsova NA
   3. Danner SM

   (2021) Computational modeling of spinal locomotor circuitry in the age of molecular genetics

   International Journal of Molecular Sciences 22:6835.

   https://doi.org/10.3390/ijms22136835

   • PubMed
   • Google Scholar
4. Book

   1. Bhattasali N
   2. Zador AM
   3. Engel T

   (2022)

   Neural circuit architectural priors for embodied control

   In: Koyejo S, Mohamed S, Agarwal A, Belgrave D, Cho K, Oh A, editors. Advances in Neural Information Processing Systems. Curran Associates, Inc. pp. 12744–12759.

   • Google Scholar
5. 1. Böhm UL
   2. Kimura Y
   3. Kawashima T
   4. Ahrens MB
   5. Higashijima S-I
   6. Engert F
   7. Cohen AE

   (2022) Voltage imaging identifies spinal circuits that modulate locomotor adaptation in zebrafish

   Neuron 110:1211–1222.

   https://doi.org/10.1016/j.neuron.2022.01.001

   • PubMed
   • Google Scholar
6. 1. Bonnot A
   2. Whelan PJ
   3. Mentis GZ
   4. O’Donovan MJ

   (2002) Spatiotemporal pattern of motoneuron activation in the rostral lumbar and the sacral segments during locomotor-like activity in the neonatal mouse spinal cord

   The Journal of Neuroscience 22:RC203.

   https://doi.org/10.1523/JNEUROSCI.22-03-j0001.2002

   • PubMed
   • Google Scholar
7. 1. Brown TG
   2. Sherrington CS

   (1911) The intrinsic factors in the act of progression in the mammal

   Proceedings of the Royal Society of London. Series B, Containing Papers of a Biological Character 84:308–319.

   https://doi.org/10.1098/rspb.1911.0077

   • Google Scholar
8. 1. Buchanan JT
   2. McPherson DR

   (1995) The neuronal network for locomotion in the lamprey spinal cord: evidence for the involvement of commissural interneurons

   Journal of Physiology, Paris 89:221–233.

   https://doi.org/10.1016/0928-4257(96)83638-2

   • PubMed
   • Google Scholar
9. 1. Callahan RA
   2. Roberts R
   3. Sengupta M
   4. Kimura Y
   5. Higashijima S-I
   6. Bagnall MW

   (2019) Spinal V2b neurons reveal a role for ipsilateral inhibition in speed control

   eLife 8:e47837.

   https://doi.org/10.7554/eLife.47837

   • PubMed
   • Google Scholar
10. 1. Cazalets J-R

    (2005) Metachronal propagation of motoneurone burst activation in isolated spinal cord of newborn rat

    The Journal of Physiology 568:583–597.

    https://doi.org/10.1113/jphysiol.2005.086850

    • PubMed
    • Google Scholar
11. 1. Cohen AH
    2. Ermentrout GB
    3. Kiemel T
    4. Kopell N
    5. Sigvardt KA
    6. Williams TL

    (1992) Modelling of intersegmental coordination in the lamprey central pattern generator for locomotion

    Trends in Neurosciences 15:434–438.

    https://doi.org/10.1016/0166-2236(92)90006-t

    • PubMed
    • Google Scholar
12. 1. Cuellar CA
    2. Tapia JA
    3. Juárez V
    4. Quevedo J
    5. Linares P
    6. Martínez L
    7. Manjarrez E

    (2009) Propagation of sinusoidal electrical waves along the spinal cord during a fictive motor task

    The Journal of Neuroscience 29:798–810.

    https://doi.org/10.1523/JNEUROSCI.3408-08.2009

    • PubMed
    • Google Scholar
13. 1. Curto C
    2. Morrison K

    (2023) Graph rules for recurrent neural network dynamics

    Notices of the American Mathematical Society 70:1.

    https://doi.org/10.1090/noti2661

    • Google Scholar
14. 1. Dale N

    (1995) Experimentally derived model for the locomotor pattern generator in the Xenopus embryo

    The Journal of Physiology 489 ( Pt 2):489–510.

    https://doi.org/10.1113/jphysiol.1995.sp021067

    • PubMed
    • Google Scholar
15. 1. Eklöf-Ljunggren E
    2. Haupt S
    3. Ausborn J
    4. Dehnisch I
    5. Uhlén P
    6. Higashijima S
    7. El Manira A

    (2012) Origin of excitation underlying locomotion in the spinal circuit of zebrafish

    PNAS 109:5511–5516.

    https://doi.org/10.1073/pnas.1115377109

    • PubMed
    • Google Scholar
16. Book

    1. Gerstner W
    2. Kistler WM
    3. Naud R
    4. Paninski L

    (2014) Neuronal Dynamics: From Single Neurons to Networks and Models of Cognition

    Cambridge University Press.

    https://doi.org/10.1017/CBO9781107447615

    • Google Scholar
17. 1. Goulding M

    (2009) Circuits controlling vertebrate locomotion: moving in a new direction

    Nature Reviews. Neuroscience 10:507–518.

    https://doi.org/10.1038/nrn2608

    • PubMed
    • Google Scholar
18. 1. Grillner S
    2. Wallén P

    (1985) Central pattern generators for locomotion, with special reference to vertebrates

    Annual Review of Neuroscience 8:233–261.

    https://doi.org/10.1146/annurev.ne.08.030185.001313

    • PubMed
    • Google Scholar
19. Book

    1. Grillner S
    2. Kozlov A
    3. Dario P
    4. Stefanini C
    5. Menciassi A
    6. Lansner A
    7. Hellgren Kotaleski J

    (2007) Modeling a vertebrate motor system: pattern generation, steering and control of body orientation

    In: Cisek P, Drew T, Kalaska JF, editors. Computational Neuroscience: Theoretical Insights into Brain Function. Elsevier. pp. 221–234.

    https://doi.org/10.1016/S0079-6123(06)65014-0

    • Google Scholar
20. 1. Grillner S
    2. El Manira A

    (2015) The intrinsic operation of the networks that make us locomote

    Current Opinion in Neurobiology 31:244–249.

    https://doi.org/10.1016/j.conb.2015.01.003

    • PubMed
    • Google Scholar
21. 1. Grillner S
    2. Kozlov A

    (2021) The CPGs for limbed locomotion-facts and fiction

    International Journal of Molecular Sciences 22:5882.

    https://doi.org/10.3390/ijms22115882

    • PubMed
    • Google Scholar
22. 1. Higashijima S
    2. Masino MA
    3. Mandel G
    4. Fetcho JR

    (2004) Engrailed-1 expression marks a primitive class of inhibitory spinal interneuron

    The Journal of Neuroscience 24:5827–5839.

    https://doi.org/10.1523/JNEUROSCI.5342-03.2004

    • PubMed
    • Google Scholar
23. 1. Ivanenko YP
    2. Poppele RE
    3. Lacquaniti F

    (2006) Spinal cord maps of spatiotemporal alpha-motoneuron activation in humans walking at different speeds

    Journal of Neurophysiology 95:602–618.

    https://doi.org/10.1152/jn.00767.2005

    • PubMed
    • Google Scholar
24. 1. Kawano K
    2. Kato K
    3. Sugioka T
    4. Kimura Y
    5. Tanimoto M
    6. Higashijima S-I

    (2022) Long descending commissural V0v neurons ensure coordinated swimming movements along the body axis in larval zebrafish

    Scientific Reports 12:4348.

    https://doi.org/10.1038/s41598-022-08283-0

    • PubMed
    • Google Scholar
25. 1. Kim LU
    2. Riecke H

    (2023) Intersegmental coordination of the central pattern generator via interleaved electrical and chemical synapses in zebrafish spinal cord

    Journal of Computational Neuroscience 51:129–147.

    https://doi.org/10.1007/s10827-022-00837-5

    • PubMed
    • Google Scholar
26. 1. Kimura Y
    2. Higashijima S-I

    (2019) Regulation of locomotor speed and selection of active sets of neurons by V1 neurons

    Nature Communications 10:2268.

    https://doi.org/10.1038/s41467-019-09871-x

    • PubMed
    • Google Scholar
27. 1. Kishore S
    2. Cadoff EB
    3. Agha MA
    4. McLean DL

    (2020) Orderly compartmental mapping of premotor inhibition in the developing zebrafish spinal cord

    Science 370:431–436.

    https://doi.org/10.1126/science.abb4608

    • PubMed
    • Google Scholar
28. Preprint

    1. Komi S
    2. Winther A
    3. Houser GA
    4. Topilko T
    5. Sørensen RJF
    6. Larsen SD
    7. Adamsson Bonfils MC
    8. Li G
    9. Berg RW

    (2024) Spatial and Network Principles behind Neural Generation of Locomotion

    bioRxiv.

    https://doi.org/10.1101/2024.10.03.616472

    • Google Scholar
29. 1. Lindén H
    2. Petersen PC
    3. Vestergaard M
    4. Berg RW

    (2022) Movement is governed by rotational neural dynamics in spinal motor networks

    Nature 610:526–531.

    https://doi.org/10.1038/s41586-022-05293-w

    • PubMed
    • Google Scholar
30. 1. McLean DL
    2. Masino MA
    3. Koh IYY
    4. Lindquist WB
    5. Fetcho JR

    (2008) Continuous shifts in the active set of spinal interneurons during changes in locomotor speed

    Nature Neuroscience 11:1419–1429.

    https://doi.org/10.1038/nn.2225

    • PubMed
    • Google Scholar
31. 1. McLean DL
    2. Fetcho JR

    (2011) Movement, technology and discovery in the zebrafish

    Current Opinion in Neurobiology 21:110–115.

    https://doi.org/10.1016/j.conb.2010.09.011

    • PubMed
    • Google Scholar
32. 1. Menelaou E
    2. VanDunk C
    3. McLean DL

    (2014) Differences in the morphology of spinal V2a neurons reflect their recruitment order during swimming in larval zebrafish

    The Journal of Comparative Neurology 522:1232–1248.

    https://doi.org/10.1002/cne.23465

    • PubMed
    • Google Scholar
33. 1. Menelaou E
    2. McLean DL

    (2019) Hierarchical control of locomotion by distinct types of spinal V2a interneurons in zebrafish

    Nature Communications 10:4197.

    https://doi.org/10.1038/s41467-019-12240-3

    • PubMed
    • Google Scholar
34. 1. Menelaou E
    2. Kishore S
    3. McLean DL

    (2022) Mixed synapses reconcile violations of the size principle in zebrafish spinal cord

    eLife 11:e64063.

    https://doi.org/10.7554/eLife.64063

    • PubMed
    • Google Scholar
35. 1. Moult PR
    2. Cottrell GA
    3. Li WC

    (2013) Fast silencing reveals a lost role for reciprocal inhibition in locomotion

    Neuron 77:129–140.

    https://doi.org/10.1016/j.neuron.2012.10.040

    • PubMed
    • Google Scholar
36. 1. Murray JM
    2. Escola GS

    (2017) Learning multiple variable-speed sequences in striatum via cortical tutoring

    eLife 6:e26084.

    https://doi.org/10.7554/eLife.26084

    • PubMed
    • Google Scholar
37. 1. Picton LD
    2. Björnfors ER
    3. Fontanel P
    4. Pallucchi I
    5. Bertuzzi M
    6. El Manira A

    (2022) Developmental switch in the function of inhibitory commissural V0d interneurons in zebrafish

    Current Biology 32:3515–3528.

    https://doi.org/10.1016/j.cub.2022.06.059

    • PubMed
    • Google Scholar
38. 1. Pujala A
    2. Koyama M

    (2019) Chronology-based architecture of descending circuits that underlie the development of locomotor repertoire after birth

    eLife 8:e42135.

    https://doi.org/10.7554/eLife.42135

    • PubMed
    • Google Scholar
39. 1. Rancic V
    2. Ballanyi K
    3. Gosgnach S

    (2020) Mapping the dynamic recruitment of spinal neurons during fictive locomotion

    The Journal of Neuroscience 40:9692–9700.

    https://doi.org/10.1523/JNEUROSCI.1885-20.2020

    • PubMed
    • Google Scholar
40. 1. Rancic V
    2. Gosgnach S

    (2021) Recent insights into the rhythmogenic core of the locomotor CPG

    International Journal of Molecular Sciences 22:1394.

    https://doi.org/10.3390/ijms22031394

    • PubMed
    • Google Scholar
41. 1. Roussel Y
    2. Paradis M
    3. Gaudreau SF
    4. Lindsey BW
    5. Bui TV

    (2020) Spatiotemporal transition in the role of synaptic inhibition to the tail beat rhythm of developing larval zebrafish

    eNeuro 7:ENEURO.0508-18.2020.

    https://doi.org/10.1523/ENEURO.0508-18.2020

    • PubMed
    • Google Scholar
42. 1. Roussel Y
    2. Gaudreau SF
    3. Kacer ER
    4. Sengupta M
    5. Bui TV

    (2021) Modeling spinal locomotor circuits for movements in developing zebrafish

    eLife 10:e67453.

    https://doi.org/10.7554/eLife.67453

    • PubMed
    • Google Scholar
43. 1. Saltiel P
    2. d’Avella A
    3. Wyler-Duda K
    4. Bizzi E

    (2016) Synergy temporal sequences and topography in the spinal cord: evidence for a traveling wave in frog locomotion

    Brain Structure & Function 221:3869–3890.

    https://doi.org/10.1007/s00429-015-1133-5

    • PubMed
    • Google Scholar
44. 1. Samsonovich A
    2. McNaughton BL

    (1997) Path integration and cognitive mapping in a continuous attractor neural network model

    The Journal of Neuroscience 17:5900–5920.

    https://doi.org/10.1523/JNEUROSCI.17-15-05900.1997

    • PubMed
    • Google Scholar
45. 1. Satou C
    2. Sugioka T
    3. Uemura Y
    4. Shimazaki T
    5. Zmarz P
    6. Kimura Y
    7. Higashijima S-I

    (2020) Functional diversity of glycinergic commissural inhibitory neurons in larval zebrafish

    Cell Reports 30:3036–3050.

    https://doi.org/10.1016/j.celrep.2020.02.015

    • PubMed
    • Google Scholar
46. 1. Sengupta M
    2. Daliparthi V
    3. Roussel Y
    4. Bui TV
    5. Bagnall MW

    (2021) Spinal V1 neurons inhibit motor targets locally and sensory targets distally

    Current Biology 31:3820–3833.

    https://doi.org/10.1016/j.cub.2021.06.053

    • PubMed
    • Google Scholar
47. 1. Sengupta M
    2. Bagnall MW

    (2023) Spinal interneurons: diversity and connectivity in motor control

    Annual Review of Neuroscience 46:79–99.

    https://doi.org/10.1146/annurev-neuro-083122-025325

    • PubMed
    • Google Scholar
48. 1. Sengupta M
    2. Bertram A
    3. Zhu SI
    4. Goodhill GJ
    5. Bagnall MW

    (2025) V2b neurons act via multiple targets to produce in phase inhibition during locomotion

    The Journal of Neuroscience 45:e1530242025.

    https://doi.org/10.1523/JNEUROSCI.1530-24.2025

    • PubMed
    • Google Scholar
49. 1. Sigvardt KA
    2. Miller WL

    (1998) Analysis and modeling of the locomotor central pattern generator as a network of coupled oscillators

    Annals of the New York Academy of Sciences 860:250–265.

    https://doi.org/10.1111/j.1749-6632.1998.tb09054.x

    • PubMed
    • Google Scholar
50. 1. Skinner FK
    2. Mulloney B

    (1998) Intersegmental coordination in invertebrates and vertebrates

    Current Opinion in Neurobiology 8:725–732.

    https://doi.org/10.1016/s0959-4388(98)80114-1

    • PubMed
    • Google Scholar
51. 1. Song J
    2. Pallucchi I
    3. Ausborn J
    4. Ampatzis K
    5. Bertuzzi M
    6. Fontanel P
    7. Picton LD
    8. El Manira A

    (2020) Multiple rhythm-generating circuits act in tandem with pacemaker properties to control the start and speed of locomotion

    Neuron 105:1048–1061.

    https://doi.org/10.1016/j.neuron.2019.12.030

    • PubMed
    • Google Scholar
52. 1. Stimberg M
    2. Brette R
    3. Goodman DF

    (2019) Brian 2, an intuitive and efficient neural simulator

    eLife 8:e47314.

    https://doi.org/10.7554/eLife.47314

    • PubMed
    • Google Scholar
53. Software

    1. Wandler FD

    (2025) ZFswim, version swh:1:rev:8ba7a68f6774d09025d600e0b7c6a484ac0be9cd

    Software Heritage.

    https://archive.softwareheritage.org/swh:1:dir:73e4bbca0d6976514e934084fa8758ad0c30ad50;origin=https://github.com/fwandler/ZFswim;visit=swh:1:snp:7af60a822b3978f29e11607fb68d0b964577e8bb;anchor=swh:1:rev:8ba7a68f6774d09025d600e0b7c6a484ac0be9cd
54. 1. Yakovenko S
    2. Mushahwar V
    3. VanderHorst V
    4. Holstege G
    5. Prochazka A

    (2002) Spatiotemporal activation of lumbosacral motoneurons in the locomotor step cycle

    Journal of Neurophysiology 87:1542–1553.

    https://doi.org/10.1152/jn.00479.2001

    • PubMed
    • Google Scholar
55. 1. Zhang K

    (1996) Representation of spatial orientation by the intrinsic dynamics of the head-direction cell ensemble: a theory

    The Journal of Neuroscience 16:2112–2126.

    https://doi.org/10.1523/JNEUROSCI.16-06-02112.1996

    • PubMed
    • Google Scholar

Author details

1. F David Wandler

   Institute of Neuroscience, University of Oregon, Eugene, United States

   Contribution

   Conceptualization, Formal analysis, Writing – original draft

   Competing interests

   No competing interests declared

2. Benjamin K Lemberger

   Institute of Neuroscience, University of Oregon, Eugene, United States

   Contribution

   Conceptualization, Formal analysis

   Competing interests

   No competing interests declared

3. David L McLean

   Discovery Brain Sciences, University of Edinburgh, Edinburgh, United Kingdom

   Contribution

   Conceptualization, Writing – original draft

   Competing interests

   No competing interests declared

   "This ORCID iD identifies the author of this article:" 0000-0001-6337-2301

4. James M Murray

   Institute of Neuroscience, University of Oregon, Eugene, United States

   Contribution

   Conceptualization, Formal analysis, Writing – original draft

   For correspondence

   jmurray9@uoregon.edu

   Competing interests

   No competing interests declared

   "This ORCID iD identifies the author of this article:" 0000-0003-3706-4895

• 485

  views

• 18

  downloads

• 2

  citations

Views, downloads and citations are aggregated across all versions of this paper published by eLife.

Citations by DOI

• 1

  citation for umbrella DOI https://doi.org/10.7554/eLife.106658

• 1

  citation for Reviewed Preprint v1 https://doi.org/10.7554/eLife.106658.1