|Home | About | Journals | Submit | Contact Us | Français|
Spinal circuits can generate locomotor output in the absence of sensory or descending input, but the principles of locomotor circuit organization remain unclear. We sought insight into these principles by considering the elaboration of locomotor circuits across evolution. The identity of limb-innervating motor neurons was reverted to a state resembling that of motor neurons that direct undulatory swimming in primitive aquatic vertebrates, permitting assessment of the role of motor neuron identity in determining locomotor pattern. Two-photon imaging was coupled with spike inference to measure locomotor firing in hundreds of motor neurons in isolated mouse spinal cords. In wild-type preparations, we observed sequential recruitment of motor neurons innervating flexor muscles controlling progressively more distal joints. Strikingly, after reversion of motor neuron identity, virtually all firing patterns became distinctly flexor like. Our findings show that motor neuron identity directs locomotor circuit wiring and indicate the evolutionary primacy of flexor pattern generation.
The mammalian nervous system is charged with the task of moving limbs—a challenge met through the construction of spinal circuits that coordinate interwoven patterns of muscle activity. Motor patterns reflect the activation of selected pools of motor neurons which, in turn, are driven by descending commands, peripheral feedback, and input from spinal premotor interneurons. Many studies have invoked the idea that local spinal circuits alone can sustain motor neuron burst firing in patterns that resemble the rhythmic alternation of antagonist muscles during locomotion (Grillner and Zangger, 1975; Kiehn and Kjaerulff, 1996; Kudo and Yamada, 1987). Yet the basic rules of spinal circuit organization that govern the rhythmicity and alternation of locomotor output remain unclear.
Attempts to delineate the spinal circuitry of mammalian locomotion have focused largely on connections among interneurons with presumed roles in pattern generation. One long-held view proposes that the premotor circuits that direct the alternation of antagonist flexor and extensor muscles exhibit an interdependence achieved through reciprocal interneuronal connections (Brown, 1914; McCrea and Rybak, 2008; Talpalar et al., 2011; Zhang et al., 2014). But the obligate role of reciprocal connectivity has been called into question by observations that rhythmic flexor or extensor motor output can, under rare circumstances, occur without activation of their antagonist pair (Burke et al., 2001; Pearson and Duysens, 1976; Zhong et al., 2012). Because spinal interneurons should be capable of distinguishing the identity of flexor and extensor motor neurons, we reasoned that insight into the organization of locomotor circuits might emerge from a focus on the recognition and selection of motor pools by premotor interneurons, rather than on the intricacies of interneuron interconnectivity.
The genetic identities, muscle targets, and functional specialization of motor neurons have diversified greatly during vertebrate evolution, suggesting the utility of addressing the influence of motor neuron identity on locomotor pattern. Within this broad evolutionary context, certain physiological findings are consistent with the idea that mammalian flexor networks evolved by co-opting a core axial motor circuit responsible for swimming in ancestral aquatic vertebrates. In primitive vertebrates, body undulations during swimming reflect the sequential recruitment of motor neurons innervating segmentally arrayed axial muscles (Grillner and Wallén, 2002). A similar wave-like sequence of motor neuron activation is evident from ventral root recordings at thoracic levels in the isolated rodent spinal cord during locomotor-like activity (Beliez et al., 2015; Falgairolle and Cazalets, 2007). This thoracic wave reflects the firing of median (MMC) and hypaxial (HMC) motor column neurons that innervate trunk and body wall muscles—the mammalian derivatives of primitive axial muscles (Kusakabe and Kuratani, 2005). Intriguingly, the firing of lumbar level flexor motor neurons represents a caudal continuation of the thoracic activity wave, whereas extensor motor neurons burst in antiphase (Falgairolle and Cazalets, 2007). This continuity of thoracic and flexor firing may reflect the reappropriation of axial circuits for flexor pattern generation and thus the evolutionary primacy of the flexor system. The idea that the basic organization of modern flexor circuits predates the emergence of extensor circuits implies that the generation of flexor-like patterns may not require interdependence between flexor and extensor circuits.
To explore the concept of flexor primacy and examine how motor neuron identity shapes the formation of locomotor circuits, we constructed cellular resolution maps of locomotor pattern in the absence of descending commands and sensory feedback. Two-photon imaging and spike inference were combined to measure the firing of hundreds of target-defined motor neurons in an isolated neonatal mouse spinal cord preparation induced to locomotor-like activity (Bonnot et al., 2002; Kwan et al., 2009). Our analysis revealed that motor pools innervating muscles with synergistic functions fire synchronously and that flexor pools are activated in a ventral-to-dorsal sequence that matches the proximodistal order of their target muscles along the limb.
This characterization of the wild-type pattern of motor neuron activation served as a reference in analyzing how the ancestral reversion of lumbar motor neuron identity modifies locomotor pattern. The concept of flexor primacy suggests that reversion of lateral motor column (LMC) neurons to an ancestral-like state will lead to their recruitment of flexor-defining premotor inputs. To address this possibility, we used genetic inactivation of the FoxP1 transcription factor to convert limb-innervating motor neurons to an HMC-like ground state (Dasen et al., 2008; Kusakabe and Kuratani, 2005; Rousso et al., 2008). In FoxP1 mutant preparations, we find that virtually all limb-innervating motor neurons—those innervating extensor as well as flexor limb muscles—are activated with the precise temporal features of flexor motor neurons. These observations show that the subtype identity of motor neurons profoundly influences the pattern of motor output. They also lend credence to the idea that a flexor-like motor pattern emerged during vertebrate evolution without reliance on an opponent extensor circuit.
We monitored Ca2+-sensitive fluorescence in hindlimb-innervating motor neurons in isolated postnatal day 2 to 5 mouse spinal cord preparations induced to a state of locomotor-like activity by glutamate and monoamine receptor agonists (5 μM NMDA, 10 μM 5-HT, 50 μM DA) (Figures 1A–1D; Kudo and Yamada, 1987). Motor neuron expression of the Ca2+ indicator GCaMP3 was achieved by crossing mice carrying a ROSA-CAG-lsl-GCaMP3 allele (Zariwala et al., 2012) with Olig2::Cre or ChAT::Cre motor neuron driver lines (Lowell et al., 2006; Sürmeli et al., 2011). Prior to imaging, groups of synergist muscles were injected with Alexa 555- or 647-conjugated cholera toxin B subunit (CTB) to identify motor neurons by their targets. Two-photon microscopy was used to acquire 90 s GCaMP3 fluorescence image sequences from 22 to 64 sagittal imaging fields (512 μm × 512 μm) that collectively spanned lumbar segments L2 to L6. Concurrent recordings of rhythmic activity from ventral roots L1 or L2 provided a reference signal for measuring motor neuron burst firing phase, with the locomotor cycle defined as the interval between peaks of L1/2 activity (peaks = 0°). Recordings from L4 or L5 (Figure 1B) and contralateral L1/L2 roots (data not shown) established that the alternating burst firing characteristic of locomotor activity was evident in each preparation.
We aimed to define firing features through the analysis of Ca2+-sensitive fluorescence from motor neuron cell bodies, but slow Ca2+ extrusion and noise in fluorescence measurements obscure prominent burst features such as duration and the phase of peak firing (Figure S1A; Helmchen and Tank, 2005). To overcome this problem, we used an improved, model-based statistical algorithm that infers the spike train most likely to underlie a somatic fluorescence time series (Figures S1B and S1C; Pnevmatikakis et al., 2014). This algorithm fits fluorescence data using a model of spike-related fluorescence fluctuations that assumes each action potential results in a fluorescence transient with instantaneous rise and exponential decay in the added presence of Gaussian noise. For each somatic fluorescence time series, the algorithm yields a relative estimate of the number of spikes that occurred during each imaging frame. These normalized spike counts were assembled into histograms that display the rhythmic burst firing of each motor neuron during the image sequence (black bars in Figure 1E). To quantify burst timing, the mean phase of each burst was calculated, and the median of these values was defined as a neuron’s phase tuning (Figures 1D and 1F).
The validity of such quantification depends on the ability of the spike inference model to capture the relationship between firing and fluorescence. The model was calibrated and its applicability evaluated by exploiting the fact that motor neurons activated antidromically by ventral root stimulation fire in patterns that follow stimulus timing (Figures S2A–S2E; Bonnot et al., 2005). For each experimental preparation, a single fluorescence transient decay time constant was computed using fluorescence measurements obtained during patterned antidromic stimulation that mimicked locomotor-like rhythmic burst firing. Use of these preparation-specific time constant values corrected for decay time variation between preparations (Figures S2F–S2H). In addition, analysis of somatic fluorescence acquired during antidromic activation indicated that models incorporating the saturation of indicator binding do not provide a more accurate prediction of spiking (Figures S1D and S1E), justifying use of a model that does not take saturation into account.
To assess the accuracy of spike inference, we examined phase tuning estimates for individual motor neurons during antidromic stimulation (Figure S2D). Tuning measurements derived from spike inference were nearly identical to values computed directly from antidromic stimuli (mean difference ± SD = −2.0 ± 10.7°, n = 367 neurons; Figures S2G and S2H). Thus, spike inference permits accurate estimation of motor neuron phase tuning.
In each spinal cord preparation, motor neurons are spread across many imaging fields, and as such, neuron-by-neuron comparisons of phase tuning require that values be stable over time. To assess tuning stability, we imaged a subset of fields in individual preparations at time points separated by 20 to 220 min. Importantly, even if tuning is stable, errors intrinsic to the measurement of burst phase from inferred spiking will result in variation in tuning estimates between time points. We estimated this error from the tuning of motor neurons imaged during antidromic activation, when all neurons fire in synchrony. The distribution of these tuning values indicated that two separate estimates of the same underlying tuning would differ on average by 10.1°. In comparison, temporally separated estimates of motor neuron firing during agonist-induced locomotor-like activity exhibited a similar difference of only 12.0° on average (n = 1,714 neuron pairs; Figures 1G–1J). Moreover, the slope of a linear regression fit indicated an incremental deviation of tuning values of only 2.8° per hr. Together, these findings establish that the phase tuning of LMC motor neurons in individual preparations is relatively stable over the duration of data collection. Thus, phase tuning estimates are both accurate and stable, enabling assessment of the relative tuning of motor neurons that innervate different limb muscles.
Pools of motor neurons that innervate muscles with similar functions at an individual joint form functional synergy groups and are clustered within the spinal cord (McHanwell and Biscoe, 1981; Vanderhorst and Holstege, 1997). To examine whether the phase tuning of motor neurons segregates with synergy group identity, we analyzed between 400 and 1,400 limb-innervating motor neurons in each preparation that exhibited phasic firing (mean = 818 motor neurons; Figures S3A–S3C; Berens, 2009). Spatial tuning maps were constructed, with the position of each motor neuron in three-dimensional space noted in a color that indicates its tuning (Figures 2A–2J and Movies S1 and S2). These maps revealed numerous motor neurons with tuning close to the reference ventral root activity peak (L1/2 = 0°), and many others with near-antiphase (~180°) tuning, at each lumbar segmental level (Figures S4A–S4D). Motor neurons with similar tuning were arranged in rostrocaudally elongated clusters that formed clear boundaries with other neuronal clusters of distinct tuning. These coherent clusters were similar in shape and extent to motor pool synergy groups, suggesting a direct correspondence between firing phase and synergy group identity. These findings contrast with prior reports of a wave-like sequence of motor neuron activation along the rostrocaudal axis of the LMC that transgresses synergy group boundaries (O’Donovan et al., 2008).
To probe further the correspondence between identity and firing phase, we measured the phase tuning of motor neurons that had been assigned to particular synergy groups (Figures 2K–2N). CTB was injected into four muscle groups: the intrinsic foot (IF; toe flexors), anterior crural (AC; ankle flexors), quadriceps (Q; knee extensor/hip flexor), and gluteal (G; hip extensor/flexor) muscles, and the tuning of retrogradely labeled motor neurons was measured. Identified IF and AC motor neuron populations exhibited unimodal tuning distributions, whereas Q and G motor neuron populations displayed bimodal distributions (Figures S4E–S4H). Among Q motor neurons, the more lateral, presumptive rectus femoris (RF) motor neurons were tuned near 0°, whereas the more medial, presumptive vastus (V) motor neurons were tuned close to 180° (De Marco Garcia and Jessell, 2008; Vanderhorst and Holstege, 1997; Figure S4G). Similarly, for G motor neurons, a more rostral, presumptive tensor fasciae latae (TFL) cluster was tuned near 0°, whereas a caudal cluster containing the three remaining gluteal motor pools (GM) was tuned around 180° (Figure S4H). These results are consistent with functional definitions of RF and TFL as hip flexors and V and GM muscles as knee and hip extensors, respectively (Platzer, 2004). The alignment of six synergy groups with phasically homogeneous clusters in tuning maps supports the view that phase tuning is organized in register with synergy group identity.
If locomotor firing is synergy group specific, then cycle-by-cycle covariation in the phase of burst firing might be stronger within than between groups. To test this possibility, we evaluated burst phase covariation using a synchrony index that reflects across-cycle consistency in phase differences between pairs of motor neurons (Figures 3A–3C; Mormann et al., 2000). We observed higher synchrony among motor neurons assigned to the same synergy group by CTB labeling (Figure 3D, mean index ± SEM = 0.51 ± 0.007, n = 517 pairs; p < 10−10, Wilcoxon test) and lower synchrony among motor neurons assigned to different synergy groups (Figure 3D, mean index ± SEM = 0.33 ± 0.026, n = 68 pairs; p = 4.2 × 10−7, Wilcoxon test; comparing with synergist pairs, p = 4.4 × 10−10, Wilcoxon test; p = 4.7 × 10−6 after controlling for differences in proximity). Thus, synergist motor neurons are preferentially synchronized.
We also assessed the degree of phase synchrony for synergist motor neuron pairs as a function of their separation. Synchrony indices did not vary significantly with proximity along the rostrocaudal axis (Spearman correlation (ρ) = −0.07, p = 0.12; Figure 3E). In contrast, we detected a shallow proximity dependence along the dorsoventral axis (ρ = −0.09, p = 0.04; Figure 3F), which may reflect slightly elevated synchrony within the motor pools that comprise each synergy group. Nevertheless, as a whole, these findings indicate that the major determinant of synchrony in motor neuron burst phase is synergy group membership.
Walking is characterized by the sequential activation of limb muscles, with a precision in recruitment that reflects their biomechanical function (Rossignol, 1996). To examine the degree to which the order of muscle recruitment can be imposed by local spinal circuits, we characterized the sequential activation of flexor synergy groups innervating different limb joints. Normalized spike histograms were used to derive an average firing rate across the locomotor cycle for individual motor neurons within defined synergy groups (Figures 4A–4F, bottom; Figures 2K–2N). Because Q and G motor neurons display bimodal tuning, we used k-means clustering (k = 2) to separate the cycle-averaged firing rates of both groups, yielding distinct RF and V pools at different mediolateral positions within the Q population, and rostrocaudally distinct TFL and GM pools within the G population (Figures S4G and S4H). Mean cycle-averaged firing rates for individual synergy groups showed that the phase of peak firing and burst duration were consistent across preparations (Figures 4A–4F, colored traces in top panels).
Strikingly, we found a tight correspondence between the dorsoventral position of synergy groups and the onset of their activation, assessed here as the time at which firing rates attained 50% of their eventual maxima (Figures 4G and 4H). The mean firing of the ventral-most motor neurons innervating the hip flexor TFL had an onset at a cycle phase of −43.8 ± 20.9° (median ± SE of median, n = 34 neurons). The firing of more dorsally positioned motor neurons innervating RF, a hip flexor with a more distal origin and insertion than TFL, had an onset at −33.0 ± 4.4° (n = 38). The firing of still more dorsally positioned motor neurons innervating ankle flexor AC muscles had an onset at −13.2 ± 2.2° (n = 106). Finally, the dorsal-most motor neurons, which innervate toe flexor IF muscles, had an onset at 19.2 ± 2.6° (n = 72). The correlations of both burst onset phase and peak firing phase with position were strong (onset: ρ = 0.70, p < 10−10; peak: ρ = 0.69, p < 10−10; Figures 4G and 4H). Thus local spinal circuits appear able to impose a motor neuron activation order that follows their settling positions and thus the proximodistal order of their target muscles.
Are locomotor firing patterns modified by reverting motor neuron columnar identity to an ancestral-like state? To test this possibility, mice harboring a conditional FoxP1 allele were crossed with an Olig2::Cre driver line to generate motor neuron selective FoxP1MNΔ mutants (Dasen et al., 2008). In FoxP1MNΔ mice, motor neurons fail to acquire LMC columnar and pool-specific identities and instead assume many of the features of thoracic HMC neurons. Transfated motor neurons in FoxP1MNΔ mice fail to exhibit a stereotyped relationship between neuronal position and muscle target (Figure S5), yet both flexor and extensor muscles are still innervated. As a consequence, muscles co-contract, limbs are rigid, and normal locomotion is precluded (Sürmeli et al., 2011).
To assess the impact of the reversion of motor neuron identity on locomotor firing, we first monitored lumbar ventral root activities. Induction of locomotor-like activity in isolated FoxP1MNΔ preparations elicited rhythmic root activity at frequencies similar to those in wild-type spinal cord (Figure S6; p = 0.66, Wilcoxon test). However, the normal ipsilateral alternation between L2 and L5 roots was replaced by near synchrony (Figures S7A–S7C), even though alternation between contralateral roots was still evident (data not shown). Phase differences between T9–10 and L2 ventral root activity peaks were also similar in wild-type and FoxP1MNΔ preparations (Figures S7D–S7F; p = 0.85, two-sample, two-tailed t test). Thus, the reversion of motor neuron columnar fate abolishes rostrocaudal alternation in motor neuron burst firing. Nevertheless, rostral lumbar ventral root activity still provides a comparable phase reference.
To probe the cellular origins of changes in lumbar locomotor activity, we performed Ca2+-sensitive fluorescence imaging of motor neurons and ventral root recording in FoxP1MNΔ preparations. Motor neuron phase tuning maps (200–900 motor neurons/map; mean = 656 motor neurons) revealed substantial differences from tuning in wild-type preparations (Figures 5A–5E and Movies S3 and S4; p = 0.0002, Kolmogorov-Smirnov [K-S] test). Motor neurons generally exhibited rhythmic firing at a common phase, close to 0° (Figure 5C), with only ~2% (29/1413) of FoxP1MNΔ motor neurons bursting closer to 180° (Figures 5F–5J). This anomalous minority likely reflects the redundant function of FoxP4 and thus the preservation of LMC identity in a small fraction of limb-innervating motor neurons (Dasen et al., 2008).
To exclude the possibility that motor neurons targeting certain muscles are not rhythmically active in FoxP1MNΔ preparations, we analyzed the activity of identified motor neurons. FoxP1MNΔ motor neurons retrogradely labeled by CTB injection into IF, AC, G, and gastrocnemius (GS, ankle extensor) muscles exhibited highly overlapping tuning distributions (IF: −3 ± 21°, mean ± SD, n = 46 neurons; AC: 20 ± 26°, n = 88; G: 13 ± 23°, n = 8; GS: −19 ± 28°, n = 33) in marked contrast to wild-type preparations. In particular, we noted a profound conversion of extensor (G and GS) motor neuron firing to a flexor-like phase (Figures 5D and 5E). In addition, IF motor neurons now fired slightly earlier than AC neurons, the opposite of their wild-type relationship. We conclude that the loss of FoxP1 erodes the normal synergy-group-specific patterns of motor neuron burst firing and promotes flexor monotony.
We also examined the precision with which motor neurons adopted flexor-like firing in FoxP1MNΔ preparations. Cycle-averaged firing rates of wild-type motor neurons could be separated into two sets using k-means clustering (Figures 6A and 6B), revealing well-separated sets within individual preparations (clustering index mean ± SEM = 3.99 ± 0.26, n = 12 spinal cords) and across different preparations (clustering index = 3.85, n = 5967 neurons). One set of firing rates was characterized by brief bursts (86.7 ± 24.0° duration, mean ± SD, n = 4,212 neurons) with phase tunings early in the locomotor cycle (13.7 ± 27.5°). The second set exhibited prolonged bursts (165.7 ± 46.5° duration, n = 1,755 neurons) tuned later in the locomotor cycle (166.2 ± 46.1°). We found that 99.4% (175/176) of identified motor neurons innervating AC and IF muscles were included within the early firing set. This finding suggests that the early-and late-firing sets are comprised of flexor and extensor motor neurons, respectively (Figures 6C and 6D).
An equivalent analysis of FoxP1MNΔ motor neurons revealed that the cycle-averaged firing rates for virtually all neurons precisely matched those of wild-type flexor motor neurons, both in phase tuning and burst duration (Figures 6E–6H). k-means clustering failed to identify well-separated sets from individual FoxP1MNΔ preparations (clustering index mean ± SEM = 0.48 ± 0.14, n = 4) or among neurons aggregated across different FoxP1MNΔ preparations (cluster index = 0.27, n = 1,413 neurons). Cluster separation was significantly less than for wild-type firing rates (p = 2.1 × 10−6, one-tailed unpaired t test). Collectively, FoxP1MNΔ motor neurons exhibited distributions of phase tuning (mean ± SD = 12.0 ± 42.2°) and burst duration (90.7 ± 29.3°) that were very similar to those of the early-firing wild-type set that comprises flexor motor neurons (Figure 6F). Consistent with this, analysis of phase tuning and burst duration distributions revealed that firing exhibited 21-fold greater similarity to that of wild-type flexors than to that of extensors (Figures S7G and S7H). Taken together, our results indicate that almost all hindlimb-innervating motor neurons fire in a precisely flexor-like pattern after genetic reversion of motor neuron columnar identity.
Our analysis reveals that the identity of motor neurons determines temporal features of locomotor activation. Most critically, the reversion of LMC neurons to an ancestral HMC-like columnar character induces essentially all limb-innervating motor neurons to fire in a flexor-like pattern, a strong indication of the primacy of flexor pattern generation. We discuss below the relevance of this relationship for the current organizational state of mammalian locomotor circuits.
The temporal features of motor neuron firing observed in neonatal spinal cord in vitro exhibit distinctions from, and commonalities with, the pattern of activation of their muscle targets in adults in vivo. The discrepancies imply an influence of descending commands or sensory feedback in shaping locomotor pattern and potentially the refinement of circuits as the spinal cord matures.
Included among the discrepancies are differences in the number and duration of bursts. We observed that TFL and RF motor neurons burst only once per locomotor cycle in vitro, yet their target muscles exhibit dual burst activity in many locomotor contexts in vivo (Rossignol, 1996; Yakovenko et al., 2002). This difference likely reflects sensory feedback, inducing a second phase of motor neuron bursting per cycle, or shifting the firing phase of a subset of neurons within the TFL and RF pools (Loeb, 1985; Perret and Cabelguen, 1980). A second distinction is that flexor motor neurons exhibit relatively brief bursts in vitro, whereas flexor muscle activation in vivo can occupy a much greater proportion of the locomotor cycle. Studies in cats and mice in vivo suggest that the duration of muscle activation is also governed by sensory feedback, in part through the regulation of muscle offset timing (Akay et al., 2014; Lam and Pearson, 2001). Together, these findings suggest that spinal circuits are sufficient to produce a basic dynamical template of locomotor activity that is subject to refinement through sensory feedback.
Nevertheless, conserved features emerge from a comparison of locomotor patterns in vitro and in vivo, most clearly in the timing of recruitment of mouse motor neurons that innervate synergist muscles acting on different joints. Our findings indicate that local circuits are sufficient to direct the activation of motor neurons innervating synergistic flexor muscles in a ventral-to-dorsal sequence that matches the proximodistal order of their muscle targets. EMG recordings from mouse hindlimb muscles during walking document the onset of hip, knee, and ankle flexor muscle activation in a similar proximodistal order (Akay et al., 2014). In cat, however, muscle activation sequences do not necessarily conform to the recruitment order we observe in vitro in mouse (Krouchev et al., 2006; Rossignol, 1996; Yakovenko et al., 2002). Such differences could reflect developmental changes, interspecies differences in local circuit wiring, or the added influence of descending commands and sensory feedback. The activation sequence we observed in vitro implies that premotor interneurons are able to recognize and select from synergy groups governing different limb joints.
We emphasize that two-photon Ca2+ imaging reveals aspects of the organization of locomotor firing across the LMC that could not have been discerned from motor nerve or muscle recordings, which conflate the activity of individual motor neurons. The high spatial resolution afforded by imaging revealed that motor neurons exhibit abrupt changes in firing at the boundaries between synergy groups. The spatial resolution and broad coverage provided by our datasets were critical in exposing spatially extended synchrony. Cellular resolution estimates of neuronal firing were also necessary to delineate the precision of flexor firing and its predominance among FoxP1MNΔ motor neurons.
Our observations also point to the inadequacy of monitoring ventral root activity alone when probing the organization of mammalian locomotor circuits. Interpretations of in vitro ventral root recordings have typically relied upon the notion that L2 and L5 root activity peaks reflect, respectively, flexor and extensor motor neuron firing phases. Our findings document sizeable populations of motor neurons that exhibit distinct flexor or extensor firing patterns at each lumbar segment. Differences in the number of flexor and extensor motor neurons across segments and/or differences in motor neuron firing rate (Yakovenko et al., 2002) could contribute to this discrepancy. Clearly, a reliance on ventral root activity peaks ignores the extent of diversity in motor neuron activities present at individual segmental levels of the spinal cord.
At first glance, the heterogeneous firing patterns across different flexor synergy groups appear inconsistent with a recent analysis of ventral root recordings from isolated neonatal rat spinal cord (Dominici et al., 2011). This study concluded that locomotor output from neonatal preparations is well approximated by two alternating rhythmic patterns, in contrast to the greater complexity seen in EMG recordings from behaving adults. This discrepancy prompted us to perform an analysis similar to that of Dominici et al. (2011) but using the cycle-averaged firing rates of the many motor neurons we recorded in individual neonatal spinal cords. Non-negative matrix factorization revealed that four components are needed to explain ~90% of the variance in locomotor firing across the neonatal LMC, as in adult EMG (Figures 7A–7E). Similar results were obtained using principal-component analysis (Figure 7F). Thus, the complexity of locomotor output from the isolated neonatal rodent spinal cord is similar to that generated in adults in vivo, contrary to the conclusion of Dominici et al. (2011).
What explains the finding that essentially all limb-innervating motor neurons fire in a flexor-like pattern after FoxP1MNΔ-mediated reversion of motor neuron identity?
One possibility is that LMC neurons play an active role in the differentiation or function of pattern-generating circuits. The reversion of motor neuron identity may undermine the formation of extensor circuits, leaving, by default, a monophasic flexor system. Mechanistically, LMC neurons could be the source of a secreted signal that instructs the assembly of extensor circuits. In fact, there is precedent for the secretion by LMC motor neurons of a signal, retinoic acid, which drives the diversification of limb-innervating motor neurons (Sockanathan and Jessell, 1998). Alternatively, synaptic feedback from LMC motor neurons may be necessary for extensor pattern generation. Recruitment of Renshaw inhibitory or equivalent excitatory interneurons by motor neuron axon collaterals might influence ongoing interneuron network activity (Alvarez and Fyffe, 2007; Machacek and Hochman, 2006; O’Donovan et al., 2010).
A second scenario is suggested by the apparent ability of premotor interneurons to discriminate flexor and extensor motor neurons. The ancestral similarity of flexor LMC and HMC motor neurons may lead to the expression of shared surface recognition features on these two motor neuron classes, permitting flexor but not extensor premotor interneurons to form connections with ancestrally reverted motor neurons. In this view, normal premotor activity would be preserved in FoxP1MNΔ spinal cord, but extensor premotor interneurons would fail to recognize HMC-like motor neurons. The finding that a small minority of motor neurons with extensor-like firing are still present in FoxP1MNΔ preparations indicates that extensor premotor circuits are at least in part preserved. In addition, the scattered distribution of the few extensor-firing motor neurons in FoxP1MNΔ preparations implies that premotor interneurons are able to select individual target motor neurons with precision.
Whether extensor pattern generation is diminished or HMC-like motor neurons recruit only flexor interneuronal input, the prevalence of flexor firing in FoxP1MNΔ preparations provides strong support for the evolutionary primacy of flexor pattern generation. In mammals, the phasic continuity evident between limb flexor and thoracic ventral root activity and the similarity between wave-like patterns in mammalian thoracic and primitive vertebrate motor output are consistent with the idea that flexor pattern generation emerged by co-opting primitive swim circuits. This implies that paired flexor and extensor patterns did not emerge jointly at the evolutionary onset of limb-based locomotion. In the direct ancestors of tetrapods, the extensor system may have evolved as a later elaboration of spinal circuitry to promote ground repulsion through limb extension.
That the basic organization of modern flexor circuits predates the evolutionary emergence of extensor circuits further implies that the generation of flexor-like pattern can occur without opponent input from extensor premotor circuits. This view agrees with the subordinate nature of extensor pattern generation suggested by certain observations. Notably, locomotor firing in mice and cats is subject to brief and sporadic periods of quiescence, termed “deletions,” that persist for several cycles. Flexor burst deletions are accompanied by tonic extensor motor neuron firing, whereas flexor motor neuron bursting continues unabated during extensor burst deletions (Duysens, 1977, 2006; Zhong et al., 2012). Other studies have indicated that the rhythm of locomotor firing may be determined by populations of interneurons that burst exclusively in flexor phase and, in turn, drive pattern-forming circuits (Brownstone and Wilson, 2008; Pearson and Duysens, 1976), which could at least partly explain how flexor dominance is imposed. Taken together with our findings, these results suggest that the late addition of extensor pattern, coupled with the need for flexor-extensor coordination, led to an asymmetric dependence in pattern-generating circuits, with flexor circuits having a dominant role.
Genetic studies have shown that locomotor firing persists after the loss of any single cardinal interneuron population (Crone et al., 2008; Gosgnach et al., 2006; Lanuza et al., 2004; Zhang et al., 2008), suggesting that the generation of locomotor firing can be achieved through a diverse array of interneuron network architectures. In addition, modeling studies have shown that locomotor-like activity patterns can be read out from neural networks permitted considerable flexibility in their connectivity, as long as the network outputs are weighted appropriately (Sussillo and Abbott, 2009). In this context, and with a new emphasis on motor neuron recognition, it is conceivable that interneuronal connectivity in locomotor circuits is only weakly constrained, whereas output connections onto motor pools are precisely specified.
All experiments and procedures were performed according to NIH guidelines and approved by the Institutional Animal Care and Use Committee of Columbia University.
Motor neurons were retrogradely labeled in vivo at P1–P3 via intramuscular injections of cholera toxin B subunit conjugated to Alexa 555 or 647 (CTB; Life Technologies) (Sürmeli et al., 2011).
Spinal cords were removed from mice, aged 2–5 days postnatal, and submerged in artificial cerebrospinal fluid (ACSF) held at constant temperature (24–25°C). Suction electrode recordings were simultaneously obtained from multiple ventral roots. Ca2+ transients were measured from GCaMP3-expressing LMC motor neurons in a single segment while the corresponding ventral root was antidromically stimulated to evoke motor neuron activity, enabling the calibration of the model of spike-related fluorescence fluctuations we used for spike inference. Subsequently, locomotor firing was induced by adding a cocktail of rhythmogenic agonists to the ACSF (5 μM NMDA, 10 μM 5-HT, 50 μM DA). Starting 1 hr later, we collected fluorescence image sequences throughout the imageable extent of the LMC.
An Ultima microscope (Prairie Technologies) with a 20× objective (1.0 numerical aperture, 2 mm working distance; XLUMPLFLN, Olympus) was used to acquire all fluorescence images (256 x 256 pixels/frame). GCaMP3 was excited using a Chameleon Ultra II laser (Coherent) tuned to 940 nm and, in 17 of 19 preparations, raster scanned across the preparation at 60 Hz using a resonant galvanometer. These signals were downsampled to 15 Hz to increase the signal-to-noise ratio. In 2 of 19 preparations, the laser was scanned at 8 Hz with conventional 6 mm galvanometers in a spiral trajectory. GCaMP3 emission was collected using a GaAsP photomultiplier tube (Hamamatsu; 525/50 emission filter).
The centroids of motor neuron somata were manually defined in ImageJ and used to demarcate a preliminary region of interest (ROI) around each soma. These ROIs were further refined using automated MATLAB scripts to include only those pixels likely to arise from each soma. Time series of ROI-averaged fluorescence fluctuations (ΔF/F) were processed using a spike inference algorithm. Phase-tuning values were computed relative to peaks in simultaneously obtained ventral root recordings using inferred spiking activity.
We are grateful to D. Wu for assistance with mouse genotyping; M. Mendelsohn and N. Zabello for animal care; H. Tucker for the conditional FoxP1 strain; Y. Ivanenko for advice about analysis; and B. Han, K. MacArthur, S. Morton, and I. Schieren for technical assistance. We thank G.Z. Mentis for the custom-built recording stage used in the electrophysiology experiments and for advice about experimental approaches. We are grateful to R. Axel, E. Azim, J. Bikoff, R. Brownstone, S. Druckmann, A. Fink, A. Karpova, A. Murray, and J. de Nooij for helpful comments on the work and manuscript. T.A.M. was supported by the NSF Graduate Research Fellowship Program. L.P. was supported by DARPA W91NF-14-1-0269, ARO MURI W911NF-12-1-0594, NSF CAREER award IOS-0641912, and Simons Foundation Global Brain Research Award 325398. T.M.J. was supported by NIH grant NS033245, the Harold and Leila Y. Mathers Foundation, and Project A.L.S. and is an investigator of the Howard Hughes Medical Institute. A.M. is a Howard Hughes Medical Institute Fellow of the Helen Hay Whitney Foundation.
AUTHOR CONTRIBUTIONST.A.M. and T.M.J. devised the project. T.A.M., T.M.J., and A.M. designed experiments. T.A.M constructed experimental apparatus, performed the experiments, and wrote core data processing scripts. T.A.M. and A.M. analyzed data. T.A.M., E.P., L.P., and A.M. developed data analytical techniques. E.P and L.P. contributed unpublished spike inference algorithms. T.A.M., T.M.J., and A.M. interpreted data and wrote the manuscript.