From our previous work studying V2a neurons in larval zebrafish we know they are organized into: (1) fast and slow subtypes based on dorsoventral (DV) soma position and birth order (McLean et al., 2007; McLean et al., 2008; McLean and Fetcho, 2009), and (2) distinct subtypes that provide stronger direct versus indirect control of motor neuron recruitment (Menelaou et al., 2014; Menelaou and McLean, 2019). Specifically, V2a neurons with descending axons (V2a-D) provide stronger inputs to other premotor interneurons, while V2a neurons with bifurcating axons (V2a-B) provide stronger inputs to motor neurons (Menelaou and McLean, 2019), in line with differential rhythm and pattern control. However, we do not know if the same mechanisms of rhythmogenesis shape the phasic activation of distinct V2a types at different speeds.

To record from distinct types of V2a neurons, we targeted green fluorescent protein (GFP)-positive neurons located at midbody in the Tg(Chx10:GFP) line (Figure 1A, B; Kimura et al., 2006), since midbody neurons would be recruited during both fast anguilliform and slow carangiform modes (Müller and van Leeuwen, 2004; Thorsen et al., 2004). V2a-D and V2a-B neurons recruited at fast speeds can be distinguished by levels of Chx10 expression in transgenic lines (Menelaou and McLean, 2019), as reported in mice (Hayashi et al., 2018). To get a better sense of differences in the number and spatial distribution of distinct types, we split GFP-positive neurons into two categories based on a bimodal distribution of fluorescence intensity and normalized their soma positions to DV and mediolateral (ML) axes in the spinal cord. Larger, dimmer V2a neurons intermingled with smaller, brighter neurons along the DV axis (Figure 1C, left), but occupied more lateral positions (Figure 1C, right). Statistical analysis confirmed these observations, with no significant difference in either DV distribution (Figure 1D, left) or number per segment between the low or high GFP-positive types (Figure 1D, right), but a significant difference in ML position (Figure 1E).

To evaluate firing patterns and synaptic drive responsible for rhythmogenesis in different types, we recorded from V2a neurons during ‘fictive’ swimming evoked by a mild electrical shock, while simultaneously monitoring rhythmic motor output using ventral rootlet recordings (Figure 1F). During real swimming escape responses decelerate (Video 1), beginning in fast anguilliform mode and ending in slow carangiform mode (McLean et al., 2008; McLean and Fetcho, 2009). Plots of ventral rootlet burst frequencies with time confirm this pattern during ‘fictive’ escape responses, where we observe the highest frequencies at the beginning of the swim episode immediately following the stimulus (Figure 1G, left). Notably, zebrafish swimming is episodic (McHenry and Lauder, 2006; Soto and McHenry, 2020) and so it can continue intermittently after the initial escape swim episode. Subsequent episodes (Swim 2+) rarely exceeded peak firing frequencies observed in the first stimulus-evoked episode (>60 Hz in Swim 1) and there was more variability in frequency (Figure 1G, right). For purposes of analysis, we pooled data from the first (Swim 1) and subsequent (Swim 2+) episodes to capture as broad a frequency range as possible.

Thus, V2a neurons can be easily distinguished for targeting purposes using levels of GFP expression and we can assess their firing and synaptic drive during fast anguilliform and slow carangiform modes distinguished by frequency and temporal sequence within fictive escape swim episodes. This set the stage for asking if the same mechanism of rhythmogenesis drives the phasic activation of distinct V2a types at different speeds.

Rhythmogenesis can originate from cellular properties or synaptic drive or some combination of the two (Marder and Calabrese, 1996; Marder and Bucher, 2001). In theory, neurons that are intrinsically rhythmogenic can create patterned outputs from unpatterned inputs. Thus, to explore the intrinsic rhythmicity of V2a neurons we performed whole-cell current-clamp recordings from high and low GFP-positive V2a neurons and tested their firing responses to stable current steps. By including dye in our recording pipette, we confirmed post hoc that all high GFP neurons we targeted for recordings were V2a-Ds and all low GFP neurons were V2a-Bs (Figure 2A), as we would expect based on our prior work (Menelaou and McLean, 2019). Differences in sizes and DV distributions of V2a-D and V2a-B neurons in our current-clamp dataset (n = 47) were consistent with our analysis of high and low GFP neurons in stable lines (Figure 2B, left). Also, measurements of input resistance and rheobase current were consistent with differences in sizes and excitability between the two types (Figure 2C, left). Thus, we were confident of a sufficient sampling of V2a neurons to explore differences in rhythmogenic capacity related to type and recruitment order.

V2a-D neurons with ‘tonic’ responses and V2a-B neurons with higher-frequency ‘chattering’ ones were distributed dorsally and had the lowest input resistances in their respective types (Figure 2B–E). In tonic V2a-Ds, firing above rheobase current was observed consistently throughout the current step (Figure 2E, top), with instantaneous spike rates increasing linearly with increasing current, starting around 40 Hz and reaching 200 Hz (Figure 2D). In chattering V2a-Bs, firing above rheobase current was intermittent throughout the step (Figure 2F, top), with instantaneous rates starting around 200 Hz and reaching 600 Hz (Figure 2D). These observations are consistent with previous recordings from fast V2a-D and V2a-B neurons (Menelaou and McLean, 2019).

On the other hand, V2a-D and V2a-B neurons with ‘burst’ responses were distributed ventrally and had the highest input resistances within their respective types (Figure 2B–E). In bursting V2a neurons, firing above rheobase was characterized by higher-frequency spikes superimposed on lower-frequency membrane potential oscillations (Figure 2E, F, bottom). While the frequency of the slower oscillations increased with current (Figure 2E, F, bottom), instantaneous spike rates were already close to peak values at rheobase (500–600 Hz; Figure 2D). Beyond 2–3x rheobase current, slower oscillations were no longer obvious and the firing pattern more closely resembled the chattering one (data not shown), as observed previously for motor neurons (Menelaou and McLean, 2012).

Next, we compared the firing properties of distinct types of V2a neurons to their recruitment patterns during fictive escape swimming. As expected (Menelaou and McLean, 2019), both tonic V2a-D and chattering V2a-B neurons were recruited at the beginning of the episode at median swim frequencies above 30 Hz (Figure 2G, H), consistent with recruitment during fast anguilliform mode (Thorsen et al., 2004; McLean et al., 2008; McLean and Fetcho, 2009). In contrast, burst V2a-Ds fired more reliably at the end of the episode at median frequencies that rarely exceeded 30 Hz (Figure 2G, H), consistent with participation in slow carangiform mode (Thorsen et al., 2004; McLean et al., 2008; McLean and Fetcho, 2009). Burst V2a-Bs were recruited just below 30 Hz, and fired more reliably throughout the episode and over a broader frequency range (Figure 2G, H), particularly compared to the narrow range of bursting V2a-Ds (two sample Kolmogrov–Smirnov test on density distributions; D = 0.157, p < 0.001; BB, n = 6, BD, n = 8). A comparison between the step- and swim-firing patterns among V2a-D neurons revealed a remarkable correspondence (Figure 2G), with median spike frequencies (tonic) or oscillation frequencies (burst) overlapping median recruitment frequencies during swimming (Figure 2I, left). However, the link between step- and swim-firing patterns was less obvious for V2a-B neurons (Figure 2G). For example, while bursting V2a-B neurons recruited at swim frequencies exceeding 60 Hz had higher-frequency oscillations, step bursting rarely exceeded 20 Hz (Figure 2I, right). For chattering V2a-B neurons, there were no clear subthreshold oscillations to calculate rhythmicity and the broad range of spike frequencies resulted in large standard deviations (±74–130 Hz, n = 5) that rendered median values less interpretable.

Thus, distinct electrophysiological properties define distinct V2a types recruited at different speeds and V2a-D neurons are better suited to providing intrinsic rhythmogenic activity during fast and slow swimming modes. This is consistent with their proposed rhythm-generating and pattern-forming roles (Menelaou and McLean, 2019), and suggests V2a-B neurons rely more on rhythmic synaptic drive to shape their phasic firing patterns during swimming.

Our current-clamp data suggest that different types would rely differently on cellular properties versus synaptic drive for rhythmogenesis. To test this idea more directly, we performed voltage-clamp recordings during fictive escape swimming. To more effectively isolate synaptic currents, we included cesium in the patch solution. Since this blocks voltage-gated channels and spiking, we needed to classify fast and slow V2a-D and V2a-B neurons subtypes by first monitoring spikes in cell-attached mode. Consistent with our current-clamp recordings, in our voltage-clamp dataset (n = 39) we found V2a-D and V2a-B neurons that fired at the beginning of the escape episode or nearer the end (Figure 3A). V2a-D neurons fired more sparsely (Figure 3A, left) and were divided into two distinct groups: (1) neurons that fired exclusively at the beginning of the episode at median swim frequencies exceeding 40 Hz, and (2) neurons that fired later on in the episode at median swim frequencies rarely exceeding 30 Hz (Figure 3B, top and bottom left; Figure 3C). On the other hand, V2a-B neurons fired more reliably (Figure 3A, right) and median values were more continuous, making it more difficult to group neurons based on location within the episode (Figure 3B, top and bottom right) or swim frequency (Figure 3D).

To better segregate both types of V2a neurons based on speed, we used the frequency-dependent transition between anguilliform and carangiform modes to define V2a neurons recruited more reliably below 35 Hz as ‘slow’ and V2a neurons recruited more reliably above 35 Hz as ‘fast’ (Figure 4A). Consistent with our current-clamp recordings, slow V2a-D and V2a-B neurons in our voltage-clamp recordings were more ventrally distributed with higher resistances than fast V2a-D and V2a-B neurons (Figure 4B). A closer examination of spike timing relative to individual ventral rootlet bursts revealed median spike distributions in advance of the center of the motor burst, as expected for premotor excitatory interneurons involved in rhythmogenesis (Figure 4C). Plots of spike densities normalized to center burst revealed that V2a-B neurons begin to fire slightly in advance of V2a-D neurons (Figure 4D), consistent with differences in membrane resistance related to size, where neurons with faster time constants respond more quickly to synaptic inputs (Menelaou and McLean, 2019).

Next, to explore recruitment probability as a function of speed in more depth we divided the full range of fictive swimming frequencies into four bins distributed around 35 Hz. For V2a-D neurons, fast neurons were rarely recruited below 35 Hz and increased to a maximum of 50% recruitment reliability per motor burst above 35 Hz (Figure 4E, top left). Similarly, slow V2a-D neurons did not exceed 50% recruitment reliability per motor burst below 35 Hz and rarely fired above it (Figure 4E, top right). When collapsed into two bins below and above 35 Hz, these differences in recruitment probability were significant (Figure 4F, top). For V2a-B neurons, fast neurons fired more reliably below 35 Hz than fast V2a-Ds and increased to 100% recruitment reliability per motor burst above 35 Hz (Figure 4E, bottom left). Similarly, slow V2a-B neurons fired more reliably above 35 Hz than slow V2a-Ds, and at 100% reliability below it (Figure 4E, bottom right). Binning of recruitment among V2a-Bs revealed significant differences related to speed among the fast subtypes only (Figure 4F, bottom). Among the slow V2a-Bs, decreases in recruitment reliability were observed in the majority of neurons (n = 5 of 6), however the difference was not significant when pooled (Figure 4F, bottom).

In further support of the speed-based modular organization of V2a neurons, in slow types higher instantaneous spike rates were observed at lower burst frequencies, while in fast types higher spike rates were observed at higher burst frequencies (Figure 4G, top). Critically, spike rates observed during fictive swimming overlap with those observed during current steps between 1 and 2x rheobase, ranging from in 200–600 Hz (Figure 4H, top), with the fastest spike rates observed in fast V2a-B neurons. This suggests that the firing patterns observed during current injections between 1 and 2x rheobase are functionally relevant. When we measured spike numbers per burst, lower numbers were consistently observed at higher burst frequencies regardless of type or speed class (Figure 4G, bottom) consistent with less time to spike during shorter cycle periods. In addition, V2a-B neurons fired significantly more spikes per burst than V2a-Ds on average (Figure 4H, bottom), consistent with their higher recruitment probabilities.

Thus, within our voltage-clamp dataset we have V2a-D and V2a-B neurons operating during anguilliform and carangiform swimming modes. While V2a-D neurons fire more sparsely and in a stricter modular fashion, V2a-B neurons fire more reliably and over a broader range of frequencies, consistent with occupation of a pattern-forming layer downstream of distinct fast and slow rhythmic signals.

To get a better sense of the relative segregation of synaptic inputs to V2a-D and V2a-B neurons based on speed, we first focused on the maximum amplitude of excitatory and inhibitory currents recorded during fictive escape swimming. Clamping V2a neurons at a potential to isolate excitatory post-synaptic currents (EPSCs) revealed faster time course, compound EPSCs superimposed on a slower, tonic inward current (Figure 5A, B). While tonic excitation was observed the entirety of the episode, fast V2a-Ds received compound EPSCs exclusively during high-frequency swimming at the beginning (Figure 5A, top), while slow V2a-D’s received compound EPSCs exclusively during slower swimming later in the episode (Figure 5A, bottom), consistent with their recruitment patterns. A similar pattern was observed among V2a-Bs, however compound EPSCs were observed over a broader range of frequencies (Figure 5B), also consistent with their recruitment patterns. We also noted differences in maximum levels of current related to type. When we plotted maximum current (tonic current + compound EPSCs) versus input resistance, we observed a significant negative correlation, with lower resistance V2a-Bs receiving orders of magnitude more excitation than higher resistance V2a-Ds (Figure 5C). This is consistent with peak excitatory drive normalized to cellular excitability, as reported previously (Menelaou et al., 2022).

Next, clamping V2a neurons at a potential to isolate inhibitory post-synaptic currents (IPSCs) revealed faster time course compound IPSCs superimposed on a slower, tonic outward current (Figure 5D, E). In all V2a types tonic inhibition persisted the entirety of the episode and compound IPSCs were larger at higher frequencies near the beginning (Figure 5D, E). Like excitation, plots of maximum current (tonic current + compound IPSCs) revealed a significant negative correlation between inhibitory current and input resistance (Figure 5F), meaning that peak inhibitory drive is also normalized to cellular excitability. Given these large differences in maximum current values, to compare frequency-dependent changes in excitation and inhibition, we normalized the currents for each cycle to the maximum currents for the neuron (Figure 5C, F). So, maximum EPSCs or IPSCs in a given cycle are expressed as a percentage of the maximum value in all cycles at any speed.

For fast V2a-D and V2a-B neurons, maximum excitatory drive increased with increasing frequency (Figure 5G, left). Differences in maximum excitation were significant for both types when data were pooled (Figure 5H), consistent with their recruitment patterns (Figure 4F). Inhibition also followed the same pattern for fast V2a neurons (Figure 5I, left), significantly increasing at fast speeds (Figure 5J). For slow V2a-D neurons excitation was significantly lower at higher frequencies (Figure 5G, top right; Figure 5H, top) and inhibition was significantly higher (Figure 5I, top right; Figure 5J, top). However, for slow V2a-B neurons neither excitation (Figure 5G, bottom right; Figure 5H, bottom) nor inhibition (Figure 5I, bottom right; Figure 5J, bottom) changed significantly when pooled, despite the decreases in excitation (n = 4 of 6) and increases in inhibition (n = 5 of 6) in the majority of neurons. These patterns of excitation and inhibition are consistent with recruitment patterns, where slow V2a-D neurons are silenced at fast speeds and slow V2a-B neurons are recruited albeit with lower reliability (Figure 4F).

While the increase in inhibition at fast speeds to slow V2a neurons makes sense in terms of their disengagement during fast anguilliform mode (McLean et al., 2008), the increase in inhibition to fast V2a neurons clearly does not have a similar impact, despite being normalized to excitability. Indeed, when we compared the maximum excitation and inhibition to V2a-Ds (Figure 5K) and V2a-Bs (Figure 5L), maximum inhibition was orders of magnitude larger than maximum excitation. Even when controlling for differences in driving forces experienced at resting potentials, it would be difficult for V2a neurons to fire if excitation and inhibition arrived concurrently.

Thus, measurements of maximum levels of excitatory and inhibitory drive that include fast phasic and slower tonic components are consistent with recruitment patterns. These data also suggest the timing of inhibition plays a critical role in shaping the rhythmic output of V2a neurons more broadly.

To compare the timing of compound EPSCs and IPSCs we normalized the amplitude to the maximum (1) and minimum (0) values for that cycle and normalized the timing to the phase of the swim cycle, with the centers of two successive motor bursts demarcating phases 0 and 1 (Figure 6A). To simplify comparisons between peaks in excitation and troughs in inhibition, we inverted the sign of excitation. By convention, synaptic currents arriving between 0.75 and 0.25 are described as ‘in-phase’ with local motor output, while currents arriving 0.25 and 0.75 are described as ‘anti-phase’. In frog tadpoles and adult lampreys, in-phase currents are assumed to arise from recurrent sources on the ipsilateral side, while anti-phase currents arise from the reciprocal sources on the contralateral side (Soffe and Roberts, 1982; Russell and Wallén, 1983). This means that differences in the timing of peaks of excitation and inhibition can be used to draw inferences about patterns of connectivity along and across the body, with phase-advanced inputs from more rostral regions and phase-delayed inputs from more caudal ones.

First, to assess differences in the timing of compound EPSCs and IPSCs, we low pass filtered the raw current trace to identify peaks in- and anti-phase (Figure 6A, black dots), which were then plotted on a circular axis and compared to cell-attached spiking activity (Figure 6B–D). For all V2a neurons, spikes were exclusively distributed in-phase with local motor output (Figure 6B), with largely overlapping distributions that were broader for fast V2a-Ds (Figure 6B, top) and slow V2a-Bs (Figure 6B, bottom), consistent with spike distributions observed in real time (Figure 4D). However, there were type-specific differences in the distributions of peak excitation and inhibition.

For both fast and slow V2a-Ds phasic excitation began early in anti-phase, while the distribution of phasic excitation in fast and slow V2a-Bs was more coherent and trailed that of V2a-Ds (Figure 6C). This is consistent with the proposed hierarchical organization of V2a neurons, with V2a-Ds occupying a position upstream of V2a-Bs, which is not as obvious based purely on spike distributions (Figures 4D and 6B). For fast V2a-Ds, phasic inhibition was also broadly distributed, with a dip that coincided with the peak distributions of both excitation and spiking activity (Figure 6D, top). For slow V2a-Ds, there were two clear peaks in phasic inhibition, one immediately trailing in-phase excitation (Figure 6D, bottom arrow) and one peaking anti-phase. Remarkably, for V2a-B neurons, we observed the opposite pattern for phasic inhibition. Fast V2a-Bs received peaks in- and anti-phase (Figure 6D, top arrow), while slow V2a-Bs received predominantly anti-phase inhibition (Figure 6D, bottom). Notably, the presence of in-phase inhibition was also associated with a more coherent distribution of spikes in slow V2a-Ds and fast V2a-Bs (cf., Figure 6B).

Next, to get a better sense of how patterns of phasic excitation and inhibition ultimately shape phasic firing patterns within different types at different speeds, we binned the data by frequency as above and examined averages of normalized raw current traces. Among the V2a-Ds, fast subtypes exhibited reciprocal inhibition at all frequencies, with phasic peaks in excitation and inhibition most obvious at higher frequencies coincident with spiking activity (Figure 6E, top). Among slow V2a-Ds, coherent peaks in reciprocal and recurrent inhibition were most obvious at slow speeds coincident with phasic excitation and spiking behavior (Figure 6E, bottom). Among the V2a-Bs, patterns of reciprocal and recurrent inhibition were consistent across all speed bins, with peaks in phasic excitation and dips in phasic inhibition coincident with spiking activity (Figure 6F).

Thus, despite differences in cellular rhythmogenic capacity, these observations suggest all types of V2a neurons rely on phasic excitation and inhibition to sculpt their rhythmic firing behavior during swimming. Peaks in phasic excitation arrive earlier among V2a-D neurons consistent with their hierarchical positioning upstream from V2a-B neurons, and the relative coherence of phasic excitatory and inhibitory drive at different frequencies is consistent with speed modularity in this population. Among V2a-B neurons, the coherence of phasic excitatory and inhibitory drive over the entire frequency range is consistent with a role integrating inputs from the different speed modules. Critically, these data also suggest reciprocal inhibition plays a key role in rhythmogenesis at fast speeds, while recurrent inhibition plays a more dominant role in rhythmogenesis at slow speeds.

Our data thus far suggest that reciprocal inhibition plays a key role in rhythmogenesis at fast speeds via V2a-D neurons, while it plays a more dominant role patterning output at slow speeds via V2a-B neurons. To test this idea, we used a genetic approach to synaptically silence commissural, glycinergic Dmrt3a-labeled dI6 neurons (Figure 7A, B), which participate in swimming over a range of speeds in larvae (Kishore et al., 2020; Satou et al., 2020). Specifically, we crossed compound Tg[Dmrt3a:Gal4;UAS:pTagRFP] fish (Satou et al., 2013; Kishore et al., 2020) with Tg[UAS:BoTxBLC-GFP] fish (Sternberg et al., 2016), to express botulinum toxin (light chain) fused with GFP (BoTx-GFP) selectively in dI6 neurons and monitored the impact on locomotor rhythmogenesis using contralateral ventral rootlet recordings. As controls, we performed recordings from heterozygous sibling larvae that lacked expression of BoTx-GFP.

First, to confirm the efficacy of synaptic silencing, we analyzed the distribution of dI6 neurons expressing GFP-tagged BoTx compared to the entire population of BoTx-negative dI6 neurons labeled by pTagRFP (Figure 7C). BoTx-GFP-labeled neurons were observed throughout the DV range of the dI6 population (Figure 7C, left and middle), with a significantly higher proportion observed more ventrally (Figure 7C, middle and right). Since dI6 neurons located more ventrally participate in faster swimming movements (Kishore et al., 2020), we expected the impact of synaptic silencing would be more obvious at higher frequencies.

Consistent with this idea, in dI6 BoTx-GFP fish we frequently observed synchronous bilateral activity at the beginning of swim episodes when frequency is typically highest (cf., Figure 1G), which could transition to rhythmic, alternating activity nearer the end of the episode (Figure 7D). In control larvae, left–right phase values in the first and subsequent swim episodes had peak distributions that rarely deviated from 0.5, as you would expect for perfect alternation during locomotor rhythmogenesis (Figure 7E, left). On the other hand, for BoTx-GFP fish, left–right phase values were more broadly distributed, especially at the beginning of the episode, which was more obvious in the initial swim episode compared to subsequent ones (Figure 7E, right). This is also consistent with differences in peak frequencies observed near the beginning of swim episodes, where they are always higher in the episode immediately following the stimulation (cf., Figure 1G).

The lack of coordinated rhythmogenic output across the body at fast speeds was also apparent when comparing ventral rootlet burst frequencies. In dI6-GFP fish, higher-frequency activity overlapping higher-frequency activity in controls was observed at the beginning of swim episodes (Figure 7D, black dots), however this was rarely associated with higher-frequency activity in the corresponding cycle on the other side (Figure 7F). In controls, cyclical frequencies on the left and right sides were symmetrical, with most of the datapoints falling along unity (Figure 7F). Spinal motor neurons exhibit a range of rhythmogenic cellular properties themselves (Menelaou and McLean, 2012), which likely contribute to the higher-frequency bursts observed unilaterally. Thus, the disorganized unilateral bursting during bilaterally synchronous output at the onset of swimming in dI6 BoTx-GFP fish is consistent with the selective disruption of rhythmogenesis at fast speeds.

Finally, we turned our attention to lower-frequency swimming observed nearer the end of swimming, which was less impacted by BoTx expression. For purposes of analysis, we divided the dataset into the initial (Swim 1) and subsequent (Swim 2+) swim episodes, and focused on ventral rootlet burst frequencies below 35 Hz. A comparison of burst frequency and phase revealed no significant differences between dI6 BoTx-GFP and control fish, consistent with a lack of impact on timing and rhythm generation at low frequencies (Figure 7G, H). However, we did observe an increase in burst duration (Figure 7I), which increased the percentage of the swim cycle occupied by motor output, or duty cycle (Figure 7J). These effects on were most obvious in the initial swim episode, where duty cycles could reach 90% (Figure 7J). Normally, duty cycles must not exceed 50% of the cycle period to avoid synchronous bilateral contractions during swimming (Grillner, 1974). Thus, at slow speeds the dominant impact of synaptic silencing of dI6 neurons is on motor output, not rhythmogenesis. Collectively, these observations are consistent with the idea that reciprocal inhibition plays a more dominant role in rhythm generation at fast speeds and a more dominant role in pattern formation at slow speeds.

Tg[Chx10:GFP] fish (ZFIN ID: ZDB-ALT-061204-2) were used to target V2a neurons in the spinal cord (Kimura et al., 2006), and compound Tg[Dmrt3a:Gal4;UAS:pTagRFP] fish (ZFIN IDs: ZDB-FISH-150901-3557; ZDB-TGCONSTRUCT-210826-11) crossed with Tg[UAS:BoTxBLC-GFP] fish (ZFIN ID: ZDB-FISH-160816-31) were used to express botulinum-toxin light chain fused with GFP in commissural inhibitory dI6 interneurons (Satou et al., 2013; Sternberg et al., 2016; Kishore et al., 2020). Fish were maintained at 28.5°C on a 14-hr light/10-hr dark cycle in an aquatics facility staffed by Northwestern University’s Center for Comparative Medicine. All experiments were performed at room temperature between 4 and 5 days post fertilization (dpf) when zebrafish larvae are free swimming. Because 4–5 dpf larvae have not yet developed secondary sexual characteristics, their sex cannot be easily determined. Therefore, experiments were performed on larvae of either sex. All experimental protocols conform to National Institutes of Health guidelines on animal experimentation and were approved by the Northwestern University Institutional Animal Care and Use Committee (IS00019359 and IS00019319).

Electrophysiological recordings were acquired using a Multiclamp 700A amplifier (Molecular Devices), a Digidata series 1322A digitizer (Molecular Devices), and pClamp 10 software (Molecular Devices). The intrinsic firing patterns of V2a neurons were assessed by performing 500ms long current steps at 1–2x rheobase in current-clamp mode. To compare the intrinsic firing properties of V2a neurons to their recruitment profiles during ‘fictive’ escape swimming, we performed simultaneous whole-cell current-clamp recordings of V2a neurons and extracellular recordings of the peripheral ventral rootlet. Synaptic currents during fictive swimming were assessed in dual whole voltage-clamp and extracellular rootlet recordings. Swimming was evoked by applying a mild and brief electrical shock (2–10 V; 0.1–0.4 ms) using a bipolar tungsten electrode placed over the tail fin.

For recordings larvae were first paralyzed using alpha-bungarotoxin (1 mg/ml in extracellular solution containing 134 mM NaCl, 2.9 mM KCl, 1.2 mM MgCl₂, 10 mM 2-(4-(2-hydroxyethyl)-1-piperazinyl)-ethanesulfonic acid (HEPES), 10 mM glucose, and 2.1 mM CaCl₂, adjusted to pH 7.8 with NaOH), immersed in extracellular solution containing anesthetic (MS-222), and secured left side up on a Sylgard lined glass bottom dish using custom etched tungsten pins through the notochord. To record the activity of the peripheral ventral rootlets, the skin over the muscle tissue at the midbody was removed using fine forceps. For access to spinal V2a neurons, the muscle tissue over one spinal segment was dissected away using a tool fashioned from sharpened tungsten wire. After the dissection, the preparation was rinsed and immersed in anesthetic-free extracellular solution. For cell-attached and whole-cell recordings, standard-wall (1 mm outer diameter) borosilicate capillaries (Sutter Instrument) were pulled to resistances between 5 and 16 MΩ using a micropipette puller (Flaming/Brown; Sutter Instrument). For peripheral motor nerve recordings, electrodes were made by pulling a patch electrode, breaking the tip, and then heat polishing and bending it using a microforge (MF-830; Narishige) to offset the approach angle for better contact between the electrode and muscle tissue. Peripheral motor nerve recordings were made no more than two segments caudal to the site of cellular recording.

For current-clamp recordings, patch electrodes were filled with intracellular solution containing (125–130 mM K-gluconate, 2–4 mM MgCl₂, 0.2–10 mM ethylene glycol-bis(β-aminoethyl ether)-N,N,N′,N′-tetraacetic acid (EGTA), 10 mM HEPES, 4 mM Na₂ATP, adjusted to pH 7.3 with KOH). For voltage-clamp recordings, the intracellular solution consisted of 122 mM CsMeSO₃, 0.1–1 mM QX314-Cl, 1 mM tetraethylammonium chloride (TEA-Cl), 3 mM MgCl₂, 10 mM HEPES, 1 mM EGTA, and 4 mM Na₂-ATP. CsMeSO₃ and QX-314 Cl were included in the solution to block voltage-dependent channels. Alexa Fluor 568 hydrazide (50 μmol/l) was also added to the intracellular solution for the post hoc confirmation of cell identity. Only neurons whose morphologies could be positively identified were included in this study.

V2a neurons were targeted on an epifluorescent microscope (Axio Examiner) under a ×40/1.0 NA water-immersion objective (W Plan-Apochromat; Zeiss) using a CCD camera (Rolera-XR; QImaging). The electrode was positioned using a motorized micromanipulator (MP-225; Sutter Instrument or Patchstar, Scientifica), while positive pressure was maintained using a pneumatic transducer (Model DPM-1B; Bio-Tek Instruments) to prevent the tip from clogging as the neuron was approached. The neuron was held at −65 mV during GΩ seal formation. Once the seal was achieved, for voltage-clamp experiments cell-attached recordings were performed to obtain the recruitment patterns of V2a neurons during fictive escape swimming before breaking into the neuron to record synaptic currents. A 5-mV hyperpolarizing pulse was then administered at a holding potential of −65 mV, and the steady-state current injected during the pulse was used to compute input resistance according to Ohm’s law.

Excitatory currents were clamped at the reversal potential for inhibition calculated from our solutions, which is −75 mV and inhibitory currents were clamped at the reversal potential for excitation, which is +10 mV. These values were corrected for a calculated junction potential of −11 mV. Thus, neurons were alternately held at the corrected potentials of −64 mV for EPSCs and 21 mV for IPSCs. To monitor series resistance throughout the experiment, recordings were alternately performed without and with 60% series resistance compensation. Standard corrections for bridge balance and electrode capacitance were applied for current-clamp recordings. Whole-cell electrical signals were filtered at 10 kHz and digitized at 100 kHz at a gain of 10 (feedback resistor, 500 MΩ). Extracellular signals from the peripheral motor nerves were amplified at a gain of 1000 and digitized with low- and high-frequency cutoffs set at 300 and 5000 Hz, respectively.

Whole-cell and extracellular recordings were annotated in Dataview to label individual spikes, ventral rootlet bursts, and swimming episodes. The onset and duration of spikes, ventral root bursts and swimming episodes were then batch exported into text files, which were then read into MATLAB (Mathworks) for further analysis. Instantaneous spike frequencies are the inverse of the period between two successive action potentials. The recruitment patterns of V2a neurons were quantified by computing the probability of observing spikes for all cycles present within each 5 Hz ventral root burst frequency bin from 10 to 65 Hz. The recruitment probability was then plotted as a function of ventral root burst frequency. The V2a neurons with an upward sloping recruitment reliability plot were classified as fast and those with a downward sloping plot were categorized as slow (cf., trendlines in Figure 4E).

Maximum excitatory and inhibitory currents reflect the maximum current received at any ventral root burst frequency. To compare frequency-dependent changes in synaptic currents across neuron types, they were normalized against the maximum synaptic current the neuron received. Normalized currents were then averaged in four discrete frequency bins spanning carangiform and anguilliform swimming modes (15–25, 25–35, 35–45, and 45+ Hz). To analyze the phasic timing of spiking and synaptic currents, they were normalized to the period of the corresponding cycle, such that the centers of successive motor bursts marked phases 0 and 1. Synaptic currents arriving between 0.75 and 0.25 were considered in-phase and those between 0.25 and 0.75 considered anti-phase. To analyze the relative timing of excitation and inhibition, we normalized amplitudes to the maximum current received in each cycle, such that the final values of normalized current ranged between 0 and 1. Currents were then averaged for all cycles present within the four frequency bins.

To quantify peaks in synaptic currents corresponding to the phasic inputs received during a swim cycle, current traces were low pass filtered at the cutoff frequency of 1000 Hz using the butterworth (second order) and zero phase digital filters in MATLAB. The latter was used to prevent the raw current values from undergoing a phase shift. Peaks were detected using the findpeaks() function in MATLAB. Counts of V2a spikes, EPSC’s and IPSC’s were plotted on a circular axis from phase 0 to 1. Data visualization and statistical analysis were performed in R.

To quantify left–right alternation in control and botulinum-toxin expressing dI6 neurons, dual ventral rootlet recordings were rectified and annotated in DataView. Owing to the highly irregular nature of the ventral rootlet activity in the dI6-silenced fish, automatic event detection parameters used to distinguish motor bursts in control conditions were unable to accurately demarcate the boundaries of motor bursts in BoTx-GFP fish. Therefore, all annotations for controls and BoTx-GFP were performed manually. Ventral root bursts at frequencies above a 100 Hz were excluded from analysis since this is the upper bound of their natural swimming range (Müller and van Leeuwen, 2004). Stimulus, burst, and bout onset and offset times for all recordings were batch exported into text files. Custom written MATLAB code was then used to classify Swim 1 and Swim 2+ episodes and identify corresponding bursts in the dual ventral rootlet recordings to compute phase values. Phase values were calculated from the center of the ventral rootlet motor burst.

After each electrophysiological recording, a series of epifluorescent and differential interference contrast images were captured using a CCD camera (Rolera-XR) and QCapture Suite imaging software (QImaging) to track the axonal trajectories of V2a neurons and confirm their morphology. Images were later analyzed in ImageJ. Confocal z-stacks of Tg[Chx10:GFP] and Tg[Dmrt3a:GFP] fish were captured on using a Zeiss LSM710 microscope using established protocols (Menelaou et al., 2014).

Soma size and DV locations of individually labeled V2a neurons were measured in ImageJ. DV and ML positions of V2a neurons in Tg[Chx10:GFP] were computed by obtaining the relevant positional coordinates from confocal images in Imaris (Oxford Instruments). Briefly, the spot detection tool was used to obtain the coordinates of the centers of all somata and orthogonal slicers were used in conjunction with measurement points to extract the coordinates of the DV and lateral boundaries of the spinal cord as well as the central canal. The positions of the somata were then, respectively, normalized against the positions of the DV and ML boundaries to get the final DV and ML positions. V2a somata were classified into two groups based on the intensity of GFP expression. The GFP fluorescence intensity of V2a somata was exported from Imaris, and its density distribution was plotted in R. Because the density distribution of GFP fluorescence intensity was bimodal, the minimum between the two modes of the distribution was used as a cutoff to split V2a neurons into classes of low and high GFP expression. DV positions of dI6 neurons in compound Tg[Dmrt3a:Gal4;UAS:pTagRFP] × Tg[UAS:BoTxBLC-GFP] larvae were calculated using the relevant positional coordinates from confocal images in Imaris and normalizing them to the dorsal and ventral boundaries of spinal cord determined using differential contrast images.

To illustrate the differences in body shape and frequency with time, we include a video of a freely swimming 5-day-old larva captured at 2000 frames/s using a FASTCAM Mini AX50 (Photron), evoked using a sharpened tungsten pin to stimulate somatosensory afferents (Video 1).

All data were first tested for normality to select the appropriate statistical test. None of the data were normally distributed, so nonparametric tests were used to assess statistical significance. The difference between the density distributions of ML and DV somatic positions was tested using the two-sample Kolmogorov–Smirnov test. Correlations were measured using Spearman’s rank correlation test. Comparisons between two groups of paired or unpaired data were performed using the Wilcoxon signed-rank test and the Wilcoxon rank-sum test, respectively. For data comprised of more than two groups, differences were assessed using the Kruskal–Wallis test, followed by pairwise comparisons using Dunn’s test with a post hoc Benjamini–Hochberg false-discovery rate correction for multiple comparisons. For circular plots, Watson’s two-sample test of homogeneity was used to determine whether the spike and post-synaptic current distributions in two conditions were significantly different from each other. Statistically significant differences are denoted as follows: *p < 0.05; **p < 0.01; ***p < 0.001. Note that it was not always possible to collect cell-attached data and voltage-clamp recordings of excitation and inhibition in the same cells, so n numbers reported in the text can vary between conditions. The same holds true for measured parameters in current-clamp recordings, where soma size, DV position, input resistance, rheobase, spike rates, and recruitment order in the same neurons were not always possible.