| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
|
|
|
|||||||
|
|||||||||
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
1Institute for Neuroinformatics University of Zurich/Swiss Federal Institute of Technology, Zurich, Switzerland; 2Bell Labs, Lucent Technologies, Murray Hill, New Jersey; and 3McGovern Institute for Brain Research, Massachusetts Institute of Technology, Cambridge, Massachusetts
Submitted 11 October 2005; accepted in final form 15 February 2006
 |
ABSTRACT |
|---|
|
|
|
INTRODUCTION |
|---|
|
; Louie 2001; Nadasdy et al. 1999; Wilson and McNaughton 1994). As sleep-related spike trains are very complex, their generation is believed to involve many neuron types and synaptic mechanisms. The songbird robust nucleus of the arcopallium RA is a premotor area (Nottebohm et al. 1976) in which replayed premotor sequences have recently been found (Dave and Margoliash 2000). RA receives motor-related projections from the higher premotor brain area HVC (Nottebohm et al. 1982; Vicario and Nottebohm 1988). During sleep, HVC exhibits sparse burst sequences that are locked to the RA burst sequences (Hahnloser et al. 2002), suggesting an important role of the premotor pathway for RA sleep burst generation.
A separate group of song-related brain areas known as the anterior forebrain pathway (AFP) indirectly connects HVC to RA (Okuhata and Saito 1987). Bilateral lesions in areas of the AFP produce profound deficits in vocal learning in juveniles but have little effect on vocal production in adults (Bottjer et al. 1984; Doupe 1993; Scharff and Nottebohm 1991; Sohrabji et al. 1990). Given that part of the song learning process might occur during sleep (Dave and Margoliash 2000; Deregnaucourt et al. 2005; Margoliash 2005), it is possible that the output of the AFP, the lateral magnocellular nucleus of the anterior nidopallium (LMAN), might play an important role for sleep burst generation in RA (Abarbanel et al. 2004).
Spiking activity in the premotor pathway and the AFP has been recently studied in an anesthetized preparation (Kimpo et al. 2003). Paired recordings in LMAN and RA showed that spontaneous bursts in RA are preceded either by bursts in HVC or in LMAN. The propagation of correlated bursts from HVC to LMAN may reflect information processing relevant to consolidation of learned vocal patterns. It was suggested that changes in the motor circuit might be controlled during sleep by LMAN-mediated synaptic plasticity in RA (Dave and Margoliash 2000; Doupe et al. 2004).
Here we describe single-neuron recordings of identified neuron pairs in the HVC, RA, and LMAN circuits in a sleeping bird. In combination with reversible pharmacological inactivation of HVC and LMAN, our data suggest that bursts in RA are driven by the synaptic input from HVC and are not driven by input from LMAN.
Part of our data have already been published previously (Fee et al. 2004; Hahnloser et al. 2002). The data and analysis reported here are more extensive with additional recordings from neurons in the AFP, i.e., from HVC neurons projecting to Area X and from LMAN neurons projecting to RA. In our previous study, we analyzed the correlations of burst events in different neuron pairs with a novel conditional correlation technique. Here we are not interested in bursts as much. We analyze spike train correlations by computing coherency functions, which is a useful technique to discount for correlations arising from bursting tendencies of neurons (Kimpo et al. 2003; Thomson and Chave 1991). This different analysis technique, in addition to confirming our previous results, reveals some interesting new features of our data. We present a simple model for these features, explaining the multiple coherency peaks we observe by sequential activity in HVC.
|
|
METHODS |
|---|
|
ANIMALS.
Zebra finches (Taeniopygia guttata) were obtained from commercial suppliers (Old Bridge, NJ and Animal Diffusion, Villarimboud, Switzerland). Animals were maintained on a day-night reversed 12-h light cycle to assist in obtaining sleep during daytime experimental sessions. Surgery began approximately at the onset of the night cycle. Birds were anesthetized with 1–3% isoflurane in oxygen and placed in a stereotaxic apparatus (Stellar, Stoelting). The flat anterior portion of the skull was set at an angle of 65° from horizontal and a thin stainless steel plate was mounted to the skull with dental acrylic (Lang Jet Acrylic). A window was made in the inner and outer bone leaflet over nuclei RA, HVC, and Area X (or RA and LMAN) of the right hemisphere. A small hole (
200 µm) was made in the dura over each area and the uncovered brain was protected with 2% low-melting agarose (Sigma). Wound margins were treated with lidocaine gel. The animal was placed in a small foam restraint and placed in the recording apparatus without further anesthesia. In most experiments, the animal was given a single dose (1–10 µg) of melatonin (Sigma) right after surgery, injected subcutaneously in phosphate-buffered saline. Melatonin is known to promote sleep in birds (Hishikawa et al. 1969; Phillips and Berger 1992). In one bird, we used 1 µg and then 2-µg doses for paired recordings of electroencephalographic (EEG) signals and RA neurons, but then switched to 10-µg doses for all subsequent experiments. The only birds to which we did not administer melatonin were two birds used to produce Fig. 2C and two birds in which we recorded RA neuron pairs and RA-HVCI neuron pairs. Because the data from the latter two birds were similar to the melatonin-related data, we included them in Fig. 8, B and E. Data were obtained from a total of 50 zebra finches.
|
|
2 mm lateral and 2 mm anterior of the bifurcation of the sagittal sinus (lambda). Signals were amplified (G = 1,000) and filtered (HP > 1 Hz, LP < 100 Hz, DAM-80, WPI).
SINGLE-UNIT RECORDINGS.
Simultaneous paired and triplet single-unit recordings were made in nucleus RA, HVC, and LMAN. High signal-to-noise (>10:1) recordings were made using sharp glass microelectrodes (5–15 M
, borosilicate, 1.0 mm OD, 0.7 mm ID) pulled on a vertical electrode puller (Model 730, Kopf Instruments) and filled with 2 M NaCl. Signals were amplified using Neurodata IR285 (Cygnus Instruments) or Axoclamp-2B (Molecular Devices) intracellular amplifiers. Custom electronics were used for additional amplification (gain of 100) and filtering (300-Hz high-pass 5-pole Bessel filter, 10-kHz low-pass constant delay filter). Extracellular signals were digitized to 16-bit precision at a sampling rate of 20 kHz and stored in blocks of 200 s on a Pentium-based PC running custom Labview software (National Instruments). For triplet recordings, spike signals recorded on a tungsten electrode in RA were amplified using a DAM-80 extracellular amplifier (World Precision Instruments). Single-unit spike trains were confirmed off-line by complete suppression of the spike autocorrelation functions at time lags <1 ms.
ANTIDROMIC ACTIVATION. Bipolar stimulating electrodes were placed in RA and/or Area X for antidromic identification of RA-projecting and Area X-projecting HVC neurons, as well as of RA-projecting LMAN neurons. Figure 1. Stimulation electrodes consisted of a pair of Teflon-insulated 50-µm-diam stainless steel wires (California Fine Wire) spaced 0.5 mm apart. For RA stimulation, one electrode was positioned within RA and the other electrode positioned more dorsally. For Area X stimulation, both electrodes were placed within Area X. Electrical stimulation was produced using an isolated stimulation unit (AMPI) with intensities in the range 50–500 µA. For most experiments, single monophasic pulses of 0.2-ms duration were used.
|
For some neurons, collision tests were done by stimulating in RA or Area X with some delay after a spontaneous spike. For putative projection neurons, stimulating at small delays (1–3 ms) in the target area resulted in 100% failure of the antidromic spike (collisions). For interneurons, stimulating at small delays did not result in antidromic spike failure (n = 10). When putative interneurons were stimulated at delays of 1–2 ms after spontaneous spikes, both average latency and latency variability decreased by 10–15% compared with that resulting from stimulation not triggered on spikes. Hence only a small fraction of the latency variability of putative interneurons can be attributed to activity-related changes in axon conduction velocity (Swadlow 1998).
All RA neurons analyzed in this study are from putative RA projection neurons even though we did not identify RA projection neurons using antidromic stimulation. We believe to have reliably distinguished RA projection neurons from other RA neuron types based on firing statistics and spike waveforms. In comparison to RA projection neurons, RA interneurons exhibit an absence of a regular firing mode in the awake bird, they have unusually high average burst rates in the sleeping bird (>5 bursts/s), and their spike waveforms are significantly narrower (Leonardo and Fee 2005). Note: we frequently recorded from RA neurons with 3–5 M tungsten electrodes, permitting us to clearly observe the narrower spike waveforms in putative RA interneurons. Due to their very infrequent occurrence, RA interneurons were discarded from analysis in this study.
REVERSIBLE LESIONS IN HVC AND LMAN. The effect of lidocaine injections in HVC and LMAN on sleep burst rate in RA was examined by micro-injection of lidocaine (2% in phosphate-buffered saline) during single-unit recordings in RA. Injections were made from pulled glass pipettes (50-µm tip size) using a pressure injection system (Pico Spritzer). Injected volumes were 10–100 nl. Control injections were made in HVC with phosphate-buffered saline. In the LMAN injection experiments, we tested three injection sites; one below, one inside, and one just above LMAN. The anatomical localization of LMAN was done by recording RA-triggered antidromic responses in LMAN prior to insertion of the injection pipette. In the HVC injection experiments, HVC location was verified by recording sleep bursts in HVC interneurons. To measure the effect of the injection in RA, we counted the number of bursts in RA neurons in the minute before and after the injection. Statistical significance of burst-rate differences before and after injection was assessed using the paired t-test with a significance level P < 0.01.
After each experiment, the animal was killed by intramuscular injection of 20% urethan or pentobarbital sodium (Nembutal). The brain was removed for histological examination of unstained slices to verify the location of stimulating and recording electrodes, and drug injection sites. All experiments were carried out in accord with protocols approved by the local IACUC and the Veterinary Office of the Canton of Zurich, Switzerland.
Data analysis
INSTANTANEOUS FIRING RATE.
Because of the large dynamic range of firing rates in RA and HVC, spike rasters are not well suited for visualizing long traces of neural activities. Therefore we represented the activity at time t of a neuron by the instantaneous firing rate R(t), a continuous function defined by the inverse of the closest inter-spike interval
 |
BURST RATE. We identified a group of at least two spikes as a burst if the instantaneous firing rate continuously exceeded 100 Hz. The burst rate was defined as the number of events per unit time in which the instantaneous firing rate crossed the 100-Hz barrier from below.
FIRING RATE IN BURST TRAINS. In Fig. 6, we computed firing rates in burst trains. These spike trains were constructed by removing all single spikes not part of a burst, i.e., all spikes that did not form an interspike interval of <100 ms with another spike. Removing single spikes was particularly effective for RA and HVCI neurons, which produced many single spikes in the awake state and during sleep (Figs. 2 and 4–6).
|
|
|
fA
/fB of average firing rate fA in neuron A and average firing rate fB in neuron B (averaged over all 3-s windows), reported in the histograms in Fig. 6C. Neuron B was always chosen to be the neuron type firing fewer spikes on average. That is, for HVCRA-RA pairs and HVCRA-HVCI pairs, neuron B is the HVCRA neuron; for RA-HVCI pairs, neuron B is the RA neuron, and for HVCX-HVCI pairs, neuron B is the HVCX neuron. Again firing-rate ratios were plotted as 
when the firing-rate correlations in Fig. 6B were significant, and they were plotted as 
when firing-rate correlations were not significant. Due to the large spread and distant outliers of measured ratios, we summarized the histograms of firing-rate ratios by the median value rm instead of the mean value (including significantly and nonsignificantly correlated pairs). By performing the same analysis but without removing single spikes, we observed much smaller firing-rate fluctuations in RA and HVCI neurons due to their tendency to produce regular spike trains. In general, not removing single spikes in Fig. 6 led to a decreased number of correlated neuron pairs and to a lower degree of correlation. Not removing single spikes in HVCRA and HVCX neurons also led to slightly decreased correlations.
SPIKE WAVEFORM CLASSIFICATION. For the HVC spike waveform classification, only waveforms of equal polarity (1st negative, then positive), very large signal-to-noise ratio and small SD were selected (interestingly, for all neurons types, >10% of the recorded spike waveforms had opposite polarity, going positive first). The waveforms shown in Fig. 3E each corresponds to the averages of 25 spontaneous spike waveforms. Only waveforms of single spikes were used to avoid the additional waveform variability associated with bursts. Each of the 25 examples was interpolated at 5-µs intervals using the fast Fourier transform, mean–subtracted, normalized to one at the peak, and aligned in time. Spike width was measured at 25% of the peak amplitude.
|
INTERSPIKE INTERVAL (ISI) PROBABILITY DENSITY FUNCTION (PDF).
We characterized the spike train of a neuron (labeled A) by the ISI pdf hA(
) (here stands for the ISI). First we computed the ISI histogram with bin centers i chosen on a logarithmic scale (i = 1,..., 100). We then derived the ISI pdf hA(i) simply by dividing the ISI histogram by the total area under the histogram, such that 
. In essence, the ISI pdf is a normalized ISI histogram. The fact that the ISI pdf sums to one is useful when we average over neurons and when we test spike trains for renewal statistics by means of autocovariance functions.
AUTO-COVARIANCE FUNCTIONS.
The auto-covariance function CAA(t) of a spike train 
A(t) (modeled as a sum of delta functions) is a measure of spike rate fluctuations. It is defined as
 | (1) |
A is the mean firing rate of neuron A over the entire recording, and T is the duration of the recording (the denominator T – |t| results in an unbiased estimation of auto-covariance at large times t after setting A(t) = 0 for t < 0 and t > T). The first term on the right side of Eq. 1 is known as the auto-correlation function. By subtracting the term A2 from the auto-correlation function, the asymptotic value of the auto-covariance function is zero. For better visibility, results in the figures are plotted on a logarithmic scale ti by integration between adjacent points, i.e., shown is the discrete function 
In Figs. 5F and 11D, we have plotted the average auto-covariance functions CAA(t)A by averaging over all neurons A of the same type.
|
). This test assumes that interspike intervals occur in completely random order and with a frequency set solely by the ISI pdf. Spike trains often do not form renewal processes, which may benefit their coding abilities (Holden 2004). To perform renewal tests on our spike trains, we computed renewal auto-covariance function CAAR(t) for each neuron. In the frequency domain, the renewal auto-covariance function CAAR(
) can be derived from the ISI pdf hA()
 | (2) |
) is the Fourier transformed ISI pdf and Re denotes the real part of a complex number. For a derivation of Eq. 2 see e.g., (Gerstner and Kistler 2002). Any deviation of CAA(t) from CAAR(t) is indicative of a memory in the spike train that causes the occurrence of a particular ISI to be conditional on previous ISIs. We tested differences between the two functions for significance by the Jackknife variance estimator (for details on the Jackknife, see SIGNIFICANCE OF COHERENCY FUNCTIONS). Average renewal functions CAAR(t)A were computed by averaging over individual renewal functions and were plotted as dashed lines in Figs. 5F and 11D.
COHERENCY ANALYSIS.
In addition to slow correlations of firing rates analyzed in 3-s windows, we quantified fast correlations between spike trains in simultaneously recorded cells by the coherency function 
AB(t). The coherency function is similar to the cross-covariance function; we computed it as follows. First we computed the cross-covariance function CAB(t) between spike trains A(t) and B(t) (each modeled as a sum of delta functions)
 | (3) |
AB() in the frequency domain was computed by Fourier transformation and normalization
 | (4) |
AB(t) in Figs. 7–9, 12, and 13 were all plotted in the time domain, by inverse Fourier transformation of AB(). Smoothing of coherency functions AB(t) was done by convolution with a Gaussian windowing function of SD of 4 ms. Note that coherency functions were computed at the temporal resolution of data samples, which is 50 µs. However, in the figures, coherency functions were plotted at a resolution of 1 ms, which was achieved by summing over coherencies in 1-ms bins. Deliberately, to obtain precise estimates of peak coherency time lags, we chose a smaller time bin for our analysis than the 10-ms bin in Kimpo et al. Note: when we computed coherency functions at a temporal resolution of 10 ms, we found qualitatively similar results as shown in Figs. 7–9, 12, and 13.
|
|
|
|
AB(t)AB for the different neuron types in Figs. 8, 9, B and D, and 12, B and C, where the average runs over all neuron pairs of the same types. Plotted in these figures is also the SD of individual coherency functions and the average significance threshold.
SIGNIFICANCE OF COHERENCY FUNCTIONS.
The significance threshold of the coherency function for a particular neuron pair was assessed by Jackknifing the data in 20 s data windows and computing the Jackknife variance (Thomson and Chave 1991)
 |
ABi(t) is the coherency function of the jackknifed data with the ith window missing, and 
AB(t) = ABi(t)i is the average jackknifed coherency function. Note that we used long 20-s windows because of the small firing rates in HVC projection neurons and because all our recordings were >200 s (typical 400–600 s). We set the significance threshold of measured coherencies to three SDs, 
, shown by dashed lines in Figs. 7, A–C, and 9, A and C. This significance threshold corresponds to a roughly 99% confidence in significance testing. Whenever the coherency function exceeded the significance threshold, a peak emerged. The distributions of peak coherency amplitudes versus peak latencies are shown in the summary plots beneath the average coherency functions in Figs. 8, 9, B and D, and 12, B and C. In these figures, we plotted the average significance thresholds of all neuron pairs of the same types as dashed lines.
NOTE ON COHERENCY ANALYSIS.
Even though our confidence threshold is high (99%) and the jackknife is a highly conservative estimator (Thomson and Chave 1991), we observed many coherency peaks at large time lags (t > 100 ms). Presumably, these peaks arose from correlated fluctuations in firing rates as seen in most neuron pairs (see e.g., Fig. 6). In general terms, whenever data derive from experiments using externally controlled trials, it is possible to account for stimulus-driven fluctuations of firing rates by replacing the average firing rates A and B in Eq. 3 by time-dependent trial averages. However, as there are no trials in our experiments, and as we did not have any control over nonstationarities of sleep-related spike trains, there is no simple way of separating correlated fluctuations on a long time scale from fast correlations on a shorter time scale. Our simple solution to this problem was thus to plot all significant coherency peaks in the bottom parts of Figs. 8, 9, B and D and 12, B and C. The largest peaks were plotted as 
, and secondary peaks were plotted as 
. Most often, the largest peaks were found close to zero time lag, whereas the secondary peaks were found at larger positive or negative time lags. This behavior was clearly visible for coherency functions involving HVCI neurons such as in Fig. 8, D and E.
We realize that our choice of using coherency functions as our main analysis tool is a tradeoff: coherencies have the advantage of being a proven standard and allowing for comparison with previous studies but might not be the optimal measure when data are nonstationary and firing rates in different neuron types highly variable.
SINGLE SPIKES AND RASTER PLOTS.
In our previous report (Hahnloser et al. 2002), we removed single spikes from RA and HVCI neuron spike trains for the raster plots and for the correlation analysis (single spikes are those spikes not part of a burst as defined in the preceding text). Here we included all spikes of all neurons; therefore raster plots as in Fig. 7A look qualitatively different from those in Fig. 3 of Hahnloser et al. (2002).
MARKOV MODEL OF SERIAL CORRELATIONS IN HVC, RA, AND LMAN. The Markov model of HVC, RA, and LMAN spike trains is based on sequential dynamics in two populations of HVCRA and HVCX neurons. The HVC sleep-burst epochs in the HVCRA-HVCX populations were modeled as follows: if St is the event that the sleep-burst epoch reaches time t, then the probability that it also reaches time t + 40 ms is given by P(St+40ms|St) = P(HVCX) = 0.8. It follows that since its initialization at time t, the HVC sequence progresses for n steps up to time t + 40n ms with probability 0.8n.
We modeled bursts in HVCI, HVCRA, RA, and LMANRA neurons at fixed delays of these 40-ms intervals. To translate burst times into spike trains (Fig. 13C), we added spikes to the burst onset times as follows: for each burst, first we randomly drew the number n of spikes from an exponential distribution (rounding the result toward infinity). The mean number of spikes per burst was three for the HVC neuron and the RA neuron, whereas it was two for the LMANRA neuron. For each burst, the interspike intervals si were determined by si = 1 + 0.5ri2 ms (i = 1,..., n), where ri is a random variable drawn from a Gaussian distribution with zero mean and 1 ms2 variance. The coherency analysis for the artificial spike trains was performed using the same Matlab scripts (Mathworks) as for the natural spike trains except that final coherency functions in Fig. 13 were smoothed with a Gaussian window of 10 ms (instead of 4 ms).
|
|
RESULTS |
|---|
|
Relationship between sleep bursts and EEG
Sleep and wakefulness in freely behaving birds can be quantitatively distinguished on the basis of spectral content of the EEG at low frequencies (1–30 Hz) (Amlaner and Ball 1994; Nick and Konishi 2001). We found that in the awake head-fixed bird, EEGs consist of a broadband noisy signal falling off at higher frequencies as f–2 (where f is the frequency, Fig. 2B). During sleep, however, the EEG signal made large low-frequency excursions, corresponding to roughly a factor of 2 more power in a band from 1 to 15 Hz; >20 Hz, the awake and sleeping EEGs were nearly identical, Fig. 2B. These spectral characteristics of head-fixed zebra finches are in agreement with earlier more exhaustive EEG measurements of slow-wave or quiet sleep in the freely behaving pigeon (Tobler and Borbely 1988). We have quantified the spectral differences between sleep and wakefulness by the ratio of energy in the 5- to 15-Hz band to energy in the 45- to 55-Hz band; in the awake state, this ratio was typically 33 ± (SD) 14 and in the sleeping state the ratio was 63 ± 14 (from 72 1-min records in 3 birds). The sleep EEG state was usually associated behaviorally with closed eyes. However, in one of three birds, sleep-like EEG signals were occasionally observed with the ipsilateral eye open.
In the awake (nonsinging) bird, RA neurons have a regular and tonic activity of low firing rate (Yu and Margoliash 1996). In brain slices, they behave remarkably like integrate-and-fire neurons with no intrinsic bursting tendencies (Mooney 1992). During singing and sleeping, these neurons produce spontaneous bursts of spikes (Dave and Margoliash 2000; Hahnloser et al. 2002). We found that sleep EEGs were strongly associated with the presence of high-frequency spike bursts in RA neurons. In three birds, 35 RA neurons were recorded simultaneously with the EEG; sleep bursts in RA tended to cluster in 1- to 2-s epochs, each comprised of 2–10 bursts. Average RA burst rate in the sleep EEG state was 0.67 ± 0.46 bursts/s (Fig. 2C) and in the awake EEG state was <0.01 ± 0.01 bursts/s. Consistent with this observation, in chronic recordings from freely behaving zebra finches, bursts in RA neurons are never observed in awake, nonvocalizing animals (Leonardo and Fee 2005).
Because bursts in RA neurons occurred only during the sleep EEG state, we used the presence of bursts in RA as an operational assay of the sleep state. We did not find any differences in the EEG signal between normal and melatonin-induced sleep. In addition, there was no melatonin dependence of RA burst rates during epochs of sleep EEG state (Fig. 2C). However, after melatonin injection, we consistently observed a 50% increase in the average burst rate of RA neurons, suggesting that the fraction of time spent sleeping is increased by administering melatonin prior to the experiment.
Classification of HVC neurons
In the following, we provide a detailed electrophysiological analysis of different neuron types in HVC, RA, and LMAN. Our goal was to obtain new insights into their differential contributions to sleep bursts in RA and into the general physiology of sleep in these areas. We first describe results from single neuron recordings, then from recordings in pairs of RA and identified HVC neurons.
Three broad classes of neurons have been previously identified in HVC: RA-projecting neurons (HVCRA), Area X-projecting neurons (HVCX), and local interneurons (HVCI). These neuronal populations have distinctive morphological and electrophysiological properties (Dutar et al. 1998; Kubota and Taniguchi 1998). RA-projecting HVC neurons are excitatory and ramify widely in RA (Mooney 2000; Stark and Perkel 1999). Interneurons are aspiny (Dutar et al. 1998; Mooney 2000) and inhibitory (Rosen and Mooney 2003). Area X-projecting HVC neurons are spiny and excitatory, just like RA-projecting cells, but they have larger somas and dendritic extension than RA-projecting cells (Dutar et al. 1998).
To identify these neuron types in extracellular recordings, we antidromically activated HVC neurons by electrical stimulation in Area X and in RA. For the purpose of classifying antidromic responses of HVC neurons, high signal-to-noise extracellular recordings were made from 98 neurons in 11 awake and sleeping head-fixed zebra finches. In six birds, stimulating electrodes were placed either in RA (n = 3 birds) or in Area X (n = 3 birds), and in five birds stimulating electrodes were placed in both areas. We identified three distinct response patterns to RA and Area X stimulation (Fig. 3A). Putative Area X-projecting neurons (n = 35) responded to low-intensity stimulation (50–500 µA) in Area X but did not respond to even high-intensity stimulation (
10 mA) in RA. Near-threshold stimulation produced latencies to the first spike between 2 and 11 ms (average 4.9 ± 2.4 ms) with small latency variability in the range 23–94 µs (43 ± 12 µs, Fig. 3B). Putative RA-projecting neurons (n = 35) responded to low-intensity stimulation in RA, but did not respond to even high-intensity stimulation in Area X (10 mA). These neurons responded with only one or two spikes even at high stimulus intensity. Near-threshold stimulation produced latencies to the first spike of 2 to 8 ms (mean 4.6 ± 2.4 ms) and a small latency variability in the range 1–52 µs (34 ± 7 µs). Putative HVC interneurons (n = 28) responded to low-intensity stimulation in RA with latencies to the first spike of 4.4 ± 1.1 ms and large latency variability in the range 220–1530 µs (580 ± 120 µs, Fig. 3C). Near stimulus threshold these neurons responded with one or two spikes; higher intensities resulted in increased numbers of evoked spiked (in some cases, 10 spikes could be evoked). We found only three examples of putative HVC interneurons that responded to stimulation in Area X.
On the basis of these results, we infer that the putative projection neurons were antidromically activated from their target areas and thus are likely to be HVCX neurons and HVCRA neurons, respectively (Fig. 3D). In contrast, the nonantidromically activated neurons were likely to be synaptically activated, and we tentatively identify these as HVCI neurons.
We did not find significant differences in the extracellular spike waveforms of the three types of HVC neurons when recorded with sharp glass electrodes. Spike widths of HVCI neurons [0.21 ± 0.03 (SD) ms, range: 0.11–0.29 ms, n = 80] were slightly (but not significantly) smaller than those of HVCRA neurons (0.27 ± 0.04 ms, range: 0.15–0.38 ms, n = 46) and slightly (but not significantly) smaller than those of HVCX neurons (0.25 ± 0.05 ms, range: 0.16–0.38 ms, n = 42). The large within-class variability of spike waveforms made a correct classification of neuron type based on waveform difficult: a nearest waveform classifier (see METHODS) misclassified the neurons in 20% of cases. Also, none of the classes could be linearly separated from the other two classes with an error rate <20%. Linear reparability was based on the first three principal components of waveform shapes, which described >97% of waveform variability.
In the subsequent experiments, we identified each recorded HVC cell by its antidromic stimulation response, distinguishing HVCI neurons from HVC projection neurons by their large differences in spike-latency variability. Nonstimulated cells are not identifiable and thus were not recorded.
Firing characteristics of different HVC, RA, and LMAN neuron types
The three neuron types in HVC differed considerably in their spontaneous firing patterns in the awake bird. HVCX neurons produced single spikes at a rate of 1.5 ± 1.2 (SD) spike/s. HVCRA neurons were not spontaneously active in the awake bird (<0.1 spike/m). Only HVCI neurons exhibited a considerable spontaneous activity in the awake bird of 7.2 ± 7.0 spike/s, in the range of 2–26 spike/s, Fig. 4, A and B.
With the exception of HVCX neurons, the firing rates of all HVC neuron types increased significantly during sleep. The main characteristic of sleep-related spike trains in HVC and RA was the emergence of high-frequency bursts.
We characterized spike trains by their interspike-interval (ISI) histograms. To average the histograms for many neurons of the same type, we normalized the histograms by their sum before averaging, resulting in estimates of ISI probability density functions (pdfs), Fig. 4C (see METHODS). RA neurons exhibited a bimodal ISI pdf, with a relatively small peak at 4 ms and a larger peak at 50 ms. The small ISI peak is reminiscent of firing statistics of RA projection neurons during singing (Leonardo and Fee 2005). The latter peak is descriptive of ISIs in awake, nonsinging birds (the average RA neuron firing rate in the awake bird is 15–25 Hz). In combination, ISI pdfs of RA neurons suggest that sleep-related spike trains in RA neurons are a mixture of spike trains during waking and singing.
The average ISI pdf of HVCI neurons also exhibited two peaks, although the large ISI peak was smaller than the corresponding peak of RA neurons. This reduced peak is in agreement with a smaller average firing rate of HVCI neurons during waking compared with RA neurons.
The average ISI pdf of HVCRA neurons exhibited a narrow peak at 3 ms due to high-frequency bursts, and a very small and broad peak at 8 s, corresponding roughly to their typical interburst interval (see also Table 1). HVCX neurons also exhibited a burst peak at 3 ms and a broader peak at several hundreds of milliseconds. The latter peak occurred at ISIs one order of magnitude smaller than the second peak of HVCRA neurons. This reduced probability of HVCX neurons to generate very large ISIs reflects the tendency of HVCX neurons to produce more single spikes than do the HVCRA neurons (see also Table 1).
|
In the following, we characterize the burst generation mechanisms in the various neuron types and their tendencies to form sleep-burst epochs by close inspection of auto-covariance functions. The average auto-covariance functions of RA and HVCI neurons reflected sleep burst epochs by their long tails of over a second, Fig. 5F (top). Note, the oscillatory appearance for auto-covariance functions of RA neurons is due to the awake-like firing mode of these neurons in between burst epochs. Auto-covariance functions of HVC projection neurons did not have long tails and fell to zero within <50 ms, Fig. 5F (bottom), suggesting the absence of sleep-burst epochs in these neurons. Auto-covariance functions of individual HVC projection neurons were often oscillatory on a small time scale (Fig. 5F, bottom, insets), illustrating their stereotyped bursting behavior.
To learn more about burst generation mechanisms in the different neuron types, we performed renewal tests of their spike trains (Cox 1962). A renewal spike train is said to have no memory because each ISI is a random variable that is independent of previous ISIs. We computed auto-covariance functions for each spike train under renewal assumptions (see METHODS). Deviations from the true auto-covariance functions are indicated by thick horizontal bars in Fig. 5F. Of the 196 RA neurons tested, 20 exhibited nonrenewal auto-covariance functions (Jackknife, P < 0.001). Of the 160 HVCI neurons tested, 18 exhibited nonrenewal auto-covariance functions (Jackknife, P < 0.001). Of the 152 HVCRA neurons tested, 34 exhibited nonrenewal auto-covariance functions (Jackknife, P < 0.001). And of the 83 HVCX neurons tested, 3 exhibited nonrenewal auto-covariance functions (Jackknife, P < 0.001).
On a population level, most neuron types displayed nonrenewal statistics. The median auto-covariance function of RA neurons was larger than the median renewal function in the intervals 5–39 ms and 250 ms to 6.9 s (Mann-Whitney rank sum test, P < 0.001, Fig. 5F). The median auto-covariance function of HVCI neurons was larger than the median renewal function in the interval 5 ms to 2.9 s (Mann-Whitney rank sum test, P < 0.001, Fig. 5F). The median auto-covariance function of HVCRA neurons was smaller than the median renewal function in the interval 8–30 ms (Mann-Whitney rank sum test, P < 0.001, Fig. 5F). However, no where was the median auto-covariance function of HVCX neurons significantly different from the median renewal function (Mann-Whitney rank sum test, P < 0.001, Fig. 5F). In conclusion, whereas RA and HVCI neurons tended to have broader auto-covariance functions than expected by renewal statistics, those of HVCRA neurons tended to be narrower. This means that HVCRA neurons are subject to some sort of fatigue or adaptation that prevents them from producing long bursts, whereas RA and HVCI neurons are subject to transient reinforcement of increased spike rates. These observation agree well with in vitro findings of spike adaptation in HVCRA neurons but not in HVCI neurons (Dutar et al. 1998).
We visualized the fluctuations of firing rates during sleep-burst epochs by histograms of firing rates measured in 3-s windows, Fig. 6A. Typically, the histograms extended from large values down to zero firing rate, illustrating the large fluctuations in spike rates on this 3-s time scale. As this analysis does not distinguish between spikes fired in bursts and spikes fired in the awake-like regular mode and because the occurrence of spike bursts was the most prominent difference between the singing and waking states in RA and HVC neurons, we recalculated firing-rate histograms with restriction to burst-related firing only. To this end, we removed single spikes from the original spike trains and analyzed the resulting burst trains The firing-rate distributions of burst trains were almost as wide as those of the original spike trains, illustrating that most periods of spiking at high rates were associated with sleep bursts (Fig. 6A, ).
By comparing the original spike trains with burst trains, we determined the percentage of spikes that were part of a burst during sleep. This number was sleep-state dependent and was different for the different neuron types. For example, on average, for HVCRA neurons, >70% of spikes were part of a burst, whereas for RA neurons, this number was <20%. The average burst rate of HVCI neurons was >2 burst/s—during the short epochs of very dense bursting, this number could be as high as 10–25 burst/s. RA neurons had an average burst rate of almost 1 burst/s with peaks as high as 5–12 burst/s. In contrast, HVCRA and HVCX neurons produced bursts extremely rarely, on average one burst every 10 s. Averages of firing rates, firing rates during bursts, burst rates, burst widths, the number of spikes per burst, and the fraction of spikes in a burst are summarized in Table 1.
To account for sleep-state-dependent variability that might have influenced burst measurements reported in Table 1, we also measured firing rates in burst trains using simultaneously recorded pairs. From these measurements, we computed the correlation coefficient of firing rates in nonoverlapping 3-s windows. The result of this analysis was that the occurrence of increased bursting in one neuron of a pair tended to be positively correlated with increased bursting in the other neuron (Fig. 6B, , significantly correlated pairs). In all 50 RA-HVCI neuron pairs, the firing rate correlation was significant, with an average correlation coefficient of 0.66. In 38 of 47 HVCRA-RA neuron pairs, the firing-rate correlation of burst trains was significant as well, with an average coefficient of 0.43. In 16 of 26 HVCRA-HVCI neuron pairs, the correlation was significant, with an average coefficient of 0.47. A similar picture also emerges for the HVCX neurons. They exhibited significant correlations in 18 of 26 paired recording with HVCI neurons, resulting in an average coefficient of 0.5.
As we have seen in Fig. 6A, the firing rates in burst trains were widely distributed. We also observed a wide distribution for the firing-rate ratios in simultaneously recorded neuron pairs, Fig. 6C. Typically, these ratios were 50% larger than the ratios inferred from single neuron recordings (Table 1, last column). Due to insensitivity to sleep nonstationarities, we believe the estimated ratios from paired recordings to be more accurate than the single neuron estimates. A simple self-consistency test reinforced us in this belief: By multiplying the median firing rate ratio of HVCRA-RA pairs (15.5) with the median ratio of RA-HVCI pairs (3.4), we obtained a value of 52.7, which was close to the measured ratio of HVCRA-HVCI pairs of 46.8.
We also used our paired recordings to estimate burst rate ratios (number of bursts per second, see METHODS). We found that RA neurons burst 13.9 times as often as did HVCRA neurons, and HVCI neurons burst 2.7 times as often as did RA neurons. These data imply that HVCI neurons should burst roughly 37.5 times as often as do HVCRA neurons, a number that is close to our measured value of 32.6. Note that the burst-rate ratio of RA neurons and HVCRA neurons of 13.9 is very similar to that found during singing [12 bursts per motif for RA neurons (Leonardo and Fee 2005) and 1 burst per motif for HVCRA neurons (Hahnloser et al. 2002)].
Analysis of burst patterns in the premotor pathway
By creating raster plots of RA neuron activity, time-aligned to the onset of bursts in HVCRA neurons, we frequently observed correlated patterns of spikes (Fig. 7A). As previously reported, RA neurons nearly always exhibit a brief sequence of bursts reliably locked to the sparse HVCRA neuron bursts (Hahnloser et al. 2002). Although well correlated to the HVCRA bursts, the bursts in RA neurons were generally not synchronized (at 0 time lag) but occurred over a wide range of time lags relative to the bursts of HVCRA neurons. We have quantified the correlated spike events by the coherency function (see METHODS). Most HVCRA-RA neuron pairs exhibited at least one significant coherency peak (35/46 pairs had 
1 significant peak, among which 21 pairs had exactly 1 significant peak, 12 pairs had 2 peaks, 1 pair had 3 peaks, and 1 pair had 4 peaks). Main peaks had a median time lag relative to HVCRA neuron spikes of 6 ms, in the range from –63 to 65 ms, Fig. 8A, bottom. All peaks including secondary peaks were dispersed over a range from about –80 to 110 ms. The average coherency function of HVCRA-RA neuron pairs exhibited a peak value of 3*10–3 at a time lag of 6 ms relative to the HVCRA neuron spikes, Fig. 8A, top. As can be seen, this peak value was above the average significance threshold for HVCRA and RA neurons (Fig. 8A). These results agree with the causal notion that HVCRA neurons drive spikes in RA neurons at a time lag within the range of measured spike latencies in Fig. 3C.
Nearly all HVCI neurons we recorded were antidromically activated from RA, confirming previous findings that they receive strong excitatory synaptic input from HVCRA neurons (Mooney and Prather 2005; Rosen and Mooney 2003). In paired recordings, we found that HVCI neurons exhibited small periods of dense bursting, time-locked to HVCRA neuron bursts, Fig. 7, B and C.
The coherency functions and peaks between different HVC and RA neuron types are summarized in Fig. 8. The average coherency function of RA neuron pairs peaked at 0 ms, with a peak coherency of 9*10–3, Fig. 8B, top. Of 31 recorded RA neuron pairs, 26 exhibited at least one significant coherency peak. Main peaks were narrowly distributed in the time interval from –10 to 8 ms (excluding 1 outlier at –140 ms). The average coherency function between RA and HVCI neurons reached a peak of 1.4*10–2 at a time lag of –4 ms (Fig. 8E, top), indicating that HVCI neurons tended to spike shortly before RA neurons did. All of 50 recorded RA-HVCI neuron pairs exhibited at least one significant (and central) coherency peak. Main peaks were distributed over an interval from –9 to 11 ms, with a median peak time lag of –2 ms, Fig. 8E, bottom. The average coherency function of HVCRA-HVCI neuron pairs peaked at a time lag of 0 ms with a peak value of 7*10–3, Fig. 8C, top. In 23 of 26 recorded pairs, we found at least one significant coherency peak. Main peaks were widely distributed, from –28 to 17 ms, with a median peak time lag of 1 ms Fig. 8C, bottom. Finally, the average coherency function of HVCI neuron pairs reached a large peak value of 3.9*10–2, Fig. 8D, top. All of the 19 recorded neuron pairs exhibited at least one significant coherency peak. Central peaks were narrowly distributed, from just –5 to 4 ms, Fig. 8D, bottom.
The large range of time lags of coherency peaks between HVCRA and RA neurons in Fig. 8A prevented us from precisely estimating the latency of HVC drive to RA. However, given the 0 ms time lag of the average HVCRA-HVCI coherency function (Fig. 8C, top) and the –4 ms time lag of the average RA-HVCI coherency function (Fig. 8E, top), we inferred a roughly 4 ± 0.5-ms latency of the HVCRA neuron drive to RA neurons (the SE of 0.5 ms was estimated by the SD of 2.3 ms for the central peaks between the 26 HVCI-RA neuron pairs and an estimated SD of 1.0 ms for the 0-ms time lag of the average HVCRA-HVCI coherency). Thus we found that the inferred latency of HVCRA neuron drive to RA neurons agreed well with the average HVCRA antidromic spike latency of 4.6 ms shown in Fig. 3C.
In conclusion, between HVC and RA neurons we find highly significant spike coherencies that peak at time lags consistent with a causal role of HVC neurons driving spikes in RA neurons.
Activity in the anterior forebrain pathway (AFP)
We examined the AFP for a possible role in the generation of RA burst sequences. First we studied the HVC input to the AFP by performing paired recordings of HVCX neurons with other HVC neuron types. Raster plots revealed correlations between HVCX and HVCI neurons, an example is shown in Fig. 9A. The average coherency function between HVCX and HVCI neurons exhibited a significant peak of 5*10–3 at a time lag of 1.3 ms, Fig. 9B, top. Accordingly, HVCX neurons tended to spike before HVCI neurons did. Of 26 recorded HVCX-HVCI neuron pairs, 17 exhibited at least one significant coherency peak. Main coherency peaks ranged from –18 to 26 ms (excluding 3 outliers), with a median peak time lag of 4 ms, Fig. 9B, bottom. In summary, HVCX neurons were less correlated with HVCI neurons than were the HVCRA neurons.
The synchrony of HVCX neurons with HVCI neurons suggests that HVCX neurons engage in the sparse HVCRA neuron sequences. The reason for this suggestion is that correlations of random variable pairs X and Y, and Y and Z, usually imply that X and Z are correlated as well. In seven paired recordings in three birds of HVCX and HVCRA neurons, we found one pair that showed strongly correlated activity, Fig. 9C: the HVCX neuron tended to burst reliably before the HVCRA neuron did, showing that HVCX neuron activity can be time-locked to HVCRA neuron sequences. We believe the reason why we found so few correlated HVCX-HVCRA neuron pairs is the sparse activity of these neurons, making it unlikely that any two of them will fire at a small enough time lag to each other. We further tested the temporal locking of HVCX neuron activity with HVCRA neuron sequences by paired recordings from HVCX and RA neurons. In 12 paired recordings, we found 7 pairs that exhibited at least one significant coherency peak. The average coherency function between HVCX and RA neurons exhibited a significant peak of 2.2*10–3 at a time lag of 4.2 ms, indicating that HVCX neurons tended to spike a few milliseconds before RA neurons did, Fig. 9D, top. Main coherency peaks ranged from –59 to 19 ms with a median peak time lag of 0 ms, Fig. 9D, bottom. The average coherency between HVCX neurons and RA neurons was smaller than the average coherency between HVCRA and RA neurons. Nevertheless, our data show that HVC sleep sequences transmitted to Area X are qualitatively similar to the sequences sent to RA. Such a similarity has also been observed during singing, in which both HVCX and HVCRA neurons were found to be sparsely bursting (Hahnloser et al. 2002).
In conclusion, the coherent activity in HVCX neurons raises the question whether the correlations between HVCRA and RA neurons arise from the direct pathway between HVC and RA or from the AFP or from both.
Reversible HVC and LMAN lesions
To determine the relative roles of the premotor pathway and the AFP in generating sleep bursts in RA, we examined the effect of reversible inactivation of LMAN while recording activity in RA, Fig. 10. When we injected the sodium-channel blocker lidocaine into HVC, the spontaneous bursting of RA neurons immediately stopped, but recovered after 2–3 min (pre: 1.29 bursts/s, post: 0.10 bursts/s, recovery: 1.21 bursts/s, 3/3 birds, Fig. 10, A and C). Saline injections into HVC did not have any significant effects on RA neuron burst rates (P < 0.01, n = 3 birds).
|
) and suggest that activity in the AFP, in contrast to HVC, has little effect on RA sleep burst rate. RA-projecting LMAN neurons
To investigate the relation of LMAN activity with RA activity in the sleeping adult bird, we recorded antidromically identified LMANRA neurons. LMAN projection neurons have spinous dendrites and bifurcating axons that project to Area X and to RA (Livingston and Mooney 1997). Unlike HVCRA neurons, RA-projecting LMAN neurons (LMANRA neurons) exhibit little firing rate adaptation in vivo to intracellular current injections (Mooney 2000; Rosen and Mooney 2000). In the singing bird, they exhibit random activity patterns and are involved in producing song variability (Kao et al. 2005; Olveczky et al. 2005). LMANRA neurons are glutamatergic, but unlike HVCRA neurons, in the adult bird, they connect onto RA neurons exclusively via NMDA-type synaptic receptors (Mooney 1992; Stark and Perkel 1999).
We found that electrical stimulation in RA produced antidromic spikes in LMAN of short latency (2.54 ± 0.81 ms, range: 1.6–4.7 ms, Fig. 11A). All neurons that responded to RA stimulation were identified by the spike collision test as LMANRA neurons, Fig. 11B. LMAN neurons that did not respond to electrical stimulation in RA were excluded from further analysis. Bursts in LMANRA neurons were rare, they had a low peak firing rate, and typically were composed of only two to three spikes. The firing properties of LMANRA neurons are summarized in Table 1: LMANRA burst rates were comparable to those of HVC projection neurons, though their average firing rates were roughly twice as high.
The average ISI pdf of LMANRA neurons showed a broad peak at 10 ms, Fig. 11C. The average auto-covariance function of LMANRA neurons peaked at short times (<10 ms) and decayed to zero only within several tens of seconds, exhibiting the longest tail of all recorded neuron types in this study (compare Fig. 11D to Fig. 6). Of 29 LMANRA neurons, all had renewal auto-covariance functions (Jackknife, P < 0.001). Nevertheless, the median auto-covariance function of LMANRA neurons was significantly larger than the median renewal function in the time interval 7–11 ms (Mann-Whitney rank sum test, P < 0.001).
To investigate correlations between LMAN and RA activity, we performed paired recordings of LMANRA and RA neurons during sleep. Figure 12A shows a typical example, in which the activity of the LMANRA neuron was highly fluctuating. There is no apparent relation between LMANRA spikes and RA neuron bursts. The average coherency function between LMANRA and RA neurons was positive over a broad range from about –200 to 70 ms but did not extend above the average significance threshold, Fig. 12B, top. The peak of 1.5*10–3 of the average coherency function was reached at a positive time lag of 13 ms. Of 11 recorded pairs, 8 exhibited at least one significant coherency peak. Main peaks were distributed over a large range from –91 to 27 ms, with a median peak time lag of –40 ms relative to LMANRA neuron spikes. The largest main peak occurred at a negative time lag (Fig. 12B, bottom), indicating that in these neurons, the LMANRA spikes were most significant when they occurred after RA neuron spikes. Further analysis showed that coherency peaks at a positive time lag in Fig. 12B were not very robust: by removing single spikes from both LMANRA and RA neuron spike trains before running the coherency analysis, we found that no significant coherency peak remained at a positive time lag, suggesting that LMANRA neuron bursts were not effective in driving bursts in RA neurons. Note, when single spikes were removed from paired recordings of HVC and RA neurons, all significant coherency peaks remained, indicating a strong causal link between bursts in HVC and RA.
To investigate correlations between LMAN and HVC activity, we performed paired recordings of LMANRA and HVCI neurons. Again, there was a broad range over which the average coherency function was positive but without extending above the average significance threshold. Of 14 recorded LMANRA-HVCI neuron pairs, 12 exhibited at least one significant coherency peak. Individual peak times ranged from –90 to 36 ms with a median peak time lag of –27 ms, Fig. 12C, bottom.
In conclusion, even though LMAN activity was not coherent on average at any time lag with either HVC or RA activity, individual neuron pairs exhibited significant coherency peaks distributed over a wide range of time lags. Coherency peaks in Fig. 12 were smaller than coherency peaks between HVC neurons and RA neurons in Fig. 8. Based on these results, we estimate the potential LMAN drive to RA to be significantly smaller than the HVC drive, in agreement with our reversible lesions results in Fig. 11.
Model: coherency peaks from serial correlations
We have found that for many neuron types, coherency functions exhibited significant peaks at both positive and negative time lags. Depending on neuron types, either the peak at positive or negative time lag is in apparent disagreement with a feedforward connection between brain areas. For example, many HVCRA neurons spiked reliably after RA neurons did (Fig. 8A) even though HVC projects to RA and not vice versa. Similarly, many LMANRA neurons spiked after RA neurons did (Fig. 12B), even though LMAN projects to RA and not vice versa. Finally, many LMANRA neurons spiked before HVCI neurons did (Fig. 12C), even though the AFP extends from HVC to LMAN (via Area X and DLM) not vice versa.
With our following model we wanted to explore whether these dual coherency peaks could be seen as a byproduct of serial correlations within HVC. The underlying assumption of our model is a discrete-time, probabilistic mechanism for the generation of HVC burst sequences. We assumed that every second, some HVCRA and HVCX neurons are stimulated from a source outside of HVC. With unit probability, the stimulated HVCRA neurons respond with a burst to the external stimulation, and with probability 0.8, the stimulated HVCX neurons respond with a burst. If and only if HVCX neurons burst, then a HVC bursts sequence (or epoch) is initiated. Forty milliseconds later, the sequence may either stop or proceed for another 40 ms (with probability 0.8). This process is repeated in discrete time intervals of 40 ms until the stimulated HVCX neurons fail to burst and the sequence comes to a halt. In other terms, we modeled HVC activity by a Markov chain such that bursts are expressed as probabilistic events, and long activity sequences are less frequent than short ones.
For simplicity, we did not model the HVC burst sequence in continuous time. We restricted our simulations of sequence dynamics to bifurcation events separated by 40-ms intervals (see METHODS). Bursts in the various HVC, RA, and LMAN neurons were modeled at fixed time lags of these bifurcation events. HVC neuron bursts had a zero time lag, the RA neuron bursts had a time lag of 4 ms, and the LMANRA neuron bursts had a time lag of 40 ms. A schematic of the model is shown in Fig. 13A.
Spike trains from a Monte-Carlo simulation of this model are shown on a large and on a small time scale in Fig. 13, B and C, respectively. Spike trains were generated by randomly selecting a small number of ISIs for each of the burst events, giving rise to fluctuating instantaneous firing rates. The spike trains of the HVCX and HVCRA neurons were produced by considering different states in the HVC sequence, assuming that HVC projection neurons burst at most once per burst epoch. In the simulations, for each burst epoch, we associated the HVCX and HVCRA neurons with a randomly selected bifurcation event. For example, in Fig. 13B, the HVCX neuron was not bursting in the first two burst epochs. Then in the third burst epoch it was part of the first bifurcation event, and in the fourth epoch it was part of the seventh event, etc.
Average coherency functions for these spike trains shown in Fig. 13D were in good agreement with the measured ones in Figs. 8A and 12, B and C. There was a single large coherency peak between the RA and the HVCI neuron centered at –4 ms. The coherency function of the HVCRA-RA neuron pair had a main peak at 4 ms and smaller peaks 40 ms apart, both at positive and negative time lags. The two coherency functions involving the LMANRA neuron clearly exhibited two main peaks, one at a negative time lag and one at a positive time lag. The location of these two peaks was quite sensitive on simulation parameters. Their location was only as shown if the burst probability of HVCX neurons was smaller than that of HVCRA neurons (0.8 vs. 1). Additional simulations showed that in the converse case, the peak close to zero time lag disappeared in favor of a peak at –80 ms; and, in the case of equal probabilities, the LMANRA coherency functions were symmetrical around a peak at –40 ms. The observation that coherency functions of the model were consistent with the data only if HVCRA neurons were more reliable than HVCX neurons for sustaining HVC burst sequences is interesting. It is in agreement with the finding that HVCRA neurons are more strongly correlated with HVCI and RA neurons than are the HVCX neurons.
A note on our model: we chose a time interval of 40 ms between bifurcation events because of correspondence with the typical interburst interval of HVCI neurons during dense burst epochs (25 burst/s). This time interval determines the time lag between coherency peaks in Fig. 13D. We leave it open as to whether this time interval also has a mechanistic interpretation for how HVC burst sequences are generated (for example by input from NIf or Uva) or whether it can be viewed as a consequence of more complex, continuous-time simulations with stereotyped sequential behaviors (i.e., where neurons are not allowed to burst at arbitrary times in the continuous sequence but only at a fixed set of discrete times). A detailed exploration of this issue is beyond the scope of this paper.
|
|
DISCUSSION |
|---|
|
). In our study, the 0.35-ms limit applied to all HVCI neurons as well. However, it also applies to most HVC projection neurons, and so it was not a feasible criterion for neuron identification. We attribute the narrowness of projection neuron spikes in our study to the large impedances of our glass electrodes: 5–15 versus 3–5 M in Rauske et al. Another reason for wider spikes in Rauske et al. might be their partial use of metal electrodes, which accentuate spike waveform differences also in our experience. Recall that the benefit of our large-impedance electrodes was our improved ability to record from HVCRA neurons, which are known to have very small somata and are hard to isolate with larger electrodes.
Virtually no HVC projection neuron could be activated by stimulation in both Area X and RA. This finding suggests that HVCRA and HVCX neurons do either have very few or weak excitatory connections between them or that a possible source of synaptic excitation between these cell types is masked by fast inhibition. In support of the latter view, in recent studies of the HVC microcircuit, strong disynaptic GABAergic inhibition has been found from HVCRA to HVCX neurons (Mooney and Prather 2005; Rosen and Mooney 2003). We have found HVCI neurons to be more reliably activated by RA stimulation than by Area X stimulation. This agrees with the finding that very few HVCX-HVCI neuron pairs show signs of excitatory connections (Mooney and Prather 2005). In agreement with the idea that the excitatory synaptic drive from HVCRA to HVCI neurons is stronger than that of HVCX to HVCI neurons, we found that HVCI neurons exhibited significant coherencies more frequently with HVCRA neurons than with HVCX neurons. Despite these possibly important differences between local HVCRA and HVCX circuits, we find qualitatively similar sparse activity patterns in these two types of projection neurons during sleep.
Similar to previous lidocaine injection experiments where song-related auditory responses in RA were found to be mediated by HVC and not by LMAN (Doupe and Konishi 1991), we here find that LMAN is not necessary for the generation of premotor-like activity in RA: burst rates in RA are unchanged after silencing LMAN, Fig. 10.
We want to compare our results to similar studies performed in the anesthetized bird (Kimpo et al. 2003; Rauske et al. 2003). Physiological differences are known to exist between sleep and anesthesia (Cardin and Schmidt 2003). We can infer some of these differences by comparing our spontaneous firing rates in Table 1 with spontaneous rates reported in Rauske et al.: HVCX neurons seem to have a somewhat lower spontaneous firing rate during sleep than during anesthesia (1.6 vs. 3.8 Hz), whereas HVCI neurons seem to have a much higher firing rate during sleep (15.0 ± 9.0 vs. 3.2 ± 3.1 Hz). These differences make it plausible that state dependence is an important factor that could affect spike generation mechanisms in RA.
Despite these differences, the average LMANRA-RA coherency function in Fig. 12B is qualitatively similar to coherency functions found in the anesthetized bird (Kimpo et al. 2003). Coherencies in Kimpo et al. are 10 times larger than the ones we depict. This apparent mismatch can be entirely attributed to our choice of bin size (1 vs. 10 ms in Kimpo et al.). To compare our measured coherency values with those in Kimpo et al. our coherency functions must be expressed in terms of 10-ms bins, which roughly mounts to multiplication by a factor of 10, resulting in good quantitative agreement.
We are able to offer a simple mechanism for our observed coherency functions. We attribute secondary coherency peaks as those in Figs. 7A, 8A, and 12, B and C, to transient locking of neural activity into stereotyped sequential patterns. Our generative model of spike train dynamics in the HVC-LMAN-RA pathways is based on sequential burst states in HVC. Interestingly, we find good qualitative agreement with the data if HVCRA neurons exhibit a stronger tendency to form sequential activation than do HVCX neurons. In agreement with this idea, adult song structure is not changed immediately after selective lesions of HVCX neurons (Scharff et al. 2000), suggesting that HVCRA neurons can produce burst sequences even in the absence of HVCX neuron input.
In the study of Kimpo et al., it was argued that the first peak at a negative time lag of LMAN-RA coherency functions was due to common input to LMAN and RA from HVC, whereas the second peak at a positive time lag was attributed to LMANRA neurons driving spiking activity in RA. In our model, coherency peaks at both positive and negative time lags arise from mere existence of a common source to RA and LMAN with serial correlations. Our model was not meant to provide a detailed account for the data, even though qualitative agreement was quite simple to achieve. In a sense, our model illustrates the well-known limitation of inferring cause from correlations. We did not make use of anatomical connections between LMANRA and RA neurons. The RA and LMAN burst dynamics in our model are conditionally independent, illustrating that no connections between LMAN and RA are necessary to explain dual coherency peaks in paired recordings in these areas. Finally, we would like to emphasize that our model is only a caricature of sleep dynamics and has several shortcomings. For example, single-neuron statistics (e.g., ISI pdfs and auto-covariance functions) are only poorly reproduced. More involved models and more realistic assumptions on HVC dynamics will be necessary for a better understanding of mechanisms for spike trains generation in different neuron types. The vast diversity of burst rates and the complexity of spontaneous activity patterns make the sleeping bird an ideal system for exploring biophysical models of population activity in neural circuits.
The AFP is known to have an important role in motor learning and song maintenance. However, a detailed understanding of whether this role unfolds in an on-line mode during singing or an off-line mode during sleep is still unknown. The fact that deafening has little effect on LMAN activity during singing (Hessler and Doupe 1999) and the lack of auditory responsiveness of LMAN neurons during singing (Leonardo 2002) suggests that if LMAN plays a role in transmitting an error-related signal to the motor system, this does not occur during singing—at least in adult birds. This raises the possibility that such processing may occur during sleep. For example, it has been proposed that LMAN activity during sleep may represent a prediction of real auditory feedback—a virtual error signal (Dave and Margoliash 2000). Our recordings were made in adult birds with a crystallized song, thus any error-related activity in LMAN is likely to be small. It would be interesting to investigate whether LMAN activity exhibits strong correlations to RA bursts in birds that are actively learning their song, for example in juveniles or decrystallized adult birds (Leonardo and Konishi 1999).
From our study, we conclude that spontaneous activity in the HVC-RA circuit of adult birds during sleep is generated by essentially the same mechanism as during singing. It is currently unclear how this finding fits into theories of memory consolidation during slow-wave sleep, where usually a dramatic difference between circuit dynamics during behavior and sleep is observed (Nadasdy et al. 1999). Possibly, further experiments in juveniles or new computational ideas are needed to resolve this conundrum.
|
|
GRANTS |
|---|
|
|
|
ACKNOWLEDGMENTS |
|---|
|
|
|
FOOTNOTES |
|---|
Address for reprint requests and other correspondence: R.H.R. Hahnloser, Institute for Neuroinformatics UNIZH/ETHZ, Winterthurerstrasse 190, 8057 Zurich, Switzerland (E-mail: rich{at}ini.phys.ethz.ch)
|
|
REFERENCES |
|---|
|
Amlaner CJ Jr and Ball NJ (Editors). Avian Sleep (2nd ed.). Phladelphia, PA: Saunders, 1994.
Bottjer SW, Miesner EA, and Arnold AP. Forebrain lesions disrupt development but not maintenance of song in passerine birds. Science 224: 901–903, 1984.
Cardin JA and Schmidt MF. Song system auditory responses are stable and highly tuned during sedation, rapidly modulated and unselective during wakefulness, and suppressed by arousal. J Neurophysiol 90: 2884–2899, 2003.
Cox DR. Renewal Theory. London: Mehtuen, 1962.
Dave AS and Margoliash D. Song replay during sleep and computational rules for sensorimotor vocal learning. Science 290: 812–816, 2000.
Deregnaucourt S, Mitra PP, Feher O, Pythe C, and Tchernichovski O. How sleep affects the developmental learning of bird song. Nature 433: 710–716, 2005.[CrossRef][Medline]
Doupe AJ. A neural circuit specialized for vocal learning. Curr Opin Neurobiol 116: 104–111, 1993.
Doupe AJ and Konishi M. Song-selective auditory circuits in the vocal control system of the zebra finch. Proc Natl Acad Sci USA 88: 11339–11343, 1991.
Doupe AJ, Solis MM, Kimpo RR, and Boettiger CA. Cellular, circuit, and synaptic mechanisms in song learning. Ann NY Acad Sci 1016: 495–523, 2004.[CrossRef][Web of Science][Medline]
Dutar P, Vu HM, and Perkel DJ. Multiple cell types distinguished by physiological, pharmacological, and anatomic properties in nucleus HVc of the adult zebra finch. J Neurophysiol 80: 1828–1838, 1998.
Fee MS, Kozhevnikov AA, and Hahnloser RH. Neural mechanisms of vocal sequence generation in the songbird. Ann NY Acad Sci 1016: 153–170, 2004.[CrossRef][Web of Science][Medline]
Gerstner W and Kistler W. Spiking Neuron Models. Single Neurons, Populations, Plasticity. Cambridge, UK: Cambridge Univ. Press, 2002.
Hahnloser R, Kozhevnikov A, and Fee MS. An ultrasparse code underlies the generation of neural sequences in a songbird. Nature 419: 65–70, 2002.[CrossRef][Medline]
Hessler NA and Doupe AJ. Singing-related neural activity in a dorsal forebrain-basal ganglia circuit of adult zebra finches. J Neuroscience 19: 10461–10481, 1999.
Hishikawa Y, Cramer H, and Kuhlo W. Natural and melatonin-induced sleep in you chickens—a behavioral and electrographic study. Exp Brain Res 7: 84–94, 1969.[Web of Science][Medline]
Holden AV. Signal processing: neural coding by correlation? Nature 428:382, 2004.[CrossRef][Medline]
Kao MH, Doupe AJ, and Brainard MS. Contributions of an avian basal ganglia-forebrain circuit to real-time modulation of song. Nature 433: 638–643, 2005.[CrossRef][Medline]
Kimpo RR, Theunissen FE, and Doupe AJ. Propagation of correlated activity through multiple stages of a neural circuit. J Neuroscience 23: 5750–5761, 2003.
Kubota M and Taniguchi I. Electrophysiological characteristics of classes of neurons in the HVc of the zebra finch. J Neurophysiol 80: 914–923, 1998.
Leonardo A. Neural dynamics underlying complex behavior in a songbird. In: Computation and Neural Systems. Pasadena CA: Caltech, 2002.
Leonardo A and Fee MS. Ensemble coding of vocal control in birdsong. J Neuroscience 19: 652–661, 2005.
Leonardo A and Konishi M. Decrystallization of adult birdsong by perturbation of auditory feedback. Nature 399: 466–470, 1999.[CrossRef][Medline]
Livingston FS and Mooney R. Development of intrinsic and synaptic properties in a forebrain nucleus essential to avian song learning. J Neurosci 17: 8997–9009, 1997.
Louie Ka and Wilson MA. Temporally structured replay of awake hippocampal ensemble activity during rapid eye movement sleep. Neuron 29: 145–156, 2001.[CrossRef][Web of Science][Medline]
Margoliash D. Song learning and sleep. Nat Neurosci 8: 546–548, 2005.[CrossRef][Web of Science][Medline]
Mooney R. Synaptic basis for developmental plasticity in a birdsong nucleus. J Neurosci 12: 2464–2477, 1992.[Abstract]
Mooney R. Different subthreshold mechanisms underlie song selectivity in identified HVc neurons of the zebra finch. J Neurosci 201: 5420–5436, 2000.
Mooney R and Prather JF. The HVC microcircuit: the synaptic basis for interactions between song motor and vocal plasticity pathways. J Neurosci 25: 1952–1964, 2005.
Nadasdy Z, Hirase H, Czurko A, Csicsvari J, and Buzsaki G. Replay and time compression of recurring spike sequences in the hippocampus. J Neurosci 19: 9497–9507, 1999.
Nick TA and Konishi M. Dynamic control of auditory activity during sleep: correlation between song respone and EEG. Proc Natl Acad Sci USA 98: 14012–14016, 2001.
Nottebohm F, Kelley DB, and Paton JA. Connections of vocal control nuclei in the canary telencephalon. J Comp Neurol 207: 344–357, 1982.[CrossRef][Web of Science][Medline]
Nottebohm F, Stokes TM, and Leonard CM. Central control of song in the canary, Serinus canarius. J Comp Neurol 165: 457–486, 1976.[CrossRef]
Okuhata S and Saito N. Synaptic connections of thalamo-cerebral vocal nuclei of the canary. Brain Res Bull 18: 35–44, 1987.[CrossRef][Web of Science][Medline]
Olveczky BP, Andalman AS, and Fee MS. Vocal experimentation in the juvenile songbird requires a basal ganglia circuit. PLoS Biol 3: e153, 2005.[CrossRef][Medline]
Phillips NH and Berger RJ. Melatonin infusions restore sleep suppressed by continuous bright light in pigeons. Neurosci Lett 145: 217–220, 1992.[CrossRef][Web of Science][Medline]
Rauske PL, Shea SD, and Margoliash D. State and neuronal class-dependent reconfiguration in the avian song system. J Neurophysiol 89: 1688–1701, 2003.
Rosen MJ and Mooney R. Intrinsic and extrinsic contributions to auditory selectivity in a song nucleus critical for vocal plasticity. J Neurosci 20: 5437–5448, 2000.
Rosen MJ and Mooney R. Inhibitory and excitatory mechanisms underlying auditory responses to learned vocalizations in the songbird nucleus HVC. Neuron 39: 177–194, 2003.[CrossRef][Web of Science][Medline]
Scharff C and Nottebohm F. A comparative study of the behavioral deficits following lesions of various parts of the zebra finch song system: implications for vocal learning. J Neurosci 11: 2896–2913, 1991.[Abstract]
Scharff C, Kirn JR, Grossman M, Macklis JD, and Nottebohm F. Targeted neuronal death affects neuronal replacement and vocal behavior in adult songbirds. Neuron 25: 481–492, 2000.[CrossRef][Web of Science][Medline]
Sohrabji F, Nordeen EJ, and Nordeen KW. Selective impairment of song learning following lesions of a forebrain nucleus in the juvenile zebra finch. Behav Neural Biol 53: 51–63, 1990.[CrossRef][Web of Science][Medline]
Stark LL and Perkel DJ. Two-stage, input-specific synaptic maturation in a nucleus essential for vocal production in the zebra finch. J Neurosci 19: 9107–9116, 1999.
Swadlow H. Neocortical efferent neurons with very slowly conducting axons: strategies for reliable antidromic identification. J Neurosci Methods 79: 131–141, 1998.[CrossRef][Web of Science][Medline]
Thomson DJ and Chave AD (Editors). Jackknifed Error Estimates for Spectra, Coherences, and Transfer Functions. Englewood Cliffs, NJ: Prentice Hall, 1991.
Tobler I and Borbely AA. Sleep and EEG sepctra in the pigeon (Columba livia) under baseline conditions and after sleep deprivation. J Comp Physiol [A] 163: 729–738, 1988.[CrossRef]
Vicario DS and Nottebohm F. Organization of the zebra finch song control system. I. Representation of syringeal muscles in the hypoglossal nucleus. J Comp Neurol 271: 346–354, 1988.[CrossRef][Web of Science][Medline]
Wilson MA and McNaughton BL. Reactivation of hippocampal ensemble memories during sleep. Science 265: 676–679, 1994.
Yu AC and Margoliash D. Temporal hierarchical control of singing in birds. Science 273: 1871–1875, 1996.
This article has been cited by other articles:
|
|
 |
|
  L. Gibb, T. Q. Gentner, and H. D. I. Abarbanel Inhibition and Recurrent Excitation in a Computational Model of Sparse Bursting in Song Nucleus HVC J Neurophysiol, September 1, 2009; 102(3): 1748 - 1762. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
S. K. Arfin, M. A. Long, M. S. Fee, and R. Sarpeshkar Wireless Neural Stimulation in Freely Behaving Small Animals J Neurophysiol, July 1, 2009; 102(1): 598 - 605. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 R. H. R. Hahnloser, C. Z.-H. Wang, A. Nager, and K. Naie Spikes and Bursts in Two Types of Thalamic Projection Neurons Differentially Shape Sleep Patterns and Auditory Responses in a Songbird J. Neurosci., May 7, 2008; 28(19): 5040 - 5052. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
J. A. Thompson, W. Wu, R. Bertram, and F. Johnson Auditory-Dependent Vocal Recovery in Adult Male Zebra Finches Is Facilitated by Lesion of a Forebrain Pathway That Includes the Basal Ganglia J. Neurosci., November 7, 2007; 27(45): 12308 - 12320. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
J. E. Miller and S. A. White The Sleeping Bird Gets the Song. Focus on: "HVC Neural Sleep Activity Increases With Development and Parallels Nightly Changes in Song Behavior" J Neurophysiol, July 1, 2007; 98(1): 3 - 4. [Full Text] [PDF]
|
|
|
|
|
|
A. A. Kozhevnikov and M. S. Fee Singing-Related Activity of Identified HVC Neurons in the Zebra Finch J Neurophysiol, June 1, 2007; 97(6): 4271 - 4283. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
Z. Chi, W. Wu, Z. Haga, N. G. Hatsopoulos, and D. Margoliash Template-Based Spike Pattern Identification With Linear Convolution and Dynamic Time Warping J Neurophysiol, February 1, 2007; 97(2): 1221 - 1235. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
R. H. R. Hahnloser and M. S. Fee Sleep-Related Spike Bursts in HVC Are Driven by the Nucleus Interface of the Nidopallium J Neurophysiol, January 1, 2007; 97(1): 423 - 435. [Abstract] [Full Text] [PDF]
|
|
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
| Visit Other APS Journals Online |
TOP
ABSTRACT
INTRODUCTION
