Unsupervised Restoration of a Complex Learned Behavior After Large-Scale Neuronal Perturbation

Abstract

Reliable execution of precise behaviors requires that brain circuits are resilient to variations in neuronal dynamics. Genetic perturbation of the majority of excitatory neurons in a brain region involved in song production in adult songbirds with stereotypical songs triggered severe degradation of their songs. The song fully recovered within two weeks, and substantial improvement occurred even when animals were prevented from singing during the recovery period, indicating that offline mechanisms enable recovery in an unsupervised manner. Song restoration was accompanied by increased excitatory synaptic inputs to neighboring unmanipulated neurons in the same brain region. A model inspired by the behavioral and electrophysiological findings suggests that unsupervised single-cell and population-level homeostatic plasticity rules can support the functional restoration after large-scale disruption of networks that implement sequential dynamics. These observations suggest the existence of cellular and systems-level restorative mechanisms that ensure behavioral resilience.

INTRODUCTION

Animal survival and reproduction requires reliable execution of behaviors. However, neuronal representations change over time, as a consequence of natural drift, or due to neuronal perturbations caused by trauma, disease, or aging^1,2. What are the mechanisms that allow brains to maintain reliable behaviors over long periods of time or after neuronal loss? To investigate this question, we studied the zebra finch, a songbird that, after learning a song as juveniles, produces stereotypical renditions of the song with minimal variability over several years. Songbirds have a series of brain nuclei dedicated for song learning and production called the song system³. HVC is the premotor nucleus in the song system that projects to the motor nucleus RA (robust nucleus of the arcopallium)⁴, involved in song production³, generating sparse sequential firing that controls timing of the song⁵. The circuit mechanism for the generation of this sequence is still under investigation^6,7.

Previous experiments have shown that complete ablation of HVC irreversibly abolishes song production³. Similarly, ablation of nucleus Uva, a major source of input to HVC, irreversibly blocks song production^3,8. In contrast, localized, partial lesions of HVC cause song degradation followed by recovery after varying amounts of time^9,10. Similarly, lesioning of nucleus Nif, one of the sources of input to HVC, also causes song degradation followed by recovery of neuronal dynamics and behavior ¹¹.

The fact that the song is able to recover from these degraded states suggests the song circuit is quite resilient. Many potential mechanisms may contribute to this recovery, including homeostatic mechanisms¹¹ and spike-timing dependent plasticity within HVC, or even relearning of the song through the song learning circuitry. The potential cellular mechanisms within HVC that may underlie this resilience, however, have not been directly investigated. Thus, we use local and selective genetic manipulation of projection neurons within HVC and intracellular recording to probe the behavioral dynamics and synaptic mechanisms of recovery after severe perturbation. To test hypotheses about how synaptic plasticity could account for the recovery, we developed a data-inspired model to explore the cellular mechanisms underlying the observed behavioral and electrophysiological findings.

RESULTS

Song degradation and recovery after selective large-scale perturbation of projection neurons

HVC projection neurons fire in an extremely sparse and precise manner when birds sing^5,12. To explore how singing may be affected by disrupting the precise firing of these neurons, we virally expressed a bacterial voltage-gated Na⁺ channel, NaChBac¹³, in HVC projection neurons via a bilaterally injected lentiviral vector (LV. Fig. 1a). This vector has been shown to selectively infect HVC projection neurons^14,15 with relatively high efficiency (about half of all HVC projection neurons, Extended Data Fig. 1). NaChBac expression in neurons perturbs their activity due to two features^16–18: first, NaChBac is activated at lower membrane potentials than vertebrate Na⁺ channels; and second, while vertebrate Na⁺ channel depolarizations last several ms, those of NaChBac expressing neurons last up to >1 s. The NaChBac transgene carried by the LVs contains the sequence of NaChBac fused in frame to the green fluorescent protein (GFP), allowing for visual identification of the infected neurons (Extended Data Fig. 1). Whole-cell patch clamp recording of HVC neurons in brain slices confirmed that GFP⁺ neurons displayed the characteristic NaChBac currents (Fig. 1c and Extended Data Fig. 2a-d).

Figure 1. Song degradation and recovery after selective large-scale perturbation of projection neurons.

(a) Schematic drawing illustrating the visually guided virus delivery into HVC (see Methods); (b) (Left) Spectrogram of a motif and the corresponding 2D projection (middle) using the UMAP algorithm from an unperturbed animal. The syllables are indicated by different colors and numbers in the spectrogram and UMAP plot. (Right) UMAP visualization containing ~100 songs of the bird, randomly sampled one day without perturbation; (c) Firing pattern of RA-projecting HVC (HVC(RA)) neurons with (NaChBac+) or without NaChBac (Control); (d) Example spectrograms and UMAP visualizations of the song of a bird injected with LV-NaChBac; (e) Example dual patch clamp traces demonstrating that expression of TeNT abolishes synaptic release from excitatory neurons in HVC; (f) Example spectrograms and corresponding UMAP visualizations of song renditions of a bird at different days post injection with LV-TeNT. The song motifs containing the syllables (12345) are marked by red lines and can be seen before perturbation and after recovery; ‘i’ stands for introductory note; ‘C’ stands for call. A segment of the song that did not fully recover is marked by red arrowheads (expanded view of the spectrogram is shown in Supplementary Fig. 2).

Following the perturbation, song degradation and recovery followed a characteristic time course (Fig. 1d). Normal songs consist of repetitions of a so-called motif, which is made up of 3 to 7 syllables, depending on the individual. In unperturbed zebra finches, the acoustic structure of the syllables, the number of syllables per motif, and the duration of each motif are highly stereotypical, with minimal variations between renditions, even across months (example shown in Fig. 1b). Several days after LV-NaChBac injection songs became highly irregular and bore little resemblance to the original songs (Fig. 1d). To visualize the dynamics of the changes in these perturbed songs, we used Uniform Manifold Approximation and Projection (UMAP), a non-linear dimensionality reduction algorithm, to project the high-dimensional acoustic representation of hundreds of songs onto a 2D plane (see Methods and Fig. 1b)^19–21. NaChBac delivery caused strong acoustic and temporal degradation across all syllables and greatly increased the variance between song renditions, which made it difficult to identify motifs (Fig. 1d and Supplementary Audio 1-2). At approximately 8 days post-injection (dpi), the songs of NaChBac animals started to recover, and gradually regained structure (Fig. 1d). By 15 dpi, the songs of NaChBac animals were highly similar to their original songs and remained so for several months (Fig. 1d and Supplementary Audio 3).

To investigate how NaChBac expression caused song degradation, we performed whole-cell patch clamp recordings. As early as 48 hours after injection, NaChBac+ RA-projecting (HVC(RA)) neurons displayed the characteristic NaChBac current with abnormal, long depolarizations (Fig. 1c). However, starting at 5 dpi, NaChBac+ HVC(RA) neurons showed a substantial increase in inhibitory synaptic inputs (Extended Data Fig. 2e-f). After the song had recovered, these synaptic changes remained in place (Extended Data Fig. 2e-f). In addition, we observed a significant decrease in excitatory synaptic inputs after recovery (Extended Data Fig. 2g-h). These combined changes in excitatory and inhibitory synaptic inputs may together serve to silence NaChBac+ cells (Extended Data Fig. 2i), as has been observed in the mammalian brain^16,18,22. Two main scenarios could explain why the song was degraded by NaChBac expression in HVC neurons. The low threshold firing of NaChBac could introduce noise into HVC neurons and cause abnormal firing. Alternatively, because NaChBac+ cells experience synaptic changes that silence them over a few days, degradation may be due to the loss of a large fraction of HVC neurons. As the mechanism of the effect of NaChBac on the circuit was ambiguous, we designed a second experiment to more directly manipulate neuronal function in HVC.

To investigate whether directly silencing HVC projection neurons is sufficient to degrade the song and if so, whether the song can recover, we delivered the same type of LV but carrying the light chain of tetanus toxin (TeNT). TeNT is an enzyme that cleaves synaptobrevin, a protein essential for the release of neurotransmitters. Thus, expression of TeNT in a neuron does not alter its electrical activity, but it abolishes its ability to communicate with its postsynaptic targets²³ (Extended Data Fig. 3). Paired whole-cell patch clamp recordings in brain slices showed that expression of TeNT in HVC(RA) neurons permanently abolished their excitatory drive onto inhibitory interneurons in HVC (Fig. 1e and Extended Data Fig. 3a-c). Furthermore, the excitatory drive from HVC received by projection neurons in the downstream nucleus RA is substantially reduced (Extended Data Fig. 3d-h).

Injection of LV-TeNT into HVC caused song deterioration and recovery with a similar pattern to that described for LV-NaChBac (Fig. 1f and Supplementary Audio 4-6). To investigate and compare the dynamics caused by these genetic perturbations, we quantified changes in acoustic features of syllables produced by birds receiving either LV-TeNT or LV-NaChBac²⁴. In both types of perturbation, degraded songs consist of shorter (Fig. 2a), weaker (Fig. 2b), and noisier (Fig. 2c) syllables than those of the original songs. Note that we use the term ‘syllable’ to refer to any continuous song segment in the degraded song, even if they are not stereotypical and they bear little or no resemblance to the original syllables. We measured the acoustic distortion of songs by calculating the distance of disordered syllables from the original syllables before perturbation, normalized to the maxima, in a high dimensional acoustic space derived from the spectrograms (Fig. 2d-e and Supplementary Fig. 1. see also Methods). We plotted the trajectories of each syllable to show the dynamics of song degradation and recovery (Fig. 2f). We found that, both for LV-TeNT and LV-NaChBac animals, the acoustic distance to all original syllables increased dramatically after the virus injection, but the onset and peak distortion occurred earlier for LV-TeNT (peak at 3.53 ± 0.39 dpi) than for LV-NaChBac animals (peak at 6.25 ± 0.66 dpi. Fig. 2f). As lentiviral expression starts at ~24 hours, and as TeNT is a powerful toxin, a handful of TeNT molecules are sufficient to mute a neuron, which likely accounts for its rapid action. Although the peak distortion happens later for NaChBac, the time to recovery was comparable for the two types of manipulations (13.71 ± 0.72 vs 14.80 ± 1.90 dpi. Fig. 2f). The recovered song was very close to the original, but in some cases, localized aspects of the song changed permanently, including fine-scale acoustical structure, and loss of some syllables (Fig. 1d,f & 2f and Supplementary Fig. 2).

Figure 2. Dynamics of song degradation and recovery after large-scale perturbation of HVC projection neurons.

(a) (Left) Density distribution of syllable durations day by day, for two birds either with LV-TeNT (upper) or LV-NaChBac (bottom). Note that the gap in the middle means the number of song bouts recorded that day was not enough for a meaningful analysis. (Right) Averaged duration of all syllables at three time points, before perturbation (control), when the song is most severely degraded (3–5 dpi), and when the song is fully recovered (25–30 dpi). Each line represents one animal; (b) (Left) Density distribution of the maximal log power of syllables day by day. (Right) Averaged syllable maximal log power at different time points; (c) (Left) Density distribution of the mean entropy of syllables day by day. (Right) Averaged entropy of all syllables at different time points; (d) Schematic showing methods for tracking the change in song over time using acoustic features sampled from the spectrograms. We constructed a single song feature vector of every syllable (left) in a high-dimensional acoustic space (right) (see also Methods for details); (e) Trajectory of songs in the first 2 principal components of the acoustic space over a series of days after LV-TeNT injection. Each dot represents one original or degraded syllable; (f) (Left) Acoustic distance to each original syllable, normalized so that the peak of the curve represents 100% degradation (see also Methods & Supplementary Fig. 1). Curves in the same color were from the same bird. (Right) Triangles represent the day when each syllable reached peak degradation (TeNT 3.53 ± 0.39 vs NaChBac 6.25 ± 0.66 dpi, nested two-way ANOVA), and round dots represent the day when each syllable achieved more than 90% recovery (13.71 ± 0.72 vs 14.80 ± 1.90 dpi, nested two-way ANOVA). Each column represents one bird. Note that a fraction of the syllables (marked by ‘no recovery’) never achieved 90% recovery. N = 24 syllables (nested) from 4 animals for TeNT. N = 18 (nested) / 4 for NaChBac; (g) Local variance across days (see Methods).

We quantified the rendition-to-rendition syllable variability across days (See Methods). For both NaChBac and TeNT, peak variability in syllable renditions occurred at around 10 dpi, approximately 4–6 days after the acoustic features of the songs were maximally degraded (Fig. 2g). By 20–25 dpi, the acoustic features were fully recovered. For the TeNT manipulation, syllable variability returned to the level before perturbation, while for NaChBac it remained somewhat higher than before (Fig. 2g). Importantly, the degradation of song was not due to toxicity or mechanical lesion, because no behavioral change beyond normal daily variability was observed when the same amount of viral vector expressing either GFP or a dead-pore NaChBac was injected (Extended Data Fig. 4. N = 4). Finally, song recovery was not due to restoration of normal firing activity in the NaChBac+ HVC(RA) neurons, as we verified with in vitro recordings that these neurons still had the same amount of NaChBac current (Extended Data Fig. 2a-d) and highly increased inhibitory synaptic input at >25 dpi, 10 days after the song had recovered (Extended Data Fig. 2e,f). Similarly, the muting of HVC(RA) neurons by LV-TeNT also persisted at >25 dpi (Extended Data Fig. 3a-c). The similar behavioral trajectories after NaChBac and TeNT delivery suggests that behavioral recovery after functional loss of a large portion of HVC projection neurons, either by the silencing of NaChBac+ cells or by expressing TeNT, possibly involves the same underlying restorative mechanisms.

Significant song recovery without practice

What mechanisms support recovery? Song learning in juvenile animals requires practice. Juvenile finches listen to the song of their fathers, memorize it, and attempt to copy it by continuous trial and error over 2–3 months and tens of thousands of practice renditions. The juvenile song is initially unstructured and improves over time, until it eventually becomes a faithful copy of the tutor’s song, “crystallizing” into a stable motif²⁵. To investigate the role of practice in adult song recovery after perturbation, we injected animals with LV-TeNT as described above, recorded the initial songs, and confirmed that they were degraded. We then prevented these animals from singing for 10 days during the putative song recovery period (Fig. 3a and Supplementary Fig. 3). Within the first hour after birds were allowed to sing, some motif renditions were already highly similar to the original songs (Fig. 3b-d). While there was considerable variability, song motifs were less degraded than in the vocalizations following perturbation prior to prevention (Fig. 3b-d). On the first day after song prevention, individual syllable durations and the timing of acoustics within syllables reemerged, though their variance remained elevated, suggesting that the backbone timing structure is largely restored without practice (Fig. 3e-h and Extended Data Figs. 5 & 6). Overall, syllable acoustic distances achieved ~56% of recovery in the first day after prevention (Fig. 4a-c). In the subsequent days of practice, the percentage of faithful copies increased until most renditions were highly similar to the original song (15 dpi, Fig. 4a,b and Supplementary Fig. 4). We measured the full trajectory of song recovery as a function of practice instead of time and observed that birds that were prevented from singing made substantial recovery with far less practice than birds who were allowed to sing freely (Fig. 4d,e). These observations indicate that much of the song restoration after perturbation can occur without practice and suggests that offline mechanisms strongly contribute to recovery in an unsupervised manner.

Figure 3. Significant song recovery without practice.

(a) Schematic illustrating the experimental procedure of perturbation with LV-TeNT and song prevention; (b) Spectrogram of the song of a bird before perturbation; (c) The same bird as in panel b was injected with LV-TeNT and prevented from singing for 10 days. Spectrograms of the last 10 songs sung before song prevention; (d) Spectrograms of the first 10 songs sung after song prevention. Compared to songs sung before prevention, songs sung after prevention are much improved and immediately resembled the original song motif. Syllables that can be identified are marked by numbers the same as in the original motif. Truncated syllables are additionally marked by an apostrophe; (e) Best-matched examples of one syllable (the syllable 3 in panel b) at different time points; (f) Time-varying acoustic features along the full duration of the same syllable (as in panel e) showing the high level of recovery right after prevention and eventual full recovery of the fine intra-syllable structure. Each trace is the average feature trajectory of the closest song segments on each example day (see Methods). (g) Best-matched examples of another syllable from a different bird; (h) Time-varying feature trajectories of the syllable in g, plotted in the same way as panel f. See Extended Data Fig 5 for individual rendition traces and additional syllable examples.

Figure 4. Songs continue to recover rapidly during the post-prevention period.

(a) UMAP visualization of the song renditions of the same bird shown in Fig.3 b-d. (b) Plots of distance to original syllables (normalized to maximum) of all birds prevented from singing (gray lines). Each color represents one bird. The 10-day prevention period is marked by a gray square; (c) At 13 dpi, the song recovered significantly both in freely-singing and song-prevented birds (respectively, 20.4 ± 6.3% and 44.2 ± 10.4% of the maximal degradation remaining, N = 4 & 3 animals, p = 6.7e-16 & 9.1e-12, one-sample t-test). Nested two-way ANOVA was used to test the difference between the two groups; (d) Example plots of syllable recovery vs number of practice epochs with (black trace) or without (red trace) song prevention. Yellow points indicate locations of peak distortion and half recovery. On average 1 practice epoch ~ 400 consecutive syllables (see Methods). (e) Group data of the number of practice epochs to reach half recovery of each syllable, each column is one bird. Free-singing, 149.4 ± 12.4, N = 24 syllables (nested) from 4 animals; Song-prevented, 44.7 ± 29.5, N = 19 (nested) / 3. p < 0.001, nested two-way ANOVA.

The cellular mechanisms and plasticity rules that may contribute to the song recovery

What are the cellular mechanisms by which the brain restores the precise execution of the song after such drastic perturbations? Dual whole-cell recordings between HVC(RA) neurons and interneurons indicated that expression of TeNT leads to permanent muting of the infected cells (Extended Data Fig. 2). Because both LV manipulations (NaChBac and TeNT) permanently perturbed the infected neurons, reducing the number of functioning HVC projection neurons, we hypothesized that the recovery may depend on changes among other neurons that were not genetically perturbed (“unmanipulated neurons”) within the same HVC. We performed whole-cell recordings to measure the synaptic and intrinsic properties of unmanipulated (GFP-negative) HVC(RA) neurons after LV-TeNT. These neurons showed no change in their intrinsic excitability or inhibitory synaptic inputs (Extended Data Fig. 7). However, they recruited a substantially higher level of excitatory synaptic inputs than neurons in unperturbed animals, as revealed by a much higher frequency of mEPSC (193.0 ± 31.5 %. Fig. 5a-c), suggesting that presynaptic mechanisms, such as formation of new synapses, activation of silent synapses, or strengthening of preexisting synapses might be responsible for these changes. Similar synaptic changes were observed in manipulated (TeNT+) neurons as well; however, because TeNT irreversibly muted these cells, it is unlikely that synaptic changes in the manipulated cells contributed to the recovery. To investigate whether these synaptic changes are caused by local perturbation of neuronal activity within the same region, we injected the virus unilaterally just in one HVC. We detected synaptic changes in unmanipulated neurons in the injected hemisphere, but not in the contralateral (unmanipulated) hemisphere (Extended Data Fig. 8). This suggests that the observed synaptic changes are induced by mechanisms local to the HVC that was perturbed.

Figure 5. Population-level homeostatic plasticity and recruitment of silent neurons in a network model contribute to the recovery of sequential activity.

(a) Schematic illustrating whole-cell recordings performed in HVC(RA) neurons in birds injected with LV-TeNT or naive controls (no virus); (b) Example traces of mEPSC recordings; (c) Group data of mEPSCs. “Degraded” indicates recordings at 5 dpi. “Recovered” indicates recordings at 25–35 dpi. mEPSC frequency (mean ± sem): Control, 9.7 ± 1.6 min⁻¹, N = 23 cells / 4 animals; Degraded GFP−, 13.1 ± 0.2 min⁻¹, N = 15 / 3; Degraded GFP+, 20.6 ± 0.3 min⁻¹, N = 16 / 4; Recovered GFP−, 18.8 ± 0.3 min⁻¹, N = 18 / 4; Recovered GFP+, 32.7 ± 0.5 min⁻¹, N = 14 / 4. mEPSC amplitude: Control, 17.2 ± 0.9 pA; Degraded GFP−, 18.3 ± 0.6 pA; Degraded GFP+, 17.1 ± 0.5 pA; Recovered GFP−, 19.5 ± 0.7 pA; Recovered GFP+, 20.8 ± 1.4 pA. One-way ANOVA followed by t-test; (d) Schematic diagram of the neuronal organization in the model; (e) Schematic illustrating the single-cell homeostatic plasticity rule implemented; (f) Spike raster plots showing the sequential dynamics generated by a subset of the modeled HVC neurons before perturbation and after recovery (with only single-cell homeostatic plasticity implemented). Color denotes cell identity; (g) (Left) Plot of the normalized total excitatory synaptic input per neuron against renditions post perturbation. (Right) Scatter and violin (shadings) plot of the normalized total excitatory synaptic input received by each HVC(RA) neuron at three time points; (h) Mean change in the time of the first spike of HVC(RA) neurons for three example networks (top) and across all networks (below, subsampled), when only single-cell homeostatic plasticity was implemented; (i) Schematic diagram illustrating that initially inactive neurons were recruited into the network; (j) Schematic diagram illustrating that a population-level homeostatic plasticity, governing the summed firing activity of all neurons, was also implemented in the model; (k) Same as panel f but with the population-level homeostatic plasticity implemented. Black denotes cells which were silent before perturbation; (l & m) Same as panels g & h, but with population-level homeostasis implemented.

We also tracked changes in interneuron synaptic inputs over time. The excitatory synaptic inputs received by HVC inhibitory neurons in LV-TeNT animals were initially greatly reduced, consistent with a reduction of the synaptic output from the projection cells expressing TeNT. However, these excitatory inputs to inhibitory neurons eventually recovered to a level that was indistinguishable from control animals (Extended Data Fig. 8g), suggesting that the recovery of the inhibitory tone was due to changes in unmanipulated neurons. Thus, our findings show that alongside song degradation and following recovery, the HVC network is dramatically restructured.

We used modeling to explore how synaptic plasticity mechanisms could contribute to unsupervised recovery of network activity after perturbation and to account for the observed synaptic changes. Song recovery in animals that were prevented from singing suggests that plasticity initially occurs largely locally within HVC rather than through behavioral feedback from practice. We sought to determine which mechanisms are consistent with restoration of the sequence as well as with our physiological findings about synaptic strength changes. We modeled HVC as an excitatory-inhibitory (E-I) network, with HVC(RA) neurons connected to each other in a feedforward, polychronous chain and recurrently connected to interneurons (Fig. 5d)^12,26,27. We then inactivated varying fractions of the HVC(RA) population (Fig. 5d) to mimic the silencing of neurons caused by TeNT expression. To explore mechanisms that may enable recovery, we first implemented spike-timing dependent plasticity (STDP) between excitatory neurons (E→E), which is integral to many models of sequence self-organization^26,28–30, and a form of downward firing rate homeostasis, which scaled down an HVC(RA) neuron’s afferents if its activity exceeded a threshold (see Methods, Eq. 1). STDP would not be expected on its own to rebuild the feedforward sequence once broken, as this learning rule requires activity in both pre- and postsynaptic neurons to strengthen synapses. It is possible that the addition of noise can drive random activity that may allow rebuilding of the sequence. We found that STDP can indeed restore sequential activity with the help of independent random inputs across the network, but that this restoration only occurs for weak perturbations (Extended Data Fig. 9 & 10a14a). Further, this mechanism does not produce the increase in excitatory connectivity relative to the level before perturbation in HVC(RA) neurons observed by our electrophysiological measurements (Extended Data Fig. 10b-d).

As STDP alone was insufficient to restore the sequential dynamics in our model of HVC, we next considered whether cell-autonomous homeostatic mechanisms may enable recovery of the network, as was previously proposed to account for recovery from lesioning in Nif, a nucleus upstream of HVC¹¹. While homeostasis may act either on intrinsic excitability or synaptic inputs, our recordings showed that unmanipulated HVC(RA) neurons did not show changes in intrinsic excitability post-perturbation (Extended Data Fig. 7), but rather in their synaptic inputs. Thus, we added a cell-autonomous homeostatic rule into our model, based on scaling excitatory synaptic inputs to individual excitatory neurons to maintain their firing rates (Fig. 5e). The implementation of this rule is consistent with reports that have found activity-dependent synaptic homeostasis in other circuits with sequential dynamics³¹. We found that adding this form of cell-autonomous synaptic homeostasis in excitatory neurons reliably restored sequential network activity, and that individual cells fired at close to their original times in the sequence (Fig. 5f-h and Extended Data Fig. 10e). In addition, implementing STDP between excitatory and inhibitory neurons (E→I) enabled the rebound of input to inhibitory neurons observed in our experiments. However, in models employing STDP alone or both STDP and cell-autonomous synaptic homeostasis, the recovered HVC outputs remained reduced in proportion to the percentage of neurons inactivated, weakening accordingly the drive to downstream regions (Extended Data Fig. 10b,f). Additionally, these models did not reproduce the relative increase in excitatory synaptic input to unmanipulated excitatory neurons as revealed by our experiments (Fig. 5c,g and Extended Data Fig. 10c,g).

A potential mechanism by which HVC might restore the strength of its output to downstream targets is the recruitment of neurons that initially do not participate in the sequential dynamics, defined here as “silent” neurons (Fig. 5i). Multiple experiments have indicated that a large fraction of HVC(RA) neurons do not fire during song, suggesting a possible redundancy ^5,32,33. We hypothesized that the presence of such silent neurons--assumed to be HVC(RA) neurons connected within the network, but whose inputs are subthreshold during song--might provide additional resilience by allowing the sequential dynamics to be partially carried by newly recruited HVC neurons when active constituents of the network fail. While such shifts in participation could be due to the loss of inhibition onto silent cells following the loss of excitatory neurons, this picture of recovery is inconsistent with our experiments in that it does not require the observed increase in the excitatory inputs onto HVC(RA) neurons. We therefore hypothesized that silent neurons may be recruited into the sequence through a form of homeostasis that is sensitive to the activity of the population, for which there is an emerging body of support ^34–37. We modeled one potential mechanism by which such population homeostasis can be achieved: synaptic scaling based on the activity-dependent release of a secreted factor (Fig. 5j) such as BDNF and TNFα, both of which have been shown to regulate local network activity in a non-cell-autonomous manner ^38–41. The recruitment of previously silent neurons by population homeostasis enabled the most complete recovery. The total activity of the network recovered to a greater degree than models without a form of population homeostasis (Extended Data Fig. 10i-k), and the dynamics of the sequence in terms of numbers of participating neurons and their temporal resolution most closely resembles the state before perturbation (Extended Data Fig. 10l). Furthermore, consistent with our experiments, E→E synaptic inputs increased by ~100% (Fig. 5k,l). This increase was primarily accounted for by strengthened inputs to silent neurons (Supplementary Fig. 5). Finally, the recruitment of silent neurons (which may be either mature cells already present or new neurons produced during adult neurogenesis⁴²) may initially add noise to song production (Fig. 5k), consistent with the observation of an increase in the song’s variability, peaking at ~10 dpi, approximately 4–6 days after the acoustic features of the songs were maximally degraded (Fig. 2g). Despite the addition of new neurons into the sequence dynamics, the timing of the originally active neurons was preserved after recovery, and timing jitter did not increase beyond that displayed by our model without silent cells (Fig. 5m). This is consistent with the reemergence of timing of within-syllable acoustic structure immediately after song prevention (Fig. 3e-h and Extended Data Fig. 5). The participation of new cells in the sequence could also contribute to noise in the acoustic structure and account for the deviations in acoustic structure observed immediately after song prevention, before the bird is able to use auditory feedback to tune the new HVC-RA synapses.

Saltatory recovery of syllable duration

We can further use our model to explore predictions for the recovery process. The core of the model is that the sequence is reliably restored through a parallel homeostatic process, in contrast to the serial regrowth one would expect in a timing-dependent process as demonstrated in previous modeling of HVC sequence organization ²⁶ (Fig. 6a and Extended Data Fig. 9a-b). A prediction of the model is that the sequence should recover in an abrupt fashion: all links are repaired simultaneously, so that by the time any broken link is restored, all other links tend to have recovered (Fig. 6b-d). STDP alone can also lead to abrupt recoveries when HVC neurons receive parallel noisy input, but as noted above, for experimentally constrained noise levels (see Methods), these recoveries typically occur for weak perturbations (Extended Data Fig. 9c,d) and are less common than abrupt recoveries under a homeostatic mechanism (Supplementary Fig. 6). Thus, we find that homeostasis is a generally more robust mechanism of recovery. In our experiments, we found many examples of such saltatory recovery in the song, whereby syllables show a bimodal temporal distribution (Fig. 6e-h and Supplementary Fig. 7), with syllables rapidly alternating between two durations from rendition to rendition until eventually they permanently recover their original duration. Syllables with these bimodal temporal distributions are only present in renditions that include song segments that bear acoustic resemblance to the original syllables. This suggests that a “skeleton” of redundant feedforward connections persists, allowing the sequence to reemerge when all links are repaired. Thus, the skeleton can act as a scaffold upon which the homeostatic plasticity can rebuild the dynamics in an unsupervised manner. As this redundancy can exist very ‘locally’ in the chain (i.e. only between adjacent links), it is still consistent with an extremely sparse network ^5,33, and may be a key contributor to resilience in HVC.

Figure 6. Saltatory recovery of syllable duration.

(a) Schematic diagrams showing two types of potential circuit recovery mechanisms. (Left) Serial - the recovery of sequential firing requires building of feedforward synaptic chains step by step from the first breaking point. (Right) Parallel - all links in a sequence are repaired simultaneously; (b) Plot of the normalized duration of modeled sequential firing against number of renditions. Note that the recovery of the full syllable duration occurred by a sudden leap from a shortened state; (c) Example raster plots showing the sequential spiking of modeled HVC(RA) neurons, picked from multiple time points indicated by red dots (shortened sequence) or triangles (full sequence) in panel b; (d) Probability density distributions of the durations of the modeled syllables at different times, ordered chronologically. Note the bimodal distribution of the duration of a single syllable during the recovery phase; (e) Example spectrograms of one syllable at multiple days before and post injection of LV-TeNT. Red arrows mark the shortened/truncated syllables; (f) Normalized distributions of the duration of syllables closest to original, ordered so that each row of panels in e and f is from the same day. Note the bimodal distribution of durations for a single syllable found during the period of song degradation and recovery; (g) Plots of the mode of the duration of the modeled sequences against rendition epochs, showing that the recovery of sequence duration occurs in a saltatory, rather than in a continuous, manner; (h) Plots of the mode of duration of actual song syllables, which show a saltatory recovery similar to that predicted by the model. The curve marked with a star is made from the syllable shown in panels e and f. 5/9 (freely singing) and 4/10 (prevented singing) syllables analyzed showed saltatory recovery patterns. Note that many syllables are already close to recovery in the song-prevented birds (see Methods and Supplementary Fig. 7).

DISCUSSION

Here we have observed the largely unsupervised restoration of a complex learned behavior following large-scale neuronal perturbation in a brain circuit essential for the production of the behavior. We perturbed neuronal activity using two distinct methods in order to probe general mechanisms of resilience in response to the multiple ways in which circuits can be disrupted. We perturbed neuronal activity by increasing neuronal excitability by NaChBac, and we blocked neurotransmitter release by TeNT. The behavioral recovery we observed was accompanied by local synaptic reorganization that can be modeled through proposed population-level homeostatic plasticity. This model recapitulates the experimental finding that restoration of sequential firing can happen largely offline (without feedback from practice) to result in song recovery, pointing to contributions from unsupervised, circuit-level reorganization. Such an unsupervised recovery of basic sequential dynamics may be necessary to scaffold feedback-dependent retuning ⁴³.

Is practice required to advance song recovery? As we showed in Results, much of the timing structure of songs was restored rapidly even without practice and significantly restored motifs were produced immediately following prevention. The remaining timing precision and acoustics were resolved over a subsequent 2–3 days of practice. This is in contrast to the freely singing birds’ recovery trajectory in which song is restored after 10–14 days of practice (Supplementary Fig. 4 & Fig. 4d,e). Together these results show that a significant number of practice renditions are not required to arrive at an equivalent state of recovery, indicating that unsupervised processes contribute to recovery. However, our findings suggest that even though the general dynamics in HVC and the backbone structure of song syllables can re-emerge without feedback, practice is still essential to re-establish the reliability and consistency of singing behavior. Additional mechanisms to completely restore the song might include further activity-dependent refinement of HVC sequential activity, regaining of muscle tone following song prevention ⁴⁴, and a potential involvement of auditory feedback through the anterior forebrain pathway to retune connections between HVC and RA.

Song recovery was accompanied by increased excitatory synaptic inputs to unmanipulated neurons within the same brain region. Based on these findings we modeled sequential dynamics to test how plasticity mechanisms can support this restoration. Single-cell homeostasis has previously been proposed as a mechanism for recovery¹¹ following song degradation after lesion of Nif, a nucleus of the song system that provides input to HVC. Here, we demonstrated through modeling that the further inclusion of population-level homeostasis is needed to account for the experimentally observed synaptic reconfiguration. This form of plasticity confers an additional mechanism for resilience as it enables the recruitment of HVC(RA) neurons that did not initially participate in song dynamics, defined here as “silent neurons”, providing a potential role for these neurons; neither STDP nor single-cell homeostasis can recruit these neurons into the active dynamics. The model predicts that the set of HVC(RA) projection neurons active in song would shift substantially following perturbation. While HVC activity is generally stable⁴⁵, such population-level activity regulation could support naturally occurring “representational drift”, which has been observed in several brain regions wherein the qualitative nature of the sequential dynamics is preserved but the neurons that participate in the dynamics change^46–50, potentially increasing circuit robustness^51,52. Our model thus suggest an important role for silent neurons in the maintenance and resilience of behavior after large scale perturbation of a brain circuit. We acknowledge that a population-level homeostasis and dynamic recruitment of silent neurons are not the only possible explanations for our experimental observations. For example, one alternative that may account for the increases in synaptic strength beyond the original level is that even when upregulated by standard homeostatic plasticity, newly recruited synaptic contacts could sum sub-linearly if they accumulate in the vicinity of the existing ones. It will be valuable in the future to test these predictions using imaging or recording methods that can reliably track the activity of individual HVC(RA) neurons in recovering animals over several weeks.

In this work we focused on the behavioral and electrophysiological consequences resulting from perturbations that resulted in the permanent and irreversible inactivation of neurons within HVC. In future experiments it would be informative to explore the resilience of HVC in response to reversible manipulations with optogenetic or chemogenetic methods. In particular, it would be interesting to explore how HVC responds to temporally controlled perturbations, such as during sleep⁵³, or selectively disrupting individual syllables using optogenetic stimulation or silencing. In addition, here we have focused our analysis on local changes that occur in HVC, the same nucleus where we perturbed neurons. In the future it will be important to investigate the contributions of other song nuclei to the completion of recovery, including changes in RA, the nucleus immediately downstream from HVC, and nuclei from the anterior forebrain pathway, which have been previously shown to be important for some forms of adult plasticity^54–56.

HVC has been shown to drive song with bilateral coordination ^57,58, raising the question of how the laterality of perturbations influences their impact. Recently Otchy et al. showed a relatively rapid recovery of song (~3 days) following unilateral perturbation of inputs to HVC via Nif ¹¹. It is possible that this rapid recovery was because the perturbation affected only one hemisphere. However, we find that with unilateral manipulation of the HVC circuit, the nature and severity of song degradation is comparable to a bilateral manipulation, but the homeostatic cellular level changes only occur in the perturbed hemisphere (Extended Data Fig. 8). This supports the conclusion that directly perturbing HVC may require a deep reorganization of circuitry within HVC, which we observed in our measurements of synaptic restructuring, further supported by our model’s need to introduce additional mechanisms beyond single-cell homeostasis.

The unsupervised recovery of a complex learned behavior after neuronal perturbation described here indicates that self-organized restoration of sequential dynamics may be key to enable brain circuits to support resilient behavior in the face of perturbations. These observations may have important implications towards understanding the mechanisms by which the brain ensures long-term preservation of information and the continuity of behavior and improves restoration of neurological function after disease or injury.

METHODS

Animals

All procedures involving zebra finches are approved by the Institutional Animal Care and Use Committee of the California Institute of Technology. All birds used in the current study were bred in our own colony and housed with multiple conspecific cage mates of mixed sexes and ages until use for experiments. Before any experiments, adult male birds (>120 days post hatch (dph)) were single housed in sound isolation cages with a 14/10 hr light/dark cycle for >5 days until they were habituated to the new environment and started singing. Thereafter, birds were kept in isolation until the end of the experiment.

No statistical methods were used to pre-determine the number of animals used for each different set of experiments, but our sample sizes are similar to those reported in previous publications ^9,10,16,17. Animals were assigned to various experimental groups randomly.

Viral vectors

Lentiviral vectors were cloned using standard procedures and were produced and titrated as described previously⁵⁹. All LVs contained the internal Rous sarcoma virus (RSV) promoter driving expression of different transgenes. LV-TeNT encoded the light chain of tetanus toxin fused to EGFP with a PEST domain in its C-terminus. LV-NaChBac encoded the open reading frame of NaChBac fused to EGFP.

Stereotaxic injection

Birds were anesthetized with isoflurane (0.5% for initial induction, 0.2% for maintenance) and head-fixed on a stereotaxic apparatus. To inject a retrograde tracer in area X or RA, craniotomies were made bilaterally and fluorescent tracers (cholera toxin b 555 0.2%, fluoro-ruby 10%, fluoro-gold 4%, or red RetroBeads, 100–300 nL) were injected through a glass capillary (tip size ~25 μm) into the corresponding nuclei (coordinates from dorsal sinus in mm - area X: Anteroposterior (AP) 3.3–4.2, Mediolateral (ML) 1.5–1.6, Deep (D): 3.5–3.8; RA: AP 1.5, ML 2.4, D 1.8–2.1). To deliver virus into HVC, a second surgery was performed 7–10 days after retrograde tracer injection, by then HVC was strongly labeled by fluorescence and visible through a fluorescent stereoscope. Because LVs only diffuse a short distance from the injection site (~100–200 μm), they were injected into multiple locations (up to 16 sites per hemisphere, ~100 nL each) to deliver the transgenes into as many cells as possible throughout HVC. All injections in HVC were performed ~20 nL/min to minimize physical damage. At the end of every surgery, craniotomies were covered with Kwik-Sil and the skin incision was closed with Gluture.

Song analysis

Song analysis was performed using Matlab (Mathworks) without blinding to the conditions of the experiments. Songs from different conditions of experiments were equally collected and processed with the same algorithms. No animals or data points were excluded from the analyses.

Song feature parameterization

Continuous audio recordings (44.1 kHz) were segmented into individual bouts manually and filtered to remove low frequency noise using the Matlab function, ‘highpass’ (cutoff frequency = 500 Hz). We used the open source Matlab software package, Sound Analysis Pro 2011²⁴ to generate spectrograms and derive non-linear, time-varying song features. The time-varying song features that we studied were: pitch, goodness of pitch, Wiener entropy, frequency modulation (FM), amplitude modulation (AM), amplitude, aperiodicity, mean frequency, and the maximum power in each quadrant of the frequency range 0–11 kHz (labeled power 1, power 2, power 3 and power 4). These features were computed across the entire bout every 1 ms. These time-varying features were the base from which various portions of the song were further characterized and were used for intra-syllabic analysis in Fig. 3 and Extended Data Figure 5.

Syllable parameterization

A high-dimensional, acoustic parameterization of individual syllables was generated by sampling from a moving average (over 10 ms windows) of individual SAP song features (pitch, goodness of pitch, mean frequency, and entropy) at 10 points across the first 50 ms of each syllable (schematic in Fig. 1b). This resulted in a 40-dimensional song feature parameterization of each syllable. If a syllable was shorter than 50 ms, the feature vectors were padded at the end with zeros. This method allowed syllables to be compared in the same parametric acoustic space despite differences in duration.

Syllable segmentation

Zebra finch vocalizations can be defined at three hierarchical levels. Syllables are elemental vocalizations with a fixed duration and spectral characteristics, and separated from each other by silences, also of fixed duration. In the adult zebra finch, a group of syllables (between 3 to 7) are produced always in the same order, and this grouping is called a motif. Bouts are defined as a group of several motifs produced without interruption. In order to study the effect of HVC perturbation on the stability of the syllable structure, we identified syllables and silences within each bout by imposing local thresholds on the time-varying log-power of the spectrogram, summed over all frequencies. We selected this parameterization of the time-varying acoustic signal because it generated sharp distinctions between syllables and silences and allowed a largely automated segmentation of syllables and silences. To consistently assign inter-syllable breath sounds as silence and to address minor variations in background noise during recordings across days, thresholds were defined by sampling from the background noise within each bout. To do this, we first applied a power threshold below which a timestep was defined as silence. This initial threshold was chosen such that intervals between syllables in which soft breath sounds occur are categorized as silences. The mean ($\mu$) and standard deviation ($\sigma$) of the log-power within silent regions for a bout were computed over timesteps from the middle portions of the silent intervals greater than 10 ms. A new threshold was defined as: $T=\mu +a\mathrm{*}\sigma$ within that bout. The multiplier, $a$, was selected upon inspection of the stereotyped song structure for each bird and then applied across all recorded bouts across all days. We then performed a smoothing step wherein periods of silence less than 15 ms were converted to song segments. Song segments less than 20 ms were excluded from analysis. Segmentation was further checked by eye by random sampling across both stereotyped motifs and degraded songs. We then applied these parameters to the entire course of song recordings for each bird. All the results involving syllable segmentation and given in the figures used this adapting threshold method. As an alternative, we tried a fixed threshold method, in which the segmenting threshold remains fixed when we switch between song bouts of the same bird. The results yielded were essentially identical when we compared these two methods of syllable segmentation.

A note on terminology: we refer to song segments to indicate elemental units of vocalizations. In the unperturbed songs, these song segments can be unambiguously identified as syllables. In the degraded songs, most of these element vocalizations do not resemble the original syllables, as they do not have a stereotypical song duration or spectral characteristics. Because this is a continuous recovery process, these terms sometimes overlap in our usage, and a vocalization identified as a song segment during the degraded stages can be later on called a syllable, as the song recovers towards its original form.

Syllable feature distributions

Syllable features (duration, log power, and entropy) were extracted from the syllable segmentation and parameterization processes. Distributions of syllable features were computed by normalizing each distribution of all syllable features within individual days such that the sum of each daily distribution over all binned durations equaled one. Distributions for individual days were then assembled into a matrix wherein the columns represented normalized distributions for individual days. This matrix was plotted as a heat map (Fig. 2a-c).

Acoustic distance trajectories

To quantify fluctuations in song, we computed k-nearest neighbor statistics of the acoustic parameterization space of song⁶⁰. We computed the average k-nearest neighbor distances of individual syllables from original, stereotyped syllable clusters in the undistorted song. This distance was calculated in the space of the first 20 principal components of the 40-dimensional, acoustic space described in Syllable feature parameterization. We defined a normative set of syllables from 1–3 days of recording pre-viral injection when the bird was singing undisrupted, stereotyped songs. This syllable set was labeled according to syllable identity. Syllable identity was defined using the Matlab clustering algorithm, dbscan. The dbscan clustering was performed on a reduced 2-dimensional acoustic space generated using the dimensionality reduction algorithm, Uniform Manifold Approximation and Projection (UMAP)^19,61. The syllable assignments were cross checked by visual examination of a randomly selected subset of bouts and found to be in strong alignment with hand-marked syllable assignments. Syllables that did not cluster into distinct groups were excluded from this analysis. Using the Matlab function, knnsearch, we measured the k-nearest neighbor distances for every syllable over the course of the experiment to each normative syllable cluster (k = 25, euclidean distance metric) and took the average over these distances.

As a measure of song degradation, we tracked the acoustic distance of the syllables which most closely resemble our original syllable set, defined within a local period of singing. We defined acoustic distance to normative syllable clusters as the 0.025 quantile of the average k-nearest neighbor distance distribution in sets of 400 consecutively sung syllables (k = 25). The acoustic distance measure in Fig. 2f and 4b,c is an average over quantile measurements of all syllable sets in a single day. When the song is highly distorted, syllables do not cluster into clearly defined groups, nor do they resemble original syllable types. This method allows us to track syllable distances even when syllables are highly distorted. We normalized the acoustic distance trajectories such that the peak acoustic distance was 100% distortion in order to compare the time course of recovery across birds and syllables.

Speed of recovery

We used the acoustic distance trajectories to measure the speed with which the song recovers. We measured the point of half recovery by tracking how much time and practice respectively are required after perturbation to achieve a 50% recovery to the pre-perturbation baseline. 50% recovery is calculated relative to the point of maximum syllable distortion for each syllable individually. The acoustic trajectory of recovery is described in the above section. This measurement is shown for the speed of recovery as a function of practice in Fig. 4d,e. Each point in the practice trajectories represents a single singing epoch of on average 400 consecutive syllables. Each practice epoch was confined to a single day. On days in which fewer than 400 syllables were sung, the entire day of song was taken to be a single practice epoch. We chose to confine epochs to single days so that (1) the distribution within each epoch would be approximately stationary and (2) epochs would not span pre-song prevention and post-song prevention days which could erase the measure of maximal distortion. We also computed this practice trajectory using epochs of 50 instead of 400 consecutive syllables, spanning days, and found comparable results (Practice epochs: free singing, 878.0 ± 252.2 vs. prevented singing, 280.7 ± 113.5. p < 0.001, nested two-way ANOVA.).

Song variability

As a measure of song variability, we tracked the local variability of syllables to other syllables that have been sung within a consecutive 400 syllable window (Fig. 2g). We quantified local variability as the average k-nearest neighbor distance of each syllable to the closest 5 syllables within the local 400 syllable window. This measure of variability quantifies how different renditions are from one another, not how similar they are to the original song.

Continuous representation of bout trajectory

We generated continuous visualizations of bouts across the entire perturbation trajectory as shown in Figs. 1b,d,f; 4a, and Extended Data Fig. 4 ^20,21. We randomly sampled 100 bouts from each day of recording to build a representative sample of the song over the course of the experiment. For each bout, we slid a 150 ms window in 3 ms steps along the bout length. We then generated a high-dimensional, acoustic parameterization of each 150 ms song window by taking the moving average in (across 20 ms windows segments every 5 ms) of seven song features (mean frequency, pitch, goodness of pitch, power 4, power 2 and summed log power of the spectrogram). We performed principal component analysis on this high-dimensional set across all bouts to normalize and reduce the feature set to 30 dimensions and applied the UMAP algorithm to project this high-dimensional representation into two dimensions^19,61. The choice of parameters was made empirically to generate the clearest visual representation of the continuous bout trajectory^20,21. This visualization gives a qualitative picture of the structure of the bout over a large set of renditions across the course of the degradation and recovery period. However, because the UMAP reduction algorithm makes nonlinear shifts in the distances between points across and within bouts, we do not use this visualization for quantitative analysis.

Duration distributions of best syllables

We tested whether the degradation and recovery duration pattern we found in our feedforward computational models after silencing a subset of neurons was present in our data. To do this, we considered the duration distributions of song segments over the course of the perturbation and recovery in our TeNT experiment that most closely resembled the original stereotyped song (defined as the 2.5% of song segments on each day that were closest to the original stereotyped syllable, measured using the k-nearest neighbor distance metric described above). In this analysis we excluded stereotyped syllables which were either less than 100 ms in duration, never recovered, significantly altered their acoustic form, or did not segment cleanly into discrete clusters. In 6/7 birds, we found syllable recovery trajectories that contained bimodal or multimodal duration distributions during days with degraded song (5/9 syllables analyzed in the TeNT freely singing experimental group and 4/10 syllables analyzed in the TeNT song prevention experimental group). See Supplementary Fig. 7 for all distribution trajectories.

Dynamic time warping

A robustness feature of our computational model is that unsupervised recovery of sequential activity occurs with minimal time distortion. When we examined the duration distributions of the best syllables (defined in the previous section) in the recovered songs immediately post-song prevention, we found this to be in strong agreement with the duration distributions of the best syllables before HVC perturbation. However, because the k-nearest neighbor distance metric was performed on a song feature space sampled in the original time basis of the acoustic recordings, there remained the possibility that syllables which were acoustically intact but strongly dilated in time were being assigned large distances by our acoustic feature distance measure, and thus we might overlook high quality acoustic syllable renditions that were warped in time. To check this, we implemented a second measure of syllable similarity using a dynamic time warping algorithm (DTW). DTW finds a nonlinear stretching transformation of two time-varying signals such that the Euclidean distance between corresponding points in the two signals is minimized. We used DTW to compute a second measure of syllable quality. We defined a new DTW distance as the average distance per time step of each song segment to 25 normative syllables of a particular type pre-HVC perturbation after the song segment had been aligned using DTW. The DTW distance was computed directly from the log spectrogram matrix of each song segment and syllable. The best syllables by this measure are again the syllables with the smallest distance.

Song prevention

Adult zebra finches 120–150 dph (N = 3) were fitted with custom made fabric vests and a 20–40-gram weight attached to their vest, pulling them towards the ground to prevent them from acquiring singing posture. All birds received LV-TeNT injection bilaterally into HVC. Before we prevented the birds from singing, we allowed them to sing a few renditions to confirm that their songs were degraded. Afterward, the birds were restricted from singing for ~10 days, during which they were monitored with a video camera to make sure they did not sing, although they could make calls. We confirmed that all the birds were able to move, perch, drink and eat freely and even allowed them to sing several renditions occasionally (1–3 every 2–3 days, which allowed us to track the degradation/recovery of songs). During the prevention phase, animals were continuously monitored by a trained experimenter via video camera during the light cycle. The experimenter observed and assessed the animals’ behavior, including their ability to eat, drink, and move freely without experiencing stress. If any animal displayed signs of high stress, the experiment was immediately halted, and the affected animal was removed from the study. The use of a live stream camera facilitated minimal stress on the animals and prevented any negative impact on their health. The experimenter maintained detailed records of each animal’s behavior, and the song prevention procedure was conducted on a batch of two animals per 10-day period to allow the experimenter to devote their full attention to the animals’ welfare. The size of the bullet weight on the birds had to be adjusted since animals got accustomed to the weight in 2–3 days and attempted to sing more frequently. The weights and vest were removed daily, 1 hour prior to the light-off period while the experimenter stood next to the chamber to closely monitor that the birds were not singing. The vest and weights were put on the birds again as soon as the lights were turned back on the next day. After the prevention period, the birds were free to sing in their respective isolation chambers. The song prevention procedures have been documented in our IACUC protocol (No. 1719) with the title song prevention (SOP #6). Please contact Caltech IACUC for further information or records.

Electrophysiology

Birds were first overdosed by an intramuscular injection of ketamine/xylazine (120/12 mg/kg) and after they became unresponsive to toe pinching, they were decapitated. The forebrain was quickly removed and kept in ice-cold slicing solution (in mM: sucrose 213, KCl 2.5, NaH₂PO₄ 1.2, NaHCO₃ 25, Glucose 10, MgSO₄ 2, CaCl₂ 2, pH 7.4). Sagittal slices (300 μm) were cut using a vibratome (Leica VT1200S) and then incubated in HEPES holding solution (in mM: NaCl 102, KCl 2.5, NaH₂PO₄ 1.2 NaHCO₃ 30, HEPES 20, Glucose 25, MgSO₄ 2, CaCl₂ 2, pH 7.35) at 34.5 °C for 30 minutes. Afterwards, slices were kept at room temperature (~ 22 °C) between 30 minutes to 5 hours before being moved to the recording chamber. Bath ACSF (in mM: NaCl 124, KCl 2.5, NaH₂PO₄ 1.2, NaHCO₃ 26, Glucose 25, MgSO₄ 1, CaCl₂ 2, pH 7.35, 33–34°C) was continuously perfused (~2 mL/min) during recording. For current clamp whole-cell recordings, glass pipettes were filled with an intracellular solution (in mM): K-gluconate 135, MgCl₂ 3, HEPES 10, EGTA 0.2, Na₂-ATP 2, phosphocreatine 14, pH 7.25. HVC(RA) cells were identified by the presence of the retrograde fluorescent tracers injected into RA.

Dual whole-cell recordings of HVC(RA) and interneurons were made between cells that were less than 100 μm apart. Interneurons were identified by their relatively big soma, not having the retrograde fluorescent tracer, and spontaneous action potential firing⁶².

Data collections were not performed blind. We used voltage clamp to record miniature EPSCs with the following chemicals added to the bath (in μM): TTX 0.5, nimodipine 5 and picrotoxin 50, and glass pipettes were filled with (in mM) Cs(CH₃)SO₃ 135, MgCl₂ 2, HEPES 10, EGTA 0.2, QX-314.Cl 5, Na₂-ATP 2, phosphocreatine 14, pH 7.25. To record mIPSCs, the following chemicals were added to the bath (in μM): TTX 0.5, nimodipine 5, CNQX 10 and APV 25, and pipettes were filled with (in mM) CsCl 120, K-Gluconate 12, MgCl₂ 2, HEPES 10, EGTA 0.2, QX-314.Cl 5, Na₂-ATP 2, phosphocreatine 14. For whole-cell NaChBac current recording, the bath solution did not contain any CaCl₂ to eliminate Calcium currents, TTX and 4-AP were added to block currents from the endogenous Na+/K+ channels, and glass pipettes were filled with (in mM) CsCl 135, MgCl₂ 3, HEPES 10, EGTA 0.2, TEA-Cl 2, Na₂-ATP 2, phosphocreatine 14. We only analyzed recordings in which access resistance was always smaller than 10 percent of the membrane resistance of the cell and no compensation was applied. Membrane potential was held at −70 mV to measure mEPSCs and −60 mV for mIPSCs.

To examine the evoked EPSCs in RA(PN) cells, we cut para-sagittal brain slices at a 20–25° angle relative to the interhemispheric fissure⁶³. We placed a clustered bipolar electrode (FHC, CE3C55) on the axonal bundle going from HVC to RA, 50–100 μm away from the RA border. Brief electric stimuli (0.1 ms duration, <0.1 Hz) were generated by an isolator (A.M.P.I., ISO-FLEX). RA(PN) cells were identified by their tonic action potential firing and spiny dendrites (visualized by filled Alexa dye)⁶⁴.

Liquid junction potential was not corrected. Electric signals were amplified and sampled at 20 kHz by an EPC-10 system (Patchmaster, HEKA). Data analysis was performed off-line using Fitmaster (HEKA), Mini-Analysis (Synaptosoft) and Matlab (Mathworks). Data distribution was assumed to be normal, but equality of variances was tested using F-test. Data were presented as mean ± s.e.m. Statistical difference was tested using one-way ANOVA followed by student’s t-test. No data points were excluded.

Histology

After the experiments were concluded, animals were sacrificed, and their brains were processed for histological analysis. Animals were first deep anesthetized by intramuscular injection of ketamine/xylazine (100/10 mg/kg) and perfused intracardially with room temperature 3.2% PFA in 1xPBS. Brains were then extracted and incubated in the same fixative for 2–4 hours at room temperature. Each brain hemisphere was cut sagittally with a vibratome into 70–100 μm thick sections. The brain slices containing HVC were collected and incubated overnight with a rabbit anti-GFP antibody (1:500. EMD Millipore, AB3080P) in 1xPBS containing 10% donkey serum and 0.2% Triton at 4 °C. Sections were washed in 1xPBS with 0.05% Triton and incubated for 2 hours at room temperature with a secondary antibody (1:1000. Abcam, ab150077). Brain slices were washed and mounted in Fluoromount (Sigma). Confocal images were taken with a LSM800 microscope (Zen Microscopy Software, Zeiss). To validate the cell-type specificity of our LVs, we performed counterstaining in a subset of the brain slices, using known markers of inhibitory neurons, specifically anti-parvalbumin (1:500. Abcam, ab11427), anti-calretinin (1:500. SWANT, 7697) and anti-calbindin (1:500. SWANT, CB-300) (Extended Data Fig. 1b). Out of 1000 counted neurons, only one GFP+ cell was double labeled with inhibitory markers. To quantify the percentage of projection neurons that were infected by our LVs (Extended Data Fig. 1d,e), we first injected tracers with different colors in area X (fluororuby (red)) and RA (flurogold (yellow)), followed by injection of LVs expressing GFP (green) into HVC. Around 7 days after LV injection, we sacrificed animals, cut their brains into sagittal slices 100 μm thick, and processed them for routine immunocytochemistry. We used a confocal microscope to image 160 × 160 × 15 μm stacks in every slice that contained retrogradely labeled HVC cells and counted all the cells in those image stacks. We observed some variability in the number of HVC cells labeled by LVs, that could be due to several factors, including differences in the size of HVC between animals, or to variability between the injection (Extended Data Fig. 1e). To quantify the singing activated HVC cells, birds with LV-NaChBac for 4 days or 28 days were first let sing freely for 2 hours after lights turned on in the morning, and then sacrificed and processed for routine double counterstaining for GFP and the immediate early gene ZENK (1:500. Santa Cruz, sc-189).

Modeling

We used network modeling to explore the role of different plasticity mechanisms.

Leaky Integrate and Fire Neurons

The membrane potentials $V_j$ of neurons in all networks were modeled using leaky integrate-and-fire dynamics:

$$C_m\frac{d{V}_j}{dt}=g^{(l)}(E^{(l)}-V_j)+g^{(i)}(E^{(i)}-V_j)+g^{(e)}(E^{(e)}-V_j)+I_{\text{ext}}$$

where $C_m=1\mu \frac{F}{c{m}^2}$ is the membrane capacitance, $g^{(l)},E^{(l)},g^{(e)},E^{(e)},g^{(i)},E^{(i)}$, the leak, excitatory, and inhibitory conductances and reversal potentials, and $I_{\mathit{ext}}\sim (0,\sigma =0.1nA)$, a white noise current input. For excitatory neurons, $g^{(l)}=0.25\frac{mS}{{cm}^2}$ and $E^{(l)}=-70mV$; for inhibitory neurons, $g^{(l)}=0.4\frac{mS}{{cm}^2}$ and $E^{(l)}=-53mV$. For all neurons, $E^{(e)}=0$ and $E^{(i)}=-90mV$. When $V_j$ reaches threshold, $V_{\mathit{th}}=-43mV$, the neuron spikes, and the voltage is reset to $V_{re}$ ($-65mV$ for excitatory, $-53mV$ for inhibitory) after a refractory period $t_r=1ms$. The excitatory and inhibitory conductances are driven by incoming spike trains, represented by delta functions at times ${t_k}^{(s)}$, where $s$ indexes all spike times for an upstream neuron $k$, filtered with a time constant, $\tau =4ms$:

$$\tau \frac{d{g}_j(x)}{dt}=-{g_j}^{(x)}+{\sum }_k{w}_{kj}{\sum }_s\delta \left(t-{t_k}^{(s)}\right).$$

In all modeling, we assume that neurons are connected with a certain synaptic strength, which does not differentiate between the number and strength of individual contact sites between a pair of neurons.

Network Architecture

We explored two different network architectures and the resulting dynamics. In the first network, we assumed all cells participate in the sequential dynamics. In the second network, we assumed that only a portion of the excitatory cells initially participate in the dynamics ^12,33. To implement this, we assumed each neuron is silent with probability $p_s=0.4$.

HVC Network

Following recent work, we modeled HVC as a feedforward, polychronous network^27,65. The network is composed of 200 excitatory (E) and 50 inhibitory (I) neurons. In order to define the approximate difference in firing times between any neuron pair, we assign a coordinate $c_i$ to each E cell, where $c_i$ is the $i^{\mathit{th}}$ sum of random variables uniformly distributed over [0, 1]. Pairs of active excitatory cells are initially connected with a fixed weight $W=4e-\frac{5}{\left(1-p_s\right)}$ if the difference in their indices $c$ is greater than zero and less than $c^{\mathrm{*}}=10$, ensuring feedforward propagation. Silent cells are assumed to be connected within the network but with much lower probability, with no sensitivity to sequence order, and with cell-by-cell heterogeneity. Specifically, if one of a pair of neurons is silent, the cells are connected with probability 0.8 with a weight $w_{ij}$ drawn from an exponential distribution with a cell-specific mean $\frac{a_{i}W}{2}$, where $a_i$ is uniformly distributed on $\left[\mathrm{0,1}\right]$. To implement recurrent inhibition observed in HVC ⁶⁶, inhibitory neurons receive connections from excitatory cells with probability $p_{e,i}=0.075$ (0.125 for networks with silent cells) and weight $w_{e,i}=3.5e-4$ (3.5e-5 if the presynaptic cell is silent) and vice versa with probability $p_{i,e}=0.5$ and weight $w_{i,e}=4e-5$.

Following observations that axonal delays between HVC(RA) projectors are relatively long ( 1 – 7.5 ms) ²⁷ and that HVC(RA) projectors typically synapse onto inhibitory interneurons close to their soma and other excitatory cells far from their soma ⁶⁷, we implemented axonal delays in our model that reflected longer E-E axonal delays and relatively shorter E-I and I-E delays. The delays between pairs of active E cells were

$$d_{ij}=3\frac{c_j-c_i}{⟨d⟩}ms,$$

where

$$⟨d⟩=\frac{1}{N_e^2}{\sum }_k{\sum }_l{c}_k-c_l.$$

Here, $k$ and $l$ sum over all active excitatory cell indices and $N_e$ is the number of active excitatory cells. When one of the pair was inactive, the delay was chosen randomly from a uniform distribution on $\left[0,\frac{3c^{\mathrm{*}}}{⟨d⟩}\right]\mathrm{m}\mathrm{s}$. This leads to E→E axonal delays that were 3 ms on average and roughly uniformly distributed. E→I and I→E delays were uniformly set to 0.5 ms. We found that a comparatively fast inhibitory pathway stabilized sequence dynamics by enabling inhibition to respond rapidly to changes in excitation.

Plasticity Rules

We then allowed networks to evolve under both firing rate homeostasis and spiking timing-dependent plasticity, which are evidenced by the synaptic reorganization among HVC neurons shown by our patch-clamp results ^31,68–70. All synapses subject to plasticity were given a lower bound $w_{\mathit{min}}=1e-8$. After each trial, synaptic strength $w_{ij}$ evolved according to

$$\mathrm{\Delta }{w}_{ij}={\beta }_f{\mathrm{\Delta }}_{ij}^f+{\beta }_{pop}{{\mathrm{\Delta }}_{ij}}^{pop}+{\beta }_{STDP}{{\mathrm{\Delta }}_{ij}}^{STDP}$$

where weights are updated due to single-cell firing rate homeostasis $({\mathrm{\Delta }}^f)$, local activity homeostasis $\left({\mathrm{\Delta }}^{\mathit{pop}}\right)$, and $\mathrm{S}\mathrm{T}\mathrm{D}\mathrm{P}({\mathrm{\Delta }}^{STDP})$, respectively, and ${\beta }_f=0.025,{\beta }_{\mathit{pop}}=0.01$, and ${\beta }_{STDP}=1.5e-4$. ${\beta }_f$ was chosen such that firing rate homeostasis could bound potentiation due to STDP. ${\beta }_{\mathit{pop}}$ was chosen to be small so that recruitment of formerly silent neurons would occur slowly. All $\mathrm{E}\to \mathrm{E}$ synaptic strengths had upper bound ${w_{e,e}}^{max}=1e-3$, all $\mathrm{E}\to {\mathrm{I}{w}_{e,i}}^{max}=2w_{e,i}$. For simplicity, we did not introduce plasticity in the inhibitory synaptic inputs to excitatory neurons. We made this simplification because in our electrophysiological data showing the TeNT perturbation’s impact to HVC’s synaptic structure, the average strength of these synapses did not change.

Firing rate homeostasis:

Single-cell firing rate homeostasis moves a neuron’s firing rate toward a set point ${r_j}^{(0)}$ according to:

$${\mathrm{\Delta }}_{ij}^f=w_{ij}\left(a-\theta \left(r_j-r_j^{(0)}\right){\left(r_j-r_j^{(0)}\right)}^2\right)$$ (1)

where $r_j$ is the average firing rate of neuron $j$ in that trial (refer to Fig. 5), $\theta$ is the Heaviside function, and $a=0.1$ is a small constant that mediates synaptic scaling when ${r_j}^{(0)}>r_j$. In our model with STDP and downward firing rate homeostasis and our model with local population homeostasis, we set $a=0$.

Local population activity homeostasis:

Taking inspiration from literature that has shown homeostasis may operate on a network scale, we included in our final model a form of homeostasis that permits individual neurons to monitor and respond to the activity of their neighbors. We implemented here one potential mechanism by which such local population activity homeostasis might be achieved, based on the TNFɑ pathway ⁴¹. We assume each E neuron secretes a chemical factor which diffuses locally in space, the concentration of which is proportional to the neuron’s own activity level. All E neurons are assumed to monitor the local concentration of this factor and adjust their incoming excitatory synapses in order to maintain a target concentration. To implement this form of local population homeostasis, each excitatory neuron was first assigned a location by uniformly sampling the space within a unit sphere. The local factor concentration that an excitatory neuron senses is then

$$m_j=\frac{1}{\sqrt{2\pi {\sigma }^2}}{\sum }_i{r}_{i}exp\left[-\frac{1}{2}\frac{{\left(x_j-x_i\right)}^2}{{\sigma }^2}\right],$$ (2)

where $x_j$ is the vector representing the location of neuron $j$, and $\sigma$ is a parameter controlling the spatial extent of each neuron’s diffuse release (${\sigma }^2=0.03$). The corresponding update to local population homeostasis is

$${{\mathrm{\Delta }}_{ij}}^{pop}=W\frac{e^{\gamma \left({m_j}^{(\mathit{0})}-m_j\right)}-\mathit{1}}{e^{\gamma {\left({m_j}^{(\mathit{0})}-m_j\right)}_{+\mathit{1}}}},$$ (3)

where ${m_j}^{(\mathit{0})}$ is the local concentration setpoint of neuron $j$ and $\gamma =\mathit{0.1}$ dictates the strength of the local population homeostasis near the setpoint. Local concentration setpoints were chosen by computing the average local concentration over 100 activations of the network prior to the perturbation.

Hebbian plasticity:

We found that an antisymmetric pairwise STDP rule with a reasonable time constant (20 ms) did not maintain the relative firing times of a sequence when the spike trains of successively firing neurons overlapped. A burst timing dependent plasticity (BTDP) rule ⁶⁵ that low pass filtered spike trains before applying pairwise STDP was able to maintain relative firing times, but could lead to contraction of the sequence if noise enabled neurons to spike earlier than their typical firing times. In contrast, we found that a triplet STDP rule well maintained the firing times of neurons in a feedforward, excitatory network. The minimal triplet rule introduced by Pfister and Gerstner depends on triplets of spikes (post-pre-post) for potentiation and pairwise interactions for depression ⁷¹. The update for the triplet rule for excitatory neurons was

$${{\mathrm{\Delta }}_{ij}}^{STDP}={\sum }_l{A}_{+}\left(w^{\mathit{max}}-w_{ij}\right){k_{ij}}^{+,1}\left({t_j}^{(l)}\right){k_j}^{+,2}\left({t_j}^{(l)}-ϵ\right)-{\sum }_m{A}_{-}{w}_{ij}{k_{ij}}^{-}\left({t_i}^{(m)}\right)$$

where $w^{max}$ differs for $\mathrm{E}\to \mathrm{E}$ and $\mathrm{E}\to \mathrm{I}$ synapse, $mm$ and $l$ index the spike times of the pre- and postsynaptic neuron, respectively, $ϵ$ is a small positive constant, and the $k$ variables keep track of relative spike timing and implement the eligibility window:

$${{\tau }_1}^{+}\frac{{d{k}_{ij}}^{+,1}(t)}{dt}={-k_{ij}}^{+,1}(t)+{\sum }_m\delta \left(t-{t_i}^{(m)}\right)$$

$${{\tau }_2}^{+}\frac{d{k_j}^{+,2}(t)}{dt}=-{k_j}^{+,2}(t)+{\sum }_l\delta \left(t-{t_j}^{(l)}\right)$$

$${\tau }^{-}\frac{d{k_{ij}}^{-}(t)}{dt}=-{k_{ij}}^{-}(t)+{\sum }_l\delta \left(t-{t_j}^{(l)}\right)$$

The constants regulating the relative strength of potentiation and depression, $A_{+}=5$ and $A_{-}=2$ (25 and 0, respectively, for E→I synapses), and the timescales of STDP, ${\tau }^{-}=33.7ms,{{\tau }_1}^{+}=16.8ms$, and ${{\tau }_2}^{+}=40ms$, were inspired by the parameters given by Pfister and Gerstner in their triplet STDP model of neurons in the visual cortex. We found sequences were most stable when when potentiation due to triplets was implemented in “nearest spike” spike fashion, i.e. ${k_{ij}}^{+,1}(t)$ and ${k_j}^{+,2}(t)$ were bounded by $\left[\mathrm{0,1}\right]$. To stabilize E→I STDP, each inhibitory neuron was assigned a total excitatory synaptic input bound determined by the neuron’s total excitatory synaptic input at the beginning of the simulation. When the bound was exceeded, $\mathrm{E}\to \mathrm{I}$ weight $w_{ij}$ was rescaled as

$$w_{ij}\to \frac{{\sum }_k{w_{kj}}^{(0)}}{{\sum }_k{w}_{kj}}{w}_{ij},$$

where ${w_{kj}}^{(0)}$ represents the size of the synapse prior to the first activation of the network. No plasticity takes place on I→E connections.

Activation of Networks

At the beginning of each trial, networks were active for 10 ms, after which each of the first 10 neurons of the network was driven by an independent burst (4 spikes, 660 Hz), with onset times drawn from a Gaussian (mean 10 ms, STD 1 ms). Input weights were chosen to produce reasonable spiking behavior in the first layers of the network. Networks were simulated for an additional 115 ms following the stimulus. The time step for all simulations was 0.1 ms.

Fluctuating Input

For models in which fluctuating input was supplied to projector neurons, each neuron received independent Poisson input ($\lambda =50Hz$) with input weight $w_p=1.6e-4$. The input rate and weight were constrained by experimental studies in which project cell activity increased a factor <3 when GABA_A antagonist gabazine was locally infused within HVC. We mimicked this procedure in silico by removing inhibition from an individual projector cell and observing the change in activity during activation of the network.

Network Initialization Procedure

To initialize networks, an activation of each network was first simulated once without plasticity. Neurons that fired during this trial were assumed to be active. Active neurons were assigned a uniform firing rate setpoint $\left(r^{(0)}=3\right)$ and were subject to cell-autonomous firing rate homoeostasis as given in Eq. (1) and STDP for 3000 additional activations of the network. For networks with population homeostasis, $\left\{m_i\right\}$, the local secreted factor concentrations were computed for activations 1100–1200, and then averaged to set $\left\{{m_i}^{(0)}\right\}$; from this point, population homeostasis was permitted to act on the network.

Simulated Tetanus Toxin Perturbation of Neurons

Tetanus toxin perturbation of the network was simulated by randomly selecting a cell with probability $p_T$ for perturbation and removing all its outgoing connections over 5 consecutive trials. $p_T$ was chosen such that $1-{\left(1-p_T\right)}^5$ equaled the perturbation percentage reported. Perturbed cells were assumed to not contribute to the population activity level. Networks were then allowed to evolve for 3000 renditions according to the plasticity rules described above. We classified networks as ‘recovered’ if for >80% of renditions 2900 to 3000 (i) each group of 20 contiguous neurons (cell indices 0–19, 20–39, etc.) had at least one active neurons, (ii) the average spike time of cells 0–19 preceded that of 20–39, and so on, and (iii) network activity persisted for >10 ms after the onset of activation (typical network sequence duration was 100 ms initially).

Characterization of graded and saltatory recoveries

To classify recoveries as saltatory or graded, we computed the variance in activation length within a running time window (20 activations) and found the maximum of this trajectory. We reasoned that graded recoveries should produce narrow distributions of activation lengths within short windows of time, whereas saltatory recoveries should contain a short period in which a wide range of activation lengths could be produced. This analysis is presented in Supplementary Fig. 6.

Extended Data

Extended Data Figure 1 – Specific infection of HVC projection neurons by LVs.

(a) Confocal image of a brain slice from a bird injected with LVs, showing the expression of the transgene (tagged with GFP) in HVC neurons. HVC(RA) neurons are labeled by a fluorescent retrograde tracer (cholera toxin b - alexa 555) injected into RA; (b) Confocal images of a brain slice showing that LV selectively target projection neurons, where only 1/1000 of the cells labeled by LV transgene (tagged with GFP, seen in green) overlap the immunofluorescent signal of pooled antibodies against some of the standard markers of inhibitory neurons (PV, parvalbumin/CB, calbindin/CR, calretinin, seen in red and blue). N = 2 animals; (c) Example of serial sagittal slices of HVC, marked by their distance relative to the most lateral side of HVC, showing the efficient labeling of HVC neurons by LV, except for the most medial posterior corner (as demonstrated by the +1300 and +1400 μm images); (d) Confocal images showing the expression of the viral-delivered transgene (tagged with GFP) in the majority of projection neurons (retrogradely labeled with fluorescent tracer, FluoroGold/cholera toxin b, injected into RA and X, respectively); (e) Histogram indicating the percentage of GFP-labeled cells among all projection neurons in HVC. Each pair of bars is generated from one animal. N = 3 animals. Error bars indicate 95% confidence interval. Numbers in each bar are number of identified GFP positive cells divided by total number of cells counted.

Extended Data Figure 2 – Expression of NaChBac in HVC(RA) neurons.

(a) Schematic drawing showing whole-cell patch clamp recordings made in HVC(RA) neurons infected with LV-NaChBac or naive control; (b) Example traces of whole-cell NaChBac currents evoked by depolarizing voltage steps (from −80 to +30 mV. Increment, 10 mV) recorded from HVC(RA) neurons; (c) I-V curve of the NaChBac current, illustrating the peak amplitude of whole-cell current at different step voltages; (d) Comparison of the maximal peak amplitude of whole-cell NaChBac currents recorded at 5 dpi (3.6 ± 0.6 nA) vs. 21–28 dpi (3.0 ± 0.3 nA). Student’s t-test. Sample size is shared between panels b and c, N = 10 cells / 3 animals (control), 15 / 2 (5 dpi), and 19 / 3 (21–28 dpi); (e) Example of current traces with mIPSC events recorded in HVC(RA) neurons expressing NaChBac at different times after injection. The bar called “Degraded” illustrates currents at 5 dpi. The bar called “Recovered” illustrates currents at 25–35 dpi; (f) Group data of the frequency and amplitude of mIPSCs in HVC(RA) NaChBac+ cells. mIPSC frequency: Control, 2.7 ± 0.2 s⁻¹, N = 23 / 4; Degraded, 6.1 ± 0.7 s⁻¹, N = 17 / 3; Recovered, 7.2 ± 0.7 s⁻¹, N = 16 / 5; (g) Example of current traces with mEPSC events recorded in HVC(RA) neurons expressing NaChBac at different time points; (h) Group data of the frequency and amplitude of mEPSCs in HVC(RA) NaChBac+ cells. mEPSC frequency: Control, 9.7 ± 1.6 min⁻¹, N = 22 / 4; Degraded, 6.8 ± 0.9 min⁻¹, N = 14 / 5; Recovered, 5.9 ± 0.9 min⁻¹, N = 17 / 4. mEPSC amplitude: Control, 17.2 ± 0.9 pA; Degraded, 18.5 ± 0.9 pA; Recovered, 16.1 ± 0.5 pA. One-way ANOVA & student’s t-test; (i) Confocal images of an HVC slice stained with antibodies against GFP and EGR-1, an immediate early gene whose expression is induced by singing related activity. Scale bar, 25 μm. Scatter plot in the bottom right represents the percentage of GFP-positive cells that are co-expressing EGR-1. Each dot represents data from one bird. N = 3 (4 dpi) and 4 (28 dpi) animals. Student’s t-test. Error bars represent SEM.

Extended Data Figure 3 – Expressing TeNT in HVC(RA) cells blocked their synaptic output.

(a) Example traces from a dual patch clamp recording in control animals, showing that when action potentials were evoked in HVC(RA) neurons (upper), excitatory postsynaptic currents (EPSCs) can be reliably detected in the connected interneuron (below); (b) Example traces similar to those in panel a, showing that no EPSC could be detected in the interneuron when the HVC(RA) neuron was expressing TeNT; (c) Summary of dual patch clamp results. The amplitude of EPSCs are shown in the scatter plot. The block of synaptic transmission by TeNT expression is long-lasting, as it could be detected even >3 weeks after viral injection, after the song had already recovered; (d) Confocal image showing axon bundles from HVC and their terminal branches within RA labeled by LV-TeNT (tagged by GFP); (e) To evoke post-synaptic responses in RA neurons, a bipolar stimulation electrode was placed in the vicinity of the axon bundles going into RA. Under DIC microscopy, RA is clearly distinguishable by strong phase contrast against the surrounding region. Scale bar, 100 μm; (f) Whole-cell patch clamp recordings were made in RA neurons. Example membrane potential traces in response to depolarizing and hyperpolarizing current steps recorded from one projection neuron in RA (RA(PN)); (g) In RA(PN) neurons, inward post-synaptic current transients were evoked by briefly stimulating axon bundles from HVC with varying intensity, measured by amplitude of stimulus current. Compared with naïve animals (black traces on top), the responses were much smaller in RA(PN) neurons recorded in birds with LV-TeNT in HVC (green traces below); (h) Summarized data of the synaptic responses in RA(PN) neurons. Stimulus – response curves were composed with data obtained at three different time points. The central marks of box plots indicate the medians, the bottom and top edges indicate the 1^st and 3^rd quartiles, and the whiskers indicate the most extreme data points that are still within the 1.5 times of the inter-quartile-region beyond the lower and upper quartiles. Outliers are individually marked. N = 11 cells / 3 animals (Control), 15 / 3 (2 dpi), and 12 / 1 (21–28 dpi).

Extended Data Figure 4 – Songs degradation and recovery was not due to mechanical lesion or inflammation.

(a) Example spectrograms of songs from a bird injected with LV-NaChBac(EtoK), a dead-pore mutant of NaChBac; (b) Distribution of syllable durations per day of the same bird shown in panel a; (c) Plots of scaled acoustic distance to original syllables. Data was from the same bird as in panels a and b. The insets are UMAP visualizations of songs at selected time points. Dashed lines are generated from the syllables of the bird injected with LV-NaChBac shown in Fig. 1d, and for comparison here they were not normalized to maximum. Experiments with LV-NaChBac(EtoK) yielded consistent results, N = 4 animals; (d) Example spectrograms of songs from a bird injected with half of the volume of LV-NaChBac used for animals shown in Fig. 1; (e) Distribution of syllable durations per day of the same bird shown in panel d; (f) Plots of scaled acoustic distance to original syllables. Data was from the same bird as in panels d and e. Each line/color represents one syllable. The insets are UMAP visualizations of songs at selected time points. Experiments with reduced volume of LV-NaChBac were repeated, N = 2.

Extended Data Figure 5 – Recovery of fine intra-syllable structure without practice.

(a-d) Individual curves of time-varying acoustic features, including mean frequency (a), entropy (b), pitch (c), and goodness of pitch (d), generated from the same syllable as shown in Fig. 3g&h. Each black curve represents one rendition of the syllable, and the red curve represents the mean. The population of renditions shown are the 2.5% of renditions closest to the original syllable cluster in a given day. (e-j) More example syllables to show intra-syllable acoustic structures on the first day after song prevention was lifted. Four time points were selected for comparison, pre-perturbation, pre-prevention, first day post-prevention, and when songs were fully recovered. Spectrograms (on the left in each panel) and averaged time-varying acoustic features (right) of each corresponding syllable were plotted in the same way as Fig. 3e-h. All spectrograms share the same time scale. Note that, on the first day after song-prevention (post-prevention), clear syllable renditions were sung and the intra-syllable timing of acoustic feature fluctuations were very similar to pre-perturbation syllables, though differences in acoustic feature magnitudes were still evident and the rendition-to-rendition variability was higher. Shading represents standard error of mean.

Extended Data Figure 6 – Recovery of syllable durations without practice.

(a) The distribution of durations of one example syllable, same as the one shown in Fig. 3g&h, before perturbation and right after song prevention (bin size, 10 ms). Note that they both peak at the same value. Two syllable-selection methods were used (see the Methods), with or without dynamic time warping, in order to test whether the k-nearest neighbor measure was missing dilated or compressed syllable renditions. The two syllable-selection methods found highly similar duration distributions; (b) Comparison of the mode of syllable durations before perturbation and post-prevention. Most syllables recovered their duration modes with high precision without song practice. Dashed lines indicate second peak of dual peak distribution; (c) Comparison of the standard deviation of duration distributions demonstrates that the variability of syllable durations was still elevated the first day post prevention. Comparable results were obtained with either type of syllable-selection method. In a-c, we performed this analysis on 2/3 birds in the song prevention experiment. In the third bird, the motif was not identifiable the first day post-song prevention, making it not feasible to construct individual syllable duration distributions on this day. One potential reason the song recovery trajectory differed is that this bird’s song was abnormally variable before the viral perturbation and song prevention experiment. We include this bird (V449) in the data and analyses presented in Fig. 4 b-e where the pre-perturbation variability and the post-prevention lag in recovery are clear. N = 8 syllables / 2 animals for both panel b and c.

Extended Data Figure 7 – Unperturbed cells did not change their intrinsic excitability or inhibitory synaptic inputs.

(a) (Left) Whole-cell patch clamp recordings were made in GFP-negative HVC(RA) neurons in naive control birds or birds injected with LV-TeNT. (Right) Membrane potential and firing pattern of HVC(RA) neurons in response to current steps; (b) Group data showing that no significant difference was found in the resting membrane potential (−75.2 ± 0.8 vs. −74.8 ± 0.8 mV), input resistance (414.8 ± 30.0 vs. 416.2 ± 33.1 MOhm), or initial F-I slope (437.9 ± 35.5 vs. 406.2 ± 37.3 Hz/A) between neurons in naive control (N = 22 cells / 3 animals) and in birds with LV-TeNT for more than 25 days (N = 30 / 4). Student’s t-test; (c) F-I curves obtained from HVC(RA) neurons showed no difference between control or animals recovered from LV-TeNT. Sample size is the same as that in panel b; (d) Group data of mIPSC recorded in HVC(RA) neurons, showing no significant difference at any time during the experiment. “Degraded” indicates the time when the song was degraded, at 5 dpi. “Recovered” indicates the time after the song had fully recovered, at 25 dpi. mIPSC frequency: Control, 2.7 ± 0.2 s⁻¹, N = 23 cells / 4 animals; “Degraded” GFP-negative cell, 2.6 ± 0.4 s⁻¹, N = 16 / 3; “Degraded” GFP-positive, 2.3 ± 0.4 s⁻¹, N = 16 / 3; “Recovered” GFP−, 2.6 ± 0.5 s⁻¹, N = 9 / 3; “Recovered” GFP+, 3.1 ± 0.8 s⁻¹, N = 9 / 3. mIPSC amplitude: Control, 37.9 ± 2.0 pA; “Degraded” GFP-negative cell, 40.6 ± 1.7 pA; “Degraded” GFP-positive cell, 43.2 ± 1.6 pA; “Recovered” GFP− negative cell, 39.0 ± 3.7 pA; “Recovered” GFP− positive cell, 41.4 ± 5.4 pA. One-way ANOVA. Error bars represent SEM.

Extended Data Figure 8 – Synaptic changes in unmanipulated neurons only occurred in the injected hemisphere.

(a) Example song spectrograms of a bird with LV-TeNT in one side of the two HVCs; (b) Acoustic distance to each original syllable when LV-TeNT was injected bilaterally (left) or unilaterally (right). Curves in the same color are from the same bird. Bilateral trajectories are similar to those shown in Fig. 2f but without normalization; (c) Comparison of the peak distortion, measured by the maximal distance from each original syllable, between unilateral or bilateral LV-TeNT perturbations. Each column represents syllables from the same bird. N = 24 syllables (nested) / 4 animals for Bilateral and 20 (nested) / 3 for Unilateral. Nested two-way ANOVA; (d) After unilateral LV-TeNT perturbation, the day when each syllable reached peak distortion compared with bilateral LV-TeNT (triangles. On average, 2.75 ± 0.46 dpi. Nested two-way ANOVA) and the day when each syllable achieved more than 90% recovery (round dots. on average, 17.82 ± 2.49 dpi, later than that for bilateral LV-TeNT, 13.71 ± 0.72 dpi, nested two-way ANOVA). Sample size is the same as that in panel c; (e) Whole-cell recordings were made in GFP-negative HVC(RA) (“unmanipulated”) neurons and interneurons in the injected and unperturbed hemispheres of birds with unilateral LV-TeNT injection; (f) Group data of mEPSC in naive control birds and birds with unilateral LV-TeNT. (Middle) Cumulative curve of inter-event intervals of mEPSCs. Dashed line represents data from GFP− neurons from animals with bilateral LV-TeNT injection for comparison and was adapted from Fig. 5c. mEPSC frequency: Naive control, 9.7 ± 1.6 min⁻¹, N = 23 cells / 4 animals; Contralateral (uninjected HVC), 9.8 ± 2.3 min⁻¹, N = 16 / 4; Ipsilateral (injected), 22.8 ± 3.3 min⁻¹, N = 18 / 3. mEPSC amplitude: Naive control, 17.2 ± 0.9 pA; Contralateral, 18.2 ± 0.9 pA; Ipsilateral, 22.3 ± 1.0 pA. One-way ANOVA followed by t-test (without adjustment); (g) Group data of mEPSC recorded in interneurons. (Left) The frequency of mEPSC in interneurons decreased after virus injection, but eventually recovered to a level comparable to that of controls. (Middle) Cumulative curve of inter-event intervals of mEPSCs. mEPSC frequency: Naive control, 1.47 ± 0.14 s⁻¹, N = 19 / 2; Degraded contralateral, 1.56 ± 0.16 s⁻¹, N = 13 / 3; Degraded ipsilateral, 0.63 ± 0.14 s⁻¹; N = 20 / 3; Recovered contralateral, 1.34 ± 0.19 s⁻¹, N = 18 / 4; Recovered ipsilateral, 1.45 ± 0.15 s⁻¹, N = 20 / 4. mEPSC amplitude: Naive control, 38.3 ± 1.2 pA; Degraded contralateral, 38.4 ± 2.6 pA; Degraded ipsilateral, 35.5 ± 1.9 pA; Recovered contralateral, 37.3 ± 2.1 pA; Recovered ipsilateral, 33.2 ± 2.0 pA.One-way ANOVA & student’s t-test. Error bars represent SEM.

Extended Data Figure 9 – Recovery of modeled sequences with STDP.

(a) Schematic showing triplet STDP implemented both in E→E and E→I synapses. Potentiation is mediated by triplets of spikes; depression is dictated by a pairwise rule. See Methods for details; (b) Spike raster plots showing the sequential dynamics generated by HVC neurons before and after perturbation with only STDP and downward firing rate homeostasis implemented (see Eq. 2 in Methods). Colors represent the firing timing of each cell before perturbation. Note that the sequence regenerates serially and all cells recapture their original firing timing; (c-e) Percentage of syllables completed prior to (pre-perturbation), just after (perturbed), and 3000 renditions after (recovered) perturbation for networks with (c) STDP alone, (d) STDP plus Poisson inputs, (e) or firing rate homeostatic plasticity, for 30% to 70% perturbations in increments of 10%. Note that networks with firing rate homeostatic mechanisms possessed larger recoverable regimes.

Extended Data Figure 10 – Recovery of modeled sequences with population-level homeostatic plasticity.

(a-d) are generated from networks with STDP and a downward, single-cell, firing rate homeostasis rule implemented; (e-h) are generated from networks with STDP and 2-sided, single-cell, firing rate homeostasis; (i-l) are generated from networks with silent cells and population-level homeostasis added to rules in (a-d); (a,e,i) Percentage of modeled syllables that successfully completed following >80% of activations in three time windows (before perturbation, perturbed and recovered) in response to different degrees of perturbation (10, 25, and 50%), when different sets of plasticity rules were implemented. Each point represents a network. Lines represent averages over all networks (N = 15); (b,f,j) The normalized total firing activity of all functional HVC(RA) (left) or interneurons (right) plotted against renditions. Solid curves show the average of multiple networks (N = 15) and shading represents SEM. RPP, renditions post perturbation; (c,g,k) Normalized total excitatory synaptic input into each unmanipulated HVC(RA) neuron after perturbations. Note that with population-level plasticity and silent neurons, an increase in E→E weight over its pre-perturbation value was introduced; (d,h,l) Distributions of single-cell firing rates per rendition for HVC(RA) (top) and interneurons (bottom) before and following 50% perturbation. Pre-perturbation distributions in black or gray. Lines represent mean and shadings represent standard error of mean, N = 15.

Data availability

All derived data in this study are included in this article. Raw datasets are publicly available online (https://doi.org/10.22002/dvhsa-h5s72) or by contacting the corresponding authors.

Code availability

Custom codes associated with this study are publicly available (for behavior analysis, https://doi.org/10.5281/zenodo.10823142; for modeling, code is available on GitHub at https://github.com/davidgbe/unsupervised_restoration_modeling and https://doi.org/10.5281/zenodo.10823218).