Recovery of a learned behavior despite partial restoration of neuronal dynamics after chronic inactivation of inhibitory neurons

Maintaining motor skills is crucial for an animal's survival, enabling it to endure diverse perturbations throughout its lifespan, such as trauma, disease, and aging. What mechanisms orchestrate brain circuit reorganization and recovery to preserve the stability of behavior despite the continued presence of a disturbance? To investigate this question, we chronically silenced inhibitory neurons, which altered brain activity and severely perturbed a complex learned behavior for around two months, after which it was precisely restored. Electrophysiology recordings revealed abnormal offline dynamics resulting from chronic inhibition loss, while subsequent recovery of the behavior occurred despite partial normalization of brain activity. Single-cell RNA sequencing revealed that chronic silencing of interneurons leads to elevated levels of microglia and MHC I. These experiments demonstrate that the adult brain can overcome extended periods of drastic abnormal activity. The reactivation of mechanisms employed during learning, including offline neuronal dynamics and upregulation of MHC I and microglia, could facilitate the recovery process following perturbation of the adult brain. These findings indicate that some forms of brain plasticity may persist in a dormant state in the adult brain until they are recruited for circuit restoration.


Introduction
Maintaining the ability to precisely execute motor behavior throughout life, despite perturbations due to trauma, disease, or aging, is crucial for reproduction and survival. To reliably execute behaviors, brain circuits require a balance of excitation and inhibition (E/I balance) to maintain physiological activity patterns. Loss of E/I balance causes abnormal patterns of neuronal activity, which can result in diseases such as epilepsy 1,2 . Given its importance, brain circuits strive to restore E/I balance once disturbed 3 . However, the mechanisms orchestrating circuit reorganization and recovery of E/I balance during the continued presence of a perturbation remain poorly understood.
Zebra finches produce a highly stereotyped song with minimal variability over extended periods of time 4 , underpinned by temporally precise neural activity 5 . This species thus provides an optimal model to simultaneously track abnormal brain activity, its effect on behavior, and changes to both over time. It serves as an excellent model for studying chronic E/I imbalance and accompanying changes over time at the behavioral, neuronal, and transcriptomic levels.
Here, we genetically block inhibitory neurons in HVC (proper name), a premotor brain nucleus of the male zebra finch involved in song production, to chronically perturb the E/I balance. HVC contains two main types of excitatory neurons that project to two main downstream targets: nucleus X (proper name) and nucleus RA (robust nucleus of the arcopallium). In addition, HVC includes several types of inhibitory neurons whose axons do not leave HVC; thus, they act locally within HVC 5 . Juvenile male zebra finches learn their song from their fathers during the "critical period" and once learned, produce a highly stereotypical song for the rest of their lives 6 . Previous work has shown that inhibition plays a role during development to close the critical period and protect learned components of the song in juvenile animals 7 . Acutely blocking interneuron signaling in adult animals by chemicals leads to abnormal behaviors that quickly return to normal once the chemical is washed out 8 . However, it is not known how long-term disruption of inhibitory neurons affects neuronal dynamics, and whether behaviors can recover after such drastic perturbation.

Results
To investigate how a complex motor behavior recovers after chronic loss of inhibitory tone, we blocked the function of HVC inhibitory neurons in adult male zebra finches by bilateral stereotaxic injection of an AAV viral vector into HVC. The AAV viral vector carried the light chain of tetanus toxin (TeNT), driven by the human dlx5 promoter, which is selectively active in inhibitory neurons 9 . TeNT blocks the release of neurotransmitters from presynaptic terminals, thereby preventing neurons from communicating with their postsynaptic partners 10 . Thus, expression of TeNT does not directly alter the ability of neurons to fire action potentials, but effectively mutes them. As a control, a second group of animals was injected with an AAV carrying the green fluorescent protein NeonGreen driven by the ubiquitous promoter CAG. Throughout various time points in this perturbation paradigm, we recorded song behavioral data, obtained electrophysiological measurements (chronic and acute within HVC), and measured changes in gene expression at single-cell resolution (Graphical Abstract).

Chronic genetic muting of inhibitory neurons in HVC leads to zebra finch song degradation and subsequent unsupervised recovery
An average male zebra finch song typically contains 4 to 7 units called syllables. These syllables are fixed in order and make up a motif, which is repeated multiple times to form a bout [11][12][13] . A bout defines the length of a vocally active period in adult animals. Throughout all experiments, the song of the individually housed animals was continuously monitored from 7-20 days pre-injection to 60-100 days post-injection (dpi) using sound-isolated chambers. In control animals, the song did not change at any time after the viral injection. For the first 24-48 hours after injection, the song of TeNT-treated animals was virtually identical to their original song. However, starting at 2-5 dpi, the song of TeNT-treated animals began to degrade, both in its acoustic and temporal features ( Figure 1). By 10 dpi, the song of TeNT-treated animals was highly degraded and bore no resemblance to the original song before viral injection (Figure 1 A).
During the first few days of song degradation, TeNT-treated animals produced instances of syllables with abnormally long durations far outside the range of any adult syllable duration observed in control animals ( Figure 1 A shows one such vocalization at 5 dpi; examples of syllable length distribution:  Figure 2). At 20-30 dpi, the song of TeNT-treated animals was highly variable between renditions and lacked the normal organization of motif and bout (Figure 1 A). However, by 40-50 dpi, the structure of the song started to re-emerge. By around 70 dpi, the song of TeNT-treated animals was highly similar to the song produced before viral injection ( Figure 1).
Nucleus LMAN is required for learning and adult song plasticity [14][15][16] . To investigate the role of LMAN in song recovery, we chemically lesioned it before the viral perturbation. We found that the LMAN lesion did not affect the song degradation and recovery trajectory after chronically muting HVC inhibitory neurons with TeNT ( Supplementary Figure 3-6). These results indicate that behavioral recovery after perturbation of HVC does not require LMAN, a nucleus necessary for song learning.
Chronic electrophysiological recordings reveal that the degree of song distortion is correlated with the presence of large-amplitude voltage deflections in HVC during nighttime To investigate electrophysiological changes in HVC that accompany song degradation due to chronic loss of inhibition and subsequent behavioral recovery, we continuously recorded electrical activity in HVC across the entire duration of song degradation and recovery (0-90 dpi).
We implanted a 4-shank electrode array containing 16 recording sites in both control and TeNT-injected animals. We simultaneously collected electrophysiological, head movement (recorded with the accelerometer built into the head stage), and audio recordings over a period of 90 dpi. We analyzed the spectral decomposition of the local field potentials (LFPs) during song production from an average of five vocalization epochs 17,18 . LFP recordings provide a readout of transient electrical signals generated by the summed and synchronous local and distant synaptic and action potential activity. In control animals, as expected, both the song behavior and LFP characteristics remained stable: the electrical activity during singing showed voltage deflections which were present throughout the song   ( Supplementary Figure 7 B). In contrast, at 5 dpi, TeNT-treated animals produced long vocalizations that were preceded by large voltage deflections, neither of which were observed in control animals ( Supplementary Figure 7 A). Between 10 to 70 dpi, the songs of TeNT-treated animals were highly variable between renditions and between days.
As a result of this instability in behavior, vocalization number and type could not be reliably characterized and aligned across trials. Hence, we focused on electrophysiological events occurring during nighttime (lights-off periods when the animals are not moving or singing), because they showed higher data consistency (Figure 2 A). The neuronal activity recorded during the nighttime period will be referred to as offline or lights-off throughout the manuscript.
Previous studies have hypothesized that HVC offline activity may reinforce the maintenance of adult song and could drive song learning [19][20][21][22][23] . Therefore, analyzing electrophysiological activity in HVC during offline periods throughout the perturbation could reveal disruptions related to degraded behavior.
The recordings in TeNT-treated animals during lights-off offline periods contained voltage deflections that were more frequent and of larger amplitude than those of controls ( Figure 2 A and B), similar to those observed during singing. In addition, in TeNT-treated animals, these large amplitude deflections at night occurred more frequently during the period of time when the song was most degraded (Figure 2 B). At 5 dpi, these stereotypical voltage deflections had an average amplitude of around -530 µV and -140 µV in TeNT-treated and control animals, respectively -a 280% increase in TeNT-treated animals (Supplementary Figure 8). In TeNT-treated animals, the rate of lights-off offline voltage deflections returned to those seen in control as the song regained stereotypy around 50 dpi (Figure 2 B).  [15][16][17][18][19][20][21][22][23][24][25][26][27][28][29][30] Hz range were approx. 27 times larger than those in control animals at 5 dpi. These differences are statistically significant between control and TeNT-treated groups, but not within the control group (p=0.7631 between controls, p<10 -35 between all other pairs). In the 30-70 Hz range, deflections were different between all animals (p=7.9*10 -5 between controls, p<10 -40 between all other pairs). The deflections across all animals and frequencies become more similar by 60 dpi (15-30 Hz: p=0.1371  To further characterize these voltage deflections, we analyzed the LFPs (1-200 Hz) during the voltage deflection events occurring during lights-off periods, and we performed spectral decomposition of the LFPs. At 5 dpi, we observed an increase in power in the low gamma band  in TeNT-treated animals, which returned to control levels as the songs regained stereotypy (Figure 2 C and D).
The highest level of song distortion coincided with the highest rate of large-amplitude voltage deflections during lights-off (Figure 2 B). These findings suggest a relationship between the behavioral timeline of recovery, and the return to control-like offline electrophysiological activity. These results suggest that the restoration of excitation/inhibition (E/I) balance and circuit function is mediated by unsupervised mechanisms, consistent with our earlier publication 24 .

Spontaneous local neuronal activity in HVC and RA during lights-off periods show opposite trends upon chronic interneuron perturbation in HVC
To record local neuronal activity at single-neuron resolution over the course of song degradation and recovery, we used high-density silicon probes (Neuropixels) 25,26 to simultaneously record neurons in HVC, RA, and intervening brain areas (Figure 3 A). As described above, the most reliable measurements of abnormal brain activity in TeNT-treated animals were obtained during lights-off periods. Therefore, we performed Neuropixel recordings in head-fixed animals when the lights were off during their natural sleep cycle. To achieve the best possible signal-to-noise ratio, which would allow single-neuron resolution, we focused on acute recordings that were obtained within a few hours after probe insertion. We performed acute recordings at 3-6, 20-22, and 70-77 dpi in TeNT-treated animals, and at 10 and 30 dpi in control animals. We also included recordings from one naïve animal (not injected with any virus) to control for potential artifacts caused by the surgery or viral injection.
We measured the overall neuronal firing rate in HVC and RA of control and TeNT-treated animals. In TeNT-treated animals, HVC started out in a heightened state of spontaneous lights-off activity (likely due to the loss of inhibition), but over the time course of the perturbation activity gradually decreased to levels below control (Figure 3 B). While activity in HVC was elevated between 3-5 dpi and eventually decreased between 20-70 dpi, RA showed the opposite trend, possibly to compensate for the observed changes in HVC activity (Figure 3 B). This suggests the existence of mechanisms that coordinate the spontaneous activity levels between connected brain areas to achieve a balanced output required for the appropriate control of behavior.
In addition, we observed a total of seven events that resemble seizure-like activity in HVC (n=3 TeNTtreated animals at 3-20 dpi) which propagated downstream into RA (Figure 3 C). These periods of elevated neuronal firing in TeNT-treated animals in the early days after perturbation suggest that the removal of inhibitory tone acutely leads to bursts of hyperactivity that resemble discharges previously reported during seizures 1,2,27 .

Firing precision of neurons in HVC specific to alpha frequency oscillations recovers as song stereotypy is regained
We used acute Neuropixel recordings to analyze the neuronal firing specifically during the offline voltage deflection events, similar to those that we initially observed using chronic recordings (

: Acute electrophysiology recordings after chronic loss of inhibitory neurons in HVC by viral expression of tetanus toxin. A Example histology of a brain slice of a control animal showing the trace of the probe (NPIX coated with DiI) insertion (red) and GFP from the viral injection (green). The raw electrophysiology traces are color coded by brain area (HVC in red, NCM in black, RA in blue). B
The average firing rate of neurons within HVC and RA in control and at 3 (n=1 animal), 4-6 (n=4), 20 (n=4) and 70 (n=4) dpi. In HVC, the increase in average firing rate at 3 dpi significantly differs from control (p-value from rank sum test: 0.4*10 -6 ), as well as the decrease in average firing rate at 20 dpi (p = 0.002) and 70 dpi (p = 0.57*10 -11 ). In RA, the decrease in average firing rate at 3 dpi significantly differs from control (p = 0.43*10 -6 ), as well as the increase in average firing rate at 20 dpi (p = 0.09*10 -9 ) and 70 dpi (p = 1.45*10 -13 ). The stars above the bar plots indicate statistical significance (* = p < 0.005, ** = p< 0.01, *** = p < 0.001). C Examples of thresholded neuronal activity in control, and at 3 and 20 dpi. The red arrows highlight the "superbursts" or extreme firing levels within HVC and RA which we observed in 3 animals (twice at [3][4]   We compared the timing of single neuron firing with the phase of the alpha and low gamma oscillations during offline voltage deflection events. To this end, we turned to methods previously described to characterize sharp wave ripple events during offline activity in the hippocampus 28 . Previous studies suggest that precise neuronal firing offline during a specific phase of sharp wave ripples underlies correct memory consolidation and recall in the hippocampus [28][29][30][31] . We found that in control animals, the neuronal firing within HVC locked tightly to specific phases of the alpha and low gamma oscillatory events during offline voltage deflections. In TeNT-treated animals, during the song degradation period, the firing precision to low gamma oscillations in HVC or to alpha and low gamma oscillations in RA during deflection events did not change compared to control (Supplementary Figure 9). However, during the song degradation period, the precision of local neuronal firing to alpha oscillations between 60 to 140 degrees was lost or broadened compared to control and song-recovered animals (Figure 3 E and F). As song stereotypy was regained, the firing precision to specific phases of the alpha oscillations during deflections re-emerged (Figure 3 E and F). These findings suggest the importance of local neuronal firing precision to alpha oscillations within HVC for correct circuit function. The observed alpha frequency LFP signals could represent inputs into HVC from other brain areas. The capacity of neurons to exhibit phase-locked firing at a specific alpha frequency oscillation may be required for integrating inputs from diverse brain regions and enables the execution of precise motor behaviors.

Single-cell RNA sequencing suggests mechanisms of neuronal plasticity driven by microglia and MHC class I genes during song perturbation
To investigate cellular mechanisms that might underlie the observed changes in neuronal activity and behavior at the transcriptomic level, we performed single-cell RNA sequencing (scRNAseq) of HVC from control (n=2) and TeNT-treated (n=2) adult male zebra finches at 25 dpi, around the time of peak song distortion. HVC from both hemispheres of all four birds were dissected based on retrograde tracer fluorescence and dissociated to prepare single-cell suspensions, which were indexed and pooled. This allowed the construction of a combined dataset, containing results from all organisms and conditions, without the need for batch correction (Supplementary Figure 12, Supplementary Table 2). Following quality control, we retained a total of 35,804 single-cell profiles spanning four individuals, consisting of two control and two TeNT-treated animals.
While cell type abundance was highly concordant between replicates of the same condition (Supplementary Figure 12, Figure 4 A), we found that animals treated with TeNT displayed a three-fold increase in the number of microglia (Figure 4 B). This increase in microglia was likely not due to an inflammatory reaction caused by the surgical procedure or the AAV injection because control animals also received a viral injection with a highly similar construct. Thus, we hypothesize that the increase in microglia in TeNT-treated animals is a consequence of the chronic muting of inhibitory neurons. Several studies have shown that microglia play a role in synaptic plasticity during early brain development and learning in mammals [32][33][34][35] . These prior observations in combination with our findings suggest that microglia might participate in the synaptic reorganization triggered by circuit perturbation. We performed in situ hybridization (ISH) using a probe against RGS10, a gene expressed in microglia, during song degradation at 25 dpi and after recovery at 90 dpi. At 25 dpi, the number of microglia increased in TeNT-treated animals compared to control (Supplementary Figure 13 A, 14 A and B), and returned to control levels by 90 dpi, when the song had recovered.
To further investigate the hypothesis that microglia are associated with circuit reorganization involved in neuronal plasticity, we counted the number of microglial cells in HVC at different times during song learning in naive (untreated), juvenile birds using ISH. The number of microglia in the HVC of juveniles was higher during the early stages of the song learning period (30-50 days post-hatching (dph)), compared to 70 dph, after the song became more stereotypic (Supplementary Figure 13 36 . This increase in MHC I and B2M is likely triggered by the genetic silencing of inhibitory neurons, and not due to an inflammatory response because it was not observed in control animals injected with a control virus. MHC I and B2M have been observed in previous studies to be highly expressed in neurons during brain development 37,38 , consistent with a hypothetical role in synaptic plasticity 39,40 . Here we observe that MHC I and B2M are upregulated in response to perturbation of neuronal activity in a fully formed circuit, indicating their possible role in the restoration of brain function.

Discussion
The brain requires a balance between excitation and inhibition to maintain physiological activity patterns and enable reliable behaviors. We found that chronic muting of inhibitory neurons in a pre-motor circuit severely disrupts a learned motor behavior for an extended period of time (30-90 days). Even after several weeks of perturbed brain activity, the brain circuit is able to regain function and recover the behavior, despite failing to restore some aspects of its neuronal dynamics. We propose that the return to controllike offline voltage deflections and the precision of local neuronal activity during alpha oscillations during these deflections are key components that accompany the behavioral recovery. Our observations indicate a putative relationship between activity dynamics offline and the restoration of circuit function. Furthermore, our data suggest that microglia and MHC I may be involved in changes caused by chronic perturbation of neuronal activity. These experiments reveal that the adult brain can overcome extended periods of E/I imbalance, potentially by processes that occur offline. The reactivation of mechanisms typically employed during juvenile learning, including night replay and activation of MHC I and microglia, could facilitate the recovery process following perturbation in the adult brain. This indicates that some forms of circuit plasticity may persist throughout adulthood, entering a dormant state until their activation is required for circuit restoration.

Animals
All procedures involving zebra finches were 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 genders and ages until used for experiments. Before any experiments, adult male birds (>120 days post-hatching (dph)) were singly housed in sound isolation cages with a 14/10 hr light/dark cycle for >5 days until they habituated to the new environment and started singing. Thereafter, birds were kept in isolation until the end of the experiment.

Behavioral recordings
Adult male zebra finches' (n=30, 130-890 dph) undirected songs were recorded 24/7 in sound-isolated chambers for 10-14 days before any manipulation to get a baseline of their song. Recordings were done with microphones (Audio-technica, AT831b) that are connected to an amplifier M-TRACK 8 and recording software Sound Analysis Pro 2011 at 44100 Hz. Animals were housed in these chambers and continuously recorded for the duration of the experiments.

Viral vectors
AAV-TeNT contained the promoter from the human dlx5 gene driving expression of the light chain of tetanus toxin fused to EGFP with a PEST domain. AAV9-dlx-TeNT was obtained from the Duke viral core facility. The control virus used was AAV9-CAG-NeonGreen, where CAG drives the expression of NeonGreen which is a GFP variant.

Stereotaxic injection
Birds were anesthetized with isoflurane (0.5% for initial induction, 0.2% for maintenance) and headfixed on a stereotaxic apparatus. First, to inject a retrograde tracer in area X, craniotomies were made bilaterally and fluorescent tracers (fluoro-ruby 10%, 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). To deliver the virus (AAV) 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. AAVs diffuse extensively (~500 µm), and a single injection (~100 nL) in the center of HVC was sufficient to label enough cells. All injections in HVC were performed at ~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.

Chronic electrophysiology recordings
Animals (n=4, 300-700 dph) were implanted in the right hemisphere HVC with 4 by 4 electrode arrays (Neuronexus A4x4-3mm-50/100-125-703-CM16LP) based on retrograde fluorescent labeling of HVC (just as for viral injections). Post-perfusion histology images were obtained to locate the electrode array within HVC for each animal (Supplementary Figure 15). Electrode implantation occurred within the same surgery as the viral injection.
This procedure follows the same surgical steps as the viral delivery protocol, until the point of electrode implantation. A small opening was cut on the dura (just big enough to fit the electrode array) to lower the electrodes manually. The reference and ground were a gold screw pin placed into the cerebellum. The skin was removed from the surface of the skull for the majority of the surface, in order to secure the implant. Before implantation, the skull and the craniotomies were cleaned with saline and dried and the skull was prepared according to the protocol of the C&B Metabond cement system. Post implantation we covered the craniotomies with kwik-sil. Once hardened, we covered the whole skull, and the part of the electrode still exposed, with metabond. The head stage (Intan RHD Part # C3335) was connected to the probe before implantation and securely metabonded to the connection between the probe and head stage in order to prevent detachment when the bird is moving. SPI interface cables (Intan Part #C3203, #C3213) were connected to the acquisition board (Open Ephys). Data was recorded at 30,000 Hz with the Open Ephys software system. Animals were freely moving with a passive counterweight-based commitator system.

Acute electrophysical recordings
Animals (n=10, 140-250 dph) went through the same surgical procedure as described for a stereotaxic viral injection. However, at the end of the surgery the skin was removed from the skull, and the whole skull was pre-treated and covered in metabond except for the craniotomies over HVC that were covered with kwik-cast until the day of the acute recording session. Shortly before the recording session, a headbar was glued on top of the frontal surface of the metabonded skull to allow the head-fixation of the bird for the recording session. Then, the kwik-cast was removed from the craniotomy over HVC (left or right hemisphere or both depending on the animal) and a small incision was made in the dure over HVC, which was identified by the retrograde tracer previously injected. The ground was placed into the cerebellum. Then the high-density silicone probe (Neuropixel) was lowered with a motorized arm over hours for 2.6-3 mm deep into the brain. The head stage and acquisition board was connected to the computer and data was recorded with the Open Ephys software. Once the probe settled in the brain, we had 4 distinct recording sessions. Post-perfusion histology images were obtained to locate electrode array within HVC for each animal (Supplementary Figure 16). Recording sessions: lights on silence (10 min), followed by playback of the bird's own song (3-10 min); lights-off silence (10 min), followed by playback of the bird's own song (3-10 min); microinjection of 100 nL 250 µM Gabazine (Hellobio, HB0901), followed by the same protocol of lights-off and on without Gabazine.

Syllable parameterization
To quantify the change in the acoustic structure of the song, we tracked song features over the course of the perturbation. In Fig 1, we took the average feature values over all song segments in the three closest days of recorded song for each day plotted. We extracted the average mean frequency, average pitch, maximum goodness of pitch, maximum log power, minimum entropy, and entropy variance over the first 50 ms of each song segment, or the entire song segment if the segment was less than 50 ms. We centered and normalized feature values by the mean and standard deviation of the distribution of feature values in syllables five days pre-perturbation to compare trends across birds.
To visualize how the population of song segments evolved relative to the original syllable structure ( Fig  1B), we plotted song segments as points in the first 2 principal components of the 7-dimensional acoustic feature space composed of: average mean frequency, average pitch, maximum goodness of pitch, maximum log power, minimum entropy, and entropy variance over the first 50 ms of each song segment, and the full duration of each song segment. 500 syllables recorded pre-perturbation were labeled by hand, and contour plots were made at the mean of each labeled stereotyped syllable type for visual reference as the song degrades and recovers. Principal component dimensions were computed on the 500 labeled syllables, and subsequent song segments were projected into this principal component space.

Syllable segmentation
We identified syllables and silences within each bout by imposing thresholds on the time-varying, total log-power of the spectrogram. We selected the threshold based on manual inspection for each bird. 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 the analysis. Segmentation was further checked by eye by random sampling across both stereotyped motifs and degraded song. We then applied the power threshold to the entire course of song recordings for each bird.
A note on terminology: we refer to song segments to indicate continuous periods of singing. In the unperturbed song, these song segments are termed syllables. Because this is a continuous recovery process, these terms sometimes overlap in our usage.
Syllable timing distributions Syllable durations were extracted from the syllable segmentation process. Discrete distributions of syllable durations were computed by normalizing each distribution of all song segment durations 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 in Supplementary  Figure 1.

Continuous representation of bout trajectory
We generated continuous visualizations of bouts across the entire perturbation trajectory as shown in the graphical abstract 42,43 . We randomly sampled 100 bouts from single days 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 20 ms segments every 5 ms of six 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 reduce the feature set to 30 dimensions.

Analysis of the chronic electrophysiology recordings
Preprocessing of Data Bad channels were removed from analysis by visual inspection of the first minute of each recording. Three of the four birds studied had one channel that had behavior that was an order of magnitudes larger than all others and data from this channel was removed for further analysis.
Detection of sharp waves Accelerometer and electrophysiology data were resampled to 125 Hz from the original 30,000 Hz using the Matlab resample function. Since accelerometer data was not zero-centered, the mean was subtracted before resampling and then added back to avoid edge effects from the lowpass filtering during resampling. Sharp waves were detected from the downsampled data using a procedure adapted from a zebra finch sleep study 44 and were only detected during periods of low movement. This was determined by setting a threshold on the moving standard deviation of the velocity as movement events appeared as large fluctuations in the velocity data. Briefly, the local field potential (LFP) traces were averaged across all electrodes and bandpassed from 1 to 40 Hz using MATLAB's bandpass function. Putative events were detected by peaks in the inverted (-1x) trace and were in the lowest 5% of all values during nonmovement, were at least 10 ms long, and were 80 ms apart. These putative events were used to construct a template by taking the average 80 ms trace centered on each event. Cross-correlating this template with the entire overnight LFP trace and taking the z-score (mean and SD from only non-movement times) revealed events that closely matched the template. Peaks were detected in the z-scored cross-correlation and defined as events if they were above a given threshold and were at least 80 ms apart. Notably, thresholds were not tuned across birds or recordings but a constant threshold was sufficient in all birds to detect events. Extraction of sharp waves. The detected events were used to index the original 30,000 Hz data. 500 ms of data was extracted for each event and recentered on the minimal value in the center 100 ms. An equal number of 500 ms non-sharp wave events were extracted from a random selection of times without movement or sharp waves. Sharp waves with clear electrical noise were omitted from further analysis.
Multi-unit activity (MUA) MUA was detected using an established method 45 . Briefly, channels were common averaged referenced, bandpassed from 600 to 6,000 Hz, and upsampled to 50,000 Hz. The square root of the local energy operator (Mukhopadhyay & Ray, 1998) was taken and then Gaussian smoothed ( = 0.4). Thresholding this trace and then downsampling back to 30,000 Hz gave the times of high MUA.
Time-frequency analysis The continuous wavelet transform was used to compute the spectra for each sharp wave. Percent increase relative to mean power spectra were averaged within each day and bird. The spectrum of average power in the center 33 ms of each event was used to compute the percent increase in power relative to non-sharp wave events at each frequency. To determine the presence of phase-amplitude coupling between lowfrequency (10-20 Hz) and gamma-frequency (30-40Hz) bursts, the phase of the low-frequency band, and power of the high-frequency band were extracted using the Hilbert transform of the band-passed signal. Determining whether the peak gamma activity occurred at a consistent phase of the low-frequency oscillation, could suggest phase-amplitude coupling. To further assess this, we asked at what time and frequency a max in the spectrum occurred for each event, relative to the minima of each sharp wave. We also wanted to ask whether the power of deflections in different frequency bands differed significantly across control and experimental birds. We computed the relative increase in power in the frequency bands of interest (15)(16)(17)(18)(19)(20)(21)(22)(23)(24)(25)(26)(27)(28)(29)(30) Hz, and 30-70 Hz) during a deflection relative to non-deflection times. We then used a t-test to statistically test whether power differed significantly across birds for a given day.

Analysis of the acute electrophysiology recordings
The low-frequency electrophysiology data was recorded at 2,500 Hz while the high-frequency data was recorded at 30,000 Hz. A MATLAB script was written to analyze the low-and high-frequency oscillations. All low-frequency data was high pass filtered (1-300 Hz) and median subtracted. All highfrequency data was common average referenced and low-pass filtered (300-7,000 Hz). For deflection event detection, we used two methods. One to use a template per animal for extraction of deflection and non-deflection events with template matching in a semi-automated manner. Second, manually picked out peaks that were 9 standard deviations away from the average raw signal, then further manually curated these events. Most of our further analysis was done using the deflection events that resulted from template matching (except Supplementary Figure 10A for the difference in power between deflection and nondeflection event calculation). For spike detection, we used two separate methods, one to threshold the high-frequency data to only consider spikes that are 9 standard deviations above the average signal in a 2-minute window. The second method used for spike detection was using the Kilosort software v2.5 46 to extract spikes from the raw data. Next, we manually curated the dataset to capture the waveform and firing rate of neurons. The thresholded spiking data was used to generate Figure 3BC. We performed statistics (Wilcoxon, rank sum test) to show statistically significant firing rates related to control in HVC and RA in Figure 3B.
To calculate the power difference between deflection and non-deflection events in the acute recording in Supplementary Figure 11 A, we used a script developed for the chronic recordings with the use of the built-in MATLAB function pwelch. To analyze the spectral decomposition of the LFP signature ( Figure  3D) in control and TeNT-treated animals during the experimental timeline we calculated the Morse continuous wavelet transform using the built-in MATLAB function. From this finding, we obtained the mean oscillation at the alpha range (1-10 Hz) and at low gamma . Then, we performed an iterative alignment process of these two oscillations and calculated the angle of the low-frequency oscillation (from obtaining the Hilbert transform) at the peak amplitude of the high-frequency oscillation (Supplementary Figure 11BC). Then we proceeded to align the single neuronal firing (obtained from Kilosort) to the alpha and the low gamma oscillations and plotted the normalized probability distribution ( Figure 3F) of neurons firing at a specific phase (angle) of the alpha or low gamma oscillations. We performed the Kolgomorov-Smirnov test on the cumulative density function (CDF) (Supplementary Figure 11D) to assess if the change in probability distribution in the relationship between alpha and low gamma oscillations is statistically significant from control distributions at 3-6, 20, and 70 dpi.
Single-cell RNA sequencing Animals All of the work described in this study was approved by California Institute of Technology and Oregon Health & Science University's Institutional Animal Care and Use Committee and is in accordance with NIH guidelines. Zebra finches (Taeniopygia guttata) were obtained from our own breeding colony or purchased from local breeders.
Dissociation and cDNA generation Animals were anesthetized with a mix of ketamine-xylazine (0.02 mL / 1 gram) and quickly decapitated, then the brain was placed into a carbogenated (95% O2, 5% CO2) NMDG-ACSF petri dish on ice. The brains were dissected on a petri dish with NMDG-ACSF surrounded by ice under an epifluorescent microscope guided by the fluoro-ruby retrograde tracing from Area X to HVC. We used the commercially available Worthington Papain Dissociation system with some minor changes and add-on steps. We followed all the steps included in the Worthington protocol with a final concentration of 50 U/mL of papain. To match the intrinsic osmolarity of neurons in zebra finches we used NMDG-ACSF (~310 mOsm) instead of the EBSS for post-dissection and STOP solution. Another modification was to add 20 µL of 1 mg/mL Actinomycin D (personal communication from Allan-Hermann Pool; 47 ) into 1 mL of the post-dissection medium and the STOP solution in which trituration occurred. Papain digestion occurred for an hour on a rocking surface with constant carbogenation in a secondary container above the sample vial at RT. We performed trituration with increasingly smaller diameter glass pasteur pipettes. Trituation was performed inside the papain solution. Then, once the tissue was fully dissociated, we centrifuged the samples at 300 g RT for 5 minutes and resuspended them in STOP solution. Next, we used a 40 µm Falcon cell strainer pre-wet with the STOP solution and centrifuged again at 300 g RT for 5 min. Finally, we resuspended the cell pellet in 60µl of STOP solution and proceeded to barcoding and cDNA synthesis. The cell barcoding, cDNA synthesis, and library generation protocol were performed according to the Chromium v3.1 next GEM single cell 3' reagent kits by Jeff Park in the Caltech sequencing facility. Sequencing was performed on an Illumina Novaseq S4 sequencer with 2x150 bp reads.
Generation of count matrices The reference genome GCA_003957565.2 (Black17, no W) was retrieved from Ensembl on March 20, 2021 (http://ftp.ensembl.org/pub/release-104/gtf/taeniopygia_guttata/). We quantified the gene expression in each of the four datasets using the kallisto-bustools workflow 48 . The reference index was built using the kb-python (v0.26.3) ref command and the above-mentioned reference genome. Subsequently, the WRE sequence was manually added to the cdna and t2g files generated by kallistobustools to allow the identification of transgenic cells. The count matrix was generated for each dataset using the kallisto-bustools count function. The resulting count matrices were compared to those generated by the 10X Cell Ranger pipeline (v6.0.1) and kallisto-bustools count with multimapping function. For all four datasets, kallisto-bustools mapped approximately 10% more reads than Cell Ranger (Supplementary Figure 17). No increase in confidently mapped reads was observed when using the multimapping function, indicating that reads align confidently to one gene in the reference genome (Supplementary Figure 17).

Quality control and filtering
The datasets were filtered separately based on the expected number of cells and their corresponding minimum number of UMI counts (Supplementary Figure 11). Following quality control based on apoptosis markers and library saturation plots (Supplementary Figure 11), the count matrices were concatenated and normalized using log(CP10k + 1) for downstream dimensionality reduction and visualization using Scanpy's (v1.9.1; 49 ) normalize_total with target sum 10,000 and log1p. Gene names and descriptions for Ensembl IDs without annotations were obtained using gget (v0.27.3; 50 ).

Dimensionality reduction and normalization
The concatenated data was mapped to a lower dimensional space by PCA applied to the log-normalized counts filtered for highly variable genes using Scanpy's highly_variable_genes. Next, we computed nearest neighbors and conducted Leiden clustering 51 using Scanpy.
Initially, this approach was performed on the control and TeNT datasets separately. This resulted in the identification of 19 clusters in the control data and 22 clusters in the TeNT data (Supplementary Figure  11). For both conditions, equal contribution from both datasets indicated that there was minimal batch effect, as expected since the data was sequenced in a pooled sequencing run. We also performed batch correction using scVI 52 which did not change the contribution of each dataset per cluster. As a result, we continued the analysis using the data that was not batch-corrected with scVI.
Next, we concatenated all four datasets and followed the approach described above. This resulted in the identification of 21 Leiden clusters, which we also refer to as cell types (Figure 4 A). Each cluster was manually annotated with a cell type based on the expression of previously established marker genes 53 . The cell type annotation was validated by the top 20 differentially expressed genes extracted from each cluster using Scanpy's rank_genes_groups (P values were computed using a t-test and adjusted with the Bonferroni method for multiple testing) (Supplementary Figure 12). Clusters identified as glutamatergic neurons were further broken down into HVC-X-and HVC-RA-projecting glutamatergic neurons using previously established marker genes (data not shown; also see https://github.com/lauraluebbert/TL_2023). We found that reclustering all cells labeled as glutamatergic neurons using the Leiden algorithm did not yield different results and we therefore continued with the initial clusters (data not shown). All results discussed in this paper were confirmed by both jointly and separately clustering the experimental conditions. Comparative analysis of clusters and conditions Differentially expressed genes between clusters were identified using Scanpy's rank_genes_groups (p values were computed using a t-test and adjusted with the Bonferroni method for multiple testing, and confirmed by comparison to P values generated with Wilcoxon test with Bonferroni correction). In the violin plots, unless otherwise indicated, a star indicates a p value < 0.05 and a fold change > 1.5 difference in mean gene expression between the indicated conditions (p value computed with scipy.stats' (v1.7.0) ttest_ind and adjusted with the Bonferroni method for multiple testing).

In situ hybridization
Animals All of the work described in this study was approved by the California Institute of Technology and Oregon Health & Science University's Institutional Animal Care and Use Committee and is in accordance with NIH guidelines. Zebra finches (Taeniopygia guttata) were obtained from our own breeding colony or purchased from local breeders. Developmental gene expression in HVC in the 20-, 50-, and 75-days post-hatch (dph) male and female zebra finches was assessed as previously described 54 . The sex of birds was determined by plumage and gonadal inspection. Birds were sacrificed by decapitation, bisected in the sagittal plane and flash-frozen in Tissue-Tek OCT (Sakura-Finetek), and frozen in a dry ice/isopropyl alcohol slurry. Brains of TeNT-manipulated finches were coronally blocked anterior to the tectum and flash frozen in Tissue-Tek (Sakura). All brains were sectioned at 10um on a cryostat and mounted onto charged slides (Superfrost Plus, Fisher).

In situ hybridization
In situ hybridization was performed as previously described 55,56 . Briefly, DIG-labeled riboprobes were synthesized from cDNA clones for RGS10 (CK312091) and LOC100231469 (class I histocompatibility antigen, F10 alpha chain; DV951963). Slides containing the core of HVC were hybridized overnight at 65°C. Following high stringency washes, sections were blocked for 30 min and then incubated in an alkaline phosphatase conjugated anti-DIG antibody (1:600, Roche). Slides were then washed and developed overnight in BCIP/NBT chromogen (Perkin Elmer). To minimize experimental confounds between animals, sections for each gene were fixed together in 3% paraformaldehyde, hybridized with the same batch of probe, and incubated in chromogen for the same duration.
Sections were imaged under consistent conditions on a Nikon E600 microscope with a Lumina HR camera and imported into ImageJ for analysis. We quantified the expression level of the gene as measured by optical density and the number of cells expressing the gene per unit area, as previously described 54 . Optical density was measured by taking the average pixel intensity of a 300x300 pixel square placed over the center of HVC. This value was normalized to the average background level of the tissue. To quantify the number of labeled cells, we established a threshold of expression that was 2.5x the background level. Binary filters (Close-, Open) were applied and the number of particles in the same 300x300 pixel square was quantified.

Histology
After cardiac perfusion with room temperature 3.2% PFA in 1xPBS we let the brains fix for 2-4 hours at room temperature. After each hemisphere of the brain was sectioned sagittally with a vibratome at 70-100 µm thickness. The brain slices containing HVC were collected and incubated at 4 C overnight with the primary rabbit anti-GFP (AB3080P, EMD Milipore) (blocked in 10% donkey serum in 0.2% Triton 1xPBS). On the second day, the brains were washed in 0.05% Triton 1xPBS and incubated for 2 hours in the dark at room temperature in the secondary goat anti-rabbit 488 (ab150077). Next, the brain slices were washed and mounted in Fluoromount (Sigma). Confocal images were taken with the LSM800.