Two ferrets were trained on a Go/No-Go delayed categorization task under appetitive reinforcement. Water-deprived ferrets had to classify click trains into two categories: target (Go) and non-target (No-Go) depending on the rates of click trains. Six rates were used, from 4 to 24 Hz in 4 Hz steps, and with a category boundary fixed at 14 Hz. To ensure the dissociation between categories and stimulus rates, one animal was trained with low rates as the Go sounds, while the second animal classified high rates as the Go sounds. Click trains were presented for 1.1 s and were followed by a 1-s-long delay in which the animals had to refrain from licking (Figure 1a). Licks during the subsequent 1-s-long response window were rewarded with water in Go trials and punished with a timeout in No-Go trials. Any licks before this response window (early licks) resulted in an aborted trial and were punished with a timeout; these were called ‘early trials’.

Figure 1 with 3 supplements see all

Download asset Open asset

Ferrets categorized the click trains (ferret P: d' = 1.2 ± 0.5, n = 35 sessions; ferret T: d' = 1.1 ± 0.2, n = 39 sessions; Figure 1b) with a bias to lick for the No-Go rates close to the category boundary (12 Hz for the animal shown in Figure 1b). This was found in both ferrets (Figure 1—figure supplement 1a and b) regardless of the mapping between stimulus rates and category, suggesting that their decision criterion was relatively liberal and impulsive. First lick probability was higher in Go than in No-Go trials throughout the entire trial duration (Figure 1c), confirming the categorical response profile. We also observed a build-up of lick probability during the delay, with early responses being quite common in Go trials. Early trials and error trials (misses and false alarms) were discarded in the analysis unless otherwise specified.

We chronically recorded neural activity from a total of 816 units in A1 (ferret P: 575; ferret T: 241), while the two animals alternately listened passively to the stimuli or actively engaged in categorizing them. Figure 1d depicts the average click responses across all cells, showing a significant increase in firing rate relative to baseline activity in all conditions. Firing rates also steadily ramped up during the delay period in Go trials compared to No-Go’s (Figure 1d and e; both animals in Figure 1—figure supplement 1c and d). Importantly here we considered only hit trials in which the animals were not licking the spout until the response window. Neurons with significant changes in firing rate during the delay period did not show any modulation of activity during the passive state (Figure 1e). Ranking neurons by delay activity during the passive condition did not reveal a similar pattern neither in passive or active states (Figure 1—figure supplement 3), reflecting a pattern of response specific to the task-engaged delay period. Additionally, we found that the firing rates of these neurons increased during the delay (Figure 1f, paired t-test active p=0.01, passive p=0.15, n = 816 units; both animals in Figure 1—figure supplement 2; ferret P: active p=0.04, passive p=0.83, n = 575 units; ferret T: active p<0.001, passive p=0.35, n = 241 units). Nevertheless, we noticed some heterogeneity in the delay activity, with a proportion of neurons showing large suppression during this period (top neurons in Figure 1e).

Because A1 activity was heterogeneously modulated during the post-Go sound delay (Figure 1f), we relied on population decoding to examine the categorical representation. To do so, we trained linear decoders to discriminate Go and No-Go trials based on single-trial population activity (Figure 2a). During task engagement, population responses steadily discriminated between Go and No-Go categories along the entire course of the trial (Figure 2a, both animals in Figure 2—figure supplement 1). Decoding accuracy diminished at stimulus offset, but subsequently increased during the delay period, ultimately peaking in the behavioral response window (Figure 2a, Figure 2—figure supplement 1). By contrast, accuracy in the passive context decreased throughout the delay period. As a control, we performed recordings in an untrained (naive) animal using the same set of stimuli (Figure 2—figure supplement 2a). The accuracy of passive context decoder was also similar to what is observed in the naive animal (Figure 2—figure supplement 2b). Note that decoding was performed at the single-session level (11.3 ± 4.9 neurons per session for ferret P; ± SD; n = 35 sessions and 5.2 ± 3.2 neurons per session for ferret T; n = 39 sessions), explaining the modest but above chance-level performance of the decoder.

Figure 2 with 5 supplements see all

Download asset Open asset

We then looked further into how population code evolves during the trial. For this, we compared decoders trained at different time bins throughout the trial. We found that for both passive and task-engaged sessions, the direction of the decoding axis was mainly preserved during the stimulus epoch (black outline in Figure 2b), and the corresponding decoders trained on passive and task-engaged data were correlated (Figure 2—figure supplement 3). Furthermore, these decoders were uncorrelated with the decoding axis during the delay period of the task-engaged condition (white outline in Figure 2b), indicating a behavior-dependent change of the population code at the sound offset. This delay activity pattern further persisted during the response window, as shown by the correlation between delay and response window decoding axes (pink outline in Figure 2b). This suggests a time-independent categorical representation in A1 population activity throughout the delay and response window epochs.

While categorical encoding is evident during task engagement after the stimulus end, it remains unclear whether it was biased toward stimulus-related or category-related responses during the stimulus. Indeed, both stimulus-related (click train rate) and category-related responses are captured by the decoders (Figure 2—figure supplement 4a and b). To disentangle sensory and categorical contributions, we computed a population-level linear regression of the neuronal firing rates at each time bin in each session (see ‘Methods’ for details). Single-neuron activity showed linear relationship with click rates (example in Figure 2c) or robust categorical encoding during and after the stimulus (Figure 2d and e). To model this type of responses, we used two regressors: (i) stimulus rates as a sensory regressor and (ii) the binary category (Go or No-Go) as a category regressor. We also fitted two additional models in which the labels of one of the regressors were shuffled (Musall et al., 2019), which allowed us to compute the relative contributions using a coefficient of partial determination (CPD) from each regressor as the decrement in variance due to shuffling one of the regressors (Fisher et al., 2019)⁠. During task engagement, we found that the CPD of the category regressor increased during the stimulus period and persisted throughout the delay period, before increasing at the response window (Figure 2f and h; paired t-test on category regressor, delay vs response window; ferret P: n = 395 neurons, p<0.001; ferret T: n = 203 neurons, p<0.01). Interestingly, both task-engaged and passive category CPD were larger than chance during stimulus presentations (Figure 2f–i), an effect absent in the naive animal (Figure 2i), which indicates training-dependent encoding of categories. Because the variance captured by the category regressor during the delay could be related to other factors than pure behavioral categories, such as reward delivery or licking, we fitted another model including lick and reward regressors (see ‘Methods’). We found similar category axes in the passive and active conditions when accounting for licks and rewards (Figure 2—figure supplement 5). We consider further the possible link of this delay activity pattern with reward expectation or motor preparation (see ‘Discussion’). For practical reasons, we will refer to the delay activity as category-related in the following sections.

We then examined how the category information during the delay emerged from the mixed representation of stimulus and category we found during stimulus presentation. Regression axes define encoding axes along which population activity carries information about stimulus click rate or category. We thus projected population activity onto the category neural axis to track the categorical representation in the population at each moment in time. This procedure sums the neuronal responses weighted by the coefficients extracted from the earlier regression analysis and reduces A1 population dynamics to one information-bearing dimension encoding category through time, independently from sensory information. We did so in both the stimulus and the delay periods for the passive and task-engaged conditions in order to compare the patterns of categorical encoding across conditions. Projections on the category neural axis (called thereafter categorical responses) revealed a stable encoding of categories during sound presentation, regardless of the stimulus identity (Figure 3a and b), indicating that these categorical responses were not contaminated by sensory activity.

Figure 3 with 3 supplements see all

Download asset Open asset

Categorization can be defined as the maximization of neural differences between categories and a minimization of the differences within a category. Therefore, we investigated stimulus discriminability along the category neural axis. We designed a category index (see ‘Methods’) which compared the neural distance between stimuli at the category boundary (12 and 16 Hz) against pairs of stimuli within categories. We found that categories were effectively present in both passive and task-engaged states during stimulus presentation with equal magnitude (Figure 3a and b, Figure 3—figure supplement 1a and b). In contrast, significant categories were found only in the task-engaged state during the delay period (Figure 3c and d).

To further determine if task engagement induced a targeted change along the categorical neural dimension, we examined the temporal profiles of the Go and No-Go categorical responses. Here, we found that there was a clear shift of population coding between the two behavioral conditions: No-Go categorical responses aligned with the pre-stimulus spontaneous activity during task engagement (Figure 3a), inducing an asymmetry between No-Go and Go categorical responses (Figure 3e, both animals in Figure 3—figure supplement 1c and d), a characteristic that is not observed in the passive projections (Figure 3b). A No-Go suppression index measuring the relative displacement of the No-Go responses closer to the projections of spontaneous activity (Figure 3f) showed that the suppression was confined to the stimulus period (Figure 3—figure supplement 1e and f, bottom row; ferret P: p<10^–6, n = 35 sessions; ferret T: p<10^–4, n = 39 sessions), and absent in the delay period (ferret P: p=0.70; ferret T: p=0.46). Crucially, projections of the passive population responses onto the active category neural axis did not show the suppression of No-Go sounds, confirming that this shift originated from a change in the structure of the population responses during task engagement, and not due to the weights of the regressor themselves (Figure 3—figure supplement 3). In summary, these findings demonstrate a task-induced asymmetry in categorical representations which specifically targets the relevant neural dimension.

The categorical signal in the stimulus period exhibits short latencies (Figure 3a) consistent with a feedforward mechanism of early categorization taking place in A1. We wondered whether this early categorical information may influence later categorical information captured in the delay period. However, the categorical axes in the early stimulus and delay periods were not correlated (white frames in Figure 4a). So how could the categorical information present in these two distinct intervals be still linked? To find out, we tested directly whether the categorical responses during the stimulus period correlated with those during the delay. Specifically, we examined the single-trial fluctuations of categorical responses in each period and discovered that the two categorical responses were correlated (one session in Figure 4b, all sessions in the inset; p<0.001, permutation test, n = 74 sessions). Note that the stimulus and delay categorical responses were computed from uncorrelated projection axes (Figure 4a, Figure 4—figure supplement 1) that were originally derived from distinct non-overlapping epochs, and hence the single-trial fluctuations in categorical responses need not have been correlated in any way.

Figure 4 with 3 supplements see all

Download asset Open asset

Another interesting relationship is that between the categorical responses and the licks. We have observed that population activity on hit trials ramped up throughout the delay period (Figures 3,43c and 4c, categorical responses during the delay). By aligning single-trial projections onto lick timings during the response window, we found that the categorical activity built up during the delay period and culminated when the animal licked (Figure 4d, categorical responses centered on licks), indicating that delay activity was anticipating behavioral responses. We further tested whether the temporal dynamics of the categorical responses correlated with the timing of the animal’s behavioral response. To do so, we took advantage of early Go trials when the animals licked before the response window (see lick histogram in Figure 1c), so that we accessed to Go trials in which first licks occurred earlier than what we had in hit trials. We then sorted all the collected trials into three groups: early trials with licks early in the delay period (first half), early trials with licks late in the delay (second half), and hits. We found significantly faster response build-up rates in early trials than in hit trials (Figure 4c). Early trials with early licks in the delay exhibited the steepest response build-ups (Figure 4c), a pattern that was not observed in the categorical responses during the stimulus period (Figure 4—figure supplement 2). We then aligned all trials to their respective reaction times and found that categorical responses of early and hit trials culminated at the licking time (Figure 4d). This alignment was done on lick times, that is, independently of neural activity, indicating a build-up of population-level dynamics anticipating behavioral response.

Finally, we searched for the neural correlates of incorrect perceptual decisions. In particular, we wanted to pinpoint which encoding stages were degraded during incorrect categorizations. We envisioned three different hypotheses causing the error trials: (a) sensory information was intact in error trials, but conversion to the correct category during the stimulus was incorrect; or that (b) sensory-to-category processing during the stimulus period is correct, but the anticipatory activity is impaired during the delay; or that (c) sensory information was in fact degraded early on during the stimulus, leading to a loss of categorical information.

We first sought to test whether the error trials correlated with an inversion of the neural responses between the Go and No-Go categories, as would be predicted by hypothesis (a). To do so, we used Go/No-Go classifiers trained with correct trials (as in Figure 2a) and reported the classifiers’ performance in predicting the expected (and not actual) behavioral categories. If false alarms were fully behaving like hits, and misses like correct rejections, we should observe below-chance decoding accuracy. That was not the case, but instead we found a decrease in decoding accuracy during the sound period and a decoding performance at chance level during the delay (Figure 5a and b, Figure 5—figure supplement 1 for separate projections of false alarm and miss trials; decoding accuracy during delay for incorrect trials: p=0.99, permutation test). As a control, we noted that decoding accuracy dipped below chance during the response window confirming that motor information was confined to this late time period (Figure 5a). This indicates that incorrect decisions could be traced back to the stimulus period, excluding hypothesis (b) in which only the delay activity was incorrect during error trials.

Figure 5 with 2 supplements see all

Download asset Open asset

We then further tested the hypothesis (c) of a degradation of sensory information during the stimulus period. In line with this interpretation, we found that sensory responses were lower in incorrect compared to correct trials (Figure 5c, Figure 5—figure supplement 2a and b; stimulus period-sensory index: ferret P: p<10^–5, n = 35 sessions; ferret T: p<10^–3, n = 39 sessions; stimulus period-category index: ferret P: p=0.01, n = 35 sessions; ferret T: p=0.35, n = 39 sessions; paired t-test across sessions, individual boxes were tested with permutation test). Categorical responses were further degraded during the delay period (delay-category index [Figure 5d and Figure 5—figure supplement 2b and c] for both animals; paired t-test across sessions; ferret P: p<10^–4, n = 35 sessions; ferret T: p<10^–5, n = 39 sessions). Altogether, this suggests that errors likely originated from an improper encoding of the stimulus that subsequently led to an incorrect categorization and behavioral response.

Previous findings have found categorical responses in A1 with a marked increase of response contrast at category boundaries (Yin et al., 2020; Xin et al., 2019)^⁠. Here, we focused on revealing the temporal dynamics and evolution of categorical encoding both in the population average and in the population code. Regression analyses allowed us to disentangle the binary categorical representation from the more graded representation of the sensory properties (i.e., click train rates) (Figure 3—figure supplement 2). Similar to a previous study in the mouse (Xin et al., 2019⁠), we found that the encoding of stimulus sensory properties within behavioral categories was stable across behavioral states. In Xin et al., 2019⁠, the authors found that categorical representations in A1 matched behavioral decision as early as at stimulus offset, and was maintained during several seconds. Such choice-related activity has been demonstrated to be the result of feedforward and feedback flow between primary sensory cortices and downstream regions (Dehaene and Changeux, 2011; Yang et al., 2016⁠). Our observations are consistent with this view as we found that perceptual choice influenced the category decoded during the response window (Figure 5a). Nevertheless, we used longer stimulus followed by a long delay in our study, which allowed us to decouple the early categorical information building up during stimulus presentation from the late choice report of the animal. In addition, categorical information declined during stimulus and delay periods, suggesting that incorrect trials were due to a lack of proper categorical information early on during the stimulus itself.

Our analyses showed that, even though categorical encoding was present in passive and active states in trained animals, task engagement induced a suppression of No-Go responses during the stimulus. This was not due to a lack of sound-evoked responses to No-Go sounds, as shown by the average population peristimulus time histogram (PSTH) (Figure 1d) and its sensory responses (Figure 3—figure supplement 2). Instead, this asymmetry in the categorical representation stems from a task-induced population-level enhancement of Go and suppression of No-Go categorical responses, consistent with our previous findings (Bagur et al., 2018) in these Go/No-Go paradigms. To elaborate, No-Go sounds instruct the animal to maintain the same behavioral output (lick inhibition) as in periods of silence when spontaneous activity is measured. We therefore proposed that the alignment of the No-Go population responses with the spontaneous activity reflects the identical behavioral meaning of these two epochs. Here, we found a similar mechanism (suppression of No-Go responses) which was confined to the categorical responses extracted through linear regression. This suggests an interesting mechanism in which population dynamics could multiplex behavior-independent sensory representations and task-modulated categorical encoding. Intriguingly, this is reminiscent of population dynamics found in prefrontal cortex (PFC) (Mante et al., 2013), where sensory information and decision formation were represented along different neural dimensions.

In line with a role of A1 in the category build-up, we found that categorical information was degraded in error trials, which is consistent with an early mapping of individual stimulus into generic categories at the level of primary sensory areas. We thus propose that A1 has an initial contribution in the feedfoward formation of behavior-dependent categories. During behavior, higher areas would access an explicit representation of the behaviorally relevant categories by reading out the asymmetrical population-level encoding of Go and No-Go sounds in A1 (Bagur et al., 2018). These areas would then utilize the A1 categorical representation to amplify and strengthen the ongoing category, possibly passing or gating it to motor-related regions depending on the behavioral state (Atiani et al., 2014).

Categorical responses changed at stimulus offset to maintain a Go-specific prolonged activity during the delay period, consistent with other studies showing choice-related activity in A1 (Bizley et al., 2013; Niwa et al., 2012b; Niwa et al., 2012a; Guo et al., 2019⁠). We do not interpret this activity as efferent copies directly sent from motor regions (Schneider et al., 2018), as one would expect such motor-related activity to have shorter latencies (~100–300 ms; Bagur et al., 2018; Orlandi et al., 2023⁠). Here, sound category was decodable throughout the entire delay period (Figure 2a), which does not match with short-latency efference copies. We have also found that the delay activity is not closely locked in time to licks since false alarm trials, in which the animals licked during the response window, were not decoded as hit trials during the delay period (Figure 5b).

Other studies investigating the role of A1 in auditory Go/No-Go tasks with a delay did not analyze neuronal responses after Go sounds (Yu et al., 2021; Huang et al., 2016), which may explain why this sustained activity has not been previously reported. Interestingly, delay activity in sensory cortices is thought not to be causally involved in the behavioral response of trained animals (causal inactivations in V1 [Goard et al., 2016^⁠] and A1 [Yu et al., 2021^⁠]). We note that this sustained activity after Go sounds can be caused by at least three different factors.

First, delay activity in A1 could be the result of a feedback signal from higher-order decision-related areas signaling the chosen category and maintaining it in memory by engaging A1 in a network of parietal and frontal areas (Goard et al., 2016)^⁠. In this framework, previous works in the somatosensory system have shown that choice-related information flows bidirectionally between primary and secondary somatosensory cortices, with choice-related information emerging in primary sensory cortex, then fed to downstream areas that further feedback enriched choice-related information to the primary field (Kwon et al., 2016⁠). Consistent with this hypothesis, we have demonstrated that trial-to-trial fluctuations of categorical responses during the sound period correlate with the amplitude of the delay categorical signal (Figure 4b), possibly suggesting that category-related information during the stimulus and delay periods is part of this communication loop.

A second interpretation is that delay activity is the result of motor preparation that would unfold over several hundreds of milliseconds (Musall et al., 2019⁠). The pattern of activity observed during early trials is consistent with this view, with faster build-up when the animals licked earlier. However, a similar pattern was not observed on false alarm trials, contrary to what would be predicted for this interpretation.

A last possibility is that delay activity signals reward expectation (Chubykin et al., 2013⁠), a type of response which increases through learning (Musall et al., 2019; Poort et al., 2015⁠). This delay activity would then reflect post-learning residual top-down feedback, or alternatively may be the signature of an eligibility trace, that is, a persistent activity necessary for bridging the gap between the sound and the response window (Raybuck and Lattal, 2014^⁠). Overall, disentangling these different interpretations will require further experiments.

Adult female ferrets (Mustela putorius furo) obtained from Marshall BioResource were used for this study. The animals were 1–3 years of age, weighing 500–800 g, and were housed in pairs or trios with a normal day–night light cycle and free access to water during weekends. Ferrets were on a water-controlled protocol in which their water intake is restricted during the weekdays. Water was delivered during behavioral sessions as a reward. To maintain a stable weight, we provided ad libitum water for 1–2 hr post behavior. The animals’ weights were daily monitored and maintained at 80% of pre-experiment weight.

Two adult female ferrets (M. putorius furo) were trained on an appetitive Go/No-Go delayed categorization task and one additional ferret was used for naive recording. The animals were head-fixed in a custom-made tube during training and recording sessions, and the stimuli were presented from a calibrated earphone (Sennheiser IE800, HDVA 600 amplifier). Ferrets had to classify click trains into two categories: target (Go) and non-target (No-Go) depending on the rates of click trains. Six rates were used, from 4 to 24 Hz in 4 Hz steps, and with a category boundary fixed at 14 Hz. To ensure the dissociation between categories and stimulus rates, one animal was trained with low rates as the Go sounds, while the second animal classified high rates as the Go sounds. Clicks were monopolar, rectangular pulses of 1 ms duration with amplitude set at 70 dB sound pressure level. A trial started with a pre-stimulus silence of 0.5 s followed by a click train (Go or No-Go) of 1.1 s (Figure 1). The animals were trained to wait for a response window that started after a delay of 1 s following stimulus offset. A hit (lick on the response window for Go stimuli) was rewarded with 0.2 ml of water. An LED attached to the water spout emphasized the delay period in which the animal had to restrain from licking. Early trials (lick during stimulus and delay period) and false alarms (lick on the response window of a No-Go sound) were punished with a timeout of 10 s. In each session, Go and No-Go stimuli were presented in a pseudo-random manner. In the absence of delay, ferrets learned to associate the categories in 1 wk, and we then slowly increased the delay between stimulus offset and response window. It took several weeks for the ferrets to be trained on the full task structure. Initially we trained with extreme categorical stimuli (4 and 24 Hz) that were easy to learn, and after reaching a consistent performance (d’ > 1), we progressively introduced other stimuli.

To head restrain during training and obtain stable neurophysiological recordings, we implanted the ferrets with a stainless steel headpost. The day prior to the surgery ferrets were injected with antibiotics (Baytril, 12.5 mg/kg subcutaneous) to minimize infections arising from the surgery. On the day of surgery, ferrets were deprived of water and food 90 min prior to the surgery. After sedation (medetomidine, 0.08 mg/kg, subcutaneous), anesthesia was induced with ketamine (5 mg/kg, intramuscular). Animals were kept under deep anesthesia (1–2% isoflurane) throughout the surgery and vitals (ECG, pulse, oxygenation, and rectal temperature) were continuously monitored. We also medicated the animals with atropine sulfate (0.2 mg/kg) to stabilize salvation and control arrhythmia arising from anesthesia. Using a complete sterile procedure, the animal skull was surgically exposed by an incision to the skin along the media crest down to the neck. The temporal muscles were carefully removed from the medial crest to the beginning of the zygomatic arch and the lateral wing at the lateral end of the nuchal crest. Using a stereotaxic apparatus, the headpost was mounted on the skull using methyl methacrylate-based dental adhesive resin cement. Stainless steel screws were anchored along the areas surrounding the auditory cortex, leaving a cavity for easy access to the auditory cortex. The typical horseshoe shape of the auditory cortex was marked with nail polish. Finally, the surrounding areas were filled with poly-methyl methacrylate-based bone cement to stabilize the implant. Antibiotics (Baytril, 12.5 mg/kg, subcutaneous) and analgesic (meloxicam, 0.05 mg/kg, oral) were administered to the animal following the surgery.

We allowed a 2-week postoperative care for the animals to recover from the surgery. Antibiotics were continued for 7 d and anti-inflammatory and analgesics were administered for 4 d. Animals were habituated to a head-restrained custom-made horizontal plastic tube a few days prior to training session. Experiments were approved by the French Ministry of Agriculture (protocol authorization: 21022) and strictly comply with the European directives on the protection of animals used for scientific purposes (2010/63/EU).

In a separate surgery, we chronically implanted 32-channel metal electrodes arrays (MEA; Pt-Ir, MicroProbes, 8 × 4, electrode of impedance 2.5 MΩ with 0.4 μm distance between the electrodes) over the prior marked auditory cortex. We custom-designed the chronic implant with MEA inserted in a drive-shuttle system having a flexible control of the array vertical movement. The base of the drive was sealed with a stretchable silicon membrane sheet to stop flowing any residues into the drive. Before the implantation, the electrodes were moved down such that the apex popped out of the silicon membrane. Under surgical anesthesia (isoflurane 1 %), we removed the cement above the location marked during the surgery. We then performed a 4 mm × 4 mm craniotomy. This craniotomy allowed us to identify the core regions of the auditory cortex (middle ectosylvian gyrus) by visual inspection of the tip of the ectosylvian gyrus. We carefully removed the transparent dura to ensure that the array penetrated the brain without additional strain. The drive-shuttle system was placed on the brain surface using a stereotaxic apparatus and the entire system was fixed to the skull using bone cement. To minimize vibrations that could be caused by shocks on the drive, the chronic implant was enclosed in a custom-made tube cemented to the implant. Immediately after the surgery we slowly lowered the electrodes and observed physiological activity, allowing us to verify the electrodes moved inside the brain.

Each recording session consisted of passive and active sessions. Recordings were performed head-fixed in a soundproof chamber. In the passive sessions, the water spout was removed. Continuous electrophysiological recordings were digitized (31,250 Hz), amplified (15,000×), and band-passed between 300 Hz and 7000 Hz using a digital acquisition system (Blackrock Cereplex). Band-passed signals were monitored online and units (including multi- and single-) were identified by spikes crossing a threshold of 3 SD of baseline noise. The data acquisition was done using an open-source suite MANTA v. 1.0 (Englitz et al., 2013). We used a custom-made open-source software Behavioral Auditory PHYsiology (BAPHY) written in MATLAB for sound delivery, recording, behavioral monitor and online analysis.

To identify units, we presented band-pass noise (0.2 s duration, 1 octave bandwidth) and pure tone stimuli to the animal using earphones (Sennheiser IE 800). Primary auditory cortical responses were identified by analyzing tuning properties to 100 ms tone pips of random frequencies spanning 4 octaves and temporally orthogonal ripple combinations (STRF) (Depireux et al., 2001). A1 responses show sharp tuning to random tones and single-peak, short-latency STRFs (Atiani et al., 2014; Fritz et al., 2003; Elgueda et al., 2019)^⁠. Finally assessing the STRFs, we continued with the experimental protocols to record the neuronal responses.

We performed offline spike sorting on thresholded signals using PCA-based customized spike sorting routines written in MATLAB. Single- and multi-unit responses were identified by spiking shape and manually adjusting the PCA clusters (Fritz et al., 2003)⁠. A total of 575 (ferret P) and 241 (ferret T) multi-units were identified and used for further analysis (Table 1). Spike sorting was done on concatenated passive and active sessions. We obtained 11.3 ± 4.9 neurons per session (± SD; n = 35 sessions) for ferret P and 5.2 ± 3.2 neurons per session (± SD; n = 39 sessions) for ferret T.

We performed both population decoding and linear regression analysis session-wise with simultaneously recorded neurons. All statistics across passive and active states were done with paired t-tests. Correlations were linearly assessed (Pearson correlation). Unless specified otherwise, error bars showed ± 1 SEM over sessions. The codes for reproducing the analysis are available on the following repository: https://github.com/rupeshjnu/A1-Category, (copy archived at Rupesh, 2023). The electrophysiological data is publicly available at https://zenodo.org/records/8371084.

The funders had no role in study design, data collection and interpretation, or the decision to submit the work for publication.

We thank Joao Barbosa and Célian Bimbard for deep reading and fruitful comments. This work was supported by ANR-17-EURE-0017 and ANR-10-IDEX-0001-02, ERC 787836-NEUME and ANR-19-CE37-0016 to SAS, and ANR-JCJC-DynaMiC to YB.

Experiments were approved by the French Ministry of Agriculture (protocol authorization: 21022) and strictly comply with the European directives on the protection of animals used for scientific purposes (2010/63/EU).

1. Preprint posted: December 20, 2022 (view preprint)
2. Received: December 20, 2022
3. Accepted: November 12, 2023
4. Accepted Manuscript published: November 16, 2023 (version 1)
5. Version of Record published: January 17, 2024 (version 2)

© 2023, Chillale et al.

This article is distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use and redistribution provided that the original author and source are credited.