Separate Functional Subnetworks of Excitatory Neurons Show Preference to Periodic and Random Sound Structures

Abstract

Auditory cortex (ACX) neurons are sensitive to spectro-temporal sound patterns and violations in patterns induced by rare stimuli embedded within streams of sounds. We investigate the auditory cortical representation of repeated presentations of sequences of sounds with standard stimuli (common) with an embedded deviant (rare) stimulus in two conditions, Periodic (Fixed deviant position) or Random (Random deviant position). We used extracellular single-unit and two-photon Ca²⁺ imaging recordings in layer 2/3 neurons of the mouse (Mus musculus) ACX of either sex. Population single-unit average responses increased over repetitions in the Random condition and were suppressed or did not change in the Periodic condition, showing general irregularity preference. A subset of neurons showed the opposite behavior, indicating regularity preference. Furthermore, pairwise noise correlations were higher in the Random condition than in the Periodic condition, suggesting a role of recurrent connections in the observed differential adaptation. Functional two-photon Ca²⁺ imaging showed that excitatory (EX), and inhibitory (IN) neurons [parvalbumin-positive (PV) and somatostatin-positive (SOM)] also had different categories of long-term adaptation as observed with single-units. However, examination of functional connectivity between pairs of neurons of different categories showed that EX-PV connected pairs behaved opposite to the EX-EX and EX-SOM pairs, with more connections outside category in Random condition than Periodic condition. Finally, considering Regularity, Irregularity, and no preference of connected pairs of neurons showed that EX-EX and EX-SOM pairs were in largely separate functional subnetworks with different preferences, not EX-PV pairs. Thus, separate subnetworks underlie coding of periodic and random sound sequences.

SIGNIFICANCE STATEMENT Studying how the auditory cortex (ACX) neurons respond to streams of sound sequences help us understand the importance of changes in dynamic acoustic noisy scenes around us. Humans and animals are sensitive to regularity and its violations in sound sequences. Psychophysical tasks in humans show that the auditory brain differentially responds to Periodic and Random structures, independent of the listener's attentional states. Here, we show that mouse ACX L2/3 neurons detect changes and respond differently to patterns over long-time scales. The differential functional connectivity profile obtained in response to two different sound contexts suggests the vital role of recurrent connections in the auditory cortical network. Furthermore, the excitatory-inhibitory neuronal interactions can contribute to detecting the changing sound patterns.

Introduction

Neural circuitry operates to maximize the information flow between the environment and its output and decrease the redundancies in the dynamically varying stimulus (Chechik et al., 2006; Gaucher et al., 2013). To do that, it becomes necessary to understand the statistical structure of the stimuli to interact with the environment efficiently (Winkler et al., 2009; Chait, 2020). The intricate statistical features present in the stimuli form a basis for the neural circuitry to identify relevant features, including expected and unexpected changes in the stimuli and enhance their representation up the cortical hierarchy (Aghamolaei et al., 2016; Parras et al., 2017). This assimilated information is further used in the hierarchy, possibly to predict future stimuli based on past experiences and interpret meaningful information. (Winkler et al., 2009; Pearce et al., 2010; Chait, 2020).

The auditory cortex (ACX) circuitry uses the varying spectro-temporal statistical features of stimuli to generate empirical distributions of multiple components present within the time-varying stimulus. Such studies have been performed across multiple modalities in humans like vision (Turk-Browne et al., 2009), motor control (Bestmann et al., 2008), speech (Saffran et al., 1996), and audition (Andreou et al., 2011; Sohoglu and Chait, 2016; Heilbron and Chait, 2018). Within the auditory domain, it is well known that probabilities of stimuli play an important role in response sensitivity (Ulanovsky et al., 2003; Yaron et al., 2012). Stimuli with dynamic probabilities have been used extensively to study neural sensitivities across different species and in multiple stations up the auditory hierarchy (Anderson et al., 2009; Malmierca et al., 2009; Taaseh et al., 2011). However, despite a few studies (Yaron et al., 2012; Aghamolaei et al., 2016; Southwell and Chait, 2018), not much is known about the temporal effects over long-time scales, especially in the context of repetitions of sound sequences. To maximize the information flow across the hierarchy, neural circuits often show adaptation at a local synaptic level (Yaron et al., 2012; Hershenhoren et al., 2014) both in feedforward and recurrent connections. Through the past studies, it is well known that stimuli with low probability of occurrence elicit the highest response, and this has been extensively studied as Deviant selectivity (Khouri and Nelken, 2015), oddball selectivity (Camalier et al., 2019; Mehra et al., 2022; Srivastava and Bandyopadhyay, 2020), and mismatch negativity (MMN). An analog of oddball selectivity, stimulus-specific adaptation (SSA), has been extensively studied at the level of single neurons. SSA refers to selective attenuated response to a stimulus with a high probability (Standard stimuli), whereas the response to a low probability stimulus (Deviant) remains unchanged or is enhanced (Ulanovsky et al., 2003).

A number of lines of evidence suggest that neurons in the ACX are sensitive to the auditory environment's statistical regularities. Previous studies on humans have focused on investigating the effect of context on deviance by violating the variety of patterns (Herrmann et al., 2015; Khouri and Nelken, 2015). However, less is known about how auditory circuitry responds to changes in patterns over a long time scale in a noisy environment. The current study tries to bridge this gap using Standard and Deviant Tones and Noise in Periodic and Random oddball stimuli, highlighting long adaptation time scales and adaptation to entire stimulus structures. Our stimulus structure is more naturalistic compared with the usual use of tones for standard and deviant, considering most mouse vocalizations of communicative significance are tonal in nature (Holy and Guo, 2005; Agarwalla et al., 2020) amid an ongoing broadband background. We further establish the differences between the neurons encoding properties in ACX in response to the different patterned and Random structures in sound sequences. A further step in our study aims to determine the role of individual inhibitory interneurons, PV and SOM, on such adaptation. With two-photon imaging, we decipher a possible role of functional connectivity among the excitatory (EX) and inhibitory (IN) neurons in carrying out the differential encoding of Random and Periodic sequences at the network level.

Materials and Methods

In vivo extracellular recordings

Surgical procedures

Adult C57BL6/J JAX mouse (Mus musculus), age postnatal day (P)30–P40, of either sex were initially anaesthetized under deep anesthesia (5% isoflurane) inside the induction chamber. After that, during the surgery, the level of anesthesia was lowered down to 1.5–2.5%. The internal body temperature was monitored and continuously maintained at 37°C using the heating pad throughout the experiment. The left temporal portion of the skull was exposed by performing a central incision and retracting the scalp. Tissue clearance and sterilization of the exposed area were performed by applying 3% H₂O₂ and alcohol, respectively. A craniotomy was made over an estimated left auditory cortical area bounded by the temporal ridge, lamboid suture, and ventral and rostral squamosal suture. Recordings were confirmed from A1 based on the direction of the tonotopic gradient.

Recording procedure

Recordings were performed using the 4 × 4 multielectrode array (Microprobes, 125-μm interelectrode spacing) of 3–5 MΩ (MicroProbes) impedance. L2/3 responses were probed between 200–300 μm from the cortical surface using a micromanipulator (MP-285, Sutter Instrument Company). Signals were acquired after passing through unity gain (1×) headstage, followed by a preamp (Plexon, HST16o25) with 1000× gain. The wideband signal (Local Field Potential or LFP, 0.7 Hz to 6 kHz) and spike signal (150 Hz to 8 kHz) were acquired in parallel through National Instruments Data Acquisition Card (NI-PCI-6259) at 20-kHz sampling rate. Further, offline/online analysis of obtained signals was performed using custom-written codes in MATLAB (MathWorks Inc.).

In vivo two-photon imaging

To prepare an animal for in vivo two-photon imaging, standard surgical procedures were performed as described before (Bandyopadhyay et al., 2010). In short, surgeries were performed under 1.5–2% of isoflurane anesthesia. The left temporalis was exposed by transecting the scalp. Further, the exposed area was cleared and sterilized using 3% H₂O₂ and alcohol. A 5-mm diameter circle was marked over the estimated auditory area. A metal plate was fixed using dental acrylic after keeping the mark in the center of the head plate. A cranial window was created by lifting a skull flap of 3-mm diameter over the left auditory area. After that, a craniotomy was filled with 1.5% low-melting agarose. Immediately after that, a 3-mm coverslip was embedded over the craniotomy to avoid brain pulsation. Animals were thereafter transferred to a soundproof chamber equipped with two-photon imaging microscopy. Throughout the experimentation, the animal temperature was monitored continuously and maintained at 37°C. During the imaging sessions, anesthesia was reduced down to 0.5–0.75% of isoflurane.

Viral injection

Adult mice (>40 d old, of either sex) were anaesthetized with isoflurane, as performed in extracellular recordings. Further, mice were injected with dexamethasone (2 mg/kg body weight) intraperitoneally and placed on a stereotaxic frame. The animal's body temperature was maintained at 37°C throughout the surgical procedure using a heating pad. An incision was made to expose the skull. A craniotomy of 3 mm diameter was performed to expose the ACX, based on landmarks (rhinal vein). Virus (AAV.Syn.GCaMP6s.WPRE.SV40) was loaded into a glass micropipette mounted on a Nanoject II attached to a micromanipulator and injected at a speed of 20 nl/min at the desired location of ACX in PV-tdTomato and SOM-tdTomato transgenic animals. The craniotomy was covered with a glass coverslip. Mice were left in their home cage to recover and virus to integrate and express for two weeks. Experiments were conducted two to three weeks after the viral injection.

Stimulus delivery

Sounds were generated through TDT RX6, attenuated using TDT attenuators (PA5), and delivered by TDT electrostatic speakers (ES1) driven by TDT drivers ED1. Sounds were presented at a distance of 10 cm to the right ear. The frequency response of the ES1 speaker measured with a microphone 4939 (Brüel & Kjær) showed a typical flat (±7 dB) calibration curve from 4 to 60 kHz. In a two-photon imaging setup, sounds were generated through TDT RZ6 multifunctional processor controlled by custom-written codes in MATLAB (MathWorks Inc.). The properties of each sound are given below.

Noise

Each 50 ms of 6- to 48-kHz white band Noise from 50 to 90 dB SPL with an interstimulus gap of ∼2 s was presented five times. The intensity of sound to be delivered was selected as 10 dB above the threshold level of noise for the majority of units recorded simultaneously.

Tone

Each 50 ms of pure frequency Tone with frequencies ranging from 6 to 48 kHz in 0.25-octave steps at 60–80 dB SPL (10 dB above the Noise threshold, see above) with an interstimulus gap of ∼2 s was presented five times. Response to these Tones was used to measure the best frequency and the tuning width of the neurons. The chosen intensity tended to be 20–30 dB above Tone thresholds.

Oddball paradigm N-f pair

A series of sound tokens (n = 15) each 50 ms long (intertoken gap of 250 ms) of the pattern SSS…SDS…SS, with S as Standard and D as Deviant, is considered as an iteration (Fig. 1A). The deviant D token occurred once in each iteration. A total of 30 iterations of such oddball sequences were presented with an interiteration gap of 2–8 s. S and D consisted of a pure Tone and broadband Noise (6–48 kHz with approximately equivalent sound level). Responses were always collected in pairs with S and D swapped (Noise as Standard, Tone as Deviant, followed by Tone as Standard and Noise as Deviant). A set of 30 iterations were presented with the Deviant in a fixed position (eighth) in the 15 tokens of each iteration (Periodic condition) or at a random position (Random condition). In the latter case, the Deviant position was anywhere between the 2nd and 15th token with uniform probability across each iteration.

Figure 1.

Differential long time scale of adaptation in response to Periodic and Random sound sequences. A, Schematic representation of the Periodic (Left) and Random (Right) oddball stimulus used in this study (Materials and Methods). For the Periodic condition (left), the S and D are shown in green and yellow, respectively, and for the Random condition (right), they are, respectively, in blue and red. The same color scheme is maintained in their respective response traces in B, C. The responses were calculated in a moving average fashion (Materials and Methods). The shaded boxes represent the tokens before the Deviant whose response was taken for moving average calculation of the Standard token. B, C, Example single-unit moving average PSTHs and LFP responses at the same site are shown as pairs (top and bottom) with irregularity or regularity preference. Examples with Tone as Standard or Deviant (B) or Noise as Standard or Deviant (C) in the Periodic (left) or Random (right) condition; Standard cases are shown in green or blue, and Deviant cases are shown in yellow or red (Periodic or Random). The black bar in PSTH responses represents 10 sp/s, and the red bar represents 50 sp/s. The black bar in LFP represents 100 µV.

Oddball paradigm f1-f2 pair

The sound structure is similar to the oddball paradigm N-f pair, as described above. However, in this paradigm, all sequences consisted of pure Tones with S and D as either f1 or f2. The two frequencies evoking high responses were selected, and the difference between f1 and f2 varied from 0.25 to 1.25 octaves.

Data analysis

Spike sorting

Spike sorting was performed using custom-written codes in MATLAB (MathWorks Inc.). Potential spikes were isolated based on fluctuations exceeding four standard deviations from the baseline from 8–16 independent channels. Further, spike waveforms were clipped from the raw data and projected into the space defined by the first three principal components. Clusters representing single-unit responses were identified by isolation distance between the cluster centroid quantified by K-mean clustering. The quality of the spike waveform was determined by visual inspection.

Rate calculation

Further analysis was performed on the firing rate calculated within the stimulus duration or multiple tokens as in the case of the oddball sequence paradigm. A single-unit was considered responsive if a response to at least one of the four stimuli (f/N or f1/f2 as Standard or Deviant) was significantly different from the baseline (300 ms preceding stimulus, two-tailed t test, p < 0.05). The inclusion criteria for the single-unit was a significant response to at least one of the conditions (Standard/Deviant in Random and Periodic condition). Always paired data (Random and Periodic sequence) was considered for further analysis.

Long time scales of adaptation

We measured the long scale of adaptation to the Standards and Deviants in Random and Periodic oddball sequences over 30 repetitions. To analyze the firing rate change across 30 iterations in Random and Periodic sequences, 21 mean firing rates were calculated by performing a moving average of firing rates across ten repetitions. A mean firing rate of all the Standard tokens preceding Deviant was considered for evaluating the behavior in mean firing rate of Standards across all the trials. The mean response of seven tokens preceding the Deviant was taken in the Periodic Deviant (8th position) case; however, the mean response of the number of tokens to be considered is variable (1–14) in Random Deviant (2nd–15th position) condition. To account for such biases in Random sequences, we performed a weighted mean average response by taking a product of the number of tokens before Deviant and the mean firing rate of all the tokens before Deviant. After that, in each moving average window, the mean firing rate over ten repetitions was divided by the sum of all Standard tokens preceding Deviant occurring in 10 repetitions of a given moving window. Normalized firing rates were computed over 21 moving averages by dividing them with the mean firing rate of the first moving average of 1–10 repetitions. For example, let the mean response of all the 30 iterations (mean response to all the standard tokens preceding the Deviant) be R₁, R₂, R₃…R₃₀ and let the number of standard tokens preceding the Deviant token in these 30 iterations be N₁, N₂, N₃…N₃₀. For the k^th moving window, the normalized mean moving average response is defined as $MR\left(k\right)=\frac{{\displaystyle {\sum }_{x=k}^{k+9}{R}_x/10}}{MR(1)}$ where $MR\left(1\right)={\displaystyle {\sum }_{x=1}^{10}{R}_x/10}$, i.e., the mean moving average response of the first window. The normalized weighted mean moving average response is defined as $WMR\left(k\right)=\frac{{{\displaystyle \sum }}_{x=k}^{k+9}\left(R_x{N}_x\right)/{{\displaystyle \sum }}_{x=k}^{k+9}{N}_x}{WMR\left(1\right)}$ where $WMR\left(1\right)={\displaystyle {\sum }_{x=1}^{10}\left(R_x{N}_x\right)/{\displaystyle {\sum }_{x=1}^{10}{N}_x}}$.

A slope across 21 normalized mean moving average firing rates for both Periodic and Random conditions in Tone and Noise was evaluated using the “regress” function in MATLAB to see how well the firing rates change over the trials. A significant positive (PS) or negative (NS) slope (F test, p < 0.05) meant that the responses increased or decreased, respectively, across iterations. All the other cases with nonsignificant slopes (F test, p > 0.05) were included in the zero slope category (ZS). We further calculated the change in slope sign(sign(slope_Random)-sign(slope_Periodic)) for both the Tone and Noise. A positive slope change meant more preference toward the Random condition, and a negative slope change meant more preference toward the Periodic condition. These were denoted as Irregularity preferring (IP) and Regularity preferring (RP). Cases, where no change was detected were termed as Non-preferring (NP).

Best frequency

The best frequency of the neuron is the frequency that evoked the highest response within the stimulus duration (50 ms) to the same sound level at which the oddball paradigm was presented.

Common selectivity index (CSI)

To quantify a neuron's selectivity to the Deviant, we used the CSI index.

$$\text{Tone}\left(\text{f1}\right)-\text{Tone}\left(\text{f2}\right)\text{case}:CSI=\frac{R_D\left(f_1\right)+R_D\left(f_2\right)-R_s\left(f_1\right)-R_s\left(f_2\right)}{R_D\left(f_1\right)+R_D\left(f_2\right)+R_s\left(f_1\right)+R_s\left(f_2\right)}$$

$$\text{Noise}\left(\text{N}\right)-\text{Tone}\left(\text{f}\right)\text{case}:CSI=\frac{R_D\left(f\right)+R_D\left(N\right)-R_s\left(f\right)-R_s\left(N\right)}{R_D\left(f\right)+R_D\left(N\right)+R_s\left(f\right)+R_s\left(N\right)}.$$

where R_S(.) and R_D(.) denote response to S and D, respectively.

Noise correlation

Noise correlation was computed between a pair of neurons recorded simultaneously in both Random and Periodic Deviant sequences. First, we represented a response to an oddball paradigm of an i^th neuron binned 50 ms at time t to the n^th iteration as $r_{i,n}\left(t\right)$. Noise correlation was calculated between the i^th and j^th neuron by calculating the Pearson correlation coefficient between two simultaneously recorded spike trains as $r_{i,n}\left(t\right)-\overline{r_i}\left(t\right)and{r}_{j,n}\left(t\right)-\overline{r_j}\left(t\right)$, where $\overline{r_i}\left(t\right)$ denotes the mean response of i^th neuron over all the trials.

LFP analysis

The raw electrical signals acquired at 20-kHz sampling rate were baseline shifted to have a mean of baseline (200 ms preceding the stimulus onset) at 0. After baseline zeroing, the acquired signals were notch filtered (50 Hz, Butterworth eighth order, to remove AC supply line Noise) and then bandpass filtered (between 1 and 300 Hz, Butterworth second order). The processed signals were further resampled at 2 kHz and stored for offline analysis. Response magnitude was quantified by considering root mean square (RMS) of the potential fluctuations within the stimulus duration or multiple tokens as in the case of the oddball paradigm. Only those channels corresponding to which single-units that evoked significant responses were considered for LFP analysis. Analysis performed on the LFP using RMS values was similar to spike analysis.

Imaging analysis

Preprocessing

The fluorescence data collected from in vivo two-photon calcium imaging was sorted to collect the mean fluorescence fluctuation within a 10-mm circle centered around the cyton body of the neurons. The spatial mean fluorescence within that circle was considered to be the activity of that cell. These data were baseline zeroed by 1–1.5 s of spontaneous fluorescence data to obtain the df/f values for every iteration separately.

Best frequency calculation

The tonal frequency that elicited the maximum significant response with respect to the spontaneous was selected as the best frequency. Paired t test was performed between the four frames preceding the onset and the mean activity within 500 ms of the Tone onset to sort out the significant responses (p < 0.05).

SSA analysis

The mean fluorescence from the onset frame to the frame just before the Deviant was taken for each iteration to find the Standard response. The Standard responses from these 30 iterations were smoothened by applying an across iteration moving average window of 10 iterations with one iteration step, making 21 data points for each cell and each stimulus. Unlike in electrophysiology data, normalization from the first window was not performed. We calculated the slope distributions (NS, ZS, and PS) and the change of slope distribution (RP, NP, and IP) like we did with the electrophysiology data.

Connectivity analysis

To assert connectivity between two cells, we selected 30 iterations Randomly from the original 30 iterations with replacement and computed the mean Pearson correlation between the two cells; this was done a total of 100 times, and a 99% confidence interval was constructed from these values. A chance distribution was also constructed by shuffling the iteration between the two cells and performing the above steps. If the mean of the bootstrapped distribution was higher than the mean of the shuffled distribution by a 99% confidence interval, those two cells were considered to be connected. In all the other cases, no connectivity was considered to be present. This method was used against the traditional Noise correlation because we wanted to remove the uncertainty in the degree of connectivity between two given cells. This connectivity was found separately for the four different kinds of stimuli (Noise-Periodic, Noise-Random, Tone-Periodic, and Tone-Random) to maintain their functional relevance.

Gini index

Having found the functional connectivity, we categorized the cells in nine different categories, depending on the combination of slopes in response to the two different kinds of stimuli (three slope types for Noise Standard multiplied by three slope types for Tone Standard). This was done separately for the connectivity of the Periodic stimuli and the Random stimuli. In both cases, if any pair was connected for at least one kind of stimuli (either Noise or Tone or both), we considered them connected. We generated a weighted graph with the nodes representing the nine different categories of cells and the edges representing the number of connections between two given categories. Since we were more interested in studying the functional connectivity within the categories, we ignored the connections shared by the neurons belonging to the same category (loops in a graph). We used the Gini index as a measure to quantify the sparsity of these graphs. To do that, we first generated a connectivity matrix “A” from those graphs with A(i,j) = A(j,i) = number of connections shared between category i and category j neurons. Since we were ignoring all the loops, the diagonal elements were all zero. We then computed the sum of this matrix along the rows, thereby computing the total number of edges emerging from all the nine nodes [note that A(i,j) = A(j,i), so the double-counting was included intentionally]. These nine values were arranged in ascending order and were divided by the total sum of all the nine values (to make sure that the sum of all these values is 1). A Lorenz curve (curve showing the cumulative sum of all these values arranged in ascending order, starting at zero and ending at 1) was constructed, and the area under this curve was computed using the trapezoidal approximation method (say B). Gini index is defined as G = 1-2B. Intuitively, it can be seen that if the graph is connected sparsely, the cumulative distribution will turn out to be very narrow, giving a lower value of B and hence a higher Gini index, thus if the neurons of this nine-category were densely connected with each other, we would see a lower value of Gini index, and if the intercategory connections were sparse, we would see a higher Gini index. The raw graph, which consisted of all the connections between the nine categories, was taken. A total of 500 Random replicates were constructed, where the edges were drawn at Random with replacements. This was done to generate a 90% confidence interval of the Gini index to compare the overall connectivity between the Periodic and Random conditions.

Relative connectivity probability

We selected all the connected pairs from the population 500 times with replacement and found out the number of connections that fell in each category. Dividing it by the total number of connections gave us the probability of different categories of connections in each of the 500 trials. Thus, we generated a 99% confidence interval of the probabilities of all the connection categories. We next constructed a chance distribution by shuffling the cells and generating pairs that may or may not be connected real. We calculated the relative connectivity probability by finding the difference between the means of these two distributions (µ_bootstrapped-µ_chance). Those connections whose mean differed from the chance mean by a 99% confidence interval (either higher or lower) were considered to be significant connections

Statistical tests

We performed χ² tests between slope distributions (NS:ZS:PS) obtained from Periodic and Random conditions to see whether the distributions were different or not and then performed a z-proportion test to compare which condition (Periodic and Random; P or R) had a higher number of slopes in the particular category (NS, ZS, and PS). We followed the same procedure to see whether the slope change distribution was different between Tone and Noise, followed by the same z-proportion test to see which stimuli (Tone or Noise, T or N) had a higher number of neurons in the particular slope change category (Regularity preferring, RP; Non-preferring, NP; Irregularity preferring, IP).

Results

We recorded single-unit, local field potential (n = 18 mice, extracellular recordings) and single neuron Ca²⁺ responses (Gcamp6s injected, n = 10 mice and Gcamp-Thy1, n = 14 mice) from the left ACX of the lightly anaesthetized mouse to oddball sound sequences presented to the right ear. We collected responses in pairs with the Standard and Deviant in a set swapped for both the Periodic (Fig. 1A, left) and Random conditions (Fig. 1A, right). Recordings were performed to collect responses to the Noise-Tone (N-f) oddball. The Noise-Tone stimuli were used to mimic a continuous noisy natural environment. Since Deviant selectivity and sensitivity in a sequence of sounds have been previously studied using two Tones (f1 and f2), we also used f1-f2 oddball stimulus with both S and D as two different Tones to investigate whether our results hold in such a case (see Materials and Methods, Stimulus delivery).

Differential adaptation of single L2/3 neurons and LFPs to random and periodic sound sequences

Extracellular single-unit recordings (200–250 µm in depth, putative L2/3) in the mouse primary ACX show distinct response adaptation patterns to the sound sequences presented in the two different contexts, Periodic and Random conditions. We analyzed the units with significant responses to the first Standard token (mean of 30 trials compared with mean of 30 baseline samples, paired t test, p < 0.05) in either the Periodic or the Random condition (Fig. 1A). In the following text, we treat each unit and its response to 30 trials of Periodic and 30 trials of the corresponding Random stimuli as a case. Out of 292 cases obtained for Noise(N)-Tone(f) stimuli from 11 mice, 205 cases showed significant responses to the first Standard in the Periodic or the Random condition and were used for further analyses. Similarly, for the Tone(f1)-Tone(f2) stimuli, of the total 283 cases obtained from seven mice, 241 cases showed significant responses (as above).

The moving average was determined by considering an average firing rate of 10 successive repetitions with a step size of one iteration (i.e., averages of repetitions 1–10, 2–11, …, 21–30, total 21 moving windows, Fig. 1A; see Materials and Methods, Long time scales of adaptation, and example plots, Fig. 1B,C). For both Periodic and Random conditions, a mean average response to Standard was obtained by considering only the Standard token's responses preceding the Deviant (Fig. 1A; see Materials and Methods). The moving average profile of rate responses of single-units (SU) and single-channel LFP response to Standard and Deviant showed different categories of response with an increase or decrease in response over repetitions (example plots: Tone, Periodic, and Random, Fig. 1B; Noise, Periodic, and Random, Fig. 1C).

Like in Figure 1B,C, Neurons show a decrease in average firing rate over repetitions to Tone or Noise as Standard and Deviant in Periodic condition, along with increase or no change in the Random condition are referred to as Irregularity preferring. Neurons could also be Regularity preferring (Fig. 1B,C) when they showed the opposite behavior over repetitions. In the Regularity preferring case, the rate response is adapted in the random condition. On the contrary, response rates increased to Tone and Noise as Standard and Deviant in the Periodic condition (Fig. 1B,C).

Layer 2/3 population responses show a preference to irregularity with adaptation depending on the component and the structure of a sound sequence

In the population of single-units and LFPs, a wide variety of long-term adaptation patterns to the sound sequences based on response to the Standard token were observed (Fig. 1). However, the normalized population moving average responses (Fig. 2) showed explicit biases. The population mean moving average response to all Standards before the Deviant position showed a strong adaptation of responses to Tone (Fig. 2A, left, green) and Noise (Fig. 2B, left, green) in the Periodic condition. In contrast, in the Random condition with an unpredictable Deviant position, there was a general increase followed by an adaptation of responses to Tone (Fig. 2A, left, blue) as well as Noise (Fig. 2B, left, blue) as Standard. We also calculated the weighted mean moving average response (see Materials and Methods, Long time scales of adaptation) to Tone (Fig. 2A, middle, green) and Noise (Fig. 2B, middle, green) to account for any possible bias because of the variable number of Standard tokens preceding the Deviant in the Random condition. The population-weighted mean moving average response showed a similar behavior as observed in the nonweighted moving average response, where adaptation of responses to Tone and Noise as Standard was observed in Periodic condition contrary to the Random condition.

Figure 2.

Population response adaptation depends on the structure of a sound sequence in Noise(N)-Tone(T) and Tone(f1)-Tone(f2) oddball paradigm. A–D, The population mean moving average response rates (A, B) and LFP response RMS value (C, D) to the N-f oddball stimuli (Materials and Methods). Moving average response to Tone (A, C) and Noise (B, D) as Standards for Periodic condition (green) and Random condition (blue) using the nonweighted (left) and weighted (middle) and as Deviant (right) for Periodic (yellow) and Random condition (red) are shown (shaded region represents standard error). Bar plots below each moving average plot show the proportion of neurons within the three slope categories (PS: Positive slope, ZS: Zero slope, NS: Negative slope) evaluated using the Standard response (left), weighted Standard response (middle), and Deviant response (right). The colored bar corresponding to “change” in each bar-plot shows the proportion of neurons in the three slope change categories (RP: Regularity preferring, NP: Non-preferring, and IP: Irregularity preferring). *p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001, and n.s. = Not significant.

The long-term adaptation also depended on the component of sound sequence considered. Mean moving average response to the Deviant (Tone and Noise) showed lesser or no adaptation compared with the Standard in the Periodic condition (Fig. 2A,B, right, yellow). Thus, there is likely a higher preference for regularity in the Deviant than in the case of Standards. However, as observed with Standards, the population mean moving average response to Tone and Noise as Deviant also showed a higher increase of firing rate over repetitions in the Random condition (Fig. 2A,B, red) compared with the Periodic condition (Fig. 2A,B, yellow). Furthermore, we analyzed RMS values from LFPs corresponding to the channels in which single-units were obtained (Fig. 2C,D, within 50-ms sound tokens) to probe for the long time scale of adaptation to Tone (Fig. 2C) and Noise (Fig. 2D) as Standard (Fig. 2C,D, left and middle) and Deviant (Fig. 2C,D, right) in the Periodic and Random conditions. We found a similar population mean moving average LFP RMS response profile to Standards and Deviant in the Periodic and Random condition as observed in single-unit activity.

We investigated the long-term adaptation by quantifying the nature of response change over repetitions. We used linear regression (slopes; see Materials and Methods, Long time scales of adaptation) to obtain positive (PS), zero (ZS), or negative (NS) slope of each single-unit and single channel LFP responses obtained in the Periodic and the Random conditions over repetitions. The same population of single-units (and LFP channels) showed marked differences in the proportion of units (or channels) in each slope category in the two conditions (single-unit, Fig. 2A,B; LFP, Fig. 2C,D, gray bars). For the single-units, we found a higher proportion of neurons showing a significant positive slope in the Random condition compared with the Periodic for both Tone and Noise as Standards with both weighted and nonweighted response (single-unit, Fig. 2A,B, left and middle gray bar plots). Similarly, we found a smaller proportion of neurons showing significant negative slopes for Standards (both weighted and nonweighted) in the Random condition compared with the Periodic condition. However, for the Deviant (single-unit, Fig. 2A, right bar plots, gray bars), the distribution slope categories for Periodic and Random were not significantly different for Noise, but for Tone, the Random condition had a higher number of neurons showing positive slope and a smaller number of neurons having a negative slope, as we found in Standard. For all the comparisons above, we performed a χ² multiproportion comparison tests followed by a z-proportionality tests (see Materials and Methods, statistical tests) to compare the slopes between the Periodic and Random conditions [single-unit, Fig. 2A,B, left bar plots; Standard response: Noise (Fig. 2B), χ² = 48.44, p < 10⁻¹¹; NS(P): p < 10⁻¹², ZS(R): p < 10⁻⁴, PS(R): p < 10⁻⁵. Tone (Fig. 2A), χ² = 16.8, p < 10⁻⁴; NS(P): p < 10⁻⁴, ZS(R): 0.400, PS(R): p < 10⁻⁴; middle bar plots, Standard weighted response: Noise (Fig. 2B), χ² = 25.63, p < 10⁻¹¹; NS(P): p < 10⁻⁷, ZS(R): 0.006, PS(R): p < 10⁻⁴; Tone (Fig. 2A), χ² = 5.47, p = 0.060; NS(P): p = 0.030, ZS(R): p = 0.360, PS(R): p = 0.020; right bar plots, Deviant response: Noise (Fig. 2B), χ² = 1.89, p = 0.380; NS(R): p = 0.130, ZS(P): 0.500, PS(P): p = 0.120. Tone (Fig. 2A), χ² = 22.26, p < 10⁻⁵; NS(P): p < 10⁻⁵, ZS(R): p = 0.400, PS(R): p < 10⁻⁵; P denotes a higher percentage in the Periodic condition, R denotes a higher percentage in the Random condition for the corresponding defined slope categories NS, ZS, and PS].

When considering LFP RMS responses, the slope distribution for both Noise and Tone presented as Standard showed similar results as in single-unit, having a higher number of positive slopes and a smaller number of negative slopes in the Random condition (Fig. 2C,D, left and middle gray bars). When presented as Deviant (Fig. 2C,D, right gray bars), neither Noise or Tone showed any difference in the slope distribution in Periodic and Random condition [LFP; Fig. 2C, left gray bars; Standard response: Noise (Fig. 2D), χ² = 21.95, p < 10⁻⁵; NS(P): p < 10⁻⁶, ZS(R): p < 10⁻⁴, PS(R): p = 0.220. Tone (Fig. 2C), χ² = 12.09, p = 0.002; NS(P): p = 0.003, ZS(R): p = 0.400, PS(R): p = 0.001; middle bar plot, Standard weighted response: Noise (Fig. 2D), χ² = 13.47, p = 0.001; NS(P): p < 10⁻⁴, ZS(R): 0.037, PS(R): p = 0.025. Tone (Fig. 2C), χ² = 7.72, p = 0.021; NS(P): p = 0.007, ZS(R): p = 0.243, PS(R): p = 0.017; right bar plot, Deviant response: Noise (Fig. 2D), χ² = 1.18, p = 0.550; NS(P): p = 0.150, ZS(R): 0.180, PS(P): p = 0.500. Tone (Fig. 2C), χ² = 1.19, p = 0.550; NS(P): p = 0.250, ZS(P): p = 0.340, PS(R): p = 0.140].

Relative change in long-term adaptation to sequence structures, regularity or irregularity preference of L2/3 neurons

We evaluated the preference for Periodic or Random stimuli by determining the change in slope signs for both the single-units and the LFPs (Fig. 2A,B and C,D, respectively, colored bars). We determined the slope change or relative preference between irregularity and regularity by calculating sign of the difference between sign of slopes obtained in the pairs of Random and Periodic conditions sign(sign(slope_Random)-sign(slope_Periodic)). Here, sign is the signum function that returns +1 for positive arguments, −1 for negative arguments and 0 otherwise. Based on the above change of slope, we obtained three categories, Regularity preferring (RP), Non-preferring (NP), and Irregularity preferring (IP). Thus, the change in the moving average profile pattern indicated each neuron's relative lack of preference or preference for regularity or irregularity over time. For single-unit data, we found the presence of all the three types of neurons described above, however, in different proportions as expected. When presented as Standard more neurons showed irregularity preference and nonpreference for Noise than Tone for the single-unit data, but for Deviant, more neurons showed regularity preference for Noise and irregularity preference for Tone. We performed the χ² multiproportion comparison tests followed by the z-proportionality tests (see Materials and Methods, Statistical tests) to compare the proportions of units with the different preferences using Tone stimuli and Noise stimuli [single-unit, Tone (Fig. 2A) vs Noise (Fig. 2B), colored bar; Standard: χ² = 4.76, p = 0.090; RP(N/T): p = 0.500, NP(T): p = 0.020, IP(N): p = 0.024; Standard weighted: χ² = 14.07, p < 10⁻³; RP(N): p = 0.231, NP(T): p < 10⁻⁴, IP(N): p < 10⁻³; Deviant: χ² = 14.41, p < 10⁻³; RP(N): p < 10⁻⁴, NP(T): p = 0.100, IP(T): p < 10⁻³; N denotes a higher percentage for Noise, T denotes a higher percentage for Tone, followed by the p values of proportionality test for the three slope change categories, RP, NP, and IP].

For the LFP data, slope change preference remained the same for both Noise and Tone when presented as Standard or Deviant [LFP, Tone (Fig. 2C) vs Noise (Fig. 2D), colored bar; Standard: χ² = 0.9, p = 0.630; RP(N): p = 0.210, NP(T): p = 0.220, IP(N/T): p = 0.500; Standard weighted: χ² = 0.41, p = 0.81; RP(N): p = 0.400, NP(T): p = 0.260, IP(N): p = 0.340; Deviant: χ² = 0.68, p = 0.700; RP(N): p = 0.330, NP(T): p = 0.200, IP(T): p = 0.340].

While differences in proportions of neurons with the different relative slopes is a measure of how the overall population behaved, it is also necessary to confirm the above results from the actual size of the slopes. Thus, we compared the population slope values from the moving average firing rates and LFP obtained in response to the Periodic and Random stimuli. For the N-f oddball stimuli, we found that in the Random condition mean population slope values were higher than in the Periodic condition both for single-unit and LFP [single-unit: one-sided paired t test; slopes from Standard response: Tone (Fig. 3A), t₍₂₀₄₎ = 3.882, p < 10⁻⁴, Noise (Fig. 3B), t₍₂₀₄₎ = 5.414, p < 10⁻⁴, left; slopes from Standard weighted response: Tone (Fig. 3A), t₍₂₀₄₎ = 2.843, p = 0.002, Noise (Fig. 3B), t₍₂₀₄₎ = 3.6684, p < 10⁻⁴, middle; slopes obtained from Deviant response: Tone (Fig. 3A), t₍₂₀₄₎ = 3.163, p < 10⁻³, Noise (Fig. 3B), t₍₂₀₄₎ = 1.668, p = 0.005, right; similarly for LFPs: slopes from Standard response: Tone (Fig. 3C), t₍₂₀₄₎ = 4.152, p < 10⁻⁴, Noise (Fig. 3D), t₍₂₀₄₎ = 5.728, p < 10⁻⁴, left; slopes from Standard weighted response: Tone (Fig. 3C), t₍₂₀₄₎ = 3.567, p < 10⁻⁴, Noise (Fig. 3D), t₍₂₀₄₎ = 5.410, p < 10⁻⁴, middle; slopes obtained from Deviant response: Tone (Fig. 3C), t₍₂₀₄₎ = 2.330, p = 0.010, Noise (Fig. 3D), t₍₂₀₄₎ = 0.663, p = 0.254, right]. The effect was absent or reduced only in the case of LFPs for Deviants.

Figure 3.

Slope comparison between Periodic and Random stimuli. A–D, Scatter plot showing slopes obtained from normalized Standard responses (left), weighted Standard response (middle), and Deviant response (right) for Tone (A) and Noise (B) in Noise-Tone (N-f) oddball stimuli form single-unit and LFP responses (C, D) data, respectively. **p < 0.01, ***p < 0.001 ****p < 0.0001, and n.s. = Not significant.

Similar long-term adaptation and regularity or irregularity preferences hold for tone(f1)-tone(f2) oddball sequences

Since oddball stimuli in the auditory system almost always have been studied with two tones (f1, f2), we also used the same to investigate whether our results with N-f oddball sequences (Figs. 1–3) hold for them. As observed with N-f stimuli, in population moving averages of single-unit rate responses to f1-f2 oddball clear biases existed (Fig. 4A,B). Rate responses increased over repetitions to the Standard (Fig. 4A, blue vs green trace) as well as the Deviant (Fig. 4A, red vs yellow trace) in the Random compared with the Periodic condition. The Deviant in the Periodic condition, like with N-f sequences, showed no adaptation (Fig. 4A, yellow, right). LFP RMS response moving average profiles, however, were different with strong adaptation after an initial increase in the Random case (Fig. 4B, blue and red traces) and mostly adaptation in the Periodic case (Fig. 4B, green and yellow traces) for both the Standard and the Deviant.

Figure 4.

Similar differential long-term adaptation patterns with Periodic and Random Tone(f1)-Tone(f2) sequences. A, B, Similar plots as in Figure 2A,C, showing the moving average responses (shaded region represents standard error) along with the gray bar plots showing slope proportions using single-unit responses (A) and LFP response RMS value (B), respectively, to Tone(f1)-Tone(f2) oddball stimulus sequence. Gray bars show the proportion of neurons within the three slope categories and the color bars show the proportion of neurons in the three slope 'change' categories as in Fig. 2A,C. C, D, Similar plots of comparison of slopes as in Figure 3A,C, for the Tone(f1)-Tone(f2) oddball stimulus sequences from single-unit (C) and LFP (D), respectively. *p < 0.05, **p < 0.01, ****p < 0.0001, and n.s. = Not significant.

The proportion of units obtained with the different slope categories in response to Standards and Deviant as f1 or f2 in Random and Periodic condition also followed a similar pattern, as observed in N-f oddball stimuli when presented as Standard (single-unit; Fig. 4A, gray bars). We found a higher number of positive slopes in the Random condition and a higher number of negative slopes in the Periodic condition [single-unit; Fig. 4A, gray bars; Standard response, left bar plots: χ² = 23.82, p < 10⁻⁶; NS(P): p < 10⁻⁷, ZS(R): p = 0.002, PS(R): p = 0.018; Standard weighted response, middle bar plots: χ² = 22.97, p < 10⁻⁷; NS(P): p < 10⁻⁴, ZS(P): p = 0.166, PS(R): p = 10⁻⁷]. In contrast, slopes obtained from Tone as Deviant were not found out to be significantly different [single-unit; Fig. 4A, gray bars, right; χ² = 0.68, p = 0.700; NS(P): p = 0.330, ZS(R): p = 0.200, PS(P): p = 0.310]. The slope distributions from LFPs were similar to that obtained from single-unit when presented as Standard [LFP; Fig. 4B, gray bars; Standard response, left bar plot: χ² = 9.8, p = 0.007; NS(P): p = 0.003, ZS(R): p = 0.303, PS(R): p = 0.010; Standard weighted response, middle bar plot: χ² = 11.14, p = 0.038; NS(P): p = 0.002, ZS(R): p = 0.370, PS(R): p = 0.004] but was different from the single-unit when presented as Deviant, having a higher number of positive slopes in the Periodic category, somewhat contrary to the general observed pattern until now [LFP; Fig. 4B, gray bars, right; χ² = 4.61, p = 0.090; NS(P): p = 0.360, ZS(R): p = 0.035, PS(P): p = 0.032].

Similarly, for the f1-f2 oddball stimuli, the mean population slope value for Random was found to be higher than Periodic for both single-units (Fig. 4C; one-sided paired t test; left scatter plot, Slopes from Standard response: t₍₄₈₁₎ = 4.366, p < 10⁻⁴; middle scatter plot, Slope from Standard weighted response: t₍₄₈₁₎ = 5.400, p < 10⁻⁴; right scatter plot, Slopes from Deviant response: t₍₄₈₁₎ = 2.200, p = 0.014) and LFP (Fig. 4D; left scatter plot, Sloped from Standard response: t₍₄₈₁₎ = 1.700, p = 0.049; middle scatter plot, Slope from Standard weighted response: t₍₄₈₁₎ = 1.17, p = 0.0121; right scatter plot, Slopes from Deviant response: t₍₄₈₁₎ = 1.565, p = 0.060).

Thus, overall, the nature of long-term adaptation to sequences with Tone(f1)-Tone(f2) oddball stimuli are similar to that observed with N-f oddball sequences, and our choice of N-f stimuli do not bias the results in any way.

Higher irregularity preference is a cortical phenomenon likely emergent in L2/3

To further evaluate whether cortical structures are necessary to incorporate such long time scales of adaptations, we additionally performed extracellular recordings in the auditory thalamus (n = 4 mice, 50 cases). The slope distribution in response to Noise as Standard was not different for Periodic and Random condition, but for the Tone Standard the Periodic condition had a higher number of positive slopes and a smaller number of negative slopes. When presented as Deviant, both Tone and Noise did not show any difference between the Periodic and Random conditions, unlike the ACX [left bar plots, Standard response: Noise (Fig. 5B), χ² = 2.16, p = 0.340; NS(P): p = 0.244, ZS(P): p = 0.207, PS(R): p = 0.071. Tone (Fig. 5A), χ² = 7.2, p = 0.027; NS(R): p = 0.013, ZS(P): 0.344, PS(P): p = 0.018; middle bar plots, Standard weighted response: Noise (Fig. 5B), χ² = 1.06, p = 0.587; NS(R): p = 0.258, ZS(P): 0.153, PS(R): p = 0.331. Tone (Fig. 5A), χ² = 6.42, p = 0.040; NS(R): p = 0.080, ZS(R): p = 0.271, PS(P): p = 0.008; right bar plots, Deviant response: Noise (Fig. 5B), χ² = 2.69, p = 0.259; NS(P): p = 0.315, ZS(R): 0.053, PS(P): p = 0.103. Tone (Fig. 5A), χ² = 1.15, p = 0.562; NS(P): p = 0.158, ZS(R): p = 0.420, PS(R): p = 0.249]. Surprisingly, when the change distribution was considered, we found a higher number of Irregularity-preferring neurons in Noise than Tone, when presented as Standard, similar to what was observed in the ACX [Tone (Fig. 2A) vs Noise (Fig. 2B), colored bar; Standard: χ² = 11.27, p = 0.004; RP(T): p = 0.110, NP(T): p = 0.026, IP(N): p < 10⁻³; Standard weighted: χ² = 9.77, p = 0.007; RP(N/T): p = 0.500, NP(T): p = 0.006, IP(N): p = 0.003; Deviant: χ² = 2.38, p = 0.304; RP(N): p = 0.061, NP(T): p = 0.205, IP(T): p = 0.262].

Figure 5.

Response adaptation in the thalamus on the structure of a sound sequence in Noise(N)-Tone(T) oddball paradigm. A, B, Similar bar plots as in Fig. 2. Proportion of neurons within the three slope categories (gray bars) and within the three slope 'change' categories (colored bars) for Tone (A), and Noise (B). * p < 0.05, and n.s. = Not significant.

When we compared these results with that of the ACX, we found that when presented as Standard in the Periodic condition both Tone and Noise show a higher number of positive slopes and a lower number of negative slopes in the thalamus as compared with the ACX [Noise: χ² = 14.94, p < 10⁻⁴; NS(A): p = 0.003, ZS(H): p = 0.085, PS(H): p = 0.001; Tone: χ² = 13.66, p = 0.001; NS(A): p < 10⁻³, ZS(H): p = 0.032, PS(H): p = 0.013; A denotes a higher percentage in the ACX, H denotes a higher percentage in the thalamus for the corresponding defined slope categories NS, ZS, and PS]. However, in the Random condition, Tone shows a higher number of positive slopes and a lower number of negative slopes as well as a higher number of positive slopes, opposite of the Periodic condition, while Noise shows no difference between the two stations [Nonweighted: Noise: χ² = 4.93, p = 0.085; NS(A): p = 0.473, ZS(A): p = 0.031, PS(H): p = 0.017; Tone: χ² = 6.65, p = 0.036; NS(H): p = 0.130, ZS(H): p = 0.124, PS(A): p = 0.005; Weighted: Noise: χ² = 2.73, p = 0.255; NS(A): p = 0.353, ZS(A): p = 0.065, PS(H): p = 0.091; Tone: χ² = 6.85, p = 0.033; NS(A): p = 0.208, ZS(H): p = 0.009, PS(A): p = 0.019]. Both Tone and Noise when presented as Deviant either in Periodic or Random did not show any differences between the two stations [Periodic: Noise: χ² = 1.94, p = 0.377; NS(A): p = 0.294, ZS(A): p = 0.203, PS(H): p = 0.082; Tone: χ² = 3.24, p = 0.198; NS(A): p = 0.046, ZS(H): p = 0.062, PS(H): p = 0.465; Random: Noise: χ² = 3.09, p = 0.210; NS(A): p = 0.040, ZS(H): p = 0.124, PS(H): p = 0.298; Tone: χ² = 2.68, p = 0.261; NS(A): p = 0.367, ZS(H): p = 0.055, PS(A): p = 0.086]. When we looked into the slope change distributions, we found that when presented as Standard both Tone and Noise register more irregularity preference and less regularity preference in the ACX as compared with the thalamus [Nonweighted:- Noise: χ² = 10.19, p = 0.006; RP(H): p < 10⁻³, NP(A): p = 0.029, IP(A): p = 0.272; Tone: χ² = 21.89, p < 10⁻⁵; RP(H): p < 10⁻⁶, NP(A): p = 0.377, IP(A): p < 10⁻³; Weighted: Noise: χ² = 6.9, p = 0.032; RP(H): p = 0.004, NP(A): p = 0.152, IP(A): p = 0.085; Tone: χ² = 12.72, p = 0.002; RP(H): p < 10⁻³, NP(A): p = 0.317, IP(A): p = 0.006]. Both Noise and Tone when presented as deviant showed no difference between the two stations [Noise: χ² = 0.45, p = 0.8; RP(A): p = 0.419, NP(H): p = 0.257, IP(A): p = 0.334; Tone: χ² = 1.8, p = 0.4; RP(H): p = 0.437, NP(H): p = 0.120, IP(A): p = 0.104].

We further used another difference metric (Slope_Random – Slope_Periodic) to quantitatively compare the results between single-unit and LFP. For both the N-f and the f1/f2 stimuli, we found this metric to be higher in case of the population single-units instead of LFPs [Tone(S): t₍₂₀₄₎ = 2.407, p = 0.008; Noise(S): t₍₂₀₄₎ = 3.544, p < 10⁻⁴; Tone(D): t₍₂₀₄₎ = 2.422, p = 0.008; Noise(D): t₍₂₀₄₎ = 1.507, p = 0.066, f1-f2(S): t₍₄₈₁₎ = 3.371, p = 10⁻⁴, p = 0.009; f1-f2(D): t₍₄₈₁₎ = 2.570, p = 0.005 where S: nonweighted normalized Standards]. These results imply that such preferential coding developing over repetitions is a cortical phenomenon which is emergent in Layer 2/3 of the ACX.

Differential response strengths of ACX neurons in random and periodic conditions

The ACX neurons are sensitive to statistical structure with higher evoked responses to less probable Tones (Ulanovsky et al., 2003; Malmierca et al., 2009; Yaron et al., 2012). In our study of adaptation to a whole sound sequence structure, we also calculated the CSI index [see Materials and Methods, Data analysis, Common selectivity index (CSI)] to probe whether differences exist between Periodic and Random conditions. However, CSI distributions in the population did not show a significant difference in Periodic and Random conditions (Kolmogorov–Smirnov test, N-f: p = 0.213; f1-f2: p = 0.505). A previous study on stimulus structure sensitivity (Yaron et al., 2012) showed higher evoked responses to Tones in a Random sequence than Periodic sequences. The observed pattern was dependent on the Deviant probability, with a stronger effect seen for Standards in Random compared with Periodic condition. We examined differences in spike rates in Random and Periodic conditions to investigate the spike response dependence at 10% Deviant probability in our experimental oddball paradigm.

Contrary to results observed by Yaron et al. (2012), we mostly found higher evoked responses to Standards and Deviant in the Periodic condition relative to the Random condition. Figure 4A,C summarizes the sound-evoked spike activity to the Noise-Tone (N-f) and Tone-Tone (f1-f2) oddball stimuli.

We calculated spike rates for Standards under two conditions. First, by considering a mean response to all the Standard tokens in a presented oddball sequence (Fig. 6A,B, left scatter plots) and secondly, by considering only the mean response to the first token (Fig. 6A,B, middle scatter plots), as robust responses were observed at first token followed by strong adaptation to successive tokens. Considering all the Standard tokens, the responses to both Standard Tones (t₍₂₀₄₎ = 3.35, p < 10⁻⁴; Fig. 6A) and Standard Noise (t₍₂₀₄₎ = 4.340, p < 10⁻⁴; Fig. 6B) in the Periodic condition were larger compared with the Random condition with 71% (146/205) and 56% (115/205) of neurons showing more response in Periodic as compared with Random condition for Tone and Noise, respectively. Similarly, considering only the first Standard token gave the same result for both Tone (t₍₂₀₄₎ = 3.120, p < 10⁻³; Fig. 6A) and Noise (t₍₂₀₄₎ = 2.890, p = 0.002; Fig. 6B) with 61% (125/205) and 47% (96/205) neurons having a higher response to the Periodic stimuli as compared with the Random stimuli for Tone and Noise, respectively. The differential response profile was observed with higher rates observed to Deviant Tones (Fig. 6A, right scatter plots) in Periodic condition (t₍₂₀₄₎ = 2.14, p = 0.016; Fig. 6A), while we saw no significant change to Noise as Deviant in Periodic or Random condition (t₍₂₀₄₎ = −0.79, p = 0.786; Fig. 6B) having 52% (107/205) and 44% (91/205) of neurons showing more response to Periodic stimuli as compared with Random stimuli for Tone and Noise, respectively.

Figure 6.

ACX L2/3 Single-unit and LFP responses to the Periodic and Random sequence. A–D, Population mean firing rate of cases to all Tone (A, C) and Noise (B, D) Standard tokens in Periodic and Random condition (left), first Standard token in Periodic and Random condition (middle), Deviant under two conditions (right) evaluated from the Noise-Tone (N-f) oddball stimuli from single-unit (A, B) and LFP (C, D) data respectively. E, F, Population mean firing rate to Tone-Tone(f1-f2) oddball stimuli similar to A, C for single-unit and LFP, respectively. The colored filled circles represent cases where a significant difference (paired t test, p < 0.05) of sound-evoked activity was observed between Random and Periodic conditions. *p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001, and n.s. = Not significant.

In the LFP recordings, the responses to both Standard Tones and Noise were mostly larger in the Periodic condition [left scatter plots, Standard Tone (Fig. 6C): t₍₂₀₄₎ = 1.653, p = 0.049; Standard Noise (Fig. 6D): t₍₂₀₄₎ = 1.862, p = 0.032], with 69% (141/205) and 56% (115/205) of neurons showing higher response to Periodic stimuli as compared with the Random stimuli. Considering only the first Standard token (Fig. 6C,D, middle scatter plots) gave the same result for both Tone (t₍₂₀₄₎ = 2.37, p = 0.009; Fig. 6C) and Noise (t₍₂₀₄₎ = 4.14, p < 10⁻⁴; Fig. 6D) with 65% (132/205) and 60% (123/205) neurons showing more response to Periodic stimuli as compared with Random stimuli for Tone and Noise, respectively. Similarly, Deviant Tone and Noise in Periodic condition [right scatter plots, Deviant Tone (Fig. 6C): t₍₂₀₄₎ = 1.658, p = 0.041; Deviant Noise (Fig. 6D): t₍₂₀₄₎ = 1.862, p < 10⁻⁴] evoked higher potential changes compared with Random condition with 53% (109/205) and 61% (124/205) neurons satisfying this condition in Tone and Noise, respectively.

Likewise, spike rates obtained using f1-f2 stimuli in Periodic and Random conditions yielded a similar response profile as observed in N-f stimuli. We found similar results for the single-unit activity for Standards taking all the Standard tokens [Fig. 6E, left scatter plot, t₍₄₈₁₎ = 4.690, p < 10⁻⁴, 64% (304/482) showing more response to Periodic condition] and only the first Standard token [Fig. 6E, middle scatter plot, t₍₄₈₁₎ = 2.560, p = 0.005, 57% (274/482)] as well as Deviant [Fig. 6E, right scatter plot, t₍₄₈₁₎ = 2.390, p = 0.008, 58% (280/482)]. Moreover, consistent results were obtained with LFP activity, showing higher evoked response to Standards [Fig. 6F, All Standard tokens, left scatter plot: t₍₄₈₁₎ = 3.37, p < 10⁻³, 57% (303/482); First Standard token, middle scatter plot: t₍₄₈₁₎ = 7.44, p < 10⁻⁴, 66% (317/482)] and Deviant [t₍₄₈₁₎ = 5.58, p < 10⁻⁴, 63% (303/482)] Tones in the Periodic condition Overall, the firing rate and input synaptic activity reflected by LFP suggests that the ACX neurons are sensitive to structures within the Random and Periodic sequences and induces higher average evoked responses to the predictable environment than to the Random context in our investigated oddball paradigm.

Random and periodic conditions differentially modulate noise correlations and functional network connectivity

The evoked Noise correlations, that is, a trial-by-trial variability between simultaneously recorded neurons (see Materials and Methods, Data analysis, Noise correlation) decrease after stimulus presentation (Oram, 2011), adaptation (Gutnisky and Dragoi, 2008), and different behavioral and attentional states (Zhang et al., 2014; Francis et al., 2018). Cortical neurons improve the efficiency of encoding stimuli by minimizing Noise correlations (Averbeck et al., 2006; Winkowski et al., 2013). Here, to consider the effect of functional recurrence in the network, we investigate the influence of stimulus structure on Noise correlations. We first computed the Noise correlations between simultaneously recorded pairs of single-units in response to the Periodic and Random stimuli. The distance between the pairs of neurons was computed based on the electrodes relative locations in the array from the respective isolated single-units. A linear fit on the Noise correlation values with distance for both these stimuli showed an expected decrease in Noise correlations as a function of distance (Fig. 7A) observed in the ACX (Rothschild et al., 2010; Winkowski et al., 2013). The mean Noise correlation decreased with distance both for the Noise (Standard)-Tone (Deviant) oddball stimuli (N-f, Periodic: p < 10⁻⁸; Random: p < 10⁻¹²; Fig. 7A, left) and Tone (Standard)-Noise (Deviant; f-N, Periodic: p < 10⁻¹⁰; Random: p = 0.017; Fig. 7A, middle). We next compared the slope between the Periodic and the Random condition, and we found that in the case of Noise (Standard)-Tone (Deviant), the drop in the value of Noise correlation with increasing distance was sharper for the Random stimuli than the Periodic stimuli (N-f, more negative slope for Random; slope comparison t test t₍₁₃₉₄₎ = 2.0567, p = 0.019; Fig. 7A, left). However, there were no differences in the case of Tone (Standard)-Noise (Deviant) stimuli (f-N, slope comparison t test t₍₁₃₉₄₎ = 1.187, p = 0.117; Fig. 7A, middle).

Figure 7.

Random and Periodic condition differentially modulates Noise correlation and network connectivity. A, The Noise correlation as a function of the distance between simultaneous recording ACX L2/3 neurons. The gray and black dot indicate the Noise correlation obtained between simultaneously recorded neurons. The solid lines represent the linear regression on the distance for both the stimulus in Periodic (black) and Random (gray) conditions in response to Noise (Standard)-Tone (Deviant), N-f; Tone (Standard)-Noise(Deviant), f-N; Tone-Tone (f1/f2-f2/f1) oddball stimuli, respectively. B, Cumulative distributions of Noise correlations for Random and Periodic conditions show higher Noise correlation in Random Deviant contexts compared with Periodic conditions. The dotted lines represent the mean Noise correlation value for Periodic (black) and Random (gray) conditions in response to the three stimuli cases, Noise (Standard)-Tone (Deviant), N-f; Tone (Standard)-Noise(Deviant), f-N; Tone-Tone (f1/f2-f2/f1) oddball stimuli, respectively. Insets, Connection probability based on Granger Causality in Periodic (black) and Random (gray) condition (z-proportionality test, **p < 0.01, ***p < 0.001, and ****p < 0.0001) for the three stimulus cases, Noise (Standard)-Tone (Deviant), N-f; Tone (Standard)-Noise(Deviant), f-N; Tone-Tone (f1/f2-f2/f1) oddball stimuli, respectively.

We next compared the Noise correlations in pairwise responses of single-units between the Random condition and the Periodic condition for both kinds of stimuli, Noise (Standard)-Tone (Deviant) and Tone (Standard)-Noise (Deviant; Fig. 7B, left and middle, respectively). We found the mean of the Noise correlations under Random conditions was higher than that under the Periodic condition for both in the Noise Standard case (paired one-sided t test t₍₆₉₈₎ = 9, p < 10⁻¹⁷; Fig. 7B, left) and Tone Standard case (paired one-sided t test t₍₆₉₈₎ = 19.54, p < 10⁻⁶⁷; Fig. 7B, right). Granger causality analysis also revealed a higher proportion of connections involved in the Random conditions compared with Periodic conditions (Fig. 7B, inset, left and middle, z-proportionality test).

Likewise, for Tone(f1/f2, S)-Tone(f1/f2, D) stimuli, mean Noise correlation for the Random stimuli was higher than that of the Periodic stimuli [paired one-sided t test t₍₁₃₉₇₎ = 31.925, p < 10⁻¹⁶⁷; Fig. 7B, right]. In a comparison of the slopes, both for the Periodic and the Random stimuli, the Noise correlation decreased with increasing distance (Periodic: p < 10⁻³; Random: p < 10⁻⁴), but there was no significant difference between the slopes (slope comparison t test t₍₂₇₉₂₎ = 0.058, p = 0.476; Fig. 7A, right). Our results indicate that the structure of stimuli modulates Noise correlations by showing higher Noise correlation in Random sequences than the Periodic stimulus. It essentially indicates that functional connectivity increases during the Randomness in structure. Thus, pairwise functional connections loosely inferred from Noise correlations suggests sparse connectivity in the network during presentations of a Periodic stimulus compared with its Random counterpart.

Two-photon Ca²⁺ imaging shows the presence of regularity and irregularity preferring excitatory as well as inhibitory neurons

The scaling up or down of the neuronal rates in response to Periodic and Random stimuli hints at a control mechanism to regulate the neural circuit's information flow. The role of inhibitory neurons has been implicated in the suppression of neuronal activity (Kato et al., 2017), gain control (Ferguson and Cardin, 2020), and correlated neural fluctuations (Okun and Lampl, 2008). Given all our observations so far, we examined the role of the EX neurons and the two most studied interneuron classes in the ACX, namely, parvalbumin (PV) and somatostatin (SOM). We performed in vivo two-photon Ca²⁺ imaging in layer 2/3 of the Gcamp-6s-injected [PV-Cre (JAX 008069), SOM-IRES-Cre (JAX 90013044), ROSA LSL-tdTomato (JAX 007908)] mouse [EX: non-PV, four mice; 11 regions of interest (ROIs) and non-SOM, six mice; 23 ROIs; Fig. 8A] from the ACX in response to Periodic and Random conditions in the Noise-Tone (N-f) and Tone-Tone (f1/f2, played only on Gcamp-Thy-1 transgenic animals) oddball stimuli. For our analyses of imaging data, for each cell, we considered the mean relative change in fluorescence with respect to the baseline (df/f) only for Standards preceding the Deviant. Because of scan rate constraints, we could not obtain the response to the Deviant embedded in the stream reliably. Since only responses to Standards are considered in the imaging results, we refer to Noise (Standard) and Tone (Standard) responses as simply “Noise” and “Tone,” respectively, for the imaging results. All neuronal populations in the GCamp-6s-injected (EX: n = 6176, PV: n = 78 and SOM: n = 102) animals, showed different effects on response adaptation over repetitions of the two types of sound sequences, as observed with single-units [Fig. 8B, Tone (above), Noise (below)].

Figure 8.

Two-photon Ca²⁺ imaging response for Periodic and Random stimuli in Excitatory and Inhibitory neurons. A, Two-photon Ca²⁺ images showing three types of interneurons, namely, Excitatory (EX), PV, and SOM. Respective neurons are marked in circles. B, Example traces of the three cells marked in A with two different types of behavior each (Regularity preferring and Irregularity preferring) are shown in a moving average fashion with a window of 10 frames and an increment of 1 frame (making a total of 21 windows). The black vertical line represents a 20% change in df/f. C, Slope distributions (gray stack bars) for the two types of stimuli (Noise and Tone) when presented in Periodic and Random fashion and the change distributions (colored stack bars) showing the proportions of Regularity preferring, Nonpreferring, and Irregularity preferring neurons for EX (left), PV (middle), SOM (right) in Gcamp-6s animals. D, Slope distribution (gray bars) and change of slope (colored bars) for two types of stimuli (Noise and Tone) in GCamp-Thy-1 animals. *p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001, and n.s. = Not significant.

As with the electrophysiology data, we computed the slopes obtained from moving average responses across iteration (without normalizing, see Materials and Methods, Imaging analysis, SSA analysis) and sorted the cells into the three slope categories (Negative (NS), Zero (ZS), and positive (PS); see Materials and Methods, Imaging analysis, SSA analysis). Within the EX population, we found that in response to the Random stimuli, higher numbers of neurons registered a negative and a positive slope as compared with the Periodic stimuli for Tone [Fig. 8C, left bar plots, gray bars, Tone, χ² = 97.36, p < 10⁻¹⁶; NS(R): p < 10⁻⁹, ZS(P): p < 10⁻¹⁶; PS(R): p < 10⁻⁸], but no such difference was found between the Periodic and Random condition for Noise [Fig. 8C, left bar plots, gray bars, Noise, χ² = 1.75, p = 0.400; NS(P): p = 0.210, ZS(P): p = 0.380; PS(R): p = 0.102]. Within the PV population, a similar pattern was found in Tone Standards where the Random stimuli generated a higher number of positive and negative slopes as compared with the Periodic stimuli [Fig. 8C, middle bar plots, gray bars, χ² = 16.11, p < 10⁻⁴; NS(R): p = 0.040, ZS(P): p < 10⁻⁵; PS(R): p = 0.002] and Noise Standards, where no change was found between Periodic condition and Random condition [χ² = 1.77, p = 0.410; NS(R): p = 0.298, ZS(P): p = 0.100; PS(R): p = 0.164]. Contrary to the behavior shown by EX and PV neurons, the SOM population had a higher number of neurons showing positive slope for Random stimuli in the case of Noise Standard [Fig. 8C, right bar plots, gray bars, χ² = 6.48, p = 0.040; NS(R): p = 0.330, ZS(P): p = 0.010; PS(R): p = 0.020] but registered no difference between Random and Periodic condition for Tone Standard [χ² = 0.6, p = 0.740; NS(P): p = 0.265, ZS(P/R): p = 0.500; PS(R): p = 0.257].

Further, we obtained distributions of the difference in sign of slopes of across iteration moving averaged response for Periodic versus Random conditions sign(sign(slope_Random)-sign(slope_Periodic)), see Materials and Methods, Imaging analysis, SSA analysis), as we did with the electrophysiology data and got the three categories, Regularity preferring (RP), Nonpreferring (NP), and Irregularity preferring (IP). Within the excitatory population, a higher number of neurons showed regularity preference, and a lower number of neurons showed nonselectivity under the Tone Standard stimuli as compared with the Noise Standard stimuli, while the number of neurons preferring irregularity roughly remained the same [Fig. 8C, left bar plots, colored bars; χ² = 38.1, p < 10⁻⁸; RP(T): p < 10⁻⁸; NP(N): p < 10⁻⁶; IP(T): p = 0.300]. In the PV population, the preference of neurons roughly remained the same [Fig. 8C, middle bar plots, colored bars; χ² = 0.286, p = 0.866; RP(N): p = 0.300, NP(T): p = 0.370, IP(T): p = 0.430] and so with SOM neurons [Fig. 8C, right bar plots, colored bars; χ² = 0.53, p = 0.760; RP(T): p = 0.380, NP(T): p = 0.320, IP(N): p = 0.230].

To rule out the possibility of the results being influenced by the Gcamp-6s longer Ca²⁺ dynamics used above, we performed additional experiments on Gcamp-Thy-1 transgenic animals (JAX 24339). Similar to the Gcamp-6s response profiles, in the N-f oddball paradigm (EX, n = 6 mice, 11 ROIs, n = 2517 cells; Fig. 8D, left), the Random condition for Tone had a higher number of positive slopes and a lower number of negative slopes [χ² = 71.4, p < 10⁻¹⁶; NS(P): p < 10⁻¹³; ZS(R): p = 0.200; PS(R): p < 10⁻¹²]. However, unlike GCamp-6s response profiles, we got a higher number of negative slopes and fewer positive slopes for Noise in the Random condition [left; χ² = 52.27, p < 10⁻¹²; NS(R): p < 10⁻¹²; ZS(P): p = 0.040; PS(P): p < 10⁻⁷]. We got the same result when we presented the f1/f2 oddball stimuli (n = 8, Gcamp Thy1 transgenic mice, 16 ROIs, n = 2120 cells; Fig. 8D, right), where the Random condition had a higher number of positive slopes and a lower number of negative slopes [χ² = 39.12, p < 10⁻⁹; NS(P): p < 10⁻⁸; ZS(P): p = 0.127; PS(R): p < 10⁻⁶]. We found consistent observation of higher positive slopes for Tone Standards in Random conditions in injected Gcamp6s and Gcamp-Thy1 transgenic animals. The population mean df/f values for all the three neurons were very slightly higher in the periodic condition both in Tone [EX: t₍₆₁₇₅₎ = 1.276; p = 0.101; PV: t₍₇₇₎ = 1.146; p = 0.128; SOM: t₍₁₀₁₎ = 0.0731; p = 0.471] and Noise [EX: t₍₆₁₇₅₎ = 1.101; p = 0.136; PV: t₍₇₇₎ = 1.967; p = 0.026; SOM: t₍₁₀₁₎ = 0.360; p = 0.360].

Differential connectivity profile of PV and SOM in response to periodic and random oddball stimuli

Next, we looked into the connectivity profiles within various subpopulations of neurons imaged. We classified each pair of simultaneously imaged neurons into two categories, either connected or not connected (see Materials and Methods, Imaging analysis, Connectivity analysis). Followed by this, we divided the connections into five different classes based on the neuron types in the pair, namely EX-EX, EX-PV, EX-SOM, PV-PV, and SOM-SOM. Since each cell could respond differently to the Noise Standard and Tone Standard stimuli, we categorized these cells based on the sign of the slope of the across iteration moving averaged response. Thus, we got nine categories (three slope types for Noise Standard multiplied by three slope types of Tone Standard), for each of the two conditions, Periodic and Random. We constructed a weighted graph consisting of nine nodes for the nine categories and the edge weights denoting the total number of connections between them. We used Gini index (Goswami et al., 2018) as a measure of sparsity (see Materials and Methods, Imaging analysis, Gini index) to quantify this connectivity between the nine different categories (a higher Gini index signifies higher sparsity in the connectivity, which implies weaker and less connectivity between neurons of different categories). We bootstrapped the connectivity graphs to generate several replicates (500) and constructed 90% confidence intervals for the Gini index. In the case of EX-EX connectivity, the sparsity of connections in the Periodic condition was significantly higher than that in the Random condition (Fig. 9A, EX-EX), implying higher functional connectivity in the Random condition. We found a similar pattern in the case of EX-SOM (Fig. 9A, EX-SOM) connections but the opposite in EX-PV connections (Fig. 9A, EX-PV), implying that PV neurons show hyperconnectivity with the EX neurons while encoding the periodic stimuli (unlike the EX and SOM neurons). The PV-PV and the SOM-SOM connections did not show significant differences between the Periodic and Random conditions within the inhibitory subpopulation (Fig. 7A, PV-PV, SOM-SOM). To see how this functional connectivity evolves across time, we computed the Gini index in moving windows of 10 successive repetitions (Fig. 9B). We see that the observed difference in Gini-indices between the Periodic and Random conditions was maintained within the EX-EX population from the initial 10 iterations to the final 10 iterations (Fig. 7B, EX-EX). Similar results were found in EX-EX connections from the data obtained in GCamp-Thy1 transgenic mice with both the Noise-Tone (N-f) and the Tone-Tone (f1-f2) oddball stimuli (Fig. 7C). However, the above was not the case with the EX-PV and EX-SOM connections, where we found that the difference was absent in the initial 10 iterations but emerged later and was evident in the last 10 trials (Fig. 7B, EX-PV, EX-SOM). The PV-PV and SOM-SOM connections did not show any difference between the Periodic and Random conditions across time. The above results indicate that the two inhibitory neuronal types change their effective influence on the network differently under the two stimulus conditions, with EX-PV connectivity becoming largely sparse under the Random condition compared with the Periodic condition and that this influence on the network emerges over time.

Figure 9.

Differential connectivity profile of PV and SOM in response to Periodic and Random oddball stimuli. A, Mean Gini coefficient calculated for all the 30 stimuli iterations within the EX-EX, EX-PV, EX-SOM, PV-PV, and SOM-SOM populations. B, Gini coefficient calculated across the iteration in a moving average window of 10 iterations, with the shaded bar representing the 90% confidence interval within the EX-EX, EX-PV, EX-SOM, PV-PV, and SOM-SOM populations. C, Moving average Gini coefficient calculated for data of GCamp Thy-1 transgenic animals for both Noise-Tone (N-f) and Tone-Tone (f1-f2) oddball stimuli. D, Network schematic showing the relative probability of connectivity w.r.t. chance between the two nodes representing the two cells with their respective slope change behavior for both Noise and Tone stimuli. The solid-colored lines represent the significant connections, with the color representing the relative probability value. The dotted lines represent connections with the nonsignificant value of relative probability, with the numbers representing their respective values on the color bar. The numbers along the arc represent the number of cells in the respective category.

The above results showed that PV and SOM neurons extend different control to the excitatory neurons. Next, we wanted to know how these neurons showing varied preferences to Periodic and Random stimuli are connected to each other. To do that, we quantified the connections between the excitatory and inhibitory neurons (EX-EX, EX-PV, and EX-SOM) of various slope change categories (RP, NS, and IR). We obtained the relative probability of occurrence of connections between pairs of neurons (type and category) with respect to chance for all the possible connection categories (six for EX-EX pairs and nine for EX-PV and EX-SOM pairs; see Materials and Methods, Imaging analysis, Relative connection probability) and if they were significantly higher (Fig. 7D, warm color edges) or lower (Fig. 7D, cold color edges) than a chance value or were at chance level (Fig. 7D, dashed edges). We found that the connections within the same slope change categories were much more probable within the EX-EX pairs than across slope change categories for both the Noise Standard (Fig. 7D, left) and Tone Standard stimuli (Fig. 7D, right). A similar pattern was found in the EX-SOM connections. In contrast, the relative connection probabilities in the EX-PV population were more spread than the EX-EX and EX-SOM, both within and across categories, with many nonsignificant connections in the Tone Standard stimuli. It is evident that significantly higher than chance connectivity was mostly present within a category (warm color edges within outer arcs representing preference categories), while across category connections were mostly significantly lower than chance (cold colored edges make outside category connections). Departure from the above was mostly because of the involvement of PV neurons. The summary of connections between slope change or regularity/irregularity preference categories (Fig. 7D) shows that separate functional subnetworks of EX-EX and EX-SOM are present that show regularity or irregularity preference. These average networks are functionally modulated by PV neurons primarily from outside category and SOM from within category over repetitions differentially under the two stimulus conditions.

Discussion

Our results show that the auditory cortical system, both at the individual neuron and local circuit levels, is susceptible to the sensory stream's statistical structures. The degree of Randomness in the Deviant positions in the two stimuli we used had far-reaching consequences, which differentially influenced the neuron's responses. Such differences may shed light on the various ways through which the brain represents structures to enhance discriminability among the various components of the auditory stimuli.

SSA corresponds to the local short-term adaptation effect on the repeated presentation of a Periodic stimulus (Ulanovsky et al., 2003; Pérez-González and Malmierca, 2014). We also found a long-scale version of adaptation resulting from a repeated presentation of a similar stream of stimulus tokens where we consider adaptation to the entire sound structure. These structural differences pertaining to the position of the Deviant token between the successive streams essentially shaped the degree and nature of adaptation occurring in the neural circuit, and the results likely indicate that previous history of the stimuli is stored at multiple timescales and is used by the neural circuit to shape its future responses.

Long-term adaptation plays an important role in memory formation in ACX

Previous studies used a continuous stream of stimuli tokens, which allowed both the short-term (because of successive tokens) and long-term (because of the probabilistic structure) synaptic depression to coexist (Ulanovsky et al., 2003; Yaron et al., 2012; Malmierca, 2014). We presented the oddball stimuli in a successive iteration fashion with sufficient gap (2–8 s) between neurons, thereby allowing them to recover from the local short-term synaptic depression and form a separate long-stream memory of the Deviant position. We seek to understand whether auditory circuitry keeps track of the previous history of structure in a stream of sounds. In an attempt to investigate changes over the long-timescale, we probed change in firing rates and synaptic input activity reflected by LFP across multiple iterations to a Periodic and Random stream of sounds in L2/3 neurons of the ACX (Fig. 2). We found higher positive slopes for responses to Standards and Deviant Tones obtained over iterations presented in the Random condition compared with the Periodic condition. Such results imply that the degree of adaptation for successive repetition of the complex Randomly-patterned stream of sounds was much lower as compared with the successive repetition of the Fixed-patterned Periodic stimulus. Furthermore, Standard responses were more affected than the Deviant responses, which reflects context-specific modulation. Taken together, these changes reflect a long-term memory mechanism, where ACX neurons continually track and compare the incoming sensory patterns to past history and update the future response accordingly.

Adaptation in long scales is different from adaptation in short scale

The general adaptation in neural responses suggests that the Periodic sequence stream responses should decrease in the successive repetitions. Conversely, if the stimuli structure changes in each iteration, such an adaptation should not occur. Fundamentally speaking, the present understanding of the adaptation phenomenon supports the existence of the Irregularity-preferring neurons (higher slope in Random, lower slope in Periodic condition). The primary result obtained with single-unit and LFP recordings is that long time scale adaptation to an entire structure of sound sequence can show decreases or increases in response to the component stimuli over repetitions. We got a considerable Regularity-preferring neurons (blue, higher slope in Periodic, lower slope in Random, colored bars, Fig. 2A,B; Noise and Tone Standard: 24.4%; Noise weighted Standard: 22.9%; Tone weighted Standard: 23.9%; Noise Deviant: 30.24%; Tone Deviant: 28.78%; Tone-Tone Standard: 23.65%; Tone-Tone weighted Standard: 23.03% and Tone-Tone Deviant: 28.84%). Such observations indicate that the understanding of adaptation cannot be easily extrapolated to stimuli consisting of a repeated sequence of sounds, and it warrants further investigation.

Higher slopes but lower rates in the random condition suggests a differential control

We further found rates for the Standards and Deviant in the Periodic condition to be higher compared with a Random condition (Fig. 6). In order to accommodate a rise in the response rates over repetitions for the Random condition, a lowering in the overall rates compared with the Periodic condition is achieved. It may be observed in the population average normalized rate response moving average profiles (Fig. 2) that the mean responses are the same in both of the conditions in the starting few moving averages, after which the responses in the two conditions diverge. Such a mechanism is likely achieved through the network with recurrence (Yarden and Nelken, 2017) or through inhibitory neurons adaptation (Natan et al., 2015) or gain control (Seybold et al., 2015), or both. In our study, the rate-specific results observed under two conditions were contrary to the Yaron et al., 2012 study, which could be because of the long-stream stimulus presented in that study in as compared with oddball stimuli used in our study with a sufficient gap between each iteration. However, our results are in accordance with higher evoked potential for the outliers in regularly-patterned (REG) stimulus compared with Random (RAND) Tone pip sequences observed in a study by Southwell and Chait, 2018. Our pieces of evidence further support results obtained in the musical expectation studies (Vaz Pato et al., 2002; Pearce et al., 2010), where they observed a higher Deviant response within the structural context compared with Random context. Since experiments were performed under lightly anaesthetized conditions, it is unlikely to consider the involvement of attention (Zhao et al., 2013) in such experimental paradigms. Thus, our results for long time scale adaptation changes and differential rates to Periodic and Random conditions suggest long-term structure learning in ACX neurons based on experience, as a new structure emerges in every iteration in the Random condition.

Functional connectivity dictates the dynamics of responses

We hypothesized that the local functional recurrent connectivity could underlie the different slope categories or slope change categories – regularity, irregularity preferring or nonpreferring. True to that hypothesis, we observed a higher degree of functional connectivity in the Random condition as compared with the periodic condition (Fig. 7B). Although Noise correlation only gives us a relative measure of the connection strength, it can be conducive to substantiation of our results comparing the effect of stimuli complexity on the interconnected local circuits (Ginzburg and Sompolinsky, 1994; Averbeck et al., 2006). We found a higher value of mean Noise correlation in the Random Deviant stimuli compared with a Periodic sequence. Although expected, it gives us a hint that complex stimuli structures recruit a larger number of neurons within different levels of hierarchy to create a “memory” and optimize their processing to show the respective adaptations. This result holds for both the Noise-Tone and Tone-Tone stimuli, essentially implying that the nature of the stimulus has little to do with the density of active circuit-level recruitment of neurons for the processing of complex stimuli. The Inhibitory neurons are known to decorrelate network activity, enhance or restrict network plasticity (Tetzlaff et al., 2012; King et al., 2013). Thus, we hypothesize inhibitory interneuron's role in understanding the mechanism underlying differential adaptive coding of sound sequences. It remains to be seen how generalized such functional connectivity-based response dynamics can be in other systems. While adaptation is known to play a role in all sensory systems, studies on such long temporal scales are missing. Regularity and irregularity is also a common determining feature in terms of stimulus parameters in visual and somatosensory systems as well but such studies on adaptation either on long temporal scales or on large spatial scales have not been done. It would be interesting to see whether such motifs of sub-networks of EX-EX, EX-INNs of different categories emerge in other sensory cortices. Thus, our results can provide a basis for general principles of adaptation on multiple time scales in cortical networks.

PV and SOM practice differential control over the excitatory responses

The local Inhibitory interaction controls the gain modulation. Previous studies have shown divisive and subtractive modulations by optogenetically manipulating PV and SOM neurons, respectively (Seybold et al., 2015). Furthermore, the role of PV and SOM neurons during context-dependent gain modulation, such as in adaptation, has been actively studied (Natan et al., 2015, 2017). Such studies still focus only on the local, short-term adaptation effects of PV and SOM on adapted and nonadapted neuron responses. We present an investigation focused on the effects of these inhibitory subtypes on the long-timescale adaptation and how statistical patterns within the stimuli affect their inhibitory control over the responses. We found differences between PV and SOM neurons in their adaptation control properties. These differences could arise because of the differential connectivity shown by PV and SOM neurons across columns and their bandwidths (Kato et al., 2017). We also found that excitatory-inhibitory connection takes longer than their excitatory-excitatory counterparts to exert their inhibitory control on the responses (Fig. 9A). Our study observed that differential Inhibitory influence emerges over time, where EX-PV connections behave differentially to Periodic and Random sequences compared with EX-EX and EX-SOM connections, showing denser functional connectivity for the Periodic stimuli, unlike the EX-EX and EX-SOM connectivity, which had denser functional connectivity for the Random stimuli (Fig. 9A). In other words, as we know that inhibition is always preceded by excitation, owing to the recurrent and columnar architecture of the cortex; similarly, the adaptation properties of these inhibitory controls take place after the excitatory rebalancing of the responses. This could point out to a possible excitatory-inhibitory rebalancing act, for which the inhibitory neurons take longer than the excitatory ones. Our results suggest that EX-PV and EX-SOM connectivity has a strong role in long scale context-specific modulation, as seen by a change in connectivity pattern emerging in the later iterations. Within the EX-PV connections, the functional connectivity was much more homogenous and nonspecific as compared with the EX-SOM and the EX-EX connections. Altogether, these results paint a vivid picture of the connection properties of the neural circuit. The findings hint that the encoding of a Randomly patterned stimuli sequence require a more specific and directed connectivity profile with EX-EX and EX-SOM connections carrying most of the computational load as compared with the encoding of Regularly patterned sound sequences, which requires relatively less specific and homogeneous connectivity and mostly rely on EX-PV connections. Such modifications throw a light on the differences in the modulating effects that are undertaken by the PV and SOM neurons.

Limitations

Our results clearly indicate different sub-networks of EX-EX and EX-INN neurons of different types to participate in different types of encoding of regularity and irregularity. However, we tested it with Tone-Tone and Noise-Tone stimuli, and it remains to be seen how well it generalizes to other types of stimuli like vocalization sequences or other complex stimulus structures, with spectral and/or temporal modulations. Our conclusions of connectivity analyses are based on significant noise correlations, which is a weak measure of connectivity. The results would be stronger with in vivo or post hoc ex vivo simultaneous whole cell recordings (Hofer et al., 2011).

Similarly, in vivo two-photon single-cell optical stimulation coupled with imaging by expressing C1V1 is another way to further strengthen the results (Packer et al., 2015). It would be very useful to be able to develop a model of the connectivity motif we obtain and study the long-term adaptation properties of such networks. However, we have insufficient data on the synapse properties of the recurrence in L2/3 to be able to completely model the behavior we observed. Such a computational network model would open up possibilities of testing and predicting behavior of small networks of neurons with different inhibitory neuronal types.