Abstract
Neurons in the mouse auditory cortex are strongly influenced by behavior, including both suppression and enhancement of sound-evoked responses during movement. The mouse auditory cortex comprises multiple fields with different roles in sound processing and distinct connectivity to movement-related centers of the brain. Here, we asked whether movement-related modulation in male mice might differ across auditory cortical fields, thereby contributing to the heterogeneity of movement-related modulation at the single-cell level. We used wide-field calcium imaging to identify distinct cortical fields and cellular-resolution two-photon calcium imaging to visualize the activity of layer 2/3 excitatory neurons within each field. We measured each neuron's responses to three sound categories (pure tones, chirps, and amplitude-modulated white noise) as mice rested and ran on a non-motorized treadmill. We found that individual neurons in each cortical field typically respond to just one sound category. Some neurons are only active during rest and others during locomotion, and those that are responsive across conditions retain their sound-category tuning. The effects of locomotion on sound-evoked responses vary at the single-cell level, with both suppression and enhancement of neural responses, and the net modulatory effect of locomotion is largely conserved across cortical fields. Movement-related modulation in auditory cortex also reflects more complex behavioral patterns, including instantaneous running speed and nonlocomotor movements such as grooming and postural adjustments, with similar patterns seen across all auditory cortical fields. Our findings underscore the complexity of movement-related modulation throughout the mouse auditory cortex and indicate that movement-related modulation is a widespread phenomenon.
Significance Statement
Throughout the sensory cortex, neural activity is influenced by behavior. It remains unknown whether primary and higher-order sensory cortical centers are similarly or differentially influenced by movement. We show that movement-related modulation in the mouse auditory cortex is locally complex and heterogeneous, but that at a more macroscopic level, the net effect of movement on primary and higher-order auditory cortex is largely conserved. These data highlight the widespread nature of movement-related modulation and suggest that movement signals may inform neural computations throughout multiple nodes of the sensory cortex.
Introduction
Neural activity in sensory cortices is influenced by movement (Niell and Stryker, 2010; Schneider et al., 2014; Stringer et al., 2019; Schneider, 2020). In the mouse auditory cortex, movement-related modulation manifests as changes in the spontaneous and sound-evoked responses of neurons during behaviors including locomotion, forelimb movements, and vocalizing (Schneider et al., 2014; Zhou et al., 2014; Rummell et al., 2016; Bigelow et al., 2019; Henschke et al., 2021; Audette et al., 2022; Vivaldo et al., 2022). Movement-related modulation of individual neurons is diverse within and across experimental paradigms, often leading to the generic suppression of spontaneous and sound-evoked activity (Schneider et al., 2014; Zhou et al., 2014; Bigelow et al., 2019), but also enhancement of responses (Henschke et al., 2021; Vivaldo et al., 2022) and acoustically selective modulation that arises with motor-sensory experience (Rummell et al., 2016; Schneider et al., 2018; Audette et al., 2022).
This diversity of movement-related modulation may reflect broad heterogeneity of neural responses throughout the auditory cortex. Alternatively, it could arise from differences in where within the auditory cortex neurons reside. The mouse auditory cortex contains multiple distinct areas that surround the primary auditory cortex (A1), including the anterior auditory field (AAF), the secondary auditory cortex (A2), and a dorsal posterior field (DP), among others. Although the specific functions of each auditory cortical field remain unresolved, different fields process sounds in distinct ways. For example, while A1 neurons often respond to individual tone frequencies (Mizrahi et al., 2014), A2 neurons are significantly more responsive to multifrequency sounds (Romero et al., 2020; Kline et al., 2023) and AAF receptive fields are biased toward faster temporal structure (Linden et al., 2003).
In addition to their different roles in sound processing, different auditory cortical fields also have distinct connectivity to motor centers (Tsukano et al., 2017, 2019; Gămănuţ et al., 2018; Henschke et al., 2021). One important modulator of auditory cortex during locomotion is a long-range projection from the secondary motor cortex, which synapses locally in auditory cortex onto both inhibitory and excitatory cells (Nelson et al., 2013; Schneider et al., 2014, 2018). Projections from the secondary motor cortex innervate much of the auditory cortex, but its connectivity to the dorsal auditory cortex is denser than in A1 (Henschke et al., 2021).
These differences in function and connectivity suggest that different cortical fields might be differentially modulated by movement. However, most previous studies that report movement-related modulation have specifically targeted the A1 (Audette et al., 2022); have not reported the auditory cortical field from which data were collected (Schneider et al., 2014); or solely used anatomical coordinates to identify cortical fields (Henschke et al., 2021), which is better accomplished through in vivo functional mapping (Romero et al., 2020; Narayanan et al., 2023). In addition, prior studies have at most investigated only two auditory cortical fields at a time. It therefore remains largely unknown whether the heterogeneity of movement-related modulation in mouse auditory cortex may be related to the specific cortical field from which neural activity was recorded.
Here, we used in vivo wide-field imaging to functionally map four distinct auditory cortical fields. We then quantified the movement-related modulation of single neurons in identified cortical fields using two-photon calcium imaging. Our findings reveal heterogeneity in the magnitude and direction of movement-related modulation across individual auditory cortical cells. Despite some subtle and significant differences among cortical fields, the average modulation driven by locomotion and other spontaneous movements was largely consistent across all areas that we sampled. We conclude that movement-related modulation is pervasive throughout the auditory cortex and that despite different roles in sound processing and differing connectivity, both primary and nonprimary auditory cortex exhibit macroscopically similar modulations by behavioral state.
Materials and Methods
Animals
Experiments were performed on 12 Camk2-Cre+/+ Ai148D+/+ [B6.Cg-Tg(Camk2a-cre)T29-1Stl/J;B6.Cg-Igs7tm148.1(tetO-GCaMP6f,CAG-tTA2)Hze/J, Jackson Laboratory, http://www.jax.org, RRID IMSR _JAX: 005359 and 030328] male mice (Table 1). Experimental animals were obtained by breeding homozygous colonies. Breeding cages were fed a doxycycline food diet for 3 weeks (625 mg/kg, ENVIGO, Teklad, TD.05125). This diet halts the Cre recombinase expression until 21 d of age, avoiding epileptic-like seizure activity in adults. Offsprings’ ear biopsies were genotyped by Transnetyx for Camk2-Cre and Ai148D transgenes (Transnetyx probes for reference lgs7-1WT, eGFP, CRE). Positive offsprings were kept in grouped cages with regular food and water available ad libitum and kept on a reverse day–night cycle (12 h day, 12 h night). All procedures were carried out in accordance with New York University's Animal Use and Welfare Committee. One mouse was excluded from the analysis because no sound-responsive cells were detected.
Neurons imaged in experimental mice
Preparation of cranial windows
Glass coverslips were cleaned and stored in 70% ethanol distilled water solution. Dry 4 mm and a pair of 3 mm coverslips (#1 thickness, Warner Instruments) were glued using a transparent, UV-cured adhesive (Norland Optical Adhesive 61). On surgery day, cranial windows were rinsed with sterile saline (Argyle Sterile Saline 0.9% or Addipak Sterile Saline).
Surgical procedure
Animals were anesthetized with isoflurane in oxygen (3% induction; 1.5%–2% maintenance) and placed in a stereotaxic holder with nonrupture, zygoma ear cups (Kopf Instruments, models 963 and 1721) with a heating pad to maintain and monitor body temperature (Harvard Apparatus). The scalp was disinfected with 70% ethanol and betadine. Mice were injected subcutaneously with an analgesic (0.1 mg/kg, Meloxicam ER, ZooPharm) and 150 µl local anesthetic under the scalp (0.25% bupivacaine hydrochloride saline solution, Sigma-Aldrich B5274-1G). Eyes were covered with Vaseline. After removing the scalp, the skull was dried and polished to remove the periosteum allowing better adhesion of the headpost. The dorsal part of the skull was covered with tissue adhesive (Vetbond, 3 M), and a Y-shaped headpost (1.5 mm thickness, 120° angle, 3 × 17 mm branches, sendcutsend.com) was aligned to the midline and glued. The headpost was cemented to the dorsal skull using Metabond dental cement (C & B, Metabond), and the area over the skull was covered with a thin layer of transparent dental acrylic (Lang Dental). Once the dental cement was cured, the ear cups were removed to allow access to the auditory cortex (∼2 mm diameter, −2.5 mm posterior, 4.2 mm left from bregma), and the headbar was clamped to stabilize the mouse head. Temporalis muscle insertion on the temporal ridge was resected and glued down. The parietal bone was dried and cleaned, and a 3 mm circular craniotomy was drilled using a dental drill (Foredom, model 1474) and micro drill bit (0.10 mm head diameter, GRAINGER, Kyocera, 07089). The posterior edge of the craniotomy borders the lambdoid suture, while the ventral edge is positioned below the squamosal suture located above the posterior end zygomatic arch. The coverslip was placed into the craniotomy with the help of a 3D manipulator-held wooden stick and hermetically sealed with Vetbond and dental cement applied between the surrounding skull and glass. The imaging window was positioned at an angle of 59.2 ± 0.4 degrees, measured in relation to the vertical axis.
Behavioral setup/apparatus
Mice were trained for 7–10 d to run head-fixed on a 3D-printed wheel before imaging experiments. Head fixation was achieved using a plate clamp (standa.it, 4PC69). The wheel was designed using SketchUp Free (https://app.sketchup.com/app; 40 mm diameter, 60 mm rod length, 1 mm rod diameters, 3 mm rod spacing) based on the previously published design (Villette et al., 2017) and affixed to a rotary encoder (US Digital, H5-1000-NE-S). Sounds were played through an electrostatic ultrasonic speaker, amplified (Tucker-Davis Technologies, ES1, and ED1), and controlled by a sound card (RME, FireFace UCX). The speaker was placed at ∼10 cm from the right ear of the mouse (contralateral ear). Videos were acquired using IR zoom lenses (Xenocam, 9–22 mm, 1.4 f) mounted on Sony CCD Camera (Amazon, AKK CA20 600TVL), and IR illumination was provided by infrared light sources (Phenas, Home 48-led CCTV IR lamp). Sound stimulation, video recording, and wheel position were controlled or recorded using custom-written code (Matlab). Sound stimuli were played using Matlab Psychophysics Toolbox Version 3 (PTB-3) sound library (PsychPortAudio; Brainard, 1997).
Auditory stimulation protocols for mapping and cellular imaging
All sounds were generated at 192 kHz using Matlab scripts.
Epifluorescence mapping auditory stimulations
Tonotopic mapping was achieved by presenting 200 ms pure tones (4, 8, 16, and 32 kHz at 80 dB) separated by 3 s and played pseudorandomly such that each frequency was played 15–20 times. Mice were awake and head-fixed and free to move on the wheel. Stimuli were calibrated before recording using an ultrasonic acoustic microphone (Avisoft Bioacoustics, CM16/CMP).
Cellular imaging auditory stimulations
A total of 38 sounds were played while head-fixed mice were free to run on the wheel. Sounds included 3 sound types: 7 pure tones (PT; 4–32 kHz spaced by 0.5 octaves), 9 amplitude-modulated white noise (AMWN) sounds (modulation frequencies: 2, 4, 8, 16, 32, 64, 128, 256, 512 Hz), and 2 chirps (4–32 kHz and 32–4 kHz). PT and AMWN were played at 40 and 60 dB. All sounds were 1 s long with a rising and falling ramp of 5 ms. All sounds were presented pseudorandomly such that each sound was played 15 times in rest and running conditions.
Calcium imaging acquisition
Wide-field epifluorescence image acquisition for auditory cortex mapping
Wide-field epifluorescence images were acquired with a 4× objective (Olympus, NA 0.10 Plan Achromat) focused 400 µm below the cortical surface. A blue light-emitting diode (Thorlabs, M470L3, 470 nm, 650 mW) excited a ∼3 mm diameter area of cortex through a filter cube (Thorlabs, DFM1T1) containing an excitation filter (Thorlabs, MF469-35), a dichroic mirror (Thorlabs, MD498), and an emission filter (Thorlabs, MF525-39). Green fluorescence was captured at 30 Hz with a 16 bit camera (sCMOS, pco.edge 3.1). The sample was illuminated only during the acquisition to reduce bleaching (trial length 2 s). Images with 384 × 512 pixel were acquired using a Matlab custom GUI and Matlab image acquisition toolbox. A blood vessel map picture was taken at the pial surface as a reference image for targeting two-photon recordings.
Cellular imaging using two-photon microscopy
We used a resonant scanning two-photon microscope (Neurolabware) focused 150–200 µm below the pial surface to image GCaMP fluorescence changes during behavior. Excitation was provided by a femtosecond pulsed laser (Spectra-Physics, InSight X3) tuned to 940 nm. Illumination power was controlled by Pockels cells (Conoptics, 350-80-02) and the beam size by a beam expander. The beam then passed through 8 kHz Galvo-Resonant Scanner (Cambridge Technology, 6215H galvo scanner and CRS 8 K resonant scanner) and 16×/0.8NA water-immersion objective (Nikon, MRP07220) to form a 512 × 796 pixel field of view (∼1 mm2). The objective was rotated between 45° and 60° off the vertical axis until perpendicular to the cranial window to obtain images of the auditory cortex while maintaining the mouse head position straight. Two-plane imaging was achieved using an electrically tunable lens (Optotune). The frame rate was 14.49 or 7.5 Hz for single- and two-plane imaging, respectively. Emission photons were filtered by photon collection optics (Semrock dichroics, 750 SWP, 562 nm LWP; emission filter, 510/84 bandpass) and were detected by a GaAsP photomultiplier tube (Hamamatsu, H11706-40 MOD2). Images were written to disk at 1×, 1.2×, or 1.4× digital magnification using an acquisition software (Scanbox software, Neurolabware).
Image and data processing
Wide-field epifluorescence for auditory cortex mapping
To localize the peak fluorescence responses to 4 and 32 kHz pure tones on the cortex below the cranial window, we processed images using Matlab scripts. First, images were cropped only to include the 3 mm diameter cranial window view and resized 300 × 300 pixels to reduce processing time. Then, movies were averaged across 15–20 trials, baselined, and Z-scored by the standard deviation for each pixel of the baseline period across all stimuli (0.2 s before stimulus onset). Frames with movement artifacts were rare and were manually discarded. The response image for each stimulus was defined as the average across frames occurring during the response window (0.1–0.3 s after stimulus onset). Fluorescence response peaks were detected on a smoothed response image (Gaussian filter radius = 100 µm) using local maxima bigger than 25% of the tallest peak.
Auditory cortex areas definition
To identify auditory cortex fields (A1, AAF, A2, DP), we used fluorescence responses to 4 and 32 kHz pure tones as landmarks (Fig. 1B). A1 was defined as a rectangle of 500 µm width spanning in length from the most posterior 4 kHz response to the most medial anterior 32 kHz response. AAF was defined as a rectangle of 500 µm width spanning from the most anterior 4 kHz responses to the most medial anterior 32 kHz response. DP was defined as a rectangle medial to A1 with a width of 600 µm. A2 was defined as the triangular area between A1 and AAF.
Mapping AC areas and cellular imaging of AC neurons during rest and locomotion across sound types. A, Schematic of the experimental setup used to map auditory cortex (AC) areas and conduct cellular imaging of AC neurons during rest and locomotion across different sound types. The setup involves a head-fixed mouse running on a treadmill while sounds, such as pure tones, AMWN, and chirps, are delivered through an ultrasonic speaker. High-speed videos are recorded using a camera, and speed is measured through a rotary encoder attached to the treadmill. Auditory cortex images are acquired either through a 4× objective for epifluorescence AC mapping or a 16× objective for two-photon imaging through a 3 mm glass window. B, Auditory cortex mapping using epifluorescence. Left, Example epifluorescence average sound responses to 4 and 32 kHz pure tones. White crosses show peaks. Right, Peak 4 and 32 kHz responses are represented by blue and red circles, respectively. Right column, top, Example auditory cortex map for one animal used to guide 2-photon imaging. Example field of view is delineated in white and shown in panel D. Right column, bottom, Average map across all animals. C, Preferred pure tone frequency map of neurons in the AC across 11 animals. Each dot on the map represents a neuron, with its color indicating its best frequency as defined in the accompanying color bar. The AC was divided into 200 μm tiles, and the color of each tile reflects the average best frequency of the neurons within it. Gray lines delineate the boundaries of the auditory cortex, estimated based on mean epifluorescence peak responses to 4 kHz (blue circles) and 32 kHz (red circles). The map is aligned to the 4 kHz responses in the A1 and AAF, and it also highlights the tonotopic gradients of A1 and AAF with the assistance of black arrows. D, Example two-photon field of view enhanced image highlighting sound-responsive neurons. Each neuron's preference for a specific sound type is color coded as indicated in the legend. The outlined field of view, delineated in panel B, is situated at the convergence point of A1, DP, and AAF's 32 kHz region. E, Example neuronal calcium and speed signals of three neurons. Normalized calcium fluorescence traces are depicted in black, while colored rectangles denote sound type, time, and duration. An asterisk (*) designates the preferred sound stimulus. Green line represents speed in cm/s. The scale bar provides reference for the speed (y-axis in purple), Z-scored fluorescence (y-axis in black), and the time (x-axis). F, Peristimulus time histogram (PSTH) show the example neurons’ individual preferred average and standard error sound responses during rest (black line) and running (green line) conditions. Sound timing is represented by a gray rectangle. The scale bar provides reference for the Z-scored fluorescence (y-axis) and the time (x-axis).
Two-photon microscopy
Two-photon calcium signals from auditory cortical neurons were extracted from imaging recordings using the toolbox Suite2p (Pachitariu et al., 2017). Briefly, suite2p corrected movement artifacts between frames with a nonrigid registration. Then, regions of interest (ROIs) and neuropil areas (an annular ring surrounding the ROI) were identified. Then, calcium signals were extracted from soma and neuropil by spatially averaging pixel intensities across those regions for each ROI. Finally, ROIs were identified as somas during a manual curation step using the Suite2p user interface. ROIs were selected based on soma-like morphological features visible on the processed version of the fluorescence image (enhanced image). Each neuropil signal was then subtracted from the raw calcium traces to limit neuropil contamination. Neuronal signals were then baselined by subtracting the mode over a 60 s sliding window. Baselined traces were then normalized by the estimated noise for each cell. The noise distribution for each cell was defined as the distribution of negative values from the baselined traces and the same negative values multiplied by −1. The noise estimate for each cell was the standard deviation of that “symmetrized” noise distribution. For neuronal response computations, peristimulus histograms for each stimulus were baselined by the average of the 1 s prestimulus baseline period. The neuronal response of a neuron is reported as the peak value over a response window (0–1.5 s after stimulus onset) in units of standard deviations of the noise distribution of that neuron, referred to as “Normalized fluorescence” in the manuscript.
Sound-responsive neurons selection
Sound-responsive cells were defined based on two criteria: (1) Significant statistical difference (Wilcoxon signed-rank test) between trial values from the baseline period (1 s prestimulus) and at least one response window (0–0.5, 0.5–1, 1–1.5 s after stimulus onset) in the rest or running condition. Using three response windows allowed to include onset and offset responses. Response magnitude was calculated as the difference between the baseline period and the best response window. A minimum of five trials was required for a cell to be considered; those with fewer trials were excluded. (2) Average response size exceeding 3 standard deviations from the neuron's noise distribution (refer to Image and data processing and Two-photon microscopy for details). For analyses comparing rest and running conditions, we only included neurons exposed to a consistent set of stimuli (4, 8, 16, 32 kHz pure tones, 9 AMWN, and 2 chirps), resulting in a sample of 1,869 neurons.
Data analysis
Alignment of auditory cortex maps across mice
To align the epifluorescence imaging maps of the auditory cortex across mice, we employed a two-step process that involved translation and rotation. First, we aligned all the maps to the 4 kHz frequency A1 epifluorescence sound response peaks. Next, we rotated individual maps such that the line joining the A1 and AAF 4 kHz epifluorescence response peaks would be at 70° relative to the abscissa.
Rest and running conditions
The conditions of rest and running were categorized by assessing the running wheel's speed within a specific time frame, which began 1 s before the onset of sound presentation and continued until the sound presentation concluded. A rest state was identified when the wheel's average speed was <1 cm/s during that window, whereas a state of running was indicated by a speed exceeding 20 cm/s.
Modulation index computation
We calculated the modulation index (MI) as MI = (Rrun − Rrest) / (Rrun + Rrest), where R represents the preferred sound responses during rest (Rrest) or run (Rrun).
Sound responses speed modulation analysis
Effect size
In Figure 4B, we compared how different recording parameters affect the MI. To do this in a standardized way, we used a statistical measure called effect size. Since our data had more than two parameters, we used a type of effect size called omega squared that estimates the variance due to an effect and dividing it by the estimated total variance, using the following formula: T ω2 = (SSbetween − (dfbetween * (SSwithin / dfwithin))) / (SStotal + (SSwithin / dfwithin))), where SS stands for sum of squares and df stands for degrees of freedom.
Decoding running speed from sound-evoked responses with support vector machines (SVMs)
To estimate how much running speed information is carried by the sound-evoked neural responses of each recording, we employed the following decoding strategy. For each recording session, we included the sound responses elicited by all sounds, regardless of the preferred sound response of each neuron. This approach resulted in a matrix of dimensions “number of neurons” by “number of trials” and a vector containing the trial running speeds, computed as the average running speed during the sound presentation. Running speeds were binned into intervals of 2 cm/s, ranging from 0 to 40 cm/s.
We trained a support vector machine with a linear kernel to predict running speed based on the sound responses. As some speeds were more represented than others, we chose the one-versus-one training strategy which better handles imbalanced datasets. To evaluate the performance of our decoding model, we employed leave-one-out cross-validation. Specifically, for each trial, we trained the SVM on the remaining trials and tested its prediction accuracy on the left-out trial. The accuracy of prediction was computed as the fraction of correct speed bin predictions when compared with the true speed bins. The regularization parameter C was 0.022.
To evaluate our results’ significance, we conducted a permutation test by reshuffling our data 100 times. In each reshuffle, we randomly altered the relationship between sound responses and running speeds, followed by training the SVM using the leave-one-out cross-validation approach as previously described, and assessed the prediction accuracy on the excluded trial.
Subsequently, we established the chance level accuracy by taking the average from the distribution of prediction accuracies across all reshuffles. We then contrasted the prediction accuracy derived from the original data with this distribution of chance level accuracies using a t test, keeping the significance level threshold at 0.05. This procedure was replicated for all recorded sessions, and the resultant session accuracies were illustrated in Figure 4A. The analysis was performed in Matlab (MathWorks) using the function fitecoc from the Statistics and Machine Learning Toolbox.
Modeling the relationship between running speed and sound responses
To analyze the modulation of sound responses by speed, we applied different models to the data depicting the relationship between speed and sound responses for each neuron. We chose four fitting functions based on observations of our data and previously published relationships between cortical sensory responses and running speed (Saleem et al., 2013; Christensen and Pillow, 2022): linear, step-function, exponential, and logarithmic (see Fig. 5C for examples). Following fitting the observed values (y), we obtained the predicted values from the model (yfit) and proceeded to assess the quality of the fit using two metrics: the coefficient of determination (R2) and the p value (Fig. 4D).
To calculate R2, we first computed the residuals by subtracting the predicted values from the observed values (yresid = y − yfit). Then, we determined the sum of squared residuals (SSresid) by summing the squared values of the residuals. Additionally, we calculated the total sum of squares (SStotal), which represents the variability of the observed values around their mean. SStotal was obtained by multiplying the variance of the observed values by the length of the data minus 1. The coefficient of determination (R2) was then calculated as 1 minus the ratio of SSresid to SStotal (R2 = 1 − SSresid / SStotal). R2 ranges from 0 to 1, where higher values indicate a better fit of the model to the data. To further assess the statistical significance of the fit, we calculated the p value. The p value was obtained using the F distribution cumulative distribution function (CDF) with the F statistic, which was computed as (R2 / (1 − R2)) multiplied by the ratio of the degrees of freedom (df) to 1. The degrees of freedom (df) were determined as the length of the data minus 2. Finally, the p value was calculated as 1 minus the CDF of the F distribution, using the F statistic, 1 as the numerator degrees of freedom, and df as the denominator degrees of freedom.
Finally, for each neuron, the best fit was determined as the fit with the biggest coefficient of determination (R2) and a significance level of the fit below 0.05 (Fig. 4E).
Video processing and behavioral state classification
Videos were recorded using custom Matlab scripts at 10 Hz and compressed as .mp4 files using custom Matlab user interface (https://github.com/lachioma/FFMpeg-compress-videos) operating with the FFMpeg algorithm (Mathis and Warren, 2018).
To visualize motion, we transformed the videos into motion frames by computing the difference between consecutive frames. We generated an RGB image where the red and blue channels contained the values of frame n, while the green channel contained the values of frame n + 1. This approach highlighted pixels with movement by showing them as pink or green (pink representing the pixels with the highest values in frame n and green representing the corresponding pixels in frame n + 1).
We manually categorized the motion frames into six different categories, which were resting, adjusting, grooming, walking, running, and sniffing + resting. For this purpose, we utilized a modified version of BENTO (https://github.com/karinmcode/bentoMAT), an open-source Matlab GUI for managing multimodal neuroscience datasets allowing browsing between recording sessions.
To streamline the labeling process, we employed a prelabeling strategy that relied on the similarity between frames (https://github.com/karinmcode/bentoMAT/tree/master/bento/plugins/Prelabel_all_similar_frames). Firstly, we computed the histogram of oriented gradients of motion frames (20 by 20 division of each frame) and stored the resulting values in a matrix of n frames by m HOG values (histogram of oriented gradients). Next, we reduced the size of the data by applying PCA (principal component analysis) to the matrix. To obtain a 2D visualization, we used UMAP (Uniform Manifold Approximation and Projection) on the PCA outputs. Finally, we assigned a cluster ID to each frame using k-means clustering, with the option of using up to 12 clusters. All prelabels were manually reviewed and reassigned to one of the six categories using the modified GUI (BENTO). Our approach involved 2 different users, each labeling frames on their own and validating frames labeled by the other.
Statistical analysis
Statistical analysis was performed using Matlab (MathWorks). Unless reported otherwise, all statistics were described as mean ± SEM. Statistical significance was defined as p < 0.05 unless stated otherwise. For post hoc pairwise comparisons, we employed Tukey's test. The significance threshold was adjusted for multiple comparisons using the Bonferroni’s correction. One-sample Kolmogorov–Smirnov test served to determine the use of a nonparametric test. If the data followed a normal distribution and exhibited homoscedasticity (uniform variance), parametric tests such as the t test and ANOVA (analysis of variance) were used for analysis. For data that were either non-normally distributed or did not maintain homoscedasticity, nonparametric tests like the rank sum test (Wilcoxon rank sum test) and Kruskal–Wallis test were used. These nonparametric methods are more suitable for analyzing datasets that do not meet the assumptions required for parametric testing. To investigate the proportions of neurons across all auditory cortical fields and pairwise comparisons, a chi-square statistical test was used with Bonferroni’s correction.
Results
Sound-evoked responses in auditory cortex L2/3 are diversely modulated by locomotion
To understand how auditory cortex activity is influenced by movement, we made large-scale optical recordings from excitatory neurons in L2/3 of primary and nonprimary auditory cortex of awake, behaving mice (Table 1). GCaMP6s was expressed in excitatory neurons using transgenic breeding strategies, followed by wide-field and cellular imaging in multiple auditory cortical fields in the same mice (see Materials and Methods). Head-fixed mice were acclimated to running on a non-motorized wheel for 7–10 d prior to imaging (Fig. 1A; Schneider et al., 2018).
Using wide-field calcium imaging, we first identified four distinct auditory fields, including A1, AAF, A2, and DP (Fig. 1B; Romero et al., 2020; see Materials and Methods). We then used two-photon calcium imaging to measure sound-evoked responses from individual neurons in different auditory cortical fields (Fig. 1C,D). On each two-photon imaging day, one or more fields of view were imaged as mice rested and ran on the treadmill while hearing experimentally controlled sounds that were uncoupled from the mouse's behavior (Fig. 1C,D). The sounds were drawn from three distinct categories: pure tones of varying frequency, AMWN of varying modulation rates (AMWN), and up- or down-sweeping chirps. Overall, we recorded 1,869 sound-responsive neurons (see Materials and Methods; Fig. 1E,F). We selected these sound categories because of their distinct acoustic features, because of their capacity to activate neurons in primary and nonprimary auditory cortex, and because we anticipated that different auditory cortical areas might exhibit varying degrees of responsiveness to different sound categories (Linden et al., 2003; Trujillo et al., 2011; Hoglen et al., 2018).
We first analyzed the sound tuning of individual auditory cortex cells while mice were at rest (Fig. 2A). Across the sound-responsive population (n = 1,869), the majority of neurons (75%) were responsive to one or more sound types during rest. Most neurons (57%) showed selectivity toward sounds from only one of the three categories, indicating a preference for specific types of sounds (Fig. 2B). Of the three categories, the majority of neurons were responsive to sounds in the AMWN category (27%), potentially because this category included more unique sound stimuli (9 AMWN frequencies vs 4 pure tones and 2 chirps) and therefore was more likely to have at least one sound within a neuron's receptive field. A smaller subset of neurons (16%) were responsive to sounds in two different categories, while fewer than 2% of neurons were responsive to sounds in all three categories.
L2/3 neurons preferably respond to one type of sound and behavioral state but their preferences do not cluster spatially. A, Example sound responses. Left, Highlighted imaged layer 2/3 cells. Right, Fluorescence responses to sound (gray shade). Pure tone, AMWN, and chirp responses are represented in consecutive columns. Black traces represent the average. Gray lines represent individual trials. B, Venn diagrams illustrating the proportions of neurons that respond to sounds during rest and locomotion for different sound types. C, Pie chart summarizing sound-type specificity changes between rest and locomotion states. The chart categorizes the neurons into four groups based on their responsiveness: those that show no difference in responsiveness, those that are exclusively responsive during rest or running conditions, and those that change their responsiveness during locomotion. D, Four example tuning curves of neurons responding to more than two categories of sounds (from top to bottom panel). The x-axis represents the sound stimuli. Stars denote significant sound responses. E, Grayscale heatmap showing all neurons responding to more than two sound categories. Neurons are sorted by their preferred sound types and stimulus frequency. The x-axis represents the sound stimuli. F, Comparison of average tuning average bandwidth across neuronal populations broken down by their selectivity to sound types: responding to PT only (PT), responding to PT and AMWN sounds (PT + AM), responding to pure tones and chirps (PT + chirps), and responding to all sound types (PT + AM + chirps). Asterisks indicate significant comparison between groups results from an ANOVA test (p < 0.001), followed by a post hoc Tukey’s test (p < 0.01). G, Map of auditory cortex cells sound class preference shown as colored dots. The color of the dots uses the same color code as in panel B. Auditory fields are delineated in gray. The map is aligned to the 4 kHz A1 epifluorescence sound response. H, Sound-type preference broken down by area. Color code is the same as in B. Chi-square test revealed that the proportions of sound class preferences were significantly different during rest (p < 0.001), but not during locomotion (p > 0.05). I, Specificity change broken down by area. Color code is the same as in C.
We next asked how sound responses were influenced by locomotion. During locomotion, the proportions of neurons responsive to each sound category were similar to the resting condition, with 2% of neurons responsive to sounds in all three categories (Fig. 2B). Twenty-four percent of neurons were not responsive to any sounds heard during rest but were responsive to at least one sound while the mouse was locomoting (Fig. 2C, run specific), indicating that some auditory cortex cells are selectively responsive only when a mouse is running. A smaller fraction of neurons (19%) were sound responsive only when the mouse was at rest. Thirty-six percent of neurons were responsive to sounds during both behavioral conditions and maintained their tuning to the same sound category across conditions. A small fraction of neurons became responsive to fewer categories during locomotion (8%), some became responsive to sounds in more categories (8%), and only 5% of neurons switched their responsiveness from one sound category to another. Taken together, these results show that locomotion leads to a slight increase in the number of sound-responsive neurons in the auditory cortex and that neurons largely retain their sound-category responsiveness across periods of rest and locomotion.
We focused our next analysis on neurons that are responsive to multiple sound categories. First, we plotted the responses of individual neurons to all sounds sound types and all individual sound tokens (Fig. 2D). Across the population, this visualization revealed many neurons that were responsive to PTs as well as AMWN and chirps (Fig. 2E). We then explored whether neurons responding to multiple sound types also responded to a broader range of PT frequencies compared with neurons only responsive to the PT category (Fig. 2F). Our findings revealed that sound-type selectivity indeed affects the bandwidth of responses to PT sounds (ANOVA test; p < 0.001). Specifically, the response bandwidth for neurons responding only to PT and those responding to both PT and AM was significantly narrower than for neurons responding to PT + chirp and PT + AM + chirp (post hoc Tukey’s test; p = 0.006; PT, 1.13 ± 0.02; PT + AM, 1.10 ± 0.03; PT + chirp, 1.32 ± 0.05; PT + AM + chirp, 1.50 ± 0.08). These findings suggest that multi-sound-category tuning may arise more readily in neurons that have broader frequency tuning.
State-dependent sound responses are consistent across auditory cortical fields
The auditory cortex comprises several distinct fields that are thought to be part of a hierarchy of auditory processing (Wessinger et al., 2001; Sharpee et al., 2011; Kline et al., 2021). Different fields receive different amounts of input from motor regions of the brain and might be influenced differentially by the behavioral state of the mouse (Henschke et al., 2021). Therefore, we next separated neurons based on their anatomical location within functionally defined auditory cortical fields (Fig. 2G), as assayed by wide-field calcium imaging (Liu et al., 2019; Romero et al., 2020).
For each mouse, we aligned wide-field imaging data to a stereotyped map of the cortical surface (Fig. 1B), and we overlaid all sound-responsive neurons onto a map which was color coded by the sound type that drove the largest (and typically only) response (Fig. 2G). By plotting only the pure tone responsive cells based on their best frequency tuning, our two-photon data confirmed the expected tonotopic gradient of A1 (Fig. 1C). In contrast, when plotting cells tuned to AMWN, we observed no periodotopic gradient across the A1 tonotopic axis (data not shown). During rest, we observed neurons responsive to all sound categories in each cortical field, and we found that individual neurons throughout the auditory cortex tended to be tuned to sounds from just one sound category (Fig. 2H). To test this more definitively, we drew boundaries around approximate cortical fields and classified neurons as belonging to each field (Fig. 2G). During rest and locomotion, the fractions of neurons that were responsive to each sound category were significantly different across auditory cortical fields (chi-square test; rest, p < 0.001; locomotion, p < 0.001; Fig. 2H). The largest fractions of neurons that responded only to tones were found in A1 (25%), consistent with more primary sensory regions encoding individual sound frequencies (Mizrahi et al., 2014). The proportion of neurons responsive to AMWN was larger in nonprimary areas (A2 = 34%; AAF = 31%; DP = 30%) compared with that in A1 (19%), consistent with higher level cortical regions encoding more spectrally broad and complex sounds (Romero et al., 2020).
In all fields of the auditory cortex, more neurons were responsive to sounds during locomotion compared with those during rest (Fig. 2I). As observed across the broader auditory cortical population, only a small subset of neurons in each cortical field changed their tuning from one sound category to another when mice were locomoting compared with resting (chi-square test; p = 0.005; Fig. 2I). Altogether, these results indicate that primary and nonprimary cortical fields process sounds in subtly yet significantly different ways but that locomotion-induced changes in category tuning of sound-evoked responses are largely consistent across auditory cortical fields.
Locomotion-related changes in response magnitude are locally diverse and globally similar
In the analyses described above, we binarized neurons as either being responsive to or unresponsive to sounds during each of the two behavioral states (resting and running, see Materials and Methods). We next more thoroughly quantified how the magnitudes of neural responses were influenced by locomotion. Individual auditory cortex neurons were diversely modulated by locomotion, with some neurons having larger responses, some neurons having largely stable responses, and other neurons having weaker responses during locomotion (Fig. 3A). Of the 1,869 neurons that were sound responsive in at least one behavioral condition, most had larger neural responses during locomotion compared with those during rest, leading to average responses that were 15% larger during locomotion (average sound response R; Rrest = 0.45; Rlocomotion = 0.57; Fig. 3A–C). To analyze responses at the level of individual neurons, we calculated a MI value for each neuron that described the ratio of responses during locomotion and rest (range, −1 to 1; see Materials and Methods; Fig. 3C,D). The distribution of MI values skewed positive, indicating that on average, sound-evoked responses were stronger during locomotion compared with those during rest (MI = 0.03 ± 0.01). While the average response was larger during running compared with that during rest (Fig. 3D), we noted that this derived primarily from a positive offset of the intercept accompanied by a slope that was <1, rather than a steeper slope of the regression line, suggesting that smaller sound responses are enhanced by locomotion while bigger responses are suppressed (Fig. 3C).
Running increases auditory cortex sound responses similarly across areas. A, Example (top 3 rows) and average (bottom row) sound responses in resting and running condition. Running MIs are reported in green. Each sound lasted 1 s and is represented by a gray shade. B, Heatmap illustrating individual normalized sound responses during two conditions: rest (left) and locomotion (right). Start and end times of sound stimuli are demarcated by the white dotted lines. To enhance visualization, we sorted all neurons based on the timing of their peak response, after being separated into rest-preferring neurons (on the left) and locomotion-preferring neurons (on the right). C, Neuronal sound responses under rest and running conditions, with each dot representing a neuron and the color indicating its MI. Small sound responses were excluded (see Materials and Methods). The dotted line represents the unity line, while the pink line shows the linear regression fit. D, Distribution of MIs with boxplot and red line indicating the median and quartiles of the distribution. The t test significance indicates a slightly positive population MI. E, Spatial organization of running modulated cells illustrated as MI map with each dot representing a significantly modulated cell. Significance was determined as sound responses in rest and running being significantly different. Gray lines delineate auditory cortex areas. F, MIs reported by boxplots for cells with significant sound responses in any conditions (gray, left) and in rest condition only (blue, right). Median is represented by the red line. Top and bottom boundaries of the box signify the 75th and 25th percentiles, respectively, and the whiskers extend to 1.5 times the interquartile range (IQR). Same color scheme applies to panels G and H. G, MIs grouped by the auditory cortex area. H, MIs grouped by quartiles of rest response sizes.
We next looked at the distribution of MI values across different auditory cortical fields. To visualize modulation across the cortical surface and across cortical fields, we overlaid all sound-responsive neurons onto a stereotyped map, color coded by their MI (Fig. 3E). Neurons that were strongly positively and strongly negatively modulated during locomotion appeared to be distributed rather uniformly across the broader auditory cortical surface (Fig. 3E). Dividing the auditory cortex into spatial bins and analyzing the mode, mean, or median of each bin did not show any spatial gradient (data not shown). To test this more definitively, we drew boundaries around approximate cortical fields and measured the distribution of MI values for neurons within each field. We found that the average MI and range of modulation indices were similar across all cortical fields, with AAF and A1 tending to be more enhanced by locomotion compared with DP and A2 (MI AAF, 0.04 ± 0.03; A1, 0.06 ± 0.02; DP, 0.01 ± 0.03; A2, −0.02 ± 0.03; Kruskal–Wallis test; p = 0.054; Fig. 3G).
Our finding of enhanced responses during locomotion is consistent with previous work using calcium imaging to monitor auditory cortex activity but deviates from findings using electrophysiology, which tend to reveal weaker responses during locomotion compared with rest (Schneider et al., 2014; Henschke et al., 2021; Vivaldo et al., 2022). A potential reason for this discrepancy is that some electrophysiology studies focus only on neurons responsive during rest, overlooking those responsive solely during locomotion. This approach could skew results toward showing reduced activity while moving (Schneider et al., 2014, 2018). Additionally, electrophysiology tends to pick up neurons with higher firing rates, whereas calcium imaging detects a broader range of neuronal activities (Bagur et al., 2022). We therefore first restricted our next analysis only to neurons that were responsive to sounds during rest, and we found that across all neurons and in each cortical field, locomotion led to suppression rather than enhancement (MIrest cells = −0.18 ± 0.01; Fig. 3F,G). We then categorized neurons into quartiles based on their sound-evoked response intensity at rest. This separation allowed us to examine weakly and strongly responsive neurons independently. Our findings revealed that only neurons in the lowest quartile showed enhanced activity during locomotion, while those in the upper 50th percentile displayed reduced activity (MIall cells Q1, 0.64 ± 0.01; Q2, −0.09 ± 0.02; Q3, −0.20 ± 0.02; Q4, −0.21 ± 0.02), confirming our linear regression analysis (Fig. 3H; Fig. 3C). A similar analysis, limited to neurons active during rest, indicated an even more pronounced effect, with no neuron group showing increased activity during locomotion (MIrest cells Q1, −0.08 ± 0.03; Q2, −0.21 ± 0.03; Q3, −0.21 ± 0.02; Q4, −0.21 ± 0.02). In summary, our results suggest that divergent findings between calcium imaging and electrophysiology studies may arise in part due to using different criteria for sound-responsive neuronal groups and partly based on method-specific cell inclusion.
Locomotion speed is encoded across auditory cortical fields
We find strong locomotion-related changes in neural activity that are heterogeneous across neurons and largely homogenous across cortical fields. But locomotion is not necessarily an all or nothing event, and mice can run fast, slow, and in between. In addition, other factors could also influence sound-evoked activity during behavior, such as how strongly a neuron responds to sounds during rest (Fig. 3H), the latency of a neuron's peak response during sound playback, the cortical field in which a neuron resides, the depth within layer 2/3 at which a neuron resides, or the sound type to which a neuron is responsive. We reasoned that these attributes could all reflect aspects of a neuron's function within the auditory cortical hierarchy and could presumably relate to whether or how a neuron's sound-evoked response is influenced by locomotion.
To determine which factors most strongly contributed to a neuron's locomotion-related modulation, we first segregated trials (i.e., sound events) based on the speed at which the mouse was running at the time of sound playback. We found that as the mouse ran faster, the MI increased significantly (Fig. 4A). We then calculated the effect size of trial speed and other variables (Fig. 4B), which revealed that running speed had the largest impact on neural activity, followed by the strength of a neuron's response to sound during rest.
Auditory cortex contains speed-modulated neurons. A, Average trial speed relationship to MI shown as boxplots for various trial speed bins. Significance test reported in gray. B, Influence of recording variables on neuronal MI, as computed by the effect size (ω2; see Materials and Methods). Error bars represent 95% confidence intervals of the effect size. C, Running speed decoding accuracy map of the auditory cortex. Circles represent neurons and their colors, how accurate speed decoding was for that recording session, as shown in the color bar. The black edge on the circle indicates that decoding of the speed of this recording was significant (SVM permutation test; see Materials and Methods). Gray lines delineated auditory cortex areas. D, Bar plot of average speed decoding accuracy for individual areas of the auditory cortex. E, Example cells’ sound responses modulated diversely by speed. The top row shows sound responses, with each line representing a trial and the color representing the trial speed. Gray background box represents sound timing. The bottom row shows the relationship between speed and sound response size. Data are fitted by functions denoted in the panel. F, Scatterplot of all neurons (dots) showing the goodness of the fit (measured by R2), the fitting p value, and the best fit shown by the color of the dots (legend in panel E). Significance alpha is symbolized by a dotted line at p = 0.05. Black dots represent neurons with no significant fit. G, Bar plot quantifying the proportions of neurons’ best fits among linear, exponential (exp), step-function (step), and logarithmic (log). The pie chart summarizes the proportion of speed-modulated cells. H, Bar plot showing the proportion of speed-modulated cells across individual areas of the auditory cortex.
Given the importance of running speed on neural responses, we next attempted to decode the mouse's running speed during sound playback from populations of simultaneously recorded neurons in each distinct auditory cortex field (Fig. 4C). For each recording session, we trained a linear support vector machine while holding out one of the trials and then predicted the running speed on the held out trial (i.e., leave-one-out cross-validation; see Materials and Methods). We considered decoding accurate if we could predict the running speed within 2 cm/s on a trial. We found that running speed could be decoded with ∼65% accuracy from most recording sessions (Fig. 4C,D). While there was heterogeneity in decoding accuracy across sessions, this variability did not map onto distinct cortical regions; decoding accuracy was equivalent in A1, A2, AAF, and DP (Fig. 4D).
Given that running speed can be decoded at the population level, we next wanted to identify how running speed was encoded at the level of individual neurons. We identified individual neurons with responses that were enhanced during locomotion, suppressed during locomotion, or unmodulated (Fig. 4E). For each neuron, we fit four different models to its sound-evoked response magnitude at different running speeds (linear, exponential, step, and logarithmic; see Materials and Methods), we identified the model that explained the most variance in neural responses across trials (Fig. 4E). Neurons for which no model could achieve a significant fit to the data were categorized as non-modulated (Fig. 4E–G). Across all of the auditory cortex, we found that the majority of neurons (62%) were not significantly modulated by speed (Fig. 4G). Of the remaining 38% of neurons, the majority were best fit by either a linear or exponential model, with roughly equal proportions (Fig. 4G; plinear = 16%; pexp = 19%; pstep = 1%; plog = 2%). This finding indicates that a large fraction of auditory cortical neurons (38%) encode a predictable (proportional or exponential) relationship between the neuron's sound-evoked response and running speed.
Lastly, we categorized neurons by cortical field and compared the fraction of neurons that were non-modulated, positively modulated, and negatively modulated by locomotion. All fields contained neurons that were both positively and negatively modulated (Fig. 4H), consistent with our earlier analyses (Fig. 3E). Area DP had significantly more locomotion-modulated neurons than did any other area, with AAF having the fewest number of locomotion-modulated neurons (Fig. 4H; DP = 51%; A1 = 35%; A2 = 32%; AAF = 27%). Our findings indicate that speed modulation is distributed throughout the auditory cortex. Although most individual neurons are not significantly modulated by locomotion, speed can nonetheless be reliably decoded from small populations of neurons throughout the auditory cortex.
Auditory cortex neurons are modulated by multiple different movements
Although our analyses thus far have focused on one behavior, locomotion, we noted that mice also perform other behaviors while head restrained atop the treadmill. These behaviors include resting, grooming, and postural adjustments, as well as finer distinctions between different types of locomotion (e.g., walking and running). This provided an opportunity to determine whether the patterns of movement-based modulation we observe for running were similar across other types of behavior. Therefore, we asked whether these different behavioral states were reflected in distinct modulation patterns across the auditory cortex.
To begin, we labeled video frames when sound playback occurred and categorized each sound playback event as rest, running, walking, postural adjusting (“adjusting”; Ramadan et al., 2021), sniffing, or grooming (see Materials and Methods; Fig. 5A). For visualization purposes, we performed dimensionality reduction (UMAP) on individual video frames, allowing us to embed each frame as a single point in a low-dimensional space. This low-dimensional embedding revealed that distinct behaviors tend to be well separated using just two dimensions (Fig. 5A). We next used a pseudocolor map to label each frame based on the speed at which the mouse was moving (Fig. 5B). We found that resting, adjusting, and grooming all occurred at speeds at or near 0 cm/s. Locomotion states largely formed a distinct island in UMAP space, and we segregated walking from running based on speed and gate. Across recording sessions, we found that mice engaged in all behaviors that we measured, albeit with different rates (Fig. 5C).
Auditory cortex sound response modulation depends on the type of movement. A, Example manually labeled session presented as a UMAP plot for visualization purposes, where motion frames are dots and color coded based on behavioral labels, such as resting, grooming, adjusting, walking, and running. Example motion frames highlight motion in the previous frame in magenta and the next frame in green. The UMAP 2D representation was employed for prelabeling motion frames using k-means clustering. Each prelabeled cluster was then inspected and relabeled with a behavioral label (see Materials and Methods). B, Similar to panel A, the UMAP plot displays motion frames colored by running speed. C, Bar plot representing behavioral occurrences measured as the proportion of labeled frames. D–G, Example sound responses of neurons that are modulated differently by behaviors. H, Population average sound responses across different behaviors. I, Bar plot illustrating the size of sound responses (baseline subtracted) across different behaviors. Two-way ANOVA revealed a significant main effect of areas (ANOVA; p < 0.001). J, Bar plot depicting neuronal activity across different behaviors (sound responses without baseline subtraction). Two-way ANOVA revealed a significant main effect of areas (ANOVA, p < 0.001). K, Heatmap displaying the activity of auditory cortex neurons under different conditions. Neurons are sorted based on their maximum sound response size, and the percentage of neurons with maximum responses is indicated in white. White in the last column of the heatmap indicates neurons that are significantly modulated by behaviors. The color represents the normalized neuronal response, as shown in the color bar. L, Violin plot showing the proportion of behaviorally modulated neurons across multiple mice. M, Spatial distribution of cells that are differentially modulated by behaviors (green dots) compared with cells that are not influenced differentially by behaviors (red dots). The dashed lines indicate the reference point of A1 at 4 kHz, against which the maps are aligned. N, Bar plot quantifying the proportion of behaviorally modulated cells in each auditory cortex area. Bootstrapping analysis shows that the proportion of behaviorally modulated cells in A1 and A2 significantly differ from other areas (ANOVA; p < 0.001; post hoc Tukey–Kramer with Bonferroni’s correction). Error bars represent 99% confidence intervals of bootstrapping distributions. O, Spatial distribution of behaviorally modulated cells depicted by dots. Dot colors represent the best sound response conditions. In the background, a density plot shows the most common best sound response condition in a square of 200 µm. Gray lines delineate the limits of the areas of the auditory cortex. P, Bar plot presenting the proportions of preferred behavioral conditions in each auditory cortex area. Significant difference between proportion of favorite modulation between areas was revealed by performing a chi-square test (p = 0.003). Q, Error bar plot showing normalized sound responses across different behaviors and auditory cortex areas. Two-way ANOVA revealed a significant main effect of behaviors and areas (ANOVA; post hoc Tukey's HSD with Bonferroni’s correction, pAreas < 0.001, pBehaviors < 0.001, pInteraction > 0.05).
Next, we compared neural responses to sounds during each different behavior. We identified many individual neurons with responses that varied across behavioral states (Fig. 5D–F) and other neurons with responses that were largely invariant to state (Fig. 5G). Compared with rest, sound-evoked responses and spontaneous fluorescence were on average enhanced during running, walking, postural adjusting, and sniffing, while neural responses were suppressed during grooming (Fig. 5H). Baseline-subtracted sound-evoked responses were significantly weaker during grooming than during all other behavioral states (Fig. 5I). When the baseline was not subtracted, we found that responses during running were significantly higher than all other behaviors and that responses during grooming were significantly lower (Fig. 5J). Across the population of sound-responsive neurons, distinct groups of neurons had their strongest sound-evoked responses during different behaviors, and for a third of the neurons, their activity for one behavior was significantly different than that measured during other behaviors (Fig. 5K–L). Together, these findings reveal that auditory cortex is modulated by many behaviors, that not all behaviors modulate the auditory cortex equally, and that sound-evoked changes and baseline changes in activity both contribute to movement-related modulation in the auditory cortex.
Finally, we asked whether distinct auditory cortical fields were differentially modulated by behaviors other than locomotion. We plotted the spatial distribution of cells, color coded by whether or not they were modulated by each behavior (Fig. 5M). Overall, we found that A1 had the lowest fraction of behavior-modulated neurons, while A2 had the highest fraction (Fig. 5N). Across the auditory cortical surface, we did not observe any clear dominant modulation by a particular movement (Fig. 5O). However, A1 and AAF neurons differed significantly in the relative distributions of movements that modulated their activity (Fig. 5P). A2 tended to have weaker responses across most stimuli (Fig. 5Q). And across all areas, sound-evoked responses during grooming were weaker than during any other behavioral state (Fig. 5Q). These data together reveal that movement-related modulation in general is widespread throughout auditory cortical fields.
Discussion
Here we demonstrate that neurons throughout L2/3 of the mouse auditory cortex are modulated by locomotion and other behaviors, including walking, grooming, sniffing, and postural adjustments. At the level of single neurons, we find that movement-related modulation can be highly heterogeneous. But this heterogeneity is distributed nearly uniformly across auditory cortical fields, revealing a homogenous influence of behavioral state at a more macroscopic level. These observations provide important insights into how the auditory cortex integrates motor- and sound-related signals during behavior.
The mouse auditory cortex contains multiple distinct regions that can be identified using functional landmarks and tonotopic gradients. However, those regions are difficult to locate using anatomical coordinates (Romero et al., 2020; Narayanan et al., 2023). Furthermore, different auditory cortical fields have different responses to simple and complex sounds (Linden et al., 2003; Kline et al., 2023) and different long-range connectivity (Tsukano et al., 2017, 2019; Gămănuţ et al., 2018; Henschke et al., 2021). Despite these differences and despite the diverse modulation phenotypes that we observed during locomotion, we found that each cortical field at a macroscopic level had largely similar movement-related modulation. This includes a slight enhancement of sound-evoked responses during movement, a stability of sound-category tuning during both locomotion and rest, reliable encoding of locomotion speed during periods of running, and modulation by multiple uncued behaviors.
In addition to these similarities, we identified significant differences across auditory cortical fields. First, we noted that the largest fractions of neurons that responded only to tones were found in A1 and A2 and the largest fraction of neurons responsive only to AMWN was observed in DP, AAF, and A2 (Fig. 2H). These observations align with a hierarchical change in the complexity of tuning curves across cortical fields (Mizrahi et al., 2014; Romero et al., 2020). Second, we saw that DP had more speed-sensitive neurons than any other field, while AAF had fewer speed-sensitive neurons than any other field (Fig. 4H). These differences in modulation may reflect underlying connectivity differences across areas, particularly with respect to long-range behavior-related inputs (Henschke et al., 2021). Finally, when we expanded our analysis to include other behaviors, we found that A2 had the highest fraction of movement-modulated cells while A1 had the lowest fraction (Fig. 5N). Our findings largely complement previous observations of diverse and distributed movement-related modulation across the auditory cortex, including subtly different modulation patterns in primary versus dorsal auditory cortex (Henschke et al., 2021).
While locomotion has been the most commonly studied behavior for sensory modulation in mice, other behaviors are also known to influence sensory cortex. Electrophysiological recordings made during forelimb behaviors reveal a net suppression of sound-evoked responses in the auditory cortex, but also heterogeneous effects at the single-cell level, including many neurons that are enhanced or unaffected by movement (Rummell et al., 2016; Audette et al., 2022). In addition, movements including running, grooming, postural adjustments, and vocalizing all lead to similar changes in subthreshold membrane potential dynamics, consistent with our observation of similar changes in calcium responses across multiple different behaviors (Schneider et al., 2014). Together with these prior reports, our findings suggest that the modulation observed during locomotion is not categorically different from the modulation observed during other behaviors. This consistency across behaviors could indicate that movement-related modulation is not related to movement per se but may instead be related to arousal (McGinley et al., 2015). However, arousal and movement tend to have dissociable effects on sensory cortical activity (Vinck et al., 2015), and different movements can be decoded from neural activity throughout much of the dorsal cortical surface (Mimica et al., 2018; Musall et al., 2019; Stringer et al., 2019). Consistent with these ideas, here we also find that different populations of auditory cortical neurons are differently modulated by distinct behaviors (Fig. 5P). We argue that rich movement-related information is a feature of sensory cortical activity that likely subserves perception while animals interact with the world.
One interpretation of the widespread modulation of auditory cortex is that behavior-related signals may be ubiquitously important for processing both simple and complex sounds. It also remains possible that behavior-related signals play different roles in distinct cortical fields. For example, while motor-related signals might be purely modulatory in some areas, they may serve as a teaching signal in others, such as when learning to anticipate the acoustic consequences of action (Schneider et al., 2018). Alternatively, movement signals may help animals learn to produce appropriate behaviors in response to different sensory input (Znamenskiy and Zador, 2013). Distinguishing among the many possible roles for movement-related modulation throughout the auditory cortex will be an important focus of future experiments.
Previous studies using electrophysiological measures (i.e., action potentials or membrane potentials) to quantify neural activity have observed suppressed sound responses during locomotion compared with rest (Schneider et al., 2014; Zhou et al., 2014; Rummell et al., 2016; Bigelow et al., 2019; Audette et al., 2022). In contrast, experiments using calcium indicators have reported enhanced responses during locomotion, consistent with our current observations (Henschke et al., 2021; Vivaldo et al., 2022). Importantly, when we restrict our analysis to neurons that are responsive to sounds during rest (regardless of whether those neurons are active during locomotion), we also observe net suppression of auditory cortex during locomotion (Schneider et al., 2018). We also found that neurons with larger responses tended to be suppressed during locomotion, consistent with potential biases in the sampling of higher firing rate neurons during electrophysiology (Fig. 3C,H; Bagur et al., 2022). Residual differences in the direction of modulation could stem from multiple factors, including sampling from different layers when monitoring calcium compared with electrophysiology; longer temporal analysis windows for calcium indicators; shorter stimuli used in electrophysiology studies (e.g., tens of milliseconds); and a nonlinear relationship between action potentials and calcium levels. We also note that there are some similarities between experiments using calcium indicators and those using electrical methods. Both experimental techniques reveal a rich heterogeneity of movement-related modulation. And when we fit a regression line to compare sound-evoked responses during running and resting, we observed a slope of less than one, consistent with the suppression observed using electrophysiological recordings (Fig. 3C). It has been speculated that this type of augmentation of receptive field tuning and bandwidth may be critical for attention-based tasks, which may be particularly important while moving (Fritz et al., 2003; Nelson and Mooney, 2016).
Finally, we note that the current experiments focus solely on excitatory neurons in L2/3. While this population has been previously shown to be strongly influenced by behavior (Keller et al., 2012; Audette et al., 2022), further differences in motor-related modulation might arise when sampling across cortical layers. Indeed, motor-related signals tend to be stronger in deep cortical layers compared with those in superficial layers (Audette et al., 2022). Moreover, excitatory and inhibitory cells are also differently influenced by behavior (Schneider et al., 2014), and signals related to violations from sensory expectations during behavior have recently been found to be concentrated in specific genetically defined populations of neurons in L2/3 (O’Toole et al., 2022). Future investigations with layer- and cell-type specificity may reveal more nuanced distinctions in how movement modulates different auditory cortical fields.
Footnotes
We thank members of the Schneider Lab for thoughtful feedback throughout this project. We extend our gratitude to Jessica Guevara for her expertise in animal care and technical assistance and to Carina Sun for her assistance in labeling video frames. We thank Alessandro La Chioma for his assistance in setting up the software for the epifluorescence imaging. This research was supported by the Swiss National Science Foundation (K.M.); the National Institutes of Health (1R01-DC018802 to D.M.S.); a Career Award at the Scientific Interface from the Burroughs Wellcome Fund (D.M.S.); fellowships from the Searle Scholars Program, the Alfred P. Sloan Foundation and the McKnight Foundation (D.M.S.); and an investigator award from the New York Stem Cell Foundation (D.M.S.). D.M.S. is a New York Stem Cell Foundation - Robertson Neuroscience Investigator.
The authors declare no competing financial interests.
- Correspondence should be addressed to David M. Schneider at david.schneider{at}nyu.edu.
