Characterization of Auditory Responsive Neurons in the Mouse Superior Colliculus to White Noise and Frequency-Restricted Sounds

Abstract

Locating the source of a specific sound in a complex environment and determining its saliency is critical for survival. The superior colliculus (SC), a sensorimotor midbrain structure, plays an important role in sound localization and has been shown to have a topographic map of the auditory space in a range of species. In mice, previous studies using broadband white noise stimuli found that high-frequency monaural spectral cues and interaural-level differences (ILDs) are used to generate a neuron's spatially restricted receptive field (RF) and that these RFs are organized topographically along the azimuth. However, in a naturalistic environment, the auditory stimuli that an animal encounters may have restricted spectral components, although these sound sources can still be localized efficiently. It remains unknown whether and how the SC neurons respond to frequency-restricted sounds and, in turn, how this changes the organization of their RFs into a topographic map. Here, we show results from large-scale in vivo physiological recordings of SC neurons from male and female mice in response to white noise, naturalistic ultrasonic pup call, and chirps. We find that mouse SC auditory neurons respond to a pup call and chirps with distinct temporal patterns and a spatial preference predominantly at ∼60° in contralateral azimuth. In addition, we categorized auditory SC neurons based on their spectrotemporal RF patterns and demonstrated that there are at least four classes of auditory-responsive neurons in the SC that lie in different locations along the anteroposterior axis of the SC.

Significance Statement

The superior colliculus (SC) receives topographically organized visual and auditory inputs used for object localization. While visually responsive SC neurons are well studied, less is known about auditory neurons. We presented white noise, pup call, and chirp stimuli to mice while recording from SC neurons. Analysis of the responses defines distinct classes of auditory neurons. Interestingly, while auditory neurons respond to pup calls, the responses are not organized topographically. Despite this, the population response to a pup call predicts the location of a stimulus, albeit not as well as that generated from white noise. These results show that the SC may use different strategies to localize sound depending on the spectral composition of the source.

Introduction

The superior colliculus (SC) is a sensorimotor midbrain structure that integrates visual, auditory, and somatosensory inputs and initiates motor commands for orienting behaviors. The SC can be divided into two layers: the superficial SC (sSC), which contains neurons that are exclusively visually responsive, and the deep SC (dSC), which contains neurons that respond to visual, auditory, and/or somatosensory stimuli (Cang et al., 2018; Ito and Feldheim, 2018; Basso et al., 2021; Liu et al., 2022; Cang et al., 2024). The visual response properties of the mouse sSC have been characterized using electrophysiology and calcium imaging to identify 24 functional subtypes (Wang et al., 2010; Inayat et al., 2015; Ito et al., 2017; De Franceschi and Solomon, 2018; Li and Meister, 2023). Much less is known about the response properties and subtypes of the dSC auditory-responsive neurons.

Studies from our lab have provided insights into the binaural and monaural cues used by SC neurons to compute sound location (Ito et al., 2020; Si et al., 2022). In mice, about half of the neurons in the dSC respond to sound, and ∼23% of these have spatially restricted RFs. These neurons are organized as a map of sound location in the dSC, with neurons in the anterior dSC responding to sound from the front of the animal and posterior dSC neurons responding to sound from the contralateral side. Interestingly, dSC neurons with frontal and lateral RFs use different cues to form their spatial RFs; neurons with frontal RFs use monaural spectral cues, while neurons with lateral RFs use interaural-level differences (ILDs; Ito et al., 2020; Si et al., 2022). One seemingly disadvantage of localizing sound using spectral cues is that the spectra of the sound that arrives at the eardrums are influenced by both the spectra of the source sound and the modulation of the spectra due to the shape of the animals’ ears and head, but the brain cannot distinguish between them. Therefore, if the spectrum of the source sound has restricted spectral components, it is hypothesized that they will be unable to provide accurate information about the sound source location; thus, the brain circuitry may need to use other strategies to localize such stimuli. In fact, many forms of naturalistic sounds that convey important information for an animal are frequency-restricted, and behaviorally, it has been shown that animals are able to localize such sounds. For example, mothers can localize their pups that are making ultrasonic vocalizations (pup calls) that have a restricted set of ultrasonic frequencies (Smotherman et al., 1974; Smith, 1976).

To better understand how SC neurons respond to and localize sounds, we characterized the auditory response properties of SC neurons to various stimuli, including white noise, a naturalistic pup call, and chirps of different pitch, presented to head-fixed mice while recording electrophysiological responses in the dSC. We characterized the temporal response patterns and found SC auditory neurons respond to a pup call with distinct patterns, suggesting functional subtypes of auditory neurons. We then determined the spatial RFs of SC neurons when responding to the pup call and found consistent tuning for spatially restricted pup calls from ∼60° contralateral azimuth. Although the organization of the sound location map is disrupted with the pup call stimulus, the population response of dSC neurons can predict its location better than chance, though not as accurately as predictions from the topographically organized RFs found in response to white noise.

To better understand the functional subtypes of auditory neurons in the SC, we determined the spectrotemporal receptive fields (STRFs) of auditory-responsive neurons and showed that STRFs can be used to categorize SC neurons into at least four classes. We also show that STRF subtype identity can predict how a neuron responds to other auditory stimuli and its location in the SC. These results provide new insights into how SC neurons localize sounds and demonstrate a way to identify subtypes of auditory-responsive neurons in the SC.

Materials and Methods

Ethics statement

All procedures were performed in accordance with the University of California, Santa Cruz (UCSC) Institutional Animal Care and Use Committee.

Procedures

We used the mouse head-related transfer functions (HRTFs) to present virtual auditory space (VAS) stimuli and recorded the activity of the collicular neurons from head-fixed, alert mice using 256-electrode multishank silicon probes. Our experimental procedures were described previously (Shanks et al., 2016; Ito et al., 2017, 2020) and are detailed below.

Auditory stimulation

HRTF measurement and VAS stimulation. The HRTFs of CBA/CaJ (Ito et al., 2020) mice were measured, and the VAS stimuli were generated as previously described (Ito et al., 2020). The measured HRTFs for CBA/CaJ mice and the explanation for their use are available in the figshare website with the identifiers https://doi.org/10.6084/m9.figshare.11690577 and https://doi.org/10.6084/m9.figshare.11691615.

VAS stimuli were generated as previously described (Ito et al., 2020). In brief, each stimulus sound was filtered by a zero-phase inverse filter of the ES1 speaker (Tucker-Davis Technologies), the eHRTF, and the measured HRTF. The VAS stimulus is then delivered pointing toward the animal's ear canals using EC1 speakers (Tucker-Davis Technologies). We tried to reproduce the physical configurations used for the HRTF measurements so that the acoustic effect induced by this setup is canceled by an inverse filter of the eHRTFs.

Full-field white noise stimulus. The baseline stimulus pattern was 50 dB SPL, 100 ms white noise containing a flat 5–80 kHz spectrum with linear tapering windows in the first and last 5 ms. After applying the filters mentioned above, the full-field white noise stimulus contains grid points of five elevations (0–80° with 20° steps) and 17 azimuths (−144 to 144° with 18° steps), totaling 85 points in the two-directional field. The stimulus was presented every 2 s and repeated 30 times per direction at the intensity of 50 dB SPL.

Random chord stimulus. To measure spectral tuning properties of the neurons with localized RFs, we used dynamic random chord stimuli and calculated the spike-triggered average (STA) of the stimuli for each neuron. The stimuli consist of 48 tones ranging from 5 to 80 kHz (12 tones per octave). Each pattern was either 10 or 20 ms long with 3 ms linear tapering at the beginning and the end. In each pattern, the tones were randomly set to either ON or OFF with a probability of 0.5. The total number of tones per pattern was not fixed. The amplitude of each tone was fixed and set to be 50 dB SPL when averaged over time. One presentation of the stimuli was 2 min long, and this presentation with the same set of patterns was repeated 20 times to produce a 40-min-long stimulus. Tones from the left and right speakers were not correlated with each other to measure the tuning to contrast between the ears. We did not use a specific HRTF for this experiment, but simply canceled the eHRTF so that the stimulus sound was filtered to have a flat frequency response near the eardrums.

Pup call stimulus. To get a representative audio sample of a mouse pup call, we recorded from Postnatal Day (P)5–P9 CBA/CaJ mouse pups with a high-frequency bat detector (Echo Meter Touch, Wildlife Acoustics) and selected an audio clip with the most typical spectrogram. To generate the baseline stimulus, filters were then applied to the recorded audio to calibrate the effect from the recording device. After applying the HRTF filters described above, the horizontal pup call stimulus contains 17 azimuths (−144 to 144° with 18° steps) on the horizontal plane. The stimulus was presented every 1 s and repeated 120 times per direction at the intensity of 50 dB SPL.

Chirp stimulus. To assess the response of auditory SC neurons to simplified synthetic stimuli, we generated a set of chirp stimuli that cover the 60 kHz frequency range of 20–80 kHz, at frequency rates of 1, 3, and 6 kHz/second, and 80–20 kHz at a rate of −60 kHz/second. After applying the HRTF filters described above, the horizontal chirp stimuli contain five azimuths (0, 36, 54, 72, 90°) that cover a quadrant of the horizontal plane. The stimulus was presented every 1 s and repeated 60 times per direction at the intensity of 50 dB SPL.

Loudspeaker experiments. CBA/CaJ strain mice were used (n = 7; five females, two males; 2–4 months old). Surgeries and recordings were done as above. Before each experiment, the probe was coated with a lipophilic dye, DiD (Invitrogen, D7757). After each recording, the mouse was perfused, and brains were dissected. Anatomical registration of the sections and probe tracks were used to map the anterior to posterior position of the neurons.

The VAS system was used to simulate white noise stimuli from −135 to 135° azimuth in 15° intervals for 30 trials at 19 azimuthal positions every second in an anechoic box. Stimulus intensity for the VAS was done at 50 dB. Two free-field electrostatic loudspeakers (Tucker-Davis Technologies, ES1 speakers) were then placed 25 cm away from the ear canal of the mouse at the 0 and 60° azimuthal, 0° elevation, and foam to attempt to reduce the acoustic noise. Stimuli were randomized between the two speakers for both pup call and white noise stimuli, 30 trials each every second. Stimulus for the loudspeaker was recorded at 54 dB SPL at the site of the ear canal.

For the analysis, we used the VAS response to determine which neurons had spatially restricted RFs as described above. Of those neurons, we determined which responded to both pup call and white noise stimuli (n = 63). We then calculated the normalized difference fraction for each neuron, defined as follows:

$${\displaystyle \frac{[(\mathrm{total}\mathrm{AP}\mathrm{at}{60}^{\circ })-(\mathrm{total}\mathrm{AP}\mathrm{at}{0}^{\circ })]}{[(\mathrm{total}\mathrm{AP}\mathrm{at}{60}^{\circ })+(\mathrm{total}\mathrm{AP}\mathrm{at}{0}^{\circ })]}.}$$

Mice

The CBA/CaJ mice used in this study were the offspring of the mice initially obtained from The Jackson Laboratory (stocks #000664) and bred in the UCSC vivarium. We used 2 to 4 month mice of each sex.

Electrophysiology

One day before the recording, a custom-made titanium head plate is implanted on the mouse's skull to fix the mouse head to the surgery rig and recording rig without damaging or touching the ears. On the day of the recording, mice were anesthetized with isoflurane (3% induction, 1.2–1.5% maintenance; in 100% oxygen), and a craniotomy was made in the left hemisphere above the SC (∼1 by 2 mm in an elliptic shape, ∼0.6 mm from the lambda suture). After recovering from anesthesia, the mouse was head-fixed onto the recording rig, where it was allowed to locomote on a rotating cylinder. A 256-channel, four-shank silicon probe was inserted into the SC with the support of a thin layer of 2% low-melting-point agarose (in saline), and a layer of mineral oil was added on top to prevent the brain from drying. The silicon probe was aligned along the anteroposterior (A–P) axis. The multiunit visual RFs were measured as the probe recorded from the sSC in all four shanks. After this measurement, the probe was lowered until most of the channels were within the dSC (approximately just past the strong visual response in the sSC). Then we lowered and raised the probe by 100–120 μm to avoid the probe movement effect of the dragged tissue during recording. Recordings were started ∼20–30 min after inserting the probe and performed in an anechoic chamber.

The probes used for our recordings were kindly provided by Prof. Masmanidis at UCLA (Du et al., 2011; Shobe et al., 2015). We use 256-channel, four-shank silicon probes (the 256A series). Each shank has a two-dimensional active recording area of ∼1 mm × 86 µm, and the shank pitch is 400 µm. The voltage signals collected from all the electrodes are amplified and sampled at 20 kHz using an RHD2000 256-channel recording and data acquisition system (Intan Technologies).

Experimental design and statistical analyses

Blind analysis. To validate our results and reduce potential false-positive findings, we performed blind analysis (MacCoun and Perlmutter, 2015). First, we analyzed the spiking activity data from half of the neurons randomly sampled from each recording. After exploring the datasets and fixing the parameters in our analysis, we performed the same analysis on the other half of the data to test whether the conclusion holds true on the blinded data. All the results reported in this study passed a significance test in both the exploratory and the blinded data unless otherwise stated.

Spike sorting. We used a custom-designed software for spike sorting (Litke et al., 2004), which was also used in our previous published work (Ito et al., 2017, 2020, 2021; Si et al., 2022). Raw analog signals were high-pass filtered at ∼313 Hz and thresholded for spike detection. The high-pass part was used for single-unit identification after motion artifact removal. Detected spikes were clustered based on the principal component analysis (PCA) of the spike waveforms on the seed electrode and its surrounding electrodes. A mixture of Gaussians models was used for cluster identification. Detailed demonstration of the silicon probe schematics, examples of raw electrophysiology data, and demonstration of the spike-sorting process can be found in our previous work (Si et al., 2022).

Estimating AP location of neurons

We estimated the relative position of the silicon probes using the visual RF positions measured in the sSC, as described in our previous work (Ito et al., 2017, 2020, 2021; Si et al., 2022). In brief, we measured the positions of the visual RFs on each shank in the sSC using multiunit activity and extrapolated the visual RFs to find an A–P position where the visual RF azimuth was 0°; then by assuming the most superficial electrode of the silicon probe was at 300 μm in depth and the insertion angle of the probe into the SC was ∼25°, we were able to estimate the A–P positions of the auditory neurons (Ito et al., 2020, 2021; Si et al., 2022). The position of a neuron relative to the probe was determined by a two-dimensional Gaussian fit to its spike amplitude across multiple electrodes (Ito et al., 2017, 2020, 2021; Si et al., 2022). We only analyzed neurons with a positive A–P position in order not to include neurons located outside the SC.

Population encoding

The support vector machine classifier (SVC) from python scikit-learn packages (Pedregosa et al., 2011) was used to predict the sound location based on a single trial. For each of the 78 neurons that had spatial RFs in response to both white noise and pup call stimuli, we determined the number of action potentials fired within the 0–20 ms time bin after stimulation. One hundred twenty repetitions of white noise and pup call stimuli across 17 azimuths create a dataset of 2,040 trials for each stimulus. The model was trained on 1,360 (67%) of the trials equally distributed from all azimuths, which predicted the sound location using the remaining 680 (33%) trials. The error was calculated by taking the absolute value of the actual stimuli location subtracted by the model's prediction. The mean and standard deviation of the error is then reported across all 680 trials. The random distribution was determined based on the distribution of all possible error outcomes. Kolmogorov–Smirnov (KS) tests were used with Bonferroni’s multitest comparison correction to determine similar distributions.

Significance test for the auditory responses

White noise response. We used quasi-Poisson statistics for significance tests of the auditory responses of individual neurons (Ver Hoef and Boveng, 2007), to deal with the overdispersion of the poststimulus firing rate because of factors such as bursting of the neural activity, as described in our previously published work (Ito et al., 2020). The overdispersion parameter was estimated by the variance of the spike count divided by the mean, which should be 1 if the spiking activity of a neuron is poissonian. To determine the significance of the response, we first estimated an overdispersion parameter and considered a response to be significant if the p value of the neuron's spike count is below 0.001 [p = 1 − CDF(N), where CDF is a cumulative distribution function of the quasi-Poisson distribution and N is the spike count of the neuron].

Pup call and chirp response. We used Poisson statistics for significance tests of the auditory responses of individual neurons to pup calls and chirp responses. We first estimated the time course of the typical response to auditory stimuli (Figs. 1C, 3) and found that the typical response of most neurons happens within 20 ms after sound onset/offset. Therefore, we evaluated the spiking activities in three time windows: onset (0–20 ms after sound onset), offset (0–20 ms after sound offset), and in-between (between 20 ms after sound onset and sound offset, to capture the atypical response patterns of some neurons). We measured the baseline firing rate and calculated the Z-score within each time window at each azimuth using Poisson error and selected neurons with a >2 Z-score at any azimuthal location as significantly responsive to the stimulus.

Figure 1.

Pup call experiment setup and example responsive neurons in the SC. A, A schematic of the experiment setup: head-fixed mice were presented with virtual pup call stimuli at 17 azimuthal locations at 0° elevation with earphone speakers pointing to the ear canals. B, The spectrogram of the pup call stimulus that was used in this study. C, The temporal response to the pup call stimulus of SC neurons that are responsive to the pup call stimulus. (The red horizontal line marks the baseline firing rate; the dashed gray lines mark the first 20 ms of the onset and offset of the stimulus.) D, The raster plot of three example neurons with onset and/or offset response to localized pup call stimuli. E, PSTH of the three examples of pup call-responsive neurons in D. (The shaded area represents the presence of the stimulus. This is an average response across all azimuths for each neuron.)

Figure 3.

The spatial RFs of the white noise and pup call response can be different. A–C, Three example neurons that are responsive to both white noise and pup call stimuli. a, b, PSTHs of the white noise response (a) and pup call response (b). c, azimuthal RFs of the onset (c, defined as 0–20 ms poststimulus onset) and offset (d, defined as 0–20 ms poststimulus offset) response to white noise (blue) and pup call (red) stimuli. D, The percentage of neurons that respond to white noise and/or pup call stimuli, out of all neurons recorded. E, The percentage of neurons that have a RF azimuth centered at <50°, between 50 and 80°, or > 80° when responding to white noise (white) and pup call (gray) stimuli, out of all neurons that significantly respond to both. F, Histograms of the correlation between the azimuthal RFs of white noise and pup call stimuli for neurons that are responsive to both in their onset (left) and offset (right) response (the red dash line marks the mean, and the red numbers are mean ± SEM of the distribution; red arrowheads in the left panel mark the two peaks in the distribution).

Dynamic random chord response. We calculated the STA from the spiking response to the dynamic random chord stimulus described above and determined whether an individual neuron had a significant STRF. First, the spikes of each neuron are discretized in time bins the same as the stimulus segment size (5 ms), and the average of the stimulus patterns that preceded each spike was calculated. The stimulus was considered as 1 if there is a tone in the frequency and 0 otherwise. The significance tests for the STRFs were performed by examining whether the difference between the mean of the stimulus and the STA is significant. We used 0.001 as a significance threshold and Bonferroni’s correction for evaluating multiple time bins, frequencies, and contra-/ipsilateral input. Neurons with <20 spikes during the stimulus were not analyzed.

Function fit to estimate the azimuth of the spatial RFs of the neurons

White noise response. We used a maximum-likelihood fit of the Kent distribution (Kent, 1982) to estimate the azimuth of an RF for the full-field white noise stimulus (17 azimuthal × 5 elevational locations), as in our previous work (Ito et al., 2020, 2021; Si et al., 2022). At each point of the directional field, the likelihood value was calculated based on quasi-Poisson statistics (Ver Hoef and Boveng, 2007). The error of each parameter was estimated from the Hessian matrix of the likelihood function. As in our previous work, the front-Z coordinate system has a discontinuity of the azimuth across the midline that gives a problem in fitting and interpretation of the data near the midline. We avoided this issue of fitting the data near the midline with the front-Z coordinate system (determined by our HTRF measurement) by switching to the top-Z coordinate system and only used neurons with their elevation smaller than 30° for azimuthal topography, to avoid an area where two coordinate systems are largely different. To achieve stable fits, we set the azimuthal range to be from −144 to +144° and set the elevation range to be from 0 to 90°.

Pup call and chirp response. We used virtual space auditory stimuli at 17 or 5 azimuthal locations on the horizontal plane for the pup calls and chirp stimuli. For the neurons' response to these stimuli, we used a maximum-likelihood fit of the Gaussian distribution to estimate the center azimuth of an RF. At each point that we presented the stimuli, the likelihood was calculated based on Poisson statistics. In parallel, we also fitted the data with a uniform distribution. We compared the goodness of the fit between the Gaussian distribution and the uniform distribution with Bayesian information criterion, and only included neurons that showed a better fitting result with the Gaussian distribution. For the stimuli for which we only tested five locations on the horizontal plane (0, 36, 54, 72, 90°; see above for details), to avoid problems in the fitting process caused by the unequal spacing, we first interpolated the data and estimated the response to a stimulus from 18° azimuth using the response to 0 and 36° azimuths. We then fitted the data with the six azimuths. There is a known defect in our HRTF in the 72° azimuth 0° elevation location, so we compensated the response to stimuli from that direction before doing any fitting analysis. After the fitting process, the neurons with a better Gaussian fitting result were determined, and the mu parameter of the optimal Gaussian fit was used as the azimuth of an RF, while the error was estimated from the Hessian matrix of the likelihood function.

Clustering analysis

Response to pup calls. We first reduced the dimensionality of the spiking response of significantly responsive neurons to pup call stimuli by doing PCA on the peristimulus time histograms (PSTHs). Specifically, for each individual neuron, we calculated the mean firing rate across trials with 1 ms bins from the onset of the stimulus to 40 ms after the offset of the stimulus over all trials (120 repetitions for the pup call stimuli). We also estimated the baseline firing rate by calculating the mean firing rate of each neuron within 400–1,000 ms of each stimulus presentation. We then subtracted the baseline firing rate for all significant neurons and concatenated the resulting response over time to get a population response matrix (number of neurons × number of time bins). After normalization (dividing each neuron's response by its Euclidean norm), we performed PCA on the baseline subtracted and normalized population response matrix. Then we estimated the least number (n) of principal components (PCs) that can explain 90% variance and performed tSNE with the first n PCs. We used k-means clustering on the high-dimensional tSNE space to cluster the response. The selection of the number of clusters and the quality of the clusters were verified by silhouette values/plots and the plot of total within-cluster sum of squares (WSS).

STRFs. We estimated the STRFs of neurons using the STAs calculated from the spiking response to the dynamic random chord stimulus described above, as was done in our previous study (Ito et al., 2020; Si et al., 2022).

To construct the data matrix for clustering, we first determined the significant “pixels” in all STRFs. For each neuron, the STRF consists of a 48 (sound frequencies) × 10 (5 ms time bins) × 2 (contra- vs ipsilateral) matrix, which can be displayed as two heatmaps (contra- and ipsilateral), each consisting of 48 × 10 pixels. Each significant neuron has significant pixels (p < 0.001) and nonsignificant pixels; when examining all significant neurons, we only selected the pixels that were significant in >3 neurons. This led to a total of 153 significant pixels.

We then concatenated each neuron's 153 significant pixels in a single row, to create a 153 by 190 (number of neurons with a significant STRF) matrix as the data matrix for clustering analysis. Then we did PCA on the clustering matrix and used the first 29 PCs that explained 90% variance of the data. We then used k-means clustering methods on the high-dimensional tSNE space of the first 29 PCs, and the quality of the clusters were verified by silhouette values/plots and the plot of total WSS.

To identify the resulting clusters of the STRF clustering, we first calculated the average within-ear cosine similarity index (cSI) of each STRF cluster; the one with the highest SI value was identified as the “spectral cue neurons” cluster, and the one with the lowest SI value was identified as the “ILD neurons” cluster. To distinguish the rest of the clusters, we then calculated the within-cluster cSI to measure how similar the STRF patterns are within each cluster, and the cluster that has a significantly lower cSI than other clusters contains the highest heterogeneity and was therefore identified as the “others” cluster. Finally, between the remaining two clusters, we calculated the average response latency, and the one with a significantly shorter latency was identified as the “fast neurons” cluster, and the other cluster was then identified as the “slow neurons” cluster. We did the same procedure of identifying these STRF clusters in the exploratory dataset and the blinded dataset, and the results turned out to be consistent between the two halves of the full dataset.

Clustering STRF responses

To statistically estimate how well the STRF clusters can be used to distinguish the response of SC neurons to all auditory stimuli tested, we used a bootstrap approach to estimate how well separated the clustered responses grouped by STRF labels are. We first used the five STRF labels (spec. ILD, fast, slow, and others) to group the neurons that are significantly responsive to both random chord and another stimulus (original pup call, pup call stretched 1.25 times, pup call stretched 1.25² times, pup call stretched 1.25³ times, 60 ms chirp that goes from 20 to 80 kHz, 20 ms chirp that goes from 20 to 80 kHz, 10 ms chirp that goes from 20 to 80 kHz, 10 ms chirp that goes from 80 to 20 kHz). For each stimulus, with these defined groups, we calculated the silhouette values (the cosine similarity difference between each neuron and other neurons in the same group vs outside the group) and calculated the average and SEM of such value for neurons that belong to each group and all neurons (Fig. 5C, black points with error bars); then, we randomly assigned the labels to neurons and calculated the average and SEM of silhouette values for each group and all neurons, which we repeated 10,000 times to get a distribution of the silhouette value for each group and all neurons combined (Fig. 5C, colorful distributions). From the relative position of the STRF grouping results to the randomized distribution, we were able to estimate how the STRF cluster labels can significantly separate the response patterns of SC neurons to various auditory stimuli.

Figure 5.

Distribution of spatial RFs of SC neurons responding to chirps. Left column, The PSTH of neurons responding to pup call (A), 60 ms chirp that goes from 20 to 80 kHz (20–80 kHz 60 ms; B), 20 ms chirp that goes from 20 to 80 kHz (20–80 kHz 20 ms; C), 10 ms chirp that goes from 20 to 80 kHz (20–80 kHz 10 ms; D), and 10 ms chirp that goes from 80 to 20 kHz (80–20 kHz 10 ms; E), with the spectrogram of each stimulus in the background. Middle column, The topographic maps of the auditory space from neurons that are significantly responsive to each corresponding stimulus on the left, measured with localized VAS stimuli for each. (Error bars indicate fitting error from the Gaussian distributions; red lines, linear fitting results, slopes, and intercepts of the lines shown in the plot.) Right column, The histogram of spatial RF centers for SC neurons responding to white noise (blue) and each corresponding stimulus on the left (red).

Results

SC neurons respond to pup calls and show several temporal response patterns

To determine whether and how SC neurons respond to ultrasonic pup calls, we presented a pup call stimulus from different azimuthal locations of VAS to awake adult CBA/CaJ mice that were allowed to locomote on a cylindrical treadmill (Fig. 1A,B) and recorded the responses of SC neurons using a four-shank (64 electrodes/shank) silicon probe. We used previously measured HRTFs (see Materials and Methods; Ito et al., 2020) to filter a prerecorded pup call stimulus (Fig. 1B) and presented it with calibrated earphones so that the mouse would hear the sound as if it came from one of 17 azimuths (−144 to 144° with 18° spacing, 50 dB SPL at the tips of the speakers) in the horizontal plane. In recordings from eight mice (P63–P111, four males and four females), we recorded the spiking activity of 1,501 neurons in the dSC (300–1,600 µm from the surface of the SC) where the bulk of the auditory-responsive SC neurons reside (Ito et al., 2020).

We found many neurons with significant responses to the onset and/or offset of the pup call stimulus. Analysis of the overall temporal response pattern reveals that, on average, neurons fire in response to the onset and/or offset of the stimulus within 20 ms. To identify the neurons that are responsive to the pup call stimulus, we compared the sound-evoked firing rate with the baseline firing rate for each neuron at each azimuthal location during the onset (0–20 ms), offset (75–95 ms), and in between (20–75 ms) time windows and identified neurons that significantly responded to the pup call stimulus from at least one spatial location during at least one time window using Poisson statistics (see Materials and Methods). We found (30.2 ± 1.2)% (453) of recorded neurons are pup call responsive, some have only an onset response (Fig. 1D,E, top), some have only an offset response (Fig. 1D,E, middle), and some exhibit both an onset and offset response (Fig. 1D,E, bottom).

To further characterize the response properties of pup call-responsive neurons, we performed clustering analysis on the response patterns of the 453 pup call-responsive neurons in our data set. To do this, we first normalized the PSTH pattern for each pup call-responsive neuron (Fig. 2A); then we concatenated the normalized response of all pup call-responsive neurons to build the response pattern matrix (Fig. 2B) for further clustering analysis using PCA and k-means clustering (see Materials and Methods). This resulted in five clusters in the PC space (Fig. 2C, shown in the first 3 PC axes) and sorted and labeled in the response pattern matrix (Fig. 2D). Within each cluster, the neurons show similar temporal response patterns; therefore, we named these clusters based on the latency of the peak of their response: “no peak” (neurons that do not have an obvious peak), “fast onset” (neurons that have an onset peak ∼10 ms), “midstimulus” (neurons that have a peak in the middle, between onset and offset response), “slow onset” (neurons that have an onset peak 10 ms or slightly later), and “offset” (neurons that have an offset peak ∼85 ms).

Figure 2.

SC neurons respond to pup call stimulus with various temporal patterns. A, B, Construction of the normalized pup call response matrix for all pup call-responsive neurons. Each pup call-responsive neuron's PSTH is normalized (A) and concatenated with all other pup call-responsive neurons to construct a matrix which is then used for clustering analysis. (Neurons are sorted by the center of mass of their PSTH in B). C, Results of clustering analysis shown with the first three PC axes. The clusters are color coded as follows: dark blue, fast onset; green, midstimulus; cyan, slow onset; pink, offset; gray, no peak. D, The normalized pup call response matrix sorted by the results from the clustering analysis, in which the five identified clusters are color coded as in C. E, The same clustering results in C generated by plotting the fraction of spikes within 0–20 ms versus 75–95 ms for each neuron color coded as in C. F, There is no difference between sex in terms of the percentage of neurons in each cluster [p values from Fisher's exact test: (fast onset, slow onset, midstim, offset, no peak) 0.3239, 0.6432, 0.3213, 0.2694, 0.2220].

To verify the temporal properties of these clusters, we then measured the number of spikes for each neuron within the onset (0–20 ms) and offset (75–95 ms) response windows and calculated the fraction of spikes from each neuron in each window (Fig. 2E). Indeed, the fast/slow onset and offset neurons stand out as having a higher fraction of spikes in one of the on/offset windows. Neurons from the same cluster are also localized close to each other in this plot, indicating that the clustering results reflect the similarity in the peak time latencies for the pup call-responsive neurons.

Pup calls are known to trigger maternal behaviors in mice, and previous studies have shown differences in the behavioral response to pup calls between male and female mice (Marlin et al., 2015; Carcea et al., 2021). To determine if the SC neuronal response to pup calls varies between sexes, we compared the fractions of neurons that belong to each of the clusters. We found that these classes of temporal response properties were present in both male and female mice with no significant difference in the percentage of each type (Fig. 2F).

Neurons that are responsive to both white noise and pup call stimuli have a topographic organization with white noise but not pup call stimulation

We next wanted to determine if pup call-responsive neurons respond exclusively to pup calls (Fig. 3A–D). Therefore, we examined the responsiveness of SC neurons to white noise and pup call stimuli (Fig. 3D; see Materials and Methods). We find that (70 ± 1.2)% of all the recorded SC neurons respond to either white noise or pup call stimulus; more than half of these [(57 ± 2)%] respond to only white noise, about a third [(33.4 ± 1.7)%] respond to both the white noise and pup call stimuli, and only a small fraction [(9.8 ± 0.9)%] respond to the pup call but not white noise stimulus.

We then determined the spatial RFs along the azimuthal axis of the neurons that responded to both the white noise and pup call stimuli. To measure the azimuthal RFs of neurons that respond to the white noise or pup call stimulus, we calculated the mean firing rate within the onset (0–20 ms) and offset (75–95 ms) time windows across all trials at each azimuthal location (Fig. 3A–C,c,d). Some neurons have matching RF locations when responding to white noise and pup call stimuli in the first 20 ms (Fig. 3A,c) and in neurons with an offset response (Fig. 3B,d). We also observed that some neurons had RF locations that were shifted along the azimuth when responding to white noise versus the pup call stimuli. For example, neurons like example Neuron 2 (Fig. 3B,c) responds to the onset of the white noise stimulus from a much more limited range of azimuths restricted to the front of the mouse (Fig. 3B,c), but its pup call azimuthal RF is much broader and centered more laterally. Alternatively, neurons like example Neuron 3 respond to the onset of the white noise stimulus from a much broader range of azimuths on the contralateral hemisphere, while the pup call onset response still favors the ∼60° contralateral azimuth (Fig. 3C,c), resulting in a pup call azimuthal RF center shifted more laterally. To quantify how much pup call azimuthal RFs match with the white noise RFs for SC neurons, we calculated the correlation between the firing rate at the 17 azimuthal locations when responding to the white noise versus the pup call stimulus (Pearson's correlation, Fig. 3F shows the probability distribution of the correlation coefficient). We found that the pup call and white noise azimuthal RFs have a much higher correlation in the onset response than in the offset response. Interestingly, we observed two peaks in the probability distribution of the correlation coefficient measured from the correlation between the white noise and pup call azimuthal RFs (Fig. 3F, left, red arrowheads). This indicates the existence of potentially two populations of neurons: a group of neurons with matching white noise and pup call azimuthal RFs (Fig. 3F, left, the right arrowhead, centered ∼0.8) and another group of neurons with shifted white noise and pup call azimuthal RFs (Fig. 3F, left, the right arrowhead centered ∼0.6).

We previously proposed that neurons in the anterior SC have frontal RFs because they respond best to the ∼25–80 kHz range of the monaural spectral cues, while neurons in the posterior SC have lateral RFs because they fire in response to ILDs. The SC neurons that have frontal RFs respond preferably to a particular “within-ear contrast” pattern of spectral cues, which is consistent with the modulation of HRTFs on frontal sounds. Because pup calls do not contain the full 40–60 kHz spectrum (Fig. 1B), we hypothesized that frontal space would not be represented by pup call stimulated neurons' RFs and, therefore, there would not be a topographic map of sound space of pup call responses. To test this hypothesis, we graphed the location of each pup call neuron's spatial RF versus its A–P location in the SC, determined the slope of the best fit line, and compared this to a similar graph generated by the white noise response. Consistent with our previous studies, we find that the spatial RFs of neurons using the white noise stimuli are organized topographically along the A–P axis of the SC (Fig. 4A) with a slope of (60 ± 3.9)°/mm and intercept of (27 ± 2.5)° (Ito et al., 2020, 2021; Si et al., 2022). This does not significantly change if we limit our analysis to the subset of neurons that responded to both white noise and pup call stimuli [Fig. 4B; slope, (61 ± 6)°/mm; intercept, (23 ± 3.7)°]. However, when we plot the azimuthal RF centers against the A–P positions of the same neurons responding to a pup call stimulus, we find that the slope of the line is (13 ± 6.3)°/mm, significantly less than that determined using a white noise stimulus (p = 9 × 10⁻¹¹; ANOVA; Fig. 4C). Instead of having a distribution of RFs from the most frontal to the back of the azimuthal axis as the white noise topographic map, we observed that a majority of the pup call RFs are centered at ∼60° azimuth (Fig. 4D) with very few monitoring sound coming from the front [(19 ± 4)% compared with (33 ± 5)% for white noise] or the more lateral space [(17 ± 4)% compared with (32 ± 5)% for white noise; Fig. 3E].

Figure 4.

Lack of topography in the organization of pup call-responsive neurons in the SC. A, The topographic map of the auditory space from neurons that are significantly responsive to white noise stimuli, measured with localized white noise stimuli. B, The topographic map of the auditory space from neurons that are significantly responsive to both white noise and pup call stimuli, measured with localized white noise stimuli. C, The spatial RF centers measured in response to localized pup call stimuli graphed as a function of the A–P SC positions of neurons significantly responsive to both white noise and pup call stimuli. D, The center of spatially restricted RFs of neurons that are significantly responsive to both localized white noise and pup call stimuli, tested with localized white noise and pup call stimuli. [Error bars indicate fitting error from the Kent (A) or Gaussian (C, D) distributions; red lines in A–C, linear fitting results, slopes, and intercepts of the lines shown in the plot; the dashed line in D marks when the centers of RFs overlap perfectly.] E, The speaker placed at 0 and 60° and either white noise (blue) or pup call (orange) is played from each speaker. F, The plot showing the ratio of the difference in the number of action potentials elicited from the 60 and 0° speaker presentations for each RF neuron using an index from −1 (responding only to the 0° presentation) to +1 difference responses at 0 and 60° for LS white noise (blue) and LS pup call (orange). The same neurons of each dataset are connected by a line, signif. ***; p < 0.0001; paired t test; df = 64. G, Normalized difference mapped to anatomical A–P SC location of each neuron for LS white noise (blue) and LS pup call (orange). H, Magnitude of change between the two normalized fractions along the A–P location of the SC. I, Seventy-eight neurons that have localized white noise (blue) and pup call (orange) responses used in the SVC encoding model sorted by peak azimuth RF location (frontal to lateral) showing that responses are generated from stimuli presented at all contralateral locations. The same neurons of each dataset are connected by a line. top right: Distribution of the location of max activity showed similar distributions (right, top; KS test, p = 0.55). bottom right: Scatterplot for the S/N ratio shows decrease in S/N in pup call responses when compared with white noise (paired t test; **p < 0.005). J, The confusion matrix shown as dot plot of SVC model prediction of sound location with respect to the actual location of the stimulus for the population encoder for each single trial for white noise (blue) and pup call (orange) stimuli. The shading of the dot represents the frequency of outcome of each model. K, Cumulative distribution of the absolute error of the difference between actual and predicted sound locations across the azimuth. Mean (vertical) and SD (horizontal) are denoted at the top of the graph. White noise (blue), pup call (orange), and random (dashed black) were compared using KS test with Bonferroni’s multitest comparison adjustments. White noise to random (p = 4.0 × 10⁻¹⁸), pup call to random (p = 1.9 × 10⁻⁵), and white noise to pup call (p = 0.0055).

To verify that the preference to ∼60° azimuth pup calls is not due to any possible artifacts from our VAS stimulation system, we tested if the spatial RFs computed from the white noise response shifted toward 60° when computed from the pup call response using free-field loudspeakers. We placed one speaker in front of the animal and another at 60° and recorded from SC neurons using white noise and pup call stimuli (Fig. 4E; see Materials and Methods). We then determined the number of spikes elicited by sound coming from each speaker for each neuron in response to white noise and pup call stimulus. We then calculated the difference in the number of spikes elicited between the 60 and 0° presentation for each neuron using an index from −1 (responding only to the 0° presentation) to +1 (responding to only the 60° presentation) for 65 neurons (Fig. 4F). We find that the distribution of this index generated from the white noise stimulus shifts positively when calculated from the responses to the pup call stimulus. This shift is true across the A–P axis in the dSC (Fig. 4G,H). This result is consistent with our VAS findings in Figure 4C.

Does this lack of topography in the SC necessarily mean that a mouse would have deficits in interpreting the location of a pup call? It is possible that despite the organization of the spatial RFs into a map, there is still plenty of information for the downstream networks to infer a pup call location using the population response. To test this hypothesis, we trained a SVC model by giving it the responses of 78 neurons from 80 trials across 17 azimuth locations for both white noise and pup call stimuli (Fig. 4I–K; see Materials and Methods) and asked the model to predict the location from untrained trials. We first assessed the population of neurons and found that they had similar peak RF locations showing that the coverage of auditory space between the pup call and white noise are similar. However, we note a decrease in the signal-to-noise (S/N) ratio of pup call responses (0.51 ± 0.40) when compared with white noise (0.62 ± 0.38; Fig. 4I). We then determined the prediction error, the absolute value of the difference between the actual and the predicted azimuthal position, for 680 (40 trials across 17 azimuths) trials. Interestingly, we find the population response can indeed localize a pup call location much better than chance (Fig. 4J, right; mean error 42.8 ± 44.9° vs chance 87.1 ± 52.9°, mean ± SD, two-sample KS test p = 1.9 × 10⁻⁵). However, we also find that the population response from a white noise trial is even more accurate (22.8 ± 31.4°; Fig. 4J, left; KS test white noise to random p = 4.0 × 10⁻¹⁸; KS test white noise to pup call p = 0.0055; Fig. 4K). These results show that the loss of topographic organization of responses in the SC does not prohibit the population from estimating the location of a sound in a single trial; however, the accuracy of this population estimate is improved when a topographical response is present.

The SC response to frequency-restricted chirps is not organized topographically

To better evaluate if frequency-restricted stimuli lead to shift in the spatial RF responses of neurons thus leading to a loss of topography in the SC, we recorded SC neurons when responding to four types of chirps with frequencies covering 20–80 kHz with different frequency/time derivatives (Fig. 5) at five azimuthal locations (0, 36, 54, 72, 90°) using the VAS system (50 dB SPL). First, we looked at the organization of the spatial RFs along the A–P axis of the SC in response to a chirp that starts at 20 kHz and progresses linearly to 80 kHz in 60 ms (Fig. 5B). We found that the topographic organization of the spatial RFs is significantly different from that found from the white noise response, with a slope of 23 ± 1°/mm (compared with 60 ± 3.9 for white noise; p = 0.0015; ANOVA) and an intercept of 47 ± 6.6°/mm (compared with 27 ± 2.5 for white noise; p = 0.0046; ANOVA). We then varied the derivative of the chirps by shortening the duration of the stimulus. Interestingly, when the chirps goes from 20 to 80 kHz within 20 ms (Fig. 5C) or within 10 ms (Fig. 5D), the resulting topographic maps have statistically the same slopes and intercepts to the white noise map (20 ms, slope (60 ± 3.9)°/mm; intercept (30 ± 7.6)°; p = 0.71; ANOVA; 10 ms, slope [57 ± 8.4)°/mm; p = 0.75; intercept (35 ± 5.2)°; p = 0.17; ANOVA] but with again a concentrated RF centers at 80° which is significantly different from the distribution for RF centers of the white noise response (20 ms, p = 1.6 × 10⁻⁷; 10 ms, p = 4.9 × 10⁻⁵; KS test). However, when we played the 10 ms chirp backward (going from 80 to 20 kHz in 10 ms), we found the topographic map [slope (46 ± 8.8)°/mm; p = 0.15; intercept (43 ± 5.5)°; p = 0.0081] missing more frontal RFs for the anterior neurons, with more neurons with a RF centered ∼85° (significantly different from the white noise distribution; p = 1.1 × 10⁻⁷; KS test; Fig. 5E). Therefore, we observe that whether the SC neurons topographically respond to spatially restricted nonbroadband sounds depends on the spectrotemporal patterns of the stimuli.

The spectrotemporal properties of the SC neurons exhibit heterogeneity within a continuum

We have demonstrated that some neurons exhibit different temporal or spatial responses depending on the frequency composition of the stimulus (white noise vs pup call vs chirps; Figs. 2–4). To determine if we can classify these neurons based on their spectral/temporal response, we used random chord stimuli to determine each neuron's STRF (see Materials and Methods). To do so, we recorded the spiking response of neurons while the animals were presented with dynamic random chord stimuli independently to each ear using frequencies ranging from 5 to 80 kHz with each random chord lasting 5 ms (see Materials and Methods). We then determined the STAs from stimuli presented at each ear. By comparing the STAs with the average probability of the presence of each given tone, we determined which pixel(s) in the 48 (number of tones) by 10 (number of time bins) by 2 (contra- or ipsilateral sides) STA images are significantly more positive or negative than chance alone. Consistent with our previous findings (Ito et al., 2020; Si et al., 2022), we find the majority of both positive (Fig. 6A, left) and negative (Fig. 6A, right) significant pixels are >20 kHz. There is also a contralateral bias for both the positive and negative significant pixels (Fig. 6A). To understand the heterogeneity of the spectrotemporal properties of SC neurons, we then performed cluster analysis on the STRF patterns. We first created a mask to only select the most significant STRF pixels to improve the S/N ratio. With the threshold of including only pixels with >3 neurons being significant, 153 pixels for each neuron were used for further analysis. We then constructed a matrix with the153 pixels from all neurons that have a significant STRF and performed dimensionality reduction using PCA (Fig. 6B). After visualizing these neurons in the first two PC space, we failed to see distinct clusters but rather we see a continuum. We therefore hypothesize that the STRFs of SC neurons fall onto a spectrum, where they display gradual rather than distinct changes of spectrotemporal properties.

Figure 6.

Clustering analysis of STRF patterns reveals at least four classes of auditory SC neurons. A–C. Clustering analysis of STRFs of SC neurons. A, Heatmap showing the probability of SC neurons showing significant positive (left) or negative (right) response at each spectrotemporal pixel in their STRFs. The red line marks the border between contralateral and ipsilateral sides. B, An example STRF before and after the application of the mask, leaving only the most significant pixels in the STRF for further processing. C, Results of clustering analysis shown with the first two PCs. D, STRFs of two example neurons from the clusters of spectral cue neurons (spec.) and ILD neurons (ILD). E, RF azimuth and AP position of neurons from the spec. and ILD clusters relative to all neurons in the topographic map of the auditory space. Red asterisks, neurons in the spec. (left) and ILD (right) clusters that also significantly respond to localized white noise stimuli; blue dots, error bars, and red lines, all neurons that significantly respond to localized white noise stimuli with errors from the Kent distribution fitting and linear fitting results as in Figure 4A. F, Average RF azimuth and average SI for neurons from each group of neurons in the STRF continuum (gray points are all neurons that have significant STRF and localized RFs; error bars indicate SEM of neurons from each group; color code as in C). G, STRFs of two example neurons from the clusters of fast neurons (fast) and slow neurons (slow). H, Average peak time in the STRF for positive or negative response on the contralateral or the ipsilateral side for the four STRF clusters: spec., ILD, fast, and slow. I, The distributions of pairwise SI values for each possible pair of neurons within the four STRF clusters. Black dots and error bars, mean ± SEM. spec., the spectral cue neurons from the STRF clustering analysis; ILD, neurons from the ILD STRF group; fast, neurons from the fast STRF group; slow, neurons from the slow STRF group.

To understand the features of the distribution of the spectrotemporal properties of the SC neurons, we used k-means clustering analysis to reveal the grouping of similarities and extremes of features within this continuum. Using the elbow method (Thorndike, 1953; Syakur et al., 2018), when calculating the inertia (or WCSS), we observed the rate of decrease in WCSS begins to slow down significantly around k = 5. Therefore, we clustered the neurons into five groups. When these five groups of neurons are visualized in the first two PC space, we find that four of them occupy distinct ends in this continuous distribution (Fig. 6C): on one end are neurons that have highly similar contra-/ipsilateral STRF patterns with prominent “within-ear contrast” (“spectral cue neurons,” example STRF shown in Fig. 6D, left), whereas on the opposite end are neurons that have anticorrelated contra-/ipsilateral STRF patterns (“ILD neurons,” example STRF shown in Fig. 6D, right). In between these neurons and stretching out in orthogonal directions are two other groups of neurons: neurons that have transient response to sound (“fast neurons,” example STRF shown in Fig. 6G, left) and neurons that have relatively prolonged response to sound (“slow neurons,” example STRF shown in Fig. 6G, right). We calculated the pairwise SI (Fig. 6I) to measure the similarity of STRFs from the same group of neurons; we found that these four groups of neurons are relatively similar to each other within groups (Fig. 6I; average pairwise SI, spec, 0.27 ± 0.17; ILD, 0.19 ± 0.14; fast, 0.31 ± 0.16; slow, 0.32 ± 0.17). However, the fifth group of neurons from the k-means analysis has a significantly lower similarity between neurons (Fig. 6I; average pairwise SI, others, 0.03 ± 0.10). This indicates that these neurons have a mixture of different STRF patterns; therefore, we have excluded this class in further analysis.

The classes of STRF responses within the continuum predict SC A–P location

Our overall hypothesis is that the auditory neurons with spatially restricted RFs in the SC use a weighted combination of spectral cues and ILDs to compute the location of broadband noise stimuli, with the neurons in the anterior dSC using predominantly “within-ear contrast” of the input spectrum which is used to provide an excitation to sound from the frontal field and the posterior neurons using “across-ear contrast” (ILD) in a broad frequency range that provides the excitation to sound from the lateral field (Ito et al., 2020). With the discovery of the four neuron groups within the STRF continuum, we hypothesize that they reflect the nature of a gradual change in the use of spectral cues and ILDs for the SC neurons along the A–P axis. To test this hypothesis, we first determined the physical location of each group of neurons in the SC. We find that the “spectral cue neurons” are predominantly located in the anterior SC and have frontal spatial RFs (Fig. 6E, left), while the “ILD neurons” are predominantly located in the posterior SC and have lateral spatial RFs (Fig. 6E, right). Consistent with their location within the STRF continuum, the “fast” and “slow” neurons were distributed throughout the A–P axis of the SC and have RF azimuth RFs that lie between the spectral cue and ILD classes. To further validate this point, we examined all neurons that have both a significant spatially restricted RF and a significant STRF and calculated the binaural cSI for each neuron (BSI; see Materials and Methods), which reflects the similarity between the contralateral and ipsilateral STRF, with the assumption that the more similar the two responses, the more likely spectral cues are used to determine the response, while the more contrast between the two responses, the more likely ILDs are used to create the response. Consistent with our previous work (Ito et al., 2020), the RF azimuth and BSI are anticorrelated (Fig. 6F). We find that the average value for RF azimuth and BSI of “spectral cue” and “ILD” neurons are well separated in the RF azimuth over the BSI plot (Fig. 6F, green and purple data points and error bars) and that neurons from the “fast” or “slow” groups have BSI and RF azimuth values in between the spectral cue and ILD classes. Consistent with our hypothesis, we therefore observe a gradual change in the use of ILDs and spectral cues for neurons with RFs along azimuths.

To understand the other dimension of the continuum that separates the “fast” and “slow” neurons, we characterized the temporal properties of the STRFs. Indeed, neurons from the fast group stand out as having the shortest latency to peak time in their STRFs (Fig. 6H). We measured each neuron's latency to peak time for both the contralateral and ipsilateral sides of their STRFs by taking the average of both the positive and negative responses across frequencies and determining the peak time for each averaged response. We found that neurons from the fast class are distinguished by their significantly faster peak time in both positive and negative responses on the contralateral side and the negative response on the ipsilateral side (Fig. 6H, yellow data points and error bars).

Stimuli with strong spectrotemporal correlations expose more heterogeneity in the responses of SC neurons than white noise

With the heterogeneity in the STRFs of the SC neurons, we predict that the temporal response for neurons within different groups of the STRF continuum will vary when they respond to other stimuli with strong spectrotemporal correlations. Therefore, we characterized the heterogeneity in the temporal response of the SC neurons when responding to white noise and nonbroadband stimuli such as a pup call (Fig. 5A, left, spectrogram) and chirps (20–80 kHz) of different frequency/time slopes (Fig. 5B–E, left, spectrograms). The left column of Figure 5 shows the average PSTH of all neurons (blue traces), and Figure 7A organizes the data such that the neurons in each group defined by their random chord response (spectral cue, ILD, fast, slow) and that also responded to the other stimuli are grouped together (white lines). Inspection of the data shows that while virtually all neurons fire in response to the onset of the stimulus, very few neurons fire in response to the offset of a stimulus, except for neurons in the ILD and to some extent the fast group (Fig. 7A,C). For example, in response to the white noise stimulus (Fig. 7Aa–Da), all neurons have a fast-onset response, which is similar to neurons in the four STRF groups but have very little response to the offset of the stimulus. After presenting a pup call stimulus, we see the fast-onset response that is present for most neurons from the four STRF classes; however, neurons from the ILD group have more of an offset response than the other classes (Fig. 7A,b; second block from the left in the heatmap). To quantify this, we calculated the following for each STRF class: (1) the percentage of neurons that significantly respond to the offset of the stimulus (Fig. 7C,a–f) and (2) for each neuron within each STRF class, the offset response amplitude measured by the z-score within 20 ms after the offset of the stimulus (Fig. 7D,a–f). We find that there is a significantly higher percentage of ILD neurons that respond to the offset of frequency-restricted stimuli (but not white noise) and that the ILD neurons have a higher response amplitude when responding to the offset of these stimuli. Specifically, neurons from the ILD group have more offset responses in the pup call and three of the four chirps (the exception being a chirp from 80 to 20 kHz with a duration of 10 ms). Indeed, the percentage of offset responsive ILD neurons and their response amplitude are significantly higher than average (Fig. 7C,D,b, p < 1 × 10⁻⁵ for both; Fig. 7C,c,d, p < 1 × 10⁻⁵; e, p = 0.0051; Fig. 7D,c,d, p < 1 × 10⁻⁵; e, p = 0.0004; ANOVA). In addition, we consistently found more offset responsive neurons in the “fast” group. Finally, we see a strikingly low percentage of offset responsive spectral cue neurons (Fig. 7C,c–e, p < 1 × 10⁻⁵ for all, ANOVA), and their offset response amplitude is also lower than average (Fig. 7D,c, p < 1 × 10⁻⁵; d, p = 0.0009; e, p = 0.0001; ANOVA) for all stimuli tested. Indeed, when we mark neurons from the four STRF groups in the first two PC space of white noise and nonbroadband stimuli response, the neurons from the four STRF classes are intermingled in the white noise PC space (Fig. 7B,a) but are more separated in the nonbroadband stimuli space (Fig. 7B,b–f). This suggests that there is more heterogeneity in the temporal response of SC neurons to stimuli with strong spectrotemporal correlations, but not white noise, and those neurons that are closer in the STRF continuum tend to have similar temporal response when presented with these stimuli.

Figure 7.

The temporal response patterns of SC neurons to white noise, pup call, and chirp stimuli separated by the STRF labels. A, The PSTH of neurons responding to white noise (a), pup call (b), 60 ms chirp that goes from 20 to 80 kHz (20–80 kHz 60 ms, c), 20 ms chirp that goes from 20 to 80 kHz (20–80 kHz 20 ms, d), 10 ms chirp that goes from 20 to 80 kHz (20–80 kHz 10 ms, e), 10 ms chirp that goes from 80 to 20 kHz (80–20 kHz 10 ms, f) with the spectrogram of each stimulus in the background. B, Temporal response of neurons in the four STRF groups (spec., ILD, fast, slow, separated by the white lines) to white noise (a), pup call (b), 60 ms chirp that goes from 20 to 80 kHz (20–80 kHz 60 ms, c), 20 ms chirp that goes from 20 to 80 kHz (20–80 kHz 20 ms, d), 10 ms chirp that goes from 20 to 80 kHz (20–80 kHz 10 ms, e), 10 ms chirp that goes from 80 to 20 kHz (80–20 kHz 10 ms, f; y axis, time/ms). B. Neurons from the four STRF groups (spec., ILD, fast, slow) highlighted on the first two PC axes of all neurons significantly responsive to both dynamic random chord and at least one of the tested variations of pup call and chirp stimuli. (Black dots, all significant neurons; color code as in Fig. 6.) C, The percentage of neurons in the four STRF neuron groups that exhibit significant offset response. (y axis, percentage/%; horizontal line, average percentage of neurons that show offset response out of all neurons that are significantly responsive to each stimulus tested.) D, The offset response amplitude shown in the Z-score for neurons in the four STRF neuron groups. (y axis, Z-score in 20 ms bin; horizontal line, average offset response amplitude of neurons that are significantly responsive to each stimulus tested; black points and error bars, mean ± SEM of each cluster; spec., the spectral cue neurons from the STRF clustering analysis; ILD, neurons from the ILD STRF group; fast, neurons from the fast STRF group; slow, neurons from the slow STRF group.)

Discussion

In this study, we show that mouse SC auditory neurons respond to pup calls with distinct temporal patterns and a spatial preference predominantly at ∼60° in contralateral azimuth, leading to a lack of topographically organized response to pup calls. Despite this lack of organization, the population response to pup calls can predict the location of the sound source better than chance, but not as well as the topographically organized response to white noise stimulus. We also demonstrate that the spectrotemporal properties of the SC neurons exhibit heterogeneity within a continuum which can be roughly divided into at least four distinct classes that predict their location along the A–P axis of the SC.

Previous studies have shown that response to pup calls in the primary auditory cortex is enhanced only in female mice with maternal experience (Marlin et al., 2015; Carcea et al., 2021). However, we did not find differences between sexes in the temporal response patterns of SC neurons that responded to pup calls. This suggests that pup calls are processed rather similarly in the SC auditory neurons of virgin female and male mice, consistent with the role of the SC in determining the location but not the identity of an auditory stimulus. One limitation of our experiments is that we only used virgin mice in this study. Given the evidence that mother mice or experienced female virgin mice exhibit different interactive behavior with pups (Carcea et al., 2021), it will be interesting to examine the differences between neuronal responses to pup calls in the SC of experienced mothers and virgin females. Another limitation of this finding is that we only compared the differences between sexes in terms of the temporal response patterns of SC neurons. It is possible that differences exist in other aspects that we did not characterize.

We did not observe a topographic map of the auditory space measured with pup calls in the SC, even though the same population of neurons produced a topographic map when measured with the white noise stimulus. We showed that neurons with white noise spatial RFs centered at the most frontal or lateral auditory space have mismatching pup call spatial RFs: the frontal white noise RFs shift laterally when the neurons respond to pup calls, while the lateral white noise RFs shift frontally. Indeed, the probability distribution of the correlation coefficient between the white noise and pup call spatial RFs captures two neuron populations: one with more correlated white noise and pup call RF azimuths and the other with less correlated RFs (Fig. 3F, left). We find that SC auditory neurons respond to pup calls with a spatial preference predominately at ∼60° in contralateral azimuth (Fig. 4B,C), which is close to the direction where ILD is maximized (Ito et al., 2020). This suggests that mouse SC neurons mainly use ILDs to localize pup calls.

Topographic maps are used to organize spatial sensory information in many areas of the brain, but theoretically they are not essential for neural circuitry to be able to localize objects (Kaas, 1997; Weinberg, 1997; Avitan et al., 2016). To determine if pup call stimuli is sufficient to localize sound without having topography, we trained a population coding model to predict the location of the stimulus given the population response to a single trial. We found that the population response can localize a pup call location much better than chance (Fig. 4I–K). This is consistent with behavioral evidence that mother mice can localize the sound source of pup calls (Ehret, 1987; Ehret, 2005). We also found that the population response from a white noise trial is even more accurate. These results suggest that although topography may not be needed for an animal to localize sound, it may enhance this ability. This would be consistent with the finding that topographic maps are found in virtually all stages of sensory processing (Luo, 2021).

We have previously shown that high-frequency spectral cues are necessary for the generation of frontal RFs when SC auditory neurons respond to white noise stimuli (Ito et al., 2020; Si et al., 2022). The pup call stimulus we used in this present study, which is also representative for pup calls in general, is a high-frequency stimulus but does not consist of the complete high-frequency range (40–80 kHz) needed to generate within-ear contrast that is required to generate frontal RFs. The lack of frontal RFs for pup call-responsive neurons indicates that the mere presence of a range of high-frequency sound is not sufficient for SC neurons to have RFs that are organized as a topographic map of the auditory space. We speculate that the restricted frequency range in a pup call makes it difficult for the brain to analyze the spectra and separate the modulation by HRTF from the spectra structure of the sound source. Consistent with this, we find similar results when presenting frequency-restricted chirps (Fig. 5). Therefore, when spectral cues cannot provide accurate information about the sound source location, the SC neurons may only have ILD information available to generate their spatial RF.

Our clustering analysis of auditory SC neurons based on their binaural STRFs did not reveal distinct clusters of response types but rather a continuum spanning the STRF space. However, neurons that reside in the four different tails of the continuum have distinct properties and can predict the location of the neuron along the A–P axis of the SC. Along one end of this continuum are the spectral cue neurons that have the most symmetric between ear STRFs, have frontal RFs, and are localized to the anterior SC. ILD neurons that have the most across-ear contrast in their STRFs and lateral RFs are on the other end of this continuum and reside in the posterior SC. In addition, two other subtypes of neurons can be detected: the “fast” neurons that are transiently activated by auditory stimulation, and the “slow” neurons that can be activated within a prolonged time window on the contralateral side. Although much work has been done to characterize subtypes of visually responsive neurons in the SC morphologically, genetically, and physiologically (Gale and Murphy, 2014, 2016, 2018; Li and Meister, 2023; Liu et al., 2023), this is the first effort to characterize the subtypes of in vivo auditory response of the mouse SC neurons.

We observed an offset response in SC auditory neurons when they respond to frequency-restricted stimuli such as the pup call and chirp stimuli. This is the first time that offset response to sound is reported in the SC. However, offset responses to auditory stimuli have been reported throughout the auditory brain across species (Kopp-Scheinpflug et al., 2018), most peripherally from the cochlear nucleus (Suga, 1964; Young and Brownell, 1976; Ding et al., 1999) to the superior olivary complex (Dehmel et al., 2002; Kulesza et al., 2003), the inferior colliculus (IC; Casseday et al., 1994; Xie et al., 2007; Kasai et al., 2012; Akimov et al., 2017), the medial geniculate body of the thalamus (He, 2002; Anderson and Linden, 2016), to the auditory cortex (Recanzone, 2000; Keller et al., 2018). Offset auditory response is important for duration discrimination (Casseday et al., 1994) and gap detection (Xu et al., 2014) in a more energy efficient way than encoding the entire duration of a sound. We do not know the source of the offset response in the SC, but because the SC receives direct inputs from the IC and auditory cortex (Edwards et al., 1979; Druga and Syka, 1984; Covey et al., 1987; Jiang et al., 1997; Doubell et al., 2000; Bednárová et al., 2018; Benavidez et al., 2021; Issa et al., 2023), it is possible that the offset response to pup calls and chirps in the SC are inherited from the IC, within the SC or from the auditory cortex. Intriguingly, we also found that ILD neurons, but not spectral cue neurons, have an offset response. This observation might indicate that the auditory neuron subtypes get inputs from distinct brain regions. In addition, we also observed spatial tuning of the offset response in SC neurons, both to white noise, pup calls, and chirps. Intriguingly, the spatial tuning of the offset response to white noise and pup calls is much less correlated than that of the onset response of SC neurons (Fig. 3F). This may imply that the onset and offset responses in the SC are inherited from different upstream sources or that the spatial auditory RFs are refined by the offset of the stimulus, which can be an interesting future direction of research.