Figures and data
![](https://prod--epp.elifesciences.org/iiif/2/98415%2Fv1%2Fcontent%2F537782v2_fig1.tif/full/max/0/default.jpg)
Experimental setup, auditory stimuli, recording strategy, and stability assessment.
A. Illustration of experimental setup. Single-unit and multi-unit recordings were obtained from the auditory cortex of awake mice using a chronically implanted 8-tetrode array. Mice were head-fixed but were able to run on a rotating cylinder. Simultaneous neuronal recordings and measurements of running speed and pupil diameter were obtained during repeated presentations of noise bursts, tone pips, and dynamic random chord (DRC) stimuli. Responses to noise bursts and tone pips were used to identify core auditory cortical areas. Responses to DRC stimuli were used to estimate contextual gain fields (CGFs) and principal receptive fields (PRFs) using the context model. B. Schematic illustrations of noise bursts (top), tones (middle), and DRC excerpt (bottom). A full DRC stimulus lasted 675 s, and consisted of 15 continuous repetitions of a 45-s-long sequence of 20-ms random chords. C. Schematic representation of the experimental design. Recordings were obtained from the same site for multiple days before tetrodes were advanced to sample new sites. Note the repetition of the full sequence of stimulus presentations (1 segment) within each session. D. Conceptual illustration of methodology used for assessing stability of the context model fits. On each day of recording, the full DRC stimulus was played twice, once in each segment (upper left). CGF or PRF estimates from different days and/or different segments (fieldj,k, where j=day and k=segment) were compared both within and between sessions to obtain a similarity matrix (right). Within-session similarities are on the diagonal (in green and yellow) and the average estimates of the across-session similarities (lower left) are on the off-diagonals (in brown).
![](https://prod--epp.elifesciences.org/iiif/2/98415%2Fv1%2Fcontent%2F537782v2_fig2.tif/full/max/0/default.jpg)
Spike waveforms matched across multiple days using pairwise waveform distances.
See text for explanation of waveform distance metrics d1 and d2. A. Null distribution. Scatterplot shows (d2, d1) spike waveform distances for pairwise comparisons (n=6574) between spike waveforms for unit recordings known to be non-matched (obtained using the same tetrode but from sites located at least 250 microns apart). The ellipse represents the 99% confidence interval (CI) for the null distribution, estimated by fitting a 2D Gaussian to the data. Marginal distributions were obtained using kernel density estimation. B. Experimental distribution. Scatterplot shows (d2, d1) spike waveform distances for pairwise comparisons (n=5594) between spike waveforms for unit recordings obtained using the same tetrode on different recording days at the same recording site. A Gaussian mixture model was fitted to the experimental data using the Expectation-Maximization (EM) algorithm with two clusters. One of the clusters was fixed to the null distribution estimated in A. Ellipses show the 99% CIs for the null (blue) and the experimental (red) distributions. We conservatively defined a waveform pair to be “matched” (i.e., likely to be coming from the same unit) if the waveform distance fell within the experimental but outside the null 99% CI. Colored dots correspond to the example matched and non-matched waveform pairs shown in C. C. Examples of spike waveform pairs. The pairs in the first two columns were identified as matches, whereas those in the latter two columns were not. D. Number of matches as a function of the temporal separation between the two recordings. Dotted gray line shows percentage of total comparisons which were matches. Note that the number of waveform pairs identified as matches was highest for recordings occurring 1–4 days apart, but this was primarily because the number of pairwise waveform comparisons was highest for recordings occurring a small number of days apart. The percentage of waveform comparisons producing a match could be just as high for recordings made weeks apart as days apart, indicating that prolonged tetrode recordings from the same site could be stable.
![](https://prod--epp.elifesciences.org/iiif/2/98415%2Fv1%2Fcontent%2F537782v2_fig3.tif/full/max/0/default.jpg)
Neuronal responses to the DRC stimulus used to estimate PRF and CGF structure in awake mice.
A. Signal power normalized by noise power (SNR) for neuronal responses to the DRC stimulus, for all units that qualified for further analysis given our selection criteria (see text). Units are sorted in order of ascending SNR. Single units are shown in red, multi-units in blue. B. Spectrographic reresentation of the final 1.5 s of the 45-s-long DRC stimulus used. Each shaded rectangle represents a 20-ms tone pulse, with darker shades corresponding to louder tone pips (see colorbar). C-F. Trial-by-trial spike rasters (top) and histograms of spiking rate (bottom), describing the responses of four example units to the DRC excerpt in B. Histogram bins are aligned with the 20-ms chords of the DRC. Units were taken from a point in the distribution in A indicated by the arrows. Time is shown relative to the beginning of the stimulus for the trial. G. Example PRFs (left) and CGFs (right) for three different units (each row is one unit). Yellow and cyan areas in the PRFs represent excitatory and inhibitory regions of the time-frequency receptive field respectively. In the CGFs, axes are time offset and frequency offset relative to a “target” tone represented by the notch at τ=0 and ϕ=0, which can be any tonal element in the DRC stimulus. Red and blue areas in the CGF indicate amplifying or dampening effects (respectively) of acoustic energy at that relative position on the gain of the neuron’s response to a target tone. In other words, the CGF depicts modulation of neuronal responsiveness by sound combinations, as a function of time and frequency differences between the tonal elements in the combinations. H. Average CGF across all units and animals (center). For units recorded across multiple days, we included in this average a single CGF estimated from all the available data for the unit. Line plots along margins show: (left) gain profile as a function of frequency offset between tone pips, averaged across time offsets; (bottom) gain profile as a function of time offset between tone pips, averaged across frequency offsets; and (right) gain profile as a function of frequency offset for the 0–20-ms time-bin alone (i.e., for near-simultaneous tone pips). Error bars indicate standard error of the estimated population means.
![](https://prod--epp.elifesciences.org/iiif/2/98415%2Fv1%2Fcontent%2F537782v2_fig4.tif/full/max/0/default.jpg)
Examples of quantification of PRF and CGF stability across recording days.
A-C. Example PRF (top row) and CGF (bottom row) pairs for neurons matched across recording sessions. The within-session repetition of the 675-s-long DRC run allowed us to estimate two PRFs and two CGFs for each session. For each example, PRFs are identically scaled to the maximum change in firing rate shown in the PRF colorbar. CGF weights at each value of (ϕ, τ) represent the change in gain induced in the response to a sound at the (0,0) notch point if one of the loudest tones of the DRC were to fall at the corresponding (ϕ, τ) location (colors correspond to gain change shown on the CGF colorbar). Like PRFs, CGFs are identically scaled within and across sessions for each example. Time runs from left to right and is in recording sessions conducted on separate but not necessarily consecutive days; numerals across the top of panel A indicate number of recording sessions following the initial session. Note the remarkable consistency of both CGF and PRF structure, which is nearly as high across days as within sessions. D-F. Heatmaps showing the normalized dot product (i.e., field correlation) between the PRFs (orange) or between the CGFs (green) shown in A-C, respectively. Diagonals indicate the within-session comparisons, off-diagonals the across-session (i.e., across-day) comparisons. Higher values indicate higher correlation in structure. The correlation color scale was set to 0.70–1.00 (rather than 0.00–1.00) to maximize visibility of small differences in the generally high correlation values.
![](https://prod--epp.elifesciences.org/iiif/2/98415%2Fv1%2Fcontent%2F537782v2_fig5.tif/full/max/0/default.jpg)
Population data on stability of PRFs and CGFs: normalized field alignment indices.
A-B. Stability of PRFs (A) and CGFs (B) quantified using a normalized field alignment index, where 1.0 indicates similarity equivalent to the field correlation observed for within-session comparisons for each unit, and 0.0 indicates baseline field correlation expected for comparisons between PRFs or CGFs obtained from different units (see text for details). Data points on day 0 represent the within-session comparison; subsequent points represent comparisons across different numbers of days. Each colored line represents a unit; the solid black line is the median across units. Insets show zoomed-in views of the bulk of the data, between days 0 and 5. Normalized field alignment remained close to 1.0 across sessions for most PRFs and CGFs, indicating that neuron-specific PRF and CGF structure was preserved for many days in most neurons.
![](https://prod--epp.elifesciences.org/iiif/2/98415%2Fv1%2Fcontent%2F537782v2_fig6.tif/full/max/0/default.jpg)
Population data on stability of PRFs and CGFs: raw correlation values.
A-B. Stability of PRFs (A) and CGFs (B) quantified using raw correlation values, where 1.0 indicates perfect alignment of fields estimated from recordings made on two different days (see Figure 4 for examples). As in Figure 5, data points on day 0 represent within-session comparisons; subsequent points represent comparisons across different numbers of days. Each colored line represents a unit, and lines are transparent so that shading darkens as multiple lines superimpose. Note that most units display high PRF (A) or CGF (B) correlation values that are stable across days or weeks. C-D. Lines of best fit to the within-session and across-session field correlation values for each unit, for PRFs (A) and CGFs (B). Each best-fit line was estimated using weighted regression, taking into account the number of within-session (0 days between recordings) and across-session (n days between recordings) comparisons available for the unit. E-F. Slope (x-axis) for each colored line in A or B respectively; units (y-axis) are ordered by increasing slope. Error bars indicate ±1 standard error of the estimated slope. For both PRFs (C) and CGFs (D), the slopes of the best-fit lines were often statistically indistinguishable from zero and rarely more negative than -0.2 (the value corresponding to loss of field correlation across 5 days).
![](https://prod--epp.elifesciences.org/iiif/2/98415%2Fv1%2Fcontent%2F537782v2_fig7.tif/full/max/0/default.jpg)
Further analysis of PRF and CGF stability.
A. Histogram of unit-by-unit differences between the PRF and CGF slope estimates from the correlation-based stability analysis shown in Figure 6. Note clustering of values near zero (dotted line). B-D. Slopes of best-fit lines from correlation-based stability analysis for PRFs (orange) and CGFs (green), plotted versus: mean firing rate evoked by the DRC stimulus (B); signal-to-noise power ratio in the neuronal response (C); and normalized predictive power of the context model fit (D). There was no apparent relationship between PRF/CGF stability and firing rate or signal-to-noise power ratio. The stability of PRFs showed a weak positive correlation with normalized predictive power (Spearman’s rho = 0.3, p = 0.014), whereas that of CGFs did not.
![](https://prod--epp.elifesciences.org/iiif/2/98415%2Fv1%2Fcontent%2F537782v2_fig8.tif/full/max/0/default.jpg)
Small but significant effects of locomotor activity and pupil dilation on context model fits.
A. Observed spiking activity of a single unit (top, black) overlaid with the context model prediction (top, red). Underneath is a trace showing the difference between the two (i.e., the residual, shown in grey, measured in spikes). Dotted black line indicates zero residual. Further below is the pupil diameter (measured as a proportion relative to eye width), and below it a trace of the animal’s running activity over the same period of time. B. Interquartile range (IQR) of the residuals for individual units recorded when the mice were still versus moving (i.e., timebins when running cylinder rotation speed was zero versus non-zero). Dotted black line indicates diagonal where the IQRs are equal. Note that more data points fall below than above the line, but that most data points are very close to the diagonal. C. Residual IQRs for units recorded when the pupil size was small versus large (i.e., pupil diameter less than or greater than the median pupil diameter for the relevant recording sessions). Conventions and observations as in B.
![](https://prod--epp.elifesciences.org/iiif/2/98415%2Fv1%2Fcontent%2F537782v2_figS1.tif/full/max/0/default.jpg)
Core auditory cortical recording sites identified using physiological criteria.
A-C. Three example units recorded from one animal. Left: raw waveform on each of the four tetrode channels. Middle: PSTH of the response of each unit to 100-ms pure tones, pooled across tone frequencies and intensities. Blue shading indicates time of tone presentation. All three units met the two criteria for classification as “core” auditory cortex: (1) robust responses to tone pips (significant difference in firing rate across trials between the 50 ms before and 50 ms after tone onset; Wilcoxon rank-sum test, p < 0.01), and (2) response latency <20 ms. Latency is indicated here with a red vertical line and was defined as the first time bin after tone onset where the firing rate fell outside the mean (dotted black line) ±3 standard deviations (grey shaded area) of the bin-by-bin firing rates in the 50 ms before tone onset. Right: Frequency-Response Areas (FRA). Top of each panel: frequency tuning profile averaged over all tone intensities. The grey and black lines indicate estimates of the frequency tuning profile obtained from two different runs of the tone-pip sequence separated by more than 20 minutes. The overlap of these two lines illustrates the consistency of frequency tuning estimates in units with the robust, short-latency responses typical of “core” auditory cortex.
![](https://prod--epp.elifesciences.org/iiif/2/98415%2Fv1%2Fcontent%2F537782v2_figS2.tif/full/max/0/default.jpg)
Effects of locomotor activity and pupil dilation on context model residual distributions and median residuals.
A. 2D histograms of the context model residuals plotted against the model predictions when mice had: (i) non-zero speeds (leftmost plot); (ii) zero speed (still: middle left); (iii) a pupil diameter above the median diameter for the unit recordings (middle right); and (iv) a pupil diameter below the median diameter (rightmost plot). Pupil median diameters were calculated for each unit based on all data available from the relevant recording sessions. Plots show pooled data from all suitable timepoints in recordings from all mice. B-C. Scatter plots showing the median of the residuals when the mice were still versus moving (B) or when the pupil size was small versus large (C). Dotted black line indicates equal values. Note that median residuals were typically slightly negative, indicating that the context model tended to over-predict firing rates. Note also that median residuals were significantly more positive (i.e., in most cases, less negative) whenthe animal was moving or the pupil was large (Wilcoxon sign-rank tests: locomotion data, p = 1.8x10−24; pupil data, p = 3.7x10−14). Effect sizes were relatively small (Cohen’s d: locomotion data 0.23; pupil data 0.29), but not as tiny as for residual IQRs.