Spontaneous activity is correlated with coding density in primary auditory cortex

A traditional view of spontaneous activity is that it is “noise,” but studies linking spontaneous activity to physiological features, anatomic features, and disorders cast doubt on this notion. We found that the amount of activation relative to suppression in response to multiple wide-band stimuli is directly related to spontaneous activity. This finding supports a novel theoretical understanding of both the purpose of spontaneous activity in primary auditory cortex and its role in tinnitus.

Abstract

Sensory neurons across sensory modalities and specific processing areas have diverse levels of spontaneous firing rates (SFRs) in the absence of sensory stimuli. However, the functional significance of this spontaneous activity is not well-understood. Previous studies in the auditory system have demonstrated that different levels of spontaneous activity are correlated with a variety of physiological and anatomic properties, suggesting that neurons with differing SFRs make unique contributions to the encoding of auditory stimuli. Additionally, altered SFRs are a correlate of tinnitus, arising in several auditory areas after exposure to ototoxic substances and noise trauma. In this study, we recorded single-unit activity from primary auditory cortex of awake marmoset monkeys while delivering wide-band random-spectrum stimuli and white Gaussian noise (WGN) to examine any divergences in stimulus encoding properties across SFR classes. We found that higher levels of spontaneous activity were associated with both higher levels of activation relative to suppression across a variety of wide-band stimuli and higher driven rates in response to WGN. Moreover, response latencies to WGN were negatively correlated with the level of activation in response to both stimulus types. These findings are consistent with a novel view of the role spontaneous spiking may play during normal stimulus processing in primary auditory cortex and how it may malfunction in cases of tinnitus.

NEW & NOTEWORTHY

A traditional view of spontaneous activity is that it is “noise,” but studies linking spontaneous activity to physiological features, anatomic features, and disorders cast doubt on this notion. We found that the amount of activation relative to suppression in response to multiple wide-band stimuli is directly related to spontaneous activity. This finding supports a novel theoretical understanding of both the purpose of spontaneous activity in primary auditory cortex and its role in tinnitus.

spontaneous neural activity in sensory areas is a widespread phenomenon, although its role in sensory processing is not well-understood. Liberman (1978) demonstrated that auditory nerve fibers (ANFs) with high spontaneous firing rates (SFRs) had lower response thresholds than low-SFR fibers, suggesting divergences in functional roles for neurons with different SFRs. Since then, evidence has mounted that spontaneous activity plays a relevant functional role in multiple auditory areas. In the auditory nerve, SFR has been linked to differences in several physiological and anatomic features, including sharpness of tuning (Temchin et al. 2008), ANF diameter (Liberman 1982; Liberman and Oliver 1984), features of the synaptic body (Merchan-Perez and Liberman 1996), and distinct projection patterns (Liberman 1991). Average spontaneous activity is significantly lower in areas downstream of the auditory nerve, with a reported mean SFR of 1.9–3.5 spikes/s in cat primary auditory cortex (A1) compared with >70 spikes/s in cat auditory nerve (Eggermont 2015). These observations suggest that differences in SFR reflect functional differences between distinct cell types within different auditory areas as well as functional differences between the sensory processing roles of different auditory areas overall.

In a variety of auditory processing areas, changes in SFR follow exposure to ototoxic substances and damaging acoustic stimuli (Brozoski et al. 2002; Eggermont 1992; Eggermont and Komiya 2000; Kalappa et al. 2014; Müller et al. 2003; Norena and Eggermont 2003). SFR changes are also linked to tinnitus (Rauschecker 1999; Syka 2002). In the auditory nerve, damaging influences typically reduce SFRs, whereas in A1, both ototoxic substances and mechanical trauma increase SFRs (Eggermont and Roberts 2004). However, the nature of the changes in spontaneous activity in A1 varies depending on the time after trauma exposure: spontaneous burst firing increases immediately after trauma and subsides several hours later, whereas overall SFR increases several hours after trauma (Eggermont 2015; Norena and Eggermont 2003). Like burst firing, tinnitus immediately follows mechanical trauma (Chermak and Dengerink 1987; Loeb and Smith 1967; Mrena et al. 2002). Burst firing has been shown to be more common in superficial layers of cortex, and higher SFR has been associated with lower probabilities of burst firing (Eggermont et al. 1993). These observations support the notion that spontaneous activity plays a significant role in tinnitus, although its role may vary by circumstance, as changes in spontaneous activity could potentially give rise to or alleviate tinnitus symptoms in different contexts.

The principle of sparse coding has garnered experimental support in several sensory cortical areas (Hromádka et al. 2008; Poo and Isaacson 2009; Vinje and Gallant 2000), including A1, in which it has been shown that only 5% of neurons elicit high activity at any given instant in response to auditory stimuli (Hromádka et al. 2008). However, the possibility that specific subpopulations of neurons within A1 exhibit a dense code relative to the remainder of the population remains plausible, and differences in stimulus coding density are likely indicative of functional differences between subpopulations. We hypothesized that high-SFR neurons exhibit a denser code than the population of neurons as a whole, which we define as a code with relatively unselective stimulus-driven activation and relatively selective stimulus-driven suppression. We tested this hypothesis by quantifying the extent to which activation or suppression was more prevalent across a wide-band random spectrum stimulus (RSS) set, which represents all frequencies equally. Using this set of stimuli enabled a comparison of frequency coding density across subpopulations with different SFR distributions. We also considered responses to white Gaussian noise (WGN) as an alternative to this metric. We hypothesized that the activation-suppression balance (A:S balance) metric determined from the RSS would be positively correlated with excitatory responses elicited by WGN because neurons with wider excitatory receptive fields and narrower inhibitory receptive fields would be more likely to be driven by WGN. Our experimental results support a novel understanding of the role of spontaneous activity in A1 neuronal function and suggest a new perspective on the link between spontaneous activity and tinnitus.

MATERIALS AND METHODS

Surgery and Recording

Adult common marmoset monkeys (Callithrix jacchus) were used as subjects. All training, recording, and surgical procedures complied with the U.S. National Institutes of Health Guide for the Care and Use of Laboratory Animals and were approved by the Washington University in St. Louis Animal Studies Committee. Subjects were trained to sit upright in a custom, minimally restraining primate chair inside a double-walled sound-attenuation booth (IAC 120a-3, Bronx, NY) for the same duration as future physiology recordings. After they acclimatized to this setup, a custom headcap for electrophysiological recording was surgically affixed to the skull of each subject. The location of the vasculature running within the lateral sulcus was marked on the skull at the time of surgery. The animals were allowed sufficient time to recover following surgery and were given pain medication to eliminate discomfort. Using the lateral sulcus as a guide, microcraniotomies (<1-mm diameter) were drilled through the skull over the temporal lobe for physiology experiments. An active recording craniotomy was partially filled with antibiotic ointment and dental cement to prevent excess tissue growth and infections after each daily recording session. The craniotomy was permanently sealed with dental cement before another active recording craniotomy was drilled. This procedure greatly preserved the cohesion of the bone and the anatomic landmarks. Daily recordings lasted ∼4 h for each animal and collectively lasted several months. The animals were head-fixed during recordings. The location of A1 was identified anatomically based on lateral sulcus and bregma landmarks and confirmed with physiological mapping (Stephan et al. 2012).

A single, high-impedance, 125-μm, tungsten-epoxy electrode (∼5 MΩ at 1 kHz; FHC, Bowdoin, ME) was advanced perpendicularly to the cortical surface within each microcraniotomy. Microelectrode signals were amplified using an alternating current (AC) differential amplifier (A-M Systems 1800, Sequim, WA) with the differential lead attached to a grounding skull screw. Single-unit action potentials were sorted online using manual template-based spike-sorting hardware and software (Alpha Omega, Nazareth, Israel). The median signal-to-noise ratio for single units recorded with this preparation is 24.5 dB, and median peak-to-peak amplitude is 1.1 mV. When a template match occurred, the spike-sorting hardware relayed a transistor-transistor logic (TTL) pulse to the digital signal processing (DSP) system [Tucker-Davis Technologies (TDT) RX6, Alachua, FL] that temporally aligned recorded spike times (2.5-μs accuracy) with stimulus delivery. Recording locations over the exposed skull were varied daily, covering all regions of interest.

Experimental Procedure

Acoustic stimuli were delivered in free field through a loudspeaker (B&W 601 S3, Worthing, United Kingdom) located 1 m along the midline of and directly in front of the animal's head. The output of this speaker was calibrated so that the maximum sound level delivered was 105-dB sound pressure level (SPL) with a flat frequency response from 60 Hz to 32 kHz (Watkins and Barbour 2011).

White Gaussian noise (WGN) and wide-band random spectrum stimuli (RSS) were used for these experiments. The randomly generated WGN stimulus used for analysis had a sound level of 75-dB SPL, and the 134 different RSS ranged in sound level from 60.77- to 71.02-dB SPL. RSS were presented in randomized order. For further details on the properties of RSS, see Barbour and Wang (2003).

Single-unit activity in A1 was recorded from 2 awake adult marmoset monkeys while they passively listened to the playback of WGN and RSS. Spontaneous spiking data from a 3rd animal was added to establish spontaneous spiking statistics. Each animal was observed with a camera throughout the recording procedure, and it was roused during any periods of apparent drowsiness. Units were isolated while delivering a variety of stimuli (e.g., tones, vocalizations, WGN, and RSS), although units were not selected for recording on the basis of a response to any particular stimulus. Once a unit was isolated, several stimulus sets were delivered repeatedly to the animal, including WGN and RSS. WGN stimuli were 500 ms in duration with at least 1-s interstimulus interval (ISI). RSS were 100 ms in duration with at least 500-ms ISI. Fifty repetitions of WGN stimuli and five repetitions of RSS were delivered.

Analysis

Spontaneous firing rate calculations.

Spontaneous firing rates were calculated using the 500- and 200-ms windows before stimulus presentations of WGN and RSS stimuli, respectively. The spontaneous firing rate was determined from dividing the total amount of spikes in this window by the total amount of time in this window. For instance, WGN experiments involved 50 presentations each of WGN stimuli at 4 different intensities, so the total time used for this calculation was determined by multiplying 200 stimulus presentations by the 500-ms prestimulus window. Inserting 500 and 300 ms of poststimulus silence in addition to the prestimulus period used for determining SFR for WGN and RSS recordings, respectively, prevented offset responses from significantly affecting the calculation of spontaneous rates. Because a single spike throughout the entire recording session (before, during, or after the stimulus) was sufficient for the unit to be included in our analysis and because we did not preselect units for analysis on the basis of stimulus responsiveness, a substantial portion of our data set consisted of low-SFR units. These efforts were intended to reduce the natural bias of extracellular recording methods toward high-SFR units.

A:S balance calculations.

Activation-suppression (A:S) balances were determined from responses to either 67 or 134 different RSS uniformly representing a bandwidth of 4 octaves. A total of 219 single units were isolated and used for analyses. All well-isolated single units that had sufficiently stable SFRs across RSS and WGN recordings and had SFRs >1 spike/s were used for A:S balance analysis (n = 118). Units with unstable SFRs across particular experiments were not considered for analysis: if the SFR calculated using the RSS experiments was both greater than 2 SD_sp (see below) less than or greater than the SFR during the WGN experiment, and if the SFR from the RSS recordings did not represent the same SFR quartile (see Spontaneous firing rate grouping), then the unit was not included in further analysis. Units with SFRs <1 spike/s were excluded from A:S balance analysis (n = 101) because we were interested in studying the degree of both activation and suppression, and spike rate suppression is difficult to measure reliably at low SFRs and finite recording time. The A:S balance is the average stimulus-induced rate change across the entire RSS stimulus set, expressed in standard deviations of the SFR:

$$\text{AS}=\frac{{\displaystyle \sum _{i=1}^n\left({\text{DR}}_i-{\text{SFR}}_i\right)}}{n\cdot {\text{SD}}_{\text{sp}}},$$

in which AS represents the A:S balance, n is equal to the number of stimuli in the RSS set, DR_i is equal to the absolute discharge rate for the ith stimulus, SFR_i is equal to the spontaneous firing rate for the ith stimulus, and SD_sp is equal to the standard deviation of the SFR measurements. The SFR was determined as described previously, whereas discharge rates were determined from the mean absolute firing rates in response to five repetitions of each RSS. Because sustained responses contain full information about stimulus identity (Wang et al. 2005), only discharge rates 50 ms after the start of the stimulus until the end of the stimulus were used to calculate A:S balances. The SD_sp values were calculated by determining spiking rates in 200-ms windows before individual RSS stimulus presentations and calculating the standard deviation of these measures. The purpose of normalizing to SD_sp was to compare changes in firing rate between neurons while also taking into account the typical amount of activity variation in the absence of stimuli. We found that SD_sp generally increased with increasing SFR; when neurons with SFRs >1 spike/s were divided into four equal quartiles on the basis of SFR, the median SD_sp values for each quartile in the order of increasing SFR were 1.60, 2.06, 2.92, and 5.71 spikes/s. These values effectively represent the 12.5, 37.5, 62.5, and 87.5% quantiles.

WGN response ratio calculations.

For each unit with SFR >1 spike/s, a WGN ratio response was calculated:

$$\text{WR}=\frac{\text{DR}}{\text{SFR}},$$

in which WR represents the WGN response ratio, DR represents the discharge rate, and SFR represents the spontaneous firing rate. Only WGN delivered at 75-dB SPL was used for this analysis to provide a stimulus with a sound level close to that of the RSS. Units with particularly low SFRs (0–1 spikes/s) were not included in WGN response ratio analysis because particularly low SFRs can lead to misleadingly high or infinite values if the SFR is close to 0 or exactly 0. The final 400 ms of the 500-ms stimulus was used for discharge rate calculations to measure only sustained responses for the same reason noted above for A:S balance calculations.

Latency calculations.

Response latencies were determined in response to 75-dB WGN. The 500-ms silent period before each stimulus presentation was divided into 5-ms bins, and firing rates were calculated for each of these bins. From these prestimulus bins, a standard deviation (SD_bin) was calculated to judge typical fluctuations in spontaneous firing patterns unrelated to stimulus presentation. The stimulus period was also divided into 5-ms bins. For any given neuron, the latency was defined as the time point at the beginning of the 5-ms bin in which the firing rate first achieved ≥5·SD_bin above the spontaneous firing rate. This measure resulted in clear delineation between responsive and unresponsive units for the number of stimulus repetitions used. If no bin in the 1st 50 ms of the stimulus contained a 5 SD_bin increase in firing rate, then that unit was not ascribed a latency. Units with SFRs <1 spike/s were not considered for latency analysis because their sparse spontaneous and driven spiking prevented accurate latency calculations. Out of 118 units with SFRs >1 spike/s, 60 met all of these criteria and were ascribed latencies. This approach is similar to methods used in previous studies of latency in primate primary auditory cortex (Bendor and Wang 2008; Camalier et al. 2012; Kuśmierek and Rauschecker 2009) but has a slightly more restrictive SD_bin requirement and a more liberal requirement for the amount of time the response must continue to include latencies from transient responses.

Spontaneous firing rate grouping.

Altogether, 219 neurons were isolated and used for analyses. SFRs were calculated as described above (Spontaneous firing rate calculations). The SFR of each particular unit was determined from its spontaneous firing rate during the WGN experiments and used for each stage of analysis to maintain consistent cell groupings organized by SFR. Neurons with SFRs >1 spike/s (n = 118) were divided into 4 approximately equal quartiles on the basis of their SFR. The subset of units with SFRs <1 spike/s (n = 101), in addition to the remainder of the neurons, was used for an alternative analysis involving RSS and WGN. For this analysis, mean driven rates across all RSS were determined for all neurons recorded (n = 219), along with both driven rates and discharge rates in response to 75-dB WGN.

Statistical tests.

The Mann-Whitney U test was used to compare A:S balance and WGN response ratio values between each of the SFR quartiles. The Spearman correlation metric was used for all correlation analyses to quantify monotonic relationships between variables. Statistical significance of Spearman correlations was determined using 1 million permutation test resamples or Mann-Whitney U tests, wherever appropriate.

RESULTS

Distribution of Spontaneous Spiking in Primary Auditory Cortex

We initially determined the distribution of SFRs in marmoset A1 to provide context for our analysis. The distribution of SFRs from our 219 units can be seen in Fig. 1. This exponential distribution has a mean of 3.46 spikes/s and a median of 1.35 spikes/s, consistent with previous studies (Wang et al. 2005; Watkins and Barbour 2011). Adding a large number of spontaneous spiking measurements from a 3rd monkey provides enough samples to reveal a nearly log-normal distribution of SFRs with some skew toward lower values (Fig. 1, inset). This finding also reflects previous observations (Hromádka et al. 2008), although our skew toward lower values may result from our sampling procedure of analyzing every single unit we isolated regardless of its responsiveness to acoustic stimuli.

Fig. 1.

Spontaneous firing rate (SFR) distribution in A1. The histogram is divided into bins with widths of 1 spike/s. The inset shows a logarithmic probability distribution of marmoset A1 SFR values. Light shading indicates units with SFR <1, which were not analyzed for suppressive responses.

Characteristic Responses of Single Units to WGN and RSS

A variety of neural responses occurred on delivery of both stimulus types. Although it is commonly believed that WGN is a poor stimulus for driving neurons in higher sensory areas (Depireux et al. 2001; Elhilali et al. 2004; Nelken 2004; Valentine and Eggermont 2004), we found that WGN very often drove a change in firing patterns for neurons with SFRs >1 spike/s. Typical neural responses to WGN included strong suppression, strong activation, onset responses followed by suppression, and temporally locked firing responses at specific times during the stimulus. For RSS, sustained firing, suppression, and onset responses followed by suppression were commonly observed. Offset responses were sometimes observed in response to both stimulus types. Responses to RSS in marmoset A1 have been extensively characterized in previous work (Barbour and Wang 2003), and our current observations reflected the same trends. Raster plots of characteristic single-unit responses to WGN and RSS stimuli are shown in Fig. 2.

Fig. 2.

Raster plots of characteristic responses to random spectrum stimuli (RSS) and white Gaussian noise (WGN). A: raster plot of a high-SFR single-unit response to 5 repetitions of 67 distinct RSS. For this and other raster plots, red dots represent spikes that were used to calculate discharge rates, and blue dots represent other spikes. The shaded region represents the duration of the stimulus presentation. Horizontal gray lines separate distinct RSS. Responses are sorted by discharge rate. B: raster plot of a high-SFR single-unit response to 50 repetitions of 75-dB SPL WGN. C: raster plot of a moderate-SFR single-unit response to RSS. D: raster plot of a moderate-SFR single-unit response to 75-dB SPL WGN.

A high-SFR (17.2 spikes/s) unit, Unit F16E445.1, was driven by many RSS to high discharge rates and suppressed by relatively few RSS. Therefore, it exhibits a high, positive A:S balance (4.99). In addition, this unit was also driven by WGN, with a WGN response ratio of 3.48. By contrast, a moderate-SFR (3.69 spikes/s) unit, Unit F16E339.2, exhibited a low, negative A:S balance (−1.43). Similarly, because of strong suppression by WGN, it has a WGN response ratio of 0. Note that only 50 presentations of 75-dB SPL WGN are displayed, whereas 335 total presentations of RSS are displayed (67 unique stimuli × 5 repetitions), leading to a visually apparent difference in spiking density between the 2 types of plots for either unit even though the SFR remained virtually unchanged.

A:S Balances and WGN Response Ratios Across SFR Subpopulations

To determine the variation across the population regarding responses to RSS and WGN, all neurons with SFRs >1 spike/s (n = 118) were split into four quartiles ordered by SFR. Median A:S balances and median WGN response ratios were then determined for each quartile. These results are shown in Fig. 3.

Fig. 3.

Median WGN response ratio (RR) and median A:S balance are related to SFR. A: median WGN response ratios for each SFR quartile. Units with higher spontaneous rates respond on average more strongly to WGN. B: median A:S balance values for each SFR quartile. Units with higher spontaneous rates respond on average more strongly to RSS.

Similar patterns existed for both median A:S balances and median WGN response ratios across SFR groups, distinguished by a noticeable increase from the first quartile to the second and third and a major increase from the middle quartiles to the highest quartile. The median A:S balances indicated levels of suppression that typically matched or exceeded levels of excitation for all quartiles other than quartile 4, consistent with sparse excitatory coding and broad inhibitory coding for the majority A1 neurons (Hromádka et al. 2008). This was not the finding for high-SFR neurons, however, which exhibited A:S balances skewed toward activation. The highest-SFR quartile exhibited significantly different A:S balances and WGN response ratios from the lowest-SFR quartile (P = 6.50 × 10⁻³ and P = 1.70 × 10⁻³, respectively, Mann-Whitney U test, Bonferroni corrected to α = 8.3 × 10⁻³). All other quartile comparisons were nonsignificant after correction. Note that for two hypothetical units with a WGN response ratio of 2, driven rates may be entirely different. For a quartile 1 unit with SFR = 1 spike/s, for example, the driven rate would be 1 spike/s (i.e., absolute discharge rate of 2 spikes/s), whereas for a quartile 4 unit with SFR = 15 spikes/s, the driven rate would be 15 spikes/s (i.e., absolute discharge rate of 30 spikes/s).

Although higher SFR units have higher median A:S balances and WGN response ratios overall, response heterogeneity exists within each quartile. This heterogeneity is reflected in Fig. 4, which contrasts quartile 1 with quartile 4. Levels of activation relative to suppression for one stimulus metric largely matched the levels determined from the other metric in both quartiles. Quartile 1 was skewed toward suppression, with 67% of neurons having A:S balances <0 and 63% of neurons having WGN response ratios <1. By contrast, quartile 4 tended toward activation, with 76% of neurons having A:S balances >0 and WGN response ratios >1. Given the traditional view of WGN as a poor stimulus for activating neurons at higher processing levels (Depireux et al. 2001; Elhilali et al. 2004; Nelken 2004; Valentine and Eggermont 2004), it is particularly interesting that 52% of units in quartile 4 had WGN response ratios >2, whereas 45% of units in quartile 1 had WGN response ratios <0.5. The large number of quartile 1 units in the (−1, 0] bin rather than the less than or equal to −1 bin is partly because these units have a limited range of suppression due to their low SFRs: for a neuron with SD_sp = 2 (approximately the mean SD_sp value for quartile 1), for example, the mean RSS driven rate value must be less than or equal to −2 spikes/s to have an A:S balance less than or equal to −1. Because these neurons have very low SFRs, this outcome is unlikely, which explains the finding that only a single neuron in this quartile has an A:S balance less than or equal to −1.

Fig. 4.

Low-SFR (quartile 1) units tend toward suppression in response to WGN and RSS, whereas high-SFR (quartile 4) units tend toward activation. A: histogram of A:S balance values for neurons in quartile 1. An A:S balance <0 means that there was net suppression across the entire RSS set. B: histogram of WGN response ratios for units in quartile 1. A WGN response ratio <1 means that the discharge rate was less than the spontaneous rate. C: histogram of A:S balance values for units in quartile 4. D: histogram of WGN response ratios for units in quartile 4.

Evaluation of the entire data set reveals trends similar to those in the group analysis. A Spearman rank-order correlation between A:S balance and SFR yields a value of ρ = 0.147. Because the A:S balance metric includes SFR in its calculation, these two measures would not be expected to be uncorrelated even if all of the contributing variables were independent. In other words, the null hypothesis is not ρ = 0. We used a permutation resampling test to establish the null hypothesis for this comparison, which amounts to forming new populations with shuffled labels and repeatedly calculating Spearman rank-order correlations between them. One million such resampling operations are depicted in Fig. 5A, which reveals a null hypothesis of ρ = −0.548. None of the shuffled values has a correlation greater than the observed value, yielding a statistically significant result (P < 10⁻⁶). Therefore, the raw correlation coefficient value of 0.147 in this case actually reveals a much stronger correlation between A:S balance and SFR than might otherwise be interpreted by assuming a null hypothesis of ρ = 0.

Fig. 5.

SFR is significantly correlated with both A:S balance and WGN response ratio. A: permutation resampling reveals that the actual correlation between A:S balance (ASB) and SFR (ρ = 0.147; vertical solid line) is significantly greater than the null correlation (vertical dashed line) for all units with SFR >1 spike/s (P < 10⁻⁶). B: permutation resampling reveals that the actual correlation between WGN response ratio and SFR (ρ = 0.213; vertical solid line) is significantly greater than the null correlation (vertical dashed line) for all units with SFRs >1 spike/s (P = 1.07 × 10⁻³).

We next examined the correlation between SFR and WGN response ratio. Once again, they are related by definition, so we used 1 million resamples of a permutation test to determine the null hypothesis correlation of ρ = −0.0525 (Fig. 5B). The actual measured correlation of ρ = 0.213 is significantly greater than this value, indicating a strong correlation between these measures (P = 1.07 × 10⁻³).

Given the relationship among SFR, A:S balance, and WGN response ratios, we next directly examined the correlation between A:S balance and WGN response ratio. These Fig. 6 two measures are significantly Spearman-correlated using the same permutation test methodology described previously (ρ = 0.664, P < 10⁻⁶). On balance, therefore, RSS and WGN appear to reveal consistent information about A1 neuron responsiveness.

Fig. 6.

A:S balances and WGN response ratios are significantly correlated. Scatterplot of the logarithm of the WGN response ratio as a function of A:S balance for all units with SFRs >1 spike/s. A significant monotonically increasing relationship can be observed in this population, verified by permutation resampling (inset; ρ = 0.664, P < 10⁻⁶).

Although A:S balance and WGN response ratio analyses were not performed on 0–1 spikes/s SFR units for reasons described previously, we were still interested in how these units responded to RSS and WGN. Various raw rate metrics comparing responses to RSS and WGN of 0–1 SFR units (n = 101) with the other quartiles are shown in Table 1. Although suppression cannot be easily inferred for 0–1 spikes/s SFR units, the raw metrics indicate that these units are mostly unresponsive to WGN and RSS. They are not entirely unresponsive to WGN and RSS, however: relative rates for preferred RSS among 0–1 spikes/s SFR units rose as high as ∼35 spikes/s, and ∼10% of units had driven rates elicited by 75-dB WGN >1 spike/s, with a maximum of ∼8.5 spikes/s. Units with 0–1 spikes/s SFR that were responsive to WGN typically responded at a specific point during WGN rather than consistently firing throughout the stimulus (data not shown). It cannot be concluded that the higher median driven rates to 75-dB WGN for 0–1 spikes/s SFR units compared with quartile 1 are representative of greater levels of suppression for quartile 1 units because the low SFRs of the 0–1 spikes/s subpopulation renders the comparison of extracellularly recorded driven rates uninformative for evaluating suppression.

Table 1.

0–1 SFR neurons are mostly unresponsive to RSS and 75-dB WGN

SFR Subpopulation	Median Mean RSS Driven Rate, spikes/s	Median WGN Discharge Rate, spikes/s	Median WGN Driven Rate, spikes/s
0–1 spikes/s	−0.040	0.150	0.000
Quartile 1 (1.10–2.31 spikes/s)	−0.410	0.825	−0.600
Quartile 2 (2.40–4.71 spikes/s)	0.442	4.45	1.32
Quartile 3 (4.74–7.26 spikes/s)	0.254	7.10	1.26
Quartile 4 (7.47–34.67 spikes/s)	4.58	26.5	17.1

Typical responses to RSS and WGN across SFR subpopulations are shown. Driven rates were calculated for RSS without including the 1st 50 ms of the stimuli. Both driven and discharge rates for WGN were calculated without including the 1st 100 ms of the stimulus.

The Relationship of A:S Balance and WGN Response Ratio with WGN Response Latency

Because previous results have linked short response latencies with other properties of cells in A1, including identity as a putative inhibitory interneuron (Wu et al. 2008), we examined relationships between response latency and our other experimental metrics. We hypothesized that units with higher WGN response ratios and A:S balances may also have shorter response latencies to WGN than units with lower WGN response ratios and A:S balances. The result of this assessment is shown in Fig. 7. No statistically significant relationship was found between SFR and response latency. However, statistically significant negative correlations existed between A:S balance and response latency (ρ = −0.292, P = 0.0235, Spearman rank-order correlation, Mann-Whitney U test) as well as WGN response ratio and response latency (ρ = −0.272, P = 0.0353, Spearman rank-order correlation, Mann-Whitney U test). These results are indicative of decreasing coding density among neurons in A1 as auditory computation proceeds, which could have functional relevance to the sparse sustained responses in A1 (Hromádka et al. 2008; Wang et al. 2005). Because only units with onset responses were usable for this latency analysis, it remains possible that different relationships with response latency exist among units without clear onset responses for which latencies are difficult to measure using extracellular methods.

Fig. 7.

A:S balance and WGN response ratio are negatively correlated with response latency. A: scatterplot of response latency as a function of A:S balance. B: scatterplot of response latency as a function of WGN response ratio.

DISCUSSION

We determined the single-unit SFRs of marmoset monkey A1 neurons and compared these spiking rates with the discharge rates elicited by RSS and WGN. We demonstrated that the A:S balance metric derived from RSS is significantly correlated with the WGN response ratio metric for individual neurons and that both of these metrics show a significant relationship with SFR, namely, that high SFR neurons tend toward stimulus-induced activation rather than suppression for both RSS and WGN. A:S balance and WGN response ratios are also negatively correlated with response latency. The varied responses of neurons to WGN and the correlation of this value with A:S balance supports the notion that WGN is a useful stimulus for quickly probing the properties of neurons in higher auditory areas, contrary to previous characterizations of WGN (Depireux et al. 2001; Elhilali et al. 2004; Nelken 2004; Valentine and Eggermont 2004). These observations are indicative of functional differences between neurons with different SFRs, as discussed below.

High SFR Neurons are Likely to be Predominantly Inhibitory Interneurons

A number of previous studies in a variety of sensory areas provide substantial evidence that cortical inhibitory interneurons are less selective than cortical excitatory neurons, including in A1 (Hromádka et al. 2008; Peters and Kara 1985; Poo and Isaacson 2009; Sohya et al. 2007; Swadlow 2003; Wu et al. 2008). Additionally, shorter response latencies are consistent with feedforward inhibition and have also been associated with inhibitory interneurons in A1 (Wu et al. 2008). As shown in this study, high-SFR neurons respond to a variety of wide-band stimuli with activation more than suppression, in contrast to the largely suppressive responses or lack of response at low SFRs. In addition, A:S balance and WGN response ratios are negatively correlated with response latency, consistent with feedforward inhibition. These observations support the hypothesis that high-SFR neurons in A1 are predominantly inhibitory interneurons. In this view, relatively unselective inhibitory interneurons with relatively high A:S balances and WGN response ratios predominate at high SFR values and become increasingly uncommon at lower SFR values.

Implications for Circuit Organization in Primary Auditory Cortex

Given that action potentials are energetically costly for the brain (Attwell and Laughlin 2001; Niven and Laughlin 2008), the existence of high-SFR neurons within sensory cortex may appear problematic from an evolutionary perspective. Our results provide insight into the functional logic of maintaining high-SFR neurons within primary auditory cortex. The majority of neurons in cortex are excitatory (Hromádka et al. 2008; Peters and Kara 1985; Prieto et al. 1994). This is consistent with the hypothesis that inhibitory neurons are generally those with high SFRs because the majority of neurons in A1 have very low SFRs; witness the exponentially distributed SFRs in Fig. 1. Stimulus-evoked responses in A1 are sparse (Hromádka et al. 2008), which can also be ascertained from the relatively low median A:S balances for the lowest three SFR quartiles and the relative unresponsiveness of 0–1 spikes/s SFR neurons to RSS. If 1) high-SFR neurons are typically inhibitory, 2) responses of inhibitory neurons are relatively dense (Hromádka et al. 2008; Peters and Kara 1985; Poo and Isaacson 2009; Sohya et al. 2007; Swadlow 2003; Wu et al. 2008), 3) low-SFR neurons are typically excitatory, and 4) responses of low-SFR neurons are sparse, then a model emerges that can explain the utility of maintaining neurons with high SFRs in A1. In this model, tonic inhibition is sent to excitatory neurons throughout the cortex even in the absence of auditory stimuli. Such an arrangement prevents high levels of firing in low-SFR, highly selective excitatory neurons during silence. Hence, the high SFRs of inhibitory neurons are useful as a means to help prevent “accidental” or “erroneous” firing in low-SFR neurons if they receive some level of excitatory input despite the absence of their preferred stimuli. If physiological features observed in the auditory nerve, i.e., that neurotransmitter vesicle quantity at the synaptic cleft is higher for low SFR neurons than high SFR neurons (Merchan-Perez and Liberman 1996), are shared by neurons in auditory cortex, then relatively few synchronized spikes from low-SFR excitatory presynaptic neurons could potentially drive significant firing in postsynaptic excitatory neurons despite high levels of presynaptic inhibitory firing. This view is consistent with intracellular studies showing that auditory cortex neurons generally exhibit rare spontaneous depolarizations that typically last a very short time (Hromádka et al. 2013), presumably because the membrane potential rapidly repolarizes after rare input spikes from low-SFR presynaptic excitatory neurons due to the continual input from high-SFR inhibitory neurons. The constant, widespread inhibition proposed for auditory cortex may also help explain why cortex does not appear to play a significant role in the generation of audiogenic seizure (Kesner 1966).

This simplified model of auditory cortical function can also explain the widespread phenomenon of burst firing in A1, which occurs when neurons fire quick bursts of several spikes within a short time frame, causing them to discharge spontaneously in a way that deviates substantially from a Poisson model (Eggermont 2015; Eggermont et al. 1993). This typically occurs in neurons with low SFRs and rarely occurs in neurons with the highest SFRs (Eggermont 2015; Eggermont et al. 1993). Within this model, burst firing occurs because presynaptic excitatory neurons briefly fire synchronously or a presynaptic neuron itself burst fires. This leads to an abnormally condensed release of excitatory transmitter that outweighs simultaneous inhibitory input for a short period. As a result, the characteristic rapid series of discharges becomes increasingly spread out toward the end of the burst firing sequence as the temporarily high excitatory neurotransmitter concentration in the synapse quickly decreases (Eggermont 2015). Low-SFR neurons require a relatively uncommon set of presynaptic events to reach threshold in the absence of preferred stimuli, but when such events occur often enough, neurotransmitter is released to cause several spikes in a short time interval. Decreased levels of inhibition, then, could lead to more spikes per burst and more burst firing, which is observed after auditory damage (Brozoski et al. 2002; Norena and Eggermont 2003). Within this model, spontaneous burst firing in the absence of preferred stimuli in low-SFR excitatory neurons is viewed as erroneous or as not contributing to accurate perception; however, the spontaneous activity of inhibitory interneurons is not considered erroneous, as it serves to decrease burst firing in low-SFR excitatory neurons in the absence of their preferred stimuli.

The Division of Function by SFR Affects the Interpretation of the Causes of Tinnitus

Tinnitus can be conceptualized as an erroneous percept that is caused by erroneous forms of neural activity. A number of forms of neural activity have been correlated with auditory system damage and tinnitus in a variety of auditory areas, including increased spontaneous activity, increased burst firing, increased neural synchrony, and reorganization of tonotopic maps (Brozoski et al. 2002; Eggermont 2015; Eggermont and Komiya 2000; Kalappa et al. 2014; Norena and Eggermont 2003). However, exactly what types of neural activity are erroneous remains controversial. Understanding which physiological correlates of auditory system damage are actual causes of tinnitus and which are representative of different processes, such as compensatory altered function, is crucial to understanding and treating this disorder.

Increased mean SFRs and burst firing are associated with damage to the auditory system, tinnitus, and old age in a number of auditory processing areas, including A1 (Eggermont 2015; Eggermont et al. 1993; Hughes et al. 2010; Norena and Eggermont 2003). The implications of our results suggest a novel way to interpret these phenomena. Increased spontaneous activity is often cited as a likely cause of tinnitus (Brozoski et al. 2002; Eggermont and Komiya 2000; Eggermont and Roberts 2004; Kaltenbach and Afman 2000). Both burst firing in A1 and tinnitus immediately follow auditory damage, whereas increased SFRs follow 2 h after auditory damage in A1 but 2 days after in the dorsal cochlear nucleus (Kaltenbach et al. 2005; Norena and Eggermont 2003). Increases in overall SFR and burst firing are negatively correlated in A1 (Eggermont 2015; Norena and Eggermont 2003). Because the percept of tinnitus immediately follows auditory damage, this time course may appear puzzling if an overall spontaneous activity increase is conceived of as erroneous. Based on our results, however, a different perspective can be taken that resolves this apparent anomaly.

If increases in spontaneous activity primarily result from an increase in spiking for high-SFR inhibitory interneurons, this increase could be viewed as a compensatory mechanism used to help reduce burst firing and erroneous perception during periods of stress for the auditory system rather than as a cause of tinnitus. Therefore, burst firing and synchrony in low-SFR neurons are more likely to be causes of tinnitus. This interpretation is consistent with the observation that in A1, SFRs increase primarily in neurons with characteristic frequencies (CFs) flanking the frequency of the tone used to damage the auditory system (Eggermont 2015; Norena and Eggermont 2003), which also is typically the frequency of the tinnitus percept (Eggermont and Tass 2015). The increase in SFR of flanking CF neurons could be attributed to an increase in high-SFR, inhibitory firing that actually serves the useful purpose of reducing erroneous burst firing in neurons with CFs closest to the damage frequency. Within this model, the high levels of spontaneous activity observed in areas of tonotopic map reorganization in A1 (Eggermont and Komiya 2000) can also be explained by the notion that higher levels of inhibitory activity are required to prevent erroneous firing in low-SFR excitatory neurons, which has a greater potential to occur during periods of increased modification of neural networks. Additionally, higher levels of both burst firing and higher SFRs in old age may be indicative of higher levels of erroneous firing and compensatory responses of the auditory system to maintain accurate perception as brain function declines.

Given that an increase in spontaneous burst firing in low-SFR excitatory neurons and greater activity in high-SFR inhibitory neurons could coexist after auditory damage, similar to old age, it may be difficult to distinguish clearly whether increases in overall spontaneous activity are predominantly due to the increased activity of inhibitory neurons or excitatory neurons. Careful studies will need to be performed to characterize more clearly the neural correlates of tinnitus directly. Our results in animals with undamaged auditory systems emphasize that close attention should be paid to the diversity of spontaneous activity when theorizing about sensory processing in both healthy and damaged neural networks.

GRANTS

This work was supported by the National Institute on Deafness and Other Communication Disorders Grant R01-DC-009215.

DISCLOSURES

No conflicts of interest, financial or otherwise, are declared by the author(s).

AUTHOR CONTRIBUTIONS

D.A.B. and D.L.B. conception and design of research; D.A.B. and R.N. performed experiments; D.A.B. and D.L.B. analyzed data; D.A.B. and D.L.B. interpreted results of experiments; D.A.B. and D.L.B. prepared figures; D.A.B. and D.L.B. drafted manuscript; D.A.B., R.N., and D.L.B. edited and revised manuscript; D.A.B., R.N., and D.L.B. approved final version of manuscript.