Identification and characterization of NA and NM units
Datasets consist of 38 NA and 53 NM units. Neurons of both nuclei respond only to ipsilateral monaural stimuli. NM axons, which were recorded in dorso-ventral penetrations within NL, alternate between ipsilateral and contralateral monaural responses, clearly distinguishing them from NL neurons, which respond to binaural stimulation (Peña et al. 1996
, Viete et al. 1997
). NA can be distinguished from NM and NL units and auditory nerve fibers by their poor ability to lock to the phase of their best frequency (Köppl and Carr, 2003
); phase locking in NA was quantified by measuring vector strength at the neurons’ estimated best frequency. The vector-strength values measured in this study are comparable to those reported by Sullivan and Konishi (1984)
and Köppl and Carr (2003)
. Neurons that lock to the phase of their best frequency will also display periodic oscillations in their SAC () whereas SACs of neurons that do not phase lock are smooth (, Louage et al., 2004
; Joris et al., 2006
; Christianson and Peña 2007
). Taking advantage of this, we used the units’ SACs to further confirm their phase locking ability for both NA and NM by comparing the frequency of the periodicity of the SAC with the neurons’ best frequency. In NM, there was a strong correlation between the frequency of the SAC’s periodicity and the STRFbf
(R= 0.99, p < 0.001). However, there was no correlation for NA units, confirming that they did not phase lock to their best frequency.
Figure 2 NA and NM units exhibit gross differences in their response properties. Representative rasters are shown for the responses of NA (A) and NM (B) neurons to repeated presentations of the same stimulus (frozen noise). (C) SAC corresponding to the NA unit (more ...)
Because the estimation of the STRF and the response reliability requires thousands of spikes, it was necessary to record exclusively from units that responded robustly throughout the 500 ms stimulus. Though this does not represent a problem in NM, our NA sample was biased towards mostly primary-like neurons and onset-type neurons with a sustained discharge. We defined onset-type neurons as those units whose ratio of peak firing rate (within the first 20 ms) to steady-state firing rate (during the last 400 ms) in response to unfrozen noise was equal to or larger than 10 (Rhode and Smith, 1986
). In our dataset, two neurons met this criterion. Köppl and Carr (2003)
classified 32% of NA neurons as primary-like and 6% as onset. Our dataset is therefore skewed towards the most common response type, which also spans the widest frequency range (Köppl and Carr, 2003
To estimate the response latency, we measured the latency of the STRF excitatory peaks. NA and NM units had significantly different response latencies (NA: 1.9 ± 0.51 ms; NM: 2.25 ± 0.34 ms; medians significantly different by Kruskal-Wallis, p<0.005). Compared to response latencies previously reported for NM and NA (Sullivan and Konishi, 1984
; Köppl and Carr, 2003
, respectively), our values are about 0.5 ms shorter. However, both previous studies used PSTHs of responses to tone and/or noise stimuli with onset ramps to estimate first-spike latency. Our method of estimating response latency differs by taking into account the latency of spikes throughout the stimulus and by construction (see methods) avoids the confounding effect of stimulus onset ramp.
As reported by Sullivan and Konishi (1984)
, we also observed lower firing rates in NA than in NM, albeit those conclusions were based on spontaneous rates, whereas our observations are based on driven rates. We observed a mean response rate of 183 ± 110 spikes/s in our sample of primary-like NA units, measured during the first 50 ms of their response to unfrozen noise. This value is lower than that reported by Köppl and Carr (2003)
. Theirs, however, is saturation rate measured with tone stimulation, which makes the difference difficult to interpret. We did not measure saturation rate with tonal stimulation routinely. An estimate of effective refractory period was made by measuring the minimum inter-spike interval from ISIHs obtained with multiple repetitions of unfrozen noise. NA and NM units’ effective refractory periods were not significantly different (1 ± 0.38 ms and 0.8 ± 0.19 ms, respectively; p > 0.05, t-test).
Response reliability in NA vs. NM
We compared the stimulus-dependent spike-timing reliability in NA (n = 38) and NM (n = 40) upon repeated presentations of the same stimulus (frozen-noise protocol). Spike-timing reliability, as viewed here, is different from phase-locking reliability. Stimulus phase changes periodically and neurons can encode sound phase with high accuracy while firing at different times during the sound. Instead, by reliable spike timing, we refer to the likelihood of spikes to fire at a given time in different presentations of the same sound (). Under the assumption that auditory neurons respond to power increases within their preferred frequency band, high reliability indicates that the firing pattern could encode information about the spectral structure and envelope of a sound.
Spike trains were analyzed using SACs (Joris et al., 2006
). The height of the SAC at zero time lag quantifies the likelihood that spikes will occur at the same time during a stimulus when it is presented repeatedly. We found that SAC peak heights were significantly larger in NA than in NM (; medians significantly different by Kruskal-Wallis test, p < 0.001). We found that SAC peak height was inversely correlated with STRFbf
in both NA and NM (NA regression: −0.00072x + 6.8, R = 0.42, p<0.05; NM regression: −0.0002x + 2.7, R = 0.72, p<0.001, within 95% confidence bounds). To assess whether the difference in SAC peak height was due to a bias in the frequency range of neurons sampled in NA vs. NM, we compared the regression lines between best frequency and SAC peak height in both populations of cells. These regressions were significantly different (p < 0.001, t-test), indicating that for a given frequency NA units SAC peaks will be significantly larger than those observed in NM.
Figure 3 Quantification of response reliability (SAC peak height) and STRFsf across population data. (A) SAC peak heights (in units of normalized number of coincidences) are significantly larger in NA units (medians significantly different by Kruskal-Wallis test, (more ...)
To verify the findings of the SAC metric using an additional measure of response reliability, we also computed the spike-train distance as described by Victor and Purpura (1996)
. Consistent with the SAC analysis, spike-train distances were significantly smaller in NA compared to NM (Supplemental Figure 1A
; medians significantly different by Kruskal-Wallis test, p < 0.001). These results indicate that the firing pattern in NA is more invariant than in NM when the same stimulus is repeated.
Comparison of STRFs of neurons in NA vs. NM
To evaluate differences between NA and NM units’ STRFs (n = 36 and n = 53, respectively), we began by quantifying the STRFbf, STRFbw and STRFtw. We found a significant positive correlation between STRFbw and STRFbf in NM and NA (NM regression: 0.089x + 37, R = 0.69, p < 0.001; NA regression: 0.15x + 220, R = 0.36, p < 0.05). When plotted together, samples of the two populations of neurons overlap (not shown), but their regressions are significantly different (t-test, p<0.001). We also found an inverse correlation between STRFbf and STRFtw in both nuclei (NM regression: − 0.0002x + 2.2, R= −0.74, p < 0.001; NA regression: −0.00019x + 2.2, R= 0.72, p < 0.001). These regressions were not significantly different.
A conspicuous difference observed is that excitatory subfields of STRFs in NA were generally preceded by a suppressive field (; STRFsf), which was generally absent in NM units’ STRFs (). When quantifying the STRFsf, we found that its magnitude across NA units is in fact significantly larger than across NM units (; medians significantly different by Kruskal-Wallis, p<0.001). This STRF property is consistent with NA neurons being sensitive to the onset of power transients within the frequency band that they are tuned to (discussed below).
We used the neurons’ STRFs, computed with unfrozen noise, to predict the response to stimuli presented during the frozen-noise protocol and compared predicted and observed PSTHs. We found that the correlation coefficients between the predicted and observed PSTHs were significantly higher for NA units’ STRFs (median: 0.7) than for NM units’ STRFs (median: 0.62, ; medians significantly different by Kruskal-Wallis, p < 0.001). This indicates that linear filters, such as STRFs, are a better descriptor of NA neurons’ response behavior than they are for NM units.
Figure 4 STRFs more accurately predict the response of NA units than NM units. Example of observed (black) and predicted (red) PSTHs, given in units of normalized spike counts (Norm. Spike Count), from NA (A, correlation coefficient = 0.85) and NM (B, correlation (more ...)
Relationship between STRF properties and response reliability
In NA, we found a significant correlation between STRFsf and SAC peak height (; regression: 17x + 1.9, R=0.62, p < 0.001), as well as between STRFbf and SAC peak height (regression: −0.00075x + 7.4, R = 0.32, p = 0.05). In NM we found that both STRFbw (regression: −0.0011x + 2.2, R=0.43, p<0.01) and STRFbf (regression: −0.0002x + 2.7, R = 0.72, p<0.001) were correlated with their SAC peak height. Because many of these parameters co-vary, we used multilinear regression analysis to determine which STRF features had the most power in modulating SAC peak height when considered together. For this analysis, STRFsf, STRFbw and STRFbf were normalized by their within-sample maxima. In NA, SAC peak was significantly influenced by STRFsf and STRFbf, where STRFsf was the dominant factor (). In NM, SAC peak was significantly affected only by STRFbf (). The lack of effect that STRFsf has on NM units’ SAC peak is likely due to STRFsf values spanning a very small range in NM. Also, it should be noted that STRFbf exerts a much more prominent effect on NA SAC peaks than on NM SAC peaks. This indicates that STRF features play a much greater role in modulating SAC peaks in NA than NM.
Coefficients of multiple linear regression analysis
Modeling the effects of STRF properties on response reliability
To assess the effect of varying specific STRF parameters on the response reliability of a neuron, we developed a simple model using artificial STRFs which were convolved with a stimulus, allowing us to generate artificial spike trains from the resulting output (). We varied parameters of interest in the modeled STRFs to observe how they affected the reliability of the neural response, which was quantified using SACs.
Our primary interest was the effect of the magnitude of STRFsf
on SACs. This suppressive field is consistent with sensitivity to the onset of power transients in the neuron’s preferred frequency band, making it more selective than a neuron without a suppressive field. To test whether greater spectrotemporal selectivity, as indicated by the presence of STRFsf
, could account for greater response reliability, we ran the model varying STRFsf
while holding STRFbw
constant. Consistent with the observations in the NA units, we found that increasing the magnitude of the suppressive field created more patterned PSTHs and rasters () and enhanced the reliability of spike trains, increasing the SAC peak height (). Although firing rates do change with varying STRFsf
, the SAC is normalized with respect to the firing rate and this does not represent a confound (see Methods). Spike-train distances also decreased with larger STRFsf
(Supplemental Figure 1B
). Overall, these results indicate that STRFsf
can account for enhanced response reliability observed in NA.
Figure 6 Modeling the effects of STRF suppressive field on response reliability. (A) Representative STRFs (top), PSTHs (middle) and corresponding rasters (bottom) at suppressive field magnitudes (STRFsf) equal to, from left to right, 0, 0.3, and 0.6. PSTHs are (more ...)
Another difference between NM and NA units which could affect reliability is the larger STRFbw
observed in NA. Our model showed that increasing STRFbw
decreased the SAC peak height (Supplemental Figure 1C
). This is consistent with our observation of an inverse correlation between SAC peak height and STRFbw
in NM (see previous section). Furthermore, spike train distances increased with increasing STRFbw
(Supplemental Figure 1D
Our model simulated neurons with different refractory periods under the same STRF parameters. The refractory period did not grossly affect the relationship between SAC peak height and STRFsf (). The refractory periods estimated from ISIHs in our dataset were not significantly different between NA and NM. However, NA units’ larger STRFbw as compared to NM should, according to the model, decrease their response reliability. We observed that despite this difference, NA units’ response reliability is still significantly larger.
Finally, we tested how well our model could predict the relationship between STRFs and response reliability in the data. We used the model to generate spike trains from the STRF of each NA unit and computed SACs on these data. We found a correlation between the SAC-peak height obtained from our in vivo data and those predicted by our model (R = 0.54, p = 0.001, ). The model, however, tended to overestimate the SAC peak height; this is to be expected, as our model did not incorporate any noise mechanism and all simulated spiking activity is purely stimulus driven. Despite the expected overestimation, this result indicates that our model captures the overall relationship between STRF shape and reliability.