Persistent Gamma Spiking in SI Nonsensory Fast Spiking Cells Predicts Perceptual Success

Abstract

Gamma oscillations (30 – 55Hz) are hypothesized to temporally coordinate sensory encoding, enabling perception. However, fast spiking interneurons (FS), a key gamma generator, can be highly sensory responsive, as is the gamma-band local field potential (LFP). How can FS-mediated gamma act as an impartial temporal reference for sensory encoding, when the sensory drive itself presumably perturbs the preestablished rhythm? Combining tetrode recording in SI barrel cortex with controlled psychophysics, we found a unique FS subtype that was not sensory responsive, and spiked regularly at gamma-range intervals (Gamma Regular Nonsensory FS, grnsFS). Successful detection was predicted by a further increase in gamma regular spiking of grnsFS, persisting from before to after sensory onset. In contrast, broadband LFP power including gamma negatively predicted detection, and did not cohere with gamma-band spiking by grnsFS. These results suggest that a distinct FS subtype mediates perceptually relevant oscillations, independent of the LFP and sensory drive.

Graphical Abstract

eTOC Blurb

Gamma oscillations are hypothesized to temporally coordinate sensory encoding. Shin and Moore found a distinct subtype of fast-spiking interneurons (FS) in SI that are non-sensory responsive and spike regularly at gamma intervals. These “gamma regular non-sensory FS” could potentially mediate perceptually relevant oscillations, independent of the LFP and sensory drive.

Introduction

Gamma oscillations (30 – 55Hz) are evident in neocortex during behaviorally relevant neural computations, including perception (Fries et al., 2001; Gray et al., 1989; Siegle et al., 2014), cognition (Cho et al., 2015; Kim et al., 2016) and action (Niell and Stryker, 2010). Decreases in gamma expression are a biomarker for numerous neurological disorders (Iaccarino et al., 2016; Lewis et al., 2005, 2012; Uhlhaas et al., 2008). These studies suggest a role for gamma in successful information processing. However, there is intense debate whether this dynamic plays a meaningful computational role, or is simply a byproduct of local resonant excitatory-inhibitory circuits that are active during such behaviors (Ray and Maunsell, 2010, 2015; Shadlen and Movshon, 1999; Xing et al., 2012).

Much of the interest in gamma stems from its hypothesized role in binding. This hypothesis proposes that gamma coordinates action potential timing across a long-range network (Gray et al., 1989; Singer, 1993; Singer and Gray, 1995). This alignment across brain areas has been hypothesized to “bind” the activity of participating cells into a coherent perception of a single entity. The binding hypothesis necessitates that gamma oscillations serve as a “clock” for coordinating sensory processing. Several predictions follow: First, this temporal reference must emerge prior to, and persist unperturbed by, stimulus presentation. Second, this oscillation should be expressed independent of specific stimulus features (Singer, 1993; Singer and Gray, 1995).

Considerable challenges have been raised to the binding-by-gamma view. Using the local field potential (LFP) as a metric for gamma, studies in visual neocortex have shown that subtle changes in visual stimulus features, such as contrast, can change gamma frequency and power (Ray and Maunsell, 2010). This sensitivity undermines the possibility that a sustained coordinating reference can be indexed by LFP gamma. Moreover, differences in LFP gamma frequency and power can be observed across receptive fields driven by a common object, further undermining its role in lateral binding (Ray and Maunsell, 2010, 2015).

Another logical challenge to the binding hypothesis concerns the mechanisms of gamma origin. Neocortical gamma is typically found to depend on coordinated spiking of fast spiking interneurons (FS) that create highly effective inhibition lasting ~20ms (Buzsáki and Wang, 2012; Cardin et al., 2009; Tiesinga and Sejnowski, 2009; Whittington et al., 2000; but see also Veit et al., 2017). A brief “window of opportunity,” permissive for temporally aligned spiking of projection neurons, is open when inhibition is at its lowest. The resulting alignment of projection neurons’ spiking can lead to an efficient relay of signals to downstream targets, and the onset of another gamma cycle (Borgers and Kopell, 2003; Cardin et al., 2009; Chen et al., 2017; Knoblich et al., 2010; Moore et al., 2010; Sohal, 2016; Sohal et al., 2009). However, in primary sensory areas of the neocortex, such as SI barrel cortex, FS are typically highly sensitive to sensory stimulation, responding at short latencies with high temporal precision (Andermann et al., 2004; Cruikshank et al., 2007; Simons, 1978; Simons and Carvell, 1989; Swadlow, 2003). These results pose a conundrum, as FS sensitivity to sensory drive suggests that the reference gamma rhythm initially established prestimulus would be perturbed upon sensory input. That is, the rhythm is disrupted by the very signal that this rhythm is seeking to temporally organize.

To better understand the relationship between FS and gamma during sensory processing, we recorded extracellularly in SI barrel cortex using chronic tetrode implants while mice performed well-controlled vibrissal deflection detection. This task was previously shown to benefit from the optogenetic induction of FS synchrony at gamma (Siegle et al., 2014). Here, we found that ~40% of FS were sensory responsive (sFS), encoding stimulus amplitude faithfully, and with a precise latency to first spike following sensory onset. The spiking behavior of sFS adheres to a Poisson process with a spike refractory period. This Poisson-like spiking behavior has been described as the typical behavior of most cortical neurons (Shadlen and Newsome, 1998). The spiking behavior of nonsensory FS (nsFS), on the contrary, categorized into two distinct clusters. One group (~30% of FS) fired in a Poisson-like manner (Poisson-like nsFS, plnsFS), similar to sFS. The other group (~30% of FS) showed inter-spike interval (ISI) distributions that peaked in the gamma range (18 – 33ms), and had higher ISI regularity (Gamma Regular nsFS, grnsFS). The gamma regular spiking behavior of grnsFS persisted from pre- to poststimulus, unperturbed by the sensory drive. On successful detection trials (hits), the grnsFS showed even higher regularity and higher occurrence of gamma interval spiking. In addition, the grnsFS showed the highest within-subtype synchrony. While these characteristics make grnsFS spiking a candidate reference rhythm for binding, grnsFS had the lowest spike-field coherence (SFC) of FS subtypes in the gamma band, and showed a negative correlation between gamma band spike spectral power and LFP power. Accordingly, the broadband LFP power, including the gamma band, predicted non-detection (misses) in the detection task. These data provide the first description of a functional FS subtype, the grnsFS: They provide a persistent oscillation independent of the LFP and the sensory input, and their gamma regular spiking may enhance sensory processing by providing an impartial temporal reference signal.

Results

~40% of FS in SI Barrel Cortex were Sensory Responsive (sFS)

To investigate the relationship between endogenous FS activity and perception, we recorded extracellularly using chronic tetrode implants in SI barrel cortex, from mice trained to respond to vibrissal deflection (Figure 1A). Data presented are from sessions with high-quality psychometric behavior (4 mice, 128 sessions; Figure 5A). Spike width, quantified as the time between peak and trough of the extracellular spike waveform, provided high separability of FS (trough to peak time ≤0.4ms; N=188) from the broader population of putative excitatory regular spiking units (RS; Methods, Figure S6A).

Figure 1. ~40% of FS in SI Barrel Cortex were Sensory Responsive (sFS): The sFS Evoked Response was Faithful to Stimulus Amplitude and had Precise Onset Timing.

A. Extracellular electrophysiology was recorded from SI barrel cortex of mice performing a head-fixed vibrissal deflection detection task (N=188 FS; 4 mice, 128 sessions). Vibrissae were clamped to a piezoelectric bender, which delivered deflections at 20Hz for 500ms at varying amplitudes on each trial. The peristimulus time histogram (PSTH) is shown for maximal stimulus amplitude trials, averaged across all sensory FS (sFS, N=69).

B. The area under receiver operating characteristic curve in this ideal observer analysis is referred to as the stimulus probability (SP); SP>0.5 indicates sensory driven rate increase, whereas SP<0.5 indicates suppression. Significance was determined by 95% bootstrapped confidence interval (CI): sFS (N=69) in gray, nsFS (N=114) in green. Only N=5 FS showed significant suppression and were not analyzed further (black).

C. Neurometric curves plot sensory evoked rate in the first 100ms relative to sensory onset (i.e., report window onset), as a function of stimulus amplitude bins. Error bars indicate standard error of the mean (SEM). The sFS neurometric curve (left) monotonically increased as the stimulus amplitude increased, whereas the nsFS neurometric curve showed a flat relationship.

D. Histogram of first spike latencies across FS in each subtype. For each FS, first spike latency is defined as the X-value at the peak of the first spike latency distribution across maximal stimulus amplitude trials (this distribution is shown in E; brightest magenta curve).

E. First spike latency distributions are shown for 4 stimulus amplitude bins (brighter colors for stronger stimulus amplitude bins). Histogram counts were calculated in 5ms bins, 1ms sliding steps, then normalized by dividing by the bin length (5ms) and the trial count. Stronger stimulus amplitude resulted in a more reliable timing of the first spike after sensory onset.

Figure 5. The Broadband LFP Including Gamma Predicted Non-Detection: FS Subtypes Showed Distinct Task-Predictive Rate Dynamics.

A. Psychometric behavior showed increased hit rate with increasing sensory stimulus amplitude (N=128 sessions; thin gray lines indicate each session, thick black the mean). Mice were trained to report detection of the vibrissal stimulus by licking for water reward within 700ms of the sensory onset (report window). For each session, trials were sorted by increasing stimulus amplitude, and hit rate was calculated for 51 trial bins, slid in 1 trial steps. Trials where the mouse was disengaged from the task (d’<1) were excluded from this analysis (Methods).

B. Left Reaction time quartiles across sessions are represented for three stimulus amplitude bins (brighter color for stronger stimulus amplitude bins). Reaction time is defined for hit trials as the time of the first lick following sensory onset. Although the reaction time accelerated with stronger stimulus amplitude (Kruskal Wallis test p=1.65×10⁻⁶¹; *p<0.05 with post-hoc Bonferroni correction for multiple comparisons), the median reaction time was similar across amplitude bins, ranging between ~250 – 280ms.

Right The cumulative distribution function (CDF) of reaction times are represented for three stimulus amplitude bins (brighter colors for stronger stimulus amplitude bins). The CDF showed a sharp reaction time probability inflection at 100ms, motivating the choice of poststimulus analysis window.

C. To compare neural activity on hits (blue) and misses (red) independent of sensory features, we matched hits and misses in stimulus amplitude and number of trials (Methods). The median stimulus amplitude of the matched trials was defined as the perceptual threshold, and was used as the cutoff between subthreshold and suprathreshold bins. Unless otherwise noted, hits and misses denote stimulus amplitude matched hits and misses.

D. The hit minus miss LFP spectrogram is shown, averaged first across every tetrode within each session, then across N=128 sessions. Multitapered spectral power was calculated in 250ms windows, sliding in 25ms steps.

E. Hit rate as a function of LFP power is shown for 3 frequency bands in the prestimulus (−250 – 0ms, left) and the poststimulus (0 – 100ms, right) period. For each session and frequency band, trials were sorted based on LFP power averaged across all tetrodes. Higher power corresponded to lower hit rate in all frequency bands.

F. (i) The PSTH on stimulus amplitude matched hit (blue) and miss (red) trials is shown, averaged across FS in each subtype.

(ii) Prestimulus (−250 – 0ms) and poststimulus (0 – 100ms) firing rate quartiles are shown, across FS in each subtype, for hits and misses (*p<0.05 Wilcoxon signed rank test, Bonferroni correction for multiple comparisons). The sFS showed lower rate prestimulus and higher rate poststimulus on hits vs misses, whereas grnsFS showed elevated rate persisting from before to after sensory onset.

We grouped FS into sensory (sFS) and nonsensory (nsFS) based on an ideal observer analysis for discrimination of 0 – 100ms poststimulus firing rate on maximal stimulus amplitude trials vs zero amplitude catch trials (Figure 1B). We chose this 100ms poststimulus window because in our detection task, mice started reporting detection ~100ms following sensory onset (Figure 5B). The sFS constituted ~40% of recorded FS (N=69/188; alternative definitions of sensory responsiveness identified largely overlapping units; Figure S1).

The firing rate of sFS encoded stimulus amplitude faithfully, increasing monotonically with intensity (Figure 1C). In addition, the latency to first poststimulus spike following sensory onset was consistent, concentrated ~10 – 15ms poststimulus (Figure 1D-E). The sFS units typically responded with only a single evoked spike above baseline firing rate before the mouse reached a decision (Figure 1C), suggesting that this initial response is most likely a crucial carrier of the feedforward detection signal. Defining sensory responsiveness by either rate (sFS, N=69) or first spike latency (slFS, N=53) yielded highly overlapping FS (N=49 overlap; Figure S1iv). In contrast, the evoked rate of nsFS showed a flat relationship with the stimulus amplitude (Figure 1C), and did not show an alignment of spike timing following sensory onset (Figure 1D-E).

Distinct Subtypes of FS Emerged from Temporal Spiking Patterns, with a Subtype of Nonsensory FS Showing Highly Regular ISIs that Occurred Most Often in the Gamma Range (grnsFS)

To investigate spontaneous spiking behavior in the absence of sensory drive, we plotted the ISI distributions of FS during the 1s prestimulus period. Most sFS showed an ISI distribution with a peak shortly after the spike refractory period (≤15ms), and an exponential fall-off with increasing ISI (Figure 2A-B gray). This distribution has been described as typical of most cortical neurons (Shadlen and Newsome, 1998), and is often referred to as a Poisson process with a refractory period (Kass et al., 2014).

Figure 2. Distinct Subtypes of FS Emerged from Temporal Spiking Patterns, with a Subtype of Nonsensory FS Showing Highly Regular ISIs that Occurred Most Often in the Gamma Range (grnsFS).

The FS subtypes are color-coded throughout: sFS (N=69, gray), grnsFS (N=60, teal), and plnsFS (N=54, olive). ISI probability distribution function (PDF) and coefficient of variation (CV) were calculated for each FS, from ISIs in the 1s prestimulus period, and pooled across all trials. ISI PDFs were calculated in 5ms bins, sliding in 1ms steps.

A. The peak of the ISI PDF was calculated for each FS. The histogram of ISI peaks (3ms bins) for nsFS revealed a gap between 15 – 18ms, separating their firing patterns into two subtypes. The subtype with peak ISIs predominantly between 18 – 33ms was named Gamma Regular nsFS (grnsFS; teal), reflecting the overlap in spike intervals and gamma frequencies (30 – 55Hz). Those with peaks at shorter ISIs (<15ms) were named Poisson-like nsFS (plnsFS; olive). The sFS (gray) also had Poisson-like ISIs.

B. The ISI PDF was averaged across FS within each subtype; revealing distributions with an exponential fall-off for plnsFS and sFS, and a distribution peaking in the gamma band for grnsFS.

C. The histogram of CV values (bin size 0.05) showed greater regularity in baseline firing for grnsFS, compared to plnsFS or sFS. Quartiles of the CV distribution across FS in each subtype is plotted on top (Kruskal Wallis test p=4.91×10⁻²⁰; *p<0.05 with post-hoc Bonferroni correction for multiple comparisons).

D. K-means clustering was performed on the two-dimensional ISI peak – CV matrix, for sFS and nsFS separately. Akaike information criterion (AIC) was calculated to determine the number of clusters that best described the data. The AIC is lowest for 1 cluster in sFS, and 2 clusters for nsFS.

E. Scatterplot shows the K-means clustering with minimal AIC for sFS (left) and nsFS (right). The clustering results of nsFS were identical to the definition in A.

When analyzing nsFS, a distinct subtype of nsFS that did not adhere to these typical patterns emerged from cluster analysis. This group is evident in the histogram of ISI peaks, where approximately half of nsFS showed a longer ISI peak concentrated in the 18 – 33ms range (N=60; teal), separated from the shorter Poisson-like ISI peaks (N=54; olive) with a gap between 15 – 18ms (Figure 2A-B). The ISI peak values did not change with the length of the analysis window, indicating robustness of this separation (Figure S3A). The group with longer ISI peaks showed distinctly higher ISI regularity (lower coefficient of variation, CV), with the majority having CV<1 (Figure 2C). Note, a Poisson process has a CV of 1, and lower CV indicates a more rhythmic processes (Kass et al., 2014).

To quantitatively address the number of subtypes that optimally describe the spiking behavior of sFS and nsFS, we performed K-means clustering on the ISI peak – CV space. The Akaike Information Criterion (AIC) was quantified subsequently, while systematically varying the number of clusters (Figure 2D). The AIC estimates the information lost by a given model, hence its minimum is the preferred model. The sFS were best modeled as a single cluster, whereas the nsFS were best modeled as two clusters (Figure 2E). The preferred clustering assignments of nsFS was identical to the two groups that were qualitatively evident in the ISI peak histogram (Figure 2A). We refer to the group with shorter ISIs and less regular firing patterns as “Poisson-like Nonsensory FS” (plnsFS; teal). The ISI peaks of the other nsFS subtype (18 – 33ms) correspond to a period of a gamma rhythm (30 – 55Hz). To reflect the criteria that separated this group, i.e., regular spiking at gamma intervals, we labeled this group the “Gamma Regular Nonsensory FS” (grnsFS; olive).

To further probe rhythmicity in grnsFS spiking, we asked whether ISIs in FS were independent and identically distributed (i.e., a renewal process; Figure S2). All three FS subtypes deviated significantly from renewal processes, indicating that FS ISIs depended on recent history (Figure S2B). Specifically, consecutive ISIs were likely to be more similar than expected by the renewal assumption, indicating a rhythmic tendency (Figure S2E-F).

We defined sensory responsiveness based on stimulus evoked rate changes, which left open the possibility that the temporal pattern of nsFS spiking could be altered by sensory drive. To test this possibility, we assessed the impact of sensory drive on various metrics, by comparing the response to maximal amplitude stimuli with prestimulus activity (Figure 3). As expected, sFS showed changes in every metric examined, including CV (Figure 3B), ISI distribution (Figure 3C-D) and broadband spike spectral power (Figure 3E-F). Likewise, LFP showed a broadband power increase in response to sensory drive (Figure 3A). In contrast, sensory drive did not alter any of these metrics in either nsFS subtype (Figure 3B-F), confirming that neither rate nor temporal pattern changed in nsFS spiking.

Figure 3. Temporal Spiking Patterns Did Not Change from Before to After Sensory Onset in nsFS.

Poststimulus (0 – 100ms, magenta) vs prestimulus (−100 – 0ms, black) comparisons were made for maximal stimulus amplitude trials. Note, the lower bound in the frequency domain was limited by the Nyquist frequency, i.e., 20Hz.

A. Left Power spectral density of the LFP is shown. The LFP power was first averaged across all tetrodes for each session: Mean ± SEM across sessions is represented in log-log scale (N=128).

Right LFP quartiles across sessions are shown for 3 frequency bands. Across all frequency bands, LFP power was higher poststimulus than prestimulus (*p<0.05 Wilcoxon signed rank test, Bonferroni correction for multiple comparisons).

B. CV quartiles across FS in each subtype are shown. CV was calculated with ISIs ending poststimulus (the second spike of each ISI occurred 0 – 100ms relative to sensory onset; magenta), compared to ISIs ending prestimulus (−100 – 0ms, black). The sFS was the only subtype with a significant difference (post>pre; *p<0.05 Wilcoxon signed rank test, Bonferroni correction for multiple comparisons).

C. ISI rate distributions for ISIs ending poststimulus (magenta) vs prestimulus (black), plotted in the logarithmic inverse scale to facilitate direct comparisons with the spike spectral power (E-F). ISI rate was calculated for each 5ms bin, slid in 1ms steps. Mean ± SEM across FS of each subtype is shown.

D. For each FS, 95% CI for post-pre difference in ISI rate distributions was determined by permuting pre and post labels amongst maximal stimulus amplitude trials 1000 times. For each 5ms ISI bin, the number of units whose post-pre difference in ISI rate distributions exceeded (magenta) or fell below (black) the 95% CI were accumulated across FS of each subtype, and plotted as percent significant units. The expected false discovery rate was 2.5% for both curves, across all ISI bins.

E. Power spectral density of spikes are shown. Mean ± SEM across FS of each subtype is plotted, with the frequency domain (X-axis) in logarithmic scale (sFS left; grnsFS center; plnsFS right).

F. Percent units that exceeded (magenta) or fell below (black) the 95% shuffled CI of post-pre difference are shown. Data are presented as in D, but for spike spectral density.

The grnsFS Showed the Highest Synchrony but had the Lowest Coherence with the LFP Gamma

A variety of evidence has linked FS synchrony to gamma band coordination of pyramidal firing and LFP power (Tiesinga and Sejnowski, 2009). The fact that grnsFS showed ISI peaks in the gamma range suggested that this subtype might show high within-subtype synchrony. To test this prediction, we examined the cross-correlogram (CCG), corrected for rate effects (Figure 4A). Intuitively, neurons with higher firing rate have a higher chance of spiking synchronously with other neurons. As such, rate effects must be corrected for in order to look at true synchrony beyond what is expected by chance. In this rate-corrected CCG, a 0ms value is equivalent to rate-excess synchrony. As predicted, rate-excess synchrony was highest for the grnsFS subtype (Figure 4B). This result suggests that grnsFS are distinctively positioned to coordinate local spiking dynamics through their synchronous activity.

Figure 4. The grnsFS Showed the Highest Synchrony but had the Lowest Coherence with the LFP Gamma.

A. Cross-correlograms (CCG) were calculated for FS pairs within each subtype (Mean ± SEM). Pairs recorded from different tetrodes were used for analyzing synchrony. Spike trains (1ms bins) of simultaneously recorded FS pairs were slid against each other, and the resulting CCG was smoothed with 5ms boxcar windows. Chance synchrony expected from slow (50ms) fluctuations in rate were corrected using the jitter-based method (50ms jitter window; Amarasingham et al., 2012).

B. The distribution of rate-excess synchrony (CCG at 0ms lag) is shown for FS pairs within each subtype (Kruskal Wallis test p=7.08×10⁻⁴; *p<0.05 with post-hoc Bonferroni correction for multiple comparisons). Rate-excess synchrony was highest for grnsFS pairs.

C-D. For all spike-LFP relationships, the LFP was averaged from 3 neighboring tetrodes relative to the tetrode where spikes were recorded.

C. Left Spike triggered LFP for each FS was averaged across all spikes that occurred in the 1s prestimulus period, across all trials (STA). The STA averaged across FS in each subtype showed a distinct pattern for grnsFS (teal) compared to sFS (gray) and plnsFS (olive).

Right The histogram of STA slopes leading up to the spike (T=0ms) is shown. Quartiles across FS in each subtype is plotted on top (Kruskal Wallis test p=4.02×10⁻⁵; *p<0.05 with post-hoc Bonferroni correction for multiple comparisons). The STA slopes leading up to spikes were negatively (<0μV/ms) skewed in sFS and plnsFS, but not in grnsFS.

D. Left Power spectral density of the spike-field coherence (SFC) was calculated for the 1s prestimulus period, then averaged across all trials. Mean across FS of each subtype is shown in log-log scale (sFS gray; grnsFS teal; plnsFS olive).

Right SFC quartiles across FS in each subtype are shown for 4 frequency bands; from left to right, 8 – 15Hz alpha band; 15 – 30Hz beta band; 30 – 55Hz gamma band; 66 – 120Hz high-gamma band (Kruskal Wallis test p=0.143 for alpha band, p=5.59×10⁻⁵ for beta band, p=9.67×10⁻⁴ for gamma band, and p=0.165 for high-gamma band; *p<0.05 with post-hoc Bonferroni correction for multiple comparisons). sFS had the highest SFC in the beta band (15 – 30Hz), whereas grnsFS had the lowest SFC in the gamma band (30 – 55Hz).

Given these observations, we further expected grnsFS to drive gamma oscillations in the LFP. To our surprise, we found that grnsFS are largely decoupled with the LFP (Figure 4C-D). First, the spike-triggered average of the LFP (STA, Figure 4C) is expected to show a negative change in LFP voltage right before a spike (Fries et al., 2001; Hasenstaub et al., 2005), as its downward slope is thought to indicate network depolarization (Buzsáki et al., 2012). The sFS and plnsFS followed this convention but grnsFS did not. The atypical shape of grnsFS STA suggests that grnsFS do not require concurrent depolarization from local sources to spike.

The decoupling of grnsFS from the LFP was also evident when examining spike-field coherence (SFC; Figure 4D). The grnsFS generally had the lowest SFC, with the subtype comparisons being significant in the 30 – 55Hz gamma band. Further, grnsFS showed a negative Spearman correlation between LFP and spike spectral power in the gamma band (Figure S3Fi). In contrast, sFS generally had the strongest relationship with the LFP, both for SFC and for power-power correlation. These results provide evidence against the prediction that grnsFS drives LFP gamma.

Several other features distinguished grnsFS from other FS (Figure S4). Differential characteristics included modest but significant differences in spike waveform and baseline firing rate (median firing rate: sFS = 10.1Hz; grnsFS = 14.0Hz; plnsFS was 8.4Hz; Figure S4B). Baseline firing rate does not correlate with CV across FS, suggesting that between-subtype differences in baseline firing rate and CV are largely independent effects (Figure S3C). The grnsFS also had a distinct spike spectral density profile, where spike spectral power had a tendency to increase with increasing frequencies (Figure S4Gi). Further, spike spectral power was highest for grnsFS in both the gamma and the high-gamma band (Figure S4Gii). On a subset of trials for which running information was available in our data, we also examined the relationship between locomotion and FS subtype. Again, significant differences were observed between subtypes: Locomotion predicted rate increases in sFS but slight suppression in grnsFS, and CV increases in grnsFS but not in sFS (Figure S5).

The nsFS separated into 2 clusters, across several alternative definitions of sensory responsiveness (Figure S1), and regardless of the order of definition (i.e., clustering first and then conditioning on nonsensory responsiveness; Figure S6A). To test the generality of other distinguishing features of these FS subtypes, we analyzed an independent dataset, also from SI barrel cortex with vibrissal deflections (202 FS; 121 recording sessions; 6 mice). Clustering results, as well as key features of the three FS subtypes, were replicated in the second dataset (see Methods and Figure S6 for details).

KEY RESOURCES TABLE

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Bacterial and Virus Strains
rAAV5-hSyn-Con/Fon--hChR2(H134R)-eYFP-WPRE	Vector Core Facility at University of North Carolina Chapel Hill	AV6149B
Deposited Data
Data deposited at Mendeley Data	This paper	http://dx.doi.org/10.17632/r5tbz5j34p.1
Experimental Models: Organisms/Strains
Mouse: C57/B6J	Jackson Laboratories	000664
Mouse: Slc32a1-IRES-Cre	Jackson Laboratories	016962
Mouse: Parvalbumin-2A-Flpo	Jackson Laboratories	022730
Software and Algorithms
MATLAB	MathWorks	https://www.mathworks.com
Simpleclust	J. Voigts	https://github.com/open-ephys/simpleclust
Chronux	P. Mitra, H. Bokil	Chronux, RRID:SCR_005547
Open Ephys Acquisition GUI	Open Ephys	https://open-ephys.atlassian.net/
Custom written behavior control software (MATLAB based)	J.H. Siegle, D.L. Pritchett, H. Shin	https://github.com/moorelab/behaviorcontrol
Custom written analysis software (MATLAB based)	This paper	https://github.com/hs13/GammaRegularNonSensoryFastSpikingNeurons
Other
Nickel-Chrome 0.012mm wire	Sandvik Kanthal HP Reid Precision Fine Tetrode Wire	https://www.amazon.com/Sandvik-Precision-Fine-Tetrode-Feet/dp/B0062MNUG6
Polyimide tubing 33ga and 37/38ga	Smallparts	http://www.smallparts2.com/Product-Categories.html
Skull Screw M0.6X0.5mm	Antrin Miniature Specialties	https://antrinminiature.com/
Open Ephys Acquisition Board	Open Ephys	http://www.open-ephys.org/store/eux9baf6a5s8tid06hk1mw5aafjdz1
NI-DAQ BNC-2110, PCI-6220	National Instruments	http://sine.ni.com/nips/cds/view/p/lang/en/nid/1865
Solenoid valve 225PNC1-11	NResearch	https://www.nresearch.com/
Rotary Encoder - 1024 P/R (Quadrature)	Sparkfun	https://www.sparkfun.com/products/11102
Piezoelectric bender CMBP09	Noliac	http://www.noliac.com/products/actuators/plate-benders/show/cmbp09/

The Broadband LFP Including Gamma Predicted Non-Detection

We next tested for predictive relationships between neural activity and behavioral outcome. To assess detection-predictive effects independent of sensory effects, we compared stimulus amplitude matched hits and misses (Figure 5A-C). First, we tested LFP power across frequency bands (Figure 5D-E, Figure S7i). We observed a generic decrease in LFP across all frequencies on hits vs misses. This negative relationship is expected for alpha and beta bands (Jones et al., 2010). However, the extension of this negative behavioral-LFP relationship to gamma and high-gamma bands contrasts with reports that have linked LFP gamma power with increased attention or perceptual/cognitive engagement (Aoki et al., 2001; Fries et al., 2008; Hoogenboom et al., 2010).

Next, we examined the differential firing rate of FS subtypes on hits vs misses (Figure 5F, Table S1). The sFS predicted hits by lower rate prestimulus (−250 – 0ms) and higher rate poststimulus (0 – 100ms). In contrast, grnsFS predicted hits with persistently higher rate from pre- to poststimulus. The plnsFS firing rate did not show significant differences between hits and misses both pre- and poststimulus. To determine the predictive strength of each metric for each FS subtype, we calculated the detect probability (DP; DP>0.5 indicates higher values predict hits, DP<0.5 predict misses; ideal observer analysis). The median, lower and upper quartile of each distribution is reported in Table S2.

The grnsFS Predicted Successful Detection with Enhanced Gamma Regular ISIs that Persisted from Before to After Sensory Onset

The grnsFS were defined as having ISIs that peak in the gamma range and are more regular (lower CV). We asked whether dynamic fluctuations in these cardinal features predicted perceptual success. Across all 3 FS subtypes, CV was significantly lower on hits than misses in the prestimulus period (Figure 6A, Figure S7Aii). Greater ISI regularity persisted in grnsFS, but not in sFS nor plnsFS (Figure 6A, Figure S7Bii). The CV values in these figures are lower compared to Figure 2 because of shorter analysis window (250ms vs 1s, Figure S3B).

Figure 6. The grnsFS Predicted Successful Detection with Enhanced Gamma Regular ISIs that Persisted from Before to After Sensory Onset.

A-B. CV (A), ISI rate distribution (log inverse scale, B) and spike spectral power (C) were computed in 250ms windows, sliding in 25ms steps. The X-axis denotes the center of the 250ms window. Sensory onset was included in the analysis window starting X=−125ms. See Figure S7 for associated statistics.

A. For each FS, CV was calculated for ISIs accumulated across all matched hits (blue) and misses (red) in each time window. Mean ± SEM across FS of each subtype is depicted. The grnsFS was the only subtype where CV was persistently lower on hits relative to misses from before to after the sensory onset. Poststimulus differences were non-significant for sFS and plnsFS.

B. The Hit - Miss difference in log inverse ISI rate distribution, averaged across FS in each subtype, is represented across time. The grnsFS showed enhanced gamma interval spiking that predicted successful sensory performance, beginning before stimulus onset and continuing poststimulus.

C. The Hit - Miss difference in the spike spectrogram, averaged across FS in each subtype, is represented. The frequency domain (Y-axis) is plotted in logarithmic scale.

Fluctuations in rate can influence fluctuations in CV (i.e., when ISI standard deviation does not scale with the mean). To test this possibility, we calculated for each FS the trial-by-trial Spearman correlation between rate and CV in the 1s prestimulus window (Figure S3D). Rate positively correlated with CV for most sFS and plnsFS, with average values of 0.16 (sFS) and 0.18 (plnsFS). In contrast, CV was not correlated with rate in most grnsFS, with the average correlation being 0.03. These differences in the rate and CV relationship are apparent when comparing rate differences in Figure 5F with CV differences in Figure 6A and Figure S7ii. Both rate and CV were lower on hits prestimulus for sFS, whereas differences in rate (hits>misses) and CV (hits<misses) predicted perceptual success in opposite directions for grnsFS pre- and poststimulus.

Next, we analyzed the hit vs miss differences in ISI rate distributions (Figure 6B, Figure S7iii-v) and spike spectral power (Figure 6C, Figure S7vi-viii). In sFS, ISI rate (count per second) and spike spectra power were lower prestimulus and higher poststimulus on hits across all ISIs and frequencies tested, mirroring rate effects (Figure 5F; see Figure S3Fii). These differences were not specific to a certain ISI range or frequency band.

For grnsFS, both ISIs and spike spectral power showed enhanced gamma band spiking on hit trials. The grnsFS ISI rate showed significant enhancement on hits persisting from pre- to poststimulus, in the gamma range (18 – 33ms) and longer (>33ms). Similarly, the hit-miss difference in spike spectral power was most prominent in the gamma range for grnsFS, with the spike spectral power from 30 – 55Hz being the only frequency band where the difference was significant both pre- and poststimulus.

In sum, across metrics, increased gamma band activity was the common feature of grnsFS firing that predicted perceptual success. The predictive behavior of gamma spiking persisted from pre- to poststimulus, and was accompanied by increased regularity in firing. Comparison of false alarms and correct rejects on catch trials are presented in Figure S8.

The grnsFS Spike Timing Relative to Sensory Onset Predicted Perceptual Performance with Gamma Rhythmicity

Prior work from our group showed that optogenetically driving FS at 40Hz can enhance perceptual success in the detection of naturalistic vibrissal motions (Siegle et al., 2014). Such FS optogenetic stimulation causes an increase in the precision of sensory-evoked spike timing in RS (Cardin et al., 2009; Siegle et al., 2014). These data offer a plausible explanation for the beneficial impact of gamma band stimulation: Reliable alignment of sensory evoked spike timing across pyramidal projection neurons would lead to coincident arrival on downstream targets, promoting effective sensory relay (Knoblich et al., 2010; Pritchett et al., 2015; Sohal, 2016).

When sensory onset was aligned to discrete phases of peristimulus 40Hz FS optogenetic stimulation, Siegle et al. (2014) further observed that optogenetic stimulation modulated perceptual success in a phase-specific manner. Perceptual performance was enhanced only when sensory onset was presented at a 12.5ms-lag relative to the last prestimulus optogenetic light pulse. An explanation for this phase-specificity is shown in Figure 7A: At 12.5ms lag (green curves), the initial lemniscal drive arriving at the SI barrel cortex (~8ms from sensory onset) coincides with the brief “window of opportunity” (green shading), where inhibition is at its lowest between 40Hz cycles of repetition. Similar phase-specificity in sensory gain modulation has been observed in non-human primates, supporting the generality of this mechanism (Ni et al., 2016).

Figure 7. The grnsFS Spike Timing Relative to Sensory Onset Predicted Perceptual Performance with Gamma Rhythmicity.

A. This schematic depicts how GABA_A-mediated inhibitory postsynaptic potentials (IPSP, rainbow colors), caused by the last prestimulus FS spike, aligns with the first poststimulus spike in SI sRS (thin dotted black line). If FS projecting to sRS spike at −12.5ms relative to sensory onset (green curve), the latency to the first sensory evoked spike in sRS coincides with the brief “window of opportunity” (green shading) between 40Hz cycles of inhibition. FS spiking before or after −12.5ms that repeats at a gamma interval would effectively suppress throughput.

B. Change in hit rate is plotted, conditioned on the last FS prestimulus spike timing. Mean ± SEM across all FS (left), and across each FS subtype (right three panels), are shown for each 5ms bin (e.g., −15 - −10ms bin for green). Friedman test was applied to test for an overall effect of last prestimulus spike timing, and the resulting p-value is denoted on each panel. Subsequently, pairwise comparisons were made between phases (*p<0.05 with Bonferroni correction for multiple comparisons).

C. Spike timing-based DP was calculated for 5ms bins, sliding in 1ms steps (see Methods). Mean ± SEM across FS of each subtype is shown. Colored lines were added at −22.5ms (purple), −17.5ms (blue), −12.5ms (green), −7.5ms (orange) and −2.5ms (red), to facilitate direct comparison with A. Prestimulus miss-prediction (DP<0.5) and poststimulus hit-prediction (DP>0.5) correspond to the rate changes described for sFS in Figure 5Fi. A peristimulus gamma-rhythmic modulation of spike timing-based DP was apparent for grnsFS.

D. Left To quantify the apparent rhythmicity of spike timing-based DP, we ran multitapered spectral analysis on the DP traces separately for each FS, for the −150 – 100ms peristimulus window relative to sensory onset. Mean ± SEM across FS of each subtype is plotted (sFS gray; grnsFS teal; plnsFS olive).

Right Spectral power of the peristimulus spike timing-based DP quartiles, across FS in each subtype, are shown for 4 frequency bands: 8 – 15Hz alpha band; 15 – 30Hz beta band; 30 – 55Hz gamma band; and, 66 – 120Hz high-gamma band (Kruskal Wallis test p=0.0079 for alpha band, p=0.0509 for beta band, p=9.59×10⁻⁴ for gamma band, p=6.35 ×10⁻⁴ for high-gamma band; *p<0.05 with post-hoc Bonferroni correction for multiple comparisons). Gamma rhythmicity in peristimulus spike timing-based DP was highest for grnsFS, compared to sFS and plnsFS.

To explore whether the timing of endogenous FS spikes relative to sensory onset influences detection outcome in a similar manner, we plotted the change in hit rate conditioned on last prestimulus spike timing (Figure 7B). To facilitate direct comparison with our prior optogenetic studies that labeled FS by Parvalbumin (PV; expressed in most FS), we plotted results for all FS, and found strong congruence (Figure 7B left). The last prestimulus spike timing showed a main effect on detection, characterized by peak enhancement when this spike occurred ~12.5ms prior (green data point). We then conducted the same analysis for each FS subtype, and found that while the pattern of modulation was similar across the subtypes, only grnsFS showed a significant main effect of last prestimulus spike timing.

Key to the framework proposed in Figure 7A is the peristimulus sustainment of gamma rhythmic spiking. Accordingly, the preferred alignment between sensory onset and grnsFS spike timing, observed for the last prestimulus spike timing in Figure 7B, is predicted to repeat with gamma rhythmicity in the peristimulus period. To test this prediction, we first calculated the spike timing-based DP (5ms bins, 1ms steps; Figure 7C). This metric quantifies whether a spike at a specific lag relative to sensory onset predicted hits (DP>0.5) or misses (DP<0.5). For a 10Hz firing rate, the probability that a spike will occur in a 5ms window is only 5%. Due to this limitation in dynamic range, the observed effects are small (i.e., DP did not deviate substantially from 0.5). Nonetheless, an oscillatory pattern in the spike timing-based DP was apparent for grnsFS, where the DP showed peristimulus local maxima at −14.5, 1.5, 18, and 34.5ms. To quantify this oscillatory pattern, we calculated spectral power of the peristimulus (−150 – 100ms) spike timing-based DP (Figure 7D): The gamma band power (30 – 55Hz) was highest for grnsFS, compared to other FS subtypes. In sum, the preferred alignment for grnsFS repeated rhythmically at gamma intervals.

Since sensory onset timing was randomized in our task, prestimulus spike alignment is not under volitional control of the mouse. Therefore, these alignment effects are likely not a mechanism that can be allocated to increase the likelihood of perceptual success. Rather, these alignment effects are better viewed as a byproduct of utilizing an oscillatory process for encoding, similar to suggestions from prior studies of brain rhythms (Fiebelkorn and Kastner, 2019; Fiebelkorn et al., 2013; Schroeder et al., 2008). Accordingly, the influence of preferred spike timing alignment relative to sensory onset (Figure 7) on detection is much smaller than the influence of the gamma regular spiking in grnsFS (Figure 6, S7 and Table S1-2). Nonetheless, we can infer from the existence and pattern of preferred spike timing alignment in grnsFS that gamma regular spiking in this subtype modulates sensory relay by evoking IPSPs in sRS.

Greater grnsFS Regularity Correlated with Hit-Predictive Dynamics in Putative Excitatory Relay Neurons

Synchronous gamma spiking in FS is hypothesized to benefit perception by influencing local excitatory projection neurons (Figure 7A; Cardin et al., 2009; Fries, 2015; Pritchett et al., 2015; Siegle et al., 2014; Sohal, 2016). Elucidating the potentially differential influence of the FS subtypes on local pyramidal neurons will require systematic analysis of many pairs of simultaneously recorded pyramidal neurons and FS in each subtype, or a method to manipulate FS subtypes selectively. Here, we began to address this question by examining sRS, and analyzing whether their task-relevant activity patterns could be predicted from simultaneously recorded FS in each subtype (Figure 8).

Figure 8. Greater grnsFS Regularity Correlated with Hit-Predictive Dynamics in Putative Excitatory Relay Neurons.

A. Spike rastergram of an example sRS on stimulus amplitude matched hits (left) and misses (right). Trials were sorted by stimulus amplitude, with stronger stimuli on top. The first poststimulus spike after sensory onset is colored in blue on the left plot, and red on the right; all other spikes are black. Stimulus amplitude matched trials were divided into subthreshold and suprathreshold based on the perceptual threshold (i.e., median stimulus amplitude among matched trials).

B. First spike latency distributions are plotted for the example sRS in A: Subthreshold hit and miss trials are shown in thin blue and thin red, suprathreshold in thick blue and thick red. Histogram counts were calculated in 10ms bins, sliding in 1ms steps, and normalized by dividing by the bin length (10ms) and trial count. The Y-value at the peak of this histogram was defined as first spike latency reliability.

C. Left First spike latency distributions, averaged across sRS (N=56), are shown for subthreshold and suprathreshold hits and misses.

Right First spike latency reliability quartiles across sRS are shown. The timing of the first spike was more reliable on hits compared to misses (⁺p<0.05 Wilcoxon signed rank test, Bonferroni correction for multiple comparisons) and on suprathreshold trials compared to subthreshold trials (*p<0.05 Wilcoxon signed rank test, Bonferroni correction for multiple comparisons).

D. Left PSTHs, as in Figure 5Fi, averaged across sRS, are shown for subthreshold and suprathreshold hits and misses.

Right Poststimulus (0 – 100ms) spike rate quartiles across sRS are shown. Sensory evoked rate increased with stimulus amplitude (*p<0.05 Wilcoxon signed rank test, Bonferroni correction for multiple comparisons). In addition, sensory evoked rate was higher on hits than misses across sRS, for subthreshold and suprathreshold stimuli (⁺p<0.05 Wilcoxon signed rank test, Bonferroni correction for multiple comparisons).

E. For simultaneously recorded pairs of sRS and FS of each subtype, trials were binned based on the CV of FS, and the first spike latency reliability of sRS was averaged across trials in each bin. Mean ± SEM across sRS are shown for each CV bin. CV was calculated on a trial-by-trial basis for ISIs within a 1s prestimulus window, and only trials where CV>0 were considered. Suprathreshold (magenta) and subthreshold (black) trials are plotted separately, to reduce the variance explained by stimulus amplitude (p-values are indicated for Friedman test in respective colors; *p<0.05 with post-hoc Bonferroni correction for multiple comparisons). The reliability of sRS first spike latency was predicted by prestimulus CV of grnsFS but not other FS subtypes, for subthreshold and suprathreshold trials.

F. As in E, but for 100ms poststimulus rate of sRS conditioned on FS CV. Significant interaction between FS prestimulus CV and sRS poststimulus rate was found for grnsFS and plnsFS on suprathreshold trials.

As a measure of onset precision in sensory evoked firing, we examined the reliability of first spike latency in sRS (i.e., the peak height of the first spike latency distribution; Figure 8A-C). We also measured changes in firing rate 0 – 100ms poststimulus (Figure 8D). Both metrics showed significant differences between hits vs misses, as well as between subthreshold (stimulus amplitude < perceptual threshold) vs suprathreshold (stimulus amplitude ≥ perceptual threshold) trials.

We tested whether these task-predictive sRS dynamics were predicted by CV of simultaneously recorded FS. Lower CV (higher regularity) of grnsFS predicted reliable alignment of sRS first spike latency on subthreshold and suparathreshold trials (Figure 8E). No consistent relationship was observed for sFS and plnsFS, despite all FS subtypes predicting hits with lower prestimulus CV (Figure S7Aii). Lower CV predicted higher sRS rate on suprathreshold trials in grnsFS and plnsFS (Figure 8F).

In sum, higher regularity in grnsFS firing was correlated with higher expression of task-predictive sRS dynamics. While such analyses are inherently inconclusive with regards to the exact connectivity patterns between FS subtypes and pyramidal neurons, these observations suggest that gamma regular spiking in grnsFS influences the sRS dynamics that lead to effective sensory relay, and in turn, perceptual success.

Discussion

The present study reports the initial description of FS subtypes with distinct temporal spiking patterns. One subtype, the grnsFS, was not sensory responsive and spiked at gamma range intervals with higher regularity. The grnsFS spiking was unperturbed by the stimulus and not coupled to the LFP, and yet showed high within-subtype synchronization. Successful trials in a detection task was associated with higher regularity and increased occurrence of gamma ISIs. A nonsensory dynamic predicting perceptual success could be interpreted as being indicative of brain state: Here, gamma regular spiking of grnsFS would indicate the brain state optimal for sensory processing. In addition, increased regularity of grnsFS predicted higher rates and more reliable alignment of first spike latency in sensory-evoked RS firing. This suggests that gamma regular spiking of grnsFS may contribute to sensory encoding through direct influence on sRS. In aggregate, gamma regular spiking of grnsFS, unperturbed by sensory input, could provide a perceptually relevant oscillation that coordinates spiking across a local population of excitatory neurons, enhancing the efficacy of sensory relay (Singer, 1993; Singer and Gray, 1995).

We and others have hypothesized that FS synchrony at gamma could facilitate perceptual relay by enabling precise alignment of local RS spiking following sensory onset (Cardin et al., 2009; Knoblich et al., 2010; Ni et al., 2016; Pritchett et al., 2015; Siegle et al., 2014). The hypothesized mechanism predicts an optimal alignment between FS spike timing and sensory onset, determined jointly by the time constant of GABAA inhibition and the first spike latency following sensory onset in SI barrel cortex. We observed the predicted preferred alignment in grnsFS. Further, grnsFS ISI regularity correlated with sRS first spike latency reliability, suggesting direct influence of grnsFS on sRS. The enhanced feedforward relay, putatively achieved by precise alignment of first spike latency in sRS, could lead to enhanced feedback from higher order areas to SI barrel cortex 50 – 100ms after sensory onset (Cauller and Kulics, 1991; Jones et al., 2007; Sachidhanandam et al., 2013; Takahashi et al., 2016). This mechanism provides an explanation for the accompanying rate increase in sRS. Alternatively, rebound from FS inhibition may underlie sRS rate enhancement (Fellous et al., 2003; Knoblich et al., 2010; Pritchett et al., 2015; Tiesinga et al., 2004).

While the behavior of grnsFS is consistent with aforementioned gamma theories, this subtype shows a marked departure from the established relationship between FS and the LFP gamma (Cardin et al., 2009; Fries et al., 2001; Hasenstaub et al., 2005). The LFP is often modeled as coordinated current fluctuations in large apical dendrites of pyramidal neurons (Buzsáki et al., 2012). As such, the relationship with the LFP can offer insights about synaptic connectivity patterns. The incoherent and weak negative correlation of grnsFS with LFP gamma suggests they are not involved in pyramidal-interneuron network gamma (Lee and Jones, 2013; Tiesinga and Sejnowski, 2009). In fact, their gamma spiking may not involve recurrent excitatory connections, given that they appear to modulate sRS but are themselves nonsensory. Further, the non-negative STA slope leading up to a grnsFS spike suggests relative scarcity of local excitatory inputs to this subtype.

Differences in grnsFS baseline spiking behavior and their relationship with the LFP, as well as other “pseudo-intrisic” properties such as the extracellular spike waveform and firing rate, suggest that grnsFS comprise an FS subtype with distinct genetic expression, biophysical properties, and/or synaptic connectivity compared to sFS. Specifically, the decoupling with the LFP suggests that grnsFS spiking may rely primarily on intrinsic mechanisms e.g., intrinsic gamma resonance (Sciamanna and Wilson, 2011). Alternatively, grnsFS spiking may be driven by long-range inputs selective to grnsFS (e.g., Kim et al., 2015). In contrast, pseudo-intrinsic properties of plnsFS were largely similar to sFS, and their identification as nonsensory may reflect their tuning (Andermann and Moore, 2006), as opposed to distinct biophysical/anatomical properties.

We identified FS primarily by their extracellular spike waveform, leaving open the possibility that some FS are not PV+ GABAergic interneurons. However, the relatively high baseline firing rate (~10Hz) of extracellularly identified FS is typical of PV+ FS, making it likely that the majority of all three FS subtypes are PV+. Further, several reports highlight the heterogeneity within PV+ FS, supporting the possibility that the distinct FS subtypes described here could all be PV+ (Bortone et al., 2014; Dehorter et al., 2015; Donato et al., 2015; Goldberg et al., 2008; Gouwens et al., 2018; Povysheva et al., 2013).

Several precedents suggest the subtypes described in the present study. Puig et al. (2008) found two distinct activity patterns of FS during up states in frontal neocortex. One FS subtype demonstrated ISI distributions with peaks in the gamma band that discharged late within an up state, strongly resembling the ISI distributions of the grnsFS described here for SI. Li and Huntsman (2014) reported two broad electrophysiological categories of FS within layer 4 of SI barrel cortex. They found that FS showing delayed spiking in response to current injection had smaller thalamocortical-evoked responses compared to early spiking FS, suggesting a link to nonsensory FS in awake behaving animals. Relatedly, Helm et al., (2013) classified subgroups of parvalbumin expressing interneurons in layer 2/3 of visual cortex based on their passive and active membrane properties using in vitro slice electrophysiology. The subgroup with a more regular firing pattern also had a higher rheobase, action potential threshold and smaller excitatory postsynaptic currents, suggesting similarities with gamma regular non-sensory FS. We note that laminar positions of recording sites were ambiguous in our data.

Testing the hypothesis that persistent gamma rhythmic spiking among nonsensory FS would enhance perception, through binding or “communication through coherence”, requires further evidence (Fries, 2015; Singer, 1993; Singer and Gray, 1995). The functional connectivity of FS subtypes, a topic we have begun to address in this study, requires a more thorough and direct investigation. Another key future direction is to investigate whether grnsFS exist in other brain areas and, if so, whether mechanisms exist for coordination of grnsFS activity across brain regions. Further, the ability to selectively manipulate grnsFS would be critical for delineating their specific role. If supported by investigation in other paradigms and brain areas, our findings provide a unique resolution to a consistent challenge for theories predicting a role for gamma oscillations in optimal brain communication.

STAR METHODS

LEAD CONTACT AND MATERIALS AVAILABILITY

Further information and requests for resources should be directed to and will be fulfilled by the Lead Contact, Christopher I. Moore (Christopher_Moore@brown.edu). This study did not generate new unique reagents.

EXPERIMENTAL MODEL AND SUBJECT DETAILS

Animals

Electrophysiology was collected from four mice (3 males, 1 female) performing a vibrissal deflection detection task. For replication of results in a second dataset (Figure S6), 6 mice (4 males, 2 females) were used; these mice were heterozygous for Slc32a1-IRES-Cre (also known as VGAT-IRES-Cre, Jackson Laboratory strain number 016962) and Parvalbumin-2A-Flpo (strain number 022730). All mice were 8-15 weeks at the time of surgery, and were recorded for up to 7 months. Animals were individually housed with enrichment toys and maintained on a 12 hour reversed light-dark cycle. All experimental procedures and animal care protocols were approved by Brown University Institutional Animal Care and Use Committees and were in accordance with US National Institutes of Health guidelines.

METHOD DETAILS

Driveable Electrode Implants

We implanted flexDrives constructed in house, loaded with either 16 tetrodes or 4 tetrodes (Voigts et al., 2013). Tetrodes were made with Sandvik Kanthal HP Reid Precision Fine Tetrode Wire, Nickel-Chrome 0.012mm diameter, and gold-plated to 200 – 400kΩ impedence. The guide tube array was made with 33ga polyamide tubes, resulting in ~250μm spacings. Two stainless-steel screws (0.6 or 0.8mm diameter, 0.5 or 1mm length) were electrically connected (soldered) to the electrode interface board (EIB) through stainless steel wire. These skull screws were implanted through skull to rest on top of the dura to serve as ground: one was positioned anterior to bregma and the other on the right hemisphere.

For the 6 mice that contributed to Figure S6, the optic fiber was included as a part of the implant (200μm multimode fiber, Thorlabs FG200UCC). The fiber was incorporated into the flexDrive such that it would occupy one of the guide tubes, which allowed the bottom surface of the fiber to be positioned perpendicular to the cortical surface. A 1.25 mm diameter stainless steel ferrule (SFLC270) was glued to the top side of the fiber, and polished to generate an even surface.

Surgical Procedure

Details of the surgical procedure and behavior control were reported in a prior study (Shin et al., 2017). Mice were induced with isoflurane anesthesia (0.5 – 2% in oxygen 1L/min) and secured in a stereotaxic apparatus. We injected slow-release buprenorphine subcutaneously (0.1mg/kg; as an analgesic) and dexamethasone intraperitoneally (IP, 4mg/kg; to prevent tissue swelling). Hair was removed from the scalp with hair-removal cream, followed by scalp cleansing with iodine solution and alcohol. Then, the skull was exposed by scalp incision. After the skull was cleaned, muscle resection was performed on the left side. A titanium headpost was affixed to the skull with adhesive luting cement (C&B Metabond). Next, a ~ 1.5mm–diameter craniotomy was drilled over barrel cortex of the left hemisphere, and subsequently a durotomy was performed. The guide tube array was centered at 1.25mm posterior to bregma and 3.25mm lateral to the midline. The drive body was angled 30 degrees relative to vertical to compensate for the curvature of barrel cortex. Once the implant was stably positioned, C&B Metabond and dental acrylic (All for Dentist) was placed around its base to seal its place. A drop of surgical lubricant (Surgilube) prevented dental acrylic from contacting the cortical surface. Mice were given ≥3 days to recover before the start of water restriction.

The surgical procedure for the 6 optogenetic mice (Figure S6) were identical to the four detection-only mice, save for the addition of the viral injection step which succeeded the craniotomy and durotomy, and preceded the flexDrive implant. We injected the rAAV5-hSyn-Con/Fon--hChR2(H134R)-eYFP-WPRE (purchased from the Vector Core Facility at University of North Carolina Chapel Hill, titer 2.3×10¹²vg/mL). This virus is a type of INTRSECT that allows expression of hChR2 in only the neurons that express both Cre and Flp (Fenno et al., 2014). In our case, the virus would express in just the GABAergic PV-positive neurons, which addresses the concern that PV is expressed in a subset of layer 5 pyramidal neurons and other non-interneuron cells (Madisen et al., 2015). We injected in 3 coordinates per mouse at depth of 0.45mm: 1.25mm posterior and 3.25mm left of bregma, 1mm and 3mm, 1.5mm and 3.5mm. For each injection coordinate, 0.5μL was injected at 0.05μL/min, and the injection pipette was left in place at the injection depth for 10 more minutes before being retracted from the brain.

Recording and Preprocessing

All electrophysiology data were collected using the Open Ephys system continuously, with a sampling rate of 30kHz. Trial alignment was achieved through a synchronizing pulse output from the computer running the behavior control software, via an analog channel of a NI-DAQ device.

Local field potentials (LFP) were defined as the continuous data down-sampled to 1000Hz. Median filtering was applied concurrent with down-sampling, i.e., down-sampling was conducted by choosing the median every 30 samples. Spike detection was conducted online during data acquisition to visualize the relevance of a recording site, with the online criterion for spikes was whether 300 – 6000Hz bandpass filtered data crossed a permissive threshold of −50μV, in at least one of the four electrodes comprising each tetrode. Offline, spike threshold for each electrode was readjusted as 4 times the standard deviation of 300 – 6000Hz bandpass filtered data. The standard deviation was approximated as the median value divided by 0.6745 (Quiroga et al., 2004). Multi-unit activity (MUA) for each tetrode was defined as spikes that crossed this threshold in at least one of the four electrodes of that tetrode (one MUA per tetrode, per session).

To obtain single unit activity (SUA), offline spike sorting was conducted manually using Simple Clust (www.github.com/open-ephys/simpleclust). Only the clusters that were well isolated were classified as SUA. After sorting, single units were classified into regular spiking (RS) and fast spiking (FS) units, based on the time between peak to trough (T_PT) in their average spike waveform (RS if T_PT>0.4ms, FS if T_PT≤0.4ms; Figure S6A-B i-ii).

Tetrodes were lowered at the end of a session if the experimenter noticed that there were no single units detected from online sorting. Tetrodes were lowered by 1/8 turns, corresponding to ~31.25μm.

Behavior Training

Mice were water restricted for ≥7 days before start of training, during which time they were acclimated to being head-fixed on a fixed-axis styrofoam ball, where they could walk or run freely. Locomotion was tracked in a subset of sessions using a rotary encoder (E6B2-CWZ3E). Water was delivered through a syringe after each acclimation session.

When training began, the vibrissae were secured through a suture loop. All macrovibrissae posterior to the 4^th arc on the right side were secured ~3mm from the mystacial pad. The suture loop, in turn, was fed through a glass capillary tube (0.8mm outer diameter) glued to a piezoelectric bender (Noliac CMBP09). On each trial, 20Hz vibratory vibrissal stimulus train (10 deflections, 500ms) were delivered through the piezoelectric bender in the caudorostral direction, with a half-sine wave velocity profile that had fast rising phase (6ms) and a slower relaxation phase (20ms). The 10 deflections within each stimulus train of a trial were of the same amplitude. On a trial-by-trial basis, the stimulus amplitude was varied between 0 (a catch trial) to maximal amplitude (~1mm deflection) in a randomized manner. Before the start of each session, the experimenter set the overall percentage of maximal trials and 0 amplitude catch trials. The rest of the trials were submaximal trials, where the stimulus amplitude was randomly drawn from a uniform distribution between 0 and maximal amplitude. Throughout training, the percentage of maximal amplitude trials was gradually lowered to 10%, and the percentage of catch trials was gradually increased to 25%.

The mice initially learned the vibrissal deflection to reward association in sessions where every vibrissal deflection trial was paired with ~3μl water delivery regardless of the animal’s response. Vacuum suction followed 500ms after every reward delivery, which removed any excess water not consumed by the animal in that trial. Mice learned the vibrissal deflection to reward association after about a week of reward-all training, at which point the animals would lick in anticipation of the reward with a stereotyped reaction time of ~250ms poststimulus onset (median 254ms; IQR 187 – 389; Figure 5B). When this behavior was observed (>4 days of training), the reward-all phase of training was concluded.

During the subsequent phase performance was required for reward, and it was only delivered on detected (hit) trials where the mouse correctly reported detection by licking 0 – 700ms of the stimulus onset (report window). Here, the consequence of non-detection (misses) was omission of reward. If the mouse falsely reported detection by licking during the 700ms report window on catch trials, the trial outcome was classified as a false alarm. The consequence of false alarms was a 15s timeout from the task, effectively delaying the next opportunity to obtain a reward. If the mouse correctly refrained from licking during the report window on catch trials, the outcome was classified as a correct reject. Inter-trial intervals (ITIs, defined as the interval between report window onsets) were randomly drawn from a uniform distribution ranging between 4.5 – 8s. There was no cue indicating the start of each trial.

If the animal developed a habit of excessive impulsive licking, an ITI-reset was implemented until that habit was eliminated. In these sessions, licking during the ITI would prolong the ITI up to 10 times.

Mice were weighed before and after each session. If the mouse consumed ≥1ml of water throughout a behavior session, the mouse would gain ~0.7g at the end of a behavior session. Hence, if the mouse did not gain ≥0.7g during the behavior session, supplementary water was given several hours after the conclusion of the session, so that the mouse would have drunk at least 1 ml each day. In addition, mouse weight was monitored throughout the duration of training and maintained ≥80% of their original weight at the time of surgery. The duration of training ranged between 3 to 7 months.

QUANTIFICATION AND STATISTICAL ANALYSIS

Behavior Analysis

A typical behavior session lasted ~2.5 hours, with ~1000 trials (median 946, IQR 817 – 1032). Even in well trained mice displaying robust psychometric behavior and stereotyped reaction time, there were periods where animals defaulted to non-optimal strategies, such as excessive impulsive licking or non-engagement. All hit, miss, false alarm and correct reject trials analyzed were trials where there was no licking activity up to 1s before the onset of the report window. Across sessions, the percent of trials filtered out due to licking in the 1s prestimulus had a median of 10%, with IQR 7 – 26%.

In addition, we employed d’≥1 criteria to filter out periods of low performance. The d’ was calculated in sliding windows of 51 trials advanced in 1 trial steps. For any blocks with d’≥1, the middle trial was included in the analysis. The first and last 25 trials were included if the first and last 51 trials had d’≥1, respectively. The d’ was defined as the following:

$${\mathrm{d}}^{’}=Z(\text{adjusted hit rate for strong amplitude stimulus})-Z(\text{adjusted false alarm rate})$$

where Z is the normal inverse cumulative distribution function. For each session, N strongest amplitude stimulus was analyzed for the Z, where N equals the number of catch trials in that session. Hit rates and false alarm rates were adjusted to avoid infinite values of d’ by adding 0.5 to the number of trials in each category (Hautus, 1995; Macmillan and Creelman, 2004; Miller, 1996), i.e.;

$$H{R}_{\mathit{adj}}=\frac{N_{\mathit{hits}}+0.5}{(N_{\mathit{hits}}+0.5)+(N_{\mathit{misses}}+0.5)}$$

$$FA{R}_{\mathit{adj}}=\frac{N_{\mathit{FA}}+0.5}{(N_{\mathit{FA}}+0.5)+(N_{\mathit{CR}}+0.5)}$$

where HR_adj denotes adjusted hit rate and FAR_adj denotes adjusted false alarm rate; and N_hits, N_misses, N_FA, N_CR denotes the number of hits, misses, false alarms and correct rejects, respectively. The d’ filtering left a median of 30% trials, IQR 10 – 45%. In many cases, this was due to satiation at the end of the session: The percentage of d’<1 trials between the first d’≥1 trial of the session and the last d’≥1 trial of the session was much lower, with a median of 5% and IQR 0 – 22%.

For each session, psychometric curves were calculated by sorting the d’≥1 trials in ascending stimulus amplitude, and smoothing the sorted stimulus amplitude vector and the corresponding trial outcome vector (1 for hits, 0 for misses) with 51-trial windows in 1-trial steps (Figure 5A). Only the behavior sessions with a psychometric curve that had a good fit with a Boltzmann distribution were selected for analysis (across the 128 sessions, R-squared median 0.96, IQR 0.93 – 0.98; root mean squared error (RMSE) median 0.067, IQR 0.053 – 0.086).

To compare hit and miss trials independent of variations in stimulus amplitude, we performed for each session stimulus amplitude histogram matching of hit and miss trials, such that matched hits and misses were equal in trial count and stimulus amplitude distribution. Matching was conducted on submaximal trials. To analyze only the trials where the animal was engaged in and motivated to perform the task, we further limited candidate trials to trials where d’≥1. Additionally, hit trials with ≤100ms reaction time were excluded (4.63% trials), allowing 0 – 100ms poststimulus window to contain only the task-predictive activity.

The histogram matching process was as follows (Figure 5C):

1. Trials eligible for matching (submaximal, d’≥1, RT≥100ms) were divided into 15 bins, equidistant in stimulus amplitude.

2. For each bin, hits and misses were matched in number; if there were more hits than misses, we randomly sub-selected hit trials equal to the number of miss trials; and vice versa for bins with more misses than hits.

Across sessions, this resulted in a median of 78 matched hits and misses each per session, where IQR was 57 – 95 trials. The median amplitude of matched trials was defined as the stimulus amplitude threshold for that session, and this value was used to demarcate stimulus amplitude bins in Figure 1, 5B, 8, and S1.

Details of Data Analysis

Data analysis was performed in MATLAB (Mathworks). Sensory responsiveness was defined based on stimulus probability (SP, Figure 1B). SP was defined as the area under receiver operating characteristic curve (AUROC) in an ideal observer analysis, where the ideal observer discriminated maximal stimulus amplitude trials from zero amplitude catch trials based on firing rate in the 0 – 100ms period poststimulus onset. Intuitively, the AUROC quantifies the performance of the ideal observer; in other words, it quantifies the extent to which the two distributions being subject to discrimination are separable. Significance was determined by 95% confidence interval of SP, constructed by bootstrapping 1000 times. MATLAB function “perfcurve” was used for this analysis.

Detect probability (DP) was similarly defined as the AUROC in an ideal observer analysis discriminating stimulus amplitude matched hits and misses (Figure 5C). DP analyses are presented in Figure 7 and Table S2.

The latency to first spike for each unit on each trial was defined as the timing of the first spike after the stimulus onset. The first spike latencies were accumulated across trials of interest and plotted as histograms, in 5ms bins sliding in 1ms steps for FS (Figure 1) and in 10ms bins sliding in 1ms steps for RS (Figure 8). A larger bin size is used in RS as RS have a lower baseline firing rate, hence there is less accumulation of spikes per bin. Note, all statistical comparisons shown for FS spike latency hold when using 10ms bins. The latency (X-value) at the peak of this distribution was defined as the first spike latency of that unit (Figure 1). The height (Y-value) at the peak of this distribution was defined as the first spike latency reliability (Figure 8).

The interval between two consecutive spike times of a SUA or a MUA is defined as the inter-spike interval (ISI). Coefficient of variance (CV) is defined as the standard deviation of the ISIs divided by the mean of the ISIs.

$$\mathrm{CV}=\frac{\mathit{std}(\mathit{ISI})}{\mathit{mean}(\mathit{ISI})}$$

Lower CV implies higher regularity in ISI’s. With the exception of trial-by-trial analyses, ISI distribution and CV for a designated time window was calculated by first pooling across trials the ISIs contained within the designated period. For trial-by-trial analyses, CV and ISI distributions were calculated for each trial (Table S2, Figure 8, S3C, S3F, S4, and corresponding figures in Figure S5). Note, trial-by-trial measures of CV and ISI distributions are noisy, as a neuron that fires at 10Hz baseline would have on average 9 ISIs in a 1s period.

Fano Factor is defined as the variance over mean of the firing rate (FR) across trials.

$$\text{Fano Factor}=\frac{\mathit{var}(\mathit{FR})}{\mathit{mean}(\mathit{FR})}$$

The Akaike information criteria (AIC) was applied to K-means clustering results to determine the number of clusters that best describes the ISI peak – CV space (Figure 2D, Figure S1A, Figure S6A-Biii). Before running K-means clustering, CV was scaled to (CV – 0.5)/2, and peak ISI was scaled to Peak_ISI/50. This allowed each dimension to have a similar numeric range. Without such preprocessing, K-means clustering would be biased towards the dimension with the larger numbers. MATLAB function “kmeans” was used, and the cluster parameter was varied systematically between 1 through 5. Because the K-means algorithm converges to a local minimum, the results can depend on the initialization. The procedure was repeated 30 times and the initialization with the lowest sum of within-cluster point-to-centroid squared Euclidean distances (SSED) was chosen, to better approximate the global minimum. AIC for this solution is defined as:

$$\mathit{AIC}=\mathit{SSED}+2\mathit{dk}$$

SSED stands for Sum of Squared Euclidean Distances; d is the dimensionality (d=2; CV and peak ISI); and k is the number of clusters (varied between 1 through 5).

For each SUA and MUA recording, spike times were converted into spike trains in bin resolution of 1ms (bin value is 1 if there was a spike at that time point, 0 otherwise). Spike trains were converted into matrices such that each column corresponded to a trial. Peristimulus time histograms (PSTH) and cross-correlograms were obtained from these matrices. The PSTH was calculated by first, averaging across trials; second, multiplying by 1000 such that the units would be in Hz; and third, smoothing with a Gaussian filter (with a standard deviation of 3ms).

Cross-correlograms were calculated as the rate of coincident spikes between two spike trains that were offset from each other, where the spike trains were from the continuous recording of the entire session, and the offsets were systematically varied between −100 – 100ms. The resulting raw cross-correlogram was then corrected for rate effects by subtracting the mean of 1000X jittered cross-correlograms (Figure 4, Figure S4A). Each jitter cross-correlogram was generated by jittering spike times of each spike train in 50ms windows, and cross-correlograms were calculated for 1000 jitter iterations (Amarasingham et al., 2012). Note, the mean of the jittered cross-correlograms asymptote to the cross-correlogram computed spike trains smoothed by 50ms boxcar window.

In Figure 7B, we plot the proportion of hit trials among trials where the last prestimulus spike fell on the query 5ms bin. The change in hit rate was calculated for each FS, for each bin. This analysis was limited to stimulus amplitude matched hits and misses; therefore, the baseline hit rate was 0.5, and the Y-values show the change from this baseline. For statistical testing, change in hit rate was defaulted to 0 (i.e., no change from baseline) if no trials had last prestimulus spike within that bin for that FS.

In Figure 7C, we calculated the spike timing-based DP for 5ms bins, sliding in 1ms steps. This analysis was intended to measure the impact of a single spike at a specific latency relative to sensory onset. As a 5ms bin had either no spike or a single spike for the vast majority of cases, spike count DP can achieve this purpose.

Spectral analysis for LFP, spike trains, and SFC was calculated using a multitapered Fourier transform, where time-bandwidth product was set to 3 and the number of tapers was set to 5. Zero padding was applied such that the LFP or spike train data being processed would have a length of 2^N; e.g., for a LFP segment with length of 250, 3 zeros were padded on either side such that the total length would be 2⁸=256. Chronux version 2.11 (http://chronux.org/, Mitra and Bokil, 2004) was used for this analysis. Spectral power in each frequency band was calculated as the geometric mean across frequencies of that frequency band.

Multitapered Fourier transform was also applied to DP traces (5ms bins, 1ms steps) in Figure 7. As the non-predictive “null” value of DP is 0.5, we subtracted 0.5 from the DP trace before executing the multitapered Fourier transform.

For STA and SFC in Figure 4, as well as spike – LFP correlations in Figure S3Fi, the LFP was taken as the average of three tetrodes neighboring the tetrode where the FS spikes were recorded. This step was taken to prevent the influence of spike contamination on the LFP (Ray and Maunsell, 2011; Waldert et al., 2013). In Figure 4Cii and Figure S6Fii, the slope of STA leading up to the spike was calculated as:

$$\{(\mathrm{mean}\mathrm{LFP}\mathrm{voltage}\mathit{0}\mathit{\pm }\mathit{2.5}ms\mathrm{relative}\mathrm{to}\mathrm{spike})-(\mathrm{mean}\mathrm{LFP}\mathrm{voltage}-\mathit{5}\mathit{\pm }\mathit{2.5}ms\mathrm{relative}\mathrm{to}\mathrm{spike})\}∕\mathit{5}$$

DATA AND CODE AVAILABILITY

Data is deposited in Mendeley Data (http://dx.doi.org/10.17632/r5tbz5j34p.1). Software is available at both Mendeley Data and Github (https://github.com/hs13/GammaRegularNonSensoryFastSpikingNeurons.git).

Highlights.

• Gamma regular non-sensory FS (grnsFS) spike regularly at gamma range intervals.

• Unlike sensory FS, grnsFS spiking is distinctly more regular than a Poisson process.

• grnsFS spike synchronously, but do not cohere with the LFP gamma.

• grnsFS predict detection at perceptual threshold with enhanced gamma regular spiking.