Narrowband gamma oscillations propagate and synchronize throughout the mouse thalamocortical visual system

Summary

Oscillations of neural activity permeate sensory systems. In the visual system, broadband gamma oscillations (30 – 80 Hz) are thought to act as a communication mechanism underlying perception. However, these oscillations show widely varying frequency and phase, providing constraints for coordinating spike timing across areas. Here, we examined Allen Brain Observatory data and performed causal experiments to show narrowband gamma (NBG) oscillations (50 – 70 Hz) propagate and synchronize throughout the awake mouse visual system. Lateral geniculate (LGN) neurons fired precisely relative to NBG phase in primary visual cortex (V1) and multiple higher visual areas (HVAs). NBG neurons across areas showed higher likelihood of functional connectivity and stronger visual responses; remarkably, NBG neurons in LGN preferring bright (ON) versus dark (OFF) fired at distinct NBG phases aligned across the cortical hierarchy. NBG oscillations may thus serve to coordinate spike timing across brain areas and facilitate communication of distinct visual features during perception.

eTOC blurb

Neural oscillations pervade sensory systems; it remains unclear how oscillations coordinate spiking across brain areas. Shin et al. show synchronized narrowband gamma (NBG) oscillations across thalamus, visual cortex, and higher visual areas. NBG improves sensory coding and aligns spikes of neurons preferring brights versus darks, a fundamental aspect of vision.

Graphical Abstract

Introduction

Oscillations of neural activity are thought to play an important role in both representing and communicating sensory information across the brain. Extensive studies in visual cortex have shown that both visual stimulus features and task performance drive broadband gamma (30 – 80 Hz) oscillations^1,2 that could facilitate neuronal communication underlying perception^3,4. However, recent work identifies several limitations for broadband gamma oscillations to both represent sensory stimuli and synchronize communication across brain areas^5,6. First, broadband gamma oscillations measured simultaneously across brain areas show high variability in frequency, amplitude, and phase. Second, there is no “central clock” that coordinates broadband gamma oscillations across brain areas at millisecond timescales. Third, broadband gamma oscillations typically emerge hundreds of milliseconds after stimulus onset and their strength fluctuates continuously with stimulus features. All these factors pose constraints for maintenance of spike timing precision and synchronization across widespread visual areas. Instead, broadband gamma oscillations may primarily reflect timescales of local cortical excitation and inhibition^7–9. If an oscillation acts primarily to coordinate visual activity across brain areas, then it should 1) show consistent frequency and phase across regions, 2) show synchronization at the visual input layers across these regions, and 3) enforce neurons to reliably spike at distinct oscillation phases according to visual feature preferences.

Recent studies have unveiled a novel narrowband gamma (NBG) oscillation in the mouse visual system. Unlike broadband gamma activity, NBG in primary visual cortex (V1) shows a highly stereotyped oscillation frequency (central peak between 50 – 70 Hz) and narrow bandwidth (5 – 7 Hz), and is not generated by visual stimulus features. NBG in mouse V1 emerges spontaneously during wakefulness¹⁰, disappears in total darkness, and arises from lateral geniculate nucleus (LGN)¹¹ and likely from upstream retinal inputs^12–14. NBG in mouse V1 varies with arousal and behavioral state in both locomoting and non-locomoting mice^{10,11,15–17}, even when pupil is pharmacologically fully dilated¹¹. Further, NBG strength accurately predicts visual perceptual task performance even in the total absence of locomotion¹⁸. This suggests that NBG could function to coordinate activity across visual areas underlying perception. This hypothesis has been extensively explored and debated for “classic” intracortical 40 Hz gamma in cats, monkeys, and humans, but a functional role for NBG in the visual system still poses many fundamental and unanswered questions. Mice present an opportunity to investigate if sub-cortically generated NBG oscillations might play a role in multi-area coordination of visual signals. However, it is unknown if NBG activity propagates beyond V1 to higher visual areas (HVAs) necessary for visual perception^19,20. It is unknown if NBG in HVAs depends on LGN and V1 activity, how synchronized NBG is across visual cortical areas, or if NBG influences the intensity and timing of spike responses to visual stimulus features. Answering these questions requires large-scale, simultaneous neural recordings from LGN, V1, and HVAs, perturbations of NBG across areas, and a framework for detecting and quantifying population and single-neuron level NBG activity.

Here we addressed these questions by analyzing the Allen Brain Observatory Visual Coding dataset of multi-area simultaneous Neuropixels recordings^21,22, and by performing recordings from HVAs during simultaneous optogenetic inactivation of V1 and LGN. We found strong evidence for NBG activity propagation and synchronization throughout the mouse thalamocortical visual system. Many neurons in LGN, V1, and HVAs showed tightly coordinated NBG spiking and high likelihood for pairwise functional interactions. NBG in HVAs aligned with and depended upon LGN activity and retinotopically aligned V1 activity. Surprisingly, NBG neurons across areas showed enhanced visual responses, and LGN neurons preferring bright (ON) or dark (OFF) stimuli fired at distinct phases of NBG oscillations; these phase preferences in LGN spikes were aligned with NBG activity in V1 and multiple HVAs according to their position along the anatomical hierarchy. Together, these findings show that NBG oscillations effectively coordinate spiking in functionally distinct groups of neurons throughout the awake mouse visual system, identifying a novel potential substrate for rapid communication of visual information across the brain.

Results

Correlated NBG spiking across LGN, V1, and HVAs

We first verified NBG communication from LGN to V1 in the Allen Brain Observatory – Visual Coding-Neuropixels dataset (Fig. 1A, top), and then found evidence for correlated NBG spiking across LGN, V1, and HVAs (Fig. 1A, bottom. Consistent with prior reports, ^11,23, many individual LGN neurons showed NBG power (between 50 – 70 Hz) in their spike autocorrelograms (ACGs; Fig. S1A–C). Further, cross-correlograms (CCGs) of many LGN-V1 pairs showed correlated spiking oscillating at NBG frequencies (Fig. 1B). We identified NBG neurons as those with CCGs that fulfilled several quantitative metrics of significant NBG power (see Methods and Fig. S1I–L for examples and population distributions). Correlated NBG spiking between LGN - V1 neuron pairs was highly specific: the same LGN neuron could show significant NBG firing coordinated at millisecond timescale with some V1 neurons (Fig. 1B), but not others recorded simultaneously just tens of microns away on the probe (Fig. 1C). In this example session, we found many significant NBG CCGs between LGN neuron pairs, LGN-V1 pairs, and LGN-HVA pairs (Fig. 1D). The appearance of “phase-tiling” in CCG peaks (Fig. 1D) likely arises from 1) spike time differences of LGN NBG neurons relative to one another (discussed in detail later) and 2) conduction delays to V1 (~2.4 ms) then HVAs (~10 – 20 ms)²². CCGs between these NBG neurons and other neurons recorded simultaneously showed no significant NBG power in any CCG (Fig. 1E), ruling out global synchronization and instead suggesting that only specific neurons show correlated NBG firing across the visual system. Across all sessions with LGN single unit recordings in the Allen dataset (n = 32 recordings), 35% of all LGN neurons were classified as NBG neurons (n = 455; Table S1).

Figure 1. Coordinated narrowband gamma (NBG) activity across thalamocortical visual areas.

A. Simultaneous multisite Neuropixels recordings from lateral geniculate nucleus (LGN), primary visual cortex (V1), and higher visual areas (HVAs) from the Allen Brain Observatory Visual Coding - Neuropixels dataset were examined for evidence of NBG oscillations from LGN to V1 (top), and across LGN, V1, and HVAs (bottom).

B. Example spike cross-correlogram (CCG) between two NBG neurons in LGN and V1 during spontaneous activity (1 session, 170 mins; ~161k LGN spikes and ~393k V1 spikes). Neurons classified as NBG neurons if CCGs showed significant 50 – 70 Hz power (inset; see Methods and Fig. S1).

C. CCG of same LGN NBG neuron but a different V1 neuron (on an adjacent contact than one in B). Neurons without significant 50–70 Hz power classified as non-NBG.

D. CCGs between NBG LGN neuron in B, and all other simultaneously recorded NBG neurons in LGN (n = 57), V1 (n = 21) and HVAs (n = 8) in same example session (see also Table S1). Each CCG normalized to peak (yellow) and sorted by peak lag.

E. CCGs between the same NBG LGN neuron as B-D and all simultaneously recorded non-NBG neurons.

We performed several additional control measures to confirm that identification of NBG neurons by CCGs depended upon spike timing, including jittering spike times and computing CCGs across sessions (Fig. S1A–H; Methods). Further, identification of NBG neurons using CCGs showed high overlap with identification based solely on spike autocorrelograms (ACGs; Fig. S1J–L), and our main results were unchanged even when analyzing only NBG neurons identified with ACGs, as will be evident later.

NBG local field potential (LFP) oscillations across V1 and HVAs were synchronized with spikes of NBG LGN neurons. We examined V1 LFPs triggered on simultaneously recorded spikes of NBG neurons in LGN (Fig. 2A). We used current source density analysis (Fig. S2A) to localize input layer 4 (L4) of V1, then calculated the spike triggered LFP (stLFP) at this site relative to NBG LGN neuron spikes during spontaneous activity. An example stLFP showed clear and significant NBG oscillations (Fig. 2A; 55.6 Hz), with LGN spikes preceding the LFP trough by 5.6 ms. We then identified the putative functional input layers of HVAs in the same way (Fig. S2 B–F) and calculated the stLFP in HVAs relative to NBG LGN spikes. As expected, the stLFP power was greatest in V1 (Fig. 2B) but also elevated in the lateral HVAs (RL, LM) versus medial ones (PM, AM; schematic in Fig. 2). Importantly, stLFP NBG power diminished significantly when triggered from non-NBG LGN neuron spikes in the same recordings (Fig. 2B dashed line; −6.9 ± 6.9 dB reduction on average), a significant difference in all areas except PM (V1: p < 1e-3; RL: p < 0.04; LM: p< 3e-4; AL: p < 6e-4; AM: p = 0.03; PM: p = 1; Wilcoxon rank sum tests). We then plotted stLFP power versus hierarchy scores that directly reflect anatomical connectivity²⁴, and found a significant correlation between the anatomical hierarchy and stLFP NBG power (Fig. 2C). Further, this steep relationship of stLFP NBG power to anatomical hierarchy (slope of −23.8 dB; Fig. 2C) was significantly weaker when triggered on non-NBG LGN neurons (slope of −11.0 dB, p < 4e-5, Student’s t-test). This hierarchical relationship of NBG LFP power was also evident with spike-field coherence (SFC) and pairwise phase consistency (PPC; see Fig. S3J–K)^15,25, and was reflected in the rhythmic structure of LGN spiking relative to cortical LFP NBG (Fig. 2D–F). Lastly, stLFP power in HVAs remained equally strong when triggered on V1 NBG neuron spikes (Fig. S4I), suggesting a tight preservation of feedforward NBG activity from LGN to V1 to HVAs.

Figure 2. NBG across HVA hierarchy depends upon LGN activity and retinotopic V1 activity.

A. Example spike-triggered LFP (stLFP) in V1 layer 4 (L4) triggered by simultaneously recorded NBG LGN neuron spikes (schematic at left). Peak LFP power at 55.6Hz. L4 identified with current source density (Fig. S2A).

B. stLFP NBG power across V1 and HVAs aligned to NBG LGN neuron spikes (n = 400 neurons, 14 sessions) sorted from highest to lowest stLFP power (mean ± SD), circles show individual session means (LM, lateromedial area; RL, rostrolateral area; AL, anterolateral area; PM, posteromedial area; AM, anteromedial area). Dashed line shows mean power triggered on non-NBG LGN neuron spikes in same sessions. See Fig. S4I for stLFP from V1 NBG neurons.

C. stLFP NBG power of V1 and HVAs (mean ± SD) sorted by anatomical hierarchy scores (see Methods). Significant correlation with anatomical hierarchy (Pearson rho = −0.90; p = 0.0142).

D. Example NBG LGN neuron spike histogram aligned to V1 NBG LFP cycle phase. Blue curve shows fitted cosine used to estimate cycle histogram signal to noise ratio (SNR, see Methods).

E. Cycle histogram SNR of V1 and HVAs across all recording sessions. Same conventions as C. Significantly greater than non-NBG neurons in all areas but PM (Methods).

F. Significant correlation of cycle histogram SNR with anatomical hierarchy scores. (Pearson rho = −0.89; p = 0.0162). Same conventions as C.

G. Recordings in V1 and HVAs during inactivation of visual thalamus by driving channelrhodopsin (ChR2) in thalamic reticular nucleus (TRN) inhibitory neurons. More than 4-fold reduction of NBG power in V1 (4.1 ± 3.1) and 2-fold reductions in HVAs (RL: 2.3 ± 1.2; LM: 2.0 ± 0.8 PM: 2.6 ± 0.6; all p < 0.004 sign tests; n = 40 experiments (V1:9, RL: 9, LM: 13; PM: 9; 12 recordings in 3 mice). NBG power reduction normalized by broadband gamma reduction (see Methods). Dashed line indicates no change with stimulation. See Fig. S2G–M for histology, inactivation efficacy, and controls.

H. HVA recordings during optogenetic inactivation of V1 (via ChR2 in V1 PV inhibitory neurons). Experiments performed when retinotopy was matched or mismatched at stimulation and recording sites. Significantly greater NBG power reduction across HVAs during retinotopically matched V1 inactivation (2.0 ± 1.0; all areas) versus mismatched inactivation (1.2 ± 0.3; p = 0.0034, Wilcoxon rank sum; n = 26 experiments, 12 recordings in 5 mice). Significantly greater NBG reduction for matched V1 inactivation in RL and LM (p = 0.04 for both, Wilcoxon rank sum). See Fig. S2S for V1 and HVA targeting via intrinsic signal imaging maps.

The strength of NBG across HVAs causally depended upon feedforward thalamic and cortical input. We confirmed strong and prevalent NBG activity in stationary, non-locomoting mice in LGN (~25% of all LGN neurons classified as NBG neurons; Fig. S2N), in V1 (Fig. S2I,M; consistent with our prior studies¹⁸), and also in HVAs(Fig. S2I–L). We expressed channelrhodopsin (ChR2) in thalamic reticular nucleus (TRN) neurons to silence visual thalamus²⁵. We verified fiber placement in TRN, and observed ChR2-expressing axons in LGN (Fig. S2G). Activating TRN rapidly silenced multi-unit activity in LGN and V1 (Fig. S2H). Across all experiments, silencing thalamus reduced residual NBG LFP power more than three-fold across V1, RL, LM and PM versus the interleaved control periods (V1: 7.5 ± 6.8 fold reduction, HVAs combined: 3.43 ± 1.8 fold reduction). Broadband gamma power (30 – 90 Hz; excluding 50 – 70 Hz) was significantly less affected than NBG (V1: 1.7 ± 0.4 fold reduction, p < 0.004; HVAs: 1.5 ± 0.3, p < 2e-6, Wilcoxon rank sum; see Methods). We accounted for these effects by normalizing the NBG reduction by broadband reduction within each experiment (Fig. 2G; NBG reduction ratio). Silencing visual thalamus caused significantly greater NBG reduction than interleaved control periods (Fig. 2G). Since ChR2 was expressed throughout TRN it could have potentially impacted activity in both LGN and secondary visual thalamus. Therefore, in a subset of these experiments, we also naturally suppressed NBG activity by transiently reducing luminance with a black full screen^11,14; this also reduced NBG in V1 and multiple downstream HVAs (Fig. S2 O–R). We next sought direct evidence for V1 contributions to NBG in HVAs.

We inactivated retinotopically defined regions of V1 by driving parvalbumin (PV) interneurons while simultaneously recording in HVAs (targeted with imaging; Fig. S2S). Remarkably, NBG power reduction in HVAs depended upon retinotopic alignment with V1: inactivating V1 sites with receptive fields (RFs) that closely matched those at the HVA recording sites (Fig. 2H; 10.2 ± 2.5° apart) led to two-fold reductions of NBG power, significantly greater than inactivation of V1 sites with RFs distant from the HVA recording sites (RFs 47.4 ± 2.6° apart; p = 0.0034, Wilcoxon rank sum). These effects were not explained by the proximity of the HVA recording to the V1 inactivation site (Fig. S4F). Taken together, these results show that the strength of NBG activity propagation across multiple HVAs depends upon LGN input and retinotopically aligned V1 input. We next sought evidence for pairwise functional connectivity between individual NBG neurons spanning visual areas in the Allen Brain Observatory data.

We found a significantly greater probability of pairwise functional connectivity among NBG neurons versus non-NBG neurons within and across areas. We first examined CCGs between pairs of LGN and V1 neurons (Fig. 3A; Table S1) and defined statistically significant peaks at 1 – 4.5ms indicative of functional connectivity per prior studies ^26,27. We controlled for the potential artefactual influence of synchronized NBG spikes by filtering out any NBG range activity in CCGs prior to statistical comparisons, and confirmed this with simulations (Methods; Fig. S5). We then compared the probability of functional interactions between NBG neurons to those between all neurons (Fig. 3B–D), and also between non-NBG neurons with high firing rates since these provide a more stringent upper bound for functional interactions via cross correlations²⁸ (Methods; Table S1). Across all LGN – V1 pairs, we found a significantly greater probability of functional connectivity among NBG neurons compared to all neurons (Fig. 3B) or compared to non-NBG neurons with high firing rates (p < 1e-30). These trends were even more pronounced within V1, where NBG neuron pairs showed a nearly 5% probability of functional connectivity, significantly greater than the 1% probability among all neurons (Fig. 3C) or the 2.7% probability among non-NBG neurons with high firing rates (p < 4e-8). We then examined pairwise interactions between LGN and HVAs (aggregated together; Table S1) and again found significantly greater probability of functional connectivity among NBG neurons relative to all neurons (Fig. 3D) or versus non-NBG neurons with high firing rates (p < 1e-28). Expanding our view of interactions to include peaks at zero lag (suggesting common input to both neurons), we found further evidence for preferential interactions among NBG neurons: nearly 30% of V1 NBG pairs (409 /1,332) showed evidence for significant common input, versus 5% all non-NBG V1 pairs (6,711 / 136,551; p < 1e-6; see Methods for zero-lag analysis), with similar trends in LGN-V1 pairs (NBG: 0.35% (15 / 4,305); all: 0.03% (28 / 82,588); p < 1e-14) and LGN-HVA pairs (NBG: 0.61% (10 / 1641); all: 0.02% (59 / 249,850); p < 1e-5). Functional interactions involved significant positive and negative CCG peaks among NBG neurons (Table S2), and involved both putative excitatory and inhibitory neuron pairs (RS and FS neurons; Table S3), consistent with prior findings in mouse visual cortex ^22,26 and other rodent neocortical areas ²⁹.

Figure 3. NBG neurons show higher probability of functional connectivity across visual areas.

A. Spike trains of LGN, V1, and HVA neurons were examined for evidence of functional connectivity.

B. Top, pairwise spike cross-correlogram (CCG, 0.5ms bins) of NBG neuron in LGN (~79k spikes) and NBG neuron in V1 (~121k spikes; 161 mins of recording). Grey line shows significance threshold for expected Poisson process (p ≤ 1e-6; see Methods). Bottom, Probability of functional connectivity among LGN-V1 NBG pairs (0.60%; 26 / 4,305 pairs) significantly greater than among all pairs (0.04%; 33/82,588 pairs; p < 1e-30, Binomial t-test, same throughout figure). 1,309 LGN and 3,694 V1 neurons in 58 recording sessions. NBG pairs also showed significantly greater probability of common input (0.35%) versus all pairs (0.03%; p < 1e-14; Table S3).

C. As in B, for pairwise correlations within V1. Probability of functional connectivity among V1 NBG pairs (4.72%; 126 / 2,664) significantly greater than all pairs (1.09%, 2,974 / 273,102; p < 1e-8). 3,694 V1 neurons in 58 recording sessions. NBG pairs also showed significantly greater probability of common input (30.70%) than all pairs (4.91%; p < 1e-14; Table S3).

D. As in B, for pairwise correlations between LGN and HVAs. Example shows CCG between NBG neurons in LGN and RL. Probability of functional connectivity among LGN-HVA NBG pairs (1.04%, 17 / 1,641) significantly greater than all pairs (0.01%, 29 / 249,850; p < 1e-28). 1,309 LGN and 12,435 HVA neurons in 58 sessions. NBG pairs also showed significantly greater probability of common input (0.61%) versus all pairs (0.02%; p < 1e-5; Table S3). All data in B – D from Allen Brain Observatory.

Thus far we have examined NBG oscillations and spike timing during spontaneous activity – how does this relate to visual processing? LGN NBG neurons showed significantly elevated spiking at the onset of gratings versus simultaneously recorded non-NBG neurons (Fig. 4A). Accordingly, downstream V1 NBG neuron populations also fired at significantly higher rates and at shorter latencies than the simultaneously recorded V1 non-NBG neurons (Fig. 4B). Sustained (seconds long) drifting gratings diminish NBG LFP power and drive “classic” 30 – 40 Hz gamma (Fig. S3L; S4J), but we found that grating onsets clearly evoked stronger responses in both LGN and V1 NBG neurons across all tested contrasts, grating orientations, drift directions and frequencies (not shown), indicating that NBG neurons generate enhanced onset responses across a wide range of stimuli.

Figure 4. ON / OFF neurons in LGN spike at distinct NBG phases across LGN, V1, and HVAs.

A. Mean normalized drifting grating responses of all LGN NBG neurons (1.99 ± 0.06, ±SEM; n = 400 cells) significantly greater than LGN non-NBG neurons (1.4 ± 0.06; n = 423; p<1e-6). Raw spike rates significantly different (p<1e-7; Results).

B. As in A, for V1 NBG neurons (n = 107; 3.7 ± 0.2) and non-NBG neurons (n = 745; 2.0 ± 0.9; p=0.02). Raw spike rates significantly different (p<1e-14; Results). Inset, first spike latencies significantly faster in NBG neurons (61 ± 38ms (SD); 82 ± 46ms, p<1e-11)

C. Significant correlation between trial-by-trial pre-stimulus NBG power and grating responses of LGN NBG neurons. Trials sorted by NBG power of all simultaneously recorded LGN NBG cells (Methods). No such relationship in non-NBG LGN neurons (Fig. S4L).

D. PSTH of LGN ON (grey, n = 147) and OFF (black, n = 77) responses to white stimulus across recordings (n = 14). Cells with ON/OFF index > 0.7 (or < 0.3) defined as ON (or OFF) cells. See Methods.

E. Same neurons as D, in response to black stimulus.

F. Scatter plot of ON/OFF index versus NBG neuron preferred phase for all LGN neurons (n = 397, 14 sessions). LGN NBG phase calculated relative to strongest NBG LGN ON neuron ACG (see Methods). Significant correlation between ON/OFF preference and NBG phase preference (Pearson rho = −0.55; p < 1e-30).

G, H. Visual responses of NBG ON (n = 156) and OFF (n = 82) LGN neurons significantly greater than simultaneously recorded non-NBG ON (n = 81) or OFF (n = 36) neurons. PSTHs normalized to mean firing per neuron, mean ± SEM response to white or black full screen flashes. Insets show distributions of peak evoked firing rate for ON (NBG: 20.2 ±16.5 spikes per s, ±SD; non-NBG: 12.8 ±15.6) and OFF groups (NBG: 29.8 ± 25.0; non-NBG: 12.2 ±19.0), both significantly different. ON/OFF neurons defined as in D.

I. LGN NBG neurons show significantly greater specificity for luminance coding of solely white or black stimuli (0.49 ± 0.28) than non-NBG neurons (0.28 ± 0.27; p<1e-22). Neurons with equal response to white and black stimuli have coding specificity = 0.

J. NBG phase of ON and OFF preference LGN neurons (same as D-E), relative to strongest NBG LGN ON neuron (see Methods). Gaussian fits to ON/OFF cell phase histograms show clear preferred phase clustering and separation.

K. Same as J, but LGN spikes referenced to simultaneously recorded V1 L4 LFP. NBG phase is identified with spike-LFP cycle histogram (Fig. 2D).

L. LGN ON versus OFF selectivity is predictable from preferred LGN NBG spike phase relative to cortical LFP. Prediction accuracy (d’) measured per HVA by calculating the distance between means of phase preference Gaussian fits (J and K) divided by the average SD of two fits (see Fig. S3A–H). Discriminability (d’) of ON/OFF neurons by NBG phase shows significant Pearson correlation with hierarchy, and d’ remains above chance level for all areas except PM.

NBG activity also predicted the strength of LGN visual responses: on trials with high pre-stimulus NBG power, drifting gratings evoked more intense spiking in NBG neurons (Fig. 4C; Methods), with no such effect on simultaneously recorded non-NBG neurons (Fig. S4L). The correlation between pre-stimulus NBG power and visual spiking was significantly different between NBG and non-NBG neurons, and persisted when we measured pre-stimulus NBG power in one half of simultaneously recorded NBG neurons and predicted responses in the other half (within session NBG r = 0.1 ± 0.1; non-NBG r = 0.01 ± 0.1; p<1e-23 between groups). Moreover, sorting the trials by NBG power in non-NBG neurons showed poorer correlation to visual responses in NBG or non-NBG neurons (r = 0.08; r = 0.007). This shows that pre-stimulus NBG power in LGN predicts improved visual responses selectively in NBG neurons on single trials.

We next examined the role of NBG neurons for encoding rapid increases and decreases in luminance (ON and OFF responses), a fundamental aspect of vision. Surprisingly, we found that NBG LGN neurons with ON versus OFF preferences (Fig. 4D, E; Methods) showed distinct NBG firing phase preferences, and these phase differences were preserved relative to NBG in the visual cortical hierarchy. We determined the preferred NBG firing phase (during spontaneous activity) for these ON and OFF neurons (Methods) and plotted the relationship between visually evoked ON/OFF index for all neurons (no selection criteria), versus their preferred NBG firing phase during spontaneous activity. We found a strong and significant correlation between the two measures (Fig. 4F).

Could NBG neurons be highly responsive yet more visually selective? Indeed, ON or OFF dominant NBG neurons showed significantly larger responses to full screen white or black flashes than simultaneously recorded non-NBG ON or OFF neurons (Fig. 4G–H). We defined a specificity index (SI) that captured if neurons were equally responsive to both black and white stimuli, or if they were solely driven by either black or white stimuli. Examining all LGN NBG and non-NBG neurons (regardless of ON/OFF index), we found significantly greater specificity for luminance coding in LGN NBG versus non-NBG neurons (Fig. 4I; p < 1e-23). We did not see the same in V1 NBG neurons (data not shown), likely due to recurrent connectivity and retinotopic dependence for ON/OFF selectivity in V1³⁰. Taken together, these findings show that NBG LGN and V1 neurons fire more spikes to rapid luminance changes, and that the improved responsiveness of LGN NBG neurons also carries greater selectivity for luminance polarity.

We then focused on the clearly ON versus OFF dominant neurons (as in Fig. 4D, E) and investigated how these fired relative to NBG phase in LGN, V1, and HVAs. ON versus OFF neuron spikes showed tight and separated NBG phase preferences (Fig. 4J), with a phase offset of π/2 (~ 5ms apart for NBG at 56 Hz). ON cells showed tighter phase locking (lower phase jitter) than OFF cells (ON: 0.85 ± 0.02 radians; OFF: 0.89 ± 0.03; mean ± SD; p=0.003, Kolmogorov-Smirnov test; Methods). Phase separability and locking were not different for locomoting versus stationary trials (Methods). We then plotted NBG LGN spike time histograms relative to the NBG LFP phase within cortical areas and found that NBG ON and OFF neurons also showed clear and distinct phase preferences relative to NBG LFP in V1 (Fig. 4K) and across the hierarchy of HVAs (Fig. S3A–H). Remarkably, the ON vs OFF visual selectivity could be accurately predicted simply from the preferred NBG spike phase of an individual LGN neuron during spontaneous activity (Fig. 4L d’ = 2.2 ± 0.06; preferred phase relative to NBG within LGN). Visual selectivity of LGN neurons was also predictable from their preferred spike phase relative to NBG in LFP of V1, LM, RL, AL, and AM (Fig. 4L), with a significant correlation across the cortical hierarchy (Fig. 4L; compare to Fig. 2C–F). Importantly, the preferred NBG spike phase of LGN neurons relative to cortical LFP remained highly consistent throughout the duration of a recording (preferred phase SD = ±0.65 radians, or ±1.72 ms for NBG at 60Hz; computed from 100 non-overlapping segments of 100s each in a single session). This suggests that LGN ON versus OFF channels maintain a consistent, millisecond timescale phase offset relative to NBG activity in the input layers across V1 and HVAs.

Finally, the results shown here were unaffected when we controlled for potential confounds of volume conduction of LFPs (Fig. S4G, H), and all major findings were replicated when recomputed using only individual LGN neurons with significant NBG ACGs (Fig. S4A–E).

Discussion

Here we showed that narrowband gamma (NBG) oscillations propagate in functionally distinct subnetworks of neurons throughout the mouse thalamocortical visual system. NBG synchronized across LGN, V1, and HVAs in both spikes and LFP. NBG neurons transmitted stronger visual responses with faster timing and greater coding specificity. Taken together, our findings suggest that thalamocortical NBG oscillations could play the role of a “central clock” to coordinate spiking across the mouse visual system.

We found NBG oscillations engaged specific subnetworks of single neurons, and these displayed enhanced functional interactions among one another. These findings relied on examination of hundreds of thousands of simultaneously recorded neuron pairs in a unique public dataset of high-density, multi-site Neuropixels probes in multiple visual areas ^21,22. One important consideration for our findings is the overall sparse prevalence of cortical NBG neurons identified via spike train analysis, particularly in HVAs. Stringent statistical thresholds identified small proportions of functionally connected NBG pairs (~5% within V1), but when we expand these criteria to include pairs with common input (significant peaks at zero lag), this reveals larger fractions of neurons that receive and share NBG activity (~30% of pairs within V1). Further, we observed clear NBG LFP power in multiple HVAs, indicative of shared, coordinated subthreshold activity giving rise to synchronized LFP oscillations. These are unlikely to be detectable if driven only by a very small fraction of functionally connected pairs. Likewise, NBG power in HVAs significantly decreased during optogenetic inactivation of both LGN and V1. All these factors suggest a concerted propagation of NBG input to cortical areas, even if detectable through spike train analysis only in smaller subsets of neurons. The laminar and cellular basis of NBG in HVAs, and its relationship to V1, LGN, and higher order thalamic projections forms an important topic for future study^24,31,32, as does further examination of LGN NBG neurons and their inputs from retina³³.

We found that the strength of NBG activity obeyed the anatomical hierarchy of visual areas and depended upon retinotopy. The simplest interpretation for hierarchical NBG activity is that it reflects the density of feedforward visual connectivity from LGN to V1 to HVAs. In support of this, we found that the strength of spike-triggered NBG LFP activity in HVAs was nearly identical when triggered on V1 versus LGN NBG neuron spikes, Further, V1 inactivation reduced NBG power in HVAs as a function of receptive field alignment. Since NBG in V1 predicts subsequent behavioral detection of spatially localized visual stimuli¹⁸, a testable prediction is that increased pre-stimulus NBG power provides an effective state to synchronize the first stimulus-evoked spikes in retinotopically aligned neurons³⁴ across LGN, V1, and HVAs, leading to improved visual perception in a variety of behavioral contexts and tasks.

Multiple visual response properties of NBG neurons– higher rates, faster latencies, and greater specificity for luminance coding– suggest functional advantages for the first spikes in V1³⁵ and activity in HVAs^19,20 underlying perceptual detection. We further revealed NBG phase coding of ON and OFF pathways; prior work^14,36,37 shows that synchronization and spike timing in ON and OFF pathways transmits the most information about the visual scene. Even though sustained visual stimulation diminishes NBG and drives intracortically generated 30 – 40 Hz gamma¹¹, the precisely timed onset firing of NBG neurons may efficiently encode visual information across neural populations³⁸. The relationship of NBG neurons to stimulus-dependent 30 – 40 Hz intracortical gamma LFP activity ^39,40 for feature coding¹⁷ and contextual processing¹⁵ is an exciting topic for future investigations. Here we established that synchronization and propagation of a subcortically generated narrowband, phase-locked, stimulus-independent oscillation throughout visual cortical areas could provide a substrate for temporal coding of first spikes in ON and OFF pathways throughout the mouse visual system. This provides a novel testbed for resolving long-standing and open questions regarding oscillations, temporal coding, and alternative modes of communication underlying visual perception^3,6,41,42.

STAR Methods

Resource Availability

Lead Contact

Further information and requests for resources and reagents should be directed to and will be fulfilled by the Lead Contact, Bilal Haider (bilal.haider@bme.gatech.edu).

Materials Availability

This study did not generate new unique reagents or materials.

Data and Code Availability

• Matlab data has been deposited at Figshare and is publicly available as of the date of publication. DOIs are listed in the key resources table.

• All original code has been deposited at Figshare and is publicly available as of the date of publication. DOIs are listed in the key resources table.

• Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request.

Key resources table.

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Deposited data
Pre-processed data (from Allen Institute) to reproduce main results	Allen Institute21	https://doi.org/10.6084/m9.figshare.19666314
Pre-processed data (Haider lab) to reproduce main results	This paper	https://doi.org/10.6084/m9.figshare.19666314
Experimental models: Organisms/strains
Mouse: C57BL/6J (Allen Inst.)	The Jackson Laboratory	IMSR_JAX:017320
Mouse: Ai32 × B6PVCre (Allen Inst.)	The Jackson Laboratory	IMSR_JAX:024109; IMSR_JAX:017320
Mouse: Ai32 × Sst-IRES-Cre (Allen Inst.)	The Jackson Laboratory	IMSR_JAX:024109; IMSR_JAX:013044
Mouse: Ai32 × Vip-IRES-Cre (Allen Inst.)	The Jackson Laboratory	IMSR_JAX:024109; IMSR_JAX:010908
Mouse: Ai32 x B6PVCre (Haider lab)	The Jackson Laboratory	IMSR_JAX:024109; IMSR_JAX:017320
Mouse: Ai32 × Scnn1a-Cre (Haider lab)	The Jackson Laboratory	IMSR_JAX:024109; IMSR_JAX:009613
Mouse: GAD2-Cre (Haider lab)	The Jackson Laboratory	IMSR_JAX:010802
Mouse: Ai40 (Haider lab)	The Jackson Laboratory	IMSR_JAX:021188
Recombinant DNA
AAV5-EF1a-doublefloxedhChR2(H134R)-EYFP-WPREHGHpA	This paper	Addgene Cat# 20298-AAV5
Software and algorithms
Matlab (2019b or later)	Mathworks	https://www.mathworks.com/
Kilosort2	Stringer et al.19	https://github.com/MouseLand/Kilosort
Custom code to reproduce main results	This paper	https://doi.org/10.6084/m9.figshare.19666314

Experimental Model and Subject Details

All procedures were approved by the Allen Institute’s Institutional Animal Care and Use Committee and Institutional Animal Care and Use Committee at the Georgia Institute of Technology.

Experimental subjects – Allen Brain Observatory

Recordings from the Allen Brain Observatory – Visual Coding database are fully detailed elsewhere ^21,22. 45 male and 13 female mice (117±13 days old) were used. 32 of 58 subjects had single unit recordings in LGN; we focused on 14 / 32 that had ≥10 NBG LGN neurons recorded simultaneously. In the 14 sessions with ≥10 NBG LGN neurons, we analyzed 316 ± 59 cells recorded simultaneously (all areas) and 51,404 ± 19,331 possible pairwise interactions per experiment. The Dataset for this study was last accessed and compiled on December 30, 2020.

Mouse Strain	n = mice	n = neurons (Total; NBG)	RRID
C57BL/6J	30	(8856; 538)	IMSR_JAX:017320
Ai32 × B6PVCre	8	(2205; 73)	IMSR_JAX:024109, IMSR_JAX:017320
Ai32 × Sst-IRES-Cre	12	(3721; 139)	IMSR_JAX:024109, IMSR_JAX:013044
Ai32 × Vip-IRES-Cre	8	(2653; 59)	IMSR_JAX:024109, IMSR_JAX:010908

Experimental subjects – Haider lab

Detailed methods for neural recording and optogenetic inactivation have been described previously ^18,43. Mice (5 – 8 weeks old; reverse light cycle individual housing; bred in house) were chronically implanted with a stainless steel headplate with a recording chamber during isoflurane (1–2%) anesthesia. After implant surgery mice recovered for 3 days before experimentation. Recordings in Haider lab used male and female mice.

Mouse Strain	n = mice	n = recordings	RRID
Ai32 × B6PVCre	5	18 (26 sessions)	IMSR_JAX:017320, IMSR_JAX:024109
C57BL/6J	6	25 (25)	IMSR_JAX:000664
Ai32 × Scnn1a-Cre	1	1 (1)	IMSR_JAX:009613
GAD2-Cre	3	13 (13)	IMSR_JAX:010802
Ai40	1	2 (2)	IMSR_JAX:021188

Across all datasets, there was no allocation strategy for selecting subjects, since there were no comparisons between experimental groups of separate subjects.

Method Details

Recordings

LGN, V1, HVA Neuropixels recordings – Allen Brain Observatory

Detailed recording procedures are described elsewhere ^21,22. Data was retrieved from the database using their proprietary software development kit (SDK). Data was compiled last on December 20, 2020. Detailed instructions for accessing the database and recording sessions with analysis code will be publicly deposited and linked from the lead contact’s institutional website.

V1 or LGN inactivation and HVA recordings – Haider lab

All procedures were approved by the Georgia Institute of Technology Institutional Animal Care and Use Committee (IACUC). A custom-built stainless steel head post with a recording chamber (11 mm inner diameter) was lightly affixed to the skull using veterinary adhesive (VetBond). Following headplate fixation, a glass coverslip (5 mm diameter, #1 thickness ~0.15 mm) was centred over the representation of V1 and HVAs (centre of window at ~ 2.4 mm lateral to midline and ~ 2.4 mm anterior to lambda) and bonded to the skull using VetBond. Mice were individually housed and monitored for full recovery for at least 3 days before intrinsic signal imaging (ISI) of V1 and HVAs, as detailed in our prior studies ⁴⁴. Briefly, ISI was performed during isoflurane anesthesia with sedation. The cortex was illuminated with green or red light to capture hemodynamic or blood oxygenation signals, respectively. Custom visual stimulation and image acquisition systems computed retinotopic maps of stimulus selectivity for azimuth and elevation, and these were used to compute visual field sign maps that defined boundaries and extent of V1 and the HVAs, consistent with established methods ^45,46. Following ISI, mice were habituated to head fixation for 3 to 5 days before undergoing awake recordings as in prior studies ¹⁸. On recording days, small craniotomies (~100–500 μm) were made in HVAs under isoflurane anesthesia, using the ISI maps and vasculature as references ⁴⁴. The skull overlying two retinotopically distant sites in V1 (~ 60 to 90° apart in azimuth) was thinned to facilitate optical stimulation. Mice were allowed to recover for >3 h, and then acute awake recordings were made with multi-site silicon probes (Neuronexus; A 1×32) spanning all layers of the cortex. Electrodes were advanced ~1000 μm below the cortical surface. The signals were acquired at 30 kHz (Blackrock Microsystems) and filtered at 0.3 – 300 Hz to acquire the LFP signal. LFP power was analyzed on the channel with greatest raw NBG power, typically in L4. Experiments reported here only examined long periods of spontaneous activity (no visual stimulation).

V1 was locally silenced by activating ChR2 expressed in parvalbumin inhibitory neurons as in our prior studies ^18,43. Laser stimulation (6.5mW power, 1s duration with 0.1s onset and offset ramps, inter-trial interval duration randomly selected from 1 – 6s per trial) was confined to a circular spot (~0.2 mm diameter half-width, measured with a beam profiler) using an array of convex and f-theta lenses. This spot was far smaller than the extent of V1 and does not directly encroach on the recording sites in HVAs (Fig. S4F). We were not concerned about our surface stimulation leaking down to TRN/LGN, since the V1 inactivation sites (−3.7 P, +2.5 – 3.5 L from Bregma) were ~ 1mm more caudal than the coordinates targeting TRN/LGN (−2.5 P, +2.5 L from Bregma, see below). Even with a high-intensity 15mW spot (more than double our 6.5mW), PV neurons can be activated laterally up to 1.75 mm away within cortex, while our spot’s 3-D distance to TRN would be 2.7mm (at minimum), making this much larger distance of spread unlikely with even weaker laser power. Further, in prior experiments we directly silenced much larger regions of V1 than here and NBG activity still remained intact in both LGN multi-unit activity and in thalamic currents in L4¹¹. The spot was targeted to portions of V1 that matched / mismatched the retinotopic coordinates of the craniotomies targeted to HVAs (using ISI maps for azimuth and elevation; Fig. S2S–T). Laser stimulation was only delivered on 25–33% of randomly selected trials with the remaining trials serving as the interleaved control trials. At the conclusion of optogenetic experiments, retinotopy was confirmed at the HVA recording sites by presenting briefly flashed bright and dark bars throughout the visual field⁴⁴. Retinotopy of the thinned V1 sites was also subsequently confirmed electrophysiologically the same way. Using azimuth coordinates from both ISI maps and the preferred azimuth of the LFP responses recorded in HVAs, we classified V1 and HVAs retinotopy as “matched” when they were both located within the binocular visual field (0 to 40° azimuth), or both located in the monocular visual field (>55° azimuth). For matched conditions, the V1 and HVA RFs were 9.6 ± 2.6° apart (mean ± SEM), within the average width of single neuron spiking RFs in mouse V1⁴⁷. “Mismatched” experiments paired recordings and stimulation across binocular and monocular locations: here, V1 and HVA RFs were 46.8 ± 3.2° apart, significantly more separated than the matched experiments (p = 2.2e-04, Wilcoxon rank sum test). Importantly, matched versus mismatched experiments varied only in azimuthal (horizontal) coordinates: differences in elevation (vertical) retinotopy between V1 and HVA sites were small and comparable across groups (Matched: 10.0 ± 1.3° apart; Mismatched 8.3 ± 1.3° apart; p = 0.32, Wilcoxon rank sum test; coordinates assigned from ISI maps).

LGN was inactivated by driving the inhibitory thalamic reticular nucleus (TRN), as described in prior studies ²⁵. Surgical procedures for were similar to those for V1 inactivation experiments. In addition to headplate implantation, a small craniotomy for virus injection and/or optical fiber implantation targeted to the left TRN was drilled at 0.3 mm posterior and 2.4 mm medial of Bregma in the left hemisphere. 200 nl of AAV5-EF1a-doublefloxed-hChR2(H134R)-EYFP-WPRE-HGHpA virus (RRID: Addgene 20298; AAV5, 2.2e12 GC/mL) was injected into TRN at 20 nl/minute using a glass pipette mounted to a Nanoject III (Drummond) on a stereotaxic manipulator arm. The pipette tip was lowered 2.9 mm from the brain surface along the axis of the stereotax arm which was tilted 20 degrees past vertical in the anterior direction such that the craniotomy coordinates were slightly anterior of TRN. The pipette was left in place for 5 minutes after the injection before withdrawal. After virus injection, a 0.2 mm diameter optical fiber (0.39NA 70–80% transmission CFML12U, Thorlabs) was inserted 2.65 mm along the same trajectory as the virus injection pipette and cemented in place. Instead of a glass coverslip, the skull surface was covered with a layer of vetbond, clear Metabond (Parkell), and cyanoacrylate glue (Zap-A-Gap Medium CA+, Zap) to maintain clarity for intrinsic signal imaging. A total of 3 GAD2-Cre mice (2 homozygous, 1 heterozygous) were used for TRN experiments. An additional mouse underwent optical fiber implantation without virus injection to serve as a non-opsin control (Ai40). ISI, habituation, and recordings targeted to V1 and the HVAs LM, RL, and PM were performed as described above. In one mouse, LGN was also targeted for recording and the electrode was painted with DiI prior to insertion. TRN was illuminated with blue light through the implanted optical fiber using a blue laser. Mice viewed a gray screen while 1 sec laser pulses of (square step, 0.5 to 1.5 mW) were delivered every 9 seconds. In the non-opsin control animal, 5mW power was used. During TRN optogenetic stimulation experiments, the eye ipsilateral to the optical fiber was constantly and diffusely illuminated to reduce the possibility of direct retinal activation by the laser⁴⁸. In a subset of recordings, full field luminance stimuli were presented consisting of 10 seconds of gray followed by 10 seconds of full screen black or white. Cortical LFP signals were processed as described above. After the final recordings, mice were deeply anesthetized (Euthasol) and intracardially perfused with 4% paraformaldehyde. Brain slices were prepared on a vibratome (50 microns thick) and imaged with a fluorescence microscope to confirm YFP expression, fiber placement, and DiI-labeled electrode tracks.

LGN recordings – Haider lab

Detailed methods for all Haider lab recordings have been described previously ³⁰. Briefly, LGN was targeted with stereotaxic coordinates (−2.5 mm posterior to bregma −2.5 mm lateral from midline). Mice were habituated to the recording environment prior to awake recordings. Single shank electrodes (Neuropixels 1.0 IMEC) were used to record from LGN (n = 23 experiments). Each probe contains 960 channels, of which a subset of 383 were used for recording. Spikes were acquired at 30 kHz and LFP at 2.5 kHz via a PXIe card, National Instruments board, and Spike GLX software. In a subset of experiments (n=3), single shank electrodes (NeuroNexus, A1×32 Poly 3) were used to record from LGN and signals were acquired through a Cereplex Direct (Blackrock Microsystems). Single units were sorted using Kilosort2⁴⁹. In a subset of experiments, multi-unit (MU) spiking activity (Fig. S2H) was computed by summing spikes across all clusters, and normalizing the PSTH by the mean firing rate in the 0.1s prior to laser onset. In all experiments, visual stimuli were shown and unit responses monitored online as electrodes advanced to LGN (~ 3.0–3.2 mm below the dura), and in most instances confirmed with histology. Spatial receptive fields were mapped with black and white squares to confirm functional responses consistent with LGN neurons.

Visual stimuli

Visual stimuli – Allen Brain Observatory

Detailed stimulus parameters are described elsewhere²¹. We analyzed responses to full-field flashes of white and black (0.25 s duration, 2 s inter-trial interval with uniform grey screen) to categorize ON and OFF neurons in LGN, and also analyzed drifting gratings (2 s duration, 1 s gray; multiple directions, orientations, spatial and temporal frequencies) to quantify spiking at the onset of visual stimulation. Although NBG activity depends upon luminance, it is not driven by screen refresh rates: within an experiment, DC or AC LED illumination (across multiple frequencies) drives the same NBG peak at 50 – 70 Hz¹¹.

Quantification and Statistical Analysis

Spike sorting

Spike sorting – Allen Brain Observatory

Neurons were pre-sorted and packaged with several pre-computed quality metrics, as detailed elsewhere ²¹. We plotted histograms of spike waveform widths and observed a clear bimodal distribution in cortical recordings with a partition at 0.42 ms that separated regular spiking (RS) and fast spiking (FS) groups. LGN recordings showed unimodal spike waveform width histograms, with a large majority having spike widths > 0.38 ms (91%; 1,187 / 1,306), including the majority of NBG neurons (91%; 415 / 455). A minority of LGN neurons showed spike widths between 0.2 – 0.38 ms (9%, 119 / 1,306), including just 8% of NBG neurons (40 / 455). We interpret this to suggest that the great majority of NBG LGN neurons correspond to classically described excitatory thalamocortical relay neurons ^50,51.

Visual response analysis

Laminar identification – Allen Brain Observatory

The earliest sink of stimulus triggered current source density (CSD) in V1 corresponds to the site of strongest LGN input in layer 4^18,43,52. This anatomical relationship is less clearly defined for CSDs in HVAs (RL, LM, AL, AM, PM). For consistency, we analyzed NBG power at the earliest stimulus triggered CSD sink channel across V1 and HVAs. All cortical channels were restricted to the hundred channels below the topmost channel expressing electrical activity (1mm total). We defined the earliest sink channel as the one that reached the half-maximum amplitude of the sink with shortest latency (Fig. S2A). This analysis identifies the site of the earliest functional visual input across all areas but cannot clearly identify the anatomical source of this input in HVAs.

NBG LFP power analysis

We quantified residual NBG power of LFP by modifying previously established methods ¹¹. We first fit a linear regression model on the logarithmic power (dB) of the frequency domain LFP (using fitlm.m function in MATLAB). To compute the residual NBG power, we first fit all data points in the frequency range spanning 30 – 90 Hz (excluding 50 – 70 Hz). We then defined the residual NBG power as the average LFP power between 50 – 70 Hz that remained after subtracting the power in the NBG range predicted from the smoothed linear fit. This method was used to calculate raw LFP residual NBG power and stLFP residual power. Residual NBG power was considered in excess when mean power in NBG frequency range was > 1 SD above the linear fit (using movmean.m and movstd.m from MATLAB with 1000 datapoints). As a control measure during optogenetic experiments (Fig. 2G–H), we calculated residual broadband gamma power in the same manner for both control (no light) and laser trials by computing linear fits from 20 – 90 Hz (but excluding 30 – 70 Hz), then measuring residual broadband gamma power from 30 – 70 Hz, consistent with the range for broadband gamma in mouse V1⁴⁰. Statistical comparisons reported in Fig. 2G treat individual sessions of inactivation plus recording as independent experiments. Since a given mouse could contribute multiple sessions, we performed an n-way ANOVA with factors for optogenetic condition (laser or control), visual area, and mouse ID. There were significant main effects for optogenetic condition and visual area (p = 0.035), but not for the mouse ID (p = 0.16).

NBG neuron identification in spike cross-correlograms and autocorrelograms

NBG neuron identification with auto-correlogram (NBG-ACG)

Spike-trains from cells in all visual areas were binned at 0.5ms resolution (‘histcount’ in MATLAB). Autocorrelograms (ACGs) and cross-correlograms (CCGs) were calculated with ‘xcorr’ in MATLAB. Units were analyzed and aggregated across cortical layers.

The first step in NBG neuron identification examined ACGs. We computed frequency domain power of the “sidebands” of the ACG function (lags between 20ms to 120ms; 201 points, 0.5ms bin size) to avoid distortion caused by the central ± 20 ms of the ACG (Fig. S1). NBG neurons identified with the ACG were those with Maximum power _{[50 – 70Hz]} > maximum power _{[40 – 300Hz]}. This bandwidth was implemented to assess the strength of NBG activity relative to a wide range of spectral content, including “classic” broadband gamma power near 40 Hz. Across the analyzed data set (n = 14 sessions, 4424 neurons), 980 (22%) exhibited significant NBG ACG power. We only analyzed NBG neurons identified with the ACGs if they also exhibited significant NBG CCGs (see below) with at least 1 other simultaneously recorded neuron (Table S1). These criteria were implemented to exclude any potentially non-physiological contamination of spike trains visible in the ACG, and to isolate NBG neurons that participated in network interactions resolvable with spike trains (272 / 980 ACG NBG neurons, 27% of those with significant ACGs). These strict criteria likely lead to an underestimate of the effects of NBG neurons exhibiting significant ACG power in the recordings. These spectral analysis methods were used for CCGs (below), and the distributions of NBG-ACG and NBG-CCG neurons is reported in Table S1.

NBG neuron identification with cross-correlogram (NBG-CCG) & controls

As described above, the main findings of this study are based on NBG neuron identification using CCGs. Our reasoning for using CCGs is twofold: 1) low firing rate neurons or those with long refractory periods may not manifest ACGs with oscillatory firing across multiple successive NBG cycles (Fig. S1I), and 2) CCGs isolate NBG neurons that participate in network interactions resolvable with spike trains. It is important to note that CCGs likely reflect a mixture of direct functional interactions and indirect interactions due to common input; this is particularly true among cell types that are known to have little local connectivity, such as among LGN neurons. The CCG analysis conditions spike histograms on joint firing at NBG frequencies across neuron pairs and thus isolates a larger population of NBG neurons than solely ACG analysis (n = 562 identified with CCG analysis; n = 272 also fulfilled ACG criteria; 14 sessions). Importantly, the identification of NBG neurons in CCGs was highly sensitive to spike timing controls. First, jittering spike times (±20ms) significantly reduced NBG power in CCGs and abolished NBG neuron classification (<0.001% false positive rate in n = 30 pairs, 15 sessions, >10k jittered CCGs per pair; Fig. S1A–H). Second, computing spike CCGs between NBG LGN neurons in one recording session (n = 34) and all neurons from a different recording (n = 242 neurons) never yielded a false positive classification of NBG neuron pairs, despite the LGN neurons showing significant NBG power in CCGs within session (>8k pairwise CCGs). Third, the main results of the study are nearly unchanged when analyzing NBG neurons identified solely with ACGs (Fig. S4A–E).

CCGs were calculated with the same resolution (0.5ms) as ACGs, but the central ±50ms of the CCG were analyzed in the frequency domain. CCGs were normalized to have maximum 1 and Fourier transformed. Both neurons of the pair that fulfilled all of the following criteria were classified as NBG neurons: 1) Magnitude _[60Hz] > 2* Max magnitude _{[150 – 1kHz]}; 2) Magnitude _[60Hz] > Mean magnitude _{[30 – 150Hz]} + 2*SD _{[30 – 150Hz]}; 3) NBG magnitude > 2 (unitless quantity, from normalized FFT). The first criterion identifies all CCGs with clear NBG peaks, while the second and third criteria enforce statistical thresholds for NBG power relative to a wide frequency range, including “classic” broadband gamma (30 – 80 Hz) and high gamma (>80 Hz) ranges. We ensured that neurons that passed both ACG and CCG criteria were not double counted as NBG neurons.

Comparisons and controls for NBG neuron identification with CCG and ACG

We tested if the identification of NBG neurons using CCG power was erroneously caused by CCG power driven by neurons with strong NBG ACGs. We computed NBG CCG power using pairings of all simultaneously recorded neurons, with one group comprised of NBG neurons identified by CCG, and another comprised of NBG neurons identified by ACG power (Figure S1K). We assigned NBG CCG power of a neuron as the highest NBG power among CCGs computed from all simultaneously recorded neurons. The distributions of NBG power in neurons identified by CCG or ACG were largely overlapping, but significantly different due to the large number of samples (Fig. S1K). Next, even when we excluded CCGs containing neurons with significant NBG ACGs, the distributions of NBG power in CCGs were still largely overlapping (Fig. S1L). Classifying NBG neurons through CCGs — when excluding neurons with significant NBG ACGs — correctly classified 56 out of 186 NBG neurons identified through CCG criteria using all pairwise CCGs and classified 161 out of 214 NBG neurons identified through ACG criteria.

We estimated the probability of false identification of NBG neurons with CCGs by computing shuffled CCGs between NBG neurons identified across different recording sessions. We performed CCG between NBG-ACG cells in LGN in one session (n = 34 significant NBG ACG neurons) and all cells in a different session (n= 242 NBG neurons). Among these 8228 CCGs, none were falsely identified as a CCG indicating NBG neurons according to our criteria.

We also verified that identification of NBG neurons with CCGs within session depended upon NBG spike timing. We jittered spike times in a NBG neuron identified through CCG, and calculated the CCG between this jittered spike train and a spike train of a neuron with significant NBG ACG power. In this example pair, 0/10k jittered CCGs were identified as containing NBG neurons (Fig. S1F), and across 30 other pairs drawn randomly across experiments, only 1/300k jittered CCGs was erroneously identified as a NBG pair. These results show that false positive identification of NBG neurons in CCGs through “power leakage” of neurons with strong NBG spike ACGs is unlikely.

We also examined the features of NBG neurons in LGN identified with ACG criteria (versus CCG criteria) and found that these were overlapping but significantly different in several regards compared to NBG neurons classified with CCG criteria (Fig S1J), including higher average firing rates (12.2 ± 4.1 vs 10.1 ± 2.9, median ± IQR) and greater amount of NBG power in CCGs (17.7 ± 2.1 vs 16.2 ± 2.6), even when excluding CCGs between pairs of neurons that both passed the ACG criteria (Fig. S1L). Nonetheless, as seen in Fig. S4A–E, the main findings of the study were nearly identical even when only analyzing NBG neurons identified solely with single neuron spike ACG criteria.

Detection of functional connectivity in cross-correlograms

We detected functional connectivity between cells with previously established methods ^26,27,29 with a few modifications. We first calculated cross-correlations between pairs of spike trains, then applied a bandstop filter between 50 – 70 Hz to minimize spurious threshold crossings caused by strong NBG correlated firing (ideal filter applied in frequency domain representation of the CCG through fft.m function in MATLAB). Next, we detected values in the filtered correlogram that deviated from the expected Poisson process significance threshold in a short time scale (1ms – 4.5ms for functional connectivity pairs, 0ms – 1ms for common input pairs). To compute the Poisson significance threshold, we convolved the filtered correlogram with the Gaussian filter (gausswin.m function in MATLAB) with SD = 7ms (total filter length 42ms at ± 3SD). The convolved correlogram approximates the hypothetical Poisson process where the spike count mean equals the variance and has been shown to detect significant deviations from the Poisson expectation at ~ p ≤ 1 e⁻⁶ threshold, as described previously ^26,27. If more than two consecutive points fell above or below the Poisson significance threshold between 1.5ms to 4.5ms (including edges), the pair was identified as one exhibiting functional connectivity. We did not factor in transmission delays (e.g., from LGN to V1) into our acceptance window in order to maintain consistent and conservative criteria for functional interactions both within and across multiple visual areas (only 1 additional LGN-V1 pair would have been included with a larger acceptance window factoring in the delay). Correlograms fulfilling the above criteria in the positive direction were classified as excitatory interactions, and in the negative direction classified as inhibitory interactions (Table S2). Correlograms fulfilling the above criteria with peaks at 0ms – 1ms were classified as common input pairs. We only considered significant common input pairs with positive crossings of the significance threshold and excluded CCGs with non-physiological profiles (i.e., exceedingly high peaks in the central bin). These enhanced functional interactions among NBG neurons were not simply explained by high firing rates: the probability of functional interactions among NBG neurons was markedly and significantly greater than among non-NBG neurons with high firing rates (those with rates greater than the mean of NBG neurons in the same area; Table S3, see ‘Conditional probability analysis‘). In V1 and HVAs, many NBG pairs with significant functional interactions involved pairs of regular spiking (RS) and fast spiking (FS) neurons (Table S3). High-density sampling with Neuropixels probes likely facilitated isolation of spikes and interactions between FS and RS neurons compared to prior studies that used less dense electrodes and found NBG signals mostly in excitatory networks¹¹. This could also be due to the enhanced detectability of functional interactions in cross-correlations of spike trains from highly active FS putative inhibitory interneurons, as in prior reports ^26,29.

Controls for potential effects of shared NBG spikes in CCGs

Using simulations, we verified that the artificial addition of synchronized NBG spikes (from a common upstream NBG neuron spike train) did not induce false positive detection of functional interactions between neuron pairs (Fig. S5). We simulated a NBG neuron spike train with firing rate, inter-spike interval statistics, and NBG oscillatory power and frequency that matched real NBG neurons identified with ACGs. We then sampled spikes randomly from this simulated NBG neuron spike train and added them to neuron pairs to assess their influence on CCGs and detection of functional interactions. First, we added simulated NBG spikes to spike trains of non-NBG neurons that showed statistically significant CCG peaks (measures of functional interactions). Adding NBG spikes to these neurons caused their ACGs to show NBG oscillations (as expected), but simply bandstop filtering the resulting CCG at NBG frequencies (50 – 70 Hz) removed the influence of the simulated NBG spikes and recovered the original CCGs along with detection of a statistically significant CCG peak (true positive). We then added NBG spikes to non-NBG neurons that did not show statistically significant peaks in their CCGs. In 3 example pairs, the artificial NBG spikes again induced NBG oscillations in ACGs and CCGs, but bandstop filtering the CCG at NBG frequencies always recovered the original (non-significant) CCGs (true negatives). We were thus reassured that this filtering procedure effectively minimized any potentially spurious effects in CCGs due to shared NBG input.

Spike-LFP cycle histogram analysis

We used two complementary methods to assess spike-LFP coherence: LFP cycle histograms ⁵³ and spike-triggered LFP. Cycle histograms were constructed by first bandpass filtering the LFP in the NBG range (50 – 70 Hz, bandpass.m function in MATLAB) and marking the location of the local maxima (findpeaks.m function in MATLAB). Then, we aligned the relative location of LGN spikes between the local maxima of the bandpass filtered LFP (spike phase; Equation below) Lastly, we normalized the histogram of these spike phases to have maximum amplitude 1 (using histogram.m function in MATLAB; binsize = 0.05 radian, from 0 to 2π)

$$\text{spike}\text{phase}=\frac{\text{spiketime}-\text{previous}\text{local}\text{maximum}\text{of}\text{LFP}}{\text{next}\text{local}\text{maximum}\text{of}\text{LFP}-\text{previous}\text{local}\text{maximum}\text{of}\text{LFP}}$$

We computed three metrics from cycle histograms by fitting a cosine (by extracting the first Fourier component of the histogram using fft.m function in MATLAB). The Cycle SNR (Fig. 2D) is the power ratio of fitted cosine and the residual (using snr.m function in MATLAB). Cycle phase is the phase of the fitted cosine (using angle.m function in MATLAB). Cycle histograms showed greatest power relative to LFP in V1, followed by lateral (RL, LM) then medial HVAs (PM, AM; Fig. 2E). Cycle histograms of NBG neurons showed significantly greater power than non-NBG neurons (Fig. 2E dashed line) in all areas (V1: p < 3e-5; RL: p < 2e-3; LM: p< 2e-4; AL: p < 4e-3; AM p < 4e-3) except PM (p = 0.14). This relationship showed a significant correlation in NBG neurons (Fig. 2F) and was significantly steeper than the same measure in non-NBG neurons (data not shown; p<5e-7).

Spike triggered LFP analysis

Spike triggered LFP (stLFP) was computed as follows:

$$stLFP=\frac{1}{n_{spike}}\sum _{s=1}^{n_{spike}}LFP\left(t+t_s\right)$$

$n_{spike}$ is the total number of spikes of a neuron used for triggering LFP vectors. stLFP vectors from raw LFP at the earliest current sink channel (identified from CSD analysis; Fig. S2A–F) were extracted surrounding the spike time from [−0.2s, 0.25s]. The time step 𝑡 was 8ms (i.e., LFP vector sampled at 1250 Hz), consistent with prior work on stLFPs ⁵⁴.

Correlation of NBG activity and visual hierarchy analysis

We computed correlation coefficients between the above-mentioned metrics and the anatomical hierarchy scores. Correlation coefficients (Pearson) and p-values were calculated with the corr.m function in MATLAB. Anatomical hierarchy scores of visual areas were taken from prior Allen Brain studies ^22,24. Numerical scores were extracted from Figure 3. C, F, I, L in ²² using GRABIT.m (MATLAB Central File Exchange).

Correlation of NBG activity in HVAs versus distance to V1 (control for volume conduction)

To control for the potential influences of LFP volume conduction ⁵⁵ from V1 to HVAs, we assessed the relationship between stLFP power (and cycle histogram SNR) in HVAs as a function of distance from V1. Here, we analyzed a subset of recordings in the Allen Brain Observatory dataset: those with ≥10 LGN NBG neurons, clear LFP / CSD in V1, and all V1 and HVA recording locations verified with the Allen Mouse Brain Common Coordinate Framework (CCF; see https://atlas.brain-map.org/). In this subset (12 out of 58 total sessions), 12/12 recordings had probes in V1, 10/12 in RL, 6/12 in LM, 8/12 in AL, 7/12 in PM, and 10/12 in AM. We computed the distance between the earliest sink channel in each HVA (Fig. S2A–F) relative to the earliest sink channel of V1 (Layer 4). We then assessed correlations between the stLFP and cycle histogram SNR for HVAs (Fig. 2B–E) relative to distance of HVAs from the V1 NBG source. The stLFP and cycle histogram SNR in HVAs was poorly correlated with distance to V1 (Fig. S4G–H). Instead, stLFP and cycle histogram SNR were both strongly and significantly correlated with the functional hierarchy position of HVAs (Fig. 2C–F).

Conditional probability analysis

We analyzed the conditional probability of functional connectivity between NBG pairs, and pairs formed by non-NBG neurons (Fig. 3). The analysis consists of counting the number of functional connectivity pairs among cells of interest then dividing by the total number of possible pairs among cells of interest. For the total number of possible pairs involving short-latency lagged interactions (1 – 4.5ms), the total possible connectivity among pairs was bidirectional n*(n−1); for the total possible number of common input pairs (0 to 1ms lagged), within area connectivity is not bidirectional among elements so was calculated as n*(n−1) / 2. Common input pairs across areas were bidirectional and the total possible connectivity among pairs was calculated as n*(n−1).

Then, we computed significance using a binomial t-test with the formula below:

$$\text{Binomial}\text{significance}=C_k^n\cdot {\text{p}}^{\text{k}}\cdot {(1-\text{p})}^{\text{n}-\text{k}}$$

where n is total number of possible pairs, C denotes combination, k is number of functional connectivity (or common input) pairs, and p denotes the null hypothesis probability. This null hypothesis was computed as the probability of functional connectivity among non-NBG neurons (Fig. 3), or among non-NBG neurons with high firing rates (Results). Since correlations increase with firing rates ²⁸ this provides a more stringent “upper bound” null hypothesis for assessing NBG neuron functional interactions. In each area under consideration, we classified non-NBG neurons as “high firing” if their rates were greater than the mean rates of NBG neurons exhibiting significant CCGs in that same area. These null hypothesis estimates found P(functional connectivity | non-NBG & high firing rate) as 0.02% (LGN to V1), 2.7% (V1 to V1), and 0.01% (LGN to HVA).

We also tested the probability of functional connectivity among the “network” of ON or OFF dominant NBG neurons from LGN to HVAs. We found that there was no significant difference for LGN ON or OFF neurons to show preferred functional interactions with ON or OFF neurons in V1 and HVAs (defined by their responses to full screen flash stimuli, described below). That is, NBG ON and OFF neurons in LGN showed equal probabilities of functional interactions with both ON and OFF neurons in downstream cortical areas (p = 0.7, Wilcoxon rank sum tests), as expected since most cortical neurons typically show a broad range of ON/OFF selectivity ³⁰.

ON/OFF classification and analysis in LGN

NBG neurons in LGN were classified as ON or OFF preference cells from the transient response to luminance increments and decrements during a full screen flash stimulus. Each cell had a total of 150 trials of stimulus responses to both luminance increments (ON; grey to white and black to grey) and decrement. (OFF; grey to black and white to grey). The number of spikes transiently responsive to the luminance change (25 – 75ms after the stimulus) were counted and used to compute ON/OFF index.

$$\frac{\text{ON}}{\text{OFF}}\text{index}=\frac{\text{Nspikes}\text{to}\text{ON}}{\text{Nspikes}\text{to}\text{ON}+\text{Nspikes}\text{to}\text{OFF}}$$

Cells with ON/OFF index > 0.7 were identified as ON dominant cells and those < 0.3 as OFF dominant cells, similar to prior studies ⁵⁶. Poorly responsive NBG LGN cells (evoked response < 2.7 spikes/sec) were excluded. Out of 400 neurons, 147 cells were identified as ON dominant and 77 as OFF dominant (14 recording sessions). NBG phase and pre-stimulus power analysis

We computed NBG phase of each NBG neuron in LGN in two different ways: through cortical LFP triggered cycle-histograms (described above in ‘Spike-LFP cycle histogram analysis’), and through spike cross-correlation analysis. For this, we performed cross-correlations (0.5ms bin; window ±50ms) among all simultaneously recorded NBG LGN neurons. We computed CCGs between each NBG LGN neuron and one reference NBG neuron in the session (the ON cell with strongest NBG power in its ACG within the session). We bandpass filtered the correlogram in the NBG range (50 −70Hz; bandpass.m function in MATLAB). NBG phase was defined as the Hilbert phase at 0ms lag. (hilbert.m and angle.m functions in MATLAB). We estimated the mean and standard deviation of ON/OFF NBG phase histograms within each area by fitting Gaussian functions (Fig S3A–H). We first normalized each ON/OFF NBG phase histogram by transforming the range of phase from [−π to π] to [(circular mean −π), (circular mean+π)] to account for circular phase. Circular mean was calculated using meanangle.m (MATLAB Central File exchange). Then we fit Gaussian distributions to the histograms using fitdist.m function in MATLAB with “normal” option. The negative correlation of preferred phase versus ON/OFF index (Fig. 4F) was broadly preserved in V1 (r = −0.1; p = 0.3) and HVAs (r = −0.5, p <1e-3), suggesting that spike timing among neurons with similar ON/OFF preference could remain broadly aligned to NBG cycles in LGN, V1, and HVAs. To assess phase-locking of ON versus OFF neurons, we computed the standard deviation of preferred phase distributions (i.e., spread of histograms in Fig. 4J,K) across sessions (n = 14), using leave-one-out estimation. Analysis of pre-stimulus NBG power effects on stimulus responses (Fig. 4C) calculated NBG power from PSTHs using spikes of all simultaneously recorded NBG neurons. Distributions of correlations between trial-by-trial pre-stimulus NBG power and stimulus responses (in both NBG and non-NBG neurons) were calculated across 300 random subsamples of 1K trials each. To compare NBG phase separability and locking (Fig. 4F) for locomoting versus stationary trials in the Allen database, we separated trials by locomotion state. Due to the increase in firing during locomotion¹⁰, we subsampled spikes during locomotion bouts to match the number of spikes between the two conditions (using randsample.m function in MATLAB). For the same average spike counts, phase separability and locking in ON and OFF neurons were not significantly different when mice were locomoting versus stationary (locomoting vs stationary preferred phase: ON, −0.2 ± 1.0 vs −0.2 ± 1.0 (mean ± SD), p = 0.94; OFF: 1.9 ± 1.0 vs 1.8 ± 1.1, p = 0.25; paired t-tests).

HIGHLIGHTS.

• Narrowband gamma (NBG; 50 – 70 Hz) spreads throughout the mouse visual system

• NBG in higher visual areas depends on retinotopically aligned V1 activity

• NBG promotes stronger, faster, and more selective visual spiking

• NBG phase precisely aligns spikes of LGN ON vs OFF neurons