Two adult male macaque monkeys (Macaca mulatta) were implanted with titanium head posts and recording chambers above areas V1/V2 under aseptic conditions. All procedures were approved by the local authorities (Senator für Gesundheit, Bremen, Germany) and followed the German Animal Welfare Act (TierSchG) and the European Council directive 2010/63/EU for laboratory animal care and use.

During recording sessions, monkeys sat in a custom-made primate chair in front of a 20-inch CRT monitor (monkey I: 93 cm, monkey II: 95 cm) with a display resolution of 1,280 × 1,024 pixels and a refresh rate of 100 Hz (ELSA Ecomo 750). Eye position was monitored by video-oculography (IScan Inc., Woburn, MA, United States). See Drebitz et al. (2018), for details of the experimental setup.

Semi-chronic recordings in cortical area V1 were performed with linear multi-contact probes. Each probe had 16 or 32 electrodes with a center-to-center spacing of 100 μm and a pointy tip (ATLAS Neuroengineering bvba, Leuven, Belgium). The iridium oxide-coated electrodes had an impedance of 0.25 MΩ at 1 kHz and a diameter of 15*15 μm. A small custom-made microdrive was fixed to a grid in the recording chamber to hold and move the probe in a direction perpendicular to the cortical surface. After each recording session, the probe was retracted until the tip was located just below or in the dura mater. It was repositioned within the cortex at the start of each subsequent session. We conducted up to five sessions before removing the microdrive from the recording chamber, replacing the linear probe if necessary, and placing it in a new position in the chamber. At the initial probe insertion, neuronal signals were monitored to detect the penetration of the dura and the entering of the individual electrodes into the cortex to control for accurate depth placement. Using an automatic visual stimulation and data analysis procedure, we simultaneously mapped the receptive fields (RF) of all recording sites along the probe based on their evoked entire spiking activity and the induced γ-band power in the LFP in response to a moving bar (see Drebitz et al., 2019 for details). Given the semi-chronic recording approach, the receptive field (RF) locations remained consistent across successive sessions for which the same probe was used at the same place. Nevertheless, prior to the recordings in these successive sessions, we confirmed the previously observed RF locations by manually mapping the RFs using a bar stimulus.

Neuronal signals were amplified 20,000-fold for monkey I and 16,000-fold for monkey II with a wideband preamplifier (MPA32I) and a programmable gain amplifier (PGA 64, 1–5,000 Hz; Preamplifier and PGA, Multi-Channel Systems GmbH, Reutlingen, Germany) and subsequently digitized with a 12-bit ADC at 25 kHz sampling rate. The signals of all recording sites were referenced against the titanium recording chamber (27 mm diameter) that was in contact with bone and intracranial tissue.

During V1 recordings, the animals performed a demanding shape-tracking task requiring the animals to attend to the continuously changing shape of a target stimulus and respond to the reappearance of its initial shape while fixating a central fixation point (FP) (Drebitz et al., 2018) in the presence of distractor stimuli. The appearance of this shape at one of the distractor stimuli had to be ignored. The stimulus configuration for the task encompassed up to four stimuli, isoeccentrically arranged around the FP (Figure 1A). One of the four stimuli was centered on the overlapping RFs of the neurons recorded in the V1 column. A second stimulus resided close to but outside of the V1 RFs. The other two stimuli were placed at point mirrored locations in the opposite visual hemifield with respect to the central FP (Figure 1A). For estimating neuronal responses without a nearby stimulus and to control that the nearby stimulus is located outside the RFs, in half of the trials either the stimulus in the RF or the stimulus nearby were not shown. Before the trial onset, a spatial cue was presented in the form of an annulus with a diameter of 1°, a line width of 0.25°, and a luminance of 3.8 cd/m². This cue indicated the specific location where the behaviorally relevant stimulus would appear during the subsequent trial. The two stimulus locations within or nearby the recorded V1 RFs were cued with 30% probability, respectively. The two stimulus locations in the opposite hemifield were cued with 20% probability, resulting in three task conditions. Attention is focused either on the stimulus within the RFs of the recorded V1 site (‘attend in’ condition), on the stimulus close to the recorded V1 site (‘attend nearby’ condition), or at a stimulus in the opposite hemifield (“attend away” condition, see Figure 1B).

The cue appeared with the FP (square, 0.15°x0.15°, 2.45 Cd/m²) on the dark monitor (background luminance: 0.03 Cd/m²) and vanished after the monkeys initiated the trial by pressing a lever. After a baseline period of 1,050 ms, the complex-shaped stimuli (~1°–1.5° diameter, line width 0.25°) appeared, and following a static period (Figure 1B), they started to change their shape, by morphing continuously into other complex shapes. One morph cycle (MC) lasted one second, i.e., the time for morphing from one shape into the next. When the initial shape of the cued target stimulus reappeared at the end of MCs 2, 3, or 4, monkeys had to release the lever within a response window (Figure 1B, red rectangle) of 550 ms (510 ms), starting 380 (310) ms before the MC ends and terminating 170 (200) ms after its end for monkey I (monkey II). Shapes for each stimulus were drawn randomly from a set of 12 shapes for monkey I, two of which could become the target shape, and from a set of 8 shapes for monkey II, all of which could become the target shape. To heighten task difficulty, the stimulus luminance changed randomly with each frame of the screen. For this, a random, integer gray pixel value was selected before each frame update [0 255] corresponding to luminance changes from 0.02 to 45.4 Cd/m², on average 3.8 Cd/m² (for details: Grothe et al., 2018; Lisitsyn et al., 2020). From trial initiation until lever release, monkeys had to hold their gaze within a window of 1° diameter centered on the fixation square. If they did so and responded within the response window to the reappearance of the target shape at the cued stimulus location, they received a water reward. Otherwise, the trial ended without reward.

All analyses described below were performed with custom MATLAB scripts running on MATLAB v. 2020a (MathWorks, Natick, MA, United States). The violin plots were created using an open-source MATLAB script available at https://github.com/bastibe/Violinplot-Matlab (Bechtold, 2016).

We recorded neuronal data in 14 sessions for monkey I and 18 sessions for monkey II. If not stated otherwise, data from MCs 2 and 3 of correctly performed trials was analyzed. If an MC terminated with the reappearance of the initial shape of the target stimulus, only data up to 150 ms before the animal released the lever were used.

For the power analysis of CSD and LFP data across the cortical depth and spectrolaminar profiles of phase coherence (PhC) between two CSD signals, all available electrodes of each session for each monkey were used.

Subsequent analyses required the signals from the electrode representing the hotspot of γ-band activity within each of the three major domains of the column, i.e., the supragranular, the granular, and the infragranular layers. For this purpose, we first computed in all 32 sessions the 1/f corrected (see Methods: Spectral analysis) power spectrum of the CSD signals in MCs 2 and 3 for each of the five electrode positions in the granular layer, the five adjacent electrode positions above in supragranular layers, and the five electrode positions below in infragranular layers across all correctly performed trials on the attend-in condition with four stimuli present. Then, we computed the mean spectrum across sessions for each of these 15 electrode positions of each monkey. Taking the mean spectrum across these spectra for the 15 electrode positions allowed defining the borders of the frequency window for γ-power calculation as the width of the PSD peak in the γ-band at half height of its maximum value (42.5–130.5 Hz in both animals). Finally, we computed the mean γ-power within these borders from each electrode’s power spectrum in each session. In the resulting depth profile of CSD γ-power of each recording session, we searched in each of the three domains for a discrete peak flanked at both sides by electrode positions of the same domain with lower γ-power. In the few sections without such a discrete peak, we searched for a peak hidden in the flank of another larger peak in neighboring regions, as indicated by a flattening in such a flank, but without a local maximum. Therefore, we calculated the gradient of the γ-power between the section’s electrodes. In the case of a local minimum of the gradient at a rising flank or a local maximum of the gradient in a falling flank, we selected the first or second electrode from the electrode pair with the gradient’s extremum, respectively. If both approaches failed, no electrode was selected for the investigated domain of the column and session.

To analyze phase coherence between CSD and ESA signals, we used the CSD signals from the hotspot electrodes of the three (or fewer) domains. To avoid analyzing phase coherence based on recording sites with no or only weak responses to the stimulus, we excluded electrodes with a CSD γ-power during MCs 2 and 3 less than 2.5 times the CSD γ-power during the baseline period. As a source for ESA signals, we selected two-thirds of all electrodes in each session with the strongest ESA γ-power in MCs two and three for the ‘attend in’ condition.

For analyzing attention-dependent modulations of the power and dominant frequency in the γ-band, we used the 26 sessions with a behavioral performance of 65% or better (disregarding trials terminated by breaking fixation).

To determine the time-resolved CSD signals for each trial, we first extracted the local field potentials (LFP) by filtering the recorded broadband neuronal signals with a low-pass FIR filter (pass: < 160 Hz, stop: 300 Hz, suppression: 80 dB) in forward and backward direction. The resulting LFPs were subsequently downsampled to 1,000 Hz and used for calculating the CSD signal with the spline inverse CSD method (iCSD), as described by Pettersen et al. (2006). This approach supposes that the recorded LFPs result from current sources within the cortical column that are evenly distributed in cylindrical discs stacked on top of each other and that the CSD varies smoothly in z-direction along the column’s center axis (perpendicular to the cortical layers) as described by a set of cubic polynomials. For the present analyses, we assumed a disc radius of 500 μm and an isotropic medium with a conductivity of 0.4 S/m (Logothetis et al., 2007). The depth profile of the CSD was calculated and spatially smoothed with a Gaussian filter (SD 200 μm, in a finite window of 1,000 μm) for each point in time using the open-access software provided by and available at Pettersen (2005). This provided for each electrode and trial a CSD signal as a function of time. For the discrimination of layers based on source and sink locations, we averaged the LFP within the first 250 ms starting after stimulus onset (see Figure 1B) across the trials of all conditions with a stimulus in the RF before computing the CSD using the spline iCSD method.

To identify the position of the individual electrodes of a probe with respect to the cortical layers and align the electrodes between recording sessions according to their cortical position, we combined two procedures:

First, we used the spatiotemporal pattern of the CSD calculated with the spline iCSD method from the average field potential response to stimulus onset as described above and identified the border between layer III and layer IV based on the inversion of the CSD’s polarity from the source closest to the cortical surface to the initial sink below (see Figure 1D, left panel for an example). Second, we calculated a correlation matrix containing for each pair of electrodes the average Pearson correlation between their CSD signals during the baseline period (Figure 1B; see also Senzai et al., 2019). For this purpose, in each trial, the time-resolved CSD signals of all electrodes were low-pass filtered with an FIR filter pass: ≤ 15 Hz, stop: 20 Hz, suppression: 35 dB) in forward and backward directions within the time window from 100 ms after the start of the baseline period until 100 ms before its end. Subsequently, the Pearson correlation coefficient between these low-frequency CSD signals was calculated for each trial and pair of electrodes. Finally, the median value of each electrode pair’s correlation coefficients across trials provided an entry into the matrix of median correlation coefficients (see Figure 1D, right panel for an example). The correlation matrix usually delivered a sharp decrease in correlation around the border of layers III/IV and helped to refine the results of the first procedure. Recording sites above the border between layers III and IV were categorized as supragranular. Recording sites located within 0.5 mm below this border were categorized as granular, and sites further below as infragranular (Maier et al., 2010; Van Kerkoerle et al., 2014; Ferro et al., 2021).

To investigate the synchronization between spiking activity and the CSD signals, we estimated the entire spiking activity (ESA), which is a continuous signal and which does not reject spikes with small amplitude by thresholding (for a comparison of MUA and ESA, see Drebitz et al., 2019). Since ESA not only encompasses spikes surpassing a set threshold but includes the full range of population spiking activity, it preserves information that might otherwise be discarded, improving spectral frequency analysis of correlated activity. ESA is particularly beneficial for data with a low signal-to-noise ratio, where spike amplitudes and background noise are difficult to distinguish (Brosch et al., 1997; Drebitz et al., 2019).

We obtained the ESA by first filtering the neuronal data with a high-pass FIR filter (pass: > 400 Hz, stop: 300 Hz, suppression: 80 dB) in forward and backward directions to obtain the high-frequency components of the signal. The high-passed signal was full-wave-rectified and low-pass filtered with the same FIR filter as used for obtaining the LFP in forward and backward directions. Subsequently, the ESA signal was downsampled to 1,000 Hz.

The spectral decomposition of CSD, LFP, and ESA signals was performed by convolving them with complex Morlet’s wavelets, Ψ(t,f₀) = A exp.(−t²/2σ_t²) exp.(2iπf₀t), with σ_f = 1/πσ_t and f0/σf=6 (Kronland-Martinet et al., 1987; Torrence and Compo, 1998; Taylor et al., 2005). The wavelets were normalized such that the total energy was 1, requiring the normalization factor A to be σtπ−1/2. The center frequencies f₀ of the daughter wavelets were set between 5 and 160 Hz based on a scheme of Torrence and Compo (1998). The time-frequency resolved power spectral density was then calculated by taking the square of the absolute value of the result of the convolution and dividing it by the Nyquist frequency of 500 Hz (Taylor et al., 2005; Drebitz et al., 2018). If not stated otherwise, we corrected for the 1f− bias of the PSD by multiplying the PSD values with the central frequency of the corresponding wavelet. Time-frequency resolved power spectral density was calculated separately for each session and attentional condition. Power spectra were calculated by taking the mean of all power values for the same frequency across time within MCs 2 and 3.

The dominant γ-frequency for the CSD signal of the hotspot electrode (see Data selection) of a given columnar domain was calculated for each trial based on the average γ-oscillatory cycle period. Only γ-oscillations with more than average amplitude during MCs 2 and 3 were considered. We first filtered the CSD signals with a broad bandpass filter (FIR-filter, passband: 35 – 120 Hz, stop frequencies: 25 Hz and 140 Hz, suppression: 20 dB) in forward and backward directions. The band-limited CSD signals were Hilbert transformed to calculate instantaneous phase φt and amplitude At of the γ-oscillations (Le Van Quyen et al., 2001; Rosenblum et al., 2001). The time course of the amplitude At was smoothed with a Gaussian filter (σ=10ms). In time spans of at least 40 ms duration in which the amplitude continuously exceeds the median amplitude value during MCs 2 and 3 of the trial, we determined the points in time when the γ-oscillations’ troughs occurred, i.e., when the instantaneous phase φt crosses π. Since this happened usually between the phase values sampled at discrete times, we determined the crossing of π using linear interpolation. The period lengths of the γ-oscillatory cycles were calculated as the time difference between successive troughs. Few periods incompatible with frequencies within the filter’s pass band (i.e., >1/35 s or < 1/120 s) were discarded. For each session and each attentional condition, we determined the median γ-cycle period based on all periods calculated as described above. We calculated the dominant frequency as the reciprocal of the median γ cycle period.

We quantified the phase coherence (PhC) at time t and frequency f using the phase locking value (PLV), as shown in Equation 1:

The phase values φCSDtf and φESAtf at times t and frequency f were obtained from the convolution of CSD- and ESA-signals with complex Morlet’s wavelets (see Spectral Decomposition).

To allow for the comparison of measurements with different numbers of trials, we corrected the PhC values for their trial-number-dependent bias by subtracting the expected value (EV) as described in Equation 2. The EV is the level of PhC expected for N trials with random phase relations (Grothe et al., 2012), which is given by:

Equation 1 describes the PhC calculation for a CSD-ESA signal pair. The calculation for two CSD signals follows the same procedure, with the second CSD signal replacing the ESA signal.

To describe the PhC between the CSD signal of a hotspot and ESA or CSD signals across the entire column, we calculated spectrolaminar PhC profiles using ESA data from all channels surviving the inclusion criteria (see Data selection). For the spectrolaminar PhC profiles based on CSD signals across the column, we excluded electrode combinations involving the same signal (i.e., the electrode representing the respective domain’s hotspot) as well as signals from the two adjacent electrodes from further analysis. Average PhC depth profiles for each domain’s hotspot were calculated across the PhC depth profiles of all recording sessions of both animals aligned according to their depth by taking the median for the PhC with corresponding depth and frequency. For the average PhC depth profiles based on CSD signals across the column, channels at a given depth with data from less than two-thirds of the total recording sessions (typically around the hotspot taken as reference) were omitted. To quantify attentional modulation of PhC between the hotspots in the three laminar domains, we calculated the average γ-PhC (42.5–130.5 Hz) during MCs two and three for the ‘attend in and ‘attend nearby’ conditions separately for each recording session and laminar domain pair in both animals. For each pair, the first CSD signal was taken from the hotspot electrode within the respective domain (for more details, see “Data Selection”). The second electrode, located in a different laminar domain, was selected based on the highest γ-PhC (average γ-band PhC during ‘attend in’ conditions, MCs 2/3) with the hotspot electrode.

Before pooling data across animals, we ensured that the data from both animals was not statistically different (p > 0.05). Wilcoxon rank-sum tests were conducted for each analysis involving pooled data to verify that the datasets were comparable.

The statistical procedures to determine significance are denoted alongside each comparison in brackets. These procedures included Kruskal-Wallis tests for assessing significant differences between multiple independent sample populations (at p < 0.05), initially conducted on unpaired data to address the multiple comparisons issue due to comparisons between three hotspots. Subsequently, Dunn’s tests were applied to identify significantly distinct sample groups (p < 0.05). For the groups where Dunn’s test indicated significance, we conducted a one-sided post-hoc Wilcoxon rank-sum test, providing the pertinent statistical parameters.

For the statistical evaluation of paired data, we initially conducted a Friedman’s test to determine whether there were statistically significant differences among the related groups. Subsequently, we employed post-hoc Wilcoxon-signed rank tests (two-sided). To control for the type I errors due to multiple comparisons, we applied the Tukey–Kramer, or Bonferroni corrections to adjust the p-values.

To examine the vertical profile of oscillatory neuronal activity in the γ-band within V1 columns, we calculated the time-resolved CSD along a linear multi-contact probe and its power spectrum during MCs 2 and 3. The depth profiles of these power spectra for the individual sessions of both monkeys indicate the presence of several well-separated hotspots of high power in the γ-band along the recorded columns’ vertical axis (see Figure 2A, left column for examples). The observations from individual sessions were confirmed after averaging the aligned and normalized power depth profiles of each monkey’s recording sessions. We observed a distinct peak within each of the three columnar domains (supragranular, granular, and infragranular) in 30 out of 32 recording sessions from both animals. However, in two sessions, either a discernible peak in the granular or infragranular domain was absent. On average, across all sessions, we found three distinct peaks of γ-power for each animal, one in each major laminar domain of the column, i.e., in the supragranular, the granular, and the infragranular layers. The laminar profile of the maximal γ-power across all electrodes confirms the presence of these three peaks in both monkeys (Figure 2B, right column, red curves). For comparison, we show the corresponding depth profiles of power spectra calculated from the LFP signals. Inspection of the profiles obtained for the individual sessions (see Figure 2A, right column for examples) shows, on the one hand, that part of the peaks observed in the corresponding CSD profiles occur at similar locations and frequencies in the LFP profiles. On the other hand, the regions with enhanced γ-power are spatially much more extended than the more localized CSD peaks. In line, the average depth profile of LFP spectra (Figure 2B, middle column) and the laminar profile of its maximal γ-power at each electrode (Figure 2B, right column, blue curve) show enhanced γ-power across extended regions along the column, with little evidence for three separate hotspots.

After identifying the three hotspots of γ-oscillatory activity in V1 columns, we asked whether they reflect three local groups of neurons operating during stimulus processing as a single, unified entity or, to some extent, functionally distinct ensembles. Therefore, we compared characteristics of neuronal γ-rhythmic activity between the three hotspots of V1 columns that could provide information on the mode of operation of the columnar networks. The precise frequency of the γ-oscillatory activity is one indicator that characterizes the dynamics of oscillatory networks. Identical γ-frequencies among the three hotspots’ CSD signals would be expected if they all reflect a shared population rhythm across the entire column, while separate local γ-oscillatory networks are less likely to have identical γ-frequencies. We computed the time-resolved CSD signal for all trials in a session. Subsequently, we calculated the dominant γ-frequency for each of the three attentional conditions and each of the three hotspots of each session of both monkeys, using trials with four stimuli shown (see methods).

The results (Figure 3A, left panel) show that the γ-frequencies at the three different hotspots differed significantly (Kruskal-Wallis test, Χ22 = 19.63, p = 5.47116*10⁻⁵, n_sup = 78, n_gra = 75, n_inf = 75). After multi-comparison correction, the observed differences between the γ-frequency at the supragranular and granular hotspots and between the granular and the infragranular hotspots were significant (post hoc Dunn’s test, α < 0.05). The median γ-peak frequency of 70.7 Hz at the granular hotspot was about 4 Hz higher than at the supragranular hotspot with 67.0 Hz and about 5 Hz higher than at the infragranular hotspot with 65.3 Hz (Wilcoxon rank sum tests, Z = 2.958, p = 0.0031 and Z = 4.1534, p = 3.2758*10⁻⁵, respectively). No significant difference was found between the γ-frequency at the supragranular and infragranular hotspot.

The dominant γ-frequencies at the hotspots of the three domains exhibit considerable variability between different recording sessions (Figure 3A). However, significant positive correlations were observed between the dominant γ-frequencies of supragranular and granular hotspots (r = 0.3264, p = 0.0051, Bonferroni-corrected p = 0.0154, n = 72) and between those of infragranular and granular hotspots, although the latter did not survive the multicomparison correction (r = 0.2712, p = 0.0212, Bonferroni-corrected p = 0.0636, n = 72). These results suggest that if γ-band activity in the granular domain oscillates at a comparably high frequency, the supragranular and infragranular domains in the same column also tend to oscillate at comparably high frequencies. No significant correlation was found between the dominant γ-frequencies of supragranular and infragranular hotspots, even before the multicomparison correction (r = 0.1487, p = 0.2125, n = 72). All comparisons were performed using Pearson’s correlation.

If neurons within V1 columns operate to some extent in functionally distinct ensembles, they might be differentially affected by the direction of selective attention. Therefore, we investigated whether γ-frequency and γ-power differed between the three different attentional conditions (“attend in,” “attend nearby,” and “attend away,” see Figure 1 for illustration). Again, we used the trial-based, time-resolved CSD signal. However, now we separately analyzed the three attentional conditions in each session of both animals and determined the dominant γ-frequency for each hotspot and attentional condition. We found an effect of attention for the supragranular hotspot (Friedman test, Χ22 = 11.62, p = 0.003, n = 26). Attending the stimulus in the RF results in a median γ-frequency of 68.2 Hz, which is significantly higher than the 66.3 Hz when attending the closely neighboring stimulus or 66.0 Hz when attending stimuli in the opposite hemifield (Wilcoxon signed-rank tests, Z = 2.4255, p = 0.0459 and Z = 2.9589, p = 0.0093, respectively; p-values are Tukey–Kramer corrected). In contrast, no significant effect of attention on γ-frequency was observed for the granular and infragranular hotspots. Thus, the results indicate that attention differently affects the frequency of the γ-oscillatory activity at the three hotspots. Its major effect is a small frequency increase of about 2 Hz at the supragranular hotspot when attention is directed to the stimulus within the RF.

In addition to the different effects of attention on γ-frequency at the three hotspots, their γ-power was also affected differently by attention. Again, we used the electrode centered on each hotspot. The relative increase of γ-power when attending the stimulus within the RF compared to attending stimuli in the opposite hemifield differed significantly between the three hotspots (Figure 3B, left panel; Kruskal-Wallis test, Χ22 = 6.1191, p = 0.0469, n = 23). In this condition, the median attention-dependent increase of γ-power was 38.1 % at the supragranular hotspot, 28.5 % at the granular hotspot, and 16.0 % at the infragranular hotspot. The difference between the attentional modulations at the supragranular and infragranular hotspots was statistically significant (post hoc Dunn’s test, α < 0.05 and Wilcoxon rank sum test, Z = 2.2848, p = 0.0223).

Similarly, directing attention to a stimulus near the RF as compared to the opposite hemifield also had differential effects on the γ-power of the three hotspots (Figure 3B right panel; Kruskal-Wallis test, Χ22 = 22.8595, p = 1.0867*10⁻⁵). In this condition, the relative changes as compared to the ‘attend away’ condition were generally smaller, with 12.8 % at the supragranular hotspot, 7.5 % at the granular hotspot, and − 4.1 % at the infragranular hotspot. Again, the difference between the attentional modulations at the supragranular and infragranular hotspots was statistically significant (post hoc Dunn’s test, α < 0.05 and Wilcoxon rank sum test, Z = 3.698, p = 2.6245*10⁻⁴).

Taken together, the three distinct hotspots of strong γ-oscillatory activity in cortical columns in V1 exhibit different characteristics not only in terms of their dominant γ-frequency but also in terms of the strength by which attention modulates their γ-frequency as well as their γ-power. This suggests that separate networks within these three major laminar domains engage in distinct γ-oscillations, which differ in frequency and the effect of attention on their spectral properties.

Suppose V1 columns comprise three distinct networks, each exhibiting its own γ-oscillatory dynamics. In that case, we expected that the activity of neurons that belong to the network would show stronger coupling to its γ-rhythm compared to the γ-rhythms of the other two networks. If, on the other hand, neurons across the entire column function as a unified ensemble with a single common γ-rhythm, then the activity of neurons throughout the column should exhibit similar coupling to this shared rhythm, regardless of where it is measured.

To test these predictions, we calculated the phase coherence (PhC) of multi-unit activity of each electrode with the three population rhythms. Subsequently, we investigated the columnar PhC patterns. As a measure of multi-unit activity, we used the entire spiking activity (ESA), which is a continuous signal and more sensitive to low amplitude spikes than measures requiring thresholding (Drebitz et al., 2019, 2020; Ahmadi et al., 2021a,b). The ESA signals vary considerably in strength across cortical depth but are similar in spectral composition (see Figure 4C for an exemplary ESA-power depth profile).

The spectral patterns of PhC across cortical depth reveal that the CSD signal of each hotspot was maximally coherent with ESA from the same laminar domain with a peak in the γ-band (Figure 4A, left three plots). Consequently, the set of ESA signals of a domain showed maximum γ-PhC with the CSD signal of the same domain and weaker γ-PhC with the CSD signals of the other two domains (Figure 4B). All but two of these differences were statistically significant (Table 1).

Since ESA and its power in the γ-band varied strongly with cortical depth, the depth profile of PhC is correspondingly biased. Therefore, we calculated the mean PhC between ESA and each of the three CSD signals at each depth and averaged values within the γ-band. We normalized these three PhC-values for each recording site by dividing them by the sum of the three values (Figure 4A, right panel).

The results reveal that ESA signals throughout the supragranular domain synchronize mostly with the γ-rhythm of the supragranular hotspot and in much smaller proportions with the γ-rhythm of the other two hotspots (Figure 4A, right panel, red curve). At the border to the granular domain, the ESA’s PhC with the supragranular hotspot steps down to a lower level, which becomes even lower for ESA in the infragranular domain. A similar pattern occurs for the γ-rhythm at the infragranular hotspot. ESA signals throughout the infragranular domain synchronize mostly in similarly high proportions with this γ-rhythm (brown curve) and in much smaller proportions with the γ-rhythms of the other two hotspots. Above the border between infragranular and granular layers, the PhC of ESA with the infragranular γ-rhythm quickly declines. The fairly similar proportions of PhC by which the spiking activity of neurons throughout both laminar domains follow predominantly the γ-rhythm measured at the hotspot of the same domain suggests that the neurons across each domain belong to a unified neuronal ensemble and follow a domain-specific, common γ-rhythm.

The depth profile of ESA’s relative PhC with the γ-rhythm of the granular hotspot has a maximum close to the upper border of the granular domain, from where it gradually declines to the lower border and stays smaller throughout the infragranular domain (Figure 4A, right panel, blue curve). Above the upper border, ESA’s relative PhC declines quickly. Although the granular hotspot’s PhC is driven predominantly by spiking within the granular layer, this proportion is very similar to the PhC proportion associated with the supragranular hotspot’s γ-rhythm. This pattern indicates that the γ-oscillatory network underlying hotspot of γ-power in the upper part of the granular domain might not extend homogeneously throughout the entire granular domain but is mainly located in its upper part.

In summary, the pattern of the PhC between the spiking activity across the column and the three hotspots of γ-power indicates three separate networks in the major laminar domains that engage in γ-oscillatory activity. Typically, the spiking activity of such a network’s neurons locks predominantly to their own network’s rhythm and less to the rhythms of the other networks.

After demonstrating that the γ-CSD signals of a hotspot mainly represent the same domain’s γ-oscillatory activity and that the power of these local γ-oscillatory activities can be modulated differently by attention, we examined whether attention also affects the strength of interlaminar γ-synchronization. Figure 5A shows the depth profiles of PhC between the CSD signal of each domain’s hotspot and the CSD signals of all other electrodes for the two conditions where attention was either directed to the stimulus within the recorded neurons’ RF (‘attend in’ left column) or to the stimulus close to, but outside, the RFs (‘attend nearby’ right column). Similar to comparisons between the CSD signals of the hotspot of each domain and ESA signals across depth (Figure 4), we observed strong γ-PhC between the granular and supragranular domains (Figure 5A, upper and middle rows) and much weaker PhC between the infragranular hotspot’s γ-activity and that of the other two domains (lower row). Furthermore, we found that PhC in the γ-band increased with attention between all three domains (Figure 5B).

The comparison of γ-PhC between the CSD signal of the hotspot of one domain and the electrode in a different domain showing the highest γ-PhC with the hotspot across all recording sites of both animals revealed significant differences between all compared pairs (Figure 5C). The median γ-PhC between the supragranular and infragranular pair increased with attention by 0.014 (Figure 5C upper plot; γ-PhC_{attend in} = 0.121, γ-PhC_{attend nearby} = 0.107, n = 32, p = 4.2204*10⁻⁰⁵, z = 4.0951), by 0.025 between the granular and supragranular pair (middle plot; γ-PhC_{attend in} = 0.283, γ-PhC_{attend nearby} = 0.258, n = 31, p = 0.0019598, z = 3.0963), and by 0.02 between the infragranular and granular pair (lower plot; γ-PhC_{attend in} = 0.09, γ-PhC_{attend nearby} = 0.07, n = 31, p = 6.9469*10⁻⁰⁵, z = 3.9781; all comparisons Wilcoxon signed-rank tests). This additive effect of attention on γ-PhC between laminar domains did not differ significantly between the different pairs (Kruskal-Wallis test, Χ22 = 0.5, p = 0.7803, n_sup/infra = 32, n_gra/supra = 31, n_inf/gran = 31).