(C) PLOS One [1]. This unaltered content originally appeared in journals.plosone.org.
Licensed under Creative Commons Attribution (CC BY) license.
url:https://journals.plos.org/plosone/s/licenses-and-copyright

------------



Multiple gamma rhythms carry distinct spatial frequency information in primary visual cortex

['Chuanliang Han', 'State Key Laboratory Of Cognitive Neuroscience', 'Learning', 'Idg Mcgovern Institute For Brain Research', 'Beijing Normal University', 'Beijing', 'Tian Wang', 'Yi Yang', 'Yujie Wu', 'Yang Li']

Date: 2022-01 

Gamma rhythms in many brain regions, including the primary visual cortex (V1), are thought to play a role in information processing. Here, we report a surprising finding of 3 narrowband gamma rhythms in V1 that processed distinct spatial frequency (SF) signals and had different neural origins. The low gamma (LG; 25 to 40 Hz) rhythm was generated at the V1 superficial layer and preferred a higher SF compared with spike activity, whereas both the medium gamma (MG; 40 to 65 Hz), generated at the cortical level, and the high gamma HG; (65 to 85 Hz), originated precortically, preferred lower SF information. Furthermore, compared with the rates of spike activity, the powers of the 3 gammas had better performance in discriminating the edge and surface of simple objects. These findings suggest that gamma rhythms reflect the neural dynamics of neural circuitries that process different SF information in the visual system, which may be crucial for multiplexing SF information and synchronizing different features of an object.

Funding: This work was supported by the National Natural Science Foundation of China Grant 32171033 (DX), National Natural Science Foundation of China Grant 32100831 (TW), The Fundamental Research Funds for the Central Universities (DX), The 111 Project Grant (BP0719032) (DX), The Open Research Fund of the State Key Laboratory of Cognitive Neuroscience and Learning. (DX). The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.

In the visual cortex, visual stimuli with a high degree of spatial frequency (SF) information such as bars [ 18 – 21 ] or contours [ 14 – 16 ] have been found to induce synchronization in the gamma band. However, more recent studies have also demonstrated that stimuli with low SF information, such as large uniform surfaces, also induce strong gamma synchronous responses [ 22 – 24 ]. The finding that both high and low SFs can generate a gamma rhythm is puzzling because high SF information, such as edges [ 25 , 26 ], is thought to be processed by different neural mechanisms than low SF components, such as surfaces [ 27 – 31 ]. There are 2 possible explanations for this contradiction. The first is that a single gamma rhythm can integrate high and low SF information. The second is that distinct gamma rhythms carry different aspects of SF information. The second hypothesis is plausible because distinct narrowband gamma oscillations have been found in cats, primates, and humans [ 9 , 20 , 32 – 35 ], among other species. However, the relationship between SF selectivity and specific gamma frequency bands remains poorly understood [ 36 , 37 ].

The gamma rhythm (30 to 100 Hz) is an oscillatory pattern of neural activity. It is commonly found in many brain regions [ 1 – 6 ] and is thought to play an important role in cognitive functions such as learning [ 7 ], memory [ 8 , 9 ], and attention [ 10 – 13 ]. Abnormal gamma band activities have been related to mental illness [ 14 – 17 ].

Materials and methods

Preparation of anesthetized cats All procedures were in accordance with the National Institutes of Health Guidelines, and the research protocol was approved by the Biological Research Ethics Committee of Beijing Normal University (ID: IACUC(BNU)-NKLCNL 2017–06) and Fudan University (ID: 2021JS0087). Twelve hours before the experiment, cats were injected with dexamethasone (0.4 mg/kg, subcutaneously). The cats (weighing 2 to 4.5 kg each) were initially anesthetized with isoflurane (5% concentration) and injected with atropine sulfate (0.05 mg/kg, subcutaneously). After we conducted a tracheotomy, the animals were artificially ventilated. During recordings, anesthesia and paralysis were maintained with propofol (2 to 6 mg/kg/h) and vecuronium bromide (0.1 mg/kg/h), respectively. The end tidal CO2 was maintained at 3.5% to 4%, and the body temperature was set at 36.5°C to 37.5°C. The eyes of the cats were treated with 1% atropine sulfate solution to dilate the pupils, and they were fitted with appropriate contact lenses with artificial pupils (1-mm diameter) to restrict the amount of luminance and ensure focus on a tangent screen. Ventilation pressure, heart rate, electrocardiogram, and blood oxygen were monitored continuously throughout the experiment. To maintain the health of the cats prior to the experiment, they were injected with dexamethasone antibiotic every day (0.4 mg/kg) and with atropine sulfate (0.05 mg/kg) every other day. One craniotomy (10 to 15 mm in diameter) was performed over the primary visual cortex (VI; centered at P0–2 and L0, mainly in Brodmann’s area 18 (A18)), and another (approximately 6 mm in diameter) was performed over the lateral geniculate nucleus (LGN) (centered at L8, A9).

Electrophysiological recordings Multielectrode arrays (Utah Array, Blackrock Microsystems, USA) were used to record signals in A18 from the horizontal direction (parallel to the cortical surface) in 12 cats. The arrays consisted of 6*8 or 10*10 grids of microelectrodes (1-mm electrode length and 400-micron electrode separation). They were inserted 0.5 to 0.6 mm into the cortex using a pneumatic insertion device. Multielectrode linear arrays (Blackrock Microsystems) were used to record signals in area 18 from the vertical direction (perpendicular to the cortical surface) in another 3 cats. The linear arrays consisted of 24 microelectrodes (spaced 0.1 mm apart, each 0.015 mm in diameter, Plexon, USA). They were inserted 1 to 1.2 mm into the cortex using an automatic controlled insertion device. The multielectrode linear arrays (Plexon) were also used to record signals in the LGN of 6 cats (in 4 of these, recordings were simultaneously conducted in the V1 using Utah arrays). They were inserted 13 to 14 mm into the cortex using an automatic controlled insertion device. To prevent pulsation, the arrays and the exposed cortex were covered with 1.5% to 2% agar. Signals were amplified using a Cerebus 128-channel system (Blackrock Microsystems). All local field potential (LFP) signals were sampled at 500 Hz per channel with a wide-band front-end filter (0.3 to 500 Hz). LFP data were postprocessed by removing channels that were deemed nonfunctional because of broken electrodes or noise.

Visual stimulation Visual stimuli were generated using a computer with a Leadtek GeForce 6800 video card and displayed on a cathode-ray tube (CRT) monitor (Sony CPD-G520, mean luminance of 32 cd/m2, 1024*768 resolution, refresh rate 100 Hz) placed 40 cm away from the animal’s eyes. Luminance nonlinearities were corrected using the luminance meter (photoresearch PR670). Stimuli were presented only to the eye contralateral to the craniotomy. For each experiment, we first mapped the receptive fields (RFs) of the multiunit activity (MUA) responses using short instances of sparse visual noise and then conducted an experiment with drifting gratings and large squares with uniform black/white surface luminance.

RF mapping After manually mapping the RFs of the recorded channels, we used sparse visual noise to precisely locate the RF center. Specifically, a white or black square (0.5 to 2 degrees of visual angle) was flashed on a gray background at different positions in a pseudorandom sequence (usually on a 13 × 13 or 11 × 11 sample grid). The stimuli were presented with an effective frame rate of 100 Hz so that each sparse noise image appeared for 40 ms. Each square was presented 30 to 60 times. The sequence was cut into small segments based on the trial length. This process enabled us to generate a two-dimensional map for each channel. The averaged responses from the x-axis and y-axis of each map were fitted with a one-dimensional Gaussian function to estimate the center position and radius of the RF (σ of the Gaussian function).

Gratings and uniform stimuli We first presented drifting gratings with different SFs (approximately 0.01 to 2.5 cycles per degree, 10 conditions) and directions (−18° to 180°, 18 conditions). The stimulus contrast was 90%, and the temporal frequency (TF) was 4 cycles per second. Afterward, all drifting gratings were enlarged (approximately 38°) to cover the RFs of all recording sites, and the stimuli were shown only to the contralateral eye. The temporal sequence of the visual experiment was prestimulus (0.4 seconds), on stimulus (2 seconds), and off stimulus (0.4 seconds). In 2 cats, a large uniform square stimulus (20° for one cat and 25° for the other) (black or white against a gray background) was then presented at different positions, so that the center of the RF for that site fell into the border (edge) or the center (surface) of the stimulus.

Laminar alignment To align the different probe placements in terms of depth, we used the laminar pattern of the MUA responses to the visual stimuli with different SFs. Because the thickness of the cortex and verticality of the probe differed between the probe placements, we used the relative depth (ReD) [38,39] to represent the cortical position of each recording site. The ReD is the normalized cortical depth, ranging from 0 to 1. Three variables were used to calculate the ReD. First, the recording site with the lowest SF selectivity was considered to be in the middle of the input layer (Cha1) and was assigned a ReD value of 0.5. Second, the cortical surface was considered to be located 0.050 mm above the uppermost channel (Cha2) with significant visually driven spiking responses (signal-to-noise ratio [SNR] > 3 and the 3 consecutive channels below Cha2 were also expected to meet the criterion of SNR > 3) and was assigned a ReD value of 0. Third, the boundary separating the gray and white matter was considered to be located 0.050 mm below the deepest channel (Cha3) with significant visually driven spiking responses (SNR > 3 and the 3 consecutive channels above Cha3 were also expected to meet the criterion of SNR > 3) and was assigned a ReD value of 1. After the 3 signatures (Cha1, Cha2, and Cha3) had been determined, they were used as references when calculating the ReD for each channel. ReD values of 0.375 to 0.625 were assigned as the input layer of A18 in the cat V1 according to the procedure used in previous studies [40,41].

Power spectrum analysis For each trial, the power spectrum of the LFP response over the period 300 to 2,000 ms after stimulus onset (yielding the stimulus power spectrum) was estimated using the multitaper method [42,43] (time-bandwidth product, 3; tapers, 5; Chronux toolbox (http://chronux.org/)), which was implemented using custom software written in MATLAB. Essentially, the multitaper method attempts to reduce the variance of spectral estimates by premultiplying the data with several orthogonal tapers known as Slepian functions. The frequency decomposition of multitapered data segments therefore provides a set of independent spectral estimates that, once averaged, provides an ensemble estimate that is more reliable for noisy data.

Coherence analysis We evaluated spike-field coherence (SFC) by calculating the coherency Cxy between different sites (x and y) as the cross-spectra between signals in x and y (Sxy), normalized by the geometric mean of their autospectra (Sxx and Syy) [8]. This was estimated using the multitaper method [44] (time-bandwidth product, 3; tapers, 5; Chronux toolbox (http://chronux.org/)) and implemented using custom software written in MATLAB. To rule out the possibility that the effect of the SFC was solely the effect of the gamma power, we conducted a control analysis in which we calculated the SFC between spike trains, shuffled relative to the LFP signals for each trial within each stimulus condition.

Model fitting and evaluation To quantify the responses of the different gamma components (GCs) in terms of different SFs, we used a two-dimensional descriptive model to fit the power spectrum. One dimension was the oscillatory frequency, and the other dimension was the SF. (1) (2) (3) (4) (5) We used a two-dimensional descriptive model (1 to 5) to fit the power spectrums of the LFP and MUA for different SFs. The LFP power in the gamma range could be modeled by a baseline and the weighted sum of the 3 frequency components (1). The frequency profile of the baseline was modeled as a 1/f function that decreased monotonically according to frequency (2), and the TF profiles of the 3 components were all modeled as Gaussian functions (3). The peak frequency μ of each GC was modeled as sigmoid function (4). The SF profile for the power W of the GC was modeled as the difference between the Gaussian function and the sigmoid function for the peak frequency (5). The power spectrums in the edge or surface condition were fitted using formula (1), and the SF tuning curve for the MUA and baseline power in the 3 gamma frequency bands were fitted by formula (5). The goodness of fit was defined as formula (6). Only the sites with a high goodness of fit (larger than 0.8) were analyzed: (6) We fitted power spectrums with the descriptive model (described by formula 1 with details in formulas 2 to 5) by minimizing the mean square error (MSE) between the model prediction and the data (power spectrum) using the fmincon function in MATLAB. The parameters that depend on SF include μ i and W i , and the parameters independent of SF include a, b, c, k, σ i , SF 0 , n i , K 0,i , K i , A 1,i , A 2,i , σ 1,i , and σ 2,i (formulas 4 to 5). Instead of determining the number of GCs before model fitting, we fixed the number of GCs at 3 for model fitting to spectrums from all recorded sites. After model fitting, we counted the number of GCs required to explain the power spectrums from each site based on the SNR of each fitted GC. The SNR of each GC was defined as the max weight of the GC (formula 5) divided by the mean power in the corresponding gamma band under the blank condition. If the SNR of a GC was larger than 3, we counted this as a required GC, and otherwise it was considered a negligible GC. In this way, we were able to determine whether a power spectrum had 0, 1, 2, or 3 GCs.

Decoding the edge and surface of a square We built a logistic regression model to decode the location within a stimulus (edge or surface) that drove a single site (or group of sites) in a given trial, either by gamma power or mean firing rate. The population response matrix was defined as X, X = (X 1 , …, X 3N )T, where X i was a vector (m by 1, m was the total number of trials) of the neural data (either gamma power or MUA response), and the gamma power was the power of the jth recording site in the low gamma (LG) (i = 3j), medium gamma (MG) (i = 3j + 1), and high gamma (HG) (i = 3j + 2) rhythms, respectively. N was the total number of sites included for decoding. As a control, we built a logistic decoder based on the MUA responses using the same procedure. To keep the size of the response matrix (X) constant, we divided the MUA responses from the jth recording site into 3 time bins (if i = 3j, mean rate 300 to 900 ms after the stimulus onset; if i = 3j + 1, mean rate 900 to 1,500 ms after the stimulus onset; and if i = 3j + 2, mean rate 1,500 to 2,000 ms after the stimulus onset). We aggregated all data together for the decoding operation. The units in the X were either all edge response or all surface response. The logistic regression model used to fit the population response to the square location was as follows: (7) Here, W is a vector (3N by 1) of weights. When p ≥ 0.5, the population is predicted to be driven by the stimulus edge; when p < 0.5, the population is predicted to be driven by the stimulus surface. We trained the decoder by optimizing the weight matrix (W) to minimize the loss function (formula 8). (8) Here, Y is an M by 1 vector representing the location of the stimulus, where 1 is the edge and 0 is the surface. We used an L1 regularization term to minimize the obtained weights and prevent overfitting. We random separate the dataset into a training set (60%), validation set (15%), and test set (25%). For a regularization parameter (λ), we chose the option that gave the lowest MSE in the cross-validation test among 20 lambda values that were equally spaced (logarithmically) between 0.001 and 40. We used the fminunc function in MATLAB to train the decoder. The results shown here (Fig 3C, S7 Fig) were obtained with the regularization parameter set at 0.1 for all sites.

Site selection The inclusion criterion for a site was as follows. The firing rate or gamma power of the site at its preferred SF was significantly higher than that of the baseline (i.e., SNR ratio larger than 3). For the MUA responses, the SNR was defined as the standard deviation of the firing rate for visual stimuli divided by that in the blank condition. For LFPs, the SNR was defined as the standard deviation of the gamma power for visual stimuli divided by the standard deviation of the power in the corresponding gamma frequency band in the blank condition.

[END]

[1] Url: https://journals.plos.org/plosbiology/article?id=10.1371/journal.pbio.3001466

(C) Plos One. "Accelerating the publication of peer-reviewed science."
Licensed under Creative Commons Attribution (CC BY 4.0)
URL: https://creativecommons.org/licenses/by/4.0/


via Magical.Fish Gopher News Feeds:
gopher://magical.fish/1/feeds/news/plosone/