(C) PLOS One
This story was originally published by PLOS One and is unaltered.
 . . . . . . . . . .



Distinct sources of decision-related signals in visual cortex are represented in different local field potential bands [1]

['Yueyue Sapphire Hou', 'Department Of Neurology', 'Neurosurgery', 'Montreal Neurological Institute', 'Mcgill University', 'Montreal', 'Quebec', 'Pooya Laamerad', 'Department Of Neuroscience', 'University Of Pennsylvania']

Date: 2026-07 

The experimental methods (i.e., behavioral training and adaptation) were detailed in previous papers [8,24,66]; this paper primarily involves a re-analysis of the associated data. Only the relevant information is reiterated here.

All procedures adhered to the regulations established by the Canadian Council on Animal Care and were approved by the Montreal Neurological Institute’s Animal Care Committee (protocol no. 5031). Daily operations were supervised by the institute’s veterinarians and trained animal health staff.

Four adult female rhesus macaque monkeys (Macaca mulatta) (Monkey Y, age: 10 years; Monkey C, age: 8 years; Monkey L, age: 10 years; Monkey A, age: 8 years; all weighted 5–7 kg) participated in this study. The animals were housed in a temperature- and pressure-controlled socially accessible environment under a 12-hour light/dark cycle. Behavioral training and experimental recordings were conducted during the light phase of the cycle. Animals sat comfortably while head-fixed in a custom-designed primate chair. Identical stimuli, timing, and rewards were used for both monkeys of the same task. A solenoid-operated reward system was used to dispense juice reward to the monkeys (Crist Instruments Co.).

Prior to the experiments, an MRI-compatible titanium head post (Hybex Innovations, Montreal) and a plastic recording cylinder were affixed to each monkey’s skull under general anesthesia. Areas MT and V4 were identified based on transitions from white matter to gray matter, as well as the electrode depth, the prevalence of robust, direction-selective visual responses, and the relationship between RF size and eccentricity. More details could be found in the respective published work [ 24 , 67 ]. Eye movements were monitored with an infrared eye tracking system (EyeLink1000, SR Research) with a sampling rate of 1,000 Hz.

For area V4, neuronal signals were recorded using an Intan Technologies system and filtered between 0.5 and 7 kHz. Real-time spike signals were thresholded by exceeding ±3 standard deviations, estimated for each channel. Segments surrounding each threshold crossing were extracted and clustered using UltraMegaSort 2000, a k-means-based clustering algorithm [ 68 ].

For area MT, neuronal signals were continuously monitored during acquisition via computer display. Spike signals were thresholded in real-time, with local field potential (LFP) signals undergoing band-pass filtering at 0.5–150 Hz, while spike signals were band-pass filtered at 150–8 kHz, with monitoring conducted on an oscilloscope and loudspeaker. Spike signals were thresholded in real-time, and spikes were assigned to single units by a template-matching algorithm (Plexon MAP System). Then, spikes were manually sorted offline using a combination of automated template-matching, visual inspection of waveforms, clustering in the space defined by the principal components, and absolute refractory period (1 ms) violations (Plexon Offline Sorter).

Single units were recorded utilizing linear microelectrode arrays (V-Probe, Plexon) comprising 16 contacts. Initially, the electrode array was lowered to the designated depth, followed by the estimation of multi-channel RFs, through the manual positioning of a moving bar within the visual field in area MT, and through the manual presentation of sparse noise stimuli at various spatial locations within the visual field in area V4. The RF mapping was then done on each animal by fitting with a 2D Gaussian function to recover the RF centers and sizes [ 24 , 67 ].

The individual trial structure of a match-to-sample form discrimination task is detailed below (also see Fig 1B ). Monkeys were trained to identify which of the two choice stimuli (a circular and a radial grating) matched a previously presented sample stimulus, which can be either a circular or radial grating. On each trial, an initial 500 ms fixation was followed by the presentation of a sample stimulus for 200 ms. This sample stimulus had Gaussian noise added at one of eight levels to encompass the animal’s psychophysical threshold range in terms of signal-to-noise (SNR) ratio (0%, 3%, 6%, 8%, 12%, 25%, 60%, and 100%). 0% SNR indicates pure noise, whereas 100% SNR is a noiseless stimulus. After the sample stimulus was removed, a randomized 250–500 ms delay period followed. Then, both noiseless (100% SNR) circular and radial grating stimuli appeared on either side of the fixation point as the response cues. They were positioned ±7° from the fixation point along the horizontal meridian. Their locations were randomly shuffled from trial to trial to prevent the animals from developing a fixed sensorimotor mapping. The monkeys should make a saccade towards either of the two cues that matched the sample stimulus and maintain fixation on it for 800 ms to exchange for a liquid reward. If the choice made was incorrect, a 1.50-s timeout followed the choosing period before the onset of the next trial. Note that few trials on 0% sample stimulus collected from animals on this task as they refused to do trials only containing pure noise. Thus, we combined 0% stimulus trials with other stimulus levels if needed for analyses.

Visual stimuli were back-projected onto a semi-transparent screen using an LED video projector (VPixx Technologies, PROPixx) with a refresh rate of 120 Hz. The screen spanned an area of 80° by 50° at a distance of 81 cm. A neutral gray color (54 cd/m 2 ) was used as the background for the task. The monkeys were trained to perform a form discrimination task, with circular and radial gratings. During each session, neuronal size and position preferences were measured using 100% contrast of circular and radial grating positioned within the RFs found at the recording site on that day.

The individual trial structure of a random dot motion discrimination task is detailed below (also see Fig 1A ). On each trial, animals established and maintained fixation for 300 ms, followed by a brief presentation of the dot coherence patch (typically for 67 ms) on the RF centers. The monkeys were then required to maintain fixation for another 300 ms, after which the fixation point disappeared, two choice targets appeared, and the monkey made a saccade to the corresponding target to indicate its perceived motion direction (preferred or anti-preferred relative to the isolated neurons) in exchange for juice rewards. Targets were typically positioned at 10° eccentricity, with each pair of angles sampled at 45° intervals on the screen. For example, if a neuron’s preferred direction was 45°, the targets were intentionally positioned to correspond with both the preferred and anti-preferred directions, meaning they would be placed at 45° and 225°, respectively. The distance between the two saccade targets was typically 20°. The appropriate saccade direction correlated with the most similar saccade target direction (i.e., the monkey performed the rightward saccade to correctly report rightward motion). The monkey was required to indicate its decision within 700 ms after the onset of the choice targets. In trials with no motion signal (0% coherence), rewards were randomly assigned in half of the instances. If fixation was disrupted at any time during the stimulus, the trial was aborted. In a typical experimental session, the monkeys performed the task on 20–40 repetitions of each stimulus.

Visual motion stimuli were presented at a frequency of 60 Hz with a resolution of 1,280 by 800 pixels. The viewing area covered 60° by 40° at a distance of 50 cm. The monkeys were trained to perform a motion discrimination task, with RDK. During each session, neuronal direction and speed preferences were measured using 100% coherent dot patches positioned within the RFs found at the recording site on that day. Specifically, the RF locations were quantified by fitting a 2D spatial Gaussian to the neuronal response measured across a 5 x 5 grid of stimulus positions in an offline analysis. The grid comprised of moving dot patches centered on the initially hand-mapped RF locations. We confirmed that all neurons included in our analysis had RF centers within the stimulus patch used for the behavioral experiments. The speed of the dots was chosen to match the preference of the MT neurons in proximity to the recording site. The direction of motion was consistently selected based on the preferred or anti-preferred direction of the neurons being examined. The stimulus size also matched the RF size of the MT neurons (mean radius = 6.3° ± 1.2°). The RDK coherence was randomly selected on each trial from nine values that encompassed the range of the monkey’s psychophysical threshold (0%, 2%, 4%, 8%, 12%, 16%, 32%, 64%, and 100%).

As described previously [ 24 , 67 ], control saline injections did not alter spiking activity ( S8 Fig ) on any channel and had no effect on behavioral performance; there was no behavioral effect of injecting saline or of injecting muscimol at sites distant from the recording site, indicating that the muscimol results were due to local inactivation, rather than to any general effects of the injection.

Behavioral testing commenced 40–50 min post-infusion at each site. Muscimol infusion resulted in complete suppression of detectable spiking activity across all electrode contacts. The probes remained at the infusion depth within the cortex for the duration of the session, allowing us to verify that spiking activity remained silenced throughout the post-muscimol period, which typically lasted 2.5–3.5 hours. The inactivated area returned to its normal state typically within two days. Experiments involving muscimol were performed at a frequency not exceeding once per week.

Muscimol, a GABA A agonist (Sigma), was dissolved in sterilized saline (pH ~= 7) to achieve a concentration of about 10 mg/mL [ 26 ]. In half of the sessions, muscimol (generally 2 μl at a rate of 0.05 μl/min) was administered via the fluid channel in the V-Probe to inhibit its adjacent neural activity. The spread of muscimol is typically less than 2 mm, so it is unlikely that the drug spread into nearby retinotopic regions of cortex [ 69 ].

The linear V-Probe contained a glass capillary with an inner diameter of 40 μm. The capillary was positioned at one end between contacts 5 and 6 of the array, with contact 16 located at the most dorsal-posterior position. The outermost end was connected via polyethylene tubing (PE20, inner diameter = 0.38 mm) to a 10 μl micro-syringe (Hamilton) to ensure continuous and smooth injection.

Analyses of neural and behavioral responses

Only recording sessions that contained pre- and post-muscimol data within the same recording session were included for analysis. We excluded trials in which monkeys failed to execute a saccade toward one of the choice targets or broke fixation during the stimulus presentation or delay periods. Overall, a total of 14 control sessions and 14 corresponding inactivation sessions were included for analyses (10 sessions for monkey Y, 6 sessions for monkey C, 6 sessions for monkey L, and 6 sessions for monkey A). On average, monkey Y completed 318 trials per session, and monkey C performed 358 trials per session, monkey L completed 1,626 trials per session, and monkey A performed 1,259 trials per session, encompassing both correct and error trials of the task. The comparable number of trials was performed after the infusion of muscimol (p = 0.992, Welch’s t test).

Psychometric curve fitting. The monkeys’ behavioral performance as a function of dot coherence or form SNR level was characterized by fitting a Weibull model to the proportion of correct responses by using nonlinear regression (fitnlm in MATLAB). The two-parameter Weibull function is, where is the predicted proportion of correct responses at stimulus level (motion coherence % for area MT and SNR % for area V4); α is the perceptual threshold, and β determines the slope of the function; and is stimulus coherence or contrast in percentage. Under a conventional parameterization, α corresponds to yielding ~81.6% correct performance. Because the fit was obtained in log-percentage space, trials with 0% stimulus strength were excluded to avoid singularities under the log transform. We plotted the resulting psychometric curves with 95% confidence intervals at each stimulus level using model-based prediction intervals. Pooled psychometric fits are shown in Fig 1C-1F, whereas Fig 1G shows statistical analyses conducted on session-wise fits. In Fig 1G, behavioral sensitivity was quantified as the inverse of the perceptual threshold at which behavioral performance reached 81.6% for each monkey in each recording session, such that larger values indicate better perceptual sensitivity.

LFP preprocessing and re-referencing. LFPs from the 16th channel, located at the most dorsal-posterior position, were utilized as the reference signal, while the LFP signals from the other channels were re-referenced accordingly so that artifacts from eye blinks and other physiological factors were eliminated. The line noise at 60 Hz was removed in LFP raw signals. We utilized the Chronux toolbox [70] in MATLAB, applying its multi-taper method for Fourier power spectral estimation, using two tapers derived from discrete prolate Slepian sequences to achieve a frequency resolution of 1 Hz. To compare LFP power before and after muscimol inactivation, we used a nonparametric statistical method. First, we calculated the log ratio of LFP power spectra (pre- versus post-inactivation) across trials from all sessions. The difference in total power before and after inactivation was quantified by computing the area under the power spectral density (PSD) curves (MATLAB’s trapz function). To assess whether this power difference varied significantly as a function of frequency, we computed Spearman’s rank correlation coefficient (ρ), testing statistical significance through a permutation test (n = 5,000 permutations).

Spectral parameterization of evoked LFPs. To evaluate the spectral structure of the stimulus-evoked LFP prior to frequency band–based CP analyses, we computed PSD during the stimulus-response epoch (50–250 ms post-stimulus) at the trial level. Each trial’s spectrum was parameterized using the FOOOF algorithm [71] over a range of 5–150 Hz, to separate broadband aperiodic structure from oscillatory peaks. To ensure robust fitting across a wide range of spectral shapes, we evaluated multiple FOOOF model configurations that varied in peak width limits (1–12 Hz), maximum number of peaks (6–8), minimum peak height (0.03–0.05), and aperiodic mode (fixed or knee). For each trial, the best-fitting model was selected based on a composite score that favored higher explained variance (R²) and lower fitting error, with a small penalty applied to excessive numbers of detected peaks. For each accepted fit, we extracted aperiodic components and oscillatory peak parameters. Peak center frequencies were pooled across trials, channels, and sessions to generate descriptive histograms of oscillatory peak distributions before and after inactivation in S7 Fig.

LFP-based choice probabilities (CP). The computation of LFP-based CP, a metric that quantifies trial-to-trial fluctuations between LFP power variability and psychophysical choices, was conducted, mostly using established methods developed for spiking activity [4]. We first subdivided trials into two choices defined by neurons’ stimulus selectivity and behavioral outcomes. We included sites for which the preferred stimulus had at least a 50% higher multi-unit spike count in the stimulus-response period (50–250 ms relative to stimulus onset) than the spikes in the baseline (−200–0 ms relative to stimulus onset). The stimulus-response interval was selected due to the significant correlation between the spikes observed during this time window and the animals’ behavioral choices [66]. Stimulus selectivity was assessed across multiple neurons at the highest dot coherence levels (64% and 100% for the MT task; 60% and 100% for the V4 task) for every electrode channel independently. In this way, a preferred choice contained trials that exhibited either a preferred stimulus with a correct response or a nonpreferred stimulus with an incorrect response; an anti-preferred choice contained trials that involved either an anti-preferred stimulus with a correct response or a preferred stimulus with an incorrect response. For the reward history analysis, current trials were additionally partitioned according to behavioral outcome of the immediately preceding trial. Specifically, for each valid trial , we identified trial in the original behavioral dataset and used its outcome to classify the current trial as rewarded if the previous trial had been correct, or as nonrewarded if the previous trial had been incorrect. Importantly, the LFP power used for CP estimation was always taken from the current trial, not from the preceding trial. Within each reward history condition, the current trials were then subdivided into preferred choice and anti-preferred choice using the same behavior-choice rules described above. Because CP is meant to be insensitive to the mean amplitude of neural responses, most analyses of CP first involve normalization of firing rate, using z-scoring or balanced z-scoring [72]. However, given the nonnormal distribution of LFP power, this approach has the potential to introduce biased estimate of CP. To assess this possibility, we used the NeuroDSP toolbox [73] to simulate LFP data from 100 trials of one-second duration (S4A Fig), each comprised of stochastic periodic and aperiodic components; these were generated using the function sim_combined with the default parameters. We implemented this procedure twice to simulate trials associated with two behavioral choices that differed in mean amplitude across frequencies, one being 100 times greater in amplitude than the other (S4A Fig). Then, we computed spectrograms to represent the simulated LFP data before normalization (S4B Fig). A proper normalization would yield CP values of 0.5, because CP computation takes into account the variation between choices, not the difference in mean. However, we found that z-scoring and balanced z-scoring failed to eliminate the mean difference between the signals associated with the two choices, largely distorting the signals that were associated with lower-frequency LFP power (S4C Fig). This distortion arises because nonGaussian distributions like LFP power are susceptible to the skewness of the PSD. Thus, z-scoring disproportionately emphasizes outliers when the distribution is nonnormal, biasing LFP-based CP estimation. We therefore used a robust scaling method that avoids such distortions. Specifically, we first resampled trials with each of the two choices to ensure that the preferred and anti-preferred choices had equal number of trials for each stimulus condition. We then normalized separately for the preferred and anti-preferred choices. For this normalization, we used the median and interquartile range (IQR) of the distribution of LFP power for each choice: where X is the original distribution and X normalized is the normalized version of the original choice by its median and IQR. This approach does not rely on any specific distributional assumptions but rather quantifies the central tendency and dispersion to better resist outliers in LFP. In our simulated example, robust scaling aligned the two choices to a uniform scale with identical values, thereby recovering the accurate CP values across all frequencies (S4D Fig). This simple simulation confirmed that robust scaling effectively returned CPs to chance when the LFP mean power is the only difference between two simulated LFP profiles, demonstrating its suitability for normalizing LFP power. Following appropriate normalization, the distributions of two normalized choices were subsequently analyzed using receiver operating characteristic (ROC) analysis to compute CP. CP was determined by calculating the area under the ROC curve. The area under the ROC curve ranges from 0.0 to 1.0, reflecting the performance of an ideal observer in determining motion direction based on the trial-to-trial LFP power. Values of 1.0 and 0.0 represent perfect classification, whereas a value of 0.5 signifies performance equivalent to random chance classification. CP can be sensitive to overall behavioral performance, and the infusion of muscimol significantly impaired the monkeys’ performance. It was therefore necessary to equalize performance pre- and post-muscimol by selecting trials with different stimulus levels. In the MT task, for Monkey Y, we used coherence levels of 0%, 4%, and 8% before muscimol injection and 0%, 8%, 12%, and 16% after injection; for Monkey C, we used coherence levels of 0%, 4%, 8%, and 12% before muscimol injection and 0%, 8%, 12%, and 16% after injection. In the V4 task, for Monkey L, we used stimulus SNR levels of 0%, 6%, 8%, and 12% before muscimol injection and 0%, 8%, 12%, and 25% after injection; for Monkey A, we used stimulus SNR levels of 0%, 6%, 8%, and 12% before muscimol injection and 0%, 12%, 25%, and 60% after injection. These stimulus levels equated behavioral performance before and after inactivation, at ~65% correct across monkeys for analyses of area MT (Monkey Y, 66%; Monkey C, 62%); at approximately 78% correct across monkeys for analyses of area V4 (Monkey L, 80%; Monkey A, 75%). However, because differences in stimulus strength could bias CP, we included trials of 0% level in both pre- and post-inactivation conditions. These ambiguous trials, on which monkeys performed at chance, controlled for any confounds related to the stimulus. We then estimated CPs based on LFP power. First, we calculated the CP for each frequency bin individually, utilizing power measurements obtained from a sliding window of 200 ms width, advancing in 20 ms increments. We calculated the averaged CPs across three distinct epochs relative to stimulus onset: baseline (−200–0 ms), which occurred during the pre-stimulus fixation, stimulus-response (50–250 ms), which spanned between stimulus onset and post-stimulus fixation, and delay (250–400 ms) spanning before the appearance of the choice targets. The stimulus-response epoch was defined by the time that LFP power response was maximal (Fig 2C), and to analyze CPs across the two areas, we kept the epoch length consistent. We statistically quantified CPs by averaging power across three frequency bands: alpha-beta (5–30 Hz), low-gamma (30–70 Hz), and high-gamma (70–150 Hz). We noted that spike waveforms have statistically little effect on any broadband LFP power [57].

CP latency. We estimated CP latency using a threshold-free change-point analysis [74,75]. Specifically, we analyzed the CP trace from −200–250 ms relative to stimulus onset using a two-segment model. In this model, the CP trace was approximated by one mean level before a candidate transition time and a second mean level after that time. The transition time was constrained to occur within the post-stimulus interval (50–250 ms relative to stimulus onset). Before fitting, each trace was aligned to its dominant post-stimulus direction of deviation from 0.5 so that the main CP excursion was positive. We defined CP latency as the transition time that provided the best fit to the data, that is, the one that minimized the residual sum of squares. This procedure quantified the onset of a sustained shift in CP without relying on a fixed amplitude threshold. The latency analysis contained a 20 ms bin size, so each segment was required to include at least two samples (i.e., an effective minimum segment duration of 40 ms). We also measured the time of peak CP deviation within the same interval, defined as the time point at which the CP trace reached its maximum deviation from 0.5; the peak magnitude was defined as the corresponding maximum deviation from 0.5. Mean ± SEM summaries for latency, peak time, and peak magnitude were computed across sessions x frequency bins.

Depth dependence of CP. To test whether CP varied in terms of spatial organization, we analyzed CP at different frequency bands as a function of relative channel position along the linear probe. Because probe angle varied across sessions and laminar boundaries were not identified, channel index was treated as a within-session measure of relative depth. For each session and frequency band, we quantified depth dependence in two ways: (1) a linear slope relating CP to relative channel depth, and (2) a superficial-deep contrast defined as the difference between mean CP in contacts closer to the probe tip versus contacts away from the tip. Channels that did not have sufficient balanced trial counts for preferred and anti-preferred choices to estimate CP were excluded. Statistical tests were performed at the session-level separately for each frequency band and condition.

Population decoding of stimulus information from LFP. We trained a simple linear decoder on simultaneously recorded high-gamma LFP spectra to classify stimulus information. For each session, we focused on correct trials and restricted to the high-evidence stimulus levels (MT coherence: 32%, 64%, 100%; V4 SNR: 25%, 60%, 100%), where stimulus information is expected to be the strongest. LFP power was computed during the stimulus-response epoch (50–250 ms after stimulus onset), and features were restricted to the high-gamma band (70–150 Hz). We averaged log power within the high-gamma band for each simultaneously recorded channel, yielding one high-gamma feature per channel. Trials were balanced within each stimulus level by random downsampling. For each condition separately, a logistic regression classifier with L2 regularization (function fitclinear, MATLAB) was trained and evaluated using repeated stratified K-fold cross validation (5 folds, 20 repeats). Feature standardization was performed using training data only and applied to held-out test data within each condition. Decoder performance was quantified as held-out classification accuracy. For each session, the classification accuracy was first pooled across folds within each repeat and then averaged across repeats to obtain a session-level estimate. We trained decoders separately for pre- and post-inactivation data because this analysis was designed to quantify stimulus information that could be recovered from the population LFP signal within each condition. For this comparison between pre- and post-inactivation conditions, the total number of retained trials was matched across conditions within each session.

[END]
---
[1] Url: https://journals.plos.org/plosbiology/article?id=10.1371/journal.pbio.3003873

Published and (C) by PLOS One
Content appears here under this condition or license: Creative Commons - Attribution BY 4.0.

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