Skip to main content

Main menu

  • HOME
  • CONTENT
    • Early Release
    • Featured
    • Current Issue
    • Issue Archive
    • Collections
    • Podcast
  • ALERTS
  • FOR AUTHORS
    • Information for Authors
    • Fees
    • Journal Clubs
    • eLetters
    • Submit
    • Special Collections
  • EDITORIAL BOARD
    • Editorial Board
    • ECR Advisory Board
    • Journal Staff
  • ABOUT
    • Overview
    • Advertise
    • For the Media
    • Rights and Permissions
    • Privacy Policy
    • Feedback
    • Accessibility
  • SUBSCRIBE

User menu

  • Log out
  • Log in
  • My Cart

Search

  • Advanced search
Journal of Neuroscience
  • Log out
  • Log in
  • My Cart
Journal of Neuroscience

Advanced Search

Submit a Manuscript
  • HOME
  • CONTENT
    • Early Release
    • Featured
    • Current Issue
    • Issue Archive
    • Collections
    • Podcast
  • ALERTS
  • FOR AUTHORS
    • Information for Authors
    • Fees
    • Journal Clubs
    • eLetters
    • Submit
    • Special Collections
  • EDITORIAL BOARD
    • Editorial Board
    • ECR Advisory Board
    • Journal Staff
  • ABOUT
    • Overview
    • Advertise
    • For the Media
    • Rights and Permissions
    • Privacy Policy
    • Feedback
    • Accessibility
  • SUBSCRIBE
PreviousNext
Research Articles, Behavioral/Cognitive

Orbitofrontal High-Gamma Reflects Spike-Dissociable Value and Decision Mechanisms

Dixit Sharma, Shira M. Lupkin and Vincent B. McGinty
Journal of Neuroscience 14 May 2025, 45 (20) e0789242025; https://doi.org/10.1523/JNEUROSCI.0789-24.2025
Dixit Sharma
1Center for Molecular and Behavioral Neuroscience, Rutgers University – Newark, Newark, New Jersey 07102
2Graduate Program in Neuroscience, Rutgers University – Newark, Newark, New Jersey 07102
  • Find this author on Google Scholar
  • Find this author on PubMed
  • Search for this author on this site
  • ORCID record for Dixit Sharma
Shira M. Lupkin
1Center for Molecular and Behavioral Neuroscience, Rutgers University – Newark, Newark, New Jersey 07102
2Graduate Program in Neuroscience, Rutgers University – Newark, Newark, New Jersey 07102
  • Find this author on Google Scholar
  • Find this author on PubMed
  • Search for this author on this site
Vincent B. McGinty
1Center for Molecular and Behavioral Neuroscience, Rutgers University – Newark, Newark, New Jersey 07102
  • Find this author on Google Scholar
  • Find this author on PubMed
  • Search for this author on this site
  • ORCID record for Vincent B. McGinty
  • Article
  • Figures & Data
  • Info & Metrics
  • eLetters
  • PDF
Loading

Abstract

The orbitofrontal cortex (OFC) plays a crucial role in value-based decisions. While much is known about how OFC neurons represent values, far less is known about information encoded in OFC local field potentials (LFPs). LFPs are important because they can reflect subthreshold activity not directly coupled to spiking and because they are potential targets for less invasive forms of brain–machine interface (BMI). We recorded neural activity in the OFC of male macaques performing a two-option value-based decision task. We compared the value- and decision-coding properties of high-gamma LFPs (HG, 50–150 Hz) to the coding properties of spiking multiunit activity (MUA) recorded concurrently on the same electrodes. HG and MUA both represented the values of decision targets, but HG signals had value-coding features that were distinct from concurrently measured MUA. On average HG amplitude increased monotonically with value, whereas in MUA the value encoding was net neutral on average. HG encoded a signal consistent with a comparison between target values, a signal which was negligible in MUA. In individual channels, HG could predict choice outcomes more accurately than MUA; however, when channels were combined in a population-based decoder, MUA was more accurate than HG. In summary, HG signals reveal value-coding features in OFC that could not be observed from spiking activity, including representation of value comparisons and more accurate behavioral predictions. These results have implications for the role of OFC in value-based decisions and suggest that high-frequency LFPs may be a viable—or even preferable—target for BMIs to assist cognitive function.

  • decision-making
  • local field potentials
  • neuroeconomics
  • nonhuman primates
  • prefrontal cortex

Significance Statement

High-frequency local field potentials (LFPs) are often assumed to be a mere proxy for local spiking activity. This study finds evidence to the contrary in the orbitofrontal cortex (OFC) of monkeys making value-based decisions. With respect to decision mechanisms, the results challenge previous findings by suggesting a role for OFC in computing value comparisons, evident in a comparison signal encoded in high gamma (HG) but not spiking. These results add to the growing evidence for spike/LFP dissociations in the prefrontal cortex and support the idea that HG is an important but overlooked resource for identifying neural computations in cognitive tasks. Additionally, single-channel HG furnished more accurate predictions about choice behavior, supporting the potential use of HG in cognitive neural prosthetics.

Introduction

Neural representations of value are essential for decision-making, learning, and other cognitive abilities. For decision-making in particular, converging evidence points to an important role for value representations in the orbitofrontal cortex (OFC). This includes evidence from electrophysiology (Tremblay and Schultz, 1999; Padoa-Schioppa and Assad, 2006), computational modeling (Rustichini and Padoa-Schioppa, 2015), human and animal lesion studies (Murray et al., 2015; Vaidya et al., 2018), and microstimulation (Ballesta et al., 2020; Knudsen and Wallis, 2020).

Much of what is known about the computations performed by OFC comes from examining spiking activity in single neurons. Many OFC neurons represent the value of items offered or chosen during decision-making (Thorpe et al., 1983; Tremblay and Schultz, 1999; Padoa-Schioppa and Assad, 2006; Hunt et al., 2018), and value representations defined by multi-neuron patterns of OFC firing can predict the outcome of value-based choices (Rich and Wallis, 2016; McGinty and Lupkin, 2023).

In comparison, far less is known about value- and decision-related information encoded by OFC local field potentials (LFPs). LFPs are an important electrophysiological signal because they reflect not only somatic and axonal currents generated by action potentials, but also the otherwise unobservable axon terminal and dendritic currents resulting from local synaptic transmission (Buzsáki et al., 2012; Einevoll et al., 2013). Thus, LFPs may furnish information about decision-related neural activity that is not observable from spiking in single neurons. In addition, LFPs represent potentially important targets for brain–machine interfaces (BMIs), which for practical reasons cannot always resolve spiking and must therefore rely on electrophysiological signals with lower spatial resolution (Stavisky et al., 2015).

LFPs in the high-gamma band (HG, typically 50–200 Hz) are a particularly attractive target for accessing neural processing without measuring spikes. This is because HG magnitude is often strongly correlated with the spiking of nearby neurons (Ray et al., 2008; Ray and Maunsell, 2011) and also reflects nearby synaptic currents and other subthreshold phenomena (Lindén et al., 2011; Reimann et al., 2013). This suggests the possibility that HG signals alone could be as—or more—informative than spiking signals for probing neural mechanisms, despite their lower spatial resolution. For example, Lundqvist et al. (2016, 2018) identified gamma-range LFP events, not reflecting local spiking, that encoded information in a working memory task.

Consistent with this view, Rich and Wallis (2017) showed that OFC spiking activity and concurrently measured HG signals both encoded the value of an anticipated reward in a non-decision-making operant task. However, they also identified differences between spiking and HG signals, in the form of distinct temporal and spatial patterns of value encoding. This dissociation between OFC spiking and HG value coding in a non-decision task suggests that a similar dissociation may also occur in a more complex decision-making context. Therefore, the goal of this study was to identify the unique properties of OFC HG and spiking during decision-making—particularly for the representation and comparison of multiple decision options (Strait et al., 2014; Hunt et al., 2018).

To address this question, we measured the modulation of OFC HG by the values of items offered in each trial of a two-alternative forced-choice decision task. We then compared three properties of these HG value representations to those of concurrently measured spiking signals. First, we examined the magnitude, time course, and sign of the modulation of HG as a function of the value of the target items. Such representations are thought to entail the input stage of a goods-based decision process (Padoa-Schioppa, 2011). Second, we asked how the representations of the values of the two targets were distributed across the population, especially the extent to which single HG channels simultaneously encoded both target values (Strait et al., 2014). Third, we quantified the choice-predictive accuracy of HG value signals, both in single channels and at the population level, and compared this to spike-based choice predictions. For all of these properties, we asked whether the HG signals merely reflected the information available from concurrently observed spiking or whether they departed from the properties of the spike signal.

The results show that compared with spiking, HG signals recorded on the same channels are more likely to be positively modulated by value, are more likely to reflect signatures of value comparison (Strait et al., 2014), and can furnish more accurate predictions about trial-by-trial choice outcomes. These findings corroborate the idea that HG in OFC does not merely reflect local spiking events (Ray et al., 2008) and suggest that value- and choice-relevant computations may be taking place at the level of synaptic interactions. In addition, these findings suggest that HG is a viable target for cognitive BMIs designed to monitor internal representations of value and choice intent and, for some applications, may furnish more informative signals than would be obtained from spiking alone.

Materials and Methods

Data source

The neural data in this study (38 sessions, 26,318 total trials before preprocessing and artifact trial removal) includes a portion of the data reported previously in McGinty and Lupkin (2023) (28 sessions, 19,936 total trials), as well as data not used in the previous publication (10 sessions, 6,382 total trials). The neural data analyses in McGinty and Lupkin (2023) use sorted single units, whereas the present study uses multiunit activity (MUA) and HG signals obtained from LFPs. Some MUA-based results will therefore be similar (but not identical) to results reported by McGinty and Lupkin (2023); HG results are entirely new.

Subjects and apparatus

All experimental procedures were performed according to the NIH Guide for the Care and Use of Laboratory Animals and were approved by the Animal Care and Use Committees of Stanford University and Rutgers University – Newark. The study subjects were two adult male rhesus macaques, referred to as Monkeys K and C, weighing ∼14 kg each at the time of the study. Data from Monkey C were acquired at Rutgers University – Newark, and data from Monkey K were acquired at Stanford University. The monkeys were implanted with MR-compatible head holders and recording chambers (Crist Instrument), and craniotomies were performed to access the OFC. Monkey K had bilateral chambers allowing recording from both hemispheres, and Monkey C had a single chamber on the left hemisphere. All surgical procedures were performed under isoflurane anesthesia using fully aseptic techniques and instruments. Analgesics and antibiotics were given pre-, intra-, and postoperatively as required. Both the monkeys had a minimum of 4 weeks of recovery time postsurgery.

Neurophysiology data were collected while monkeys were head-restrained and seated in a custom-built chair, with their eyes 57 cm away from a CRT monitor displaying the task stimuli (120 Hz refresh rate, 1,024 × 768 resolution). Three response levers (ENV-612M, Med Associates) were placed in front of the subjects within their reach. One lever was placed 21 cm below the display center, and the other two levers were located ∼8.5 cm to the left and right of the center lever. Stimulus presentation, reward delivery, and monitoring of lever presses and eye position were controlled through a set of custom scripts written by R. Kiani for the MATLAB computing environment (MathWorks) and Psychtoolbox-3 (Kleiner et al., 2007). Eye movements were recorded noninvasively at a sampling rate of 250 Hz (EyeLink, SR Research). Juice rewards were delivered via a gravity-fed reservoir and solenoid valve. Neural activity, eye movement, and task event data were acquired and stored using a Plexon OmniPlex system (Plexon). Analyses were performed using custom code and standard toolboxes in MATLAB 2019b.

Behavioral task

A condensed behavioral task sequence is illustrated in Figure 1a. The complete task sequence is described as follows. Monkeys initiated a trial by fixating on a central point on the display and manually depressing the center lever. After holding fixation and simultaneously pressing the center lever for a variable duration of 1–1.5 s, the fixation point disappeared, and two target arrays appeared, centered 7.5 degrees of visual angle to the left and right of the display center. Once the arrays appeared, the monkeys were allowed to shift their gaze to view the targets. Targets were initially masked and became visible only when the monkeys started moving their gaze toward one of the targets (see below). They were allowed to view the targets in any order and for any amount of time until they initiated their choice. The monkeys initiated their choice by lifting their hand from the center lever and then indicated their choice by pressing the left or right lever within a 400 ms (Monkey K) or 500 ms deadline (Monkey C). This deadline discouraged monkeys from deliberating (e.g., changing their mind) after lifting their hand from the center lever. After the left/right lever press, a reward of 1–5 drops of juice was delivered according to the value of the chosen target (Fig. 1b).

Figure 1.
  • Download figure
  • Open in new tab
  • Download powerpoint
Figure 1.

Economic decision-making task and behavioral performance. a, Abbreviated task sequence (not shown to scale), in which monkeys initially fixated on a central dot and then freely viewed the decision targets (the yellow and blue glyphs) before choosing one target by pressing a lever. The “+” shapes indicate the position of visual crowders around the targets; the actual crowders are shown in panel c. b, Example set of decision targets used in a typical recording session; the associated juice reward for each target (number of drops) is shown on the right. c, Close-up view of an example target (blue, center) surrounded by six multi-colored crowders, approximating their appearance on the task display. d, Choice performance: proportion of left choices as a function value difference between left and right targets (n = 12,379 and n = 15,839 trials for Monkeys C and K, respectively). Lines show a logistic fit. e, Mean reaction time for each monkey as a function of trial difficulty, defined in terms of the absolute difference in target values; lines show a linear fit.

Target properties, visual crowders, and initial mask

The choice targets were colored glyphs, ∼0.75 × 0.75 degrees of visual angle in size. An example set with associated rewards are shown in Figure 1b. New target sets with different color and shape combinations were generated every 1–5 sessions to minimize over-learning of the stimuli. Each target was associated with a juice reward ranging from 1 to 5 drops, with at least two unique targets representing identical reward volumes at each reward stratum (Fig. 1b, the five horizontal boxes). Note that the term “target value” refers to the juice volume associated with a given target, whereas “target identity” refers to a target's unique appearance (color/shape conjunction).

Two methods were used to encourage the monkeys to fixate directly on both targets before making a choice. First, each target was surrounded by six “crowder” stimuli placed in close proximity, as shown in Figure 1c. The crowders consisted of multicolored “×” or “+” shapes that had no association with reward and that were randomly generated for each target array. Crowders reduce the effectiveness of peripheral vision (Whitney and Levi, 2011; Crowder and Olson, 2015), which encourages the monkeys to identify the targets using high-acuity foveal vision—i.e., by shifting their gaze directly onto the targets. Second, the task display was programmed to initially obscure the targets (with a randomly chosen crowder stimulus) until the monkey initiated their first eye movement outside of the centrally located initial fixation window. This ensured that the monkey remained unaware of the target values and identities until they initiated a saccade. Once the initial masks were removed, the targets were visible and remained on the screen until the monkey initiated a choice by lifting its hand from the center lever, at which point all targets and crowders were extinguished.

The nominal patch luminance for all the colors used for targets and crowders in Monkey C's sessions was ∼2.7 cd/m2 with negligible background luminance and was ∼22 cd/m2 for Monkey K's sessions with a background luminance of ∼4 cd/m2, as measured by a Tektronix luminance J17 photometer with J1820 head. Note that during data collection, the luminance of the target stimuli was reduced by up to ∼50% of the nominal values, with no change in the crowders, to better obscure the targets (Fig. 1c). This reduction in the luminance did not prevent monkeys from identifying the targets because the choice performance was nearly perfect for the easiest trials (Fig. 1d).

Neurophysiological recordings

Linear recording arrays (Plexon V-probes) were introduced into the brain through a sharpened guide tube whose tip was inserted 1–3 mm below the dura. Probes had 16, 24, or 32 channels spaced either 50 or 100 µm apart. Depending on the number of probes used, between 32 and 80 channels were recorded per session (mean 50.0 and 55.6 channels for Monkeys C and K, respectively). OFC was identified based on gray/white matter transitions and by consulting a high-resolution MRI acquired from each animal. We targeted the lateral bank of the medial orbital sulcus and the laterally adjacent gyrus, with anterior-posterior coordinates ranging from +32 to +38 mm anterior to the interaural landmark (Öngür and Price, 2000; Saleem and Logothetis, 2012), corresponding approximately to Walker's areas 11 and 13. Arrays were placed to maximize the number of channels in gray matter.

Spike and LFP processing

For all the neural analyses in this study, except for the spectrogram shown in Figure 2, we defined two types of neural signals for every channel: MUA and the HG signal. Spike times were detected online by bandpass filtering the raw signal between 300 and 1,000 Hz and detecting negative deflections exceeding a threshold of 4 standard deviations below the mean signal amplitude. The MUA signal at each channel is, therefore, unsorted spiking, which reflects the collective activity of multiple nearby neurons. Using unsorted MUA, instead of sorted single-unit activity, permits spiking and HG to be directly compared on a channel-by-channel basis. This is particularly important for interpreting analyses in which HG shows different properties from MUA; if only well-isolated single units were used, then any discrepancy between spiking and HG could potentially be attributed to the influence of unsorted (and therefore unobserved) spike activity. Spike times were counted in 200 ms bins, time locked to the viewing of the decision targets in each trial.

Figure 2.
  • Download figure
  • Open in new tab
  • Download powerpoint
Figure 2.

Normalized power spectrograms. Normalized spectral power was measured in time windows aligned to the first target viewing time in each trial (x-axis). Normalized spectrograms were first computed for each channel in individual trials and then averaged across trials and channels (n = 783 channels for Monkey C, n = 1,183 for Monkey K; see Materials and Methods). Dotted lines indicate the 50–150 Hz frequency band defining HG in this study.

HG signals were obtained by first downsampling the wideband signals from 40 to 1 kHz using an 8th-order Chebyshev type I low-pass filter with a cutoff frequency of 400 Hz. The downsampled signal was then notch filtered at 60, 120, and 180 Hz to remove noise attributable to alternating current power sources. To isolate the HG signal, we bandpass filtered the notch-filtered data between 50 and 150 Hz using a 1,000 order bidirectional FIR filter (MATLAB functions designfilt.m and filtfilt.m). The signal was then rereferenced using common average referencing method across all channels within a probe (Ahmadi et al., 2021). A Hilbert transform was applied to the band-passed, rereferenced signal to obtain the instantaneous analytic amplitudes, which we refer to as HG. The 50–150 Hz band was chosen because the amplitude within this frequency band increases relative to baseline after target viewing (Fig. 2) and because this band is comparable with the range for high gamma used in previous studies (Ray et al., 2008; Rich and Wallis, 2017). The HG analytic amplitudes for each channel were segmented into trials aligned to the viewing of the decision targets. The trial-aligned data was binned in 200 ms windows at 40 ms increments for further analyses.

Artifact channel and trial removal

Prior to data analysis, we identified and removed channels and trials based on artifacts observed in LFP activity (Falaki et al., 2024). LFP activity was obtained by downsampling and notch filtering the wide-band 40 kHz signal, using the same filters as used for HG signal. The resultant signal was then high-pass (3 Hz) and low-pass (300 Hz) filtered using FIR filter of order 1,000 to obtain the LFP signal, which was referenced across channels on each probe using common average referencing (Ahmadi et al., 2021).

To detect artifact channels, we used the time-averaged LFP signal from the baseline period (−400 to −200 ms from the target onset). The artifact channels were defined as follows: (1) any channel with unrealistically high LFP amplitude (>400 µV) in at least 5% of trials and (2) any channel showing LFP activity deviating more than 3 standard deviations (SD) away from other simultaneously recorded channels in at least 5% of trials. We detected and removed 57 artifact channels (17 and 40 for Monkeys C and K), resulting in 1,966 channels for further analysis (783 and 1,183 for Monkeys C and K).

To detect and remove artifact trials, we used the LFP signal during the trial period (−320 to 720 ms relative to first target view) in artifact-removed channels. The artifact trials were detected using three criteria, as follows: (1) For every trial we calculated the absolute pairwise correlation between the time series of all recorded channels and then found the grand mean and SD of all pairwise correlations across trials. We rejected trials in which 25% of channels had correlations >4 SD above the grand mean. (2) In each session we computed the signal mean and SD at each time sample across channels and trials. We rejected trials in which any channel had an LFP signal that was >4 SD above the mean in at least 2% of time samples. (3) We rejected trials showing unrealistically high absolute LFP amplitudes (>400 µV) in any channel in at least 2% of time samples. Based on these criteria, we removed 23 and 140 artifact trials in Monkey C and K, resulting in 22,519 decision trials for further analysis (10,832 or 667.0 per session for Monkey C and 11,687 or 531.2 per session for Monkey K).

The spectrograms shown in Figure 2 were generated with Chronux toolbox in MATLAB using trial-level LFP signals (Mitra and Bokil, 2008), using three tapers and a time-bandwidth product of 2. Within each trial, the spectrogram was normalized within each frequency bin with respect to activity during that trial's baseline period (−400 to −200 ms before target onset). We then averaged the normalized spectrograms across trials to obtain the session-wise average spectrogram for each channel. Figure 2 shows the mean spectrogram averaged across channels for each monkey.

Data analysis

Behavior, task events, and key task variables

We quantified choice performance by a logistic regression in which the fraction of left choices was explained by the value difference between left and right targets (Fig. 1d). Decision reaction times were quantified as a function of choice difficulty, defined as the absolute difference in value between the two targets, with a difference of 0, (equal-value targets) representing the greatest difficulty, and 4, the lowest (Fig. 1e).

In this task, monkeys viewed the targets in each trial sequentially using self-paced saccadic eye movements. The viewing order, total number of target views, and total viewing times were determined entirely by the monkeys. For a detailed description of viewing and choice behavior, see Lupkin and McGinty (2023).

Analyses were organized with reference to the viewing order of the targets in each trial. Accordingly, the value associated with the first-viewed target is referred to as “value1,” and the value of the second-viewed target is “value2.” So that both variables would be defined in every trial, we only used the trials in which monkeys viewed both targets at least once in a trial (96.4% and 90.4% of trials for Monkeys C and K, respectively).

Because the monkeys' eye movements were unrestricted, the trial duration and event timing varied across trials. Therefore, the data in each trial were aligned to both the first- and second-target viewing times, and all major analyses were performed in reference to these time points. For some analyses, we use “early” and “late” analysis epochs, defined as activity 200–400 ms following the onset of fixations onto the first- and second-viewed targets in each trial.

Single-channel encoding of value and other decision variables

To quantify the encoding of task variables in single channels, we fit the following ordinary least squares linear model at every 200 ms time bin in 40 ms increments (Wilkinson notation):Y∼value1+value2+chosenvalue,(1) where Y is a trials-by-1 vector of MUA spike counts or HG analytic amplitudes for a given channel and time bin, value1 and value2 are the values of the first- and second-viewed targets in each trial (range 1–5 drops), and chosen value is the value of the target chosen in each trial (range 1–5 drops). All variables were z-scored across trials before performing the regression. Encoding was quantified by computing for each variable the coefficient of partial determination (CPD) using the methods of Hunt et al. (2018), which captures the unique variance explained by each variable after accounting for the variance explained by all the others. Within each time bin, the mean and SEM of the CPD were calculated across channels, separately for MUA and HG signals; the results are shown in Figure 4. We use the full regression model in Equation 1 to calculate the proportion of channels encoding the value variables and to obtain regression estimates used in Figures 5a, 6⇓⇓–9.

To quantify the unique contribution of MUA and HG to explaining variance in value at the single channel level, we calculated and compared the adjusted r2 of three models:value∼MUA(2a) value∼HG(2b) value∼HG+MUA,(2c) where value is either value1 or value2. These regressions were performed separately on data from the “early” and “late” analysis epochs (200–400 ms after viewing the first and second targets, respectively). These time windows were chosen based upon the time course of value representations observed in Figure 4 (gray and cyan shaded regions). The difference in adjusted r2 between the combined model (Eq. 2c) and the two single-predictor-variable models (Eq. 2a or 2b) was interpreted as the unique variance explained by each variable; adjusted r2 is used to account for the extra predictor variable in Equation 2c.

Choice probability analysis in single channels

We used choice probability (CP) analysis to examine the trial-to-trial relationship between neural variability and the variability in choice behavior, an approach first developed for perceptual decisions (Britten et al., 1996) and recently adapted for value-based decision studies (Conen and Padoa-Schioppa, 2015; McGinty and Lupkin, 2023). At an intuitive level, this analysis asks if the activity of value coding signals (MUA or HG) can classify upcoming choices on a trial-to-trial basis while holding all other related variables constant (such as the value or identity of the targets). For in-depth details, see McGinty and Lupkin (2023).

We performed CP analysis separately for MUA and HG at each channel and time bin. For each channel and time bin, the CP was computed twice, once using the encoding sign of value1 to set the positive class and once using the sign of value2. In this way, the CPs are computed with respect to each channel's encoding of the target values.

To compute CP, we first used half the trials as training trials to find the sign of value encoding for value1 and value2, by fitting two one-term linear models. Then, the remaining half of trials were used for classifying choice outcomes (computing CPs). To compute CPs in the second half of trials, we first normalized the neural signal within each channel and bin by taking the residuals of a two-way factorial ANOVA that explained neural signals as a function of identities of the two targets on each trial. (Identity refers to the 12 unique color/shape combinations.) This procedure removes variance attributable to the target identities and values from the neural signal. The normalized signals were then submitted to a receiver operating characteristic (ROC) analysis, which quantifies the ability of an ideal observer to distinguish between trials in which the first or second target was chosen. Classification was only performed over trials in which the target values differed by 0 or 1, because only these trials showed appreciable variability in choice (Fig. 1d).

Classification accuracy was quantified by the area under the ROC curve, where 0.5 indicates chance-level classification. The positive class was set to first-target choices when the value encoding sign was positive and set to second-target choices when the sign was negative. With this parameterization, an area under the ROC >0.5 indicates better than chance prediction of choice congruent with a channel's value1 sign, and an area <0.5 indicates better than chance predictions congruent with the value2 sign (Fig. 10a,c).

Choice probability analysis using population neural signals

We also classified choice outcomes using population-based, single-trial estimates of the target values, obtained by pooling neural activity recorded simultaneously on all channels of a probe (Yates et al., 2020). This method used a two-stage, regression-based linear encoder and decoder. In the encoding step, we fit a regression model to explain value1 or value2 as a function of neural activity over a set of training trials. In the decoding step, the fitted model is used to estimate value1 and value2 using neural activity in held-out test trials. Choice behavior is then predicted in the held-out trials by the submitting neurally derived value estimates to an ROC classifier. Because this approach requires simultaneously recorded neural signals, these steps were performed using simultaneously recorded signals on all channels of the same probe. In some recording sessions, we recorded multiple probes simultaneously. To address potential violations of statistical independence in averages and t test results, we confirmed results by randomly selecting one probe per session (see below, Bootstrap analysis).

In the encoding step, the data are first stratified into an equal number of training and test trials (chosen at random); the training trials are then used to fit the following L1-regularized linear regression (LASSO) model:Y=b0+b1S1+b2S2+…+bnSn.(3) Here Y is either value1 or value2. Si is the neural signal, (MUA or HG) from channel i, bi is the regression estimate for channel i, and n is the total number of simultaneously recorded channels on a probe (16, 24, or 32). The estimates {b1…bn} indicate the degree to which the signal of each channel uniquely contributes to explaining the variance in the value variable Y. The model was fit by LASSO regression, which minimizes the mean squared error plus a penalty according to the L1-norm (Hastie et al., 2015).

Before fitting Equation 3 to value1 or value2, the effects of the nonrelevant value variable were removed from the neural data (in only the training trials). This was necessary because the distribution of the value2 variable was not guaranteed to be identical across every stratum of the value1 variable; hence, changes in the neural signal attributable to value2 could be inappropriately attributed to value1, leading to inaccurate regression estimates. This step is particularly important for time bins in which more than one value variable is simultaneously represented (Fig. 4b, 200–400 ms). Therefore, before fitting the model for value1, neural data were mean centered with respect to the identity of the second target; likewise, before fitting the model for value2, the neural data were mean centered according to the identity of the first target.

For decoding, we estimated single-trial population-based value representations by computing the weighted sum of neural signals in the held-out trials:Y^=b0+b1S1′+b2S2′+…+bnSn′.(4) Here Y^ is the estimated value signal for all held-out trials, bi indicates the regression estimate from Equation 3 for channel i, b0 is the intercept, Si′ is the neural signal (MUA or HG) from the i-th channel in the held-out trials, and n is the total number of channels recorded simultaneously on a probe.

We then normalized Y^ by taking the residuals from two-way factorial ANOVA, in which Y^ is explained by the identity of the first target and the second target without an interaction term. This removes from Y^ variance attributable to the trial conditions, leaving only the residual variability in the neurally derived estimates of value1 and value2. The normalized Y^ was then used to classify trials according to the decision outcome (first or second target chosen) using an ROC analysis; the normalized Y^ calculated for held-out trials is referred to as “population value signal” or “population decode” for that trial. ROC classification was performed only over the test trials with target value differences of 0 or 1. The positive class was always defined as choices in favor of the first-viewed target. Therefore, the AUC was above 0.5 when population value signals corresponded to a greater tendency to select the first target and was below 0.5 when population value signals were associated with a tendency to select the second target.

To compare CPs calculated using single-channel and population value signals, we used 2 × 2 ANOVA with signal-type (MUA/HG), analysis-type (single-channel/population), and their interaction as factors. Post hoc tests were then performed between relevant pairs using the Tukey–Kramer procedure (MATLAB function multcompare.m).

In a follow-up analysis, we compared the variance in value explained by population-level MUA, HG, and combined MUA and HG signals. To do so, we computed for every probe the adjusted r2 for Equation 3 when Y was value1 and Si was MUA on all channels, HG on all channels, or the concatenated MUA and HG on all channels. The time bin used was 200–400 ms after first-target viewing. Results were averaged across probes and are shown in Figure 12.

Correlated variability

We calculated pairwise noise correlations across all channels in a session, using Pearson's correlations over the baseline signals measured −400 to −200 ms from target onset (i.e., during the initial fixation period). These correlations were performed separately for MUA and HG signals.

Bootstrap analysis

Because multiple channels were recorded in a session, and because the noise correlations between channels of each probe are positive on average for both MUA and HG, the results from individual channels are not strictly independent, violating the assumptions of many statistical tests. Therefore, to confirm key results we repeated selected analyses using a bootstrap procedure in which we randomly selected one channel at a time from each recording probe (n = 86). Results are reported as the median and 95% confidence intervals (CI) of 1,000 bootstrap samples. We used a similar procedure to confirm selected CP analysis results, by randomly sampling one probe per session (n = 38), and reporting the median and 95% CI over 1,000 samples.

Results

Task performance

Rhesus monkeys performed a two-alternative, forced-choice decision task in which they chose between two visual targets based on their associated fluid reward (Fig. 1a,b). Visual crowding and a gaze-contingent display (see Materials and Methods) were used to prevent the monkeys from discriminating the two targets with peripheral vision, requiring them to fixate directly on the targets to identify them and their associated values in each trial. All neural analyses were therefore performed in reference to the two target viewing times. The choice in each trial was indicated manually by pressing either the left or right response lever located below the task display.

The monkeys looked at both targets at least once in the large majority of trials (96.4% for Monkey C and 90.4% for Monkey K), and their choices and reaction times were consistent with a deliberative process dependent on both target values (Fig. 1d,e). For a detailed analysis of decision and target viewing behavior in this task, see Lupkin and McGinty (2023).

OFC spiking and high-gamma signals explain unique variance in value

We simultaneously recorded spiking and wideband signals from the OFC during the task. The MUA on each channel was obtained from unsorted threshold crossings, and the HG signal on each channel was defined as the analytical amplitude of the LFP signal between 50 and 150 Hz (see Materials and Methods). In this task, the monkeys viewed the two target stimuli in sequence, and all analyses were performed on data time locked to the viewing of the first and second targets.

A prior study characterized the value-encoding properties of HG signals in a non-decision, one-option reward expectation task (Rich and Wallis, 2017), but how HG encodes competing value targets during decision-making is unknown. Figure 3 shows an example of a single channel with activity modulated by the value of the first item the monkey viewed in each trial. In this channel, both the MUA firing rate (Fig. 3a,c) and HG amplitude (Fig. 3b,d) show positive modulation with value ∼200–400 ms after viewing the target.

Figure 3.
  • Download figure
  • Open in new tab
  • Download powerpoint
Figure 3.

A representative channel showing neural signals modulated by target value. a, Mean MUA firing rate measured on the channel after viewing the first target, segregated into three trial types: low value (1 or 2 drops of juice), mid value (3 drops), and high value (4 or 5 drops). Shaded regions show SEM across trials (n = 233, 234, and 218 for low, mid, and high, respectively). b, Mean analytic amplitude of the HG signal for three trial types. c, d, Same data as a and b but using neural signals that were z-scored across trials for each time bin.

To quantify the relationship between the neural signals on each channel and the values of the targets, we fit linear regression models at each time bin, separately for MUA and HG (Eq. 1). The models explained the neural signals as a function of the values of the two targets in each trial (variables value1 and value2), as well a value variable known to be encoded in OFC spiking activity, chosen value, defined as the value of the target chosen in each trial. Encoding for each variable was quantified by the coefficient of partial determination (CPD), which gives the variance explained by that variable after accounting for the effects of all other variables in the model (see Materials and Methods).

We found that MUA and HG encoded the value variables with similar time courses. As illustrated in Figure 4, the variance explained by value1 increased ∼200 ms after the first target was viewed for both MUA (Fig. 4a,e) and HG (Fig. 4c,g). Likewise, for both MUA and HG, the variance explained by value2 increased ∼200 ms after the second target was viewed (Fig. 4b,d,f,h).

Figure 4.
  • Download figure
  • Open in new tab
  • Download powerpoint
Figure 4.

Encoding of value variables in MUA and HG. The y-axis in each panel gives the encoding of three value variables (measured by CPD; see Materials and Methods) and the x-axis gives the time relative to viewing the first (a, c, e, g) or second (b, d, f, h) target, separately for Monkey C (a–d) and Monkey K (e–h). Shading shows SEM across channels (n = 783 for Monkey C and n = 1,183 for Monkey K). Shading indicates early (gray) and late (cyan) epochs which are used in subsequent analyses. Horizontal box plots show the distributions of 2nd target viewing times and reaction times (RT) as indicated.

However, one notable difference between MUA and HG was the strength of value1 encoding following the viewing of the second target. For the MUA signal, the explained variance for value1 peaked in the early epoch (mean CPD of 0.65% SEM 0.05% and 0.64% SEM 0.04 for Monkeys C and K; Fig. 4a,c, shaded region) and then remained similar or decreased in the late epoch (Monkey C 0.69% SEM 0.05, p = 0.43 by paired t test; Monkey K 0.46% SEM 0.02, p < 1 × 10−5; Fig. 4a,b,e,f, shaded regions). In contrast, for the HG signal, value1 encoding increased in the late epoch compared with the early epoch (Fig. 4d,h, arrow; Monkey C 1.20% SEM 0.05, p < 1 × 10−55 by paired t test; Monkey K 0.66% SEM 0.03, p < 1 × 10−5). A similar pattern was evident in the fraction of channels that had statistically significant encoding of these variables (not illustrated). Thus, both MUA and HG show a significant representation of value1 after viewing the second target, but only HG showed a significant increase in the explained variance compared with the epoch following the first target.

We then asked whether encoding was similar in the HG and MUA signals on the same channels. To do so we computed the correlation between the estimates from the regression model (Eq. 1), using data from both the early and late epochs (Fig. 4, shaded regions). We found a significant positive correlation between the MUA and HG estimates for value1 measured on the same channels in the early epoch (r = 0.37, p < 1 × 10−63; Fig. 5a). In the late epoch, similar correlations were seen for value2 (r = 0.39, p < 1 × 10−71, not illustrated) and value1 (r = 0.38, p < 1 × 10−67, not illustrated).

Figure 5.
  • Download figure
  • Open in new tab
  • Download powerpoint
Figure 5.

Variance in value explained by MUA and HG. a, Scatterplot comparing value encoding in MUA (x-axis) and HG (y-axis) signals. Each dot represents the regression estimates for value1 for a single channel (Eq. 1a,b) using data from the shaded epoch in Figure 4a,c,e,g (n = 1,966, data combined across monkeys). The black line denotes the least squares regression line. b, Box plots showing the distribution of explained variance in value1 by MUA, HG, or both signals (Eq. 2). Data are from the early epoch in Figure 4a,c,e,g. Only channels with significant effects in either MUA or HG were used (p < 0.005 uncorrected in Eq. 1, n = 547, data combined across monkeys); results were similar when using all channels (not illustrated). p values give the results of a paired t test across channels.

These modest positive correlations indicate similar, but not identical, modulation by value, suggesting that each signal may explain a unique portion of variance in value relative to the other. We quantified this by computing for each channel the variance in value1 explained by MUA alone, by HG alone, or by the combination of MUA and HG (Eq. 2), using early epoch activity. The combined signals explained a greater fraction of value1 variance than either signal alone, as measured by adjusted r2 (Fig. 5b). During the late epoch, the combined signals also explained more variance than either signal alone, for both value1 and value2 (not illustrated). Taken together, this indicates that MUA and HG reflect similar encoding of value at each channel but that each signal also captures a unique portion of the variance in value not explained by the other signal. Although the variance explained by MUA and HG signals were low in absolute terms, the results are consistent with previous studies (Kennerley et al., 2011; Hunt et al., 2018; Saez et al., 2018; Zhang et al., 2022) and are explained by the high across-trial variability in OFC neurons and the tendency for OFC spike counts to be overdispersed (McGinty et al., 2016).

High-gamma increases on average as a function of value

OFC neurons can either increase or decrease firing as a function of value, and we and others have observed that value-increasing and value-decreasing cells are present in roughly equal proportions (Padoa-Schioppa and Assad, 2006; McGinty et al., 2016; McGinty and Lupkin, 2023). However, it is unknown whether OFC HG signals have this same property. We quantified the sign of value encoding at each channel using the regression estimates obtained from Equation 1, where positive estimates indicate channels that increase signal magnitude with value (Fig. 3a, between 200 and 400 ms), and negative estimates indicate the opposite.

Consistent with prior studies of spiking activity, the average regression estimates for value1-encoding MUA channels were no different from zero after viewing the first target (early epoch, −0.018 SEM 0.009, p = 0.0502, n = 338, for t test comparison to 0; Fig. 6a,c). In contrast, for concurrently measured value-encoding HG channels, value1 estimates were positive on average (0.029 SEM 0.007, p = 0.00001, n = 404; Fig. 6b,d) and were significantly more positive than the estimates measured in MUA (p < 1 × 10−9, n = 338, paired t test). Similarly, value2 estimates measured after the second target were on average larger for HG than for MUA (late epoch, 0.097 SEM 0.005 vs 0.044 SEM 0.007, n = 365, p < 1 × 10−14; Fig. 6e,f). Finally, for the representation of value1 after the second target, both MUA and HG showed negative encoding on average (MUA: −0.060 SEM 0.005, p < 1 × 10−33; HG: −0.148 SEM 0.003, p < 1 × 10−210; Fig. 6g,h); however, the HG estimates were significantly more negative than MUA estimates (n = 546, p < 1 × 10−64).

Figure 6.
  • Download figure
  • Open in new tab
  • Download powerpoint
Figure 6.

Sign of value coding in MUA and HG. a, b, Regression estimates for value1 (Eq. 1) observed in MUA (a) and HG (b), plotted in 200 ms bins time locked to the viewing of the first target; only channels with significant value encoding are shown (uncorrected p < 0.005, n = 256 for MUA and n = 161 for HG, combined across monkeys). Red indicates positive regression estimates (signal increases as a function of value), and blue indicates negative. c–h, Average regression estimates across value-encoding channels. Panels on the left show average estimates measured in MUA for (c) value1 after viewing the first target (n = 338), (e) value2 after viewing the second target (n = 365), and (g) value1 after viewing the second target (n = 546). Panels on the right (d, f, h) show the same, but for HG signals. Data were combined across monkeys. Data from the shaded “early” and “late” epochs were used for analyses reported in the main text.

In sum, compared with concurrently observed spiking signals, HG signals were more likely to show net positive or net negative modulation by value, with positive modulation for targets immediately after they are each viewed and negative modulation for the persistent representation of the value of first target after viewing the second target. We confirmed these findings using all channels separately for each monkey in early and late epochs (Fig. 7). We also confirmed these findings using a bootstrap procedure that ensures independent sampling of activity from each probe (not illustrated).

Figure 7.
  • Download figure
  • Open in new tab
  • Download powerpoint
Figure 7.

Regression estimates across all channels. Average regression estimates across all channels are shown for Monkey C (left panel, n = 783) and Monkey K (right panel, n = 1,183) during early and late epochs. Light bars indicate MUA, dark bars indicate HG, and error bars indicate SEM. Significance indicators (*) give the results of a paired t test.

High-gamma reflects a signature of value comparison

Value comparison is an essential computation for economic decision-making. Single neurons can express a value comparison signal by encoding the signed difference between two offer values (e.g., value1 minus value2) or more generally by encoding the two offer values with opposite signs of modulation (Strait et al., 2014). This form of value comparison signal has been identified in neurons in the ventromedial prefrontal cortex and ventral striatum (Strait et al., 2014; Maisson et al., 2021) but appears to be weak or absent in OFC cells (McGinty and Lupkin, 2023). Here, we asked if the same is true for OFC HG signals.

Consistent with prior studies, we found that after the monkeys view both targets, the values of both targets are encoded simultaneously (Fig. 4b,d,f,h). We computed correlations between regression coefficients for value1 and value2 after viewing the second target (late epoch). A negative correlation between the regression estimates would indicate that single neurons tend to encode values with opposite signs, consistent with a value comparison signal as described above. For MUA, we found a weak but statistically significant negative correlation between regression estimates (rpearson = −0.09, p = 0.00014; rspearman = −0.03, p = 0.16 for data combined across monkeys), consistent with previous reports (Hunt et al., 2018; McGinty and Lupkin, 2023). In contrast, a much larger negative correlation was found for the HG signal (rpearson = −0.23, p < 1 × 10−24; rspearman = −0.22, p < 1 × 10−21), indicating that HG channels that were positively modulated by value2 also tended to show negative modulation by value1 and vice versa. These results were consistent across monkeys (Fig. 8). We also confirmed these findings using a model that uses only value1 and value2 as regressors (MUA: rp = −0.09, p = p < 1 × 10−4; rs = −0.05, p = 0.03; HG: rp = −0.21, p < 1 × 10−19; rs = −0.22, p < 1 × 10−22). Thus, the comparison between offer values is reflected in HG to a much stronger extent than in concurrently observed spiking.

Figure 8.
  • Download figure
  • Open in new tab
  • Download powerpoint
Figure 8.

Signature of value comparisons were stronger in HG than in MUA. The scatterplots show the channel-by-channel relationship between regression estimates for value1 and value2 (Eq. 1) measured in the late epoch in Figure 4, separately for MUA (left panels) and HG (right) in Monkey C (a) and Monkey K (b). Each dot represents a single channel (n = 783 and n = 1,183 for Monkeys C and K, respectively). The “r” and “rho” statistics indicate Pearson’s and Spearman's coefficients, respectively.

For both HG and MUA signals, individual channels showed similar encoding for value1 in the early epoch as they did for value2 in the late epoch, evident from positive correlations between regression estimates (Fig. 9a). This suggests a “view-centered” value coding scheme in which individual channels have a consistent pattern of value encoding for whichever item is being viewed (Grabenhorst et al., 2023). In both signals we also found a tendency for individual channels to change their sign for the encoding of value1 between the early epoch and late epoch. This was evident from a small negative correlation in the value1 regression estimates measured at the two time points (Fig. 9b).

Figure 9.
  • Download figure
  • Open in new tab
  • Download powerpoint
Figure 9.

a, Relationship between regression estimates of value1 after the first target and value2 after the second target. b, Relationship between value 1 after the first target and value1 after the second target. Early and late epoch indicate 200–400 ms after viewing the first and second targets, respectively (Fig. 4, shaded regions). Pearson's and Spearman's correlation coefficients are indicated by “r” and “rho,” respectively. Data are combined across monkeys (n = 1,966 channels).

Choice-predictive value representations in MUA and HG

Decision neuroscience seeks to explain choice behavior as a function of underlying neural activity. One classic approach to this question is “choice probability” (CP) analysis, which explains trial-by-trial variability in choice according to variability in neural activity (Britten et al., 1996). This analysis uses the residual variability in these measures by first removing other sources of variance in the data (i.e., the stimulus values in each trial), meaning that it gives the relationship between neural activity and choices over and above the activity directly related to the task conditions. When CP effects are evident well before the choice is made in each trial, it means that neural activity can predict what the monkey will choose; this is considered an important signature of neural decision mechanisms, because it is a necessary (though not sufficient) condition for demonstrating a causal relationship between neural activity and choice.

In a recent study we showed significant choice-predictive activity from population-level representations of value decoded from many simultaneously recorded OFC cells—but no significant choice-predictive activity from single cells (McGinty and Lupkin, 2023). Here, we use a similar approach to compare the choice-predictive power of MUA and HG value signals.

We first computed the CP (i.e., choice-predictive accuracy) effects for the HG and MUA signals of individual channels. Intuitively, for a channel that encodes value1 with a positive sign, increased firing would be expected to correspond to an increased subjective value for the first offer and an increased likelihood of choosing the first offer. The opposite would be expected for a channel encoding value1 with a negative sign. Therefore, for each channel we computed CP with respect to the sign of encoding for value1; a separate sign-based CP was computed with respect to each channel's encoding of value2.

Following the first target, neither the MUA nor HG signals showed consistently better than chance prediction of choice, evident in the mean ROC values (see Materials and Methods) that were not consistently different from 0.5, for both value1- and value2-based analyses (Fig. 10a, left). In contrast, after the second target, both MUA and HG showed better than chance prediction accuracy for both value1 and value2 representations (Fig. 10a, right, open circles and filled squares). The accuracy for HG was consistently higher than MUA for value1-related classification, but not for value2 (Fig. 10a,b, right). We confirmed the results using a bootstrap-based sampling method to ensure statistical independence of data from channels recorded on the same probe (mean and SEM of difference between MUA and HG AUC's: early epoch, value1: 0.0274 SEM 0.003, p = 0.48 by paired t test across probes; late epoch, value1: −0.106 SEM 0.004, p = 0.005; late epoch, value2: 0.0040 SEM 0.004, p = 0.28).

Figure 10.
  • Download figure
  • Open in new tab
  • Download powerpoint
Figure 10.

Choice probability (CP) analysis for individual channels and population-level value signals. a, Choice predictive accuracy is quantified by the mean area under the ROC curve (AUC), separately for MUA and HG and computed with reference to neural representation of either value1 or value2 (see Materials and Methods). The x-axes give the time relative to the viewing of the first target (left panel) and second target (right panel); shaded early and late epochs indicate time windows used for analyses in panels b, d, and e. Shapes at the bottom and top of each panel indicate significance (one-sample t test against chance, p < 0.01, FWER corrected across time points), where empty dots indicate significance for MUA and filled squares indicate significance for HG, separately for value1 (top) and value2 (bottom). In this panel, AUCs represent the accuracy of the ROC analysis for predicting choices in favor of the first target; hence, the AUC for value1-based predictions is >0.5, and the AUC for value2 is <0.5. b, Bar plots showing predictive accuracies relative to chance in MUA and HG in early (left panel) and late epochs (right panel). In panels b, d, and e, the data for are rectified so that chance performance is 0 and better-than-chance performance is >0. Significance indicators show results of paired t tests across channels. c, Same as a but using population-based value decodes for each probe over n = 86 probes (see Materials and Methods). d, Same as b but using data from the population-based analysis in panel c. e, The data in b and d are re-arranged to directly compare single-channel to population-level predictive accuracies for value1 in the early epoch (left), value2 in the late epoch (middle), and value1 in the late epoch (right). Significance indicator shows results of post hoc tests from a two-way ANOVA with an interaction term (see main text). Shaded contours and error bars indicate SEM across channels for single-channel analyses (n = 1,966 in panel a, b, and indicated bars in panel e) or across probes for population analyses (n = 86 in panels c, d, and indicated bars in panel e).

We next assessed the predictive accuracy of population-level value signals, obtained by taking the weighted sum of activity from all the simultaneously recorded channels in a probe [see Materials and Methods and McGinty and Lupkin (2023)]. The overall accuracy of population-level signals was greater than single channels (compare y-axis in Fig. 10a–c; see also Figs. 10e, 11), which was expected because population signals pool information across multiple channels. Within the population-level results, MUA-derived signals showed equal or greater accuracy than those derived from HG (Fig. 10d), in contrast to the single-channel data, where HG showed equal or higher accuracy than MUA (Fig. 10b).

This pattern of results suggests that for MUA, population-level predictions were significantly more accurate than single-channel–level predictions but that this was not the case for HG (Fig. 10). To directly compare population CP with single-channel CP in both signals (Fig. 10e), we used a two-way ANOVA with signal type (MUA/HG), analysis type (single-channel/population), and their interaction as factors. For data in the late epoch, the interaction term was near the significance threshold for both value1 (F(1,4098) = 3.37, p = 0.06; Fig. 10e, right) and value2 (s(1,4098) = 3.29, p = 0.07; Fig. 10e, middle), suggesting that the improved performance of the population decoder compared with single channels was larger in MUA than in HG. A post hoc comparison between single channel and population CP values for MUA confirmed these findings (MUA: value1 p = 2.8 × 10−5, value2 p = 8.2 × 10−5; HG: value1 p = 0.20, value2 p = 0.28). For CP effects related to value1 after the first target (early epoch; Fig. 10e, left) the interaction term was not significant (F(1,4098) = 1.24, p = 0.26), but post hoc tests revealed a significant difference between single-channel and population-based CP for the MUA data (p = 0.027) but not for HG (p = 0.6). These results were consistent across monkeys (Fig. 11).

Figure 11.
  • Download figure
  • Open in new tab
  • Download powerpoint
Figure 11.

Comparison of single-channel and population-level choice predictions for individual monkey. Analysis and conventions are the same as in Figure 10e, except data are shown individually for Monkey C (panel a, n = 783 channels and 25 probes) and Monkey K (panel b, n = 1,183 channels and 61 probes.

We observed a similar pattern when comparing the average explained variance in value1 in single channels (Eq. 2, Fig. 5b) to the variance explained by multiple simultaneous channels (Eq. 3, Fig. 12). For example, using MUA data from the early epoch, the average explained variance in value1 for single channels was 3.1% SEM 0.1 (Eq. 2a; results in Fig. 5b), and was much larger for the population-level model (11.9% SEM 1.3, p < 1 × 10−10, by two-sample t test; Eq. 3). In contrast, using the HG data over that same time interval, the mean explained variance for single channels was 1.8% SEM 0.1 (Eq. 2b) but was only slightly more for the population-level model (4.4% SEM 0.6, p < 1 × 10−10; Eq. 3).

Figure 12.
  • Download figure
  • Open in new tab
  • Download powerpoint
Figure 12.

Explained variance in value using pooled neural signals. Each box plot shows distribution of explained variance in value1 across probes (n = 86) calculated using a model (Eq. 3) that pools only MUA, only HG, or both MUA and HG signals across channels in a probe. p values give the results of a paired t test across probes. Data are from 200 to 400 ms after viewing the first target (early epoch).

Taken together, these results show that for both MUA and HG, pooling value signals across multiple channels increases the decodable information about choices and values; but this effect is greater for MUA signals compared with HG. Linear decoders like the one used for the population-based model above perform best when the individual signals are independent and contribute unique information about the variable to be decoded (Kohn et al., 2016). The fact that pooling HG signals over many channels results in only a modest increase in decoding performance over single channels suggests that compared with MUA, HG signals measured on different channels of the same probe are more similar (less independent) from trial-to-trial (Hwang and Andersen, 2013).

To test this hypothesis, we measured the noise correlations between pairs of simultaneously recorded channels using activity during the initial fixation period and compared the correlations for MUA to those for HG. On average, the noise correlations for MUA on pairs of simultaneously recorded channels was 0.065 SEM 0.001 (n = 61,370 pairs), whereas for HG signals it was 0.186 SEM 0.001 (p < 1 × 10−100, paired t test). We found the same pattern when considering only pairs of adjacent channels (MUA: 0.336 SEM 0.006; HG 0.683 SEM 0.004, n = 1,988 pairs, p < 1 × 10−100) as well as pairs of channels separated by more than four array channels (minimum 200 µM distance, MUA: 0.044 SEM 0.000; HG 0.144 SEM 0.001, n = 53,656 pairs, p < 1 × 10−100).

Discussion

High-frequency LFPs are often regarded as a mere proxy of local population spiking (Belitski et al., 2008; Rasch et al., 2008; Ray et al., 2008; Ray and Maunsell, 2011). However, recent studies have challenged this view, showing differences in the information encoded in prefrontal LFPs and spiking in primates performing cognitive tasks (Lundqvist et al., 2016, 2023). In this study, we investigated encoding of value- and decision-related information in primate OFC, focusing on the differences between HG and concurrently measured MUA. While we identified similar features of HG and MUA value representations, we also identified significant differences: First, in MUA the value of a just-viewed item was encoded by either increases or decreases in activity in roughly equal numbers of channels, consistent with earlier studies (Padoa-Schioppa and Assad, 2006; McGinty et al., 2016). In contrast, HG channels were much more likely to encode values with increasing activity. Second, while both MUA and HG encoded the value of a previously viewed item, many more HG channels significantly encoded this variable. Third, we identified a signature of value comparison that was clearly evident in HG, but negligible in MUA. Finally, trial-wise variability in choices were better predicted with HG than with MUA recorded from the same channels.

These results add to the growing evidence that HG is not merely a proxy for local spiking, especially with regard to cognitive variables encoded in the prefrontal cortex. Moreover, the results also potentially explain differences in data obtained in humans and monkeys and provide new insights into economic decision mechanisms, as we explain below.

Dissociable representation of cognitive variables in spikes and HG

Across species and in different brain areas, many studies report that spiking and HG activity are correlated (Belitski et al., 2008; Rasch et al., 2008; Ray et al., 2008; Ray and Maunsell, 2011) and that spikes and HG exhibit similar encoding of task variables (Donnelly et al., 2014; Tremblay et al., 2015; Rich and Wallis, 2017). However, growing evidence from prefrontal cortical recordings suggests differences in how HG and spiking encode variables in cognitive tasks. For example, Rich and Wallis (2017) showed that in OFC both spikes and HG encode economic value but that HG signals had large-scale spatiotemporal patterns that were not evident in spiking. In the dorsolateral prefrontal cortex, Lundqvist and colleagues identified brief bursts of HG-range signals (50–100 Hz) that were separable from spiking activity in working memory tasks (Lundqvist et al., 2016). Moreover, in a follow-up study they show that HG (but not spiking) preferentially encoded higher-level “control” signals related to the task rules, whereas both HG and spiking encoded lower-level task variables related to the task stimuli (Lundqvist et al., 2023).

Economic value is an abstract, subjective variable that in our task must be relearned every 2–5 sessions when a new set of stimuli are introduced (neural data is collected only after animals learned the stimuli well; see Materials and Methods). In addition to representing economic values, the decision task also requires two other cognitive operations: after the second offer is viewed in each trial, the monkeys must remember the value of the first offer; and they must also compare this remembered value to the second offer. Interestingly, we find clear dissociations between HG and MUA related to all three operations: Compared with MUA, HG was more likely to encode values with positive modulation (Figs. 6, 7), had many more channels encoding the value of the remembered offer (Fig. 4), and showed a stronger signature of value comparison (Fig. 8). Thus, our results are consistent with recent observations showing a dissociation between cognitive functions reflected in spikes and HG.

The neurophysiological basis of this dissociation is unclear. Spiking signals by definition reflect only suprathreshold excitation, whereas LFPs (including HG) emerge from a combination of sources, such as synaptic transmission, dendritic potentials, and suprathreshold and subthreshold somatic potentials (Mitzdorf, 1985; Buzsáki et al., 2012; Einevoll et al., 2013). Therefore, a natural hypothesis is that subthreshold processes, such as NMDA-mediated dendritic potentials (Leszczyński et al., 2020), are modulated by economic values in a manner that does not immediately affect spiking—leading to the dissociation in value coding properties of HG and MUA.

Another possibility, not mutually exclusive, is that the spike/HG dissociation in our data reflects the differential contribution of excitatory and inhibitory neurons. Thresholded spiking activity measured in MUA is primarily driven by excitatory pyramidal neurons (Supèr and Roelfsema, 2005), whereas emerging evidence suggests that gamma-range signals are primarily driven by parvalbumin (PV) and somatostatin (SOM) inhibitory interneurons (Verret et al., 2012; Chen et al., 2017; \Veit et al., 2017; Guyon et al., 2021; Wagatsuma et al., 2023). Inhibitory interneurons are abundant in cortical layers 2 and 3 (Llorca and Deogracias, 2022) and HG in layers 2/3 is stronger (Mendoza-Halliday et al., 2024) and more dissociable from spiking than in deeper layers (Leszczyński et al., 2020). Thus, layer 2/3 inhibitory interneurons may have an especially important role in the subthreshold processes that produce the spike-dissociable cognitive signals we observed in HG. Supporting this idea, it has been suggested that the signature of value comparison identified in our data (Fig. 8, anticorrelated value coding) arises from a “mutual inhibition” mechanism involving push–pull-like interactions between local inhibitory interneurons and pyramidal neurons (Wang, 2002; Strait et al., 2014). Additionally, Ballesta et al. (2022) found that behavioral signatures of value comparison—but not other decision processes—were uniquely disrupted by low-current stimulation of OFC; they speculated that low currents preferentially activated inhibitory interneurons, thereby disrupting the excitation–inhibition in local circuitry and value comparison computations.

Future experiments could test these hypotheses by identifying putative excitatory and inhibitory neurons and resolving cortical layers in neural recordings. The critical question would be whether interneurons in layers 2/3 have distinct neurophysiological signatures that could account for the properties of HG, such as preferential encoding of task variables or underlying cognitive processes (e.g., value comparison).

Implications for neural decision mechanisms

Economic choice requires assigning values to decision options and then comparing these values. In a recent study, we concluded that OFC is involved in assigning values but that OFC did not compute value comparisons (McGinty and Lupkin, 2023). This study examined only spiking in OFC, and its conclusions were based on two properties of the spike-based value code: First, single neurons represent values with either increasing or decreasing firing rates, in roughly equal proportions. But at a behavioral level, value influences behavior in a monotonic fashion (e.g., we pursue large rewards more vigorously than small rewards). Second, the neurophysiological signatures of value comparison that have been identified in some prefrontal regions (Strait et al., 2014; Maisson et al., 2021) are negligible in OFC spiking, suggesting that these comparisons are computed in a downstream region. Together, these prior spike-based findings suggested that the value signals in OFC reflect an early stage of information processing and that later stages downstream of OFC compare the values and transform the value signals into a monotonic code.

However, the value-coding properties we now show in HG suggest a more nuanced view of the value- and choice-related computations in OFC. HG signals tend to be monotonic and net positive (for currently viewed items) and tend to encode relative values. This suggests that OFC is not only involved in the “early stage” processing of values but rather may be involved in later stage processes such as value comparisons. Consistent with this idea, low-current microstimulation in OFC selectively disrupts value comparisons in simple economic choices (Ballesta et al., 2022).

More generally, these results indicate that spiking is not the only informative measure of information processing in economic decisions and that future studies should consider information encoded in both spiking and HG signals. An important open question is whether the processes driving comparison signals in HG are mediated entirely by local activity within OFC, by inputs or feedback to OFC from other cortical areas, or a combination of these. These questions could be addressed by future experiments resolving cortical layers and neural identities, as discussed above. A second open question is how HG signals ultimately influence computations in downstream areas, which could be addressed by multiregion recordings in potential OFC targets.

Implications for human neurophysiological and imaging studies

Intracranial value signals are often studied in animal models, but recordings in humans are available only in limited clinical settings. In humans, value signals in high frequency LFPs in the OFC consistently show net positive modulation with value (Saez et al., 2018; Lopez-Persem et al., 2020; Shih et al., 2023). In contrast, single neuron value signals in nonhuman primate OFC tend to show net neutral value encoding across populations of many neurons (McGinty et al., 2016; McGinty and Lupkin, 2023). Our results showing net positive encoding in HG but net neutral encoding in spikes from the same electrodes strongly suggests that this discrepancy is due to differences in the electrophysiological signals rather than differences between species.

Human functional imaging studies typically find that BOLD signals in vmPFC are positively related to value [Lebreton et al., 2009; see Bartra et al. (2013) for a meta-analysis]. BOLD signals are more coupled to gamma than spiking (Nir et al., 2007; Goense and Logothetis, 2008; Magri et al., 2012), suggesting that positive value encoding was observed in human fMRI, human intracranial HG, and monkey HG index similar neural processes. More broadly, these results suggest that HG may be an especially useful signal for bridging observations across species and modalities in the study of higher cognitive processes.

Implications for cognitive neuroprosthetics

BMIs typically use neuronal spiking activity to decode behavior (Andersen et al., 2010, 2022). However, using spiking signals for chronic implants is not ideal, because the isolation of single units deteriorates over time (Dickey et al., 2009; Chestek et al., 2011; Perge et al., 2014). Hence, there is growing interest in decoding from LFPs, which are more stable over time than spikes (Perge et al., 2014; Tremblay et al., 2015; De Sousa et al., 2021; Prakash et al., 2021, 2022; Proix et al., 2022).

With regard to values and choices, HG was as or more informative than spiking signals in nearly every respect, consistent with recent studies showing accurate decoding of abstract cognitive variables from LFPs (Wilson et al., 2020; Prakash et al., 2021). In addition, because HG is stronger in superficial cortical layers (Mendoza-Halliday et al., 2024), our results could potentially extend to less invasive modalities, such as electrocorticography (ECoG). Together, our results add to the growing evidence that high-frequency LFPs are a viable, and perhaps even preferable, target for chronic BMIs to assist higher cognitive function in humans.

However, there was one analysis for which spiking activity was more informative than HG: When decoding the monkeys' choices, single HG channels were more accurate on average than single spiking channels, but a multivariate spike decoder was more accurate than a multivariate HG decoder. This result is in line with previous studies showing that population decoders with spikes perform better than LFP bands (Hwang and Andersen, 2013). This suggests that while HG signals alone are highly informative, optimal performance may require a multimodal decoder with separate processing streams for spikes and LFPs. Important directions for future work are the nature of HG signals that can be measured in more complex cognitive tasks, as well as the potential spike–HG dissociations in other prefrontal areas that have a high degree of heterogeneous encoding.

Footnotes

  • We thanks W.T. Newsome for funding and material support; J. Brown, E. Carson, S. Fong, A. McCormick, M. Ortiz, J. Powell, J. Sanders, and D. Siegel for technical assistance; and D. Headley and B. Krekelberg for helpful comments on the manuscript.

  • This work was supported by the Howard Hughes Medical Institute (W.T. Newsome), National Institutes of Health Grant K01-DA-036659-01 (V.B.M.), the Busch Biomedical Foundation (V.B.M.), and a Whitehall Foundation Fellowship (V.B.M.).

  • The authors declare no competing financial interests.

  • Correspondence should be addressed to Vincent B. McGinty at vince.mcginty{at}rutgers.edu.

SfN exclusive license.

References

  1. ↵
    1. Ahmadi N,
    2. Constandinou TG,
    3. Bouganis C-S
    (2021) Impact of referencing scheme on decoding performance of LFP-based brain-machine interface. J Neural Eng 18:016028. https://doi.org/10.1088/1741-2552/abce3c
    OpenUrlCrossRef
  2. ↵
    1. Andersen RA,
    2. Aflalo T,
    3. Bashford L,
    4. Bjånes D,
    5. Kellis S
    (2022) Exploring cognition with brain–machine interfaces. Annu Rev Psychol 73:131–158. https://doi.org/10.1146/annurev-psych-030221-030214
    OpenUrlCrossRefPubMed
  3. ↵
    1. Andersen RA,
    2. Hwang EJ,
    3. Mulliken GH
    (2010) Cognitive neural prosthetics. Annu Rev Psychol 61:169–190. https://doi.org/10.1146/annurev.psych.093008.100503 pmid:19575625
    OpenUrlCrossRefPubMed
  4. ↵
    1. Ballesta S,
    2. Shi W,
    3. Conen KE,
    4. Padoa-Schioppa C
    (2020) Values encoded in orbitofrontal cortex are causally related to economic choices. Nature 588:450–453. https://doi.org/10.1038/s41586-020-2880-x pmid:33139951
    OpenUrlCrossRefPubMed
  5. ↵
    1. Ballesta S,
    2. Shi W,
    3. Padoa-Schioppa C
    (2022) Orbitofrontal cortex contributes to the comparison of values underlying economic choices. Nat Commun 13:4405. https://doi.org/10.1038/s41467-022-32199-y pmid:35906242
    OpenUrlCrossRefPubMed
  6. ↵
    1. Bartra O,
    2. McGuire JT,
    3. Kable JW
    (2013) The valuation system: a coordinate-based meta-analysis of BOLD fMRI experiments examining neural correlates of subjective value. Neuroimage 76:412–427. https://doi.org/10.1016/j.neuroimage.2013.02.063 pmid:23507394
    OpenUrlCrossRefPubMed
  7. ↵
    1. Belitski A,
    2. Gretton A,
    3. Magri C,
    4. Murayama Y,
    5. Montemurro MA,
    6. Logothetis NK,
    7. Panzeri S
    (2008) Low-frequency local field potentials and spikes in primary visual cortex convey independent visual information. J Neurosci 28:Article 22. https://doi.org/10.1523/JNEUROSCI.0009-08.2008 pmid:18509031
    OpenUrlPubMed
  8. ↵
    1. Britten KH,
    2. Newsome WT,
    3. Shadlen MN,
    4. Celebrini S,
    5. Movshon JA
    (1996) A relationship between behavioral choice and the visual responses of neurons in macaque MT. Vis Neurosci 13:87–100. https://doi.org/10.1017/S095252380000715X
    OpenUrlCrossRefPubMed
  9. ↵
    1. Buzsáki G,
    2. Anastassiou CA,
    3. Koch C
    (2012) The origin of extracellular fields and currents—EEG, ECoG, LFP and spikes. Na Rev Neurosci 13:Article 6. https://doi.org/10.1038/nrn3241 pmid:22595786
    OpenUrlPubMed
  10. ↵
    1. Chen G,
    2. Zhang Y,
    3. Li X,
    4. Zhao X,
    5. Ye Q,
    6. Lin Y,
    7. Tao HW,
    8. Rasch MJ,
    9. Zhang X
    (2017) Distinct inhibitory circuits orchestrate cortical beta and gamma band oscillations. Neuron 96(6):1403–1418.e6. https://doi.org/10.1016/j.neuron.2017.11.033 pmid:29268099
    OpenUrlPubMed
  11. ↵
    1. Chestek CA, et al.
    (2011) Long-term stability of neural prosthetic control signals from silicon cortical arrays in rhesus macaque motor cortex. J Neural Eng 8:045005. https://doi.org/10.1088/1741-2560/8/4/045005 pmid:21775782
    OpenUrlCrossRefPubMed
  12. ↵
    1. Conen KE,
    2. Padoa-Schioppa C
    (2015) Neuronal variability in orbitofrontal cortex during economic decisions. J Neurophysiol 114:1367–1381. https://doi.org/10.1152/jn.00231.2015 pmid:26084903
    OpenUrlCrossRefPubMed
  13. ↵
    1. Crowder EA,
    2. Olson CR
    (2015) Macaque monkeys experience visual crowding. J Vis 15:14. https://doi.org/10.1167/15.5.14 pmid:26067532
    OpenUrlAbstract/FREE Full Text
  14. ↵
    1. De Sousa C,
    2. Gaillard C,
    3. Di Bello F,
    4. Ben Hadj Hassen S,
    5. Ben Hamed S
    (2021) Behavioral validation of novel high resolution attention decoding method from multi-units & local field potentials. Neuroimage 231:117853. https://doi.org/10.1016/j.neuroimage.2021.117853
    OpenUrlCrossRefPubMed
  15. ↵
    1. Dickey AS,
    2. Suminski A,
    3. Amit Y,
    4. Hatsopoulos NG
    (2009) Single-unit stability using chronically implanted multielectrode arrays. J Neurophysiol 102:1331–1339. https://doi.org/10.1152/jn.90920.2008 pmid:19535480
    OpenUrlCrossRefPubMed
  16. ↵
    1. Donnelly NA, et al.
    (2014) Oscillatory activity in the medial prefrontal cortex and nucleus accumbens correlates with impulsivity and reward outcome. PLoS One 9:e111300. https://doi.org/10.1371/journal.pone.0111300 pmid:25333512
    OpenUrlCrossRefPubMed
  17. ↵
    1. Einevoll GT,
    2. Kayser C,
    3. Logothetis NK,
    4. Panzeri S
    (2013) Modelling and analysis of local field potentials for studying the function of cortical circuits. Nat Rev Neurosci 14:770–785. https://doi.org/10.1038/nrn3599
    OpenUrlCrossRefPubMed
  18. ↵
    1. Falaki A,
    2. Quessy S,
    3. Dancause N
    (2024) Differential modulation of local field potentials in the primary and premotor cortices during ipsilateral and contralateral reach to grasp in macaque monkeys. J Neurosci 44:e1161232024. https://doi.org/10.1523/JNEUROSCI.1161-23.2024 pmid:38589229
    OpenUrlAbstract/FREE Full Text
  19. ↵
    1. Goense JBM,
    2. Logothetis NK
    (2008) Neurophysiology of the BOLD fMRI signal in awake monkeys. Curr Biol 18:631–640. https://doi.org/10.1016/j.cub.2008.03.054
    OpenUrlCrossRefPubMed
  20. ↵
    1. Grabenhorst F,
    2. Ponce-Alvarez A,
    3. Battaglia-Mayer A,
    4. Deco G,
    5. Schultz W
    (2023) A view-based decision mechanism for rewards in the primate amygdala. Neuron 111:3871–3884.e14. https://doi.org/10.1016/j.neuron.2023.08.024 pmid:37725980
    OpenUrlCrossRefPubMed
  21. ↵
    1. Guyon N,
    2. Zacharias LR,
    3. de Oliveira EF,
    4. Kim H,
    5. Leite JP,
    6. Lopes-Aguiar C,
    7. Carlén M
    (2021) Network asynchrony underlying increased broadband gamma power. J Neurosci 41:2944–2963. https://doi.org/10.1523/JNEUROSCI.2250-20.2021 pmid:33593859
    OpenUrlAbstract/FREE Full Text
  22. ↵
    1. Hastie T,
    2. Tibshirani R,
    3. Wainwright M
    (2015) Statistical learning with sparsity: the lasso and generalizations. New York: Chapman and Hall/CRC. https://doi.org/10.1201/b18401
  23. ↵
    1. Hunt LT,
    2. Malalasekera WMN,
    3. de Berker AO,
    4. Miranda B,
    5. Farmer SF,
    6. Behrens TEJ,
    7. Kennerley SW
    (2018) Triple dissociation of attention and decision computations across prefrontal cortex. Nat Neurosci 21:1471–1481. https://doi.org/10.1038/s41593-018-0239-5 pmid:30258238
    OpenUrlCrossRefPubMed
  24. ↵
    1. Hwang EJ,
    2. Andersen RA
    (2013) The utility of multichannel local field potentials for brain–machine interfaces. J Neural Eng 10:046005. https://doi.org/10.1088/1741-2560/10/4/046005 pmid:23744624
    OpenUrlCrossRefPubMed
  25. ↵
    1. Kennerley SW,
    2. Behrens TEJ,
    3. Wallis JD
    (2011) Double dissociation of value computations in orbitofrontal and anterior cingulate neurons. Nat Neurosci 14:1581–1589. https://doi.org/10.1038/nn.2961 pmid:22037498
    OpenUrlCrossRefPubMed
  26. ↵
    1. Kleiner M,
    2. Brainard D,
    3. Pelli D
    (2007) What’s new in Psychtoolbox-3? https://pure.mpg.de/rest/items/item_1790332/component/file_3136265/content
  27. ↵
    1. Knudsen EB,
    2. Wallis JD
    (2020) Closed-loop theta stimulation in the orbitofrontal cortex prevents reward-based learning. Neuron 106:537–547. https://doi.org/10.1016/j.neuron.2020.02.003 pmid:32160515
    OpenUrlCrossRefPubMed
  28. ↵
    1. Kohn A,
    2. Coen-Cagli R,
    3. Kanitscheider I,
    4. Pouget A
    (2016) Correlations and neuronal population information. Annu Rev Neurosci 39:237–256. https://doi.org/10.1146/annurev-neuro-070815-013851 pmid:27145916
    OpenUrlCrossRefPubMed
  29. ↵
    1. Lebreton M,
    2. Jorge S,
    3. Michel V,
    4. Thirion B,
    5. Pessiglione M
    (2009) An automatic valuation system in the human brain: evidence from functional neuroimaging. Neuron 64:431–439. https://doi.org/10.1016/j.neuron.2009.09.040
    OpenUrlCrossRefPubMed
  30. ↵
    1. Leszczyński M,
    2. Barczak A,
    3. Kajikawa Y,
    4. Ulbert I,
    5. Falchier AY,
    6. Tal I,
    7. Haegens S,
    8. Melloni L,
    9. Knight RT,
    10. Schroeder CE
    (2020) Dissociation of broadband high-frequency activity and neuronal firing in the neocortex. Sci Adv 6:eabb0977. https://doi.org/10.1126/sciadv.abb0977 pmid:32851172
    OpenUrlFREE Full Text
  31. ↵
    1. Lindén H,
    2. Tetzlaff T,
    3. Potjans TC,
    4. Pettersen KH,
    5. Grün S,
    6. Diesmann M,
    7. Einevoll GT
    (2011) Modeling the spatial reach of the LFP. Neuron 72:859–872. https://doi.org/10.1016/j.neuron.2011.11.006
    OpenUrlCrossRefPubMed
  32. ↵
    1. Llorca A,
    2. Deogracias R
    (2022) Origin, development, and synaptogenesis of cortical interneurons. Front Neurosci 16:929469. https://doi.org/10.3389/fnins.2022.929469 pmid:35833090
    OpenUrlCrossRefPubMed
  33. ↵
    1. Lopez-Persem A, et al.
    (2020) Four core properties of the human brain valuation system demonstrated in intracranial signals. Nat Neurosci 23:664–675. https://doi.org/10.1038/s41593-020-0615-9
    OpenUrlCrossRefPubMed
  34. ↵
    1. Lundqvist M,
    2. Brincat SL,
    3. Rose J,
    4. Warden MR,
    5. Buschman TJ,
    6. Miller EK,
    7. Herman P
    (2023) Working memory control dynamics follow principles of spatial computing. Nat Commun 14:1429. https://doi.org/10.1038/s41467-023-36555-4 pmid:36918567
    OpenUrlCrossRefPubMed
  35. ↵
    1. Lundqvist M,
    2. Herman P,
    3. Warden MR,
    4. Brincat SL,
    5. Miller EK
    (2018) Gamma and beta bursts during working memory readout suggest roles in its volitional control. Nat Commun 9:394. https://doi.org/10.1038/s41467-017-02791-8 pmid:29374153
    OpenUrlCrossRefPubMed
  36. ↵
    1. Lundqvist M,
    2. Rose J,
    3. Herman P,
    4. Brincat SL,
    5. Buschman TJ,
    6. Miller EK
    (2016) Gamma and beta bursts underlie working memory. Neuron 90:152–164. https://doi.org/10.1016/j.neuron.2016.02.028 pmid:26996084
    OpenUrlCrossRefPubMed
  37. ↵
    1. Lupkin SM,
    2. McGinty VB
    (2023) Monkeys exhibit human-like gaze biases in economic decisions. eLife 12:e78205. https://doi.org/10.7554/eLife.78205 pmid:37497784
    OpenUrlCrossRefPubMed
  38. ↵
    1. Magri C,
    2. Schridde U,
    3. Murayama Y,
    4. Panzeri S,
    5. Logothetis NK
    (2012) The amplitude and timing of the BOLD signal reflects the relationship between local field potential power at different frequencies. J Neurosci 32:1395–1407. https://doi.org/10.1523/JNEUROSCI.3985-11.2012 pmid:22279224
    OpenUrlAbstract/FREE Full Text
  39. ↵
    1. Maisson DJ-N,
    2. Cash-Padgett TV,
    3. Wang MZ,
    4. Hayden BY,
    5. Heilbronner SR,
    6. Zimmermann J
    (2021) Choice-relevant information transformation along a ventrodorsal axis in the medial prefrontal cortex. Nat Commun 12:4830. https://doi.org/10.1038/s41467-021-25219-w pmid:34376663
    OpenUrlCrossRefPubMed
  40. ↵
    1. McGinty VB,
    2. Lupkin SM
    (2023) Behavioral read-out from population value signals in primate orbitofrontal cortex. Nat Neurosci 26:2203–2212. https://doi.org/10.1038/s41593-023-01473-7 pmid:37932464
    OpenUrlCrossRefPubMed
  41. ↵
    1. McGinty VB,
    2. Rangel A,
    3. Newsome WT
    (2016) Orbitofrontal cortex value signals depend on fixation location during free viewing. Neuron 90:1299–1311. https://doi.org/10.1016/j.neuron.2016.04.045 pmid:27263972
    OpenUrlCrossRefPubMed
  42. ↵
    1. Mendoza-Halliday D, et al.
    (2024) A ubiquitous spectrolaminar motif of local field potential power across the primate cortex. Nat Neurosci 27:547–560. https://doi.org/10.1038/s41593-023-01554-7 pmid:38238431
    OpenUrlCrossRefPubMed
  43. ↵
    1. Mitra P,
    2. Bokil H
    (2008) Observed brain dynamics. New York: Oxford University Press.
  44. ↵
    1. Mitzdorf U
    (1985) Current source-density method and application in cat cerebral cortex: investigation of evoked potentials and EEG phenomena. Physiol Rev 65:37–100. https://doi.org/10.1152/physrev.1985.65.1.37
    OpenUrlCrossRefPubMed
  45. ↵
    1. Murray EA,
    2. Moylan EJ,
    3. Saleem KS,
    4. Basile BM,
    5. Turchi J
    (2015) Specialized areas for value updating and goal selection in the primate orbitofrontal cortex. eLife 4:e11695. https://doi.org/10.7554/eLife.11695 pmid:26673891
    OpenUrlCrossRefPubMed
  46. ↵
    1. Nir Y,
    2. Fisch L,
    3. Mukamel R,
    4. Gelbard-Sagiv H,
    5. Arieli A,
    6. Fried I,
    7. Malach R
    (2007) Coupling between neuronal firing rate, gamma LFP, and BOLD fMRI is related to interneuronal correlations. Curr Biol 17:1275–1285. https://doi.org/10.1016/j.cub.2007.06.066
    OpenUrlCrossRefPubMed
  47. ↵
    1. Öngür D,
    2. Price JL
    (2000) The organization of networks within the orbital and medial prefrontal cortex of rats, monkeys and humans. Cereb Cortex 10:206–219. https://doi.org/10.1093/cercor/10.3.206
    OpenUrlCrossRefPubMed
  48. ↵
    1. Padoa-Schioppa C
    (2011) Neurobiology of economic choice: a good-based model. Annu Rev Neurosci 34:333–359. https://doi.org/10.1146/annurev-neuro-061010-113648 pmid:21456961
    OpenUrlCrossRefPubMed
  49. ↵
    1. Padoa-Schioppa C,
    2. Assad JA
    (2006) Neurons in the orbitofrontal cortex encode economic value. Nature 441:Article 7090. https://doi.org/10.1038/nature04676 pmid:16633341
    OpenUrlPubMed
  50. ↵
    1. Perge JA,
    2. Zhang S,
    3. Malik WQ,
    4. Homer ML,
    5. Cash S,
    6. Friehs G,
    7. Eskandar EN,
    8. Donoghue JP,
    9. Hochberg LR
    (2014) Reliability of directional information in unsorted spikes and local field potentials recorded in human motor cortex. J Neural Eng 11:046007. https://doi.org/10.1088/1741-2560/11/4/046007 pmid:24921388
    OpenUrlCrossRefPubMed
  51. ↵
    1. Prakash SS,
    2. Das A,
    3. Kanth ST,
    4. Mayo JP,
    5. Ray S
    (2021) Decoding of attentional state using high-frequency local field potential is as accurate as using spikes. Cereb Cortex 31:Article 9. https://doi.org/10.1093/cercor/bhab088 pmid:33866366
    OpenUrlPubMed
  52. ↵
    1. Prakash SS,
    2. Mayo JP,
    3. Ray S
    (2022) Decoding of attentional state using local field potentials. Curr Opin Neurobiol 76:102589. https://doi.org/10.1016/j.conb.2022.102589 pmid:35751949
    OpenUrlCrossRefPubMed
  53. ↵
    1. Proix T, et al.
    (2022) Imagined speech can be decoded from low- and cross-frequency intracranial EEG features. Nat Commun 13:48. https://doi.org/10.1038/s41467-021-27725-3 pmid:35013268
    OpenUrlCrossRefPubMed
  54. ↵
    1. Rasch MJ,
    2. Gretton A,
    3. Murayama Y,
    4. Maass W,
    5. Logothetis NK
    (2008) Inferring spike trains from local field potentials. J Neurophysiol 99:1461–1476. https://doi.org/10.1152/jn.00919.2007
    OpenUrlCrossRefPubMed
  55. ↵
    1. Ray S,
    2. Crone NE,
    3. Niebur E,
    4. Franaszczuk PJ,
    5. Hsiao SS
    (2008) Neural correlates of high-gamma oscillations (60–200 Hz) in macaque local field potentials and their potential implications in electrocorticography. J Neurosci 28:11526–11536. https://doi.org/10.1523/JNEUROSCI.2848-08.2008 pmid:18987189
    OpenUrlAbstract/FREE Full Text
  56. ↵
    1. Ray S,
    2. Maunsell JHR
    (2011) Different origins of gamma rhythm and high-gamma activity in macaque visual cortex. PLoS Biol 9:Article 4. https://doi.org/10.1371/journal.pbio.1000610 pmid:21532743
    OpenUrlPubMed
  57. ↵
    1. Reimann MW,
    2. Anastassiou CA,
    3. Perin R,
    4. Hill SL,
    5. Markram H,
    6. Koch C
    (2013) A biophysically detailed model of neocortical local field potentials predicts the critical role of active membrane currents. Neuron 79:375–390. https://doi.org/10.1016/j.neuron.2013.05.023 pmid:23889937
    OpenUrlCrossRefPubMed
  58. ↵
    1. Rich EL,
    2. Wallis JD
    (2016) Decoding subjective decisions from orbitofrontal cortex. Nat Neurosci 19:973–980. https://doi.org/10.1038/nn.4320 pmid:27273768
    OpenUrlCrossRefPubMed
  59. ↵
    1. Rich EL,
    2. Wallis JD
    (2017) Spatiotemporal dynamics of information encoding revealed in orbitofrontal high-gamma. Nat Commun 8:Article 1. https://doi.org/10.1038/s41467-017-01253-5 pmid:29074960
    OpenUrlCrossRefPubMed
  60. ↵
    1. Rustichini A,
    2. Padoa-Schioppa C
    (2015) A neuro-computational model of economic decisions. J Neurophysiol 114:1382–1398. https://doi.org/10.1152/jn.00184.2015 pmid:26063776
    OpenUrlCrossRefPubMed
  61. ↵
    1. Saez I,
    2. Lin J,
    3. Stolk A,
    4. Chang E,
    5. Parvizi J,
    6. Schalk G,
    7. Knight RT,
    8. Hsu M
    (2018) Encoding of multiple reward-related computations in transient and sustained high-frequency activity in human OFC. Curr Biol 28:2889–2899.e3. https://doi.org/10.1016/j.cub.2018.07.045 pmid:30220499
    OpenUrlCrossRefPubMed
  62. ↵
    1. Saleem KS,
    2. Logothetis NK
    (2012) A combined MRI and histology atlas of the rhesus monkey brain in stereotaxic coordinates. London: Academic Press.
  63. ↵
    1. Shih W-Y,
    2. Yu H-Y,
    3. Lee C-C,
    4. Chou C-C,
    5. Chen C,
    6. Glimcher PW,
    7. Wu S-W
    (2023) Electrophysiological population dynamics reveal context dependencies during decision making in human frontal cortex. Nat Commun 14:7821. https://doi.org/10.1038/s41467-023-42092-x pmid:38016973
    OpenUrlCrossRefPubMed
  64. ↵
    1. Stavisky SD,
    2. Kao JC,
    3. Nuyujukian P,
    4. Ryu SI,
    5. Shenoy KV
    (2015) A high performing brain–machine interface driven by low-frequency local field potentials alone and together with spikes. J Neural Eng 12:036009. https://doi.org/10.1088/1741-2560/12/3/036009 pmid:25946198
    OpenUrlCrossRefPubMed
  65. ↵
    1. Strait CE,
    2. Blanchard TC,
    3. Hayden BY
    (2014) Reward value comparison via mutual inhibition in ventromedial prefrontal cortex. Neuron 82:1357–1366. https://doi.org/10.1016/j.neuron.2014.04.032 pmid:24881835
    OpenUrlCrossRefPubMed
  66. ↵
    1. Supèr H,
    2. Roelfsema PR
    (2005) Chronic multiunit recordings in behaving animals: advantages and limitations. Prog Brain Res 147:263–282. https://doi.org/10.1016/S0079-6123(04)47020-4
    OpenUrlCrossRefPubMed
  67. ↵
    1. Thorpe SJ,
    2. Rolls ET,
    3. Maddison S
    (1983) The orbitofrontal cortex: neuronal activity in the behaving monkey. Exp Brain Res 49:93–115. https://doi.org/10.1007/BF00235545
    OpenUrlCrossRefPubMed
  68. ↵
    1. Tremblay S,
    2. Doucet G,
    3. Pieper F,
    4. Sachs A,
    5. Martinez-Trujillo J
    (2015) Single-trial decoding of visual attention from local field potentials in the primate lateral prefrontal cortex is frequency-dependent. J Neurosci 35:9038–9049. https://doi.org/10.1523/JNEUROSCI.1041-15.2015 pmid:26085629
    OpenUrlAbstract/FREE Full Text
  69. ↵
    1. Tremblay L,
    2. Schultz W
    (1999) Relative reward preference in primate orbitofrontal cortex. Nature 398:704–708. https://doi.org/10.1038/19525
    OpenUrlCrossRefPubMed
  70. ↵
    1. Vaidya AR,
    2. Sefranek M,
    3. Fellows LK
    (2018) Ventromedial frontal lobe damage alters how specific attributes are weighed in subjective valuation. Cereb Cortex 28:3857–3867. https://doi.org/10.1093/cercor/bhx246 pmid:29069371
    OpenUrlCrossRefPubMed
  71. ↵
    1. Veit J,
    2. Hakim R,
    3. Jadi MP,
    4. Sejnowski TJ,
    5. Adesnik H
    (2017) Cortical gamma band synchronization through somatostatin interneurons. Nat Neurosci 20:951–959. https://doi.org/10.1038/nn.4562 pmid:28481348
    OpenUrlCrossRefPubMed
  72. ↵
    1. Verret L, et al.
    (2012) Inhibitory interneuron deficit links altered network activity and cognitive dysfunction in Alzheimer model. Cell 149:708–721. https://doi.org/10.1016/j.cell.2012.02.046 pmid:22541439
    OpenUrlCrossRefPubMed
  73. ↵
    1. Wagatsuma N,
    2. Nobukawa S,
    3. Fukai T
    (2023) A microcircuit model involving parvalbumin, somatostatin, and vasoactive intestinal polypeptide inhibitory interneurons for the modulation of neuronal oscillation during visual processing. Cereb Cortex 33:4459–4477. https://doi.org/10.1093/cercor/bhac355 pmid:36130096
    OpenUrlCrossRefPubMed
  74. ↵
    1. Wang X-J
    (2002) Probabilistic decision making by slow reverberation in cortical circuits. Neuron 36:Article 5. https://doi.org/10.1016/S0896-6273(02)01092-9
    OpenUrlCrossRef
  75. ↵
    1. Whitney D,
    2. Levi DM
    (2011) Visual crowding: a fundamental limit on conscious perception and object recognition. Trends Cogn Sci 15:160–168. https://doi.org/10.1016/j.tics.2011.02.005 pmid:21420894
    OpenUrlCrossRefPubMed
  76. ↵
    1. Wilson GH,
    2. Stavisky SD,
    3. Willett FR,
    4. Avansino DT,
    5. Kelemen JN,
    6. Hochberg LR,
    7. Henderson JM,
    8. Druckmann S,
    9. Shenoy KV
    (2020) Decoding spoken English from intracortical electrode arrays in dorsal precentral gyrus. J Neural Eng 17:066007. https://doi.org/10.1088/1741-2552/abbfef pmid:33236720
    OpenUrlCrossRefPubMed
  77. ↵
    1. Yates JL,
    2. Katz LN,
    3. Levi AJ,
    4. Pillow JW,
    5. Huk AC
    (2020) A simple linear readout of MT supports motion direction-discrimination performance. J Neurophysiol 123:682–694. https://doi.org/10.1152/jn.00117.2019 pmid:31852399
    OpenUrlCrossRefPubMed
  78. ↵
    1. Zhang, W.,
    2. Xie, Y., &
    3. Yang, T.
    (2022). Reward salience but not spatial attention dominates the value representation in the orbitofrontal cortex. Nat Commun, 13:6306. https://doi.org/10.1038/s41467-022-34084-0 pmid:36273229
    OpenUrlCrossRefPubMed
Back to top

In this issue

The Journal of Neuroscience: 45 (20)
Journal of Neuroscience
Vol. 45, Issue 20
14 May 2025
  • Table of Contents
  • About the Cover
  • Index by author
  • Masthead (PDF)
Email

Thank you for sharing this Journal of Neuroscience article.

NOTE: We request your email address only to inform the recipient that it was you who recommended this article, and that it is not junk mail. We do not retain these email addresses.

Enter multiple addresses on separate lines or separate them with commas.
Orbitofrontal High-Gamma Reflects Spike-Dissociable Value and Decision Mechanisms
(Your Name) has forwarded a page to you from Journal of Neuroscience
(Your Name) thought you would be interested in this article in Journal of Neuroscience.
CAPTCHA
This question is for testing whether or not you are a human visitor and to prevent automated spam submissions.
Print
View Full Page PDF
Citation Tools
Orbitofrontal High-Gamma Reflects Spike-Dissociable Value and Decision Mechanisms
Dixit Sharma, Shira M. Lupkin, Vincent B. McGinty
Journal of Neuroscience 14 May 2025, 45 (20) e0789242025; DOI: 10.1523/JNEUROSCI.0789-24.2025

Citation Manager Formats

  • BibTeX
  • Bookends
  • EasyBib
  • EndNote (tagged)
  • EndNote 8 (xml)
  • Medlars
  • Mendeley
  • Papers
  • RefWorks Tagged
  • Ref Manager
  • RIS
  • Zotero
Respond to this article
Request Permissions
Share
Orbitofrontal High-Gamma Reflects Spike-Dissociable Value and Decision Mechanisms
Dixit Sharma, Shira M. Lupkin, Vincent B. McGinty
Journal of Neuroscience 14 May 2025, 45 (20) e0789242025; DOI: 10.1523/JNEUROSCI.0789-24.2025
Twitter logo Facebook logo Mendeley logo
  • Tweet Widget
  • Facebook Like
  • Google Plus One

Jump to section

  • Article
    • Abstract
    • Significance Statement
    • Introduction
    • Materials and Methods
    • Results
    • Discussion
    • Footnotes
    • References
  • Figures & Data
  • Info & Metrics
  • eLetters
  • PDF

Keywords

  • decision-making
  • local field potentials
  • neuroeconomics
  • nonhuman primates
  • prefrontal cortex

Responses to this article

Respond to this article

Jump to comment:

No eLetters have been published for this article.

Related Articles

Cited By...

More in this TOC Section

Research Articles

  • Increased neuronal expression of the early endosomal adaptor APPL1 replicates Alzheimer’s Disease-related endosomal and synaptic dysfunction with cholinergic neurodegeneration.
  • Presynaptic mu opioid receptors suppress the functional connectivity of ventral tegmental area dopaminergic neurons with aversion-related brain regions
  • Change of spiny neuron structure in the basal ganglia song circuit and its regulation by miR-9 during song development
Show more Research Articles

Behavioral/Cognitive

  • Increased neuronal expression of the early endosomal adaptor APPL1 replicates Alzheimer’s Disease-related endosomal and synaptic dysfunction with cholinergic neurodegeneration.
  • Presynaptic mu opioid receptors suppress the functional connectivity of ventral tegmental area dopaminergic neurons with aversion-related brain regions
  • Change of spiny neuron structure in the basal ganglia song circuit and its regulation by miR-9 during song development
Show more Behavioral/Cognitive
  • Home
  • Alerts
  • Follow SFN on BlueSky
  • Visit Society for Neuroscience on Facebook
  • Follow Society for Neuroscience on Twitter
  • Follow Society for Neuroscience on LinkedIn
  • Visit Society for Neuroscience on Youtube
  • Follow our RSS feeds

Content

  • Early Release
  • Current Issue
  • Issue Archive
  • Collections

Information

  • For Authors
  • For Advertisers
  • For the Media
  • For Subscribers

About

  • About the Journal
  • Editorial Board
  • Privacy Notice
  • Contact
  • Accessibility
(JNeurosci logo)
(SfN logo)

Copyright © 2025 by the Society for Neuroscience.
JNeurosci Online ISSN: 1529-2401

The ideas and opinions expressed in JNeurosci do not necessarily reflect those of SfN or the JNeurosci Editorial Board. Publication of an advertisement or other product mention in JNeurosci should not be construed as an endorsement of the manufacturer’s claims. SfN does not assume any responsibility for any injury and/or damage to persons or property arising from or related to any use of any material contained in JNeurosci.