Behavioral responses before and after inactivation

We recorded neural activity from cortical area MT in two macaque monkeys that were trained to perform a motion discrimination task (Fig 1A) with random dot kinematograms (RDKs) [4,24]. In separate experiments, we recorded from area V4 [25] in two other monkeys trained to perform a form discrimination task (Fig 1B). These tasks were chosen to best match the functional selectivity of the recorded areas.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 1. Behavioral performance before and after cortical inactivation in MT and V4.

(A) Two rhesus macaques performed a random dot motion discrimination task. The stimulus was positioned in the receptive field (RF, pink circle) of recorded MT neurons. Animals reported the perceived motion direction by making a saccade to one of two choice targets. Task difficulty was varied by adjusting motion coherence, defined as the percentage of dots moving coherently in one direction. (B) Two other macaques performed a form discrimination task using circular or radial gratings placed in the RF of recorded V4 neurons. Animals reported which of the two targets matched the stimulus. Task difficulty was varied by adjusting the signal-to-noise ratio (SNR) of the form stimulus. (A–B) Pink elements and arrows are for illustration only and were not shown to the animals. The colored shades mark the epochs used to compute choice probability (CP) (blue: baseline; yellow: stimulus-response; green: delay). This period excludes the first 100 ms of fixation and the final 100–400 ms preceding the Choice phase, thereby isolating stimulus-driven neural activity without contamination from sensorimotor signals. (C–F) Psychometric curves for individual monkeys before (black) and after (blue) muscimol inactivation of local spiking activity pooled across experimental sessions: (C) Monkey Y and (D) Monkey C from the MT task; (E) Monkey L and (F) Monkey A from the V4 task. Curves show model-based Weibull fits with 95% confidence intervals. Small semi-transparent dots indicate correct percentages computed separately for each session at each stimulus level. Inactivation of MT or V4 caused a rightward shift in the psychometric functions, reflecting reduced sensitivity and confirming the causal contribution of these areas to their respective tasks. (G) Behavioral sensitivity (1/threshold) for each session and monkey. The underlying numerical data are provided at https://doi.org/10.5281/zenodo.20583873. Gray lines link paired sessions recorded on the same day. Each semi-transparent dot represents the sensitivity estimate obtained from a single session-level psychometric fit. Bars represent mean ± standard error of the mean (SEM) across sessions. “n.s.” denotes no significant differences that were observed between animals within either area MT or V4, indicating comparable behavioral performance across subjects.

https://doi.org/10.1371/journal.pbio.3003873.g001

During each trial of both tasks (Fig 1A and 1B), the monkeys fixated a central point on the screen while a motion or form stimulus appeared within the receptive fields (RFs) of concurrently recorded neurons. Following a fixed delay, they reported their perceptual decision by making a saccade to one of two choice targets. In the MT task, these targets were positioned along the motion axis, and the monkeys were rewarded for choosing the direction of the RDK’s coherent motion. In the V4 task, one of the two targets matched the presented form, and monkeys were rewarded for selecting the matching target. Task difficulty was titrated on each trial by randomly varying the signal-to-noise (SNR) ratio of each stimulus [4,25].

To assess the causal contributions of these areas to perception, we pharmacologically inactivated the corresponding cortical site using muscimol, a GABA_A receptor agonist that silences local spiking activity [24,26]. Crucially, all included sessions (N = 14) contained both pre- and post-inactivation blocks recorded on the same day, ensuring a concurrent measure of neural and behavioral changes. Figure 1C–1F shows the behavioral results for each animal before and after inactivation pooled across all experimental sessions. As expected, inactivation consistently shifted psychometric functions to the right, indicating reduced performance and confirming a causal role of these areas in their respective tasks. For the most difficult stimulus condition (2% motion coherence for MT and 3% SNR for V4), performance approached chance levels (55%–60% correct).

For each task, we quantified session-wise behavioral performance by fitting the Weibull function to generate psychometric curves (see Methods for details). Figure 1G shows the sensitivity values for each monkey before and after inactivation, illustrating the consistency of behavioral performance across animals performing the same task. Specifically, sensitivity did not differ significantly between monkeys in either area MT (monkey C: 0.034 ± 0.014 versus monkey Y: 0.034 ± 0.003; p = 0.492, Wilcoxon rank-sum (WRS) test) or area V4 (monkey A: 0.090 ± 0.021 versus monkey L: 0.096 ± 0.014; p = 0.589, WRS).

CP as a function of LFP frequency and behavioral epoch

For both experiments, we defined the preferred stimulus for each recording site based on multi-unit spike counts (Fig 2B), which exhibit greater selectivity than LFP signals [17,27] (see Methods for details). We then computed CP values from the distribution of LFP power across trials [16,17]. CP values above 0.5 indicate higher LFP power on trials in which behavioral choices aligned with the recorded neurons’ preferred stimulus; values below 0.5 indicate an inverse relationship between LFP power and preferred stimulus choices. Due to the nonnormal distribution of LFP power, we computed CP in the LFPs using robust scaling—normalization based on median and interquartile range—rather than the pooled z-scoring method that is commonly used for spike data. We validated this approach through simulations (see Methods and Supplementary Information).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 2. Measuring and computing LFP-based choice probability (CP).

(A) A 16-channel linear electrode array was used to simultaneously record multi-unit spiking activity and local field potentials (LFPs) from the same cortical site. Within each session, the GABA_A receptor agonist muscimol was injected locally to silence spiking activity while preserving LFP signals. Example single-unit spike rasters across task trials and LFP traces from area MT illustrate the loss of spiking activity and persistence of LFP responses after inactivation. (B) The preferred and anti-preferred stimulus conditions for each recording site were determined from multi-unit spike counts during the stimulus-response period prior to muscimol injection (pink shades). The stimulus was either the neuron’s preferred or anti-preferred motion direction (MT) or form (V4). (C) Example time-frequency representations of LFP power before and after inactivation in area MT. Colors indicate power (in dB) normalized to the pre-stimulus baseline across frequencies from 0 to 150 Hz. LFP power decreased overall after inactivation, with a consistent change across frequencies.

https://doi.org/10.1371/journal.pbio.3003873.g002

In each session, we computed LFP-based CP before and during muscimol inactivation of spiking activity. Although muscimol consistently reduced spiking rates to zero [24], LFP activity persisted (Fig 2A), suggesting that it contained information about input to the MT and V4 recording sites [27]. A polarity analysis further shows that the evoked peak sign reversal in the lower panel of Fig 2A only occurred in a subset of channels, whereas response magnitude was consistently reduced (S6 Fig). Indeed, following muscimol inactivation, LFP power in response to the visual stimuli decreased by an average of 76.0%, with no consistent difference across frequencies (Fig 2C, Spearman’s ρ = 0.067, permutation test, p = 0.420).

Because neural-behavioral correlations have been shown sensitive to task difficulty [28,29], we equated behavioral performance between pre- and post-inactivation sessions by matching stimulus difficulty levels (see Methods for details). We then computed LFP-based CP across three behavioral epochs: the baseline period (−200–0 ms, pre-stimulus fixation), the stimulus-response period (50–250 ms after stimulus onset), and the delay period (250–400 ms, prior to the presentation of two choices). Among these, the stimulus-response epoch is the most informative for interpreting CP, as this is when sensory evidence first reaches the recorded cortical sites and before the motor response is initiated. We considered CP in three canonical LFP frequency bands commonly reported in prior literature: alpha-beta [5–30 Hz 17,30,31], low-gamma [30–70 Hz 32–34], and high-gamma [70–150 Hz 30,35]). Analysis of the stimulus-evoked LFP power spectra confirmed a largely broadband spectral structure before and after inactivation, with no change in narrowband oscillatory peaks (S7 Fig).

High-gamma CP is consistent with a feedforward influence

For high-gamma oscillations (70–150 Hz), CP peaked in the stimulus-response epoch, and the average CP value was significantly above chance before inactivation in both areas MT and V4 (Fig 3A). In MT, the mean CP was 0.509 ± 0.001 (mean CP ± SEM, p < 0.001, two-sided Wilcoxon signed-rank (WSR) test against 0.5), and in V4, 0.532 ± 0.006 (p < 0.001, WSR test). Thus, consistent with previous studies of spiking activity [4,36], there was a small but consistent relationship between trial-to-trial LFP fluctuations and perceptual decisions.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 3. Frequency- and epoch-specific choice probability (CP) before and after cortical inactivation.

(A–C) LFP-based CP traces (mean ± SEM) are shown for each frequency band across time from stimulus onset. CP was computed in 200 ms sliding windows advanced in a 20 ms step size. Each row represents one frequency band: (A) high-gamma (70–150 Hz), (B) low-gamma (30–70 Hz), and (C) alpha-beta (5–30 Hz). Within each row, panels display area-wise averages for area MT (left) and area V4 (middle), followed by the aggregated data across all four monkeys (right). Black and blue traces denote pre- and post-inactivation sessions, respectively. The pink shaded region indicates the stimulus-response epoch (50–250 ms after stimulus onset), during which statistical comparisons were conducted. Significance markers reflect results from Wilcoxon rank-sum (WRS) tests comparing pre- and post-inactivation CP values within this epoch (p < 0.05, * p < 0.01, ** p < 0.001, ***, “n.s.” denotes nonsignificant difference). Shaded regions in the left and middle panels indicate the SEM of time-resolved CP across sessions at each time point; smaller SEM values reported in the right panel and text reflect epoch-averaged CP collapsed across time. In the aggregated data, during the stimulus-response epoch, (A) high-gamma CP before inactivation was again significantly above chance (CP = 0.521 ± 0.005, p < 0.001, WSR test) but decreased to chance levels after inactivation (CP = 0.501 ± 0.001, p = 0.237, WSR test), with a significant reduction between drug conditions (p < 0.001, WRS test). (B) Low-gamma CP before inactivation was 0.526 ± 0.009 (p < 0.001, WSR) and 0.521 ± 0.006 (p < 0.001, WSR) after inactivation, with no reliable difference between drug conditions (p = 0.622, WRS). (C) Alpha-beta CP showed heterogeneous effects across areas and individuals. The underlying numerical data are provided at https://doi.org/10.5281/zenodo.20583873.

https://doi.org/10.1371/journal.pbio.3003873.g003

With inactivation and at an equivalent behavioral performance level, average CP values decreased significantly following inactivation (two-sided WRS test to compare before and after inactivation; area MT: p < 0.001; area V4: p < 0.001) and were no longer reliably different from chance, despite a marginal deviation in V4 (area MT: 0.499 ± 0.002, p = 0.342; area V4: 0.503 ± 0.001, p = 0.032; WSR test), which did not survive correction for multiple comparisons and was negligible in magnitude. These findings were consistent across individual monkeys (S1 Fig). In contrast, inactivation had no effect on CP measured during the baseline (p = 1.000, WRS) or the delay (p = 0.597, WRS) epochs (Fig 3A). These results were further confirmed by a linear mixed-effects ANOVA (S1Table), in which the epoch x inactivation interaction was marginally significant in the high-gamma band (p = 0.049) with no other significant main or interaction effects (all other p > 0.05).

In comparison to the decrease in CP, inactivation had a smaller effect on stimulus coding. A linear decoder trained on to classify stimulus identity from the population of high-gamma LFP signals (see Methods) maintained comparable accuracy before and after inactivation in MT (pre: 0.566 ± 0.029; post; 0.519 ± 0.027; p = 0.156, session-wise paired WSR test; S5A Fig) and V4 (pre: 0.626 ± 0.053; post: 0.616 ± 0.039; p = 1.000, paired WSR test; S5B Fig). Thus, the disruption of high-gamma CP with muscimol inactivation was likely not entirely due to a loss of stimulus encoding, but rather to a block in the transmission of sensory signals from visual cortex.

These results indicate that decision-related high-gamma activity depends critically on local spiking output [22,37]. When spiking was silenced, the corresponding CP signal disappeared, consistent with the causal hypothesis that high-gamma LFP power reflects sensory signals contributing directly to perceptual decisions. This interpretation aligns with prior findings that high-gamma oscillations closely track local spiking rates and the feedforward propagation of sensory information to downstream cortical areas [21,31].

Low-gamma CP reflects correlated decision information

Low-gamma oscillations (30–70 Hz) also carried decision-related signals during the stimulus-response epoch (Fig 3B). Before inactivation, mean CP values were significantly above chance in both areas (MT: 0.517 ± 0.009, p < 0.001; V4: 0.535 ± 0.019, p < 0.001, WSR). Critically, these decision-related signals persisted after muscimol inactivation. Post-inactivation CP values remained significantly above chance (MT: 0.518 ± 0.011, p < 0.001; V4: 0.524 ± 0.009, p < 0.001, WSR), and no significant differences were observed between pre- and post-inactivation sessions (MT: p = 0.429; V4: p = 0.437; WRS test). This pattern was consistent across monkeys (S2 Fig), with all individuals showing low-gamma CP values that did not differ significantly before and after inactivation (Monkey Y, p = 0.984; Monkey C, p = 0.950; Monkey L, p = 0.192; Monkey A, p = 0.384; WRS). As with the high-gamma band, low-gamma CP values were at chance levels during the other trial epochs, with no change after inactivation (Fig 3B; baseline, pre-inactivation CP = 0.499 ± 0.010, p = 0.922, post-inactivation CP = 0.494 ± 0.007, p = 0.432; delay, pre CP = 0.507 ± 0.008, p = 0.275, post CP = 0.503 ± 0.007, p = 0.846, WSR). A mixed-effects ANOVA (S2 Table) confirmed these findings, showing a significant main effect of epoch (p < 0.05), but no inactivation or interaction effects (all other p > 0.05). Thus, despite the elimination of local spiking, the continued presence of low-gamma decision signals strongly suggests a component of CP that is not causally connected to behavior [38].

Spectral weighting on decomposition of LFP population signals

Previous work has suggested that, at the population-level, some decision-related signals are distinct from those that encode the relevant sensory stimuli [39]. This suggests a parcellation of population activity into stimulus-related and decision-related activity. Given the different roles for decision signals described in the previous sections, we wondered whether a similar distinction would be found in a population analysis of LFP signals.

For each session, we trained a linear classifier to discriminate stimulus identity, using the full LFP power spectrum (see Methods for details). The weight vector recovered by the decoder was taken as the stimulus axis. We then projected this axis out of the population activity and applied principal component analysis (PCA) to the residual matrix. The first principal component was taken as the leading null dimension orthogonal to the decoder-defined stimulus axis [39,40]. We compared the spectral composition of the stimulus axis with that of the null dimension in the same population activity space (Fig 4A).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 4. Frequency-specific decomposition of decision-related LFP activity.

(A) Schematic illustration of the population activity space. (B) Frequency band loading energy for the stimulus axis, shown separately for low-gamma and high-gamma before (gray) and after (blue) inactivation. (C) Frequency band loading energy for the leading null dimension, shown separately for low-gamma and high-gamma before (gray) and after (blue) inactivation. (B-C) Bars indicate mean ± SEM across sessions, each dot represents one session, filled circles indicate before inactivation, open circles indicate after inactivation, and gray lines connect matched sessions. (D–G) Baseline (−300 to 0 ms) CP in the alpha-beta range (5–30 Hz) for trials preceded by rewarded and nonrewarded outcomes. (D) Monkey Y and (E) Monkey C, recorded in area MT. (F) Monkey L and (G) Monkey A, recorded in area V4. Black and blue bars indicate before and after inactivation sessions, respectively. Semi-transparent dots overlaid on each bar represent individual session-level CP values computed within the specified time-frequency window (−300 to 0 ms relative to stimulus onset, 5-30 Hz). The underlying numerical data are provided at https://doi.org/10.5281/zenodo.20583873. Statistical significance markers reflect WRS tests comparing rewarded and nonrewarded trials aggregated across pre- and post-inactivation conditions within this epoch (p < 0.05, * p < 0.01, ** p < 0.001, ***); n.s. denotes nonsignificant differences between reward conditions. Error bars represent the SEM across sessions.

https://doi.org/10.1371/journal.pbio.3003873.g004

Across areas and subjects, the stimulus axis was consistently high-gamma weighted (Fig 4B). High-gamma loading energy on that axis was before inactivation and after inactivation, whereas low-gamma loading was substantially smaller (before: , after: ; one-sided WSR test, p < 0.001 in both conditions). By contrast, the null dimension showed a different spectral tendency (Fig 4C): low-gamma loading exceeded high-gamma loading both before and after inactivation (low- versus high-gamma; before: versus before inactivation; versus after inactivation), reaching significance after inactivation (post: , one-sided WSR test). This low-gamma weighting along the null dimension did not change significantly after inactivation (p = 0.787, two-sided WSR test).

Together, these results indicated that stimulus-aligned structure remained concentrated in high-gamma, whereas the persistent low-gamma signal was preferentially expressed in a dimension orthogonal to the stimulus axis. This pattern was consistent with prior work showing that the dominant activity patterns in a source population do not necessarily coincide with the most task-relevant subspace [41,42], and that choice-related signals can occupy dimensions misaligned with stimulus encoding [8,39].

Alpha-beta CP and reward history

Decision-related signals in the alpha-beta band (5–30 Hz) were heterogeneous across monkeys, showing no consistent pattern within areas (S3 Fig). In area MT, monkey C exhibited persistent below-chance CP both before and after inactivation (before: CP = 0.494 ± 0.003, p < 0.001, WSR; after: CP = 0.492 ± 0.008, p < 0.001, WSR; p = 0.217, WRS), whereas monkey Y showed a strong decline from above chance CP before inactivation (CP = 0.521 ± 0.018, p < 0.001, WSR) to below-chance CP after inactivation (CP = 0.493 ± 0.011, p < 0.05, WSR); In area V4, monkey A maintained consistently positive CP values before and after inactivation (before: CP = 0.543 ± 0.034, p < 0.001, WSR; after: CP = 0.532 ± 0.015, p < 0.001, WSR; p = 0.197, WRS), while monkey L showed a reversal from CP at chance before inactivation (CP = 0.506 ± 0.025, p = 0.179, WSR) to significantly below-chance after inactivation (CP = 0.491 ± 0.021, p < 0.001, WSR). A mixed-effects ANOVA (S3 Table) showed no significant main or interaction effects (all p > 0.05), indicating that, even after accounting for individual differences across monkeys, alpha-beta CP still did not vary significantly with epoch or inactivation condition. Such inconsistency implies that this frequency band might integrate multiple influences, such as internal states or task strategies [21,31,43,44].

One possibility raised by previous studies is that CP on one trial could be influenced by the reward (or lack thereof) on the previous trial; this could affect arousal, which in turn could modulate both neural responses and behavioral performance [7,45,46]. In that case, CP during the baseline period in the alpha and beta bands might capture this covariation between neural and behavioral signals, without necessarily being related to the decision per se. To assess this possibility and to further disambiguate among individuals, we separated CP for trials that were preceded by a successful (rewarded) trial from those that had been preceded by an unsuccessful (nonrewarded) trial (see Methods for details).

Across monkeys, we found that CP in the alpha-beta frequency range was often higher following a nonrewarded trial than a rewarded one, a pattern observed in three of the four monkeys (Fig 4D–4F). When averaged across both pre- and post-inactivation sessions, mean CP for nonrewarded trials exceeded that for rewarded trials in Monkey Y (nonrewarded: 0.533 ± 0.001; rewarded: 0.513 ± 0.001; p < 0.001, WRS), Monkey L (nonrewarded: 0.522 ± 0.003; rewarded: 0.487 ± 0.001; p < 0.001, WRS), and Monkey C (nonrewarded: 0.524 ± 0.001; rewarded: 0.485 ± 0.000; p < 0.001, WRS), whereas Monkey A showed no significant difference (Fig 4G, nonrewarded: 0.480 ± 0.004; rewarded: 0.475 ± 0.001; p = 0.731, WRS). This reward-related component of CP persisted after muscimol inactivation. Averaging across post-inactivation sessions demonstrated a consistent pattern across all four monkeys: higher CP for nonrewarded than rewarded trials (nonrewarded versus rewarded; Monkey Y: 0.534 ± 0.001 versus 0.515 ± 0.001; Monkey C: 0.532 ± 0.001 versus 0.484 ± 0.001; Monkey L: 0.533 ± 0.004 versus 0.488 ± 0.002; Monkey A: 0.521 ± 0.004 versus 0.471 ± 0.002; all p < 0.001, WRS). Together, these results suggest that the alpha-beta decision signals carry a component reflecting prior reward history, which becomes more dominant when sensory-driven encoding is suppressed by inactivation of local spiking activity.

CP latency across frequency bands

Intuitively, one might expect feedforward and feedback influences on decisions to appear at different time points in the neural response [6,47]. Indeed, feedforward and feedback inputs can occur with different latencies [48], though this is not always the case [14]. To address this issue, we quantified the latency of the emergence of CP in our data (see Methods). Before inactivation, the mean latencies were 100.7 ± 2.1 ms in high-gamma, 109.7 ± 3.6 ms in low-gamma, and 116.3 ± 4.1 ms in alpha-beta. However, paired comparisons across monkeys did not reveal reliable differences between bands (high-gamma versus low-gamma, p = 0.655; high-gamma versus alpha-beta, p = 0.179; low-gamma versus alpha-beta, p = 0.180, WSR tests). After inactivation, the mean change-point latencies were 112.1 ± 2.2 ms in high-gamma, 101.9 ± 3.0 ms in low-gamma, and 102.2 ± 4.4 ms in alpha-beta; again, none of the paired comparisons were statistically reliable (all p > 0.1). Thus, the relative temporal ordering across frequency bands was not significantly different across conditions.

Consistent with this result, the time of peak CP deviation was also similar across frequency bands. Mean peak times ranged from 144.2 to 154.1 ms before inactivation and from 138.0 to 150.1 ms after inactivation, and all paired comparisons were nonsignificant (all p > 0.1 before and after inactivation, WSR tests). Similarly, the magnitudes of peak CP were not statistically different across frequency bands before or after inactivation (all p > 0.1). Together, neither latency nor magnitude could reliably separate mixed sources of decision-related signals in our dataset.

Comparison with previous work

Previous work has used CP to investigate the neural bases of decision-making across various brain regions, including the parietal cortex [13,29,49] and early visual areas [43,50,51], using behavioral paradigms such as visual discrimination [3,4,52] and change-detection tasks [53,54]. Recent work in this area has hypothesized different time courses for feedforward and feedback influences on CP [6,11]. Feedforward components of CP are thought to arise early in the stimulus-response period, with feedback influences arising later [6,9,39,47]. These correlational studies are, however, limited by the fact that the relationship between the timing of signals and their anatomical source is not entirely straightforward in visual cortex [14]. Moreover, fluctuations in neural activity are not strictly attributable to task performance [8,24,49]. Thus, there is increasing emphasis on causal methods for probing the circuitry of decision-making [39,50].

For example, neurons in parietal cortex often exhibit strong choice-related signals, and reversible inactivation can lead to significant perceptual deficits [29]. But other studies have found little effect of parietal inactivation on behavioral performance, suggesting a noncausal component of these decision signals [13,49,55,56]. A similar pattern has emerged in area MT, with most studies finding large perceptual deficits with inactivation [13,49,55] and others finding little or none [24]. It is clear from these studies that the causal influence of a cortical area on perceptual decisions cannot be predicted from CP alone, but rather depends on the precise nature of the task, the training procedure, and the stimuli [8,24,29].

A few previous studies have attempted to evaluate CP based on LFP power [16,17]. In one study, the authors recorded spikes and LFPs from V1 and V4 while monkeys performed a two-alternative forced-choice orientation discrimination task [16]. Similar to our findings, they reported decision-related information in the alpha and beta frequencies, as well as in high-gamma frequencies during the stimulus-response period. Liu and Newsome recorded LFPs from area MT during a speed discrimination task and found significant LFP-based CP in frequencies above 40 Hz, while LFP power below 30 Hz actually exhibited below-chance CP values [17].

Frequency-specific neural computations in decision-making

The strong reduction of high-gamma CP signals following muscimol inactivation (Fig 3A) supports the hypothesis that these signals represent feedforward processing [20,21,57]. This might result from the fact that higher-frequency LFPs are generally found in the upper layers of cortex, which send feedforward projections to higher-level areas [30], although we were unable to verify the laminar locations of our recordings. We attempted to examine the laminar dependency of CP signals in our data, but we found no reliable depth gradients or superficial-deep contrasts in any frequency band before and after inactivation (see Methods for details), presumably because we did not have a reliable way to estimate the laminar position of our recording channels [58].

The persistence of low-gamma CP signals (Fig 3B) after muscimol inactivation could reflect feedback signals from decision-related areas [6,9,47]. Previous studies found that low-frequency LFP components in V4 could remain intact when local spiking was strongly reduced, which suggests that such signals primarily reflected long-range synaptic input rather than local output [38,59]. Alternatively, they could reflect shared noise correlations between neural activity in MT (or V4) and other areas, such as V3A [60], MST [61], or IT [25]. These could be feedforward in nature, in the sense that their sensory activity might inform decisions, rather than the other way around. We have taken a simplified view of feedforward and feedback contributions here, but feedback needs not be noncausal; in recurrent decision circuits it can help drive commitment itself, for example, by accelerating ramping toward the bound [62], and the very short trials in our task likely limited such effects.

CP signals in the alpha and beta bands were also found before and after inactivation (Fig 3C). Given that the results were not consistent across monkeys, this alpha-beta component of CP can hardly be attributed to any mechanisms related to sensory processing. Rather, they seem to reflect the reward outcome of previous trials, as has been found for spiking activity [7]. One interpretation of this finding is that absence of reward transiently boosts arousal or attention, which influences both stimulus coding and task performance [7,44]. We could not test this possibility directly without quantifying arousal by pupil size, but we took an indirect approach assuming that arousal would affect broadband, epoch-invariant modulation [63,64]. However, we still found no evidence to favor the interpretation that arousal mediated the relationship of alpha-beta CP and reward history. Alternatively, such reward history-dependent modulation may reflect changes in internal state, which would be better addressed in paradigms designed to study learning.

Interpretation of CP values

Although the absolute magnitude of our LFP-based CP values was typically near 0.5, such values are consistent with the literature on decision-related activity in sensory cortex [65]. Early work in area MT demonstrated that perceptual decisions could be accounted for by the activity of a relatively small number of neurons [36]. Subsequent analyses showed that CPs below 0.6 are pervasive because only a small fraction of neurons need to be activated to yield decisions [65]. From this perspective, the modest but statistically significant deviations from 0.5 observed in CP values reflected the relationship between population-level neural activity and perceptual decision.

On average, our MT results exhibited relatively lower CP magnitudes than our V4 dataset. This difference likely stemmed from the temporal structure of our behavioral paradigms. In the MT task, the brief presentation of a visual stimulus in our experiments might not have exerted a sufficiently strong influence on local neural activity, whereas the two-fold longer presentation of a visual stimulus in our V4 experiments might have allowed more time to integrate sensory information.

Ethics statement

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

Observers

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

Experimental setup

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

Neural recording

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

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

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

Visual stimuli and behavioral paradigm

Infusion of muscimol

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

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

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

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

Analyses of neural and behavioral responses

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

Psychometric curve fitting.

The monkeys’ behavioral performance as a function of dot coherence or form SNR level was characterized by fitting a Weibull model to the proportion of correct responses by using nonlinear regression (fitnlm in MATLAB). The two-parameter Weibull function is,

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

LFP-based choice probabilities (CP).

The computation of LFP-based CP, a metric that quantifies trial-to-trial fluctuations between LFP power variability and psychophysical choices, was conducted, mostly using established methods developed for spiking activity [4].

We first subdivided trials into two choices defined by neurons’ stimulus selectivity and behavioral outcomes. We included sites for which the preferred stimulus had at least a 50% higher multi-unit spike count in the stimulus-response period (50–250 ms relative to stimulus onset) than the spikes in the baseline (−200–0 ms relative to stimulus onset). The stimulus-response interval was selected due to the significant correlation between the spikes observed during this time window and the animals’ behavioral choices [66]. Stimulus selectivity was assessed across multiple neurons at the highest dot coherence levels (64% and 100% for the MT task; 60% and 100% for the V4 task) for every electrode channel independently.

In this way, a preferred choice contained trials that exhibited either a preferred stimulus with a correct response or a nonpreferred stimulus with an incorrect response; an anti-preferred choice contained trials that involved either an anti-preferred stimulus with a correct response or a preferred stimulus with an incorrect response.

For the reward history analysis, current trials were additionally partitioned according to behavioral outcome of the immediately preceding trial. Specifically, for each valid trial , we identified trial in the original behavioral dataset and used its outcome to classify the current trial as rewarded if the previous trial had been correct, or as nonrewarded if the previous trial had been incorrect. Importantly, the LFP power used for CP estimation was always taken from the current trial, not from the preceding trial. Within each reward history condition, the current trials were then subdivided into preferred choice and anti-preferred choice using the same behavior-choice rules described above.

Because CP is meant to be insensitive to the mean amplitude of neural responses, most analyses of CP first involve normalization of firing rate, using z-scoring or balanced z-scoring [72]. However, given the nonnormal distribution of LFP power, this approach has the potential to introduce biased estimate of CP. To assess this possibility, we used the NeuroDSP toolbox [73] to simulate LFP data from 100 trials of one-second duration (S4A Fig), each comprised of stochastic periodic and aperiodic components; these were generated using the function sim_combined with the default parameters. We implemented this procedure twice to simulate trials associated with two behavioral choices that differed in mean amplitude across frequencies, one being 100 times greater in amplitude than the other (S4A Fig). Then, we computed spectrograms to represent the simulated LFP data before normalization (S4B Fig). A proper normalization would yield CP values of 0.5, because CP computation takes into account the variation between choices, not the difference in mean. However, we found that z-scoring and balanced z-scoring failed to eliminate the mean difference between the signals associated with the two choices, largely distorting the signals that were associated with lower-frequency LFP power (S4C Fig). This distortion arises because nonGaussian distributions like LFP power are susceptible to the skewness of the PSD. Thus, z-scoring disproportionately emphasizes outliers when the distribution is nonnormal, biasing LFP-based CP estimation.

We therefore used a robust scaling method that avoids such distortions. Specifically, we first resampled trials with each of the two choices to ensure that the preferred and anti-preferred choices had equal number of trials for each stimulus condition. We then normalized separately for the preferred and anti-preferred choices. For this normalization, we used the median and interquartile range (IQR) of the distribution of LFP power for each choice:

where X is the original distribution and X_normalized is the normalized version of the original choice by its median and IQR. This approach does not rely on any specific distributional assumptions but rather quantifies the central tendency and dispersion to better resist outliers in LFP. In our simulated example, robust scaling aligned the two choices to a uniform scale with identical values, thereby recovering the accurate CP values across all frequencies (S4D Fig). This simple simulation confirmed that robust scaling effectively returned CPs to chance when the LFP mean power is the only difference between two simulated LFP profiles, demonstrating its suitability for normalizing LFP power.

Following appropriate normalization, the distributions of two normalized choices were subsequently analyzed using receiver operating characteristic (ROC) analysis to compute CP. CP was determined by calculating the area under the ROC curve. The area under the ROC curve ranges from 0.0 to 1.0, reflecting the performance of an ideal observer in determining motion direction based on the trial-to-trial LFP power. Values of 1.0 and 0.0 represent perfect classification, whereas a value of 0.5 signifies performance equivalent to random chance classification.

CP can be sensitive to overall behavioral performance, and the infusion of muscimol significantly impaired the monkeys’ performance. It was therefore necessary to equalize performance pre- and post-muscimol by selecting trials with different stimulus levels. In the MT task, for Monkey Y, we used coherence levels of 0%, 4%, and 8% before muscimol injection and 0%, 8%, 12%, and 16% after injection; for Monkey C, we used coherence levels of 0%, 4%, 8%, and 12% before muscimol injection and 0%, 8%, 12%, and 16% after injection. In the V4 task, for Monkey L, we used stimulus SNR levels of 0%, 6%, 8%, and 12% before muscimol injection and 0%, 8%, 12%, and 25% after injection; for Monkey A, we used stimulus SNR levels of 0%, 6%, 8%, and 12% before muscimol injection and 0%, 12%, 25%, and 60% after injection. These stimulus levels equated behavioral performance before and after inactivation, at ~65% correct across monkeys for analyses of area MT (Monkey Y, 66%; Monkey C, 62%); at approximately 78% correct across monkeys for analyses of area V4 (Monkey L, 80%; Monkey A, 75%). However, because differences in stimulus strength could bias CP, we included trials of 0% level in both pre- and post-inactivation conditions. These ambiguous trials, on which monkeys performed at chance, controlled for any confounds related to the stimulus.

We then estimated CPs based on LFP power. First, we calculated the CP for each frequency bin individually, utilizing power measurements obtained from a sliding window of 200 ms width, advancing in 20 ms increments. We calculated the averaged CPs across three distinct epochs relative to stimulus onset: baseline (−200–0 ms), which occurred during the pre-stimulus fixation, stimulus-response (50–250 ms), which spanned between stimulus onset and post-stimulus fixation, and delay (250–400 ms) spanning before the appearance of the choice targets. The stimulus-response epoch was defined by the time that LFP power response was maximal (Fig 2C), and to analyze CPs across the two areas, we kept the epoch length consistent. We statistically quantified CPs by averaging power across three frequency bands: alpha-beta (5–30 Hz), low-gamma (30–70 Hz), and high-gamma (70–150 Hz). We noted that spike waveforms have statistically little effect on any broadband LFP power [57].

Decomposition of LFP population signals.

To compare stimulus-aligned and orthogonal structure in population LFP activity, we represented each session as a single-trial feature matrix , where each row corresponded to one trial and each column to one channel-by-frequency feature (). Features were the log-transformed stimulus-epoch LFP power values (5–150 Hz) concatenated across simultaneously recorded channels. Using this matrix, we defined two distinct types of axes.

First, we defined a decoder-based stimulus readout axis by training a linear classifier to discriminate stimulus identity. The resulting weight vector was normalized to unit length, , and taken as the stimulus axis, corresponding to the direction in population activity that best separated the two stimulus classes.

Second, to characterize activity structure not captured by that axis, we removed the projection of each trial onto the stimulus readout axis,

and applied PCA to the standardized residual matrix . The first residual principal component was taken as the leading null dimension [39,40]. To verify the geometry of this decomposition, we quantified the relationship between the null dimension and the stimulus axis using the absolute cosine similarity, , and the corresponding angle, , in degrees. As expected, the null dimension was orthogonal to the stimulus axis across sessions, with cosine alignment values on the order of and corresponding angles of .

To assess the spectral contribution of each axis, we quantified loading energy as the fraction of squared component weight assigned to features within each frequency band. For axis with weight vector , loading energy in band was defined as

where denotes the set of features belonging to band . Because this metric depends on squared weights, it is invariant to the arbitrary sign of the component vector.

Quantification and statistical analyses

We used nonparametric and ANOVA techniques for the statistical evaluation of our data. All statistical studies were performed using MATLAB’s and Python’s core tools and custom programs.

We employed the two-sided WSR test to assess the significance level of LFP-derived CPs relative to 0.5 when the sizes of the datasets were equal. The difference in LFP-derived CPs before and after muscimol infusion was evaluated using the WRS test. Significance was determined at p < 0.05.

To test whether the effect of inactivation interacted with behavioral epoch per frequency band, we fitted a linear mixed-effects model:

CP ~ epoch × condition + (1|monkey),

where CP was treated as a continuous dependent variable; epoch (baseline, stimulus, delay) and inactivation condition (pre, post) were fixed factors; and monkey was included as a random grouping factor to account for repeated measures across epochs. Coefficients represent changes in CP relative to the baseline epoch and post-inactivation condition, which together define the model’s reference level.

All analysis code, including the LFP-CP pipeline and validation simulations, is publicly available (see Code and Data Availability).