Surgical procedure and FEF mapping

Two male rhesus monkeys (Macaca mulatta) weighing between 6 and 8 kg underwent a single surgery during which they were implanted with two MRI compatible PEEK recording chambers placed over the left and the right FEF hemispheres respectively, as well as a head fixation post. Gas anesthesia was carried out using Vet-Flurane, 0.5%–2% (Isofluranum 100% at 1,000 mg/g) following an induction with Zolétil 100 (Tiletamine at 50 mg/ml, 15 mg/kg and Zolazepam, at 50 mg/ml, 15 mg/kg). Post-surgery pain was controlled with a morphine painkiller (Buprecare, buprenorphine at 0.3 mg/ml, 0.01 mg/kg), 3 injections at 6 hr-intervals (first injection at the beginning of the surgery) and a full antibiotic coverage was provided with Baytril 5% (a long-acting large spectrum antibiotic, Enrofloxacin 0.5 mg/ml) at 2.5 mg/kg, one injection during the surgery and thereafter one each day during 10 days. A 0.6 mm isotropic anatomical MRI scan was acquired post surgically on a 1.5T Siemens Sonata MRI scanner, while a high-contrast oil-filled 1 mm × 1 mm grid was placed in each recording chamber, in the same orientation as the final recording grid. This allowed a precise localization of the arcuate sulcus and surrounding gray matter underneath each of the recording chambers. The FEF was defined as the anterior bank of the arcuate sulcus and we specifically targeted those sites in which a significant visual, premotor and/or oculomotor activity was observed during a memory guided saccade task at 10–15° of eccentricity from the fixation point. In order to maximize task-related neuronal information at each of the 24-contacts of the recording probe, we only recorded from sites with task-related activity observed continuously over at least 3 mm of depth. All surgical and experimental procedures were approved by the local animal care committee (C2EA42-13-02-0401-01) approved by the French Ministry of Research and in compliance with the European Community Council, Directive 2010/63/UE on Animal Care.

Behavioral tasks and experimental setup

In all behavioral tasks, monkeys were seated in front of a computer screen (1920 × 1200 pixels and a refresh rate of 60 Hz) at a distance of 45 cm with their head fixed. To initiate a trial, the monkeys had to hold a bar in front of the animal chair, thus interrupting an infrared beam. They were required to hold the bar throughout the trial until a target stimulus was presented, otherwise the trial was aborted. The eccentricity of the visual stimuli was adjusted from day to day between 10–15°, to match the preferred spatial location of the multi-unit activity recorded on the 48 contacts of both recording probes.

Exogenous 100% validity cued luminance change detection task (Exo).

The task is a 100% validity cued luminance change detection task with temporal distractors (Fig 1A). To initiate a trial, the monkeys had to hold a bar in front of the animal chair, thus interrupting an infrared beam. A blue fixation cross (size: 0.7 × 0.7°) appeared in the center of the screen. Monkeys were required to hold fixation throughout the entire trial, within a fixation window of size 4° × 4°. Failing to do so aborted the trial and another trial started again. Four gray landmarks (0.5 × 0.5° for monkey M1, 0.68 × 0.68° for monkey M2) were presented simultaneously and equidistantly with the fixation cross. These landmarks were located, in the upper right, upper left, lower left and lower right quadrants of the screen, thus defining the corners of an illusory square. Their specific eccentricity was adjusted from day to day between 10–15°, to match the preferred spatial location of the multi-unit activity recorded on the 48 contacts of both recording probes. After a variable delay from fixation onset, ranging between 700 and 1900 ms, a green square was presented for 350 ms, indicating to the monkey in which of the four landmarks the rewarding target change in luminosity would appear. This green square (from now on, the cue) was presented on top of the peripheral landmark to be attended to. After the cue presentation, the monkeys needed to orient their attention to the target landmark in order to monitor it for an increase in luminosity while maintaining eye fixation onto the central blue cross. This increase in luminosity (from now on, the target) could occur anywhere between 500 and 2,800 ms from cue onset. The increase in luminosity was titrated during preliminary behavioral experiments, independently for each of the four possible target positions, such that behavioral performance during the Exo task would be in the range of 85–90% at each location. In order to receive their water or juice reward, the monkeys were required to release the bar (thus restoring the infra-red beam) in a time window of 200–700 ms following the target onset. This event accounted as a hit trial, while failing to respond to the target resulted in miss trials. In order to assure that the monkeys were correctly orienting their attention towards the cued landmark, unpredictable changes in the luminosity identical to the awaited target luminosity change could take place at the uncued landmarks (from now on, distractors). On each trial, from none to three such unpredictable distractors could take place, no more than one per uncued landmark position. Monkeys were trained to avoid responses to these distractors as a response to a distractor interrupted the trial and was counted as a false alarm trial.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 1. Description of the tasks and main behavioral results.

(A) Monkeys performed three behavioral tasks: The endogenous (Endo) and exogenous (Exo) attention tasks consisted of 100% validity cued target-detection task with distractors. To initiate the trial, monkeys had to hold a bar with the hand and fixate their gaze (white dashed line, first screen) on a central cross on the screen during the whole trial. A cue (green square, second screen) was presented indicating to the monkey in which of the four landmarks the target change in luminosity would appear. Monkeys received a liquid reward for releasing the bar 200–700 ms after target presentation onset. Monkeys had to ignore any uncued event (distractors). Importantly, they had to detect the stimuli using selective spatial attention (red dashed lines). In the memory-guided saccade task (Mem), a peripheral stimulus appeared 300 ms following fixation (500 ms). The monkeys had to hold the position in memory, throughout a variable delay of 700 to 1900 ms, until the fixation cross disappeared. At the extinction of the fixation cross, the monkeys had 500 ms to execute a saccade to the previously presented stimulus location. At the end of the saccade, the peripheral stimulus reappeared and the monkeys had to wait and respond when a target stimulus was presented. (B) Behavioral performance. On the top part, boxplots represent the distributions of hit rates and reaction times (RT) over all sessions (n = 12). On the bottom part, boxplots show distributions of the false alarm rate (FA rate) and reaction time (RT) of responses to distractors over all sessions (n = 12, non-parametric Wilcoxon signed rank test *p < 0.05, **p < 0.01). Data and code to Fig 1B can be found at https://osf.io/z8eh9/.

https://doi.org/10.1371/journal.pbio.3003353.g001

Neural recordings

Simultaneous recordings in the two FEFs were carried out using two 24-contact Plexon U-probes. The contacts had an interspacing distance of 250 µm. Neural data was acquired with PlexonOmniplex neuronal data acquisition system (Omniplex). The data was amplified by a factor of 400 and digitized at 40,000 Hz. The neuronal data was low-cut filtered at 300 Hz. All analyses were performed on the MUA recorded from each of the 48 recording contacts (M1: number of sessions = 7, total number of channels = 336; M2: number of sessions = 5, total number of channels = 240). A threshold was applied independently for each recording contact and before the actual task-related recordings started, such that the crossing of this threshold resulted in multiunit spike trains. This threshold corresponded to 3 standard deviations from mean activity during unconstrained eye movements by the monkeys while exploring a blank screen. This threshold was confirmed during the offline analyses to optimize single unit activity extraction for each channel. This readjustment impacted MUA final thresholding values. Spike sorting was performed, on a subsample of channels (M1: n = 94; M2: n = 69), using the Offline Sorter (Plexon), on the broadband time signal (bandpass filtered between 300 and 3,000 Hz) of the entire session prior to assigning the signal to the different tasks. Only spikes that could be unambiguously assigned to a same neuron throughout all the tasks were used for the SUA-based analyses (M1: n = 43; M2: n = 15). The “Clusters vs. Time” view was used (using the most informative principal component pairs) to check how the waveform changed as a function of time throughout the session. Only units that could be continuously tracked through the entire session, with no drift or a smooth continuous drift in PCA cluster attributes were used for the SUA-based analyses. Examples of spike sorted waveforms tracked in the three different tasks are presented in S1 Fig. All further analyses of the data were performed in Matlab (The Mathworks, Natick, Massachussetts).

Data pre-processing

In each task, MUA/SUA activity was epoched per trial. Spike trains were filtered with a Gaussian kernel (δ = 60 ms). First, MUA/SUA activity previous to the cue onset was epoched from −200 ms to 0 ms with respect to the onset of the Cue for the Pre-Cue analysis. Then, MUA/SUA activity was epoched between 500 ms to 0 ms prior to Target onset for the Delay analysis. To extract the averaged firing rates for dPCA analysis, these epoched MUA/SUA data were baseline corrected (independently for each channel and each trial) relative to the time period from −500 ms to −200 ms before the cue. This reduced trial-to-trial variability in baseline activity levels and reduced individual trial differences.

MUA selectivity

Spatially selective channels during the delay epoch were identified as follows. First, a non-parametric Wilcoxon test was used to identify responsive channels, i.e., channels for which the activities observed during the baseline and during the post-cue period (800–900 ms following cue onset, only considering trials for which target or fixation off event happened 1,000 ms after cue presentation) were significantly different for at least one of the four possible spatial cued positions (p < 0.05). For each responsive channel, a non-parametric two-way ANOVA Task × Position (David Stern, https://www.mathworks.com/matlabcentral/fileexchange/44308-randanova2, this method uses the Still and White’s [33] approach to calculate P-value for interaction effect) was then implemented on the average activity in the post-cue period in order to identify the Task related channels, the spatially selective channels as well as the mixed selectivity channels (p < 0.01). All p values were corrected for multiple comparisons using a Holm-Bonferroni correction. Note that the proportion of selective cells did not change substantially across different analysis time windows (S5 Fig).

Demixed PCA analysis

Recent studies have shown that neurons in the prefrontal cortex show simultaneous significant tuning to multiple task-related parameters, a property called mixed selectivity [12]. This feature is a hallmark of high dimensionality of the neuronal population. Due to this property, the activity structure of the population can be estimated by applying a dimensionality reduction method to the recorded activity such as Principal Component Analysis (PCA). Using this method, it is possible to extract a number of latent variables (principal components) that capture independent sources of data variance providing a description of the statistical features of interest [34]. However, the PCA method is blind to the sources of variability of the data and does not take into account the task related parameters, mixing these sources of information within each of the extracted latent variables [32]. Here, we wanted to study and compare the population activity structure in the three different tasks and describe how much the variance in the neuronal population can be explained by task-related aspects such as task, spatial target location, and their interaction. To address this question, we applied dPCA (demixed principal component analysis [31,32]) on the data which captures the maximum amount of variance explained by predefined sources of variability in each extracted component and reconstruct the time course of the category-specific response modulation. A recent publication indicates that dimensionality reduction methods are accurate in the absence of spike sorting [35]. In other words, neural dynamics and scientific conclusions are comparable when using multiunit threshold crossing or sorted neurons when using dimensionality reduction methods. This is particularly relevant to the present study. Indeed, 10–15 dimensions capture over 90% of the variability of trial-averaged neural population activity when monkeys perform the different tasks. This is associated with low error in manifold estimation [36]. As a result, due to the theory of random projections, neural population dynamics can be accurately estimated from our recordings using MUA instead of SUA, since the former is a low-dimensional representation of the latter. In the present work, the dPCA is applied on MUA activities.

Demixed PCA was applied in the two different time intervals (Pre-Cue and Delay). The purpose of the first dPCA analysis was to identify specific task-related components during the Pre-Cue period. Then, another dPCA was performed during the delay period, with the aim to extract task and spatial-related components, as well as components associated with their interaction (non-linear mixed selectivity). Both analyses were performed using appropriate Matlab based toolbox developed by Kobak and colleagues [32]. We considered the 20 first demixed components that accounted for >90% of the total variance.

We used the decoding axis of each dPC (assigned to task in the Pre-Cue analysis, or task, position and interaction in the Delay analysis) as a linear classifier to decode the different types of trials (Endo, Exo, and Mem Sacc for task, and Up Left, Up Right, Down Left, and Down Right for position), following the procedure described by Kobak and colleagues [32]. This method allows to quantify the capacity of each demixed component to classify a trial between the classes of a given category. To extract the statistical significance of this accuracy, we shuffled 100 times all available trials between classes and we thereby computed the distribution of classification accuracies expected by chance. For each session, the chance-level was considered as the maximal accuracy value obtained across all randomizations.

Monkey performance is task-dependent

The present study aims to describe and compare, in the PFC, the electrophysiological underpinnings of spatial information encoding during the performance of different tasks. To this end, we recorded MUA from bilateral monkey frontal eye field (FEF), an area involved in the control of covert spatial attention [5,7,37,38] and saccades [16,39–41], of two macaque rhesus monkeys while they performed three different tasks: An endogenous attention task (Endo), an exogenous attention task (Exo) and a memory-guided saccade task (Mem). We used two probes (one in each recording site) containing 24 contacts each in both monkeys, which were implanted in each recording session (number of sessions: M1, n = 7; M2, n = 5). The median number of trials per task was as follows: Exogenous task, M1: 318 (iqr = 59); M2: 380 (iqr = 62); Endogenous task, M1: 357 (iqr = 67); M2: 589 (iqr = 171); Memory guided saccade task, M1: 244 (iqr = 39); M2: 302 (iqr = 184). These three tasks were presented consecutively in the same session in a block design. The schematic of these tasks was very similar (see Fig 1A and “Methods” section for an extensive description of these tasks): At the beginning of each trial, a visual cue was presented in one of the four possible screen quadrants, and after a variable and randomized time interval monkeys had to produce a response. However, the actual task differed, and the monkeys needed to adapt their cognitive processes to perform each task. The exogenous and endogenous attentional tasks consisted of a 100% validity cued luminance change detection task in which monkeys had to respond manually to a target located peripherally in the cued quadrant. In the exogenous version, the cue was shown in the peripheral landmark to be attended to (thus, in the same position as the target). In contrast, in the endogenous attention task the cue was presented close to the fixation cross pointing towards the landmark to be attended. In both endogenous and exogenous attention tasks, distractor stimuli were presented during the cue to target interval (CTOA), and monkeys needed to ignore them (otherwise, the trial was cancelled). These two tasks were used to recruit attentional processes [29]. In the memory-guided saccade task, monkeys were required to hold the position of a spatial cue in memory during a variable amount of time and, then, to perform a saccade towards that memorized spatial location on the presentation of a go signal. Importantly, and in contrast with the attentional tasks, monkeys were required to perform a spatially oriented oculomotor response rather than a manual non-oriented response. This task thus recruited spatial working memory processes [29,16,41]. Although these three tasks showed some similarities in their structure, they recruited different cognitive processes. While monkeys needed to enhance visual attention in the cued quadrant to respond to the target onset in the exogenous and endogenous attentional tasks [29,42,43], they needed to memorize the location of the cue to produce an oriented eye movement towards that location in the memory-guided saccade task. Sensory conditions were also different. Indeed, in the attentional tasks, peripheral landmarks were present all throughout the trial, while they were absent in the memory guided saccade task. While this probably did not impact pre-cue neuronal activity, it possibly impacted post-cue task differences observed in the activity.

Both monkeys showed different behavioral patterns as a function of the type of task they performed. Non-parametric Friedman test (Tasks as repeated measures, df = 2, performed across sessions) revealed that hit rates (M1: χ²(5) = 12.29, p = 0.002; M2: χ²(5) = 7.6, p = 0.0224) and reaction times (RT) (M1: χ²(5) = 12.29, p = 0.0021; M2: χ²(5) = 7.6, p = 0.0224) were different across tasks (Fig 1B). This reflects different task demands.

Both monkeys were able to discriminate the change in luminosity of the cued landmark while ignoring distractors in the spatial attention tasks, though their performance differed depending on the type of task. Indeed, both monkeys showed a higher hit rate when performing Exo than Endo, (Fig 1B M1: 92% versus 78%, Wilcoxon ranksum test, Z = 3.07, p = 5.8e − 04; M2: 87% versus 60%, Wilcoxon ranksum test, Z = 2.09, p = 0.0079). M1 showed a lower hit rate in memory-guided saccade task than in Exo and Endo, while M2 showed significantly lower hit rate in Endo than in Mem (Fig 1B M1: 72% versus 92%, Wilcoxon ranksum test, Z = 3.07, p = 5.8e − 04, 72% vs.78%, Z = 2.17, p = 0.0233; M2: 60% versus 78%, Z = 2.51, p = 0.0079). Regarding the RT, and expectedly, both monkeys had slower manual response times in attentional tasks than saccadic response times in the Mem task (M1, Friedman test, χ²(5) = 12.29, p = 0.0021; M2, Friedman test, χ²(5) = 7.6, p = 0.022. M1: Exo versus Mem, Wilcoxon ranksum test, Z = 3.07, p = 5.8e − 4, Endo versus Mem, Wilcoxon ranksum test, Z = 3.07, p = 5.8e − 4; M2, Exo versus Mem, Wilcoxon ranksum test, Z = 2.51, p = 0.0079; Endo versus Mem, Z = 2.51, p = 0.0079). Regarding the attentional tasks, although no significant difference was found, RTs tended to be slower for the Endo than for the Exo task as reported in other studies [29].

False alarm rates (measured as the ratio between the number of responses to distractor and the total number of distractors shown) and reaction time in false alarms were analyzed only for the attentional tasks, since no distractors were used in Mem. Both monkeys showed lower false alarm rates in Exo than in Endo (Fig 1B M1: 7% versus 13%, Wilcoxon ranksum test, M1: Z = 2.56, p = 0.007; M2: 10% versus 21%, Wilcoxon ranksum test, Z = 0.63, p = 0.0079). Monkeys did not show differences in reaction times in the false alarms between the two attentional tasks. Notably, responses were slower in false alarms than in hits for M1 in Endo (Wilcoxon ranksum test, Z = 1,79, p = 0.0175).

Spiking activity in the FEF encodes task- and sensory- related information simultaneously

We examined the FEF neural responses during the performance of these three tasks. To this end, we averaged the smoothed firing rates in the time interval −500 ms to 1,000 ms locked to the cue for each of the three tasks. As reported in previous studies [7,42,44,45], these recorded neurons showed enhanced responses when the cue was presented in or was cueing towards their preferred spatial position, relative to when it was presented in the least preferred spatial location, and this was true for all three tasks (Fig 2A and 2B, see S2 Fig for SUA analysis). Similar results were found when data were locked to the target onset (S3 Fig). Spatial selectivity however varied between the three tasks. Indeed, we found a significant task-effect, across the entire population, on the response to the preferred position (non-parametric Friedman test, performed across MUAs, with tasks as repeated measures, χ²(2) = 39.3, p = 2.89e − 09).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 2. Mixed selectivity corresponds to the ability of a neuronal population to encode simultaneously independent sources of information: (A) Firing rate of a representative MUA aligned on the cue (200 ms before the cue onset to 1,000 ms after the cue onset) in exogenous task (blue), endogenous task (green) and memory guided saccade task (red) for preferred (solid line) and non-preferred (dash line) position.

The data were baseline corrected (−500 to −200 ms before cue onset) on individual trials and channel. The gray box corresponds to the delay period during which neuronal activities were analyzed for subplots C and D. (B) Mean firing rate of the MUA neuronal population in the three tasks aligned on cue presentation (200 ms before the cue onset to 1,000 ms after the cue onset) in the exogenous task (blue), the endogenous task (green) and the memory guided saccade task (red) for preferred (full line) and non-preferred (dash line) position. The data were baseline corrected (−500 to −200 ms before cue onset) on individual trial and channel. Gray box corresponds to the delay period during which neuronal activities were analyzed for subplots C and D. (C) Venn diagram representing the proportion of cells showing a spatial tuning in each of the three tasks. (D) Pie chart corresponding to the proportion of neurons tuned to only position (red), task (blue) and both (Additive mixed-selectivity; yellow; Mixed-selectivity with interaction; green). Non-selective neurons are plotted in gray. Wilcoxon ranksum test was performed between baseline (−100 to 0 before the cue onset) and time interval (800 to 900 ms after the cue onset) for each channel and position across trials. A non-parametric 2-Way-ANOVA was performed on selective channels with Bonferroni–Holm correction on spatial and task factors on time interval from 600 to 900 ms after the cue onset (p < 0.01). Data and code to Fig 2 can be found at https://osf.io/z8eh9/.

https://doi.org/10.1371/journal.pbio.3003353.g002

Specifically, we observed that some cells were recruited during the performance of the three tasks during the delay period (Fig 2A, gray box). We measured the proportion of cells that expressed a significant spatial selectivity during the delay period of the three tasks. Results of this measure are represented in Fig 2C as a Venn diagram. Interestingly, we found that 17% of cells were tuned to position in all three tasks, indicating that these cells encoded spatial information both during spatial attention and spatial working memory. In addition, we found neurons that showed mixed selectivity for both task and spatial information, like the one represented in this figure. Therefore, we classified all recorded cells based on either their selectivity to position, task, or to simultaneous position and task selectivity. Based on the statistical significance of the firing rate modulation for these two parameters (spatial selectivity and task selectivity), we found that 32% of cells were uniquely selective for the task, 6% were uniquely selective for position, and 60% showed mixed selectivity for both parameters: 38% showed no interaction between task and position (corresponding to linear mixed-selectivity) and 22% showed an interaction between task and position (corresponding to non-linear mixed-selectivity – Fig 2D). S2 Fig duplicates the analyses presented in Fig 2 for SUAs. S4 Fig represents an example and the population activity for each category of MUAs/SUAs. This categorization was robust against possible sources of noise in the data (S1 File Supporting information: A, MUA; B: SUA). For MUA, selectivity did not depend on the post-cue analysis window (MUA: S5A Fig), and only marginally did so for SUA (SUA: S5B Fig). As expected, no position-related nor mixed-selectivity MUAs or SUAs were identified during the pre-cue epoch (−200 to 0 ms prior to cue presentation, S5 Fig). These results indicate that the FEF encodes spatial information and the underlying cognitive processes recruited by each task through mixed-selectivity neurons. Importantly, we found that 94.6% of channels that were exclusively task selective during the delay, were also task-selective in the pre-cue period.

Fig 3 aims to characterize the impact of mixed-selectivity, with and without interactions, on attention selectivity spatial indicators prior to target presentation. Some MUAs had a firing rate that was significantly higher for a well-identified preferred position compared to the three other positions in at least one of the three tasks. Such a spatially organized response pattern could be related to the match between our task configuration and the MUAs’ receptive field location, where one of the target positions coincided with the center of the receptive field. However, such a response pattern was predominantly identified in the Exo task (Mixed-selectivity without interaction: Exo: 18%, Endo: 5%, Sacc: 5%; Mixed-selectivity with interaction, Exo: 46%, Endo: 24%, Sacc: 18%). This indicates that, independently of the relative match between task configuration and MUA receptive field, the response pattern changed between the Exo task and the other tasks and became spatially less resolved. This is exemplified in Fig 3, which shows representative single MUAs (Fig 3A, mixed-selective MUA finely tuned receptive field; 3B, mixed selectivity MUA with broad receptive field; MUA and SUA examples showing finely-tuned receptive fields or broad receptive fields are presented in S6 Fig) and population-level responses (Fig 3C, average population mixed-selective MUA without interaction; 3D, with interaction). To further analyze the impact of the task on the reconfiguration of the MUAs’ spatial selectivity, we identified, for all mixed-selectivity MUAs, the two spatial positions eliciting the highest firing rates. We then compared the degree of partial overlap across tasks, pairwise. For the vast majority of mixed-selectivity MUAs without interaction, the receptive fields shared at least one spatial position across each pair of tasks (Exo–Endo: 98%; Exo–Sacc: 95%; Endo–Sacc: 94%). This was also the case for the receptive fields of mixed-selectivity MUAs with interaction, with the marked exception of the Endo-Sacc comparison (Exo–Endo: 96%; Exo–Sacc: 93%; Endo–Sacc: 69%). Because the reconfiguration of receptive fields between the Endo and Sacc tasks was specific to this MUA category, it strongly indicates that these MUAs play a distinct role in these tasks. It is worth noting that task order cannot account for these selective changes, not only because they are specific to this MUA mixed-selectivity category, but also because the order of the Exo and Endo tasks was randomized across different sessions. This possibly relates to previous observations describing shifts of receptive fields in the parietal and prefrontal cortex around saccade execution [46–50]. Likewise, as these shifts are mostly evident when comparing the Endo and Sacc tasks, it would be relevant to understand in what tasks and context such shifts arise [50].

Download:

• PNG
  larger image
• TIFF
  original image

Fig 3. Spatial tuning of mixed-selectivity cells.

(A) MUA example, with mixed-selectivity, and finely-tuned receptive field. Plots represent firing rates (average across trials ± s.e.) baseline corrected (−500 to −200 ms before cue onset) on individual trial: Exo: blue, Endo: green, Sacc: red. Firing rates at preferred spatial position are significantly different from the other spatial locations for all tasks (p < 0.005 or better). (B) MUA example, with mixed-selectivity, and board receptive field. All else as in (A). Firing rates at the two preferred spatial positions are significantly different from the other spatial locations for all tasks (p < 0.05 or better), except for the Ipsi vs. Contra and Diagonal comparison in the Sacc task. (C) Population spatial tuning of mixed-selectivity MUAS with no interaction. Plots represent firing rates (average across MUAs ± s.e.) baseline corrected (−500 to −200 ms before cue onset) on individual trial: Exo: blue, continuous: MUAs with broad receptive fields (RF), dashed: MUAs with finely-tuned receptive fields, Endo: green, Sacc: red. Firing rates at preferred spatial position are significantly different from the other spatial locations for all tasks and conditions (p < 0.005 or better). (D) Population spatial tuning of mixed-selectivity MUAS with interaction. All else as in (D). Firing rates at preferred spatial position are significantly different from the other spatial locations for all tasks and conditions except for the firing rates of the MUAs with broad receptive fields in the Exo task (p < 0.001 or better). Firing rates at Ipsi spatial position are significantly different from the Contra and Diag spatial locations for all tasks and conditions except for the Sacc task (p < 0.005 or better). Preferred indicates the spatial location eliciting the highest variation in firing rate from baseline. Ipsi indicates the stimulated spatial position that falls in the same hemifield as the Preferred location. Contra indicates the stimulated spatial position that falls in the contralateral hemifield as the Preferred location. Diag indicates the stimulated spatial position that falls in the contralateral hemifield as the Preferred location, diagonally to it. Data and code to Fig 3 can be found at https://osf.io/z8eh9/.

https://doi.org/10.1371/journal.pbio.3003353.g003

Task-related information accounts for FEF neuronal variability before cue presentation

In the previous section, we described that the averaged firing rate pattern in response to the cue and during the delay epoch differed as a function of the type of task that monkeys were performing. We hypothesize that such different firing rate patterns as a function of the type of task, result from a context-dependent functional reorganization of the neural population. To test this hypothesis, we used dimensionality reduction to identify the low-dimensional representation of the neural recordings. Dimensionality reduction consists of projecting the high-dimensional neuronal recordings onto a low-dimensional space minimizing information loss. To this end, we applied principal component analysis (PCA) to our data pooled along task category in a pre-cue time interval. At this time of the trial, task-related events soliciting the processing of spatial information have not been presented yet. Because trials from any given task are organized in blocks, at this time in the trial, the only prior information available to the monkey is task-related information. The projection of this pre-cue neuronal population data onto the principal components of the PCA showed a mixing of task-information across the different components, as task-related neural variability was projected onto multiple PCA axes (S7 Fig). In the following we will use the term task-related information to refer to the information available in the neuronal population about task identity, i.e., about which task the monkey is engaged in.

Other methods of dimensionality reduction allow the labelling of the explained variability by projecting specific sources of variance held in data onto each principal component. One of these methods is the demixed principal component analysis (dPCA) [32]. The objective of this method is to capture the majority of the variance in data in only few latent variables or components (as PCA does) while demixing the dependencies of the population activity on specific parameters by not constraining the components to be orthogonal. This method thus allows to directly interpret the impact of a related feature (e.g., task, position) on the neural population encoding. We applied dPCA on FEF MUA neuronal population activity in order to extract components related to task-related variability, unmixed from other types of variability. To do this, we pooled data across task category (Exo, Endo, and Mem). This analysis was firstly performed during the pre-cue period (200 ms before cue onset to cue onset) and will be later performed during the cue to target interval (500 ms to 0 ms before target onset) to unmix task- from sensory- specific variance.

The dPCA indicated that 90% of the variance of the recorded brain activity prior to cue presentation was associated with the task (Fig 4A, see S8 Fig for duplication on SUA data). The overall variance explained by dPCA is similar to the variance explained by PCA (Fig 4B) indicating that dPCA reduces dimensionality of the data with a similar performance as PCA, thus minimizing the neuronal information loss. The three first dPCs (i.e., the two first task-related components (#1 and #2), and the first task-independent component (#3)) explained 80% of the variance, indicating that with only few components we recovered the majority of data variance. Therefore, population activity was accurately represented by the dPCA components. The angle between the two first components related to the task, did not significantly deviate from orthogonality (88°). This was confirmed by a second analysis performed on individual sessions (mean = 82°, std = 4°). This thus indicated that these two first components represented independent task-related information.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 4. Demixed PCA unmixes task and independent sources related to variance.

(A) Pie chart shows how the total signal variance is split among parameters: task (blue) and condition independent (red). (B) Cumulative variance explained by PCA (black) and dPCA (red). Demixed principal component show similar explained variance as PCA. (C) Demixed principal components. In each plot, normalized MUA firing rate averaged per task (in arbitrary units, exogenous (blue); endogenous (green) and memory-guided saccade (red) in the interval from −200 ms to 0 ms locked to the cue onset are projected onto the task-related dPCA components. Thick black lines show time intervals during which the task-information can be reliably extracted from single-trial activity (assessed against 95% C.I.). (D) True positive decoding rate for PC1 task-related. (E) True positive decoding rate for PC2 task-related. Dashed lines in (C) and (D) represent the 95% C.I. See S8 Fig for duplication on SUA data. Data and code to Fig 4 can be found at https://osf.io/z8eh9/.

https://doi.org/10.1371/journal.pbio.3003353.g004

Firing rates projected onto the first dPCA component associated with the task show a dissociation between the two attentional tasks and the working memory task (Fig 4C). Notably, such differences could be due to a reconfiguration of the prefrontal population computations between the different tasks. This reconfiguration could be associated with higher order cognitive factors segregating attentional processes from working memory processes. Alternatively, it could be associated with other task parameters such as effector specificity, the attentional task requiring a manual response, while the memory guided saccade task requires a saccadic response. Further experiments will be needed to precisely identify the computational underpinnings of these task-related differences. The projection of the firing rates onto the second dPCA associated to the task shows a dissociation between the exogenous (blue) and endogenous (green) tasks, suggesting a dissociation as a function of the attentional task-demands specifically within the attentional tasks. S9 Fig, shows, per each session, the projection of the dataset onto each of these components per task (trial-based). The third dPCA is associated with task-independent variance.

In order to assess whether the tuning of each of these two demixed components was statistically significant, we used these components as linear decoders to measure their ability to encode task-related information. To this end, we used cross-validation to measure time-dependent classification accuracy, and shuffling procedure was used to assess the significance of the accuracy measure (see “Methods” for more details about decoding). Confusion matrix of each classification was extracted to analyze the true positive rates (TPR) of the classifier for each task. Fig 4D shows that for the first dPC, the TPR in classifying the memory task was 80% (iqr 7.3), while the TPR for both attentional tasks were significantly lower and not significantly different one from each other (1-way ANOVA, task as main factor: p = 1.1835e − 15, Exo versus Mem: post-hoc Bonferroni test: p = 5.4277e − 14; Endo versus Mem: post-hoc Bonferroni test: p = 1.8701e − 14; Exo versus Endo: post-hoc Bonferroni p > 0.05) suggesting a dissociation between attentional and working memory processes. In contrast, TPR in predicting the memory task using the second dPC was lower than predicting both exogenous and endogenous attention (1-way ANOVA, task as main factor: p = 0.002, Exo versus Mem: post-hoc Bonferroni: p = 0.0218; Endo versus Mem: post-hoc Bonferroni: p = 0.0023; Exo versus Endo: post-hoc Bonferroni: p > 0.05) suggesting a dissociation in the way the attentional processes (exogenous versus endogenous) were implemented.

In order to confirm that dPC1 reflected a genuine task-identity coding, rather than session order effects, we performed task-dPCA on the independent sessions. For each session and each task, we projected the population activity onto the first task component. We then plotted the average of this activity not as a function of task, but as a function of task order in the session. While the memory guided saccade task was always the last task, half of the sessions started with the Exo task and the other half started with the Endo task. There was no statistically significant difference between the distributions of average activity recorded during the first and second task projected onto the first dPC (S2 File Supporting information). Thus, the task-dPCA component cannot distinguish between first and second task in the session, while it can distinguish between Endo and Exo attentional tasks. These precue effects were not driven by recording instability (S10 Fig and associated text).

Task-related and spatial information segregate in orthogonal demixed components

Mixed selectivity corresponds to the ability of neurons to simultaneously encode different sources of information [12,13]. The aim of the present analysis was to first investigate whether these different sources of information could be assigned to separate principal components and then to quantify the level of their interaction. Here, we were interested in the task and target spatial information, thus, we focused on the delay period between the cue and the target in order to include spatial information, as opposed to the pre-cue interval that had been selected in our previous analysis, in which there was no spatial information available yet. We applied dPCA on neuronal population activity during the delay epoch, respectively organized by position independent of task, and by task independent of position, in the time interval corresponding to cue to target interval (500 ms to 0 ms before target onset) (Fig 5, see S11 Fig for replication of analyses on different time intervals and S12 Fig for replication on SUA data). Demixed PCA split variance in position-related variance (52% of the explained variance), task-related variance (26% of the explained variance); and interaction between task information and spatial information (20% of the explained variance) (Fig 5A). The rest of the variance (2%) was related to independent sources of information. The overall variance explained by dPCA was similar to that explained by PCA, both accounting for 70% of the variance by the five first components. As for the pre-cue period (see previous section), we found that the first task-related dPC did not significantly deviate from orthogonality relative to the second task-related dPC during the cue-to-target interval (87.9°), and both of these did not significantly deviate from orthogonality relative to the first position-related dPC (81.6°; and 89.8°). We reproduced these observations when performing this analysis on the independent sessions (dPC1-task versus dPC2-task: mean = 83°, std = 5°; dPC1-pos versus dPC1-task: mean = 74°, std = 8°; dPC1-pos versus dPC2-task: mean = 83°, std = 7°). Therefore, population activity was accurately represented by different orthogonal dPCA components (Fig 5B). Similarly to what we found in the previous section, the two first components associated with the task showed, respectively, a dissociation between attentional task-related processes and working memory task-related process (component 3), and a trend of separation between the two attentional tasks (component 4). In addition, the first component associated with position exhibited a dissociation of each position corresponding to the four expected target locations, whereas the projection onto the second spatial-related component showed two different states, possibly corresponding to the segregation of information across hemifields and recording probes (i.e., probes respectively located into the right or left FEF) (S11 Fig). The distribution of weights attributed to the different units in the first position-related dPC as a function of the first task-related dPC were unimodal, such that most units contribute to the coding of both task- and position-related information (S13 Fig). This supports the fact that mixed-selectivity contributes to the lower dimensional coding of these two variables. Last, the interaction component (Fig 5C, bottom panel) describes the non-linear mixed selectivity between information related to task and position. We used this interaction component to decode trials based on the cued position and task. Fig 5D shows the average confusion matrix of this decoder across all sessions. We found that true positive rates for each of the 12 classes were significantly above the 95% C.I. (S14C Fig). Overall, this indicates that part of the spatial information represented in the FEF neuronal population is independent (orthogonal) to the task-related information, while another part of this spatial information interacts with the task-related information.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 5. Demixed PCA unmixes task and spatial sources related to variance.

(A) Pie chart shows how the total signal variance is split among parameters: Task (blue), Position (red), Interaction (purple) and Condition independent (gray). (B) Cumulative variance explained by PCA (black) and dPCA (red). Demixed principal components show similar explained variance to PCA. (C) Demixed principal components. In each plot, MUA firing rate averaged per task (exogenous (blue shades); endogenous (green shades) and memory guided saccade (red shades)) and cued position in each task, in the interval from 500 ms to 1,000 ms locked to cue onset, are projected onto the task-related, position-related and interaction dPCA components. Thick black lines show time intervals during which the parameter information can be reliably extracted from single-trial activity (assessed against 95% C.I.). (D) Confusion matrix of decoding task and position using the demixed interaction components, averaged across all sessions. Absolute chance level of the confusion matrix is 1/12 = 0.08. All elements along the diagonal are significantly above the 95% C.I. Data and code to Fig 5 can be found at https://osf.io/z8eh9/.

https://doi.org/10.1371/journal.pbio.3003353.g005