Abstract
To better understand how prefrontal networks mediate forms of cognitive control disrupted in schizophrenia, we translated a variant of the AX continuous performance task that measures specific deficits in the human disease to 2 male monkeys and recorded neurons in PFC and parietal cortex during task performance. In the task, contextual information instructed by cue stimuli determines the response required to a subsequent probe stimulus. We found parietal neurons encoding the behavioral context instructed by cues that exhibited nearly identical activity to their prefrontal counterparts (Blackman et al., 2016). This neural population switched their preference for stimuli over the course of the trial depending on whether the stimuli signaled the need to engage cognitive control to override a prepotent response. Cues evoked visual responses that appeared in parietal neurons first, whereas population activity encoding contextual information instructed by cues was stronger and more persistent in PFC. Increasing cognitive control demand biased the representation of contextual information toward the PFC and augmented the temporal correlation of task-defined information encoded by neurons in the two areas. Oscillatory dynamics in local field potentials differed between cortical areas and carried as much information about task conditions as spike rates. We found that, at the single-neuron level, patterns of activity evoked by the task were nearly identical between the two cortical areas. Nonetheless, distinct population dynamics in PFC and parietal cortex were evident. suggesting differential contributions to cognitive control.
SIGNIFICANCE STATEMENT We recorded neural activity in PFC and parietal cortex of monkeys performing a task that measures cognitive control deficits in schizophrenia. This allowed us to characterize computations performed by neurons in the two areas to support forms of cognitive control disrupted in the disease. Subpopulations of neurons in the two areas exhibited parallel modulations in firing rate; and as a result, all patterns of task-evoked activity were distributed between PFC and parietal cortex. This included the presence in both cortical areas of neurons reflecting proactive and reactive cognitive control dissociated from stimuli or responses in the task. However, differences in the timing, strength, synchrony, and correlation of information encoded by neural activity were evident, indicating differential contributions to cognitive control.
Introduction
Cognitive control is the process by which the brain uses changes in internal state, encoded by neural activity representing such variables as rules, goals, or behavioral context, to modify responses to environmental stimuli (Miller and Cohen, 2001). Understanding the neural and cortical basis of cognitive control is essential to understanding the rapid adaptability of human behavior. In contrast to slower, incremental forms of behavioral adaptation (e.g., reinforcement learning), cognitive control alters input–output relationships in the brain rapidly, in the time needed to establish a new pattern of activity over neurons encoding a new rule or goal. Functional imaging studies in humans have demonstrated that cognitive control jointly activates the PFC and parietal cortex (Yeung et al., 2006; Woolgar et al., 2015; Sheu and Courtney, 2016; Tschentscher and Hauk, 2016; Kwashie et al., 2022), and neural recording studies in nonhuman primates performing cognitive control tasks have compared the physiological activation of prefrontal and parietal neurons to delineate their respective contributions to this function (Nieder and Miller, 2004; Buschman and Miller, 2007; Goodwin et al., 2012; Crowe et al., 2013; Qi et al., 2015; Ramirez-Cardenas et al., 2016; Jacob et al., 2018; Panichello and Buschman, 2021; Zhou et al., 2021). In apparent contradiction to the different functions attributed to PFC (e.g., executive control and working memory) and parietal cortex (spatial processing for attention and motor control), neurons within these areas typically exhibit parallel changes in firing rate during cognitive control tasks, perhaps as a consequence of the reciprocal, corticocortical projections that extend between PFC and parietal cortex (Cavada and Goldman-Rakic, 1989; Medalla and Barbas, 2006). From a functional standpoint, this confounds functional dissection of PFC and parietal cortex in cognitive control. One potential resolution may relate to the level of scale at which the comparison between cortical areas is made. Namely, differences between prefrontal and parietal areas, while subtle at the level of individual neurons, clearly emerge at the level of neural populations (Qi et al., 2010; Goodwin et al., 2012; Swaminathan and Freedman, 2012; Meyers et al., 2018), particularly in relation to the temporal dynamics of neural population activity (Buschman and Miller, 2007; Crowe et al., 2013; Katsuki and Constantinidis, 2013; Phillips et al., 2014; Astrand et al., 2015; Antzoulatos and Miller, 2016; Jacob et al., 2018; Meyers et al., 2018; Bastos et al., 2020; Panichello and Buschman, 2021; Dang et al., 2022). The analysis of cortical dynamics may therefore provide the tools necessary to functionally dissect distributed cortical systems.
Neural dynamics in cortical systems, however, strongly depend on the computations distributed networks are tasked to perform (Buschman and Miller, 2007; Meyers et al., 2018). For example, in categorization tasks, activity patterns encoding categories can either emerge first in parietal neurons and later in prefrontal neurons (Swaminathan and Freedman, 2012), or vice versa (Goodwin et al., 2012), depending on the task used. This suggests that cognitive control is not a unitary or static process. Rather, prefrontal networks appear to dynamically reconfigure to support cognitive control under different task demands. To understand the contributions of prefrontal and parietal neurons to forms of cognitive control that are specifically disrupted in neuropsychiatric populations, we have translated a cognitive control task that measures deficits in schizophrenia (Jones et al., 2010) to nonhuman primates (Blackman et al., 2013, 2016; Kummerfeld et al., 2020). In this task, the response to a probe stimulus is contingent on a preceding cue stored in working memory. On some trials, the cue countermands a habitual response to the probe (proactive control). On others, the probe countermands a habitual response associated with the cue (reactive control). Here we characterize how computations that support specific forms of proactive and reactive cognitive control recruited by the AX continuous performance task (AX-CPT) are mediated by neural activity patterns and circuit dynamics in the prefrontal–parietal network.
Materials and Methods
Subjects
We recorded neural activity in the dorsolateral PFC (Brodmann area 46) and posterior parietal cortex (Brodmann area 7a) of 2 male monkeys (8-11 kg) (see Fig. 1G,H) performing the dot-pattern expectancy (DPX) task (Jones et al., 2010; Blackman et al., 2013) (see Fig. 1A–D). All animal care and experimental procedures complied with National Institutes of Health guidelines and were approved by the Animal Care and Use Committee at the University of Minnesota and Minneapolis Veterans Administration Medical Center.
Experimental design and statistical analysis
The DPX Task
Monkeys sat in a primate chair 77.5 cm in front of a back-projection screen. Visual stimuli were back projected by an LCD projector (Dell 5100MP DLP) onto a translucent screen. Eye position was recorded at 60 Hz using an infrared eye tracking system (ISCAN). Monkeys initiated each trial by acquiring and maintaining gaze fixation (within 3.3°) of a central target. Central gaze fixation was required throughout the trial up to the delivery of reward. After 500 ms of fixation, a cue stimulus, consisting of a pattern of dots, was presented for 1.0 s at the center of the display. One cue dot pattern was designated the A-cue, whereas five alternative cue dot patterns were collectively designated B-cues (see Fig. 1C). After the offset of the cue, a 1 s delay period followed, after which a probe stimulus, also consisting of a pattern of dots, was presented for 0.5 s at the center of the display. One probe dot pattern was designated the X-probe, whereas five probe dot patterns were collectively designated Y-probes (see Fig. 1D). Cue and probe stimuli subtended 2.7°-4.4° of visual angle. The motor response required by the task was a right or left joystick movement, which monkeys made with their right hand, while maintaining central gaze fixation. Responses made in the correct direction within a 1.5 s response window starting at the onset of the probe stimulus were rewarded with a drop of juice (∼0.2 ml). The AX cue-probe sequence indicated a target trial requiring a leftward joystick movement. All other cue-probe combinations indicated nontarget trials requiring a rightward joystick movement.
The DPX task was administered in two trial sets. In “balanced” trial sets, the four trial types defined by cue-probe sequence (AX, AY, BX, BY) were presented in equal proportion. Balanced trial sets included 80 or 300 trials total (in 38 and 4 neural ensembles, respectively). In “prepotent” trial sets, most of the trials were target trials presenting the AX cue-probe sequence (69%), whereas the remaining minority of trials (31%) were nontarget trials presenting the remaining cue-probe combinations (AY 12.5%, BX 12.5%, BY 6%). Prepotent trial sets replicate the proportion of trial types in the DPX task and AX-CPT as these tasks have typically been administered to patients (MacDonald, 2008; Jones et al., 2010). Prepotent trial sets included 301 or 400 trials total (in 11 and 29 ensembles, respectively; two ensembles were recorded on balanced trial sets only). In prepotent trial sets, AX trials outnumbered BX trials ∼5:1. Consequently, presentation of the X-probe was associated with a prepotent target response, countermanded by the B-cue on the minority of trials it was presented.
Monkeys received an intramuscular injection of saline (0.41-0.89 ml) before recording neurophysiological data for 14 of the 42 neural ensembles included in this study as part of a study investigating the effects of the NMDA antagonist phencyclidine on neural activity in PFC (Zick et al., 2018). (All neural data included in the present study were recorded before monkeys were first exposed to phencyclidine during recordings.)
Neural recording
3T structural MR images of monkeys were obtained to localize target recording regions to the banks of the principal sulcus in PFC (Brodmann area 46) and the inferior parietal gyrus in posterior parietal cortex (Brodmann area 7a) (see Fig. 1G,H). To prepare monkeys for acute neural recording, they underwent an aseptic surgical procedure under isoflurane gas anesthesia (1%-2%). Craniotomies were made in the skull overlying the dorsolateral prefrontal and posterior parietal cortex in the left cerebral hemisphere. Five titanium posts were fixed to the surface of the skull using titanium screws. Plastic recording chambers (13 mm i.d.) were set in place over the craniotomies and fixed to the screws and posts using surgical bone cement. A halo for head stabilization was attached to the posts using metal tabs. Analgesia was maintained for several days after surgery (Buprenex, 0.05 mg/kg, twice a day). Neurophysiological recordings were obtained using dual 16-microelectrode motorized drives (Thomas Recording) that independently advanced glass-coated platinum iridium microelectrodes (70 µm o.d.) through the dura and into the PFC (16 microelectrodes) and posterior parietal cortex (16 microelectrodes) simultaneously. Recorded signals were amplified (gain: 2500×) and bandpass filtered (between 2 and 5 kHz). The action potentials of individual neurons were isolated online using a spike waveform template-matching algorithm (Alpha Omega Engineering) and time-amplitude window discriminators (BAK Electronics). Typically, the action potentials of ∼20 prefrontal neurons and ∼20 posterior parietal neurons were isolated simultaneously. Digital output pulses from the waveform template-matching system indicating detected action potentials were sampled with 40 µs resolution (DAP 5200a Data Acquisition Processor; Microstar Laboratories). The time of isolated action potentials, as well as the timing of stimulus onset and offset, and also digitized horizontal and vertical eye and joystick position were saved to disk. One to four of the 16 electrodes were used to record local field potential (LFP) data concurrently with spiking data. LFPs were sampled at 2000 Hz, bandpass filtered (1-100 Hz), notch filtered (60 Hz), demeaned, and resampled (400 Hz) using the ft_preprocessing function in the FieldTrip open source Matlab toolbox (Oostenveld et al., 2011). An LFP channel was rejected if artifacts were present on visual inspection, or if it contained strong residual 60 Hz noise after notch filtering. Individual trials were rejected if the signal saturated or showed abrupt discontinuities related to the motor response and/or reward. Channels were also rejected if they did not contain at least 200 trials after trial-level rejection.
Data analysis
ANCOVA-based classification of neural activity
To identify task-responsive neurons, we applied ANCOVA to firing rates measured each trial in the fixation, cue, delay, probe, and response periods. (The response period was defined as ±200 ms relative to the time of the joystick movement.) Trials from prepotent and balanced trial sets were combined. We coded the cue (A or B), probe (X or Y), and response (target or nontarget) as binary factors in the analysis. We applied one-way ANCOVAs to firing rates in the cue and delay periods using the cue as the single factor. We applied a two-way ANCOVA to firing rates in the probe period using the cue and probe as factors. Firing rates during the fixation period were treated as a covariate in these analyses. We conducted a one-way ANCOVA using firing rates in the response period, and response direction as the single factor. Firing rates in the delay period served as the covariate for this analysis. To identify neurons with selective responses to specific dot patterns making up B-cue (see Fig. 1C) and Y-probe (see Fig. 1D) stimuli, we conducted one-way ANCOVAs on firing rates during the cue and probe periods on the subset of trials in which B-cues and Y-probes were presented. We used the specific dot pattern presented as the single factor in these analyses (5 levels corresponding to the 5 B-cues and Y-probes).
Neurons were classified based on the trial periods and task factors that influenced firing rate (p < 0.05), in combination with the stimulus and response preferences of these neurons. Early B-cue and Delay B-cue neurons exhibited activity that was modulated by the cue stimulus with activity that was greater on B-cue than A-cue trials during the cue and delay periods, respectively. Late A-cue neurons exhibited activity that was modulated by the cue during the probe period with activity that was greater on A-cue than B-cue trials. Nontarget Response neurons exhibited activity that was greater on nontarget than target trials during the response period. Finally, Nonselective Response neurons exhibited activity during the response period that was greater than activity during the delay period (paired t test, p < 0.05), but were not selective for response direction by ANCOVA. This classification scheme is a simplified version of the one adopted in our prior report (Blackman et al., 2016) that captures the strongest and most prevalent population activity patterns related to DPX task performance.
Sliding-window regression analysis
To evaluate the strength of the relationship between neuronal activity and task variables as a function of time in the trial, we measured firing rates of single neurons on correct trials of prepotent trial sets within a sliding 100 ms window (20 ms steps), and at each time step regressed the trial-by-trial firing rates onto the cue, probe, and response by fitting the following linear models.
In which R is firing rate, Cue (A vs B), Probe (X vs Y), and Response (target vs nontarget) are dummy-coded independent variables, and E is the error (residual). Neurons were ranked according to the magnitude of the peak significant regression coefficient, and an equal number of neurons with the lowest p values were included from each cortical area in the analysis (adjusted to not exceed the number of neurons in the cortical area with the smaller number of neurons). The results of the regression analysis were then expressed as the proportion of explainable variance (PEV) (Brincat and Miller, 2016) associated with each regressor, computed using the formula of Olejnik and Algina (2003) as follows:
Decoding analysis
We applied time-resolved pattern classification analysis (Klecka, 1980; Johson and Wichern, 2008; Crowe et al., 2010) to spike counts measured within a sequence of 50 ms time bins over the population of recorded neurons in PFC and parietal cortex. In separate analyses, we decoded the value of the cue (A or B), probe (X or Y), and response (target, left, or nontarget, right) as binary variables from the time series of spike counts. At each time step, we performed the classification using spike counts in a sliding window of 3 consecutive 50 ms time bins (150 ms width). We applied decoding both to spike counts of simultaneously recorded neurons (ensembles) and also neurons aggregated over days of recording (populations). For population decoding, we constructed pseudo-trials of population activity by matching trials of neural activity recorded on different days according to the repetition number of the decoded variable. (For example, when decoding the cue, we aggregated all trials having the same repetition number of the A- or B-cue in the randomized trial sequence into a single trial of population activity.) Population decoding accuracy can strongly depend on the numbers of neurons included in the population and the task-defined signals they exhibit. To compare neural representation in PFC and parietal cortex in an unbiased manner, we constructed populations of prefrontal and parietal neurons that contained equal numbers of the most “significant” neurons in each of the five primary response categories from the two areas as indicated by the p value obtained in the ANOVA (above). To carry out the decoding, we used the classify MATLAB function using the “diaglinear” discriminant type, and “empirical” priors. When decoding behavioral variables from each time bin, we trained the classifier on activity patterns in the same time bin as we advanced through the trial, using leave-one-out cross validation (removing the test trial being decoded from the training data) at each time step.
In addition, to determine whether neural representations of cue and probe stimuli were categorical or feature-selective, we performed analyses taking advantage of the fact that multiple visual patterns collectively signified B-cues and Y-probes (see Fig. 1C,D). We decoded the cue (A vs B) as before, but on B-cue trials, we deleted trials with the same B-cue stimulus as the test trial from the training data. This ensured that population responses to individual B-cues did not contribute to the ability to discriminate A from B cues. Similarly, we decoded the probe (X vs Y) as before, but on Y-probe trials, we deleted trials with the same Y-probe as the test trial from the training data. This similarly ensured that population responses to individual Y-probes did not contribute to the ability to discriminate X from Y probes. Successful decoding in these cases would provide evidence that the neural representation of stimuli was categorical rather than feature-selective in nature, because the neural representation of cues and probes generalized over the different B-cue and Y-probe stimuli presented. Conversely, if it proved that removal of training trials with the same visual stimulus as the test trial strongly degraded decoding performance, this would provide evidence that neural representations of stimuli were feature-selective. As a further test of feature selectivity, we decoded the specific B-cue dot pattern presented on B-cue trials, and the specific Y-probe dot pattern presented on Y-probe trials.
Signal correlation analysis
We performed a signal correlation analysis (Crowe et al., 2013) measuring the degree to which fluctuations in the strength of neural representations of specific task variables covaried over time in PFC and parietal cortex at different lags. The analysis was based on ensembles of simultaneously recorded neurons in the two areas. We first decoded the cue, probe, and response from single trials of ensemble activity patterns measured in a sequence of 50 ms time bins (using neurons in which firing rate varied significantly in relation to the decoded variable). This converted a sequence of ensemble activity patterns in PFC and parietal cortex into a corresponding sequence of single trial posterior probabilities. The posterior probability time series captured variation in the strength of neural representation of the same behavioral variable in the two cortical areas. We then sought to determine the degree to which fluctuations in coding strength over time covaried between cortical areas. Estimates of the strength of the relationship between time series can be inflated by autocorrelation structure in the time series (Georgopoulos et al., 2007). To account for this, we fit autoregressive integrated moving average (ARIMA) models (of order [10, 2, 2]) to the posterior probability time series. The residuals from the fits isolated the proportion of variance in the probability time series that could not be explained by their own prior histories (autocorrelation structure), allowing us to isolate the proportion of variance in the probability time series in each cortical area that might reflect external input from the other. To evaluate signal covariation between PFC and parietal cortex as a function of time in the trial, we regressed the residual posterior probabilities in one cortical area onto the other within a sliding window. (The window was 500 ms in width, spanned 10 residual probabilities per trial, and was advanced in 50 ms steps.) We aggregated probabilities within the window over trials to perform the regression at each time step. To evaluate simultaneous signal correlation, we conducted the regression with the residual probability time series aligned in the two areas. To evaluate top-down signal correlation, we regressed the parietal data onto the prefrontal data shifted one 50 ms time bin earlier. To evaluate bottom-up signal correlation, we regressed the prefrontal data onto the parietal data shifted one 50 ms time bin earlier. The analysis produced a time series of F statistics quantifying the strength of the relationships between fluctuations in neural representations in the two cortical areas. The F statistic at each time step was considered significant if it exceeded the 95th percentile (p < 0.05, false-discovery rate [FDR]-corrected over the time series) of a bootstrap distribution of F statistics obtained by trial-shuffling the probability time series between prefrontal and parietal neurons and repeating the sliding window regression analysis using the trial-shuffled data (1000 iterations). Trial-shuffling broke the simultaneity of the neural signals in PFC and parietal cortex but otherwise left the underlying probability time series intact. Signal correlation in excess of that present in the trial shuffled data were likely to reflect real-time neural interactions, between prefrontal and parietal neurons directly or between neurons in these areas and other neural structures. To measure transmission of cue and probe information, we included Early B-cue and Late A-cue neurons that were simultaneously recorded in PFC and parietal cortex in the decoding analysis (see Fig. 2A,B). To measure transmission of response information, we included Delay B-cue and Nontarget Response neurons in the decoding analysis (see Fig. 2C,D).
LFP time-frequency analysis
To isolate oscillations in LFPs that reflect internal brain dynamics and not deflections in the LFP that were entrained to external events, we subtracted the trial-averaged LFP (event-related potential) at each recording site from each single trial of LFP data. We then convolved the resulting single trial LFP data with a set of 49 complex Morlet wavelets with peak frequencies ranging from 1 to 100 Hz, and 3 (at 1 Hz) to 10 (at 100 Hz) cycles per wavelet. Peak frequencies and cycles per wavelet spanned their ranges logarithmically. We used the ft_freqanalysis function in the open source FieldTrip MATLAB toolbox (Oostenveld et al., 2011) to perform the convolution. At each time point in the LFP data, the function computed the dot product between the 49 complex wavelets (sampled at 400 Hz) and a segment of the filtered, demeaned, and resampled LFP time series of equal duration. The resulting time series of complex dot products quantified the power (and phase) of oscillatory components at each wavelet frequency that were present in the LFP data at that time point (Cohen, 2014). In general, power relates to frequency inversely as 1/f in LFP recordings. To offset that scaling, on each trial, we divided the power value at each time point by the baseline power averaged from −400 to −100 ms before onset of the cue stimulus during the preceding gaze fixation period. We averaged single-trial spectral data over recording sites in each cortical area, and trials with the same cue-probe sequence, to produce time-frequency surfaces (see Fig. 5).
LFP decoding analysis
We decoded task variables (cue, probe, and response) from single trial LFP data. For each trial of LFP data, we computed mean power in a sequence of 50 ms time bins in each of five frequency bands (δ, 1-4 Hz; theta, 4.1-8 Hz; α, 8.1-15 Hz; β, 15.1-35 Hz; and γ, 35.1-100 Hz). We then constructed pseudo-trials of binned mean LFP power data by matching trials over recording days based on trial repetition number of the decoded variable. Decoding was otherwise performed by applying time resolved pattern classification to pseudo-trials of population LFP data as described above for pseudo-trials of population spike count data.
Results
Behavioral performance
Monkey 1 performed 95% of DPX trials correctly, and Monkey 2 performed 97% of DPX trials correctly. Both monkeys performed >90% of each of the four trial types defined by the four possible cue-probe combinations correctly (Fig. 1E). Response time (RT) varied across trial types (Fig. 1F), both in Monkey 1 (Kruskal–Wallis test; χ2 = 5715.14, df = 3, p < 0.0001), and Monkey 2 (χ2 = 178.77, df = 3, p < 0.0001). Post hoc testing indicated that RT on AY trials was longer than RT on all other trial types in Monkey 1 (multiple comparison of mean ranks, p < 0.01), and was longer than RT on BX and BY trials in Monkey 2 (p < 0.01). On AY trials in the prepotent set, the A-cue established the expectancy that an X-probe would follow requiring a target response. When the relatively infrequent Y-probe followed the A-cue on AY trials, this expectation had to be overridden to produce the required nontarget response, a process associated with an increased RT.
DPX task, performance, recording locations, and distribution of task-related neural responses between prefrontal and parietal cortex. A, Event sequence on a target DPX trial presenting the A-cue followed by the X-probe, requiring a leftward joystick movement. Correct responses made within a 1.5 s response window starting at probe onset were rewarded (Resp win in B). B, Event sequence on a nontarget DPX trial presenting a B-cue followed by the X-probe, requiring a rightward joystick movement. C, Cue stimuli. One dot pattern constituted the A-cue, five other dot patterns collectively constituted B-cues. D, Probe stimuli. One dot pattern constituted the X-probe; five other dot patterns collectively constituted Y-probes. E, Mean (± 2 SEM) proportion of trials performed correctly by Monkey 1 (filled circles, solid line) and Monkey 2 (open circles, dashed line) separated by trial type (cue-probe sequence). F, Mean (± 2 SEM) reaction time of Monkey 1 (filled circles, solid line) and Monkey 2 (open circles, dashed line) separated by trial type (cue-probe sequence). G, H, Recording locations (blue regions) in PFC and parietal cortex in Monkey 1 (G) and Monkey 2 (H). PS, Principal sulcus; IPS, intraparietal sulcus. I, J, Venn diagrams represent the numbers of neurons in prefrontal (I) and parietal (J) cortex in which firing rate varied as a function of the cue (A vs B), probe (X vs Y), and/or response (target vs nontarget) in an ANCOVA (p < 0.05).
Neural database
We recorded the activity of 1491 neurons during task performance. In PFC, we recorded the activity of 993 neurons, in 42 neural ensembles (Monkey 1: 32 ensembles and 740 neurons; Monkey 2: 10 ensembles and 253 neurons). Prefrontal ensembles contained 23 individual neurons on average (range 9-40 neurons). In posterior parietal cortex, we recorded the activity of a total of 498 neurons, in 25 ensembles (Monkey 1: 19 parietal ensembles and 382 neurons; Monkey 2 6 parietal ensembles and 116 neurons). Parietal ensembles contained 20 individual neurons on average (range 8-32 neurons). Prefrontal neurons were located within the banks of the principal sulcus in Brodmann area 46 (Fig. 1G,H). Posterior parietal neurons were located in Brodmann area 7a in the inferior parietal gyrus (Fig. 1G,H). Between one and four LFP recordings were made in each area concurrently with single-unit recordings. We obtained usable LFP data from 53 recording sites in parietal cortex and from 61 recording sites in PFC.
Comparing the types of neural signals evoked by the DPX task in prefrontal and parietal neurons
Venn diagrams illustrate the numbers of prefrontal (Fig. 1I) and posterior parietal (Fig. 1J) neurons with a minimum firing rate of 0.5 Hz through the trial and in which mean firing rate differed (p < 0.05) in relation to the cue (A vs B), probe (X vs Y), or response (target vs nontarget) in the ANCOVA. Cue-selective activity was most common among task-related neurons (87% of prefrontal and 91% of parietal task-related neurons). Probe-selective activity was notably less common among task-related neurons (32% of prefrontal and 41% of parietal task-related neurons). Response-selective activity was also less common (53% of prefrontal and 47% of parietal task-related neurons). These trends were present in both cortical areas (Fig. 1I,J).
We classified neurons (Fig. 2) according to the timing of significant modulations in firing rate (based on the ANCOVA) in cue, delay, probe, and response periods, coupled with each neuron's preference for specific cues (A or B), probes (X or Y), and responses (target or nontarget). Early B-cue neurons exhibited activity during the cue period that was greater on B than A-cue trials. (Fig. 2A). Late A-cue neurons exhibited activity during the probe period that was greater on A- than B-cue trials (Fig. 2B). The two neural populations exemplified a “switch” neuron pattern (Blackman et al., 2016) by switching their stimulus preference from B-cues during the cue period to A-cues in the probe period (Fig. 2A,B). In Early B-cue neurons, activity modulation based on the cue was strongest in the cue period, whereas in Late A-cue neurons, it was strongest during the probe period. In both neural populations, presentation of the Y-probe following the A-cue was associated with further increase in firing rate (Fig. 2A,B). Presentation of the B-cue is associated with proactive cognitive control (inhibition of a prepotent response to a future stimulus). Presentation of the Y-probe following the A-cue is associated with reactive cognitive control (inhibition of a prepotent response associated with a prior stimulus). Thus, the activity of switch neurons appears to reflect both proactive and reactive aspects of cognitive control. Delay B-cue neurons (Fig. 2C) exhibited sustained delay period activity that was greater on B-cue trials, accompanied by an additional increase in activity around the time of the motor response that was particularly prominent in parietal cortex (top row). “Nontarget response” neurons exhibited activity time-locked to the motor response that was stronger on nontarget than target trials (Fig. 2D). This activity was comparatively weak in parietal cortex (Fig. 2D, top). “Nonselective response” neurons exhibited response activity that did not vary as a function of response direction (Fig. 2E). Population activity patterns bore striking resemblance in parietal (Fig. 2A–E, top row) and prefrontal (Fig. 2A–E, bottom row) cortex, with some differences. Nontarget response activity was particularly robust and more frequently combined with delay period activity in parietal cortex (Fig. 2C, top), whereas nontarget response activity was more frequently present without delay period activity in PFC (Fig. 2D, bottom).
Population activity patterns in PFC and parietal cortex. Population spike density functions (SDFs) (δ = 40 ms) show average population activity on subsets of trials defined by cue-probe sequence (gold represents AX; red represents AY; purple represents BX; blue represents BY). Neurons were divided into five functional groups based on the results of the ANCOVA applied to firing rates in the cue, delay, probe, and response periods. Activity in Early B-cue neurons varied as a function of the cue during the cue period and was higher on B-cue trials. Activity in Late A-cue neurons varied as a function of the cue during the probe period and was higher on A-cue trials. Activity in Delay B-cue neurons varied as a function of the cue during the delay period and was higher on B-cue trials. Activity in Nontarget Response neurons varied as a function of response direction during the response period (response movement onset ±200 ms) and was higher on nontarget response trials. Activity in Nonselective Response neurons was significantly elevated in the response period compared with the delay period but did not vary as a function of response direction. (The parietal activation on BX trials, E, top, combined with activity on BY and AY trials, all requiring the nontarget response, did not yield an overall significant difference in activity on target vs nontarget trials.) The number of neurons meeting these criteria and contributing to each population is indicated in the corresponding panel. A–E, Population activity in parietal (top row) and prefrontal (bottom row) cortex aligned to cue onset in (A) Early B-cue, (B) Late A-cue, (C) Delay B-cue, (D) Nontarget response, and (E) Nonselective response neurons. F, G, Population activity in parietal (top row) and prefrontal (bottom row) cortex realigned to the time of the initiation of the motor response (vertical line) in Early B-cue and Late A-cue neurons combined (F), and in Delay B-cue and Nontarget Response neurons combined (G). H, Bars indicate the proportion of Early cue neurons (selective for cues during the cue period) and Late cue neurons (selective for cues during the probe period) that preferred A-cues and B-cues in parietal cortex (light blue) and PFC (dark blue).
Alignment of population activity to the time of the motor response revealed two successive peaks in population activity occurring near the end of the trial (Fig. 2F,G). In “switch” neurons (Early B-cue and Late A-cue neurons combined), population firing rate on A-cue trials peaked during the probe period ∼300 ms before the time of the motor response in both parietal cortex and PFC (Fig. 2F). Activity in these neurons had returned to near baseline levels by the time that the response was initiated, suggesting their activity was more closely related to computations for cognitive control triggered by the probe stimulus than the response itself. In neurons with response-related activity (Delay B-cue and Nontarget Response neurons combined), in contrast, population activity peaked at the time that the response was initiated in both cortical areas (Fig. 2G), although more prominently in parietal cortex (top). Most neurons encoding the identity of the cue during the cue period preferred the B-cue (Fig. 2H, Early cue), whereas most neurons encoding the identity of the cue during the probe period preferred the A-cue (Fig. 2H, Late cue), consistent with the switch pattern. These biases in cue preference in the cue and probe periods were evident and comparable in both the prefrontal and the parietal cortex (Fig. 2H). However, these biases were not absolute in the sense that there were a minority of neurons in which firing rate varied significantly in relation to the cue but with a stimulus preference that was opposite to the predominant “switch” pattern (i.e., Early cue neurons with A-cue preference, and Late cue neurons with B-cue preference; Fig. 2H).
Comparison of single neuron responses in PFC and parietal cortex
To compare the strength and timing of firing rate modulation in single prefrontal and parietal neurons throughout the trial, we regressed the firing rate of each neuron onto the cue (A or B), probe (X or Y), and response (target or nontarget) within a sliding 100 ms window advanced in 20 ms steps. We included all task-related neurons assigned to 1 of the 5 functional groups (Fig. 2), and used neural activity recorded in prepotent trial sets. Each row of the heat maps (Fig. 2A,D,G) illustrates regression results from a single neuron, either in parietal cortex (left column) or PFC (right column), as a function of time in the trial. Warmer colors represent a greater proportion of explainable variance in firing rate (PEV) by each predictor variable (cue, probe, and response). Neurons were sorted according to the magnitude of the peak PEV, and equal numbers of neurons with the strongest responses were included from PFC and parietal cortex. Neurons were then ranked according to the time to peak PEV in the plots. Bands of color indicate recruitment of neurons to encode each variable throughout the trial. Periods of peak PEV were typically brief and staggered in time across the population, suggestive of a dynamic, sequential recruitment of neurons (Crowe et al., 2010). Some neurons exhibited more persistent periods of encoding (indicated by longer periods of elevated PEV). Such periods of persistent encoding were more apparent in the PEV attributable to the cue (Fig. 3A), compared with the probe or response (Fig. 3D,G). To compare the timing of neural recruitment across areas, we compared cumulative distributions of the time to peak PEV attributable to the cue (Fig. 3B), probe (Fig. 3E), and response (Fig. 3H). Cue signals emerged significantly earlier in parietal cortex (Fig. 3B, light blue) than PFC (dark blue; Kolmogorov–Smirnov test, p < 0.01, n = 180 parietal and prefrontal neurons). Probe signals also emerged significantly earlier in parietal cortex (Fig. 3E, light blue) compared with PFC (dark blue; Kolmogorov–Smirnov test, p < 0.001, n = 120 parietal and prefrontal neurons). Time to peak PEV attributable to the response (target vs nontarget) did not significantly differ between cortical areas (Fig. 3H, Kolmogorov–Smirnov test, p = 0.96, n = 100 parietal and prefrontal neurons).
The proportion of explainable variance (PEV) in firing rate attributable to the cue, probe, and response. A, D, G, Color in the heat maps represents the PEV attributable to the (A) cue, (D) probe, and (G) response obtained by regressing the firing rate of single neurons within a sliding 100 ms window advanced in 20 ms steps onto each task factor. Left column, parietal data. Right column, prefrontal data. Equal numbers of the most “significant” neurons (with lowest p values) were included in each cortical area. B, E, H, Cumulative distributions of time to peak PEV attributable to the (B) cue, (E) probe, and (H) response in parietal neurons (light blue) and prefrontal neurons (dark blue). C, F, I, Time courses of population average PEV attributable to the (C) cue, (F) probe, and (I) response in parietal neurons (light blue) and prefrontal neurons (dark blue).
Comparison of population decoding in PFC and parietal cortex
To compare representation of task variables by population activity in PFC and parietal cortex, we applied a decoding analysis to firing rates measured in a sliding window advanced through the trial. At each time point, the analysis computed the trial-averaged posterior probability reflecting the certainty with which population activity patterns represented the correct value of the cue, probe, and response (based on stimuli presented and the response made) each trial. Neural signals representing the cue (A vs B) increased in parallel in PFC and parietal cortex shortly after cue onset but persisted at significantly higher levels through the delay into the probe period in PFC (Fig. 4A, red dots). The difference between cortical areas was particularly prominent on the subset of trials in the analysis in which the B-cue was presented (Fig. 4D). Neural signals representing the probe (X vs Y), in contrast, increased shortly after probe onset and decreased shortly after probe offset with near overlapping time courses in PFC and parietal cortex (Fig. 4B,E). Neural signals representing the response direction (left, or target, vs right, or nontarget) increased shortly after cue onset in both cortical areas, reflecting partial information about response direction afforded by the cue, then increased further at the time of the response after presentation of the probe determined response direction with certainty (Fig. 4C,F). In PFC, information about the forthcoming response direction was maintained at a higher level during the delay and probe periods leading up to the response than in parietal cortex (Fig. 4C,F, red dots). Thus, PFC carried stronger information about the prior cue and forthcoming response than parietal cortex, whereas information about the probe was comparably strong in the two cortical areas.
Cue, probe, and response decoding in PFC and parietal cortex. Functions plot the trial-averaged posterior probability associated with the neural representation of the cue (A or B), probe (X or Y), and response (target, left, or nontarget, right), each trial derived by a decoding analysis applied to patterns of population activity in prefrontal (dark blue) and parietal (light blue) cortex. The same two neural populations of equal size in prefrontal (n = 336 neurons) and parietal (n = 336 neurons) cortex were used for all decoding analyses illustrated (A–L). Prefrontal and parietal populations were constructed by including equal numbers the most “significant” neurons (lower p value) in each of the five functional groups in each cortical area (Early B-cue, Late A-cue, Delay B-cue, Nontarget Resp, and Nonselective Resp neurons). Firing rates were measured in a sliding window and the decoding analysis applied to activity patterns observed at each time step. Differences in the trial-averaged posterior probability between cortical areas were deemed significant (as indicated by red circles) if they exceeded the 95th percentile of a bootstrap distribution of differences between areas obtained after randomizing firing rates between cortical areas and repeating the decoding analysis (p < 0.05, FDR-corrected). Horizontal dashed lines indicate the prior probability (and hence chance decoding) of the cue, probe, and response based on their frequencies in the trial set. Deviations from the prior probability reflect additional information conveyed by neural activity patterns. A–C, Decoding time courses plot the mean posterior probability associated with the cue (A; A vs B), probe (B; X vs Y), and response (C; left vs right) averaged over all trials. Neural data were recorded on prepotent trial sets. D–F, Decoding time courses plot the mean posterior probability of the cue, probe, and response averaged over the subset of trials in prepotent trial sets with greater cognitive control demand (B-cue trials, Y-probe trials, and nontarget responses, in D–F, respectively). G–I, Influence of stimulus features and cognitive control demand on cue decoding. G, Decoding time courses plot the mean posterior probability associated with the cue (A vs B) obtained when trials with the same B-cue dot pattern as the one displayed each trial were removed from the training data (cross validate Bs). Neural data were recorded on prepotent trial sets. H, Same as in G, but using neural data recorded on balanced trial sets. I, Decoding time courses plot the mean posterior probability associated with the specific B-cue stimulus dot pattern displayed each trial. Neural data recorded on prepotent and balanced trial sets combined. J–L, Influence of stimulus features and cognitive control demand on probe decoding. J, Decoding time courses plot the mean posterior probability associated with the probe (X vs Y) obtained when trials with the same Y-probe dot pattern as the one displayed each trial were removed from the training data (cross validate Ys). Neural data were recorded on prepotent trial sets. K, Same as in J, but using neural data recorded on balanced trial sets. L, Decoding time courses plot the mean posterior probability associated with the specific Y-probe stimulus dot pattern displayed each trial. Neural data recorded on prepotent and balanced trial sets combined.
Neural population activity patterns differentiating A- from B-cues, or X- from Y-probes, may have represented these stimuli in a categorical manner, according to their significance to cognitive control (i.e., whether stimuli required overriding a prepotent response or not). In this case, activation of switch neurons (Fig. 2A,B) by B-cues during the cue period would reflect engagement of proactive control, and by Y-probes (following A-cues) during the probe period engagement of reactive cognitive control. Alternatively, population activity patterns differentiating cues and probes may have reflected neural selectivity for the various dot patterns making up these stimuli. Preferential activation by B-cues and Y-probes at the population level would in this case reflect the fact that multiple distinct dot patterns comprised B-cues and Y-probes, increasing the chances on these trials that individual neurons with feature-selective responses to individual dot patterns would be preferentially activated, driving population firing rates to higher levels on B-cue and Y-probe trials. To differentiate between these possibilities, we analyzed the degree to which population activity patterns generalized across the various stimuli comprising B-cues and Y-probes (consistent with a categorical representation), or varied over these dot patterns (consistent with feature-selective representation). Both cue decoding (Fig. 4G, compare Fig. 4D) and probe decoding (Fig. 4J, compare Fig. 4E) proved to be robust against the removal of the specific dot pattern shown each trial from the data used to train the classifier (see Materials and Methods). This provided evidence that decoding did not depend on neural responses to specific dot patterns shown as B-cue and Y-probe stimuli. In addition, decoding was at chance levels when attempting to decode which specific dot pattern was presented as the B-cue (Fig. 4I) or the Y-probe (Fig. 4L) directly. These decoding results were supported by the observation that individual dot patterns rarely influenced single neuron firing rates. Whereas 651 of the 1491 neurons recorded exhibited activity that differentiated between A- and B-cues during the cue period (∼40% of the neural sample), only 91 neurons exhibited activity during the cue period that differentiated between the B-cue stimuli considered individually (∼6% of neurons, near the expected Type I error rate at p < 0.05). Similarly, whereas 420 of the 1491 neurons recorded exhibited activity that differentiated between X and Y probes during the probe period (∼28% of the sample), only 108 neurons exhibited activity that differentiated between the Y-probes during the probe period considered individually (∼7% of neurons). Collectively, these findings provide convergent evidence that the neural representation of cues and probes was primarily categorical rather than feature-selective in nature.
To evaluate the impact of cognitive control demand on neural activity, we compared decoding results on prepotent and balanced trial sets. Prepotent trial sets require increased proactive control on B-cue trials, and reactive control on Y-probe trials, to override the prepotent target response, since most trials (69%) are AX requiring the target response. The markedly stronger representation of cues in PFC that was evident on prepotent trial sets (Fig. 4G) largely disappeared on balanced trial sets (Fig. 4H). The neural representation of Y-probes, in contrast, was comparably strong in PFC and parietal cortex regardless of whether neural activity was recorded on prepotent (Fig. 4J) or balanced (Fig. 4K) trial sets. These observations suggest that increasing proactive, but not reactive, cognitive control demand biases processing toward the PFC.
LFPs encode task information in PFC and parietal cortex
To relate computations for cognitive control to oscillatory dynamics in the prefrontal–parietal network, we applied a Morlet wavelet spectral analysis to LFPs. Separating power spectral density by trial type (cue-probe sequence) revealed that A- and B-cues evoked distinct time-frequency patterns of power modulation in PFC and parietal cortex (Fig. 5). In parietal cortex, A-cue trials were associated with increases in power in the delta, theta, and alpha ranges during the cue and delay periods (Fig. 5A,B, region 1) that was terminated by strong desynchronization in these frequencies during the probe period (Fig. 5A,B, region 2, dark blue). The period of desynchronization approximately corresponded to the prominent rise in population firing rate that occurred during the probe period on A-cue trials (Fig. 5A,B, light blue spike density function [SDF]). A modest increase in beta power was also evident during the latter half of the delay period (Fig. 5A,B). On B-cue trials, the increase in beta power was more prominent, and distinct periods of elevated β and alpha power were seen extending through the delay period (Fig. 5C,D; region 3). A modest increase in beta power (Fig. 5A–D; region 4) and a more pronounced increase in gamma power (Fig. 5A–D, region 5) were evident around the time of the response that did not clearly vary appreciably as a function of trial type (Fig. 5A–D).
Local field potentials (LFPs) encode task variables in parietal cortex and PFC. Power spectral density of LFPs separated by cue-probe sequence (AX, AY, BX, BY). Spike density functions (light blue) are overlaid to illustrate average changes in the firing rate of neurons that were recorded concurrently with LFPs (firing rate scale at right). Numbered regions identify oscillatory features described in text. A–D, Parietal LFP power spectral densities and average population firing rates. E–H, Prefrontal LFP power spectral densities and average population firing rates. I, J, Results of a decoding analysis applied to mean LPF power values in each frequency band (delta, theta, alpha, beta, and gamma) in a sliding window. Functions plot the trial-averaged posterior probability associated with the correct cue (A vs B), and probe (X vs Y) presented in each trial. Dashed line indicates chance decoding accuracy based a priori on the frequency of cues and probes in the trial set.
In PFC, power modulations during the cue and probe periods (Fig. 5E–H) were generally modest compared with parietal cortex (Fig. 5A–D). A-cue trials were associated with an increase in alpha power (Fig. 5E,F, region 6), and B-cue trials were associated with an increase in beta power late in the delay period (Fig. 5G,H, region 7). Desynchronization during the probe period was generally weaker in PFC than parietal cortex (Fig. 5E–H, dark blue regions) but corresponded to peaks in population firing rate (Fig. 5E–H, light blue SDFs). Around the time of the response, increases in beta power (Fig. 5E–H, region 8) and gamma power (Fig. 5E–H, region 9) were evident in PFC that did not vary appreciably as a function of trial type. The beta power modulation around the response was more pronounced in PFC (Fig. 5E–H, region 8) than parietal cortex (compare Fig. 5A–D, region 4).
To measure the amount of information about task stimuli encoded by modulations of LFP power, we applied the decoding analysis to mean LFP power values measured in each frequency band within a sliding window (Fig. 5I,J). The mean posterior probability associated with the cue (A vs B; Fig. 5I) and the probe (X vs Y; Fig. 5J) peaked at values approaching 1.0 (near perfect decoding). Therefore, task stimuli could be decoded with approximately the same certainty whether LFPs (Fig. 5) or individual neuron spike rates (Fig. 4) were used as the neural activity feature vectors input to the decoding analysis.
Functional communication between PFC and parietal cortex
Parietal and prefrontal neurons exhibited in several instances remarkably similar patterns of population activity during behavior (Fig. 2). To determine whether parietal and prefrontal neurons may physiologically interact during behavior also, we measured the degree to which fluctuation in the information about specific task variables (cue, probe, and response) that were encoded by patterns of activity in ensembles of simultaneously recorded prefrontal and parietal neurons covaried over time (Crowe et al., 2013). We first measured spike rates in 50 ms time bins and then applied pattern classification to the firing rates to extract two time series of posterior probabilities that captured rapid fluctuation in the strength of the neural representation of the cue, probe, or response in prefrontal and parietal ensembles. We then fit ARIMA models to the time series to measure their autocorrelation structure and retained the residuals of the model fits to isolate the component of variation in each time series that could not be predicted by its own history. We next passed a sliding window through the residual posterior probability time series, regressing one time series onto the other within the window at each time step. We performed the regression with the two time series aligned to capture simultaneous correlations, or offset by one 50 ms time bin to capture lagged correlations. The produced time-series of F statistics associated with the regression capturing the degree to which signals encoding the cue in PFC, for example, could predict signals coding the cue in parietal cortex one time bin later (capturing top-down signal transmission), or vice versa (capturing bottom-up signal transmission). Finally, to determine whether any detected relationships between probability time series in the two cortical areas depended on the simultaneity of the underlying neural signals, we compared the results obtained with the original neural data against a permutation distribution of regression results obtained after trial-shuffling prefrontal and parietal residual probability time series to break the simultaneity of the neural signals.
Temporal fluctuations in cue (Fig. 6A), probe (Fig. 6E), and response (Fig. 6I) signals were coupled between prefrontal and parietal ensembles. The interaction between cue signals in the two cortical areas was restricted largely to the cue (and early delay) period (Fig. 6A–D), although information about the cue was sustained for longer periods of time in the decoding analysis (Fig. 4A,D,G; accurate decoding extends through the delay and into the probe period of the trial). This suggests that periods of signal communication between areas can be independent of the duration of the signals within areas. Top-down (Fig. 6B, dark blue time course) and simultaneous (Fig. 6C, green time course) cue signal F statistics clearly exceeded the FDR-corrected significance threshold based on the bootstrap distribution generated by trial-shuffling neural data between PFC and parietal cortex (Fig. 6B,C, solid black time courses). This provides evidence that signal interaction depended on the simultaneity of the underlying neural activity in PFC and parietal cortex, suggesting that signal interactions reflected real-time neural interactions, either between neurons in PFC and parietal cortex, or between these areas and other neural structures. With respect to probe signals, we detected top-down (Fig. 6F), simultaneous (Fig. 6G), and bottom-up (Fig. 6H) signal interaction that occurred in the probe and subsequent response periods. Top-down interactions (Fig. 6E,F, blue) preceded bottom-up interactions (Fig. 6E,G, green). Interactions between response signals in PFC and parietal cortex were only evident in the simultaneous analysis — lagged correlations were essentially absent (Fig. 6I–L). Therefore, lagged correlations were present in cue and probe signals but not response signals, although individual neurons and neural populations encoded the three variables with comparable strength (Figs. 3A–F, 4A–C).
Temporal correlation of neural signals encoding the cue, probe, and response between prefrontal and parietal neurons. We evaluated correlation over time in the information about the cue, probe, and response encoded by patterns of activity in simultaneously recorded prefrontal and parietal ensembles. Decoding was first applied to a time series of firing rate measurements (50 ms time bins) in simultaneously recorded prefrontal and parietal neurons that exhibited activity modulation in relation to the cue, probe, or response. This generated a prefrontal and parietal time series of posterior probabilities capturing rapid fluctuations in the strength of neural signals that specifically encoded the cue, probe, and response in each cortical area. We then removed the proportion of variance in each probability time series attributable to their independent histories (by fitting ARIMA models) and regressed the residual probabilities (potentially reflecting extrinsic input) in one cortical area onto the other within a sliding window (500 ms, 50 ms steps). Signal correlation functions plot the F statistic of the regression at each time step. The regression was performed with the prefrontal and parietal time series either aligned in time (simultaneous signal correlation; green), or shifted by one 50 ms time bin with prefrontal data leading parietal data (top-down signal correlation; dark blue), or parietal data leading prefrontal data (bottom-up signal correlation, light blue). To further isolate the component of signal correlation likely to reflect real time neural interactions, either directly between prefrontal and parietal neurons or between them other neural structures, we compared the magnitude of F statistics obtained from the original data (with concurrent trials of prefrontal and parietal activity) against a permutation distribution after trial shuffling (with trials of prefrontal and parietal activity recorded at different times). Thin dashed black lines indicate the 95th percentile of the permutation distribution of F statistics obtained from the trial shuffled data at each time step. Thin solid black lines indicate the significance threshold at p < 0.05, FDR-corrected for the number of tests in the time series. A–D, Covariation in cue signals between prefrontal and parietal cortex as indexed by the F statistic associated with the top-down (dark blue), simultaneous (green), and bottom-up (light blue) regression analyses, respectively. B–C, Comparison of the F statistic time courses from each analysis to the corresponding 95th percentile and FDR-corrected significance threshold from the corresponding permutation distribution. E–H, Corresponding data for probe signal correlation. I–L, Corresponding data for response signal correlation.
To determine whether interactions between prefrontal and parietal neurons were modulated by cognitive control demand, we compared signal correlation between prepotent trial sets, in which the AX cue-probe sequence and the target response predominated, and balanced trial sets, in which cue and probe stimuli were equiprobable. On prepotent trial sets, robust top-down, bottom-up, and simultaneous cue signal correlation was evident during the cue period (Fig. 7B). On balanced trial sets, cue signal correlation was modest during the cue period and strong simultaneous signal correlation shifted to the probe and response periods (Fig. 7A). On prepotent trial sets, probe signals exhibited robust top-down, bottom-up, and simultaneous correlation during the probe and response periods (Fig. 7F). On balanced trial sets, probe signal correlation was largely absent (Fig. 7E). Simultaneous correlation of response signals was more prominent on prepotent trials (Fig. 7J) than balanced trials (Fig. 7I). These observations indicate that signal interaction between PFC and parietal cortex increased with increasing cognitive control demand.
Influence of cognitive control demand on temporal correlation of neural signals encoding the cue, probe, and response between prefrontal and parietal neurons. Top-down (dark blue), simultaneous (green), and bottom-up (light blue) correlation of task-defined signals in prefrontal and parietal neural ensembles. A–D, Cue signal correlation on (A) balanced trial sets, (B) prepotent trial sets, (C) A-cue trials in prepotent trial sets, and (D) B-cue trials in prepotent trial sets. E–H, Probe signal correlation on (A) balanced trial sets, (B) prepotent trial sets, (C) X-probe trials in prepotent trial sets, and (D) Y-probe trials in prepotent trial sets. I–L. Response signal correlation on (A) balanced trial sets, (B) prepotent trial sets, (C) target-response trials in prepotent trial sets, and (D) nontarget response trials in prepotent trial sets.
Discussion
We sought to better understand the neural mechanisms of cognitive control in the prefrontal–parietal network, particularly in relation to forms of cognitive control that are impaired in schizophrenia (MacDonald, 2008; Jones et al., 2010; Carter et al., 2012). For that purpose, we compared neural activity in PFC and parietal cortex of monkeys performing a dot-pattern variant (Jones et al., 2010) of the AX-CPT task that measures specific (rather than generalized) cognitive control deficits in the disease (MacDonald, 2008; Jones et al., 2010; Carter et al., 2012). The task recruits both proactive cognitive control, when cues stored in working memory countermand habitual responses to subsequent probes, and reactive cognitive control, when probes countermand habitual responses associated with prior cues. Here, we characterize neural activity and dynamics in the prefrontal–parietal network related to these forms of cognitive control.
Neural correlates of cognitive control
Neural activity patterns in prefrontal and parietal neurons reflected cognitive control in ways that could not be easily attributed to stimuli or movements in the task. This included activity in “switch” neurons we previously identified in PFC (Blackman et al., 2016) that we presently report exist in posterior parietal cortex also (Fig. 2A,B). Switch neurons are so named because they exhibit preference for B- over A-cues during the cue period, but A- over B-cues during the probe period, thereby switching their cue preferences as a function of time in the trial. Activation of switch neurons by the B-cue during the cue period is associated with proactive control, the necessity to countermand a habitual response to a stimulus in the future (the probe). Activation of switch neurons by Y-probes following A-cues is associated with reactive control, the necessity to countermand a habitual response associated with a stimulus in the past (the cue). Our data suggest that both proactive and reactive control are mediated by dissociable firing rate modulations in the same population of neurons. In addition to switch neurons, other forms of neural activity in both prefrontal and parietal neurons exhibited a strong bias for infrequent events that countermanded the habitual response in the task. For example, most delay neurons selective for the cue preferred the B-cue (Fig. 2C). Most response neurons selective for the direction of the response preferred the nontarget response (Fig. 2D). In combination with the switch neurons described above (with joint B-cue and AY selectivity), a picture emerges in which the prefrontal–parietal network is predominantly activated on trials when habitual responses are countermanded by surprising (or infrequent) environmental inputs, recruiting cognitive control. Presentation of the AX, expected, cue-probe sequence minimally activated the prefrontal–parietal network. The habitual motor response required on AX trials is likely mediated by other motor circuits (possibly recruiting the basal ganglia).
Do switch neurons reflect attention?
An alternative interpretation of switch neurons is that their activity reflects control of either spatial or feature (Martinez-Trujillo and Treue, 2004) attention during the DPX task. In PFC, population activity exhibits switching dynamics analogous to those observed in the present study when attention shifts between visual hemifields according to a cognitive rule (Lennert and Martinez-Trujillo, 2013). In addition, prefrontal population activity that reflects the storage of spatial information in working memory also reflects the control of spatial attention when the two functions are independently tested (Panichello and Buschman, 2021). In parietal cortex, population activity during a spatial cognitive task correlates with the deployment of spatial attention according to the cognitive demands of the task (Chafee et al., 2007). However, it seems unlikely that switch neuron activity reflects redirection of spatial attention in the present study as the location of spatial attention was not required to shift during the DPX task (as the cue and probe stimuli remained centered on the gaze fixation target on all trials). Switch neuron activity may reflect control of feature rather than spatial attention to discriminate the dot patterns making up cue and probe stimuli. For example, ramping activity of switch neurons on A-cue and not B-cue trials could reflect the need to discriminate between probe stimuli in the former and not the latter case to select the correct response. However (as discussed below), switch neuron activity was not feature-selective, arguing against a feature attention account. It remains possible that switch neuron activity may reflect control of attention in some form, consistent with the link between cognitive neural signals in the prefrontal–parietal network and attention that has been reported in prior studies. Even in that case, however, switch neurons provide a neural correlate of cognitive control because their activity modulates in accordance with the logical demands of proactive and reactive control in the DPX task.
Parallels with human neuroimaging
We found that proactive and reactive control are combined into the activity of switch neurons. Functional imaging in humans performing the same task have identified voxels within PFC in which BOLD activity is elevated for the same combination of conditions preferred by switch neurons (B > A in the cue period, and AY > AX in the probe period) (Kwashie et al., 2022). This suggests that cognitive control may be mediated by similar computations in human and nonhuman primate PFC. In addition, human EEG studies of cortical dynamics during cognitive control in the AX-CPT have revealed a pattern of cortical activation that may be broadly analogous to the switch neuron activity we report here. These signals include early components of ERPs that are stronger on B-cue trials and late components that are stronger on A-cue trials, source localize to the PFC (Dias et al., 2003), are dampened in schizophrenia (Dias et al., 2011), and have parallels in the laminar distribution of evoked LFPs in monkey PFC (Dias et al., 2006).
Categorical coding of stimuli
Another alternative account of switch neuron activity is that it reflects neural selectivity for the features of the visual stimuli used. Population activity may have reflected a bias for B-cues over A-cues because there were more B-cue dot patterns. (Similarly, population activity may have reflected a bias for Y-probes over X-probes because there were more Y-probe dot patterns.) The larger number of B-cues and Y-probes increases the chance of evoking strong responses from feature-selective neurons. To address this possibility, we quantified the degree to which population activity differentiated among the B-cue and Y-probe stimuli shown. Neural activity in both PFC and parietal cortex conveyed little to no information about the specific dot patterns making up the visual stimuli shown. We tested this in three ways. First, only 6%-7% of all neurons recorded exhibited activity that differed significantly between the different B-cue and Y-probe stimuli presented (by ANCOVA, near the anticipated Type I error rate of the test at p < 0.05). Second, removing trials with the same dot pattern as the test trial from the training data had little impact on decoding performance either for the cue (Fig. 4G) or the probe (Fig. 4J). This is evidence that neural activity patterns generalized over the specific dot patterns shown (i.e., training on some dot patterns supported accurate decoding of neural responses to other dot patterns). Third, decoding of specific B-cue dot patterns (Fig. 4I) and Y-probe dot patterns (Fig. 4L) was near chance levels. This combination of findings argues that neural representation of A- versus B-cues was categorical in nature and not strongly influenced by the visual features of the stimuli, as has been observed in other task contexts (Goodwin et al., 2012; Nieder, 2012; Swaminathan and Freedman, 2012; Crowe et al., 2013).
Distributed neural activity in the prefrontal–parietal network
Across the spectrum of different functional classes of neurons we encountered, patterns of firing rate modulation at the population level were strikingly similar in prefrontal and parietal cortex (Fig. 2). Parallel and matching modulations of firing rate in prefrontal and parietal neurons have been observed in tasks requiring cognitive control (Goodwin et al., 2012; Crowe et al., 2013; Qi et al., 2015; Panichello and Buschman, 2021), working memory (Chafee and Goldman-Rakic, 1998, 2000; Qi et al., 2010), attention (Buschman and Miller, 2007; Katsuki et al., 2014; Meyers et al., 2018), and categorization (Swaminathan and Freedman, 2012; Zhou et al., 2021), as well as quantitative judgments about numerosity (Nieder, 2012; Viswanathan and Nieder, 2013), and proportion (Vallentin and Nieder, 2010). Direct, reciprocal axonal projections between PFC and parietal cortex (Cavada and Goldman-Rakic, 1989; Medalla and Barbas, 2006) may entrain neurons in the two areas to the same patterns of activity during behavior. If so, the sharing of neural signals throughout distributed networks may be a general principle of cortical function, given the prevalence of such connections between cortical areas (Felleman and Van Essen, 1991).
Localized neural dynamics and modulation by cognitive control demand
We applied an analysis we developed previously (Crowe et al., 2013) to measure the degree to which neural signals in PFC and parietal cortex encoding the same item of task-defined information covaried over time. Application of this analysis revealed robust functional coupling in real time between cue, probe, and response representations in PFC and parietal cortex (Figs. 6, 7). This suggests that coactive neural populations in PFC and parietal cortex functionally communicate. Generally, top-down, bottom-up, and simultaneous transmission of task-defined signals rose and fell together during task performance. Top-down interactions in cue signals (Fig. 6B) and probe signals (Fig. 6F) either predominated or preceded bottom-up interactions (compare Fig. 6D,H). Other differences between cortical areas in relation to the temporal dynamics of neural activity were evident. At the single-neuron level, responses evoked by cue and probe stimuli emerged earlier in parietal than prefrontal neurons (Fig. 3B,E). At the population level, categorical signals reflecting proactive control as instructed by B-cue were stronger and more persistent in the PFC than parietal cortex (Fig. 4A,D,G). Signals encoding the forthcoming response were stronger in PFC also (Fig. 4C,F). Finally, time-frequency patterns of oscillatory dynamics in LFPs differed between PFC and parietal cortex. For example, periods of elevated alpha and beta power during the delay period on B-cue trials were more prominent in parietal cortex (Fig. 5C,D, region 3), whereas elevated beta power around the time of the response was more prominent in PFC (Fig. 5E–H, region 8). Collectively, these findings suggest robust differences in neural dynamics between cortical areas.
Finally, we observed that cortical dynamics in the prefrontal–parietal network varied with cognitive control demand. Greater cognitive control was required to override the target response on prepotent trial sets (in which most trials were AX) than balanced trial sets (in which all cue-probe combinations were equiprobable). We observed that cue decoding was markedly stronger in PFC than parietal cortex on prepotent trial sets (Fig. 4G) but not balanced trial sets (Fig. 4H). This suggests that proactive cognitive control selectively recruits PFC. Probe decoding, in contrast, was comparable between PFC and parietal cortex regardless of cognitive control demand (Fig. 4J,K). Additionally, both top-down and bottom-up transmission of cue and probe signals between prefrontal and parietal neurons was stronger on prepotent trials (Fig. 7B,F) than balanced trials (Fig. 7A,E). These findings show that neural dynamics in the prefrontal and parietal network reflect differential contributions of the two areas to cognitive control that are sensitive to cognitive control demand. This may offer insight into neural dynamics that are disrupted by schizophrenia to produce cognitive control deficits in the disease, warranting further investigation of prefrontal–parietal network dynamics in models of the disease state.
Footnotes
This work was supported by National Institutes of Health R01MH077779, R01MH107491, and P50MH119569; the Department of Veterans Affairs; the American Brain Sciences Chair; the Wilfred Wetzel Graduate Fellowship; and National Institute of General Medical Sciences T32 GM008244 and T32 HD007151. This work was performed while R.K.B. was employed at the University of Minnesota. The opinions expressed in this article are the author's own and do not reflect the views of the National Institutes of Health, the Department of Health and Human Services, or the U.S. Government. This material is the result of work supported with resources and the use of facilities at the Minneapolis VA Health Care System. The contents do not represent the views of the U.S. Department of Veterans Affairs, the National Institutes of Health, the Department of Health and Human Services, or the U.S. Government. We thank C. Dean Evans for technical assistance during surgeries and neural recordings as well as exemplary animal care; Dale Boeff for assistance with computer programming as well as design and construction of neurophysiological recording equipment; and A.D. Reddish for insightful suggestions regarding the decoding analyses.
The authors declare no competing financial interests.
- Correspondence should be addressed to Matthew V. Chafee at chafe001{at}umn.edu