Top-down control of exogenous attentional selection is mediated by beta coherence in prefrontal cortex

Summary

Salience-driven exogenous and goal-driven endogenous attentional selection are two distinct forms of attention that guide selection of task-irrelevant and task-relevant targets in primates. Top-down attentional control mechanisms enable selection of the task-relevant target by limiting the influence of sensory information. Although the lateral prefrontal cortex (LPFC) is known to mediate top-down control, the neuronal mechanisms of top-down control of attentional selection are poorly understood. Here, we trained two rhesus monkeys on a two-target free-choice luminance-reward selection task. We demonstrate that visual-movement (VM) neurons and not visual neurons or movement neurons encode exogenous and endogenous selection. We then show that coherent-beta activity selectively modulates mechanisms of exogenous selection specifically during conflict and consequently may support top-down control. These results reveal the VM-neuron-specific network mechanisms of attentional selection and suggest a functional role for beta-frequency coherent neural dynamics in the modulation of sensory communication channels for the top-down control of attentional selection.

eTOC Blurb

Dubey et al. demonstrate inhibitory network mechanisms of top-down control of attentional selection in prefrontal cortex. A specific neuronal response-class - visual-movement neurons encode exogenous and endogenous selection. Coherent-beta activity selectively suppresses sensory information flow in these neurons to favor endogenous selection during conflict.

Introduction

In primates, selection of task-relevant targets is guided by goal-driven (“top-down”) endogenous attentional processes whereas selection of task-irrelevant distractors is guided by salience-driven exogenous (“bottom-up”) attentional processes^1–6. Exogenous selection is fast and occurs earlier in time whereas endogenous selection is slow and occurs later in time^2,7–9. Therefore, trial-by-trial flexible selection behavior depends on the dynamic interplay between exogenous and endogenous attentional mechanisms^2,7,9,10. However, how attentional selection is controlled when exogenous and endogenous attentional mechanisms are in conflict remains unclear. How is the task-relevant target selected when in conflict with salient target?

Endogenous attentional selection relies on a top-down control process that enables the selection of task-relevant targets by limiting the influences of automatic-salience selection^11–13. Neural mechanisms that support top-down control are distributed throughout the fronto-parietal regions and rely heavily on the lateral prefrontal cortex (LPFC)^9,13–18 as evident from lesion experiments^19–23. Thus, LPFC-mediated top-down control mechanisms may support selection of the task-relevant target when in conflict with task-irrelevant salient target.

Information flow about task-relevant and irrelevant targets during conflict must be mediated by multiregional communication and, specifically, competition between convergent information streams. Since exogenous attentional selection is fast and processes sensory streams of information while endogenous attentional selection is slow and processes information about goals, each attentional process operates across distinct neural pathways, i.e. communication channels (Fig 1A). Consequently, selective filtering of information flow across sensory and reward-based communication channels may support the top-down control of attentional selection.

Figure 1: Experimental design.

(A) Attentional selection involves filtering information flow across communication channels. The luminance channel communicates the task-irrelevant sensory information. The reward channel communicates task-relevant goal information. During conflict, the luminance and reward communication channels compete to guide exogenous or endogenous selection. The top-down control may support exogenous or endogenous selection by filtering either or both of the luminance and reward channels. (B) Luminance-reward-selection (LRS) task events. (Ci) congruent and conflict LRS trials - Mean value of reward associated with each target is varied in blocks of 40–70 trials (top panel). Luminance value associated with each target is randomly selected on each trial (bottom). (Cii) Same as Ci except for luminance-only (left) and reward-only (right panel) trials. (D) Neural recording locations (white dots). Areas 8 and 46³⁹. as: arcuate sulcus; ps: principal sulcus.(E) Spike rasters and peri-stimulus time histogram (PSTH) for an example visual-movement (VM), visual and movement neuron for congruent and conflict LRS trials aligned to target presentation. Saccade onset (red dots). Target onset (dotted lines). (F) PSTH for VM (N=139, left), visual (N=57, middle) and movement (N=65, right) neurons on congruent and conflict LRS trials when target selection was InRF and OutRF. The s.e.m of firing rates (shaded). Average RT (red arrows).

Neuronal coherence measured by local field potential (LFP) activity in specific frequency bands reveals the correlations in the timing of neural activity across populations of neurons²⁴, and is generally interpreted in terms of multiregional communication^25–27. Many studies highlight the importance of multiregional communication and neuronal coherence to attentional selection. Attentional selection involves interactions between populations of LPFC neurons^28,29. In LPFC, cue-triggered LFP activity in the beta-frequency (15–35 Hz) band reflects exogenous selection^9,29. Beta frequency activity after the cue also reflects endogenous selection and suppression of sensory information during working memory and attention tasks^30–36. This suggests that beta frequency neuronal coherence may support top-down control and the trial-by-trial interplay between endogenous and exogenous attentional selection during conflict. However, prior work has not dissociated endogenous and exogenous selection during conflict to understand how beta frequency coherence biases information flow across communication channels to guide attentional selection. Whether beta frequency neural coherence acts on communication channels carrying salience-driven or goal-directed information is not known.

Here, we test the neural mechanisms of top-down control of attentional selection and the role of beta-frequency neuronal coherence.

Results

We trained two Rhesus macaque monkeys (macaca mulatta) to perform a luminance-reward-selection (LRS) task (Fig 1B, STAR Methods). The LRS task dissociates exogenous and endogenous attentional selection by independently-manipulating reward and luminance (Fig 1Ci) to yield either congruent or conflict trials. On congruent trials, luminance and reward value were both high for one target (Rich-Bright) and low for the other (Poor-Dim). On conflict trials, one target was Rich-Dim while the other was Poor-Bright (Fig 1Ci). The LRS task also featured luminance-only trials with similar relative target reward values and reward-only trials with similar relative target luminance values (Fig 1Cii).

We recorded neural activity from 32 electrodes in LPFC during LRS task performance (Fig 1D. Monkey 1: N=39 sessions, Monkey 2: N=42 sessions) yielding 409 task-responsive single units (M1: 179; M2: 230 neurons). We also recorded the activity of each neuron during a single-target oculomotor delayed response (ODR) task. Spiking activity increased in a spatially-selective manner during the ODR trials³⁸. Of 409 neurons, 261 neurons were ODR-task-responsive with an excitatory response field (64%). Most neurons showed elevated firing activity following target onset and the saccadic response, which we term visual-movement (VM) neurons (N=139, 53%; Fig S1A). Other neurons increased firing after target onset alone, termed visual neurons (N=57, 22%; Fig S1B), or around the saccade and not target onset, termed movement neurons (N=65, 25%; Fig S1C).

To further analyze the LPFC neuronal population dynamics, we performed a principal component analysis of activity during the luminance-only trials which revealed visual and movement modes that explained ~75% firing variability (Fig S1D). We projected firing rate activity of each group of neurons onto the visual and movement modes (Fig S1F,G). While visual-movement neurons showed activity for both visual and movement modes, visual neurons showed activity mainly for the second mode. Movement neurons showed activity mainly for the movement mode and not the visual mode.

LPFC neurons involved in attentional selection should fire more spikes on trials when the target in the RF is chosen (inRF) compared with trials when the target outside the RF is chosen (outRF; Fig 1E,F for LRS task and Fig S1H,I for luminance-only and reward-only trials). VM neurons responded significantly more on InRF trials compared to OutRF trials (LRS: p=4.2×10⁻³; Luminance-only: p=5.8×10⁻⁴ Reward-only: p=4.5×10⁻³, rank-sum, 50–200 ms). Movement neurons also responded significantly more during movement on InRF trials compared with OutRF trials, but not immediately after target onset (LRS: p=5.7×10⁻⁵; Luminance-only: p=1.1×10⁻³, Reward-only: p=3.8×10⁻⁵, rank-sum, 150–250 ms). Visual neurons responded similarly for InRF and OutRF trials (LRS: p=0.61, rank-sum, 0–100 ms). Visual neurons fired similarly for InRF and OutRF trials on reward-only trials but responded significantly more on InRF, luminance-only trials (Luminance-only: p=0.03, Reward-only: p=0.62, rank-sum, 0–100 ms). Visual neuron responses are not necessarily due to attention because they are only selective on luminance-only non-conflict trials. These results show that VM neurons and not visual and movement neurons play a more direct role in attentional selection.

Visual-movement neuron spiking reflects attentional selection

Conflict trials may reveal endogenous, reward-driven selection or exogenous, stimulus-driven selection (Fig 2A). We specifically predicted that RT should be longer on conflict trials when endogenous selection is expressed and the Rich-Dim target is chosen not the Poor-Bright target, endo-conflict trials, compared with conflict trials when exogenous selection is expressed and the Poor-Bright target is chosen, exo-conflict trials. RTs were significantly greater for endo-conflict trials compared to exo-conflict trials (M1: Endo-conflict RT=191+/−29ms, Exo-conflict RT=169+/−26ms; p=5.3×10⁻²⁶ M2: Endo-conflict RT=192+/−28ms, Exo-conflict RT=185+/−36ms; p=8.2×10⁻²¹, rank-sum, mean+/−sem). Consequently, conflict trials revealed whether endogenous selection or exogenous selection was expressed trial-by-trial.

Figure 2: Visual-movement (VM) neurons reflect exogenous and endogenous attentional selection.

(A) On congruent trials, luminance and reward value is high for one target- Rich-Bright and low for other target Poor-Dim. When both luminance and reward drive are in congruent, they both favored the selection of the Rich-Bright target. On conflict trials, one target has high-reward and low-luminance Rich-Dim while the other has low-reward and high-luminance Poor-Bright. When luminance and reward drive are in conflict, luminance driven choices result in exogenous (Exo) selection of the Poor-Bright target and reward driven choices result in endogenous (Endo) selection of the Rich-Dim target. (B) Schematic. On Exo-InRF trials the selected Poor-Bright target is in the RF while on Exo-OutRF trials the selected Poor-Bright target is out of the RF. On Endo-InRF trials the selected Rich-Dim target is in the RF while on Endo-OutRF trials the selected Rich-Dim target is out of the RF. (C) Spike rasters and PSTHs for exo and endo selection of an example visual-movement, visual and movement neuron on conflict trials shown aligned to the target presentation. Saccade RT (black dots). Target onset (dotted line). (D) PSTH for VM, visual and movement neurons on Exo-InRF, Exo-OutRF, Endo-InRF and Endo-OutRF conflict trials. The s.e.m of firing rates (shaded). (E) Firing rate difference for selection into and out of the RF for three groups of neurons on conflict trials (top panel). Mean +/− s.e.m. (F) Permutation test p-values against a null hypothesis that there is no difference in InRF and OutRF firing rates (bottom panel). False-discovery-rate (FDR) corrected p-values for alpha=0.01 (black). Arrow: selection time (ST) when first time separation becomes significant (VM: Exo ST=49 ms, Endo ST= 116 ms; Visual: Exo ST = 61 ms ; Movement: Exo ST=153 ms, Endo ST=156 ms). Average RT for exogenous and endogenous selection trials (dotted lines).

Behavioral choice variations with RT also revealed exogenous and endogenous selection. On conflict trials, shorter RTs reflected exogenous selection whereas longer RTs choices reflected endogenous selection (Fig S2A-D). On luminance-only trials, shorter RTs reflected exogenous selection and bright target choice probability approached chance for longer RTs. On reward-only trials, longer RTs reflected endogenous-driven selection and rich target choice probability approached chance for shorter RTs. Therefore, only conflict trials reveal exogenous and endogenous selection.

We next investigated the underlying neural mechanisms. On exo-conflict trials when the Poor-Bright target is selected and the target is in the RF, exo-InRF trials, neuronal firing should differ from exo-conflict trials when the Poor-Bright target is selected and the target is out of the RF, exo-OutRF trials (Fig 2B). Firing supporting endogenous selection on endo-conflict trials when the Rich-Dim target in the RF is selected, endo-InRF trials, should differ from firing on endo-conflict trials when the Rich-Dim target out of the RF is selected, endo-OutRF trials (Fig 2B). Since exogenous selection occurs earlier than endogenous selection, neuronal selectivity on exo-conflict trials should occur before endo-conflict trials.

Consistent with a role in attentional selection during conflict, VM neuron selectivity on exo-conflict trials occurred before endo-conflict trials (Example: Fig 2C). VM neurons responded more when the InRF target was selected compared to when the OutRF target was selected for both exo-conflict trials and endo-conflict trials. Interestingly, VM neuron firing on inRF trials differed from OutRF trials substantially earlier on exo-conflict trials compared to endo-conflict trials. After the target onset, VM neuron firing rate during exogenous selection separated ~50 ms earlier than during endogenous selection (Exo-ST=49ms, Endo-ST=116ms. Fig 2D,2E,S2E,S2F). Therefore, VM neurons process both exogenous and endogenous selection during conflict.

VM neuron firing for In-RF selection of Poor-Bright target was driven by exogenous attention and was not simply due to physical brightness of the target (Fig S2G,H). On luminance-only trials, VM neurons responded more when the bright-target was selected rapidly compared to when the bright-target was selected more slowly (Fig S2G, In-RF). Thus, exogenous attention and not physical brightness drives firing.

The firing rates of visual neurons for InRF and OutRF conditions significantly differed on exo-conflict but not endo-conflict trials (p<0.01, permutation; Exo-ST=61 ms. Fig 2E). Consequently, visual neuron activity likely reflects exogenous selection alone and not conflict with endogenous selection. Movement neurons, however, showed elevated responses on InRF trials compared to OutRF trials for both exo-conflict and endo-conflict trials (Fig 2E). But movement neuron firing rates for two conditions separated at a similar time after the target onset (p<0.01, permutation, Exo-ST=153 ms, Endo-ST=156 ms, Fig S2E). Hence movement neuron activity does not reflect conflict and likely reflects subsequent response preparation and movement. Therefore, only VM neuron firing reflects attentional dynamics during conflict.

LPFC neuron spiking activity contains beta-frequency bursts

We investigated the role of neuronal coherence in LPFC in the control of exogenous and endogenous selection. In the pre-target period, LFP activity on individual electrodes displayed clear bursts of beta-frequency activity, 15–30 Hz, which we term beta-bursts (Fig 3A). Pre-target beta-bursts were clearly and reliably visible in LFP activity on individual trials. When present, beta-bursts tended to occur in the pre-target period and not after the target onset, and typically occurred for several hundred milliseconds.

Figure 3: Beta-frequency bursts and coherent neuronal dynamics.

(A) Raw extracellular recordings at an example recording site during several LRS task trials. Shaded area denotes the window of interest used for calculating beta amplitude values. (B) Pre-target beta-burst amplitude at the example site (same as A) on an example experimental session. Dark-green: high-beta trials (HB, ~33% highest beta-bursts); Light-green: low-beta trials (LB, ~33% lowest beta-bursts). (C) Spike-field coherence (SFC) between an example unit and field recorded on a neighboring electrode (same as A and B). (D) Population average SFC of coherent pairs (N=176) and not-coherent pairs (N=233). The s.e.m of SFC is shown in lighter shades. (E) Scatter plot of HB SFC versus LB SFC at 20 Hz. Plot limits are zoomed to improve visibility. Inset: all the SFC electrode pairs. Each dot denotes a recording pair. Red dot denotes the example SFC in C. Marginal histograms denote the SFC distribution for HB and LB trials.

For each trial, we estimated the amplitude of pre-target beta-bursts at a single site from 200 ms before target onset until target onset, a duration long enough to sample several cycles of activity at the beta-frequency. Beta-burst amplitude varied significantly from trial-to-trial (example site:Fig 3B). Across the population, beta bursts were reliably present across LPFC recording locations in each animal (M1: 1108/1152 sites; M2: 1299/1344 sites, 96% of electrodes, p<0.05 permutation). We grouped the trials with the highest ~33% and lowest ~33% beta-burst activity to yield high-beta (HB) trials and low-beta (LB) trials.

We first sought to assess whether beta-bursts in LFP activity could reflect a local source in LPFC. To help answer this question, we looked for evidence of coherent activity in the spiking activity of 409 single units in LPFC (M1:N=179; M2: N=230) by correlating spiking with nearby LFP activity (within approx. 1.5 mm) using spike-field coherence (SFC, Fig 3C). During the pre-target period, of the 409 neurons, 176 neurons significantly fired spikes at times predicted by nearby LFP activity in the beta-frequency range (15–35 Hz) (p<0.05 cluster-corrected, permutation; M1:N=59 and M2:N=117, Fig S3A). This suggests that beta-burst LFP activity involves LPFC neuron firing and is not simply due to activity propagating from other regions that do not necessarily involve LPFC neuron firing. SFC amplitude in LPFC was greatest for activity in the beta-frequency range, compared with frequencies greater than 35 Hz. The number of LPFC neurons that fired coherently in the gamma (40–70 Hz) frequency range was not significant (< 5%, Fig S3B).

Trial-to-trial variability in beta-burst amplitude may reflect trial-to-trial changes in the timing of spiking activity across the population of LPFC neurons. If so, spiking during HB trials should display greater coherence than spiking during LB trials. The dependence of neural coherence on beta-burst events should specifically be observed in the neurons that participate in the coherent activity. Neurons that do not participate, firing spikes at times that cannot be predicted by beta-frequency neural activity, should not show differences in coherence with beta-burst events. To test this, we estimated SFC immediately before target onset separately for the HB and LB trials for coherent and not-coherent neurons (Fig 3D-E, Fig S3C-E). Consistent with a strong relationship between spiking and beta-burst events, SFC was significantly stronger during HB trials than LB trials for coherent neurons (p=9.3×10⁻²⁷ rank-sum). The change in coherence for between high and low beta trials was smaller if not absent for non-coherent neurons (p=0.1, rank-sum), with the caveat that the lack of an effect in the population of non-coherent neurons may be a flooring effect. While the presence of SFC in coherent neurons may not be due to HB versus LB, the increase in SFC for coherent neurons between HB and LB could be due to higher power in HB. These results demonstrate that when high-amplitude beta-bursts occur during the pre-target period they reflect increased coherent spiking in LPFC neurons.

Beta-bursts selectively modulate exogenous attentional selection during conflict

We asked whether beta-bursts modulate attentional selection in general, or modulate either exogenous or endogenous attentional selection. We used conflict trials to test the relationship between beta-bursts and the neuronal mechanisms of attentional selection, and to ask whether beta-bursts exhibit specificity for endogenous or exogenous selection. We focused on LRS conflict trials for which the choices were made into the response field of each neuron under study.

We examined three hypotheses. First, since the firing rate of VM neurons reflects both endogenous and exogenous attentional selection, if LPFC beta-bursts modulate attentional selection, the rate of VM neuron firing should differ when beta-burst amplitude was high compared to when beta-burst amplitude was low. Second, since visual and movement neuron activity does not reflect attentional selection, if beta-bursts mediate control of attentional selection, the firing rate of these neurons should not differ on HB and LB trials. Finally, if beta-bursts do not modulate selective attention in general and modulate either endogenous or exogenous selection, the relationship between beta-bursts and VM neuron firing rate should be present for either endogenous or exogenous selection trials and not both sets of trials.

We observed that VM neuron firing on InRF conflict trials involving exogenous selection significantly differed when pre-target beta-burst amplitude was high compared to when the amplitude was low (Fig 4A, VM-exo: p<0.01, permutation). Visual and movement neuron firing did not differ between HB and LB trials during exogenous selection (InRF-conflict: Visual-exo:p>0.01, Movement-exo:p>0.01, permutation). This demonstrates that beta-bursts in LPFC can modulate attentional selection and do not modulate LPFC firing rates more generally.

Figure 4: Beta-bursts selectively modulate VM neuron firing for exogenous selection during conflict.

(A) PSTH of the VM, visual and movement neurons for exogenous selection when pre-target beta-burst is high (HB trials) and low (LB trials). Mean +/− s.e.m. are shown for InRF conflict trials when selection was in the RF of the units. Dotted lines denote target onset. (B) Same as A but for endogenous selection. (C) Difference in firing rates for pre-target low and high beta-bursts. Mean +/− s.e.m. are shown for exogenous and endogenous selection.(D) Permutation test p-values under a null hypothesis that there is no difference in firing rates for high-beta and low-beta trials for exogenous selection and endogenous selection. FDR corrected p-values for alpha=0.01 (black).

Beta-bursts selectively modulated exogenous selection and LPFC neuron firing in the response field for HB and LB trials. Beta-bursts did not significantly modulate firing activity during endogenous selection and this was true for all three classes of neuronal response (InRF-conflict: VM-endo:p>0.01; Visual-endo:p>0.01; Movement-endo:p>0.01, permutation). Beta-bursts did not alter VM neurons firing out of the response field (OutRF-conflict trials: VM-exo:p>0.01. VM-endo:p>0.01. Visual-exo:p>0.01. Visual-endo:p>0.01. Movement-exo:p>0.01. Movement-endo:p>0.01, permutation, Fig S4A-D).

We compared VM neuron firing across all trials with choices into the response field against HB and LB trials (Fig S4F,G). On exo-HB trials, VM neuron firing was lower than the average firing rate, and on exo-LB trials VM neuron firing was higher than the average firing rate. This was not observed for endogenous selection trials. Therefore, an increase in pre-target beta activity could suppress luminance processing and a reduction in pre-target beta could facilitate luminance processing selectively on exogenous selection trials.

Beta-activity modulated exogenous selection irrespective of the location of RF in the visual field (Fig S4H-J). Beta-burst modulation effect was not selective to 33% grouping used for selecting high-beta and low-beta trials and was observed significantly for other percentiles as well (Fig S5). The modulation of VM neurons firing for exogenous selection was specific to beta frequency range and was not present for alpha (8–13Hz) and gamma (40–70Hz) activity (Fig S6). These results demonstrate that pre-target beta activity selectively modulates VM neuron firing for exogenous selection into the response field.

We next asked whether beta-burst modulation alters the timing of exogenous selection. Consistent with the suppression of VM neuron firing rate, higher-amplitude beta-bursts suppressed the exogenous process in time and delayed the selection by 10 ms on HB trials compared to LB trials (exoHB ST=54ms, exoLB ST=44 ms, p=0.03 permutation, Fig S7A-C). This was not observed for endogenous selection trials (endoHB ST=121ms, endoLB ST=117ms, p=0.26, permutation, Fig S7D-F). The difference in beta-burst amplitude values for HB exogenous and endogenous trials further supported the differences for exogenous and endogenous selection (p=0.02, rank-sum, Fig S7G). These results demonstrate that beta-burst activity modulates both the strength and timing of exogenous attentional selection.

Beta bursts do not modulate exogenous selection in the absence of conflict with endogenous selection

To further investigate beta-burst-related modulation of exogenous selection, we analyzed luminance-only trials (Fig 1Cb). In these trials, there was no conflict present and selection for fast RTs was predominantly guided by the exogenous selection. If beta-bursts inhibit exogenous selection in general, then VM neuron firing rate on InRF trials should differ on trials when beta-burst amplitude is high compared to when beta-burst amplitude is low. Alternatively, if beta-bursts specifically inhibit exogenous selection when there is conflict with endogenous selection, the rate of VM neuron firing should not differ on HB and LB trials.

Unlike during LRS conflict trials, the rate of VM neuron firing did not significantly differ for HB and LB trials when the Bright target was selected in the presence of the Dim target and the reward contingencies were the same (Fig 5A, InRF. VM-exo:p>0.01, permutation, Fig S7H). Visual and movement neuron firing did not differ between HB and LB trials on these trials (Fig 5A, InRF. Visual-exo:p>0.01; Movement-exo:p>0.01, permutation). This demonstrates that pre-target beta-bursts in LPFC specifically inhibit exogenous selection when in conflict with endogenous selection and do not modulate exogenous selection in general.

Figure 5: Beta-bursts do not modulate exogenous selection when not in conflict with endogenous selection.

(A) PSTH of three groups of neurons on luminance-only trials. Mean +/− s.e.m. are shown for InRF trails when the Bright target is selected in the RF. Dotted lines denote target onset. (B) Same as A, but for reward-only trials. Mean +/− s.e.m. are shown for InRF trails when the Rich target is selected in the RF. (C) Difference in firing rates for HB and LB trials. Mean +/− s.e.m. are shown for exogenous and endogenous selection. (D) Permutation test p-values under a null hypothesis that there is no difference in firing rates for HB and LB trials for exogenous selection and endogenous selection.

We also analyzed reward-only trials, the results confirmed that pre-target beta bursts did not modulate LPFC neuron firing rate in the absence of conflict between exogeneous and endogeneous selection (Fig 5B, InRF trials. VM-endo:p>0.01; Visual-endo:p>0.01; Movement-endo: p>0.01, permutation).

Pre-target beta-burst exogenous attentional modulation is transient in time

Pre-target beta-bursts (−200 ms to 0 ms, where 0 is target onset) selectively inhibit the neuronal mechanisms of exogenous attentional selection and not endogenous attentional selection. Since exogenous selection occurs earlier in time (49ms) compared to endogenous selection (116ms), the results may simply be due to the proximity of beta-bursts in time to exogenous-selection mechanisms. If so, beta-bursts that occur later in time, and hence closer to the time of endogenous selection, may instead modulate endogenous-selection not exogenous-selection. We analyzed beta-bursts during six time epochs and studied VM neuron firing patterns for InRF trials involving endogenous-selection when beta-burst amplitude was high compared to when the amplitude was low (Fig 6B). VM neuron firing rates on InRF trials were not significantly different for high-beta and low-beta trials during the [−100 100] epochs (Fig 6C,D, Endo [−100 100]: p>0.01, permutation).

Figure 6: Beta-burst exogenous attentional modulation is transient.

(A) VM neurons firing activity modulated by high (HB) and low (LB) beta-burst computed during six different time-windows. Mean +/− s.e.m. are shown for exogenous selection when the Poor-Bright target is selected in the RF. Dotted lines denote target onset. (B) Same as A, but for endogenous selection. Mean +/− s.e.m. are shown for InRF trials when the Rich-Dim target is selected in the RF. (C) Difference in firing rates for HB and LB trials. Mean +/− s.e.m. are shown for exogenous and endogenous selection. (D) Permutation test p-values under a null hypothesis that there is no difference in firing rates for HB and LB trials for exogenous selection and endogenous selection. FDR corrected for p-values for 0.01 alpha (black).

Examining the time course of beta-burst related modulation also revealed that early beta-bursts do not tend to modulate exogenous-attentional selection (Fig 6A). The strongest modulation of VM neuron firing was observed for beta-bursts that occurred immediately before target onset (Fig 6C).

Pre-target beta-bursts modulate exogenous selection reaction times

If the selective modulation of VM neuron firing with coherent-beta activity reflects attentional selection, then the modulatory effect of beta activity on conflict trials should be present during exogenous choice behavior more than during endogenous choice behavior. Since behavioral RTs reflect the underlying mechanism of attentional selection, we specifically predicted that RTs should vary trial-by-trial with coherent beta-activity on exo-conflict trials more than on endo-conflict trials. RTs were correlated with coherent beta-activity on exo-conflict trials (M1:rho=0.32,p=0.02; M2:rho=0.79,p=0 Spearman-correlation) as well as on endo-conflict trials (M1:rho=−0.38,p=4.1×10⁻³; M2: rho=0.78,p=0 Spearman-correlation). For each monkey, changes in coherent-beta activity were associated with changes in RTs on exogenous choice trials more than on endogenous choice trials (Fig 7A). For the exo-conflict group of trials, the reaction times significantly differed with coherent-beta activity (M1: Normalized-RTrange:4.65%. Absolute-RTrange: 8.59ms. p=0.02; M2: Normalized-RTrange:2.97%. Absolute-RTrange:5.54ms. p=0.01 Permutation; Normalized RT range refers to percent range (maxRT - minRT) of variation of RT with beta values. Absolute RT range refers to the range maxRT - minRT of variation in millisecond durations). For the endo-conflict group of trials, RTs did not significantly differ with coherent-beta activity (M1: Normalized-RTrange:1.25%. Absolute-RTrange:2.76 ms. p=0.35; M2: Normalized-RTrange:0.95%. Absolute-RTrange:1.34 ms. p=0.12. Permutation). Finally, since VM neuron firing effects are not present on non-conflict trials, the relationship between beta-activity and RTs should not be present on non-conflict trials. Pre-target coherent-beta activity and RTs did not significantly differ when sorting on luminance-only non-conflict trials, and significantly differed when sorting on reward-only non-conflict trials, albeit across a small range of RT compared with exogenous conflict trials (Fig 7B. Luminance-only: M1: Normalized-RTrange:1.21%. Absolute-RTrange:2.14ms. p=0.47. M2: Normalized-RTrange:3.67%. Absolute-RTrange:3.78ms. p=0.11. Reward-only: M1: Normalized-RTrange:2.63%. Absolute-RTrange:5.09ms. p=0.05. M2: Normalized-RTrange:2.32%. Absolute-RTrange: 4.54ms. p=0.02. Permutation).

Figure 7: Beta-bursts selectively modulate exogenous selection reaction times.

(A) Saccade RTs on conflict trials as a function of pre-target beta burst amplitude for Monkey 1 and Monkey 2. Exogenous selection choices (red). Endogenous selection choices (blue). (B) Same as A, but for luminance-only and reward-only trials.

Therefore, the role of coherent-beta activity in attentional selection is specifically present during conflict, is generally consistent with the pattern of results observed for the VM neurons and consequently may mediate top-down control of attentional selection by modulating sensory, cue-driven responses in the VM neuron subpopulation.

Discussion

We make two specific contributions that demonstrate a role for beta-frequency neural coherence in attentional selection through inhibitory mechanisms (Fig 8). We propose that attentional selection involves filtering of luminance and reward channels that communicate information to visual-movement neurons in LPFC in order to select a response (Fig 8A). When beta bursts are not present, target onset drives LPFC to select information in the luminance-channel before information in the reward-channel is available (Fig 8B). When beta bursts are present, information in the luminance-channel is inhibited and the response tends to be selected based on information in the reward-channel (Fig 8C).

Figure 8 : Channel modulation hypothesis.

(A) LPFC VM neurons receive luminance and reward information from two distinct communication channels, which compete to guide behavior. Exogenous selection of ‘Bright’ and endogenous selection of ‘Rich’ target depends on inhibitory modulation of the luminance channel. (B) In absence of beta-bursts, the luminance channel is open, communicating sensory-driven salient information earlier than goal-driven information. Information in the luminance channel drives the exogenous selection of the ‘Bright’ target. (C) In presence of beta-bursts, the luminance channel is close, inhibiting communication of sensory information. Information in the reward channel drives the endogenous selection of the ‘Rich” target.

We then demonstrate that coherent neuronal activity in the beta frequency range (15–30 Hz) selectively modulates exogenous selection by suppressing the luminance channel that carries salient sensory information. Beta activity observed in the pre-target period is associated with the inhibited post-target, sensory-driven firing by LPFC neurons when selection is driven by exogenous attention but not by endogenous attention. Consequently, our results are consistent with the top-down control view of attentional selection. According to the top-down control view, selection of task-relevant endogenous targets relies on mechanisms of multiregional communication that limit the influence of sensory inputs^11–13. Since top-down control mechanisms operate under the knowledge of task-relevance¹³, the beta-activity effect was observed on conflict trials but not on non-conflict trials. On conflict trials, selection of the task-relevant target yielded high reward whereas on non-conflict trials the task did not prioritize one target over another based on reward-value. Since we show the role played by coherent-beta activity could be to modulate information flow due to sensory inputs, our work provides new evidence for how coherent-beta activity in LPFC could mediate the top-down control of attentional selection.

We show how coherent-beta-activity could bias the mechanisms of attentional selection in LPFC by influencing the flow of sensory information during target selection. We specifically show that a subgroup of LPFC neurons, VM neurons and not visual and movement neurons, encodes both exogenous and endogenous selection. The timescales underlying exogenous and endogenous selection have been a major focus of behavioral work which has shown that reaction times are typically ~30 ms faster for exogenous selection^1,2,4,7,40. Here, we go further and also measure the timescales of exogenous and endogenous selection by analyzing the spiking patterns of populations of individual LPFC neurons. Consistent with previous recordings in LPFC⁹, we find that VM neuron spiking activity in response to target onset encoded exogenous selection ~50 ms before endogenous selection. This difference in timing means that coherent-beta-activity in PFC can have a substantial influence on the direction of sensory information flow and bias the selection of relevant targets in the presence of irrelevant distractors.

In the following, we discuss the mechanisms of top-down attentional control and how coherent-beta activity may support the selection of task-relevant targets.

Top-down attentional control mechanisms are mediated by coherent-beta-activity

In LPFC, beta-activity reflects exogenous and endogenous attentional processes^9,29,31,41. The emergence of LPFC-beta-activity before selection during different goal-defining tasks further suggests a role for beta-activity in the top-down control of attention selection^11,31,41. Here, we more closely examine the strong relationship between spiking and coherent-beta-activity in LPFC immediately before presenting relevant and irrelevant targets to reveal mechanisms of top-down attentional control. The central aspect of top-down control is inhibition with knowledge of what needs to be controlled, i.e. relevance¹³. We show that LPFC-beta-activity is associated with the inhibition of LPFC neural firing during exogenous selection and not endogenous selection, and so is grounded in task-relevance. Importantly, LPFC-beta-activity mediated selective inhibition was only observed in presence of conflict i.e, when sensory and reward drive each favored the selection of different targets (Fig 4). In absence of conflict, when sensory information was absent, LPFC firing rates were not modulated with beta-activity (Fig 5).

On conflict trials, reward-drive favored the selection of the task-relevant target whereas, on non-conflict trials, absence of reward-drive diminished the task relevance of one target over other. Therefore, we propose that LPFC performs top-down control of attentional selection by deploying beta-frequency coherent neural activity to selectively limit or bias the flow of sensory information specifically when conflicting information drives target selection.

The posterior parietal cortices also process exogenous sensory information^9,17,42,43. LPFC coherent-beta-network that selectively inhibits sensory information likely operates across frontal-parietal projections. Indeed, frontal and parietal areas both reflect coherent-beta activity indexing stimulus selection in attention and working memory^9,29,44. LPFC may selectively inhibit PPC information flow through a long range beta network. If so, prefrontal areas need to generate a sufficiently reliable and impactful neural stimulus to influence posterior parietal areas. The firing of bursts as compared to single isolated spikes offer a candidate mechanism^45,46. For-example, long-range beta-burst synchronization between anterior cingulate cortex and LPFC exists during selective attention⁴⁷. Our observations of pre-target beta-bursts highlight a potential mechanistic role for how information is routed through PPC during the top-down control of attentional selection.

LRS task reveals the timescale of exogenous and endogenous selection mechanisms

The LRS task revealed the time-course of attentional selection mechanisms in LPFC. Use of a non-cue binary-choice task, in which selection immediately followed target onset and both target locations were spatially-randomized trial-by-trial, revealed distinct timescales for each form of attentional selection. Previously-used behavioral tasks have often manipulated spatial attention in a delayed design by presenting an attentional cue before the onset of a target^9,29,48–51. In such paradigms, spatial attention is allocated to the cue location before exogenous or endogenous attention is recruited by the target. Previous work has also used tasks in which targets are presented at spatial locations in a predictable manner, which can generate spatial biases in behavior that may also be confounded with attentional selection mechanisms⁴⁹.

While the LRS task was used to dissociate exogenous and endogenous attentional selection mechanisms, previous work has shown that a non-salient target previously associated with reward may also capture attention^52,53, resulting in an interaction between salience-drive and involuntary-value-driven automatic attention. Whether coherent-beta-activity is implicated in the neural mechanisms of interactions between other forms of attention capture is an interesting direction for further work.

One concern is that the difference in VM neuron firing between exo and endo conditions is simply due to physical brightness of the target and not attentional selection (Fig 2). However, the luminance-only trials control for physical brightness effects. In these trials, physical brightness is the same as in the conflict trials but the recruitment of exogenous and endogenous attention differs (Fig S2B,D). Further, the physical brightness value of the selected target was not correlated to the pre-target beta value (Fig S8A-H). Therefore, the beta-modulation effects that we report are most consistent with an attentional effect.

Dynamic interplay of exogenous and endogenous attentional selection

By dynamically shifting between more active and less active coherent states, high-beta and low-beta, our results show that VM neurons in the coherent-beta subnetwork may flexibly modulate multiregional communication across a sensory information channel that carries visual target information into the association cortices. We report behavioral effects in which the influence of coherent-beta-activity on saccade RTs is consistent with the effects observed in VM neuron firing (Fig 7). Changes in coherent-beta activity were associated with changes in RTs on exogenous choice conflict trials more than on endogenous choice conflict trials. On trials when the choice was to the endogenous target, RTs were more similar across trials with beta bursts before target onset that differed in strength. This pattern of results mirrors that for the variations of VM neuron firing with pre-target beta activity across conflict and non-conflict trials. Thus, neural and behavioral results reinforce the flexible interplay between exogenous and endogenous selection associated with beta-mediated modulation of a sensory-driven information channel.

Comparison with previous studies

We categorized the responses of the population of LPFC neurons in terms of three response groups-visual, movement and visual-movement neurons. To assess the presence of multiple response types, we examined LPFC neuronal response during the LRS task and demonstrated the presence of non-random task-relevant structured selectivity (Fig S1F,G). Whether the selectivity of frontal cortical neurons is random or non-random, as well as whether non-random selectivity forms distinct clusters in a high-dimensional space of responses, as opposed to simply a continuum in two-dimensions as we observe, is an interesting topic. Previous studies report neuronal populations express both random mixed-selectivity^54,55 as well as non-random mixed-selectivity that forms clusters in a high-dimensional space of neuronal responses^56,57. Consequently, our data are less consistent with random selectivity and are more broadly consistent with⁵⁶ who also show that responses co-vary with task-relevant features. Overall, these analyses support our approach to characterize the population response in terms of multiple response groups.

Previous studies have associated beta activity with inhibition and reach movement initiation^58–61. In the sensorimotor cortex, beta amplitude increases at rest and for stable postures and reduces during movement^35,62,63. For example, Kilavik et al. showed increased beta in both pre-cue and pre-go epochs of reach movement tasks, with a temporary drop in beta amplitude after cue⁶⁴. The post-cue suppression of beta-amplitude for movement planning and initiation may be related to PFC beta before oculomotor selection that we report. However, the detailed pattern of our results does not suggest that PFC beta is related to the saccade itself. We only observe the beta modulation effects on trials involving conflict. If the results were due to movement suppression we would also observe them on reward-only and luminance-only trials (Fig 5). In addition, we do not observe beta effects on those PFC neurons whose activity is most tied to the movement- the movement neurons (Fig 4). Furthermore, beta activity altered VM neuron firing for ipsilateral and contralateral saccade selection and was not a lateralized motor effect (Fig S4H-J). Finally, we do not observe the effects on trials captured by endogenous attention and only observe the effects on trials with luminance driven exogenous responses (Fig 4).

Prefrontal beta-activity observed during attention and working-memory reflects information about the task-relevant rules that determine stimulus-response mapping^9,31,44 using tasks that employ a task-rule to match a sample either in object or space feature following a cue/delay period. In comparison, the LRS task employs a non-cue binary choice task design to examine attentional selection mechanisms. The beta-burst modulation effect we report could be due to differences in involvement of beta-activity in the preparatory period and cue-triggered delay-period. However, we propose that conflict between the two sources of information influences the attentional selection process irrespective of the beta activity period- cue-triggered/preparatory beta. The channel modulation top-down control model (Fig 8) suggests that beta-activity is involved in resolving conflict. Therefore, cue-period beta-activity may also influence the selection process in the presence of conflict.

Previous studies have suggested that LPFC neuron firing activity is modulated by reward-value^65,66. We did not observe a value-based modulation of VM neuron firing activity (Fig S2I,J). However, we did observe value-based modulations in movement neuron firing. Movement neuron firing was greater when the rich target was selected inRF compared to poor target selection inRF. Movement neuron firing during our task likely does not reflect endogenous attention processing and may instead reflect a form of reward expectancy. Note, however, the increased firing for the rich target contradicts work by Kaping et al. which reports enhanced LPFC activity when a low value target is selected over a high value target. This discrepancy could arise from our use of an immediate saccade unlike other work involving covert attentional cues.

In conclusion, we reveal the mechanisms of top-down control of attentional selection in LPFC involve the inhibition of luminance information to facilitate reward-guided behavior. We show that the dynamics of a population of VM neurons that fire coherently with beta activity may mediate top-down control of attentional selection, consistent with a role in inhibitory multiregional communication. We further show that coherent beta-activity selectively modulates exogenous responding compared with endogenous responding resulting in the flexible interplay between exogenous and endogenous selection necessary to resolve conflict.

STAR METHODS Text

RESOURCE AVAILABILITY

Lead contact

Further information and requests for resources should be directed to and will be fulfilled by the Lead Contact, Bijan Pesaran ( pesaran@upenn.edu )

Materials availability

The study did not generate new unique reagents

Data and code availability

• The electrophysiological and behavioral data reported in this study have been deposited at: Figshare. The DOI is listed in the key resources table.

• All original code to analyze the electrophysiological and behavioral data has been deposited at Figshare and is publicly available as of the date of publication. DOIs are listed in the key resources table.

• Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request.

KEY RESOURCES TABLE

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Deposited data
Experimental data	Recorded experimental data	https://doi.org/10.6084/m9.figshare.23567439.v1
Experimental models: Organisms/strains
Rhesus macaque (Macaca mulatta)	Covance and Charles River Laboratories	N/A
Software and algorithms
Analysis code	Custom software analysis code	https://doi.org/10.6084/m9.figshare.23503767.v2
MATLAB R2017	Mathworks	https://www.mathworks.com/products/matlab.html
FSL	Analysis Group, FMRIB, Oxford, UK	https://fsl.fmrib.ox.ac.uk/fsl/fslwiki/
Other
Microelectrodes 0.7–1.4 MΩ impedance	Alpha Omega single electrodes	https://www.alphaomega-eng.com/
Neural recordings and amplifier for monkeys 1 and 2	NSpike NDAQ system, Harvard Instrumentation	http://nspike.sourceforge.net/#Overview
Eye tracking	ISCAN, MA	http://iscaninc.com
Task controller	Custom LabView software with a real-time embedded system NI PXI-8820	https://www.ni.com/en-us/shop/labview.html

EXPERIMENTAL MODEL AND STUDY PARTICIPANT DETAILS

All surgical and animal care procedures were done in accordance with National Institute of Health guidelines and were approved by the New York University Animal Care and Use Committee. Two adult male rhesus monkeys (Macaca mulatta) participated in the experiments (Monkey 1, 9.5 kg and Monkey 2, 8.4 kg). Both animals had been previously used in other eye-movement experiments.^7,67

METHOD DETAILS

Experimental preparation

Once trained on behavioral tasks, each animal was implanted with a low-profile recording chamber (Gray Matter Research, MT). The craniotomy was made over the right pre-arcuate cortex of each animal using image-guided stereotaxic surgical techniques (Brainsight, Rogue Research, Canada). A semichronic microelectrode array microdrive (SC32–1, Gray Matter Research, MT) was inserted into the recording chamber and sealed. The SC32–1 system has 32 microelectrodes, spaced 1.5 mm away (Fig 1D). The SC32–1 is a modular, replaceable system capable of independent bidirectional control of 32 microelectrodes.

Behavioral experiments

Experimental hardware and software:

Eye position was constantly monitored with an infrared optical eye tracking system sampling at 120 Hz (ISCAN). Visual stimuli were presented on an LCD screen (Dell Inc) placed 34 cm from the animal’s eyes. The visual stimuli were controlled via custom LabVIEW (National Instruments) software executed on a real-time embedded system (NI PXI-8184, National Instruments).

Experimental design:

Each monkey first performed a visually-guided oculomotor delayed response (ODR) task to map the spatial response fields of neurons. Each monkey then performed the luminance-reward-selection (LRS) task to study the flexible control of attentional selection. Behavior and neural data was recorded across 39 (Monkey 1) and 42 (Monkey 2) experimental sessions.

ODR task:

Each trial began with a visual fixation target presented at the center of the screen. Each animal maintained fixation for a variable 500–800 ms baseline period. After the baseline period, a red square appeared in the periphery to indicate target location of the saccade. There were eight possible iso-eccentric target locations spaced 10 deg around central fixation. Target location was randomized over trials so that animals could not predict where the cue would appear on any given trial. Each monkey maintained fixation for a variable 1000–1500 ms delay period. After the delay period, the central fixation square was extinguished, providing the Go signal for the animal to move his eyes to the target location. A fluid reward was awarded on successful completion of the trial. A trial was aborted if the animal failed to align his gaze within 2deg of the center of fixation or periphery target. On a given experimental session, on average M1 95 +/− 24 trials performed trials and M2 performed 248 +/− 32 ODR trials (mean +/− sd).

LRS task:

Each trial again started with fixation at a visual target at the center of the screen for a variable 500–800 ms baseline period. After the baseline period, the center fixation target was extinguished, and two red targets (T1 and T2) were presented at random locations in the visual periphery at a 10 deg eccentricity from the central fixation. Two targets were constrained to be at least 90 deg apart on each trial. The randomized spatial location of targets controlled for the influence of spatial attention at the start of each trial. Onset of targets provided the animal Go signal to perform a saccade to one of the targets. Each animal was required to maintain a fixation of 300 ms at the chosen target, after which appropriate juice reward was delivered. Each trial lasted 890–1400 ms, and only one choice could be made per trial. A trial was aborted if the animal failed to align his gaze within 2deg of the center of fixation or choice targets. On a given experimental session, on average M1 performed 1276 +/− 348 and M2 performed 1677 +/− 139 (mean +/− sd) LRS trials.

T1 and T2 were two identical in size rectangular stimuli (3-to-1 aspect ratio) with different orientation (Fig 1B). T1 was oriented so that the long axis was vertical and T2 was oriented so that the long axis is horizontal. Long axis of each target subtended 2 deg of visual arc. Two targets were associated with different liquid reward values. Each animal was motivated to select the target associated with the highest value of liquid reward. Mean value of the liquid reward associated with each target was kept constant for blocks of 40–70 trials (Fig 1C). The block transition was unsignaled. Mean reward values varied between 0.04 ml/trial and 0.21 ml/trial. On each trial, a Gaussian-distributed variability (SD = 0.015 ml) was added to the value with each target. Variable reward values further increased animal’s uncertainty about the times of reward block transitions. Since the choice behavior around each reward block transition was more exploratory (Supp Fig 8I), we performed all the analysis after excluding the first 10 trials after the block transition. This ensured that the animals followed the reward contingencies.

On each trial, target luminance values were randomly assigned. T1 luminance was randomly assigned from a log-uniform distribution of values ranging from 0.01 to 12.15 cd/m². The minimum luminance value was set above the psychophysical threshold for stimulus detection titrated during the ODR task. After the T1 luminance was assigned, the luminance of T2 was assigned such that mean luminance across both targets was 6 cd/m². On each trial, target luminance values were assigned independently from the rewards associated with T1 and T2. Additionally, the randomized spatial locations of two targets ensured that the target location of the high-reward and low-luminance target could not be determined from the low-reward and high-luminance target.

Trial-by-trial independent manipulation of luminance and reward values randomly yielded either congruent or conflict set of trials. On a given experimental session, on average Monkey 1 performed 322 +/− 74 congruent trials and 317 +/− 81 conflict trials; Monkey 2 performed 392 +/− 33 congruent trials and 392 +/− 39 conflict trials (mean +/− sd).

On congruent trials, luminance and reward values were both high for one target (Rich-Bright) and were both low for the other target (Poor-Dim). Each monkey showed a strong preference for selecting Rich-Bright target compared to Poor-Dim target (M1: 84% total trials: 9881; M2: 72% total trials 15615: across 39 and 42 experimental sessions).

On conflict trials, however, one target had high-reward and low-luminance (Rich-Dim) and the other target had low-reward and high-luminance (Poor-Bright). Conflict trials, when endogenous selection was expressed and Rich-Dim target was selected were termed as endo-conflict trials (on average each monkey performed M1=211 +/− 62, M2= 301+/− 44 endo-conflict trials per experimental session, mean +/− sd). Similarly, conflict trials when exogenous selection was expressed and Poor-Bright was selected, were termed exo trials (on average each monkey performed M1=107 +/− 47, M2= 91 +/− 25 exo-conflict trials per experimental session, mean +/− sd). Each monkey followed rewards and showed preference for selecting Rich-Dim target compared to Poor-Bright target (M1: 68% total trials: 9751; M2: 77% total trials: 15652 trials, across 39 and 42 experimental sessions).

Interestingly, M2 behavioral performance was better on congruent trials compared to conflict trials. We attribute the difference in M2 performance on conflict vs congruent trials to the role of a shape bias. On conflict trials, M2 preferred T1 compared to T2 (pT1= 82%, pT2= 73%) but this shape preference was not as strong on congruent trials (pT1 = 76%, pT2=73%). Conflict-specific shape preference explains the effect because the preference results in better performance for T1 on conflict trials compared with congruent trials (pT1-conflict = 82%; pT1-congruent = 76 %) but not for T2 (pT2-conflict = 73%; pT2-congruent = 73 %). On each trial, there are three sources of choice information: shape bias, luminance, and experience-dependent reward. On congruent trials, the luminance and reward information agree and so the shape bias does not tend to alter decisions and is not revealed on average- the effect is subthreshold. On conflict trials, the luminance and reward information disagree and so the shape bias tends to alter the decisions and is revealed on average- the effect rises above the threshold. Consequently, the difference in performance between conflict and congruent trials is due to the presence of shape bias in a manner that is consistent with the conceptual framework of the LRS task.

On each experimental session, on a subset of trials, the LRS task featured non-conflict reward-only and luminance-only trials. On reward-only trials, the luminance values of two targets were kept the same for blocks. On luminance-only trials, the average reward values associated with two targets were kept the same for blocks. On a given experimental session, on average Monkey 1 performed 220 +/− 98 reward-only trials and 294 +/− 120 luminance-only trials and Monkey 2 performed 245 +/− 29 reward-only trials and 337 +/− 57 luminance-only trials (mean +/− sd).

Neurophysiological experiments

Recording protocol and data acquisition:

Neural recordings were made with glass-coated tungsten electrodes (Alpha Omega, Israel) with impedance 0.7–1.5 M measured at 1 kHz (Bak Electronics, MD). Neural signals were preamplified (10 x gain; Multichannel Systems, Germany), amplified and digitized (16 bits at 30 kHz; NSpike, Harvard Instrumentation Lab), and continuously streamed to disk during the experiment (custom C and Matlab code). Neural recordings were referenced to a ground screw implanted in the left occipital lobe, with the tip of the screw just piercing through the dura mater.

In each animal, electrodes were advanced in each recording session to maximize the yield of isolated single units. Electrodes were advanced through a silastic membrane in the recording chamber, the dura mater and pia before entering the cortex. Each electrode was advanced sequentially in increments of 15 microns, 10 minutes apart to give the electrode time to settle in the tissue. Initial action potentials were recorded at a median depth of 3 mm (2.23 mm in M1; 3.04 mm in M2). Electrodes were gradually advanced across sessions (on average 34 μm/day in M1 and 100 μm/day in M2) until action potentials were no longer present, indicating passage into white matter. Neural recordings were made up to a median distance of 6 mm from their initial position.

Local field potential (LFP) activity was obtained offline by low-pass filtering the broadband raw recording at 300 Hz using a multitaper filter with a 1.5 ms time window. The low-pass filtered LFP activity was further downsampled to 1 kHz from 30 kHz. Multiunit activity (MUA) was obtained by high-pass filtering the raw recordings at 300 Hz and maintaining the original 30 kHz sampling rate. Single unit activity (SUA) was isolated by thresholding MUA activity at 3.5 standard deviations below the mean, performing a principal component analysis of putative spike waveforms, over-clustering these waveforms in PCA using k-means and then merging clusters based on visual inspection. Spike-sorting was performed for each recording session using custom Matlab code (Mathworks). Non-stationarity in recordings were accounted for by performing spike-sorting in 100 ms moving windows. Trials on which spike-clusters were not isolated were removed from further analysis.

Neuronal databases:

We advanced electrodes to isolate and record 746 units (M1: 384; M2: 362 units) during the ODR task. Out of 746 units, we further selected 409 (M1: 179; M2: 230 units) single units that were responsive to the LRS task. We selected units with firing rates greater than 5 sp/s in 0 to 200 ms epoch after onset of targets for the LRS task.

Each neuron’s response-field (RF) was mapped using the ODR saccade task to eight possible target locations. LPFC neurons showed increased firing in response to target onset alone, saccadic eye movement alone or both target onset and saccadic eye movement (Fig 1). Therefore, we computed each neuron’s trial-averaged baseline subtracted firing rate in response to eight target locations around target onset and saccade onset (Target onset: baseline epoch = [−200 0ms], stimulus epoch = [0 100ms] and [75 200ms] where 0ms is targets onset; Saccade onset: baseline epoch = [−400 200ms], stimulus epoch = [−50 70ms] where 0ms is saccade onset). We used these epochs to accommodate the firing activity of visual, visual-movement (VM) and movement neurons (Supp Fig 1-3). Each neuron’s RF was estimated against the null hypothesis that there is no difference in response firing rate with respect to baseline, using a permutation test. The baseline-subtracted firing rate at each target’s location was compared with the null distribution. Null distribution was generated by shuffling firing rate across eight target locations 1000 times (p<0.05, permutation test). Since this procedure involves multiple comparisons, we corrected the p values by controlling for the false discovery rate (FDR, ⁶⁸. Units with significant p-values either for target onset or saccade onset epochs were used for further analysis. Out of the 409 single units, we selected 216 neurons that showed an excitatory response inside the RF and had greater than 5 Hz firing rate either around target or saccade epoch (M1 = 122; M2 =139 neurons).

The ODR task further revealed the firing patterns of different LPFC neurons. We classified each unit that had an excitatory RF response into visual, visual-movement (VM) and movement neurons based on their firing patterns around target onset and saccadic eye movement. The delay period of the ODR task separated the visual and saccade related neuronal activity and allowed us to examine each neuron’s firing patterns in response to target and saccade onset. Around target onset, visual and VM neurons showed an increase in firing activity and not movement neurons. Additionally, visual neurons reflected an increase in firing rate immediately after the target onset whereas VM neurons showed a delayed response (Fig S1A-C, and Fig 1). Around saccade onset, VM and movement neurons showed an increase in firing activity and not visual neurons. Single unit responses at preferred target location were tested for selectivity around target onset and saccade onset through permutation testing. To classify between visual and VM neurons we compared each unit’s baseline-subtracted firing rate around target onset epochs (0 to 100 ms and 75 to 200 ms, where 0 ms is target onset). To classify between movement and VM neurons we compared each unit’s baseline-subtracted firing rate around saccade onset epoch (−50 to 70 ms where 0 ms is saccade onset). Units with significant p-values in target-onset (0 to 100 ms) epoch and not saccade-onset epoch were classified as visual neurons. Units with significant p-values in saccade-onset epoch and not target-onset epoch were classified as movement neurons. Units with significant p-values in both target-onset (75 to 200 ms) and saccade-onset epochs were classified as VM neurons. We further confirmed each unit’s classification label by visual inspection. Out of 261 units, N=139 (M1=54, M2=85) were VM neurons, N=57 (M1=27, M2=30) were visual neurons and N=65 (M1=41, M2=24) were movement neurons.

QUANTIFICATION AND STATISTICAL ANALYSIS

LRS task selectivity:

On the LRS task, the two targets were presented simultaneously. Therefore, on each trial, the location of both the targets with respect to a LPFC neuron’s RF was identified. For further analysis, we pooled the data across two monkeys to increase the statistical power. For each neuron, we selected the subset of trials on which one target was inside the RF and the other was outside the RF. Trials on which both the targets were inside the RF or both the targets were outside the RF were removed from further analysis. We examined each neuron’s selectivity to the LRS task based on the saccade response and the target properties. Trials on which saccade response was inside the RF were termed InRF trials and trials on which saccade response was outside the RF were termed OutRF trials. Fig 1F shows the population data of 139 VM neurons across 36864 InRF trials and 36455 OutRF trials. Similar to the ODR task, VM neurons responded significantly more on trials when the InRF target was selected compared to trials when the OutRF target was selected (p=4.2 × 10⁻³, Wilcoxon rank-sum test, epoch=50 to 200 ms). Firing rate increased soon after target onset and extended through the saccade. Movement neurons (N=65) also responded significantly more on the InRF (N=16450) trials compared to OutRF (N=16020) trials (p=5.7 × 10⁻⁵, Wilcoxon rank-sum test, epoch=150 to 250 ms). Visual neurons (N=57) however, showed comparable firing rates for InRF (N=14416) and OutRF (N=15651) trials (p=0.61, Wilcoxon rank-sum test, epoch=0 to 100 ms). The results were similar if different time-windows around the peak-firing rates were used ([56 304], [0 182], and [139 301] for VM, movement and visual neurons. These time-windows are determined based on half-firing rate, when the firing rates were half of the peak firing rate).

The InRF and OutRF trials were further subgrouped on the basis of attentional selection. Exo-InRF trials are exo-conflict trials on which Poor-Bright target was selected and target was in the RF, whereas Exo-OutRF trials are exo-conflict trials on which Poor-Bright target was selected and target was out of the RF. Similarly, Endo-InRF trials are endo-conflict trials on which Rich-Dim target was selected and the target was in the RF, whereas Endo-OutRF trials are endo-conflict trials on which Rich-Dim target was selected and the target was out of the RF. The subgrouping of InRF and OutRF trials based on attentional selection yielded the following number of trials for each subgroup. The trials were pooled across neurons in three cell-type (VM, visual and movement neurons) groups. VM neurons: Exo-InRF=5459, Exo-OutRF=5379, Endo-InRF=16842, and Endo-OutRF=16470 trials. Visual neurons: Exo-InRF=2308, Exo-OutRF=2567, Endo-InRF=644 and Endo-OutRF=6740 trials. Movement neurons: Exo-InRF=2548, Exo-OutRF=2540, Endo-InRF=7388 and Endo-OutRF=7230 trials.

Selection-time (ST) analysis:

We estimated the onset of selectivity in firing rates as the time after target onset when firing rates differed significantly for InRF and OutRF selection. We did this by first calculating the firing rates using a 15 ms smoothing window and then computing the difference in InRF and OutRF firing rates for each neuron. We tested the mean difference in firing rate for each group (Fig 2E) against a null hypothesis that there is no difference in firing rates using a permutation test. A null distribution of firing rate differences was generated by shuffling the InRF and OutRF firing rates across neurons in each group 1000 times. We detected ST as the first time-point when InRF firing rates were significantly greater than OutRF rates (p<0.01, permutation test). Since this procedure involves multiple comparisons, we corrected the p values by controlling for the false discovery rate.

Spike-field coherence analysis:

We estimated spike-field coherence (SFC) as a function of frequency using multitaper spectral estimation ^69,70 with 10 Hz smoothing, and an estimation window spanning 200 ms before the target onset. The SFC was estimated between spiking and nearby LFP activity (within approx. 1.5 mm) to account for spiking activity bleeds into the LFP recording (Fig S3F). There was no spike amplitude for the broad-band recording on the LFP electrode when the activity was triggered on the spike times recorded on the other electrode.

The significance of SFC for each spike-field pair was tested against a null hypothesis that there was no SFC using a permutation test (1000 permutations, p<0.05). Null distribution for no SFC was generated by randomly permuting the order of trials for the spiking data compared to the LFP data. Raw coherence values were converted to z-scores by subtracting the mean and then dividing by the standard deviation of the null distribution. We applied cluster correction to identify the significant clusters of p-values while accounting for multiple comparisons ⁷¹. The significant cluster in beta (15–35 Hz) and gamma (40–70 Hz) frequency range was selected after performing a permutation test (1000 permutations, p<0.05). The coherent and not-coherent spike-field pairs in beta and gamma frequency ranges were identified based on the presence of a significant cluster in respective frequency bands. We identified 176 (M1=59, M2=117) coherent and 233 (M1=120, M2=113) not-coherent pairs in beta frequency range. A small number of spike-field pairs (10 out of 179 in M1 and 11 out of 230 in M2) were coherent in the gamma frequency range.

Beta-amplitude analysis:

At each recording site, we tested whether beta bursts are specifically present in the 200ms prior to target onset against the null hypothesis of activity at other times during the trial using a permutation test. Across the population, beta bursts were reliably present in ~96 % of recording sites in each animal (M1: 1108 out of 1152 sites; M2: 1299 out of 1344 sites; p<0.05, permutation test). We estimated amplitude of pre-target beta-burst for each trial at a single site, using multitaper spectral estimation ^69,70. We used 5 Hz smoothing, and an estimation window from 200 ms before target onset until target onset. The power values in beta (15–30 Hz) frequency range were converted to amplitude by taking square root. The logarithm transform of beta amplitude values were normalized with respect to mean across trials. We used these normalized beta-burst amplitude values for further analysis. Beta values varied trial-by-trial and observed a gaussian distribution at a given site (Fig 3B).

We examine the time-course of beta-burst related modulation in firing rates for exogenous and endogenous selection, by computing the beta values in six different 200 ms long time-windows (Fig 6). If otherwise mentioned beta related modulations were referred to beta values computed in 200 ms time-window before the target onset.

For a given site, we grouped the trials with the highest ~33% and lowest ~33% beta values to yield high-beta (HB) trials and low-beta (LB) trials. We calculated the SFC separately for HB and LB trials for coherent and not-coherent neurons.

Beta-bursts and attentional selection:

We compared the firing responses of VM neurons on high-beta and low-beta trials for exogenous and endogenous selection. We further subgrouped the exo/endo InRF and OutRF trials based on beta values to yield the following number of trials for each subgroup. VM neurons -high-beta: Exo-InRF=1818, Exo-OutRF=1801, Endo-InRF=5599, Endo-OutRF=5458 trials, VM neurons-low-beta: Exo-InRF=1808, Exo-OutRF=1815, Endo-InRF=5524, Endo-OutRF=5513 trials. Similarly, for visual neurons we yielded, high-beta: Exo-InRF=792, Exo-OutRF=866, Endo-InRF=2166, Endo-OutRF=2222 trials and low-beta: Exo-InRF=772, Exo-OutRF=867, Endo-InRF=2110, Endo-OutRF=2306 trials. And for movement neurons we yielded, high-beta: Exo-InRF=852, Exo-OutRF=863, Endo-InRF=2459, Endo-OutRF=2345 trials and low-beta: Exo-InRF=849, Exo-OutRF=843, Endo-InRF=2468, Endo-OutRF=2445 trials.

Permutation test:

We tested the difference in firing rates on HB and LB trials for each group in Fig 4C, 5C and 6C. We computed the difference in firing rates between LB and HB trials for each neuron and tested the mean difference across neurons against a null hypothesis that there is no difference in firing rates using a permutation test. A null distribution of firing rates difference was generated by shuffling the HB and LB firing rates across neurons in each group 1000 times (p<0.01, permutation test). Since this procedure involves multiple comparisons, we corrected the p values by controlling for the false discovery rate.

We tested the difference in RTs after stratifying trials according to beta value for endogenous-conflict and exogeneous-conflict trials as shown in Fig 7A. We performed this test separately for each monkey. For each group of trials, we computed the test statistic given by the maximum difference in RT (range = max RT - min RT) after stratifying trials by beta value. We then tested the hypothesis that the difference in RT across beta values differed for the specific group of exogenous or endogenous trials using a permutation test. A null distribution of the test statistic was generated by shuffling the trial labels within each group (exogenous-conflict and endogenous-conflict) separately. We then computed the RT for trials stratified by beta values as for the original data set. RT for each group was standardized to be mean 1 before permuting by dividing by the mean RT in each group. Beta values for each group were standardized to be mean zero for each group before permuting by subtracting the mean beta value in each group. We performed this permutation 1000 times and compared the maximum difference in RT for each permutation with the test statistic (p<0.01, permutation test). We used an analogous procedure to test for a significant difference in RTs after stratifying trials according to beta value for luminance-only and reward-only trials, as shown in Fig 7B. Since this procedure was performed once per monkey and group, it was not necessary to control for multiple comparisons.

Highlights.

• Visual-movement neurons encode exogenous and endogenous attentional selection

• Coherent-beta activity selectively modulates mechanisms of exogenous selection

• Beta-activity suppresses exo-selection when in conflict with endo-selection

• Beta-activity-mediated mechanisms selectively inhibit visual-movement neurons