Cortico-subcortical dynamics in primate working memory

Goal-directed behavior relies heavily on working memory, the capacity to retain sensory or internal infor- mation and use it to act according to current rules and task demands1–4. Working memory is a dynamic process with multiple stages from memory encoding to action that in- volves numerous interacting neural elements. Despite exten- sive research, the mechanisms by which the brain supports working memory are still not fully understo od. This gap in understanding is partly due to the fragmented study of both the dynamics of working memory and the neural elements that jointly support it. Regarding the dynamics of working memory, many stud- ies have concentrated on mnemonic aspects such as memory encoding and maintenance, often overlooking memory uti- lization and the processes connecting these stages. While re- cent research has begun to address this gap5, invasive studies examining the neural basis of working memory from encod- ing to action are still rare. In terms of neural elements, pioneering work has ex- plored the neural correlates of working memory in various cortical and subcortical regions. Prefrontal and parietal cor- tical areas are known to play crucial roles in working memory6–10. Additionally, subcortical structures, particu- larly higher-order thalamic areas and basal ganglia regions, also likely contribute to working memory 11–17. However, their fragmented study prevented a direct quantitative char- acterization how memory information is distributed across the cortico-subcortical network. Furthermore, working memory, like other cognitive functions, fundamentally depends on interactions between multiple brain regions 18–22. Despite this widely accepted view, with a few notable exceptions21,23,24, electrophysiologi- cal studies of working memory in primates have been con- ducted using recordings from one or at most two regions at a time. Studies using simultaneous recordings from cortical and subcortical areas are even more scarce. Consequently, due to the lack of simultaneous recordings, the interactions between areas that support working memory remain largely unknown. In sum, three central issues and related questions regard- ing the neural basis of working memory persist. First, what are the dynamics of memory information from encoding to utilization? More specifically, is working memory infor- mation reloaded towards its use? If so, in which regions and is such an enhancement of information task specific? Sec- ond, what specific roles do key cortical and subcortical areas play? H ow is working memory information quantitatively distributed across the cortico -subcortical network? And, to what extent do higher-order thalamic nuclei and basal gan- glia support other working memory features than spatial memory? Third, what is the structure of cortico-cortical and cortico-subcortical interactions supporting working memory? In particular, what is the dominant direction of in- teractions within the frontoparietal network or between cor- tical and subcortical regions? And how are network interac- tions modulated during working memory from encoding to action? .CC-BY-NC-ND 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted July 25, 2024. ; https://doi.org/10.1101/2024.07.24.605013doi: bioRxiv preprint Cortico-subcortical dynamics in primate working memory von Nicolai & Siegel 2024, bioRxiv 2 To answer these questions, we simultaneously recorded neural activity from prefrontal, parietal, thalamic, and basal ganglia regions in macaque monkeys during color and spa- tial working memory tasks. These large -scale recordings provided us with the unique opportunity to directly compare mnemonic information in all these structures and to quan- tify their interactions. We found that color and spatial memory information are widely distributed across all investigated cortical and sub- cortical structures and that they exhibit dynamics that vary by area and in a task-dependent manner. Notably, the utili- zation of working -memory information for action is pre- ceded by the reloading of task-relevant information in infor- mation-specific cortico-subcortical networks. Reciprocal in- teractions between cortical and subcortical areas are ubiqui- tous, dynamic, and partially task-specific, with interactions again being particularly strong during memory use for ac- tion. Overall, cortical areas LIP and FEF appear to serve a coordinating role in this network by driving other regions.

Task, recordings, and behavioral results We trained two macaque monkeys in a working memory task in which the animals had to memorize either the color or the position of a peripheral cue over a delay period of 1.25 s and to respond with a saccade either to the matching one of two colored peripheral targets or to the position of the pre- ceding cue (color and spatial task, respectively; Fig. 1A and Methods). Importantly, the visual input was the same for both tasks until the beginning of the response period. Fur- thermore, in the spatial task, animals could in principle plan the required upcoming eye movement to the target to be memorized from the cue presentation onwards, whereas in the color task they could not. The conditions were presented in blocks without explicit cueing of the current task rule, i.e., the task at hand was implicit. Both animals performed well on both tasks, with percentage correct performance being 83% (monkey V) and 80% (monkey E) for the color task and 94% (monkey V) and 92% (monkey E) for the spatial working memory task. Once the animals had mastered the task, we implanted recording chambers over large parts of one hemisphere (Supplementary Figure S1). This provided simultaneous ac- cess to seven frontal (A46, A9/46, frontal eye field - FEF), parietal (lateral intraparietal area - LIP), higher order tha- lamic (mediodorsal thalamus - MD and Pulvinar) and basal ganglia (Striatum) brain regions (Figure 1B, Supplementary Figure S1 and Methods). We simultaneously recorded extra- cellular spiking activity from these regions using multi-con- tact microelectrodes. Dynamics of spiking activity We found that spiking activity was substantially modu- lated by task events in all recorded areas (Figure 1C and D). This temporal modulation was strongest in LIP and weakest in the Striatum and prefrontal areas (Figure 1C). The tem- poral profile of spiking d ynamics was generally similar across brain regions. Spiking transiently increased following stimulus onset and then decreased towards the memory de- lay. Interestingly, during the delay, spiking remained ele- vated only in area LIP, but dropped below baseline in all other areas. Towards the response period, spiking signifi- cantly increased again in all regions. Figure 1. Behavioral tasks, target re- gions and spiking activity (A) Behav- ioral tasks. Monkeys performed a color and a spatial working memory task in block-wise manner. For both tasks, a colored cue (blue or red) was presented at one of four spatial locations (90, 180, 270, 360 degrees) for 0.5 s. Following a 1.25 s delay and fixation spot offset, monkeys either had to sac cade to the matching color target presented together with the non-matching target in a random spatial configuration across both main diagonals (color task) or towards the re- membered stimulus location (spatial task). For correct responses, a liquid re- ward was delivered, and a green square was flashed. ( B) 3D -rendering of the brain and target regions for recordings. (C) Spiking activity modulation quanti- fied as the coefficient of variation across trial time (standard deviation relative to the mean). (D) Average normalized spik- ing activity during both tasks. Shaded re- gions denote SEM. Numbers in brackets and on the right indicate number of elec- trodes and baseline firing rates, respec- tively. B C A46 A9/46 FEF Striatum LIP MD Pulvinar Time Fixation Cue Delay Color Task Spatial Task Go Response = 2 s Hold Feedback ITI A D 0 0.3Spiking modulation (norm.) Time (s) 0 0.1 0 0.5 1 1.5 0 Cue Resp. A46 (752) A9/46 (1048) FEF (928) LIP (864) MD (952) Pul (784) Area Str (664)Spiking activity (norm.) 8 Hz 7 Hz 17 Hz 12 Hz 18 Hz 12 Hz 4 Hz 0 0.1 0 0.1 0 0.8 0 1.5 0 0.6 0 0.3 A46 A9/46 FEF LIP MD Pul Str .CC-BY-NC-ND 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted July 25, 2024. ; https://doi.org/10.1101/2024.07.24.605013doi: bioRxiv preprint Cortico-subcortical dynamics in primate working memory von Nicolai & Siegel 2024, bioRxiv 3 Dynamics of memory information To address the first two questions raised above, about the dynamics of working memory information and how it un- folds in cortical and subcortical areas, we quantified the in- formation about critical task variables encoded by spiking activity. Specifically, f or each time point throughout the task, we determined the variance in spiking that could be ex- plained by the color and spatial location of the memory cue across trials (color and spatial information) Importantly, our approach ensured that information about each task variable was quantified independently from all other variables (see Methods). Figure 2 shows the time course of neural information during correct trials about the color and spatial position of the cue from the prestimulus baseline through the memory delay to response onset (Figure 2A) as well as the average information and correspon ding statistics during three suc- cessive time windows: cue presentation (0 -500 ms post cue onset), early memory delay (250-750 ms post cue offset), and late delay before the response (500-0 ms before response on- set) (Figure 2B). To quantify and statistically assess the tem- poral dynamics of information, we computed differences of information between successive task periods (Figure 2C). We found that, following a variable rise upon cue onset, during cue presentation, color and spatial information was present in all recorded cortical and subcortical areas (Figure 2A and 2B, first row). Following cue offset, information dropped again in most areas, while notably MD and A9/46 showed stronger color information during the early delay of the color and the sp atial task, respectively, as compared to the cue period (Figure 2C, first row). Throughout the entire memory delay, all cortical and subcortical areas significantly encoded color or spatial information with variable strength (Figure 2B, middle and bottom row). Thus, despite the wide- spread suppression of spiking activity, during the memory delay (Figure 1D), task -relevant memory information was broadly distributed across cortical and subcortical struc- tures. Also, in the late delay interval, when the memory-based response was imminent, color or spatial information was sig- nificant in all recorded cortical and subcortical regions (Fig- ures 2B, bottom row). Moreover, there was a marked in- crease of memory information towards the response in dis- tinct sets of regions depending on the task and information at hand (Figure 2A and C). In sum, during cue presentation and the following memory interval, visual color and spatial information was broadly encoded across parietal (LIP), frontal (FEF) and pre- frontal cortex (A46, A9/46), the thalamus (MD, Pulvinar) and Striatum. Furthermore, in t his network, working 0 5 0 0 2 A46 A9/46 FEF LIP MD Pul Str Color 1.5 0 25 0 5 0 10 A46 FEF LIP MD Pul Str Space A9/46 Cue Early Delay Late Delay Information (%EV) 0 0.8 0 0.8 0 4 0 1.5 Time (s) 0 0.5 1 1.5 0 0 Information (%EV) Color task Spatial task 0 1.5 0 1.5 0.8 Color Space Cue Resp. 0 1 0 0.8 0 10 0 30 0 4 0 3 0 0.5 1 1.5 0 1 0 A B A46 A9/46 FEF LIP MD Pul Str -.04 0 .01 -.015 0 .015 A46 A9/46 FEF LIP MD Pul Str ∆ Information (%EV) Early Delay - Cue Late - Early Delay ColorC -.2 0 .05 -.06 0 .06 A46 A9/46 FEF LIP MD Pul Str Space Figure 2. Dynamics of memory in- formation. (A) Time courses of infor- mation about the color and spatial position of the memory cue during correct trials of the color (light) and spatial (dark) working memory tasks. (B) Fixed -window statistics on color and spatial information during 3 task periods. (C) Fixed-window difference statistics on temporal changes of color and spatial information. Shaded regions and error bars de- note SEM. Colored crosses denote significance (yellow, purple, green: p < .05, .01, .001, respectively; paired and unpaired t -tests, FDR -cor- rected). .CC-BY-NC-ND 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted July 25, 2024. ; https://doi.org/10.1101/2024.07.24.605013doi: bioRxiv preprint Cortico-subcortical dynamics in primate working memory von Nicolai & Siegel 2024, bioRxiv 4 memory information was prominently reloaded towards the memory-guided behavioral response. Task dependence of memory information How did the behavioral task that the animals had to per- form (report either memorized color or space) affect the en- coding of visual information? We found that particularly the reloading of memory information towards the response was significantly modulated by the task at hand (Figure 3). Dur- ing cue presentation and the early delay, both color and spa- tial information were stronger almost exclusively in the color task (Figure 3A and B). This changed markedly during the late delay when utilization of mnemonic information for the response was imminent. At this point, not only did color and spatial information increase relative to the early delay (Figure 2C bottom row), but color and spatial information were specifically enhanced in the task for which the corre- sponding information was behaviorally relevant. In the late delay, color information was significantly stronger when color had to be reported and spatial information was signifi- cantly stronger when space had to be reported (Figure 3C). This task-dependence was due to a selective reloading of in- formation. The irrelevant information showed no significant enhancement towards the response (Figure 2C). This task - dependent modulation towards the response was significant in distinct sets of cortical and subcortical regions depending on the task and information at hand (Figure 3C). Despite this abundance of the late task-dependent mod- ulation, the time -course of this effect differed between re- gions. MD appeared to be the first structure to exhibit a clear task-specific enhancement of color information already dur- ing the early delay period (Figure 2A and 3B). A weaker early enhancement during the delay was also seen in in the Pulvi- nar, Striatum and LIP (Figure 2A and 3B). In sum, we found that encoding of visual information was modulated by the task at hand. Towards the behavioral response, specifically the memory information that was relevant for the task at hand was enhanced in spe- cific cortico-subcortical networks. Interactions between space and color The above analysis quantified information about color and space as independent variables. In another set of analysis, we investigated the dynam- ics of information about the interaction of color and space, i.e. neural activity that encoded a specific color at a specific spatial location (Figure 4A and B). Do only visuomotor areas such as LIP or FEF carry retinotopically specific memory information? Or is memory information also spatially specific in other cortical or subcortical regions? We found evidence for the latter. Throughout the trial spatially specific color in- formation was strongest in LIP and FEF. However, during cue presentation, also A46 exhibited a sig- nificant interaction effect (Figure 4A and B). In ad- dition, in the early delay the interaction between color and space was also significant in the color task in the Pulvinar and Striatum whereas in the spatial task, it was significant only towards the response in all areas except for MD and the Striatum. Thus, although spatially specific color information was most prominent in visuo-motor areas LIP and FEF it was also broadly distributed across the cor- tico-subcortical network. Choice information During the color task, animals made errors, i.e. they did not always choose the cued color (17% for monkey V; 20% for monkey E). This allowed us to investigate neural infor- mation about the animals’ color choice independent of the sensory input (Figure 4C). Importantly, as the color response targets were spatially randomized, color choice information was also independent of the spatial motor response. We found significant choice information, i.e. neural ac- tivity that predicted the animals’ color choice independent of the sensory cue and motor response, during all trial inter- vals including the fixation baseline. Choice information was generally strongest in the late delay before the behavioral re- sponse, but the distribution of choice information across brain regions changed markedly throughout the trial. Dur- ing the pre-cue baseline, choice predictive information was strongest in prefrontal (A46, A9/46) an d subcortical (MD, Pulvinar, Striatum) regions and not significant in LIP and FEF (Figure 4A). During cue presentation, choice infor- mation was strongest in the thalamus (MD, Pulvinar) and significant in LIP. During the delay, choice information in- creased such that during the late delay before response on- set, choice predictive information was broadly distributed and significant across all cortical and subcortical areas. In sum, choice information showed a dynamic remapping from being confined to associative and subcortical regions before the sensory input to a broad distribution across the entire sensorimotor pathway and subcortical regions towards the choice report. 0 12 x10-3 x10-3 A46 A46 A46 A9/46 A9/46 A9/46 FEF FEF FEF LIP LIP LIP MD MD MD Pul Pul Pul Str Str Str Cue 0 6 Early Delay -0.08 0 0.02 Late DelayA B C ∆ Information (%EV) Color task - Spatial task Color information Spatial information Figure 3. Task dependence of memory information. (A) Fixed-window statistic on task-related differences (color task - spatial task) of color and spatial information in the cue interval. Positive and negative values indicate more information in the color and spatial task, respectively. (B) Task-related differences in the early delay interval. (C) Task-related differences in the late delay interval. Error bars denote SEM. Col- ored crosses denote significance (yellow, purple, green: p<.05, .01, .001, respec- tively; unpaired t-tests, FDR-corrected). .CC-BY-NC-ND 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted July 25, 2024. ; https://doi.org/10.1101/2024.07.24.605013doi: bioRxiv preprint Cortico-subcortical dynamics in primate working memory von Nicolai & Siegel 2024, bioRxiv 5 In sum, we found widespread dynamic correlates of working memory for action in cortical and subcortical re- gions. Memory information was broadly distributed across the cortex, thalamic nuclei and basal ganglia. Task-relevant memory information was enhanced when memory -based action was impending in distinct but overlapping sets of cor- tical and subcortical areas with thalamic nuclei (MD and Pulvinar) showing an early task -specific enhancement of memory information. Furthermore, color choice predictive information was already present in subcortical and associa- tive cortical regions before the memory cue and was then widely distributed across all cortical and subcortical regions before the choice report. Directed interactions between cortical and subcortical ar- eas We next asked how cortical and subcortical structures in- teracted during the working memory task. To investigate this, we computed Granger causality between spiking activ- ity from all simultaneously recorded areas (Figure 5). We ap- plied strict statistical cr iteria to rule out spurious interac- tions (see Methods). Specifically, we required a given area pair to fulfill three requirements: first, significant shuffle - corrected Granger causality; second, significant interareal asymmetry of Granger causality, indica ting a dominant direction of information flow; and third, a reversal of the interareal asymmetry for time -reversed Granger causality 25,26. In addition to computing Granger causality between pairs of areas, we also determined each areas’ average interaction with all other structures. We found widespread, reciprocal interactions between areas in both tasks (Figure 5A, small full matrices). For each pair of areas, we subtracted both directions of interactions to quantify the asymmetry or dominant direction of information flow. FEF and LIP were by far the strongest send- ers, interacting with almost all other areas in a driving manner. In contrast, all other structures were by and large net receivers, including thalamic areas MD and Pulvinar as well as the Striatum (Figure 5A, large triangul ar matrices). Interest- ingly, interactions between prefrontal areas A46 and A9/46 were very rare. They were net receivers of information from FEF and LIP but had a net driving influence on MD and Pulvinar. Most im- portantly, directed interactions were strongest dur- ing the late delay when memory information was needed for response generation. As with information encoding, we next tested whether the structure of directed interareal inter- actions was dynamic (Figure 5B). We computed differences in Granger causality between the same consecutive temporal window pairs as above. In addition, given the substantial interactions already during the fixation baseline, we also quantified the change in Granger causality from fixation to memory cue presentation. Directional coupling in- creased with cue presentation, particularly in area pairs with LIP and FEF in both tasks (Figure 5B, small full matrices). Also, the net drive exerted by FEF and LIP on thalamic areas MD and Pulvinar increased with cue presentation (Figure 5B left, large triangular matrices). Interestingly, some pairs of areas in the color task also showed reduced interaction dur- ing cue presentation. From cue presentation to early delay, interactions mostly decreased, and the net driving influence of FEF and LIP on prefrontal and thalamic areas also de- creased. From early to late delay, interactions became ubiq- uitously stronger and the net driving influence of FEF and LIP on MD and Pulvinar increased again. Finally, we examined whether the structure of directed interactions was task -specific (Figure 5C). During fixation, raw Granger causality (small full matrices) was stronger in the color task than in the spatial task, whereas directional asymmetries (large triangular matrices) were hardly task - specific, except for a stronger drive from LIP to FEF in the spatial task. Interestingly, during cue presentation, the task- specificity of directional asymmetries was limited to interac- tions involving the largely visual area LIP. During the early delay, raw interactions were generally stronger in the spatial task, and this task-specificity was particularly pronounced in area pairs involving frontal areas A46, A9/46, and FEF. 0 3 0 0 1 1.5 A46 A9/46 FEF LIP MD Pul Str Color x Space Color choice Cue Early Delay Late Delay Information (%EV) 0 0.2 Fixation 0 0.1 Cue 0 0.2 Early Delay 0 0.4 A46 A9/46 FEF LIP MD Pul Str Late Delay Information (%EV) Information (%EV) Color x Space 0 0.8 0 0.8 0 0.8 0 4 0 0.8 0 0.8 0 0.5 1 1.5 0 0.8 0 A B C Time (s) Cue Resp. A46 A9/46 FEF LIP MD Pul Str Color task Spatial task Figure 4. Interaction between color and spatial information and color choice in- formation. (A) Time courses of information about the interaction between the color and spatial position of the memory cue during color (light) and spatial (dark) working memory tasks. ( B) Fixed -window statistics on information about the interaction be- tween color and space during 3 task periods. ( C) Fixed-window statistics on chosen color information during four periods of the color working memory task. Shaded re- gions and error ba rs denote SEM. Colored crosses denote significance (yellow, pur- ple, green: p<.05, .01, .001, respectively; paired and unpaired t-tests, FDR-corrected). .CC-BY-NC-ND 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted July 25, 2024. ; https://doi.org/10.1101/2024.07.24.605013doi: bioRxiv preprint Cortico-subcortical dynamics in primate working memory von Nicolai & Siegel 2024, bioRxiv 6 However, there were hardly any task -specific directional asymmetries during this period. This changed markedly dur- ing the late delay, when the animals prepared for action based on working memory. Towards the behavioral re- sponse, many pairs of areas showed task-specific interac- tions and directional asymmetries. In particular, MD and Pulvinar showed stronger sending relationships with many areas during the color and spatial tasks, respectively. Over- all, frontal, parietal, and thalamic structures showed sub- stantial task-specific modulation before the response. In summary, we found ubiquitous cortico -cortical and cortico-subcortical directed interactions that were clearly structured and dynamically modulated during the working memory tasks. Frontoparietal areas FEF and LIP played ma- jor driving roles during all ta sk phases, whereas prefrontal and subcortical areas were predominantly receivers of infor- mation. Overall, directed interaction increased significantly during preparation for action. During this period, and espe- cially between frontoparietal and thalamic regions, directed interactions were also most task specific.

Our study offers new and comprehensive insights into the large -scale circuit mechanisms that underlie working memory. We leveraged an advanced experimental approach with simultaneous large-scale recordings from several corti- cal and subcortical regions of the primate brain during a flex- ible working memory task. Our approach revealed that color and spatial memory information is pervasive in prefrontal, parietal, thalamic, and basal ganglia structures, exhibiting area-specific and task -dependent dynamics. We o bserved that the utilization of working memory information for ac- tion is preceded by the reloading of task -relevant infor- mation in overlapping but distinct cortical and subcortical regions. Additionally, reciprocal interactions between these areas were widespread, dynamic, and partially task-specific, with the strongest interactions occurring during memory use for action. Specific cortical areas, particularly LIP and FEF, played a dominant role in driving activity in other re- gions. Neural dynamics of working memory processing Our findings enhance the limited understanding of the neural dynamics involved in working memory -dependent action. It has been suggested that internal selective attention functions are crucial for both mnemonic and action-related aspects of working memory4. This involves enhancing rele- vant over irrelevant working memory representations for ac- tion27. Our observation of task -specific enhancement of ei- ther color or spatial information in terms of a reloading of Fixation Cue Early Delay Late Delay A46 A9/46 FEF LIP MD Pul Str All A46 A9/46 FEF LIP MD Pul Str Color - Space C +4.6 -4.6 t-score Triangles +19 -19 t-score Row Column Column RowFull B A46 A9/46 FEF LIP MD Pul Str All A46 A9/46 FEF LIP MD Pul Str Cue - Fixation Early Delay - Cue Late - Early Delay Color Space +19 -19 t-score Triangles +37 -37 t-score Full Row Column Column Row Late Delay A46 A9/46 FEF LIP MD Pul Str All A46 A9/46 FEF LIP MD Row Column GC Full .01 0 Row Column Column Row +24 -24 t-score Triangles Pul Str Fixation Cue Early Delay Color Space A Figure 5. Directed cortico -subcortical interac- tions. Granger causality (full matrices) and asym- metry of Granger causality (triangular matrices) between spiking activity in pairs of brain regions. Full matrices show interactions in both directions. Triangular matrices show the asymmetry of inter- actions, i.e. the difference between both direc- tions. The last ro w and/or column depict the cor- responding marginal interactions. ( A) Full di- rected interactions and directional asymmetries during four periods of the working memory tasks. Asymmetries were computed by contrasting inter- actions of both directions of a given area pair. Asymmetries are shown as t -scores across elec- trode pairs. For full matrices, rows and columns correspond to sources and targets of interactions, respectively. For triangular matrices, asymme- tries are from row to column for positive values and re versed for negative values. ( B) Temporal differences of interactions across three succes- sive task periods. Plots show t -scores of window differences across electrode pairs. ( C) Differ- ences of directed interactions between color and spatial working memory t asks. Plots show t - scores of task differences (color - space) across electrode pairs. Green dots denote area pairs with significant Granger causality in either direction af- ter shuffle correction (p<.05; FDR -corrected) and a significant asymmetry (p<.05; FD R-corrected) that reverses for time-reversed data. .CC-BY-NC-ND 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted July 25, 2024. ; https://doi.org/10.1101/2024.07.24.605013doi: bioRxiv preprint Cortico-subcortical dynamics in primate working memory von Nicolai & Siegel 2024, bioRxiv 7 information towards the response supports this idea. Fur- thermore, our results indicate that the cortico -subcortical networks prioritizing task-relevant information are specific to the type of information. While all areas supported infor- mation encoding and maintenance, reloading of task -rele- vant information towards the response was task- and infor- mation-specific. During the late delay, color information increased in FEF, MD, and Pulvinar in the color task, but not in the spa- tial task. Conversely, spatial information increased only in the late delay of the spatial task in A9/46, FEF, LIP, MD, Pul- vinar, and the Stria tum. Memory information notably in- creased late during the delay in both tasks, suggesting an ac- tion-bound and task -specific selection of memory infor- mation in large-scale cortical-subcortical networks4,22. Our study also revealed intriguing dynamics of stimulus- independent neural activity predicting the animals’ color choice. Choice information was present during the fixation baseline in many areas, excluding FEF and LIP, but became broadly distributed across all cortical and subcortical areas towards the response. These findings align with the distrib- uted encoding of choices observed in cortical and subcortical areas in rodents28 and primates29. Moreover, the choice sig- nal dynamics demonstrated in the present study suggest that interactions between rostral prefrontal regions, along with associated higher-order thalamic and basal ganglia regions, may influence activity in structures more dedicate d to memory-dependent action preparation. Cortico-subcortical interactions during working memory Our large-scale simultaneous recordings allowed us to, for the first time, characterize directed interactions between key cortical and subcortical regions during working memory in the primate brain. Critically, we directly quantified inter- actions at the level of spiking activity, which avoids potential interpretational ambiguities at the level of populations sig- nals such as e.g. local field potentials. We identified strong reciprocal interactions, with many area pairs showing direc- tional asymmetry, indicating a dominant direction of inter- action. These interactions were dynamic, with a prominent enhancement towards the behavioral response. They were also task-specific, with distinct connections being enhanced during either color or spatial working memory. Overall, LIP and FEF were primarily net senders of infor- mation, while prefrontal and subcortical regions predomi- nantly received information. These findings align with pre- vious research that demonstrated a stronger driving influ- ence of the parietal cortex o n the prefrontal cortex during working memory21, supporting the idea that frontoparietal network interactions are crucial for working memory pro- cessing21,30–32. Our approach also enabled the examination of interac- tions between frontoparietal areas and their primary higher- order thalamic partners. The MD and Pulvinar thalamic nu- clei are known to facilitate interactions between cortical ar- eas during cognitive tasks 13,14,33–39. While previous studies have shown that the thalamus, such as MD on the prefrontal cortex13,14,38,39 or the Pulvinar on parietal areas 35,36, can exert a strong driving influence, this seems to depend on the spe- cific task and stage examined. Reciprocal interactions ap- pear to be the norm rather than the exception33. Indeed, we observed significant reciprocal interactions between cortical and thalamic areas, with cortical drive often exceeding tha- lamic drive in both working memory tasks. Additionally, thalamic areas demonstrated substantial task-specific infor- mation processing, sometimes even earlier than cortical ar- eas, such as MD during the color task. These results high- light the important role of corticothalamic interactions in working memory and suggest an early, task -dependent se- lection of memory information in subcortical structures, in- dependent of dominant cortical influence. Specific roles of cortical and subcortical regions in working memory Our study offers novel insights into the specific roles of various cortical and subcortical regions in working memory, uncovering task -dependent dynamics and emphasizing their contributions to processing and utilizing memory in- formation for action. Notably, the simultaneous recordings during the same behavior allowed us to directly quantify the distribution of memory information across a broad cortico - subcortical network. We did not observe a particularly prominent role of pre- frontal areas A46 and A9/46 in working memory infor- mation coding and directional coupling. At first glance, this may seem contrary to their known involvement in working memory40,41. However, many prior investigations on prefron- tal mechanisms featured recordings from caudal PFC re- gions like FEF (area 8Ad). Indeed, we found that FEF exhib- ited strong task -dependent memory encoding and exerted robust influences on all other areas except LIP. This con- firms previous studies suggesting that FEF, as a mixed visuo- motor structure, is crucial for preparing eye movements 42 and is a key component of attention and working memory networks, at least in visual tasks requiring eye move- ments43,44. Our results suggest that FEF plays a central role in transforming memory information into eye-movement re- sponses through interactions with cortical (LIP) and subcor- tical (MD, Pulvinar) partners. Future studies are required to test if more demanding tasks rely more on coordinating ac- tivity in rostral frontal areas like A46 and A9/4644,45. Parietal area LIP exhibited the strongest color and spatial information coding during cue presentation, which aligns with its role in visual and spatial cognition46. Unlike FEF or thalamic areas, LIP’s color information was not enhanced during the late delay when memory information was used for action. However, LIP was the strongest source of directed interactions during both cue presentation and the late delay of both tasks, even compared to FEF, and had a net driving influence on thalamic areas throughout the delay. Our re- sults suggest that LIP significantly contributes to infor- mation maintenance during mnemonic delays and provides essential input to areas like FEF for generating behavioral responses44,47,48. Thalamic areas, particularly MD, were first to show task- dependent differentiation of cue information coding during .CC-BY-NC-ND 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted July 25, 2024. ; https://doi.org/10.1101/2024.07.24.605013doi: bioRxiv preprint Cortico-subcortical dynamics in primate working memory von Nicolai & Siegel 2024, bioRxiv 8 the memory delay and towards the behavioral response. MD has long been associated with working memory processing, similar to the prefrontal cortex, its main input and output region10,11,13,49,50 (but see 3,51–53). Our findings align with previ- ous evidence that MD supports sensory -to-motor transfor- mations and coordinates frontal cortex computations in a task-dependent manner 12–14,33,49,54–56. Prior studies on MD have primarily focused on spatial working memory tasks13,57– 59. Our results indicate that MD supports working memory more generally by maintaining and reloading also non-spa- tial memory information needed for action. The Pulvinar, unlike MD, has been primarily linked to visual sensory and attentional processing. However, it is a higher-order thalamic nucleus with extensive connections to occipital, temporal, parietal, and frontal cortex areas, includ- ing FEF and dorsolateral PFC60. This suggests that in partic- ular the dorsomedial Pulvinar studied here is involved in cognitive processes like working memory 61,62. Indeed, we found that the Pulvinar exhibits working memory -related coding and connectivity similar to MD, with differences in interaction partners during specific task stages. Given their overlapping connections, this can be expected, but the Pul- vinar’s broader connections might lead to different dynam- ics. Future studies should clarify how MD and the dorsome- dial Pulvinar support cognition together. In any case, our re- sults emphasize the important roles of higher -order tha- lamic nuclei and thalamocortical loops in cognition17,52,54,63. The Striatum, a significant site of dopaminergic influ- ence15,16,64, has been linked to working memory processing65 because it is the primary input structure of the basal ganglia for cortical projections, establishing cortico -basal ganglia - thalamocortical loops18,66. Early studies showed striatal ac- tivity modulation during different stages of working memory tasks67, pointing to a possible role as a subcortical gate- keeper65,68. Indeed, recent fMRI studies found evidence for corticostriatal output gating during working memory tasks69. The Striatum was the least prominent sender and receiver in the interaction analyses, possibly due to challenges in target- ing interconnected sites in this large structure or the nature of the tasks used70,71. However, we found significant cue-re- lated color and spatial memory information in the Striatum, which suggests its contribution to both spatial and non-spa- tial working memory for action in the primate brain.

Our findings indicate that task -relevant information is widely distributed across key cortical and associated subcor- tical areas during color and spatial working memory tasks, with dynamics that vary by task and area, primarily bound to action. Both the fro ntoparietal cortex and higher -order thalamic regions play significant roles in working memory processing with frontoparietal areas FEF and LIP serving central, coordinating functions. Extensive interactions be- tween cortical and subcortical areas reflect task engagement and adapt to task demands, particularly during memory for action. These insights clarify how working memory inte- grates sensory information with its use for action bridging the perception-action cycle4,22,72.

1. A. Baddeley, Working memory: theories, models, and controversies. Annu Rev Psychol 63, 1–29 (2012). 2. M. D’Esposito, B. R. Postle, The cognitive neuroscience of working memory. Annu Rev Psychol 66, 115–142 (2015). 3. E. K. Miller, M. Lundqvist, A. M. Bastos, Working Memory 2.0. Neuron 100, 463–475 (2018). 4. F. van Ede, A. C. Nobre, Turning Attention Inside Out: How Working Memory Serves Behavior. Annu Rev Psychol 74, 137–165 (2023). 5. F. van Ede, S. R. Chekroud, M. G. Stokes, A. C. Nobre, Concurrent visual and motor selection during visual working memory guided ac- tion. Nat Neurosci 22, 477–483 (2019). 6. B. Pesaran, J. S. Pezaris, M. Sahani, P. P. Mitra, R. A. Andersen, Tem- poral structure in neuronal activity during working memory in ma- caque parietal cortex. Nat Neurosci 5, 805–11 (2002). 7. M. Siegel, M. R. Warden, E. K. Miller, Phase-dependent neuronal cod- ing of objects in short -term memory. Proc Natl Acad Sci U S A 106, 21341–6 (2009). 8. S. Funahashi, C. J. Bruce, P. S. Goldman -Rakic, Mnemonic coding of visual space in the monkey’s dorsolateral prefrontal cortex. J Neuro- physiol 61, 331–349 (1989). 9. G. Rainer, W. F. Asaad, E. K. Miller, Selective representation of rele- vant information by neurons in the primate prefrontal cortex. Nature 393, 577–579 (1998). 10. J. M. Fuster, G. E. Alexander, Neuron activity related to short -term memory. Science 173, 652–4 (1971). 11. J. M. Fuster, G. E. Alexander, Firing changes in cells of the nucleus medialis dorsalis associated with delayed response behavior. Brain Res 61, 79–91 (1973). 12. Y. Watanabe, S. Funahashi, Thalamic mediodorsal nucleus and work- ing memory. Neurosci Biobehav Rev 36, 134–42 (2012). 13. S. S. Bolkan, J. M. Stujenske, S. Parnaudeau, T. J. Spellman, C. Rauf- fenbart, A. I. Abbas, A. Z. Harris, J. A. Gordon, C. Kellendonk, Tha- lamic projections sustain prefrontal activity during working memory maintenance. Nat Neurosci advance online publication (2017). 14. S. Parnaudeau, P.-K. O’Neill, S. S. Bolkan, R. D. Ward, A. I. Abbas, B. L. Roth, P. D. Balsam, J. A. Gordon, C. Kellendonk, Inhibition of me- diodorsal thalamus disrupts thalamofrontal connectivity and cogni- tion. Neuron 77, 1151–1162 (2013). 15. R. Cools, Dopaminergic control of the striatum for high -level cogni- tion. Curr Opin Neurobiol 21, 402–7 (2011). 16. A. Westbrook, M. J. Frank, R. Cools, A mosaic of cost –benefit control over cortico-striatal circuitry. Trends in Cognitive Sciences 25, 710–721 (2021). 17. M. Wolff, M. M. Halassa, The mediodorsal thalamus in executive con- trol. Neuron 112, 893–908 (2024). 18. X. J. Wang, Synaptic reverberation underlying mnemonic persistent activity. Trends Neurosci 24, 455–63 (2001). 19. C. Constantinidis, E. Procyk, The primate working memory networks. Cogn Affect Behav Neurosci 4, 444–465 (2004). 20. S. Liebe, G. M. Hoerzer, N. K. Logothetis, G. Rainer, Theta coupling between V4 and prefrontal cortex predicts visual short -term memory performance. Nat Neurosci 15, 456–62 (2012). 21. R. F. Salazar, N. M. Dotson, S. L. Bressler, C. M. Gray, Content-specific fronto-parietal synchronization during visual working memory. Sci- ence 338, 1097–1100 (2012). 22. T. B. Christophel, P. C. Klink, B. Spitzer, P. R. Roelfsema, J.-D. Haynes, The Distributed Nature of Working Memory. Trends Cogn Sci 21, 111– 124 (2017). 23. N. M. Dotson, S. J. Hoffman, B. Goodell, C. M. Gray, Feature -Based Visual Short-Term Memory Is Widely Distributed and Hierarchically Organized. Neuron 99, 215-226.e4 (2018). 24. M. F. Panichello, T. J. Buschman, Shared mechanisms underlie the control of working memory and attention. Nature 592, 601–605 (2021). 25. S. Haufe, V. V. Nikulin, K. -R. Müller, G. Nolte, A critical assessment of connectivity measures for EEG data: a simulation study. Neu- roimage 64, 120–133 (2013). 26. M. Vinck, L. Huurdeman, C. A. Bosman, P. Fries, F. P. Battaglia, C. M. A. Pennartz, P. H. Tiesinga, How to detect the Granger-causal flow di- rection in the presence of additive noise? Neuroimage 108, 301 –318 (2015). .CC-BY-NC-ND 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted July 25, 2024. ; https://doi.org/10.1101/2024.07.24.605013doi: bioRxiv preprint Cortico-subcortical dynamics in primate working memory von Nicolai & Siegel 2024, bioRxiv 9 27. A. Heuer, S. Ohl, M. Rolfs, Memory for action: A functional view of selection in visual working memory. Visual Cognition 28, 388 –400 (2020). 28. N. A. Steinmetz, P. Zatka -Haas, M. Carandini, K. D. Harris, Distrib- uted coding of choice, action and engagement across the mouse brain. Nature 576, 266–273 (2019). 29. M. Siegel, T. J. Buschman, E. K. Miller, Cortical information flow dur- ing flexible sensorimotor decisions. Science 348, 1352–1355 (2015). 30. S. N. Jacob, D. Hähnke, A. Nieder, Structuring of Abstract Working Memory Content by Fronto -parietal Synchrony in Primate Cortex. Neuron 99, 588-597.e5 (2018). 31. E. Rezayat, K. Clark, M.-R. A. Dehaqani, B. Noudoost, Dependence of Working Memory on Coordinated Activity Across Brain Areas. Front Syst Neurosci 15, 787316 (2021). 32. E. G. Antzoulatos, E. K. Miller, Synchronous beta rhythms of frontopa- rietal networks support only behaviorally relevant representations. Elife 5, e17822 (2016). 33. A. L. DeNicola, M. -Y. Park, D. A. Crowe, A. W. MacDonald, M. V. Chafee, Differential Roles of Mediodorsal Nucleus of the Thalamus and Prefrontal Cortex in Decision-Making and State Representation in a Cognitive Control Task Measuring Deficits in Schiz ophrenia. J Neu- rosci 40, 1650–1667 (2020). 34. Y. B. Saalmann, M. A. Pinsk, L. Wang, X. Li, S. Kastner, The Pulvinar Regulates Information Transmission Between Cortical Areas Based on Attention Demands. Science 337, 753–756 (2012). 35. I. C. Fiebelkorn, M. A. Pinsk, S. Kastner, The mediodorsal pulvinar coordinates the macaque fronto -parietal network during rhythmic spatial attention. Nat Commun 10, 215 (2019). 36. H. Zhou, R. J. Schafer, R. Desimone, Pulvinar -Cortex Interactions in Vision and Attention. Neuron 89, 209–220 (2016). 37. L. I. Schmitt, R. D. Wimmer, M. Nakajima, M. Happ, S. Mofakham, M. M. Halassa, Thalamic amplification of cortical connectivity sustains attentional control. Nature 545, 219–223 (2017). 38. T. Staudigl, T. Zaehle, J. Voges, S. Hanslmayr, C. Esslinger, H. Hin- richs, F. C. Schmitt, H. -J. Heinze, A. Richardson -Klavehn, Memory signals from the thalamus: Early thalamocortical phase synchroniza- tion entrains gamma oscillations during long -term memory retrieval. Neuropsychologia, doi: 10.1016/j.neuropsychologia.2012.08.023 (2012). 39. B. J. Griffiths, T. Zaehle, S. Repplinger, F. C. Schmitt, J. Voges, S. Hanslmayr, T. Staudigl, Rhythmic interactions between the mediodor- sal thalamus and prefrontal cortex precede human visual perception. Nat Commun 13, 3736 (2022). 40. J. Fuster, The Prefrontal Cortex (Academic Press, Amsterdam ; Boston, ed. 5, 2015). 41. E. K. Miller, J. D. Cohen, An integrative theory of prefrontal cortex function. Annu Rev Neurosci 24, 167–202 (2001). 42. D. Jonikaitis, B. Noudoost, T. Moore, Dissociating the Contributions of Frontal Eye Field Activity to Spatial Working Memory and Motor Preparation. bioRxiv, 2023.06.12.544653 (2023). 43. Z. Bahmani, K. Clark, Y. Merrikhi, A. Mueller, W. Pettine, M. Isabel Vanegas, T. Moore, B. Noudoost, “Prefrontal Contributions to Atten- tion and Working Memory” in Processes of Visuospatial Attention and Working Memory, T. Hodgson, Ed. (Springer International Publishing, Cham, 2019; https://doi.org/10.1007/7854_2018_74) Current Topics in Behavioral Neurosciences, pp. 129–153. 44. R. E. Passingham, H. Lau, Do we understand the prefrontal cortex? Brain Struct Funct 228, 1095–1105 (2023). 45. C. Constantinidis, X. -L. Qi, Representation of Spatial and Feature In- formation in the Monkey Dorsal and Ventral Prefrontal Cortex. Front Integr Neurosci 12, 31 (2018). 46. R. A. Andersen, Encoding of intention and spatial location in the pos- terior parietal cortex. Cereb Cortex 5, 457–469 (1995). 47. J. W. Gnadt, R. A. Andersen, Memory related motor planning activity in posterior parietal cortex of macaque. Exp Brain Res 70, 216 –220 (1988). 48. W. E. Mackey, C. E. Curtis, Distinct contributions by frontal and pari- etal cortices support working memory. Sci Rep 7, 6188 (2017). 49. K. Oyama, Y. Hori, Y. Nagai, N. Miyakawa, K. Mimura, T. Hira- bayashi, K. -I. Inoue, T. Suhara, M. Takada, M. Higuchi, T. Mina- mimoto, Chemogenetic dissection of the primate prefronto - subcortical pathways for working memory and decision -making. Sci Adv 7, eabg4246 (2021). 50. J. M. Fuster, Unit activity in prefrontal cortex during delayed-response performance: neuronal correlates of transient memory. J Neurophysiol 36, 61–78 (1973). 51. M. Wolff, S. D. Vann, The Cognitive Thalamus as a Gateway to Mental Representations. J. Neurosci. 39, 3–14 (2019). 52. B. A. L. Perry, E. Lomi, A. S. Mitchell, Thalamocortical interactions in cognition and disease: The mediodorsal and anterior thalamic nuclei. Neurosci Biobehav Rev 130, 162–177 (2021). 53. A. H. Lara, J. D. Wallis, The Role of Prefrontal Cortex in Working Memory: A Mini Review. Front Syst Neurosci 9, 173 (2015). 54. R. V. Rikhye, R. D. Wimmer, M. M. Halassa, Toward an Integrative Theory of Thalamic Function. Annu. Rev. Neurosci. 41, 163 –183 (2018). 55. J. M. Phillips, N. A. Kambi, M. J. Redinbaugh, S. Mohanta, Y. B. Saal- mann, Disentangling the influences of multiple thalamic nuclei on prefrontal cortex and cognitive control. Neurosci Biobehav Rev 128, 487–510 (2021). 56. S. Funahashi, Thalamic mediodorsal nucleus and its participation in spatial working memory processes: comparison with the prefrontal cortex. Front Syst Neurosci 7, 36 (2013). 57. S. Funahashi, K. Takeda, Y. Watanabe, Neural mechanisms of spatial working memory: contributions of the dorsolateral prefrontal cortex and the thalamic mediodorsal nucleus. Cogn Affect Behav Neurosci 4, 409–20 (2004). 58. Y. Watanabe, K. Takeda, S. Funahashi, Population vector analysis of primate mediodorsal thalamic activity during oculomotor delayed -re- sponse performance. Cereb. Cortex 19, 1313–1321 (2009). 59. I. Tanibuchi, P. S. Goldman -Rakic, Dissociation of spatial -, object -, and sound-coding neurons in the mediodorsal nucleus of the primate thalamus. J Neurophysiol 89, 1067–77 (2003). 60. J. Homman -Ludiye, J. A. Bourne, The medial pulvinar: function, origin and association with neurodevelopmental disorders. J Anat 235, 507–520 (2019). 61. M. M. Halassa, S. Kastner, Thalamic functions in distributed cognitive control. Nat. Neurosci. 20, 1669–1679 (2017). 62. J. Jaramillo, J. F. Mejias, X. -J. Wang, Engagement of Pulvino -cortical Feedforward and Feedback Pathways in Cognitive Computations. Neuron 101, 321-336.e9 (2019). 63. D. N. Scott, A. Mukherjee, M. R. Nassar, M. M. Halassa, Thalamocor- tical architectures for flexible cognition and efficient learning. Trends in Cognitive Sciences, doi: 10.1016/j.tics.2024.05.006 (2024). 64. T. Ott, S. N. Jacob, A. Nieder, Dopamine receptors differentially en- hance rule coding in primate prefrontal cortex neurons. Neuron 84, 1317–1328 (2014). 65. O. Hikosaka, Y. Takikawa, R. Kawagoe, Role of the basal ganglia in the control of purposive saccadic eye movements. Physiol Rev 80, 953– 78 (2000). 66. G. E. Alexander, M. R. DeLong, P. L. Strick, Parallel organization of functionally segregated circuits linking basal ganglia and cortex. Annu Rev Neurosci 9, 357–81 (1986). 67. O. Hikosaka, M. Sakamoto, S. Usui, Functional properties of monkey caudate neurons. I. Activities related to saccadic eye movements. J Neurophysiol 61, 780–98 (1989). 68. J. Vitay, F. H. Hamker, A computational model of Basal Ganglia and its role in memory retrieval in rewarded visual memory tasks. Front Comput Neurosci 4 (2010). 69. C. H. Chatham, M. J. Frank, D. Badre, Corticostriatal output gating during selection from working memory. Neuron 81, 930–942 (2014). 70. O. Hikosaka, H. F. Kim, M. Yasuda, S. Yamamoto, Basal ganglia cir- cuits for reward value -guided behavior. Annu. Rev. Neurosci. 37, 289– 306 (2014). 71. A. Pasupathy, E. K. Miller, Different time courses of learning -related activity in the prefrontal cortex and striatum. Nature 433, 873 –6 (2005). 72. J. M. Fuster, Cognitive Networks (Cognits) Process and Maintain Working Memory. Frontiers in Neural Circuits 15 (2022). 73. W. F. Asaad, E. N. Eskandar, A flexible software tool for temporally - precise behavioral control in Matlab. J Neurosci Methods 174, 245–258 (2008). .CC-BY-NC-ND 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted July 25, 2024. ; https://doi.org/10.1101/2024.07.24.605013doi: bioRxiv preprint Cortico-subcortical dynamics in primate working memory von Nicolai & Siegel 2024, bioRxiv 10 74. G. Paxinos, M. Petrides, X. -F. Huang, A. W. Toga, The Rhesus Monkey Brain in Stereotaxic Coordinates (Elsevier Science, 2008). 75. E. Calabrese, A. Badea, C. L. Coe, G. R. Lubach, Y. Shi, M. A. Styner, G. A. Johnson, A diffusion tensor MRI atlas of the postmortem rhesus macaque brain. Neuroimage 117, 408–416 (2015). 76. K. S. Saleem, N. K. Logothetis, A Combined MRI and Histology Atlas of the Rhesus Monkey Brain in Stereotaxic Coordinates (Academic Press, ed. 2, 2012). 77. S. Olejnik, J. Algina, Generalized eta and omega squared statistics: measures of effect size for some common research designs. Psychol

8, 434–447 (2003). 78. L. Barnett, A. K. Seth, The MVGC multivariate Granger causality toolbox: A new approach to Granger -causal inference. J. Neurosci.

223, 50–68 (2014). 79. Y. Benjamini, Y. Hochberg, Controlling the False Discovery Rate: A Practical and Powerful Approach to Multiple Testing. Journal of the Royal Statistical Society: Series B (Methodological) 57, 289–300 (1995). 80. R. Oostenveld, P. Fries, E. Maris, J. -M. Schoffelen, FieldTrip: Open source software for advanced analysis of MEG, EEG, and invasive elec- trophysiological data. Comput Intell Neurosci 2011, 156869 (2011).

We thank Peter Dicke for technical advise , assistance and helpful discussions. This research was supported by the European Research Council (ERC; https://erc. europa.eu/) StG 335880 (M.S) and CoG 864491 (M.S), Deutsche Forschungsgemeinschaft (DFG; German Research Foundation; https://www.dfg.de/) project SI1332/3 -1 (M.S.) and project 276693517 (SFB 1233) (M.S.) and the Centre for Integrative Neuroscience (DFG, EXC 307) (M.S.). The authors acknowledge support by the state of Baden-Württemberg through bwHPC, by the German Research Foundation (DFG) through grant no INST 39/963 -1 FUGG (bwForCluster NEMO), and by the Open Access Publishing Fund of the University of Tübingen. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript. Author contributions CvN: Conceptualization, Methodology, Investigation, Formal analysis, Visualization, Writing – original draft, Writing – Review & Editing MS: Conceptualization, Methodology, Investigation, Supervision, Resources, Project administration, Funding acquisition, Writing – original draft, Writing – Review & Editing Competing interest statement All the authors declare no competing interests. Data availability statement Data and code to reproduce all the reported results are available from the lead author upon reasonable request.

Subjects and behavioral task All procedures involving non -human primates were carried out in agree- ment with national and international law, rules and regulations on the hu- mane care and treatment of research animals and were approved by local authorities. We trained 2 macaque monkeys ( Macaca mulatta ) in two tasks that probed color and spatial working memory, respectively (Figure 1A). Im- portantly, both tasks had identical, fixation baseline, cue presentation and memory delay intervals. Tasks were presented in a block -wise manner (at least 64 correct trials per block), i.e., the current task rule was implicit. An- imals initiated a trial by acquiring and holding fixation of a central white spot (diameter: 0.1 º visual angle) presented on a black background for 1 s. Following fixation, a blue or red cue stimulus (diameter: 1 º) was presented for 0.5 s at 45, 135, 225, or 315 º circular angle and 6 º eccentricity. Depend- ing on the current task, animals had to memorize either the color or the location (color and spatial task, respectively) of the stimulus over a delay of 1.25 s during which they had to maintain central fixation. Following the delay and the offset of the fixation point, in the color task, a pair of periph- eral targets colored blue and red was presented at two randomly chosen, opposing (180 º circular distance), possible prior stimulus locations. In the spatial task, fixation point offset was followed by a blank screen. Animals had to make a saccade either to the target that matched the previously cued color (color task) or to t he previously cued location (spatial task). Thus, in the spatial task, animals could prepare the motor response during the delay whereas in the color task, they could not. In both tasks, monkeys had to hold their gaze at the correc t target location for at least 50 ms. Then, for correct trials, a green central square (edge length: 6 º) was shown and ani- mals received a liquid reward. An incorrect response was followed by a 100 ms flash of a grey square of the same size. A trial was ab orted if monkeys did not keep their gaze within a 1.5 º radius around the fixation spot before the end of the delay period. Both correct and error trials were followed by an inter-trial period of at least 2 s. Behavioral task control was accomplished with the Monkeylogic toolbox73 for MATLAB (The Mathworks, Natick, MA, USA). Stimuli were presented using CRT monitors with a vertical refresh rate of 100 Hz. Mon- keys sat in a primate chair (Crist Instruments) at a viewing distance of 42 cm from the screen and their heads were immobiliz ed via titanium head- posts throughout the course of behavioral training and recording sessions. Gaze control was accomplished with infrared eyetracking systems (View- Point EyeTracker, Arrington Research Inc., Scottsdale, Arizona, USA; and Eyelink 1000, SR Research, London, Ontario, Canada) at a sampling rate of 220 and 1000 Hz; we used the Eyelink system for neural recording sessions. We measured the luminance of monitor background, fixation spot, cue and target stimulus, and feedback square colors and equated their brightness for behavioral training and recording sessions. Neural recordings Electrodes were lowered into and removed from the brain anew on each recording day. We used custom-designed, 3D-printed grids for placement of electrode drives and precise, reliable targeting of cortical and subcortical structures (Fig. 1B and Figure S1). We recorded broadband neural activity using linear, multi-contact Plexon (Plexon Inc.; Dallas, Texas, USA) or Neu- ronexus (NeuroNexus Technologies, Inc.; Ann Arbor, Michigan, USA) probes at a sampling rate of 24414 Hz with a TDT (Tucker -Davis Technolo- gies; Alachua, Florida, USA) multichannel recording system. We used Plexon U - and V -probes with 24 contacts and vertical interelectrode dis- tances of either 0.1, 0.15, 0.25, or 0.5 mm (single electrode configuration) or 0.15 mm vertical and 0.05 mm horizontal int er-electrode distance (stere- otrode configuration); Plexon S -probes with 32 contacts and vertical inter - electrode distances of 0.1 or 0.25 mm; and Neuronexus probes with 32 con- tacts and vertical inter-electrode distances of 0.1 mm. Area targeting For chamber implantation and recording target planning, we resliced T1 - weighted structural MR images of both animals such that the interaural zero point represented the origin of the recording space 74 (Figure S1). We used a publicly available, standardized rhesus macaque brain atlas 75 to per- form successive, combined linear and non -linear coregistriations between this atlas, regions of interest (ROI) label maps derived from it and the indi- vidual MR images of the animals. This provided region label volumes that we combined with MR images for target planning, in addition to consider- ing other standardized rhesus macaque brain atlases 74,76 and various litera- ture sources on the anatomical connectivity between our target regions. We custom-designed titanium chambers based on the individual animal’s skull surface above the desired target brain areas using CAD -software, with the construction space being equal to the interaural space used for target plan- ning. Headpost and chamber implantations were performed in separate ses- sions under general anesthesia and aseptic conditions. For chamber im- plantations, a stereotaxic arm was used to position the chamber at its target location in interaural space. Following implantation, we measured of the exact position of the chamber in 3D -space and realigned the final posit ion of the chamber with the individual monkey’s MRI images to allow for pre- cise target planning using 3D -printed grids placed inside the chamber. Signal preprocessing From the broadband recordings, we derived spiking signals (multi -unit ac- tivity; MUA) by 500 Hz high -pass filtering. For analyses of neural infor- mation and directed interactions of spiking activity, we made use of the con- tinuous time course (envelope) of MUA, so-called analog multi-unit activity (AMUA), which provides a robust estimate of neural population activity. We computed AMUA by rectification, 250 Hz low -pass filtering and loga- rithmic scaling of the high-pass filtered data. To account for slow drifts and .CC-BY-NC-ND 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted July 25, 2024. ; https://doi.org/10.1101/2024.07.24.605013doi: bioRxiv preprint Cortico-subcortical dynamics in primate working memory von Nicolai & Siegel 2024, bioRxiv 11 offsets of spiking activity over time we subtracted a median filtered AMUA signal (10 trials) prior to information and Granger causality analyses. Spiking information coding We computed AMUA information about variables of interest by means of two separate n -way ANOVAs across trials for each recording site using w2 as an unbiased measure of variance of spiking activity explained 77. The first ANOVA served to compute information about cue color (blue vs. red), cue location (4 locations at 45, 135, 225, 315 º circular angle) and their interac- tion for correct trials. It included the factors cue color, cue location, re- sponse location and a running count of the respective task. We conditioned it on the factors task (color vs. spatial task) and block. The second ANOVA served to compute information about the color of the chosen target. It in- cluded the factors task, cue color, cue position, chosen target color and re- sponse location; we conditioned it on the factor block. For each ANOVA, block-wise estimates of information where averaged before further anal- yses. For information time courses, we used a sliding window approach (50 ms window, 25 ms steps). For fixed temporal windows, we used non -over- lapping 250 ms windows. These windows were used for further statistical analyses. Granger causality analysis We quantified broadband Granger causality between pairs of AMUA sig- nals by integrating the frequency -domain estimates across the spectrum 78. Frequency-domain estimates were based on Fourier -transforms computed for frequencies between 0 and 256 Hz in 4 Hz steps for 0.25 second segments of AMUA after multiplication with 3 discrete prolate spheroidal sequence (DPSS) tapers. To avoid erroneous i nferences of statistical causality due to volume conduction effects, we computed Granger causality on trial -shuf- fled and on time-reversed AMUA signals25. We subtracted the trial-shuffled Granger causality from all raw Granger causality estimates. Statistics on neural information We used one-sided t-statistics to evaluate the significance of explained var- iance against the null hypothesis of zero information in separate time win- dows spanning 500 or 250 ms each. For the fixation, stimulus presentation, and early delay periods, we ave raged omega square values of two adjacent 250 ms segments to comply with the 250 ms window of the late delay period. For time-window differences, we subtracted omega square values of the cor- responding condition-specific windows and, for normalization, div ided the difference by their sum. Similarly, for condition differences, we subtracted omega square values of condition-specific windows and normalized by their sum. We used two -sided t-statistics to evaluate significance of window dif- ferences from 0 of individual c onditions, condition differences and differ- ences between condition -specific window-differences. Alpha levels of all t - tests were set to p < 0.05. All p-values were FDR-corrected according to the

proposed by Benjamini and Hochberg 79 to account for multiple hy- potheses testing. Statistics on Granger causality We applied several criteria to Granger causality to be taken as reflecting true interaction between a given pair of areas. First, Granger causality from ei- ther direction A to B or B to A had to be significantly non-zero (one-sided t- test across electrode p airs, p < 0.05 FDR corrected) after subtracting a sur- rogate estimate obtained by shuffling trials before Granger causality com- putation. Second, there had to be an asymmetry of directional coupling when subtracting Granger causality from A to B and B to A, i.e., their dif- ference had to be significantly different from zero as assessed by two -tailed t-tests across electrode pairs (p < 0.05 FDR corrected). Third, when con- trasting results from real (i.e., time -forward) and time -reversed data, the sign of the int eraction had to flip. We used this conservative approach to preclude spurious interactions due to, e.g., volume conduction effects 25. If all three criteria were fulfilled, a given functional connection between areas A and B was considered as valid in the sense that it statistically exhibited true interaction. This assessment of Granger causality also lay the grounds for all further analyses: for window differences, we subtracted Granger causality of a given pair of temporal windows that were the same as those used for spiking in- formation analysis (fixation, sti mulus presentation, early and late delay pe- riods), and for condition differences, we subtracted Granger causality from a given pair of conditions for each of those temporal windows. Window and condition differences were assessed for significance by two -tailed t-tests in addition to fulfilling the above criteria. As for neural information analyses, alpha levels for Granger causality statistics were set to 0.05 and p -values were FDR-corrected to account for multiple comparison. All data analyses were performed in MATLAB using custom scripts and the FieldTrip80 toolbox. .CC-BY-NC-ND 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted July 25, 2024. ; https://doi.org/10.1101/2024.07.24.605013doi: bioRxiv preprint Cortico-subcortical dynamics in primate working memory von Nicolai & Siegel 2024, bioRxiv 12 A46 IA: +40,08 mm A9/46 IA: +36,75 mm Striatum IA: +27,75 mm FEF IA: +24,60 mm MD IA: +10,20 mm Pulvinar IA: +05,70 mm LIP IA: +03,00 mm Left LeftRight Right C PS AS CS IPS A46 A9/46 FEF Striatum LIP Pulvinar MD B A Figure S1: Recording approach. We used custom -designed titanium chambers spanning large parts of the anterior -posterior (AP) extent of one hemisphere, custom-designed microdrives and grids to record broadband neuronal activity from multiple corti- cal and subcortical structures simultaneo usly. (A) 3D-rendering of the titanium chamber model and of the headpost -base on the skull and brain of one of the monkeys. ( B) Lateral view of a monkey brain from the standard monkey brain atlas of Paxinos and Watson illustrating the position of frontal sections used in (C) and corresponding most closely to the AP positions of the probes used in a representative recording session from one monkey. Colors correspond to those used for ROI illustrations in (C). ( C) We mapped electrode trajectories (blue lines) onto the individual animal’s structural MRI including colored regions of interest (ROI). Frontal MRI sections from one monkey and most closely matching frontal sections from the standard monkey brain atlas of Pax- inos and Watson. AP coordinates are given with reference to interaural (IA) zero. Colored ROIs indicate target areas. We rota ted MRI images by 50 degrees to account for the chamber’s inclination and to ease comparison with the standard atlas sections. Red lines indicate trajectories blocked by the curved parts of the chamber. Supplemental Information .CC-BY-NC-ND 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted July 25, 2024. ; https://doi.org/10.1101/2024.07.24.605013doi: bioRxiv preprint