Bi-directional command of cognitive control by distinct prefrontal cortical output neurons to thalamus and striatum


 The medial prefrontal cortex (mPFC) steers goal-directed actions and withholds inappropriate behavior. Dorsal and ventral mPFC (dmPFC/vmPFC) circuits have distinct roles in cognitive control, but underlying mechanisms are poorly understood. In this study, we provide anatomical, behavioral, and neurophysiological evidence for distinct roles of four distinct prefrontal projection populations in behavior. We used neuroanatomical tracing techniques, chemogenetics and fiber photometry in freely behaving rats, and in vitro electrophysiology to characterize dmPFC and vmPFC outputs to distinct thalamic and striatal subdomains and show that they have dissociable roles in cognitive control. We identify four spatially segregated projection neuron populations in the mPFC. Chemogenetic silencing shows that dmPFC and vmPFC projections to lateral and medial mediodorsal thalamus subregions oppositely regulate cognitive control. In addition, superficial and deep layer dmPFC neurons projecting to striatum and thalamus divergently regulate cognitive control. Using fiber photometry, we show that these projections distinctly encode behavior. Finally, we show that postsynaptic striatal and thalamic neurons differentially process synaptic inputs from dmPFC and vmPFC, highlighting mechanisms that potentially amplify distinct pathways underlying cognitive control of behavior. Collectively, we show that mPFC output circuits targeting anatomically and functionally distinct striatal and thalamic subregions encode bidirectional command of cognitive control.


Introduction
Cognitive control involves the ability to suppress undesirable actions and remain attentive to relevant stimuli. The medial prefrontal cortex (mPFC) is highly involved in these processes, as shown in lesion, pharmacological, optogenetic and chemogenetic experiments (1-4). Distinct neuronal activation patterns across mPFC subregions, cell types, and behavioral subdomains often underlie cognitive control (3,5,6). However, there is substantial heterogeneity in timing, location and origin of brain activity associated with behavior (7). For instance, the dorsomedial PFC (dmPFC; defined here as dorsal prelimbic and anterior cingulate cortex) has been associated with longer windows of activity than the ventral mPFC (vmPFC; as the infralimbic and ventral prelimbic cortex) during cognitive control tasks (4). Cortical cells, including dmPFC and vmPFC neurons, can further be classified based on their projection target and transcriptomic profile (8)(9)(10). Functional studies have established a role for projection-specific mPFC populations in goal-directed behavior (11,12). This suggests that studying the function of projection-specific neurons may lead to better understanding of the role of specific neural populations and circuits in cognitive control.
Several downstream targets of the mPFC are associated with cognitive control. The mediodorsal thalamus (MD) contains medial and lateral subregions (MDM/MDL), which are reciprocally connected to the and dmPFC, respectively. These circuits maintain activity during cognitive control tasks, and are thought to guide correct behavioral output by maintaining a representation of a task rule (13)(14)(15)(16)(17). Likewise, the dorsomedial and ventromedial striatum (DMS/VMS) have both been linked to inhibitory control and attention (18)(19)(20) and receive input from the dmPFC and vmPFC, respectively. Moreover, specific mPFC-DMS projections are linked to development of cognitive control and show ramping during preparatory attention (12,21), whereas mPFC-VMS projections are associated with anticipation and reward processing during cognitive control tasks (11,22,23). This indicates that prefrontal populations can be separated based on projection target and that they are distinctly involved in behavior. However, the exact role and timing of activity of these projections in cognitive control is unknown.
We provide evidence for the existence of four distinct prefrontal projection populations. Neuroanatomical tracing using retrograde virus and retrobeads was used to identify corticothalamic and corticostriatal projection neurons. Next, we measured cognitive control performance in rats using a self-paced 5-choice serial reaction time task (SP-5-CSRTT; 23), which specifically allows investigation of attention and inhibitory control. We then tested the causal role of each projection in cognitive control using chemogenetics, which suggested distinct roles in inhibitory control depending on projection target and population location in the mPFC. In a separate group of rats, we investigated temporal dynamics of brain activity during task performance, and found that projection populations had distinct activation patterns. Finally, we report distinct postsynaptic responses to prefrontal stimulation in striatal and thalamic neurons. Based on literature, we expected differential involvement of two corticostriatal and two corticothalamic projection populations.
Collectively, we here demonstrate a distinct role for each projection population in cognitive control.

Distinct distribution of prefrontal projection neurons
Pyramidal neurons projecting to the MD and striatum are located across the dmPFC and vmPFC (8,9). However, whether these neurons have distinct targets is unclear.
Therefore, we first expressed eYFP in the dmPFC or vmPFC and observed axonal eYFP expression in MD and striatum subdomains (Figure S1A-D). We next infused retrobeads in the MD and striatum subdomains with a high degree of eYFP-positive axons. Quantification of labeled mPFC somata across three anterior-posteriorlocations revealed a gradient of retrobead-positive neurons along the dorso-ventral axis, as well as a gradient across cortical layers. We found that 90±2.25% of MDLprojecting neurons were located in dmPFC areas, with the remaining cells located in the vmPFC (Figure 1A), and 81±3.07% of all MDM-projecting neurons were found in the vmPFC with the remaining neurons situated in the dmPFC ( Figure 1B). MDprojecting mPFC neurons were primarily found in deep layers, while striatumprojecting mPFC neurons were located in layers 2/3 and 5. Of all DMS-projecting neurons 82±1.6% were located in the dmPFC with the remaining part being in the vmPFC (Figure 1C), and of all VMS-projection neurons 75±1.98% were located in the vmPFC with the remaining neurons located in the dmPFC ( Figure 1D) Cortical neurons can project to multiple target regions through axon collaterals (24,25). Moreover, while projection neuron location was biased to layers, the layers did not exclusively contain neurons projecting to a single target area (Figure 1A-D). To test whether single neurons project to both the MD and striatum, we separately injected CAV2-cre and retro-FLPo in the MD and striatum combined with credependent eYFP expression and FLPo-dependent mCherry expression in the mPFC ( Figure 1E-F). Only a minority of dmPFC neurons (0.8%) and vmPFC neurons (0.6%) were positive for both mCherry and eYFP ( Figure 1E-F), suggesting that the vast majority of neurons specifically project to either MD or striatum. Additionally, no eYFP-or mCherry-positive neurons were positive for GAD67, excluding the possibility that long-range interneurons significantly contribute to our projection groups (26). Together, these data suggest that the majority of MD-and striatumprojecting mPFC neurons form distinct pyramidal neuron populations.

Bi-directional control of impulsivity by mPFC projection neurons
The mPFC, MD and striatum regulate cognitive control of behavior (4,(16)(17)(18), but the role of specific mPFC projections to MD and striatal subdomains is incompletely understood. We expressed the inhibitory DREADD-receptor hM4D(Gi) in each projection population to test whether they are causally involved in cognitive control (Figure 2A). Clozapine-N-Oxide (CNO) elicited membrane potential hyperpolarization, increased rheobase, and decreased spike frequency under current step injections in acute mPFC brain slices of hM4D(Gi)-expressing animals ( Figure   S2). We then expressed hM4D(Gi) in each projection population and trained animals in the SP-5-CSRTT (20). Animals could earn food rewards by withholding responses during a delay period until a visual cue appeared randomly in one of five cue holes ( Figure 2B). Premature responses made before cue onset were used as a measure for inhibitory control, whereas the ratio of correct and incorrect responses, as well as omissions, were used as a measure for attention ( Figure 2B-C). After reaching stable baseline task performance, animals started testing sessions. On test days, we either varied cue durations or delays between trial start and cue presentation, to increase cognitive load and avoid overtraining, and to more specifically test attention and inhibitory control, respectively (20,27). Animals were injected with each CNO dose in a randomized order, and performed 402±10 trials (mean ± SEM) per 2.5-hour session in these conditions ( Figure 2D). Premature responding consistently increased with longer delay duration ( Figure 2D-F, Table S1; F [2,20] = 51.13, p<0.001), while trials with shorter cue duration decreased accuracy and increased omissions ( Figure S3A, Table S1).
CNO-mediated inhibition of MDL-projecting mPFC neurons decreased premature responding, especially in trials with long delays (Figure 2G; values relative to saline condition. MDL: group x dose x delay: F [8,128] = 9.31, p<0.001). Additionally, we observed a delay-independent increase in omissions ( Figure 2H,   Table S1, group x dose x delay: F [8,128] = 9.31, p<0.001). CNO-mediated inhibition of MD-projecting neurons during variable cue duration sessions increased omissions, independent of cue duration, but did not affect accuracy or any other measured behavioral parameter ( Figure S3M-P, Table S1). To exclude for confounding effects of motivation and motor control we tested for effects of CNO on 5-CSRTT parameters such as response latency or number of started trials. These were unaffected in all sessions (Table S2). Finally, no effect of CNO was observed in the eYFP control group (Figure 2G-H Inhibition of DMS-projecting dmPFC neurons in the variable delay protocol increased premature responding, but did not affect omissions ( Figure 2I, Table S3; DMS: group x dose x delay: F [8,124]=2.72, p<0.01), while inhibition of vmPFC-VMS projections did not affect premature responses, omissions or any other behavioral parameter in the task ( Figure 2J, Table S3-4). During variable cue duration sessions, CNO had no effect on accuracy ( Figure S3, Table S3-4), and additional behavioral parameters such as premature responses, response latencies, and number of started trials were also unaffected (Table S1-2), suggesting that inhibitory control, motivation and task engagement of animals were unaltered.
Altered premature responding can reflect changes in temporal strategies or perception (28). However, we found no effect of CNO on the temporal distribution of premature response latencies in long delay trials ( Figure

Activity of mPFC projection neurons encodes behavioral trial outcome
To test whether activity patterns of specific mPFC projection neuron populations encode trial outcome, z-scored mean fluorescence traces were compared to a randomly resampled population using a bootstrap approach to determine periods of elevated activation during long-delay trials ( We then asked whether these projection population-specific activation profiles contain predictive information on behavioral trial outcomes. We compared activation dynamics within each population during trials with different behavioral outcomes  A) Average GCaMP6 fluorescence of each mPFC projection neuron population in individual rats during correct response, omission and premature response trials (δF/F is z-scored to trial baseline). Plot is capped at -2 and 2 z-scores. Baseline marked in top left. B) Group activity during behavioral trials. C) Upper: windows of significantly increased activity during delay and around cue and response. Bootstrap parameters: 5000 iterations, α=0.001. Lower: windows with significant difference in activity between indicated projection populations. Bars represent significant permutation test results. (see Supplemental Methods for permutation test procedures). Double colored bars represent populations that were compared. Iterations: 5000, α<0.01 D) Average activity during different trial outcomes. E) Statistical evaluation of activity in (E). Upper: time windows with significant elevated activity during delay, and around cue and response. Bootstrap parameters same as in (C), see Supplemental Methods section for detailed procedure. Lower: windows with significant difference between activity during different behavioral trial outcomes. Bars represent significant permutation test results. Permutation test parameters same as in (C). Singleton significant data frames were discarded. Double colored bars as in (C). MDL n=8, MDM n=4, DMS n=7, VMS n=6.

Distinct functional properties of mPFC output pathways
Our neuroanatomical tracing, behavioral and fiber photometry results indicate that the mPFC has differential output pathways to the subregions of the MD and striatum in support of cognitive control. However, whether this is also reflected in the These data show that the passive and active electrophysiological or synaptic input properties of mPFC-targeted MD or striatum neurons, respectively, differ between their subregions. Together, this indicates that the dorsal or ventral mPFC pathways to the subregions of the MD or striatum are part of differential information streams in support of cognitive control.

Discussion
In this study, we provide anatomical, behavioral, and neurophysiological evidence for distinct roles of four distinct prefrontal projection populations in behavior. Projection neuron populations are spatially segregated in the mPFC, and inhibition of these projection neurons disrupts both inhibitory control and attention. We show for the first time that mPFC projection neurons targeting distinct MD subregions have opposite roles in inhibitory control. We also show that thalamus-and striatum-projecting mPFC neurons have distinct roles in inhibitory control. Moreover, projection neuron populations showed distinct temporal dynamics that predicted behavioral trial outcome. Finally, we show that postsynaptic neurons in target regions respond to prefrontal input in a distinct manner. Taken together, we present four distinct lines of experimental evidence for distinct roles for mPFC projection neuron populations in cognitive control.
Prefrontal neurons are known to project to the striatum or thalamus (9,11,23) in a dorsal-to-ventral and layer-based distribution (8,12,29). Our data further specifies prefrontal afferents into populations of excitatory neurons that preferentially target subdomains of the thalamus and striatum. Additionally, although inhibitory frontostriatal projections have been reported in mice (26), we found no GAD-67 expression in projection populations. Possibly, inhibitory projections target more posterior regions of the caudate-putamen (30). While projection neurons were located mostly in specific prefrontal layers and subregions, we also show that they can be situated outside the regions we described. Projection-specific transcriptomic analysis of mPFC neurons (10,23,29) may resolve this issue. Furthermore, while we find little evidence for axon collaterals to both the MD and striatum, projection populations could be interconnected within the mPFC. This could potentially result in projection-unspecific effects on behavior and could be resolved by doing manipulations or recordings at axon terminals. Finally, we used both retro-AAV and CAV vectors, which can have distinct viral tropisms (31)(32)(33). While, to our knowledge, no such effects have been reported for the populations we investigated, a method to circumvent this is to use an enhanced CAV vector (32). We perturbed physiological activity of subpopulations within these larger mPFC regions. Our results suggest that inhibiting small and specific populations of projection neuron disentangles specific aspects of cognitive control such as inhibitory control and attention, which are collectively affected when manipulating entire mPFC subregions in a non-specific manner. Silencing multiple projection populations using multiple-wavelength optogenetics could provide further insight into the role of each projection and redundancy of information sent through projections.
Activation of mPFC neurons during the delay period of cue detection paradigms has frequently been reported, but timing and amplitude of activity varies between trial outcomes, target area, and task parameters (3,5,21,22,34). In all projection populations, we observed that activity followed delay duration, indicating that each population was activated in support of cognitive control over behavior. We find that dorsal mPFC projection neurons were recruited faster than ventral mPFC projection neurons, and that dmPFC projection neurons were active for a larger proportion of the delay period. This is in line with previous findings of different activity kinetics between non-identified vmPFC and dmPFC units, and suggest a more proactive role of the dorsal mPFC and a more reactive role of the ventral mPFC (4,5). We also report differences in population activity between projection populations, and between activity levels during delay periods leading up to different trial outcomes. Hence, fiber photometry recordings indicate that there is population activity that significantly deviates from baseline, and chemogenetic inhibition showed that disruption of such physiological levels of activity caused deficits in inhibitory control and attention.
Optogenetic identification combined with optogenetic manipulation has been used to further characterize the role of projection populations in behavior (35), and would be a suitable technique to combine activity recordings with targeted inhibition.
Connections between the mPFC and MD are organized in recurrent loops, through which the MD can amplify local connectivity in the mPFC (13,24). Both mPFC-MD projections and MD neurons have been associated with behavioral flexibility and working memory, and drive correct behavioral output in different paradigms by maintaining a representation of a task rule (13,14,36). Additionally, mPFC input to MDL neurons are required for proper rule encoding (37). In this study, we show an opposite effect of manipulation of mPFC neurons projecting to MDL-or MDM when delay duration was unpredictable. Additionally, inhibiting mPFC to MDL neurons increased omissions. Possibly, inhibition of this projection affects rule encoding, which thereby reduces response readiness and manifests behaviorally with both a reduction of premature responses and increased omissions. Hence, activity in the MDL-projecting population could drive responsive action, while MDM-projecting mPFC neurons could relay a signal to withhold a response until a sensory event occurs. Alternatively, mPFC-MD projections could be responsible for maintenance of rule representation, rather than encoding. Perturbation of these projections specifically during rule encoding phases of the task could further unravel their exact role.
We find that prefrontal inputs elicit a facilitating response in both MDL and MDM neurons, which may be a potential mechanism through which recurrent activity in corticothalamic circuits is maintained during a delay period, and through which these projections regulate rule encoding or maintenance (37,38). We also find that MDLprojecting neurons are recruited earlier during the delay than MDM-projecting neurons, supporting earlier evidence that MD subregions likely have distinct roles and are part of distinct circuits, and that dmPFC activity precedes vmPFC activity in cognitive control paradigms (4,5). Our findings that MDL and MDM differ in basic electrophysiological features further support distinct roles in cognitive control.
Additionally, we show differences between population activity before premature and correct responses in the MDL, and between omissions and correct responses in the MDM. This difference suggests that different levels of activity in this circuit can underlie distinct types of behavior.
The mPFC has been shown to exert top-down control over the DMS (21), but chemogenetic and optogenetic inhibition of the dmPFC did not affect premature responding (2,4). However, these manipulations were not targeted to a specific population, and covered both MDL-and DMS-projecting populations, which have an opposing effect. Hence, our data support a role of dmPFC-DMS projections in inhibitory control. A potential mechanism could involve striatal dopamine. It was shown that optogenetic enhancement of mPFC excitability diminishes the striatal response to dopamine and suppresses reward seeking behavior (39), while infusions of both D1 and D2-like receptor agonists specifically in the DMS increase premature responding in the 5-CSRTT (40). Thereby, dopamine in the DMS may increase reward seeking and impulsivity, which can be controlled by mPFC inputs in a topdown fashion. Our results show an increase in population activity in DMS-projecting neurons during the delay period. Changes in firing rate have been reported in both mPFC and DMS during the delay before a response (3,22,34,41), as well as during cue presentation (12). Premature responses have been associated with reduced amplitude of neuronal activity in the dmPFC (5). Our data show that this population is active during the delay period and during the cue presentation before a correct response. While we did not see a reduced amplitude in the delay period before premature responses, we do see a shorter active window compared to correct responses. We also found a mixed synaptic input response in the DMS, which could be due to projection neurons differentially innervating D1-and D2-receptor expressing MSNs (42,43), or by specific topological innervation patterns seen in corticostriatal projection neurons (44). Previous work suggests that this projection may also be important for accuracy of responding (45). In our study, we targeted a more specific neuronal population, which could account for distinct behavioral effects.
We found no behavioral effect of inhibition of VMS-projection neurons. Previous functional disconnection studies targeting the mPFC and NAc shell, but not core, showed increased premature responding (46). Additionally, mPFC and contralateral NAc lateral core lesions increased premature responses after an error in the 5-CSRTT, suggesting a role of this pathway in adaptive control (47). However, we did not target a specific NAc subregion. Our neuroanatomical data shows axon terminals in the medial ventral caudate-putamen and NAc. Additionally, the NAc core receives top-down glutamatergic inputs from several other brain regions, such as the ventral hippocampus or insula (48,49). It has been shown that fast-spiking interneurons in the NAc core have different levels of activity leading up to correct and premature responses in the 5-CSRTT, indicating that this area is active during the task (19). We also find that NAc neurons show a depressing response to vmPFC input, and that VMS-projecting neurons do show delay-dependent kinetics of population activity, even though signal amplitude was not significantly increased from baseline. Hence, the vmPFC does project to the NAc, but it likely does not drive the behavior we studied. While activity parameters at times do not significantly differ from other projections, it is likely that this activity is not synchronized enough to yield significantly elevated activity windows during the delay. Whether sparse mPFC-NAc activity are involved in cognitive control remains to be tested.

Lead contacts and material availability
Further information and requests for resources and reagents should be directed to and will be fulfilled by the Lead Contact, Huibert D. Mansvelder (h.d.mansvelder@vu.nl).

Animals
A total of 172 rats (Charles River, Den Bosch, The Netherlands; Janvier, Le Genest-Saint-Isle, France, control groups were vendor matched) were used across all experiments (overview in Table 1). For neuroanatomical tracing experiments, and ex-vivo electrophysiological validation, 29 male Long Evans rats (8 weeks old) were housed in pairs with food and water available ad libitum. For chemogenetic experiments, 84 male Long Evans rats (8 weeks old) were initially housed in pairs with food and water available ad libitum one to two weeks before surgeries, after which they were separated for training and testing in CombiCages (20). Rats were housed under a 12h light/dark cycle (lights off at 12PM). For fiber photometry experiments, 29 male Long Evans rats were housed in pairs until surgery. After surgery for these experiments, animals were housed individually in CombiCages until finishing the testing protocol. For electrophysiology experiments, 26 male Long Evans rats were used, which underwent surgery at 8 weeks of age, and were then housed in pairs until the start of experiment. All experimental procedures were in accordance with European and Dutch law and approved by the central committee animal experiments and local animal ethical care committee of the VU University and VU University Medical Center (Amsterdam, Netherlands).

Surgery
For all experiments, rats were anaesthetized with 2.5% isoflurane gas mixed with air and oxygen and delivered with a flow rate of 1.2L/min. The rats were placed on a heating pad in a stereotaxic frame (Kopf, USA) and their skin of the scalp was retracted to expose the skull. A craniotomy was made at the location stated below and the virus/Retrobead infusion was done using a Nanoject II (Drummond Scientific, USA) via a glass micropipette. After the infusion, we held the pipette in place for 8 min to allow for diffusion, retracted it for 100 µm, waited 1 min, repeated this procedure, and then finally slowly retracted the pipette to minimize virus/Retrobead leakage. For the fiber photometry experiments, we implanted the fiber optic cannulas (pre-assembled from Doric lenses, NA 0.51, core diameter 400μm, fiber length 4.5mm for dmPFC targets, 5.5mm for vmPFC targets) directly after, at the same location as the virus infusions. Additionally, we attached stainless steel screws (0.7mm diameter, Jeveka) to the skull to improve headcap stability. Fibers were fixed to the skull using UV-cured dental cement (RelyX, 3M). To minimize suffering from surgeries, as an analgesic, Rimadyl (carprofen, 5 mg/kg), was administered a day before the surgery, on the day of the surgery and two days afterwards. Also, the analgesic temgesic (buprenorphine, 0.05 mg/kg) was administered once, 30-60 min before the surgery. During surgeries, lidocaine (xylocaine) was used as local anesthetic. Immediately after the surgery, before waking up, animals received 1 ml 0.9% saline.

Histology and Immunofluorescence
Rats were anesthetized with Euthasol (AST Farma, The Netherlands) and perfused transcardially, first with 200 ml 0.9% saline followed by 300 ml of 4% paraformaldehyde. Brains were removed and kept in the same fixative for 24 h and were then transferred to PBS with 0.02% NaN3. Coronal sections of 50 µm were cut on a vibratome. Sections from the Retrobead experiments were directly mounted on glass slides using 2% Mowiol. Immunofluorescent stainings were performed for either NeuN, mCherry, GAD-67, and GFP. We used the following antibodies: mouse anti-NeuN (Abcam, 1:1000) with Alexa Fluor 647 donkey anti-mouse (Thermo Fisher Scientific, 1:400), rabbit anti-RFP (Rockland, 1:1000) with Alexa Fluor 546 donkey anti-rabbit (Thermo Fisher Scientific, 1:400), mouse anti-GAD67 (Millipore, 1:1000) with Alexa Fluor 647 donkey anti-mouse (Thermo Fisher Scientific, 1:400), and rabbit anti-GFP (Abcam, 1:1000) with Alexa Fluor 488 donkey anti-rabbit (Thermo Fisher Scientific, 1:400). The sections were washed and permeabilized in PBS with 0.25% Triton X before being incubated for 3 h with blocking solution containing PBS, 0.3% Triton X and 5% normal goat serum. Next, sections were incubated overnight with primary antibodies in blocking solution at 4° C. The following day, the sections were rinsed with PBS and incubated with secondary antibodies in blocking solution for 2 h at room temperature. Images were acquired with a Nikon Eclipse Ti confocal microscope.

SP-5-CSRTT task
Elaborate descriptions of the self-paced (SP-5)-CSRTT have been described previously (20). Briefly, we constructed CombiCages by connecting a macrolon home-cage to an operant chamber (Med Associates Inc., St. Albans, VT, USA) using a custom-made polymer tube with a diameter of 10 cm. Operant chambers were equipped with five cue holes containing LED stimulus lights and infrared beam detectors on one side. A food magazine, a red magazine light and a yellow house light were positioned on the opposite wall.
We placed the rats in the CombiCages (20) two days before the training in the task started. During training, animals earned their food in the form of precision pellets in the task (Dustless Precision Pellets, grain-based, F0165, 45mg, Bio-Serve, USA). To maintain the rats' weight to an 85-90% food restriction regime, we provided additional standard food chow.
Acquisition of SP-5-CSRTT performance was established by different training phases. First, animals learned to associate pellet delivery with reward. In this phase, for 50 trials a pellet was delivered after a variable delay. Reward was signaled by the magazine light, and a magazine response started the next trial. In the subsequent phase, rats needed to nose poke in one of five illuminated cue holes to earn reward for 50 trials. Next, only one of the 5 cue holes was illuminated and responses into this hole after a delay of 5 s led to reward delivery. During this phase, incorrect or premature nose pokes were not punished. Animals needed to complete 100 trials in this stage.
In the final phase, the animals needed to respond to the cue after a fixed delay of 5 s. The cue hole was lit for a specific cue duration which was initially 16 s and was reduced to 1 s in five steps. The rats had to nose poke during the cue within a 2 s limited hold period after cue presentation. A lack of response was considered an omission and resulted in a timeout period of 5 s. Premature responses, nose pokes during the delay, or incorrect responses were also punished with a 5 s timeout period. Correct responses were always rewarded with a pellet.
After a correct response, animals could start the next trial 5 s after reward collection of the pellet. Importantly, animals could only initiate during the first 2.5 h of the dark cycle (20). In this final phase, the performance criterion to reach a following stage with shorter SD was a minimum of 50 started trials, accuracy (ratio of correct and incorrect responses, see below) > 80% and either omissions < 20% or correct trials > 200 in the current stage. The program monitored these parameters online using a sliding window of 20 trials. If rats passed performance criterion, the program automatically moved to the next shorter cue duration (20).
Chemogenetic inactivation was performed in cognitively challenging sessions in which either the delay was randomly varied between 5, 7.5 or 12.5 s to test inhibitory control, or sessions in which the cue duration was varied between 0.2, 0.5 or 1 s to test attentional aspects of the task (27).
Fiber photometry sessions were performed in similarly cognitively challenging sessions. In addition, rats were also retrained to baseline performance in an operant cage without homecage attachment, which was more suited to tethered recordings.

Drug administration
Two weeks before testing, animals were habituated to injections by giving them several saline injections. Directly before a testing session, Clozapine N-oxide (CNO) dihydrochloride (Hello Bio, UK) was dissolved in 0.9% saline and injected intraperitoneally (i.p.) 30 min prior to the start of the dark phase. Solutions were freshly prepared on each test day and doses were administered using a Latin square design. Animals received either 0, 5 or 10 mg/kg CNO per testing session, in randomized order, based on recent work in rats (25).

Fiber photometry
Rats used for fiber photometry were trained in CombiCages until baseline performance, and then recorded for 4-6 sessions, each lasting up to 150 minutes. We used a setup based on the one used by Lerner et al. (50) (Figure S4B for schematic overview), centered around a lock-in amplifier (RZ5P, Tucker-Davis Technologies, USA) that controls two excitation LEDs (405nm at 531Hz, and 490nm at 211Hz; Thor Labs M490F1 and M405F1). This setup allowed us to use the isosbestic wavelength of GFP as a control for motion-induced and other systemic noise, since the 405nm channel will contain all incoming signal except specifically GCaMP-emission. Light was then led through a filter cube (FMC4_AE(405)_E(460-490)_F(500-550)_S, Doric Lenses) into the fiber optic rotary joint. Rats were tethered to the recording setup with a patch cord (MFP_400/440/LWMJ-0.53.FC-ZF2.5, Doric Lenses) and a fiber optic rotary joint (FRJ_1x1_FC-FC, Doric Lenses). Emitted light from GCaMP6m was led back to the filter cube into a photodetector (Newport Femtowatt 2151), which then transmitted signal back to the lock-in amplifier which demodulated both incoming channels into separate signal traces. Data was then recorded on a dedicated recording PC using Synapse (Tucker-Davis Technologies). Incoming behavioral signals were also transmitted from the operant chamber to the lock-in amplifier using a MedPC SuperPort card (DIG-726, MedAssociates) and corresponding cable (CMF, Tucker-Davis Technologies). Using this system, we could reliably perform chronic recording experiments for over 3 months.

Electrophysiology
After obtaining a stable giga seal, a ramp current was injected from 0 to 500pA to assess baseline rheobase current. Spike frequency was determined both by increasing steps of current injection and by constant supra-threshold current injection. ACSF with 10 µm CNO was washed in for at least 5 min before rheobase current and spike frequency were determined again.
For voltage-and current-clamp experiments borosilicate glass patch-pipettes (3-5 MΩ, resulting in access resistances typically between 7 and 12 MΩ) were used with a Kgluconate-based internal solution containing (in mM): K-gluconate 135, NaCl 4, MgATP 2, Phosphocreatine 10, GTP (sodium salt) 0.3, EGTA 0.2, HEPES 10 at a pH of 7.4. Reciprocally connected MD neurons were targeted using somatic expression of red retrobeads and striatal medium spiny neurons were targeted based on morphology. Data was sampled using a Multiclamp 700B amplifier (Axon Instruments) and pClamp software (Molecular Devices) at 20 kHz and low-pass filtered at 2 kHz. Neurons were filled with 2-4% biocytin for reconstruction.
Chronos-induced post-synaptic currents (PSCs) were recorded in voltage clamp at -60mV. Chronos was activated by blue light (470 nm, 10 sweeps, 10Hz, 5 pulses of 1 ms) using a DC4100 4-channel LED-driver (Thorlabs, Newton, NJ) as light source. The light source was directed as far away from the soma as possible (typically > 200 um) and the illumination area was limited using a diaphragm such that reliable but minimal activation was achieved. Light intensity was adjusted to elicit a half maximum amplitude (typically > 10pA) of the first EPSC to prevent overstimulation of the axon boutons (Collins et al., 2018).

Exclusion criteria
Animals with misplaced virus or retrobead infusions were excluded, as were rats for the chemogenetic experiments that had unilateral virus expression or that did not establish stable baseline performance. Additionally, for the photometry experiments, rats with misplaced fibers or no virus expression were excluded. For slice electrophysiology experiments, three outliers were removed, one had a capacitance above 500 pF, and two had a Rinput above 340 MΩ, exclusion did not affect outcome.

Cellular quantification
For the Retrobead experiments, maximum intensity Z projections of 5 z-planes were made using ImageJ. Next, images were overlayed with a rat brain atlas at AP + 3.00 mm, +2.76 mm or +2.52 mm. Subregions of the mPFC were included as ROIs. Layers of the PFC were determined using the Swanson brain atlas and were validated with NeuN sections. Cells were counted manually using ImageJ per ROI and area of the ROIs was determined. For the double labeling experiments, composite images were created for signal from eYFP, GAD-67 and mCherry. Cells were counted manually. For the DREADD experiments, the areas of virus expression were selected as a ROI in ImageJ. The area of the ROI was calculated and cells within the ROI were counted manually.

Chemogenetics and behavior
Behavioral data were acquired with MED-PC software (Med-Associated, USA). All data analyses and statistics were done with custom written scripts in MATLAB (Mathworks, USA). We calculated the percentage accuracy as: #correct/ (#correct + #incorrect) * 100. Premature responses and omissions were expressed as a percentage of the total number of trials. All latencies were expressed in seconds. Trials with a magazine latency > 10 s were excluded from further analysis (20). Normality of the data was tested with the Shapiro-Wilk test. Timedependent effects of CNO were analyzed by splitting the 2.5 h session in five blocks of 30 minutes. Two-way mixed repeated measures ANOVAs were employed with time and dose as within-subject factors (20). To compare the effects of CNO in the different projection groups, three-way mixed repeated-measured ANOVAs were employed with dose and delay or cue duration as within-subject factors and group as between-subjects factor. Additional parameters, such as number of started trials, were not dependent on delay or cue duration and effects of CNO were tested with two-way mixed repeated-measures ANOVAs with dose as within-subject factor and group as between-subject factor. Post hoc testing was done using Wilcoxon rank-sum tests or t tests with Benjamin-Hochberg false discovery rate (FDR) to adjust p values for multiple comparisons. For the neuroanatomical data, a Chi-Square independence test was used to test differences in mediolateral distributions between projection populations in dorsal and ventral mPFC. In the ex-vivo electrophysiological experiments, a non-parametric Mann-Whitney-U test was used to assess the effects of CNO on mCherry (putative DREADD)-positive cells versus control neurons. To test the effects of CNO on the distribution of premature responses, paired Kolmogorov-Smirnov tests were performed between the doses and p-values were corrected for multiple testing. In all cases, the significance level was set at p < 0.05. Data are presented as mean +/-SEM throughout the main text and figures and as mean +/-SD in the supplementary tables.

Fiber photometry
Fiber photometry data was analyzed using custom made MATLAB scripts. In short, raw data from the TDT RZ5P recording system was first corrected for motion and other systemic noise by fitting the 405nm-channel to the 470nm-channel and dividing, resulting in a raw δF/F (F being the adjusted 405nm-channel). We then lowpass filtered the signal on 1Hz and highpass filtered on 30Hz. We then performed a spectral analysis to correct for remaining low frequency noise. Finally, we down sampled the signal by a factor of 64, yielding a final frame rate of around 16Hz, which was our final δF/F. For all subsequent analyses, we used small time windows around the trial. To be able to standardize signals and look only for changes in population activity associated to the task, we aligned every trace to a baseline period between -5 and -1 seconds before the start of each trial. Since we included a 10 second inter-trial interval after each trial where rats could not initiate a new trial, the baseline should not include any trial-related signals. To test differences in signal between delay periods, we only looked at signal between trial initiation and the cue presentation time of the longest delay (12.5s). We either used Friedman test (comparison between trial outcomes within group), or Kruskal Wallis one-way ANOVAs (comparison between groups), with post hoc Dunn-Sidak multiple comparison tests and Benjamini-Hochberg false discovery rate to adjust p values. Significance for ANOVAs was set on p < 0.05. To assess difference from baseline, we calculated bootstrapped confidence intervals with 5000 iterations and alpha of 0.001. In short, we randomly sampled mean signal traces for each outcome type for each rat and took the mean of each random sample (each random sample being the same as the total number of rats in the group), and repeated 5000 times. We then took a confidence interval with an alpha of 0.001 of all 5000 mean traces of a given trial outcome, yielding an interval between the 99.9 th and 0.01 st percentile value for each data frame, which we considered as boundaries between which the signal could be. We then took averages of the upper and lower confidence interval bounds of all rats to construct the group confidence interval. To study differences between signal traces of two experimental groups or two outcomes, we performed permutation tests that compared distributions at every data point. For each data point, we considered the distributions significantly different if the alpha was < 0.01. For both the bootstrapping and permutation tests, singleton significant points (i.e. data points with no neighbors that were also significant) were filtered out of the data set. One data frame corresponded to approximately 125ms.

Electrophysiology
Chronos-evoked PSCs were calculated by taking the median over 10 sweeps that were corrected for drift using a robust regression fit. Paired-pulse-ratios were calculated by dividing the peak of PSCN by PSC1. Chronos-evoked PSC latency was calculated as time to reach 80% of peak value from light onset. Input resistance was calculated using the slope of the linear fit to the current-voltage curve using negative current steps between 0 and -100 pA (15 or 20 pA increments, 0.5 or 1 s duration), using the steady state voltage in the last 200 ms of the step. The membrane time constant tau determined by the median over fitting a first order exponential function (only goodness of fit > 0.8 used) to the first 300 ms to the voltage trace in response to three negative current steps between 0 and 50 pA (15 or 20 pA increments, 0.5 or 1 s duration). Capacitance was calculated as input resistance over membrane time constant. Sag was calculated as the percentage difference between the Δ peak voltage and Δ steady state (last 1/5 th of the step duration) from baseline in response to a negative step current (0.5 or 1 s) that elicited a Δ peak voltage closest to -20mV. Burst and steady state firing frequency were calculated based on the number APs (threshold at 0 mV) in the 50 ms after the first AP (burst) or the last 200 ms (steady state) of positive current steps between 0 and 200 pA (50 pA increments, 0.5 s duration). Some neurons were recorded with 15 pA increments, here steps with less than 5pA difference from the 50 pA increments steps were used. Biocytin-filled neurons were reconstructed in Neuromantic software (V1.6.3) and plotted for illustrative purposes using the Neuroanatomy toolbox in ImageJ. Offline data analysis was performed in Graphpad Prism 6 and Matlab 2019a. No assumptions were made about the data distribution and all analyses were done using nonparametric Friedman with post-hoc Dunn's and Benjamini-Hochberg's false discovery rate corrected Mann-Whitney U-tests for repeated measures and Mann-Whitney U-tests for simple comparisons, significance set at P < 0.05.

Data and code availability
Datasets are available upon reasonable request.  Bottom: Examples of GCaMP6m expression in axon terminals in target areas. White dashed lines: Brain atlas overlay. Scale bars: 500μm. E) Example heatmaps for animals with either GFP (left), or GCaMP6m expressed in projection populations. Only sessions with variable delay, but all trial outcomes and delays have been pooled together. Data are z-scored to baseline (t-5 to t-1 from trialstart). F) Cumulative density plot for bins that are >2std above or below baseline. X axis is number of bins, Y axis is cumulative proportion of trials. (B, D) Sag, calculated as percentage difference between Δ steady state (SS) and Δ peak from a hyperpolarizing current step resulting in a peak voltage change closest to -20mV. Left: example trace, right: summary plot. Boxplots: center line, median; box edges, 1st and 3rd quartile; whiskers, data range without outliers.

Supplementary Tables
Table S1 Effects of chemogenetic cortico-thalamic inactivations on var delay and var cue duration parameters. Summary of behavioral parameters from variable delay and cue duration sessions shown per delay /cue duration. Green colored cells represent significant increases compared to saline after dose x group interaction. Red/green colored and underlined parameters reflect significant decreases/increases compared to saline respectively after dose x group x delay /cue duration interaction. Post-hoc tests are FDR-corrected for multiple testing. CNO5: CNO 5 mg/kg. CNO10: CNO 10 mg/kg injection. Data are expressed as mean ± SD.