Thalamic circuits for independent control of prefrontal signal and noise

Interactions between the mediodorsal thalamus and the prefrontal cortex are critical for cognition. Studies in humans indicate that these interactions may resolve uncertainty in decision-making1, but the precise mechanisms are unknown. Here we identify two distinct mediodorsal projections to the prefrontal cortex that have complementary mechanistic roles in decision-making under uncertainty. Specifically, we found that a dopamine receptor (D2)-expressing projection amplifies prefrontal signals when task inputs are sparse and a kainate receptor (GRIK4) expressing-projection suppresses prefrontal noise when task inputs are dense but conflicting. Collectively, our data suggest that there are distinct brain mechanisms for handling uncertainty due to low signals versus uncertainty due to high noise, and provide a mechanistic entry point for correcting decision-making abnormalities in disorders that have a prominent prefrontal component2–6.


Supplemental Introduction:
In decision making, the ability to estimate uncertainty is critical for optimizing outcomes 1 .
Uncertainty can be about sensory inputs (input uncertainty 2 ), their mapping onto internal/behavioral variables (rule uncertainty 3 ), or their likelihood of predicting reward (outcome uncertainty 4 ). Input uncertainty is mostly studied in sensory systems 5,6 but executive systems are also confronted with uncertain inputs that need to be turned into discrete control signals 7 . For example, at an event where multiple languages are spoken, it may take time to decide whether spoken words belong to one of two closely related languages, a form of uncertainty that may prohibit one from joining their preferred conversation. Conversely, uncertainty may be due to silence. In either scenario, resolving uncertainty is a prerequisite for directing attention, which is thought to involve the prefrontal cortex (PFC) 8 . Prefrontal dysfunction is found in disorders like schizophrenia, where patients may fail to efficiently and optimally deploy cognitive resources 9 .
Recent studies have shown that the mediodorsal thalamus (MD) is a critical partner for the PFC in generating attentional signals, and cognitive control more broadly [10][11][12] . While earlier work in animals has shown that the MD enhances task relevant prefrontal activity patterns 13,14 and suppresses task irrelevant ones [15][16][17] , recent studies in humans show that MD engagement in decision making scales with the degree of input uncertainty 18,19 . In this study, using mice, we found that the enhancement of activity required for maintaining working memory and executive control signals 14,21,22 , and suppression of activity required for task switching 23,24 . A clear case for this functional segregation comes from an attentional control task that was previously developed and validated 14 . Briefly, in this task a freely behaving mouse selects between a visual and auditory target whose spatial locations are pseudorandomized (Extended Data Fig. 5a). On each trial, the selection process is guided by a 100 ms-long learned cue; either a high-pass (HP) or low-pass (LP) filtered white noise pulse, which correspond to one of two rules -attend to audition and attend to vision respectively. A mouse is required to hold the presented cue in mind over a delay prior to the simultaneous presentation of the two targets. The targets correspond to the spatial location of the reward being delivered through a right or left reward port (e.g. left LED flash on an attend to vision trial or an upsweep on an attend to audition trial signaled a response on the left reward port). Mice perform at chance level when the PL is optogenetically suppressed either during the cueing or delay periods of the task 14 . In contrast, when an animal achieves stable performance on the task, MD inactivation on single interleaved trials is only disruptive when delivered during the delay period 14 . However, MD inactivation during the cueing period has a detrimental effect when the cues are switched, and the MD exhibits unique electrophysiological signatures associated with prefrontal suppression 24 . These findings give rise to the notion that MD-dependent PL suppression is required for setting up the appropriate task-relevant connection patterns 24 which MD-dependent PL activation can later work to maintain 14,22 .
Could these two functions be subsumed by the two identified MD cell types? To test this idea, we trained D2-cre and GRIK4-cre mice on the attentional control task described above (Extended Data Fig. 5a). Interestingly, we found that animals required several trials to achieve stable performance on this task, a quality that was not associated with a visual detection task that we previously showed was not PL dependent (Extended Data Fig. 5b-d). Therefore, we hypothesized that this initial 're-learning' requirement is the consequence of the PL needing to be configured for the attentional control task, and that this process may be MD dependent. Indeed, cue-specific MD suppression during this 're-learning' period resulted in animals taking longer to become engaged in the task (Extended Data Fig. 5e, i). Equivalent MD suppression in the PL-independent, visual detection, task had no effect on trials taken to get task engaged (Extended Data Fig. 5f).
Importantly, we found that MDGRIK4 but not MDD2, suppression resulted in a similar behavioral effect (Extended Data Fig. 5g-k). In contrast, MDD2 but not MDGRIK4 inactivation reproduced the total MD inactivation effects on behavior when delivered during the delay period of the task (Extended Data Fig. 5l-n). These experiments confirmed that the functional segregation of the identified thalamic cell types have distinct behavioral effects on our previously established PLdependent attention control task. 3. Cue-specific MD inactivation diminishes task performance in the first few trials of task switching (See panels 3e-g, Supplementary Fig. 11 in Rikhye et al.) 24 , which we now demonstrate is also evident during task engagement as a proxy for switching (See panels a-f in Extended Data Figure 5 in the current manuscript).

MD inactivation in the cueing sequence tasks (introduced in the current manuscript):
1. In the current manuscript, MD inactivation was never performed during the delay period.
If it was, we expect it to have similar effects to the original task but since the focus on the neural activity was limited to evidence integration, we limited our manipulations to that period.
2. All manipulations were done in the steady-state behavior.
3. Under those conditions, we find that MD inactivation has minimal impact on the behavior when the cueing ambiguity is low (akin to what we find in the single cue version of the task). We interpret these findings as evidence of consistency in the role of the MD across these different tasks.
4. When the cueing ambiguity is increased, cue-specific MD inactivation has a detrimental effect on behavior. This is an original finding reported in this manuscript.

Supplemental Discussion:
Decision making is a topic of broad relevance, impacting several fields such as neuroscience, psychology, and economics 25,26  Our findings are broadly consistent with previous findings of frontal cortical activity patterns encoding task-relevant inputs (cues) and outputs (rules) [30][31][32] , and MD thalamic responses being 'contextual' in nature 14,16 . These differences in encoding regimes may be related to differences in architectural features across these areas; although both structures are predominantly composed of excitatory glutamatergic neurons, the thalamus is devoid of local excitatory recurrent connections, a prominent feature of cortical organization 33 . The lack of local recurrence may actually be a design feature that allows these circuits to control cortical activity patterns that carry signals and noise independently. In addition, the nature of MD encoding of input uncertainty-related signals is also consistent with 'summary statistic' type responses seen in the Pulvinar of non-human primates performing a perceptual decision making (random dot motion) task 34,35 .
Our data may speak to how fronto-thalamic circuits handle uncertainty in input statistics, categorizing it to low signals or high noise. It will be interesting to see how these circuits operate under different types of task-relevant uncertainty, such as those related to mapping (onto a rule) or outcome (probability of reward). This is particularly pertinent given the recent pioneering work on MD-PFC interactions in outcome uncertainty 31 . Such investigations are likely to yield important insights into how decision processes are organized, and how uncertainty at different hierarchical levels interact, perhaps through cortico-thalamic loops. In addition, it will likely be relevant for linking the computational abnormalities seen in schizophrenia to the underlying circuits to identify therapeutic targets in the process.