Mixed selectivity encoding and action selection in the prefrontal cortex during threat assessment

Abstract

The medial prefrontal cortex (mPFC) regulates expression of emotional behavior. The mPFC combines multivariate information from its inputs, and depending on the imminence of threat, activates downstream networks that either increase or decrease the expression of anxiety-related motor behavior and autonomic activation. Here, we selectively highlight how subcortical input to the mPFC from two example structures, the amygdala and ventral hippocampus, help shape mixed selectivity encoding and action selection during emotional processing. We outline a model where prefrontal subregions modulate behavior along orthogonal motor dimensions, and exhibit connectivity that selects for expression of one behavioral strategy while inhibiting the other.

Introduction

The medial prefrontal cortex (mPFC) is a critical region for input integration and higher order computations in many cognitive domains, including anxiety. When incorporating information about the relative threat that a stimulus poses in the face of other factors – for example one’s motivational state, stimulus history, and its similarity to previous threats - the mPFC plays an important role in generating and suppressing defensive motor behavior, as is warranted by each individual situation. This is achieved by coordinating a multitude of downstream projections to areas such as the thalamus, amygdala, and hypothalamus [1,2], which together orchestrate the particulars of the behavioral and autonomic repertoire. Consistent with its role as an integrative control center, loss of mPFC function in patients with Post-Traumatic Stress Disorder (PTSD) and Generalized Anxiety Disorder (GAD) is associated with indiscriminate anxious behaviors, such as avoidance, and autonomic arousal, which can lead to self-medication with alcohol or drugs of abuse [3–5]. Using a combination of experimental and computational approaches, we are beginning to piece together how cortical and subcortical inputs shape prefrontal population activity to encode different aspects of a stimulus and its environment, resulting in a rich repertoire of neural activity that represents different behavioral strategies for dealing with threat.

Mixed selectivity encoding of conditioned emotional stimuli in the mPFC

Widespread avoidance and defensive responding to relatively safe stimuli are cardinal features of PTSD. From a neurobiological perspective, displaying the same behavioral repertoire to threatening and non- threatening events is evidence of deficient neural mechanisms that should be driving divergent behaviors [6]. Whereas many regions partake in encoding disparate features of incoming stimuli, the mPFC serves as an important convergence point that is well-placed to integrate cortical, sensory and subcortical inputs about various aspects of an emotional stimulus, and select an appropriate response. Some example inputs the mPFC receives about a cue include signals from sensory cortices bringing auditory and visual information [7,8], afferents from basolateral amygdala (BLA) providing information about stimulus valence and salience [9–13], the hippocampus providing contextual and spatial information [14–19], thalamic inputs improve communication within the mPFC [20,21], and neuromodulators like acetylcholine contribute to shaping mPFC neural activity [22]. As a result of this vast system of inputs, the mPFC, in concert with other regions, mediates expression and suppression of anxiety related behaviors, including defensive freezing [9,26–29], threat avoidance [30], and stimulus discrimination [31,32]. To study the underlying neural mechanisms of PTSD in the laboratory, stimulus contingencies of the real world are modeled via tasks such as differential fear conditioning, extinction, avoidance training, and presentations of competing motivational stimuli, tasks that collectively drive observable coping strategies, such as passive and active avoidance, reward seeking or defensive freezing [10,23]. While these tasks are impoverished relative to the full spectrum of stimuli and behavioral repertoires displayed by animals in the wild [24,25], they have been useful for studying neural communication during isolated, but clinically-relevant, aspects of PTSD. Thus, proper mPFC function is critical for managing many of the behavioral components of anxiety, as the mPFC parses incoming information about stimulus valence in the environment and communicates this information downstream.

Given the richness of its inputs, the mPFC is proposed to use mixed selectivity encoding during threat assessment and action selection, as has been previously shown in working memory and categorization tasks [33–36]. This encoding scheme postulates different permutations of neural response profiles in mPFC cell populations, which provide a multi-dimensional space for encoding anxiety-related information, depending on the particular combination of simultaneously occurring input patterns (Figure 1). This strategy allows for efficient, non-linear, and non single-feature selective encoding of highly processed input [33,34,37]. Interestingly, in operant tasks, mPFC neurons have been shown to form different mixed selectivity ensembles as different actions are selected, but keep similar overall numbers of cells, even as single neurons drop in and out of the encoding population [38,39]. Thus, in an actively engaged mPFC, the contribution of each neuron to the population appears to be relatively stable. It is likely that in the compromised mPFC, as in PTSD, there is a decreased pool of neurons readily available to fill in when some drop out of an ensemble.

Figure 1. Schematic of mixed selectivity encoding and action selection in the prefrontal cortex.

As stimuli are processed, a mix of sensory, subcortical, cortical and neuroendocrine inputs establish neural ensembles in the mPFC that incorporate multivariate information using a mixed selectivity encoding strategy. Neural ensembles in the PL drive CMI-based defensive responses, whereas populations in the IL drives CME-based responses (behavioral and autonomic) that are particularly evident during fear suppression. When one behavioral strategy is selected, mutual inhibition between the IL and PL inhibits competing strategies.

Notably, the mPFC is not the only brain region that has been proposed to demonstrate features of integrative or mixed-selectivity encoding. For example, some hippocampal cells have been shown to encode temporal and contextual information [40–42], and population activity in amygdala neurons has been shown to be best explained by a combination of spatial attention and stimulus valence [37,43,44]. It is therefore possible that with further investigation, these and other regions will also demonstrate features of multi-dimensional feature encoding. Thus far, however, feature encoding in these regions was shown to have a reduced dimensionality relative to the mPFC, and to be restricted to subsets of cells, although this may change with further study, given the high afferent connectivity of these structures [45,46].

Action selection in the prelimbic and infralimbic cortices of the mPFC

An important consequence of mixed selectivity encoding is action selection via downstream activation of stimulus-relevant motor output [34,37,38,47]. In animal models, this type of behavioral selection can be tested during differential fear conditioning recall, after two stimuli are first conditioned to either always (CS+) or never (CS-) co-terminate with an inherently aversive unconditioned stimulus (US) [32]. Likewise, during extinction of conditioned fear, a CS+ is made less aversive via repeated presentations without the previously accompanying US, resulting in the subject learning a new and relatively less aversive contingency for the CS. In each case, neural activity is compared between the relatively more and less aversive versions of the CS, which drive different motor action selection. These comparisons demonstrate that activation of the infralimbic subregion (IL) of the mPFC in rodents (and ventromedial [vmPFC] in primates and humans) is associated with decreased expression of defensive freezing [5,28,31,48,49], whereas activation of the more dorsally located prelimbic (PL) subregion is involved in expression of defensive freezing when assayed with the same paradigms [9,10,26,27].

In a mixed selectivity model of mPFC function during threat assessment, neural population activity in the PL during defensive freezing or in the IL during suppression of freezing, is likely to reflect the associative learning about a particular CS combined with activation of different behavioral strategies to deal with that CS, such as active escape or defensive freezing [23,50,51]. In keeping with this, unit recordings have shown that activity in a large proportion of mPFC cells predicts a particular behavioral strategy better than a single-feature identity of the CS, such that the same aversive CS+ elicits firing in some cells during conditioned motor inhibition (CMI, e.g. freezing) and in other cells during conditioned motor excitation (CME, e.g. active movement [50]). Furthermore, pharmacological inhibition of the IL decreases CME during the CS+, whereas activation of the IL has the opposite effect, suggesting that population activity in the IL is associated with selecting for CME behavioral strategies [52] whereas changes in CMI are controlled via projections from the PL (Figure 2). Thus, each subregion is likely to use mixed selectivity encoding to integrate multi-dimensional feature information about a CS while also activating the conditioned excitatory or inhibitory motor repertoire to that CS. These findings are in line with previous work on mixed selectivity encoding in memory tasks, which show that higher-dimension encoding predicts behavior more accurately than lower-dimension encoding [33–34].

Figure 2. IL and PL control behavioral output along orthogonal motor dimensions.

IL activation increases conditioned excitatory motor (CME) output, whereas PL activation increases conditioned inhibitory motor (CMI) output. These two types of output are proposed to move along orthogonal dimensions, from high-to-low excitatory as IL becomes engaged and disengaged, and from high-to-low inhibitory as PL becomes engaged and disengaged. Activation of IL-driven excitatory behaviors inhibits PL-driven behaviors and vice versa. When both subregions are inactive, the mPFC is not engaged in conditioned emotional responding (lower left quadrant). Due to proposed mutual inhibition, the top right quadrant is represented by a null set, as simultaneous CME and CMI does not occur.

Recent work on action selection in females show that female rodents learn to discriminate between a CS+ and CS- faster than males (within 1 day), and are prone to generalization with longer training (after 3 days of training) whereas males generalize initially and discriminate better with increased training [53]. These findings are consistent with females showing faster learning rates on extinction training [53,54], suggesting a different temporal evolution of neural changes or possibly alternate mechanisms for those changes in females. For example, impaired extinction in male but not in female mice is associated with dendritic retraction and increased presence of thin spines in the IL [55,56]. Such differences may indicate that the IL is modulated on different time scales in males and females, or that poor extinction in females is better correlated with changes in different regions. One area of interest in this regard is the Bed Nucleus of the Stria Terminalis. An in-depth discussion of this region is beyond the scope of this review, but has recently been shown to regulate tonic levels of anxiety, and to interact with the mPFC, neuroendocrine and autonomic systems (for a review, see [57,58]).

Subcortical contribution to mixed selectivity encoding and action selection in the mPFC

To gain a better understanding for how mixed selectivity encoding may occur in the mPFC, we take a closer look at the organization of two example inputs, the basolateral amygdala (BLA), and the ventral portion of the hippocampal formation (vHPC). Given the fast latencies of BLA responding to sensory stimuli [59], evidence has been accumulating that BLA input could prove to be a crucial contributor to mixed selectivity encoding and action selection in the mPFC, especially during the initial stages of acquisition [9,10,13,60]. The BLA has been shown to project to the PL and IL, and monosynaptically excite mPFC pyramidal cells that in turn project to the periaqueductal grey (PAG), nucleus accumbens, and also mPFC cells that project back to the amygdala [61–63]. In addition to excitation of the mPFC, BLA input was also shown to drive strong inhibitory effects in the mPFC. In vivo and in vitro recordings demonstrate that the BLA briefly excites pyramidal neurons but then exerts a strong inhibition [61,62,64–66] via feedforward activation of parvalbumin-expressing (PV+) and somatostatin-expressing (SOM+) local interneurons [62]. Simultaneous recordings in the PL and IL of conditioned rats show that firing rates in the IL dramatically decrease right after conditioning (possibly due to BLA input), while BLA input to the PL is important for production of CMI in the form of defensive freezing [9,10,67–69]. Furthermore, BLA-mediated inhibition of the IL was shown to be particularly strong onto layer 2 pyramidal cells, which project back to the BLA [62]. Critically, the IL-to-BLA projection is known to shut down amygdala output and decrease defensive freezing during extinction and safety discrimination [28,70–72]. Thus, BLA activation of the mPFC could bias prefrontal selection of CMI in several ways: (1) by contributing to mixed selectivity encoding in the PL [9,10,67], while (2) simultaneously suppressing IL outputs that orchestrate CME, including the suppression of a reciprocal projection from the IL back to the amygdala that would otherwise inhibits amygdala output [62]. Notably, a quiescent mPFC and overactive amygdala have been described in patients with PTSD and GAD [73,74], suggesting that in anxiety disorders, amygdala could bias for action selection of CMI in the mPFC irrespective of associative learning about an aversive CS [9,10,60,67,75], but on a more permanent basis.

The vHPC is another important subcortical input for encoding CS-related information for mixed selectivity encoding in the mPFC [76–78]. The vHPC innervates both the BLA and the mPFC, with single vHPC neurons sending axon collaterals that can depolarize both structures simultaneously [79,80]. This arrangement creates the anatomical framework for synchronous communication and Hebbian plasticity within the vHPC-BLA-mPFC network [16,81]. Afferents from the vHPC have been shown to contribute spatial and contextual information to the multi-dimensional encoding in the mPFC during threat assessment. In rodents, mPFC-projecting vHPC neurons were demonstrated to be driven by innately anxiogenic contexts, such as the open arms of an elevated plus maze and the aversive context where the CS-US pairing originally occurred [16,78,82,83]. As in the case of BLA input to the mPFC, the vHPC is thought to gate action selection of CMI and CME by modulating interneurons and pyramidal cells in the mPFC [67].

Mutual inhibition in the mPFC

Action selection during threat assessment suggests that while CME is actively driven by population activity in the IL, patterns of neural activity associated with CMI in the PL should be inhibited. In keeping with this, IL and PL cells selectively fire during CME or CMI, respectively. Furthermore, there is evidence for inhibition of PL neurons when IL neurons are active [29,50,68,84,85].

Mutual inhibition between the PL and IL during CMI and CME suggests active participation of prefrontal interneuron populations during CS encoding and behavioral strategy selection. In keeping with this, evidence for a prominent role for interneurons in appropriate action selection comes from multiple avenues of research. For example, fast spiking, PV+ GABAergic cells in the PL play a key role in generating a low theta (4-6 Hz) oscillation that is associated with production of CMI behavior such as defensive freezing [26,86]. Notably, the low theta oscillation is an important physiological vehicle for synchronizing mPFC output with other structures, such as the BLA and vHPC, during aversive learning [16,26,70,71,86,87]. Thus, PV+ interneurons act as potential pacers of a communication channel that is widely used by the mPFC to synchronize its output with structures downstream. Furthermore, an anatomical framework for mutual inhibition between the IL and PL subregions of the mPFC is beginning to emerge. Bipolar IL GABAergic cells that express neuropeptide Y have been shown to inhibit pyramidal neurons of the PL [88]. Likewise, BLA-projecting PL cells are under tight inhibitory control from local circuit interneurons in the mPFC [89]. Additionally, as mentioned above (see, Subcortical contribution to mixed selectivity encoding and action selection in the mPFC), mPFC afferents drive inhibition as well as excitation. The importance of inhibitory signaling for shaping the prefrontal response to external stimuli is underscored by the fact that dramatically upregulated inhibition in the mPFC is a prominent feature of stress-induced anxiety, with evidence that female mice are more affected than male mice [90,91]. It has been shown that in response to stress, increased inhibition is likely to drive dendritic shortening and reduced plasticity in the mPFC [55,92]. Thus the interplay of inhibition within and between mPFC subregions integrates information about the CS, participates in selecting for a specific behavioral strategy for dealing with the CS, and inhibits competing behavioral responses from adjacent subregions.

Conclusions

Combining experimental and computational approaches to tackle cognition is yielding insights into general principles of stimulus encoding and action selection in the mPFC. The PL and IL integrate inputs from many sources to establish flexible population codes that use mixed selectivity to drive conditioned motor inhibition or excitation. It’s important to determine how afferents interface with the mPFC to establish ensembles that can learn new associations and suppress anxiety-like behaviors when they are not warranted.

Figure 3. Contributions of basolateral amygdala and ventral hippocampal inputs to mixed selectivity encoding in the mPFC.

The IL and PL receive processed subcortical information about the valence and salience of the CS from the BLA and spatial/contextual valence information from the vHPC. Their inputs are crucial for shaping mixed selectivity encoding as they drive CME and CMI across the PL and IL. Amygdala input contributes to action selection in the mPFC by exciting neurons in the PL that drive CMI and inhibiting IL neurons that drive CME.

Highlights.

• The medial prefrontal cortex uses a mixed selectivity model for encoding emotional conditioned stimuli

• The prelimbic and infralimbic subregions of the medial prefrontal cortex use mixed selectivity to drive conditioned motor inhibition and conditioned motor excitation, respectively

• Subcortical inputs from the amygdala and ventral hippocampus are two areas that contribute to mixed selectivity coding in the medial prefrontal cortex

• Prelimbic-mediated behaviors are inhibited when the infralimbic cortex is active