Manipulating neural activity in physiologically classified neurons: triumphs and challenges

Abstract

Understanding brain function requires knowing both how neural activity encodes information and how this activity generates appropriate responses. Electrophysiological, imaging and immediate early gene immunostaining studies have been instrumental in identifying and characterizing neurons that respond to different sensory stimuli, events and motor actions. Here we highlight approaches that have manipulated the activity of physiologically classified neurons to determine their role in the generation of behavioural responses. Previous experiments have often exploited the functional architecture observed in many cortical areas, where clusters of neurons share response properties. However, many brain structures do not exhibit such functional architecture. Instead, neurons with different response properties are anatomically intermingled. Emerging genetic approaches have enabled the identification and manipulation of neurons that respond to specific stimuli despite the lack of discernable anatomical organization. These approaches have advanced understanding of the circuits mediating sensory perception, learning and memory, and the generation of behavioural responses by providing causal evidence linking neural response properties to appropriate behavioural output. However, significant challenges remain for understanding cognitive processes that are probably mediated by neurons with more complex physiological response properties. Currently available strategies may prove inadequate for determining how activity in these neurons is causally related to cognitive behaviour.

1. Introduction

Our understanding of brain function has been greatly advanced by the development and application of methods for manipulating the activity of neurons in awake experimental subjects. The suppression of neural activity through lesions, pharmacological inactivation or genetic manipulations can powerfully demonstrate the necessity of targeted neural elements in mediating specific functions. However, a detailed understanding of the neural mechanisms underlying different cognitive, emotional and behavioural processes requires knowing more than the identity of the critical neural elements. A full understanding demands characterization of the information encoded by the activity of neurons, and of how this information produces meaningful behavioural output.

The manipulation of neural activity based solely on neurons' anatomical location, anatomical connectivity, morphology or molecular phenotype may activate neurons that do not share physiological response properties. Thus the effects of activating neurons targeted by these features may not be predicted by the information that these neurons encode. By contrast, the ability to manipulate neural activity in populations of neurons targeted by virtue of their physiological response profile provides a powerful approach for linking neuronal response properties to behavioural effects.

In this review, we highlight studies focused on manipulating neural activity in physiologically classified neurons to understand the relationship between neural activity and behaviour. Perhaps the most influential method for activating physiologically defined neurons in experimental subjects has been the application of electrical microstimulation to targeted neuronal elements [1–7]. The earliest studies using electrical stimulation focused on stimulating motor systems to elicit movements [8,9]. Penfield and co-workers [10,11] then established that applying stimulation to many brain areas outside of the motor system could produce sensory and cognitive effects. Indeed, this approach was used to verify the presence of topographic maps in sensory cortices; for example, stimulation of primary visual cortex produces phosphenes at the location of the receptive field of the stimulated neurons [12–15].

The discovery that electrical microstimulation could produce sensory effects led to approaches that used the power of electrophysiological recordings in combination with carefully designed psychophysical tasks to test causal hypotheses about the role of physiologically classified neurons. These studies typically exploit the fact that many sensory areas exhibit functional organization whereby neurons with similar physiological properties are clustered together anatomically (e.g. in cortical columns). This organization permits the positioning of microelectrodes in the middle of a cluster of neurons that encode a similar sensory parameter. The goal is to preferentially activate neurons that encode this feature. In a series of experiments, Salzman, Newsome and co-workers [4,5,16–20] provided causal evidence that the activity of direction selective neurons in visual area MT is related to perceptual judgements of motion direction. The effects of stimulation were closely tied to the physiological properties of neurons at the stimulation site, as moving the electrode very small distances (100–300 μm) could alter the effect of stimulation dramatically [16,19]. Furthermore, increasing stimulating current to levels that would directly activate neurons well beyond the targeted cortical column resulted in a loss of directional influence on perceptual judgements [16]. Instead, performance deteriorated, as if noise had been injected into the sensory representation. Taken together, these experiments illustrate that when neurons sharing physiological response properties are anatomically clustered, electrical microstimulation provides a powerful approach for establishing a predictable and causal link between neural activity and behaviour.

These classical studies have now been extended to investigate other aspects of neuronal response properties in MT. For example, MT neurons also exhibit selective responses to binocular disparity, and neurons with similar preferred disparities are clustered anatomically [21]. Electrical microstimulation targeted to neurons categorized according to two parameters, direction and disparity preference, has revealed a causal link between activity and performance in tasks in which subjects must judge disparity or discriminate structure from motion [5,21,22]. Similar approaches have also been used to study causal mechanisms mediating different aspects of perceptual decision-making and attention. Stimulation has been applied to physiologically classified neurons in somatosensory cortex [23,24], inferotemporal cortex [25], the frontal eye fields [26,27], the lateral and ventral intraparietal areas [28,29] and visual area MST [30–32], among other areas. These experiments all took advantage of the physiological response properties recorded at a stimulation site to test whether activity predicts behavioural effects according to the stimulated neurons' response preference.

The success of the microstimulation experiments outlined above is critically dependent on the anatomical clustering of neurons that share response properties. However, in many brain areas neurons with different and even opposing physiological response properties are anatomically intermingled. Thus, the application of microstimulation cannot selectively activate neurons that share a particular physiological response feature. Establishing causal mechanisms for physiologically classified neurons in these brain systems therefore requires new experimental approaches. The remainder of this review summarizes recent studies that use activity-dependent expression of molecules sensitive to chemical or optical stimulation. These approaches allow the manipulation of neuronal activity in physiologically classified neurons that may be anatomically intermingled. Although these strategies have proved powerful in elucidating the neural circuits that mediate several innate and learned behaviours, significant challenges await. Many cognitive functions likely rely on algorithms that integrate information from neurons with complex physiological response properties. Current strategies for manipulating physiologically classified neurons may therefore provide insufficient specificity to study the mechanisms underlying complex cognitive behaviours.

2. Genetic strategies for activating neurons based on physiological response properties

Our current understanding of the molecular, morphological and anatomical features of neurons is often insufficient to deduce the specific physiological response properties encoded by neurons (with the potential exception of dopaminergic neurons [33]). Increasing attention is therefore being paid to novel genetic strategies that permit the labelling and manipulation of populations of neurons based upon their responses to individual stimuli [34–38].

Immediate early genes (IEGs), including Arc, c-fos and zif268, are transiently expressed in response to cellular depolarization [39–41]. IEGs therefore provide a potential tool for identifying neurons that respond to a specific stimulus. Indeed, immunostaining for the protein products of IEGs has identified circuits activated by numerous sensory stimuli and events [42,43]. In addition, exploitation of the temporal dynamics of c-fos and Arc RNA migration from the nucleus to the cytoplasm has permitted the identification of cells that respond to different stimuli in the same animal [44]. These studies have revealed distinct ensembles that are activated by mating and fighting in the ventromedial hypothalamus [45], and distinct populations of neurons that are activated by an appetitive and aversive unconditioned stimulus in the basolateral amygdala (BLA) [46]. IEG immunostaining therefore provides a powerful genetic means to identify neurons activated by specific stimuli or events. However, the transient nature of expression limits its use over prolonged time periods. Moreover, the visualization of stimulus representations using IEGs offers no indication as to the causal role of these representations in behaviour. Recent efforts have therefore focused on using IEG promoters to drive the expression of reporters to facilitate the prolonged labelling and manipulation of cells activated by specific stimuli (figure 1). These emerging technologies have afforded novel insight into the neural circuits that mediate a range of behavioural responses to complex sensory stimuli. Here we will survey some of the advances these approaches have made in our understanding of the circuitry mediating innate olfactory behaviours, contextual drug conditioning, contextual fear conditioning and classical cued conditioning.

Figure 1.

Genetic strategies for manipulating the activity of physiologically classified neurons. (a) Schematic of Arc:CreER^T2 transgenic approach for activating physiologically classified neurons. Arc:CreER^T2 transgenic mice are injected with an adeno-associated virus (AAV) encoding a Cre-dependent channelrhodopsin fused to enhanced yellow fluorescent protein (EF1α:DIO-ChR2-EYFP). In response to cellular activity, the Arc promoter drives the expression of CreER^T2. In the presence of tamoxifen, CreER^T2 migrates to the nucleus and effects recombination between loxP sites in EF1α:DIO-ChR2-EYFP. This results in the inversion of the ChR2-EYFP sequence and persistent expression of ChR2-EYFP in active neurons. (b) Schematic of Daun02 approach for inactivating physiologically classified neurons. Daun02 is infused into target brain regions of c-fos:lacZ transgenic rats. In response to cellular activity, the c-fos promoter drives expression of LacZ, resulting in production of β-galactosidase. β-galactosidase converts Daun02 to daunorubicin, which diminishes calcium-dependent action potentials and induces apoptosis. This results in silencing of active neurons. (c) Schematic of c-fos:tTA transgenic approach for activating physiologically classified neurons. c-fos:tTA transgenic mice are injected with an AAV encoding ChR2-EYFP under the control of the tetracycline response element (TRE). In response to cellular activity the c-fos promoter drives expression of the tetracycline transactivator (tTA). In the absence of doxycycline, tTA binds to TRE to drive expression of ChR2-EYFP. However, in the presence of doxycycline, tTA binds to doxycycline preventing expression of ChR2-EYFP. This results in labelling of active neurons with ChR2-EYFP in the absence of doxycycline. (d) Schematic of lentiviral c-fos:ChR2-EYFP-2A-mCherry approach for activating physiologically classified neurons. In response to cellular activity, the c-fos promoter drives expression of ChR2-EYFP and mCherry. This results in labelling of active neurons with ChR2-EYFP and mCherry.

3. Innate olfactory behaviour

Odourants in the external environment bind to receptors on sensory neurons in the olfactory epithelium [47]. Each sensory neuron expresses 1 of over 1000 sensory receptors. Neurons expressing a given receptor project to two specific glomeruli in the olfactory bulb [48]. Individual glomeruli send spatially stereotyped projections to the cortical amygdala, with each glomerulus innervating a different region of the cortical amygdala. Glomeruli also send diffuse, apparently random projections across the entire piriform cortex, such that the projection pattern of one glomerulus in piriform cortex is indistinguishable from another [49]. This connectivity has been proposed to provide an anatomical substrate for innate and learned olfactory behaviours, respectively. Root et al. labelled and manipulated neurons responsive to innately appetitive and aversive odours by using the Arc promoter to drive expression of the light-activated cation channel, channelrhodopsin [50]. The authors injected an AAV encoding a Cre-dependent channelrhodopsin fused to enhanced yellow fluorescent protein (ChR2-EYFP) into the cortical amygdala of a transgenic mouse in which the Arc promoter drives the expression of a tamoxifen-dependent Cre-recombinase (CreER^T2). In this system, neuronal activity induces expression of CreER^T2. In the presence of tamoxifen, Cre mediates recombination between loxP sites in the Cre-inducible ChR2-EYFP, resulting in persistent expression of ChR2-EYFP in active neurons (figure 1a). Using this strategy, the authors were able to identify distinct ensembles that responded differentially to an innately appetitive and an innately aversive odourant. Moreover, photoactivation of these ensembles elicited approach and avoidance behaviours, respectively. Thus, the authors were able to identify neuronal representations in cortical amygdala that comprise determined neural circuits for the generation of innate olfactory responses [50].

These data extend a neural circuit that mediates behavioural responses to innately attractive and aversive odourants from the nose to the amygdala. Cortical amygdala sends projections to multiple structures implicated in the generation of behavioural responses, including the nucleus accumbens and the extended amygdala [51,52]. Optogenetic manipulations of the outputs of these appetitive and aversive ensembles may provide insight into the systems-level interactions that mediate innate valence-specific behavioural responses.

4. Contextual drug conditioning

Drugs of abuse can become associated with stimuli present in the environment during drug administration. These contextual cues can thereby acquire the ability to modulate behaviour [53–55]. For example, contextual cues associated with cocaine administration will both potentiate cocaine-induced increases in locomotion (context-specific sensitization) and potentiate lever press responses to obtain cocaine (context-induced reinstatement). Electrophysiological studies have identified neurons in multiple brain structures, including the nucleus accumbens, medial prefrontal cortex and orbitofrontal cortex, that respond to drugs of abuse and their associated cues [56–58]. However, these representations are often sparse and distributed. This has rendered traditional inactivation and stimulation protocols ineffective in determining their causal role in drug-related behaviours.

A series of recent studies used the c-fos:lacZ transgenic rat to silence the distributed neurons that respond to drugs of abuse and their associated cues. The c-fos:lacZ transgenic rat contains a transgene in which the c-fos promoter regulates expression of the bacterial LacZ gene. LacZ encodes the protein β-galactosidase, which converts an infusible compound Daun02 to daunorubicin, a product that diminishes calcium-dependent action potentials (figure 1b). Infusion of Daun02 into the nucleus accumbens of the c-fos:lacZ transgenic mouse following cocaine-context pairing will therefore silence the neurons activated by the conditioning paradigm [59].

Koya et al. exploited the Daun02 inactivation method to explore the neural mechanisms mediating context-specific sensitization [59]. Electrophysiological studies previously identified a sparse and distributed subset of neurons in the nucleus accumbens that respond to cocaine in a context previously paired with cocaine; these neurons do not respond to cocaine in an unpaired context [60]. Koya et al. determined the necessity of these neurons for context-specific sensitization [59]. Animals were repeatedly injected with cocaine in context A to form a cocaine–context association. Seven days later, animals were exposed to context A and injected with cocaine or saline before Daun02 was infused into the nucleus accumbens to silence the neurons activated by the cocaine-conditioning paradigm. Three days later, locomotor activity was assessed in response to cocaine injection. Silencing neurons that responded during the cocaine conditioning paradigm largely abolished the increase in locomotion evoked by the paired context. This was not observed in animals that received saline prior to Daun02 infusion nor in animals that were exposed to a distinct context B and injected with cocaine before Daun02 infusion. These data suggest that the activity of neurons that respond to the training context and neurons that respond to the reinforcing drug is necessary for context-specific sensitization. Thus, a critical component of the cocaine–context association is encoded in this distributed ensemble in the nucleus accumbens [59].

In addition to modulating behavioural responses to drugs of abuse, contextual cues can also potentiate drug seeking. For example, exposure to a drug-associated contextual cue will reinstate a previously extinguished drug-seeking action (context-induced reinstatement). Studies employing IEG immunostaining have identified neurons in the amygdala, nucleus accumbens and prefrontal cortex that respond to drug-associated contexts [61,62]. Using the Daun02 inactivation method in c-fos:lacZ transgenic rats, recent studies demonstrated that the activity of a context representation in the nucleus accumbens shell is necessary for context-induced reinstatement of cocaine-seeking [63], while the activity of context representations in the orbitofrontal and ventral medial prefrontal cortex is required for the context-induced reinstatement of heroin-seeking [64,65]. Thus context representations in the nucleus accumbens, orbitofrontal and ventral medial prefrontal cortex play critical roles in modulating drug-seeking behaviours. Elucidating how this circuitry interacts to mediate drug-related behaviours might ultimately contribute to an understanding of addiction that can guide future therapeutic strategies.

5. Contextual fear conditioning

Experimental strategies using IEG promoters to drive the expression of optogenetic and chemogenetic tools in physiologically classified populations of neurons have also been implemented to illuminate the neural circuits that mediate aversive contextual conditioning [66–73]. Lesions of the hippocampus impair the acquisition and expression of contextual fear learning [74–76]. In addition, electrophysiological and imaging studies, as well as reports examining the expression of IEGs, have identified distributed but specific ensembles of neurons activated by contextual fear learning in the dentate gyrus, CA1 and CA3 [69,77–81]. Liu et al. used a TetTag transgenic mouse to selectively label and photoactivate neurons incorporated into a contextual fear engram in the dentate gyrus [69]. This transgenic mouse expresses the tetracycline transactivator (tTA) under the control of the c-fos promoter. tTA is therefore expressed transiently in active neurons. The tTA protein will bind to the tetracycline response element (TRE) to drive the expression of a downstream target. However, the binding of tTA to TRE can be blocked by the administration of doxycycline. Thus, the absence of doxycycline delineates a window in which activated cells can be marked by a reporter protein (e.g. ChR2-EYFP, figure 1c) [69,82].

The TetTag transgenic mouse system has been used effectively to show that the activity of neurons activated by a contextual fear conditioning paradigm is sufficient for the expression of freezing behaviour to a learned aversive context [69]. Moreover, experiments using the Arc:CreER^T2 transgenic mouse demonstrated that the activity of these neurons is also necessary for the expression of learned contextual fear [67]. However, both of these experiments manipulated the activity of neurons in the dentate gyrus that responded when the contextual conditioned stimulus (CS) and the footshock unconditioned stimulus (US) were presented together. The neurons in the dentate gyrus responsible for eliciting freezing behaviour might therefore represent one or both of context and footshock. A subsequent study began to resolve this issue by expressing ChR2 specifically in an ensemble of cells in the dentate gyrus representing a novel context. They then asked whether this context representation could act as a CS by pairing photoactivation of this representation with footshock delivery in a distinct context. Subsequent exposure to the original context was sufficient to elicit freezing behaviour. In addition, subsequent photoactivation of the original context representation alone was sufficient to elicit freezing behaviour. This effect was not observed when the context representation in CA1 was photoactivated [70]. This suggests that through temporal pairing with footshock, the context representation in the dentate gyrus becomes connected to circuitry mediating defensive behavioural output such that both natural cues and optical stimulation can elicit appropriate responses. Moreover, photoactivation of the context representation in the dentate gyrus induced c-Fos expression in the basolateral and central nuclei of the amygdala. This raises the interesting possibility that contextual learning is mediated through changes in functional connectivity between hippocampal context representations and the amygdala. Support for this notion was provided by recent experiments revealing that a single hippocampal engram activates distinct amygdala representations to elicit behavioural responses of opposing valence [71].

Complementary experimental approaches have also identified neurons in dorsal CA1 that are activated by contextual conditioning. These CA1 neurons are necessary for the expression of conditioned fear [73]. In addition, these neurons are required for the activation of specific cortical and amygdala circuits that may mediate the expression of learned fear, such as the retrosplenial cortex. Photoactivation of neurons in the retrosplenial cortex that are activated by contextual fear conditioning elicits freezing behaviour and activates amygdala circuitry. Notably, although subsequent inactivation of the hippocampus abolishes the expression of conditioned fear to contextual cues, it has no effect on the freezing behaviour elicited by photoactivation of retrosplenial cortical neurons that were previously activated by contextual fear conditioning [66]. This suggests that the retrosplenial cortex may receive contextual information from the hippocampus that is sufficient to elicit behaviour. Taken together, these data suggest that the hippocampus to retrosplenial cortex pathway may play a role in connecting representations of context in the hippocampus to valence-specific circuits in the amygdala that mediate the expression of learned behaviours.

6. Cued conditioning

The BLA, like the hippocampus, lacks a discernable topography, with unconditioned and conditioned stimuli of differing valence and modality activating intermingled populations of neurons [43,46,83–93]. Many studies have yielded behavioural effects by manipulating neuronal populations in the BLA based upon their molecular identity or anatomical projections [94–100]. However, the populations of manipulated neurons may possess heterogeneous responses to various stimuli. The relationship between their physiological response properties and behaviour, therefore, remains to be established.

Gore et al. recently characterized representations of unconditioned stimuli in the BLA and assayed their role in innate and learned behaviours by manipulating the activity of physiologically defined subsets of neurons [46]. This study used a lentiviral vector encoding ChR2-EYFP and a nuclear localized mCherry under the control of the c-fos promoter (c-fos:ChR2-EYFP-2A-mCherry). This approach labels active neurons with ChR2-EYFP and mCherry, enabling the identification and manipulation of cells activated by a stimulus (figure 1d). The authors identified distinct neuronal ensembles that responded differentially to an appetitive US (nicotine) and an aversive US (footshock) (figure 2). Moreover, photoactivation of these ensembles elicited innate valence-specific physiological and behavioural responses (figure 3). These data suggest that representations of unconditioned stimuli in the BLA connect earlier-stage sensory representations of unconditioned stimuli to valence-specific innate responses.

Figure 2.

An appetitive and an aversive US activate distinct ensembles in the BLA. (a–c) Animals injected with lentivirus expressing ChR2-EYFP-2A-mCherry under the control of the c-fos promoter received two footshock treatments separated by 18 h and were stained for mCherry (a), c-Fos (b) and merged with DAPI (a nuclear counterstain) (c). (d–f) Animals injected with lentivirus expressing ChR2-EYFP-2A-mCherry under the control of the c-fos promoter received two intraperitoneal (i.p.) nicotine injections separated by 18 h and were stained for mCherry (d), c-Fos (e) and merged (f). (g–i) Animals injected with lentivirus expressing ChR2-EYFP-2A-mCherry under the control of the c-fos promoter received footshock followed by nicotine 18 h later and were stained for mCherry (g), c-Fos (h) and merged (i). (h–j) Animals injected with lentivirus expressing ChR2-EYFP-2A-mCherry under the control of the c-fos promoter received nicotine followed by footshock 18 h later and were stained for mCherry (j), c-Fos (k), and merged (l). (m) Quantification of per cent overlap ((c-Fos⁺ + mCherry⁺)/mCherry⁺) of cells expressing c-Fos and mCherry in the BLA (shock–shock 84.07 ± 4.46, n = 6; nicotine–nicotine 76.02 ± 4.90, n = 5; shock–nicotine 8.22 ± 1.40, n = 6; nicotine–shock 9.28 ± 2.94, n = 5. One-way ANOVA, F_3,18 = 130.43, *p < 0.0001). Reproduced with permission from [46].

Figure 3.

Photoactivation of footshock- or nicotine-responsive cells in the BLA elicits innate responses. (a) Per cent change in heart rate upon optical stimulation of footshock- or nicotine-responsive cells. We also injected animals with a virus expressing ChR2-EYFP-2A-mCherry under the control of the synapsin promoter to determine the effects of photoactivating a random ensemble of BLA neurons (shock −7.88 ± 1.93% n = 5; nicotine 9.49 ± 4.52% n = 5; synapsin −2.55 ± 3.40%, n = 6. Two-way ANOVA, group × optical stimulation interaction, F_2,26 = 6.24, p < 0.01, *p < 0.05). (b) Per cent change in respiration rate upon optical stimulation of footshock- or nicotine-responsive cells, or optical stimulation of a random ensemble (shock −9.20 ± 2.55%, n = 5; nicotine 10.63 ± 1.21%, n = 5; synapsin −2.19 ± 4.05%, n = 6. Two-way ANOVA, group × optical stimulation interaction, F_2,26 = 10.14, p < 0.001, *p < 0.01). (c) Per cent of time spent freezing during optical stimulation or during the intertrial interval (ITI) in animals expressing ChR2 or green fluorescent protein (GFP) in US-responsive neurons (shock ChR2 optical stimulation 26.05 ± 2.83%, ITI 8.99 ± 1.66% n = 7; nicotine ChR2 optical stimulation 6.93 ± 1.58%, ITI 6.94 ± 1.44 n = 6; shock GFP optical stimulation 7.70 ± 1.37%, ITI 9.32 ± 1.65, n = 6; nicotine GFP optical stimulation 11.07 ± 2.42%, ITI 9.02 ± 1.62%, n = 6; synapsin ChR2 optical stimulation 7.43 ± 2.19%, ITI 6.95 ± 1.46%, n = 5; no US optical stimulation 8.96 ± 1.42%, ITI 10.05 ± 1.40, n = 5. Two-way ANOVA, group × optical stimulation interaction, F_5,58 = 8.28, *p < 0.0001). Methods: animals were injected with lentivirus encoding c-fos:ChR2-EYFP-2A-mCherry in the BLA. Nine days later, animals were exposed to footshock or nicotine to generate an ensemble of cells expressing ChR2 in the BLA. The following day, the footshock- and nicotine-responsive ensembles were photoactivated and physiological and behavioural responses were recorded. Reproduced with permission from [46].

Unconditioned stimuli both elicit innate responses and reinforce learning. The BLA has been implicated extensively in aversive learning. Pharmacologic silencing and lesions of the BLA prevent the acquisition and expression of auditory, olfactory and visual fear conditioning [101–105]. Neurons in the BLA can respond to conditioned stimuli and unconditioned stimuli [83,84,89,106,107]. Moreover, responses to conditioned stimuli in the BLA are modulated by aversive learning [89,108–111]. These data suggest that after learning, US representations in the amygdala may connect sensory representations of conditioned stimuli to circuitry mediating defensive behavioural output. Gore et al. used the c-fos:ChR2-EYFP-2A-mCherry viral approach to determine the role of US representations in the BLA in learned behaviour [46]. Photoactivation of US-responsive cells was able to act as an US and reinforce valence-specific learning about otherwise neutral auditory or olfactory conditioned stimuli (figure 4). In addition, photoactivation of nicotine-responsive cells was able to reinforce instrumental behaviour. Notably, photoactivation of footshock-responsive cells did not suppress instrumental responding. This may be owing to the small number of nosepokes performed by animals expressing ChR2 in footshock-responsive cells, or it may indicate that conditioned punishment is mediated by circuitry independent of the BLA (but see [112]).

Figure 4.

Photoactivation of footshock- or nicotine-responsive cells in the BLA can reinforce olfactory conditioning. (a) Olfactory conditioning paradigm. (b) Approach–avoid index (difference between time spent in CS+ and CS− compartments, divided by the total time spent in both compartments) for animals trained to associate optical stimulation of the BLA with a CS+ odour (shock without odour 0.13 ± 0.12, with odour −0.47 ± 0.14, n = 6; untreated without odour 0.08 ± 0.03, with odour 0.02 ± 0.07, n = 6; nicotine without odour −0.09 ± 0.05, with odour 0.57 ± 0.13, n = 6; synapsin without odour 0.01 ± 0.12, with odour 0.00 ± 0.14, n = 5. Two-way ANOVA, group × conditioning interaction, F_3,38 = 12.65, p < 0.0005, *p < 0.05). Methods: animals were injected with lentivirus encoding c-fos:ChR2-EYFP-2A-mCherry in the BLA. Nine days later, animals were exposed to footshock or nicotine to generate an ensemble of cells expressing ChR2 in the BLA. The following day, animals received paired presentations of one CS odour that co-terminated with optical stimulation (CS+), and randomly interleaved presentations of a distinct CS odour (CS−). Animals were placed in a three-compartment chamber and the CS+ and CS− were infused from opposite ends of the apparatus. Time spent in each compartment was recorded. Reproduced with permission from [46].

After aversive learning, a CS can elicit appropriate defensive behaviour. As the activation of US-responsive neurons can generate a defensive behaviour (freezing, figure 3), a neural representation of a learned CS may elicit appropriate behaviour by activating US-responsive cells in the BLA. Gore et al. determined how conditioned stimuli generate appropriate learned defensive responses [46]. First, they asked whether CS representations within the BLA become connected to behavioural output as a result of learning. Photoactivation of a learned aversive CS representation in the BLA was able to elicit freezing behaviour. By contrast, photoactivation of a neutral CS representation in the BLA failed to elicit freezing behaviour. These data suggest that, after learning, a CS representation in the BLA becomes connected to circuitry that generates appropriate behavioural output. As learning increased the overlap of CS- and US-responsive neurons in the BLA [46], the authors questioned whether conditioned stimuli generate learned behavioural output through the activation of US-responsive cells in the BLA.

A lentiviral vector was generated that expressed the neural silencer halorhodopsin fused to EYFP (NpHR-EYFP) under the control of the c-fos promoter. This approach permits the optical silencing of cells activated by a stimulus. Photoinhibition of footshock-responsive cells prevented the expression of learned auditory or olfactory fear, whereas photoinhibition of nicotine-responsive cells had no effect on the expression of learned fear (figure 5). These data suggest that the expression of learned aversive behaviour is mediated through the convergence of CS representations onto US ensembles in the BLA. It remains to be ascertained whether learned appetitive behaviour is mediated via similar circuitry. Indeed, although the glutamatergic projection from BLA to nucleus accumbens has been implicated in appetitive learning [98,113], classical lesion studies suggest the BLA may play a more nuanced role in appetitive behaviour [114–117]. This notwithstanding, these data suggest that conditioned stimuli generate learned defensive behavioural output through the activation of US-responsive cells in the BLA. US representations in the amygdala therefore link earlier-stage representations of sensory stimuli to neural circuitry that generates appropriate innate and learned behavioural output.

Figure 5.

Photoinhibition of footshock- but not nicotine-responsive cells in the BLA prevents the expression of learned fear. (a) Protocol for silencing US-responsive cells during olfactory CS presentation. (b) Approach–avoid index in the presence and absence of photoinhibiton (shock NpHR with yellow light 0.16 ± 0.28, without yellow light −0.51 ± 0.15, n = 6; nicotine NpHR with yellow light −0.83 ± 0.07, without yellow light −0.69 ± 0.15, n = 6; shock GFP with yellow light −0.78 ± 0.13, without yellow light, −0.72 ± 0.17, n = 6; nicotine GFP with yellow light −0.76 ± 0.15, without yellow light −0.71 ± 0.22, n = 6. Two-way ANOVA, group F_3,40 = 5.36, p < 0.005, *p < 0.05). Methods: animals received paired presentations of a CS+ odour with footshock and received randomly interleaved presentations of a CS− odour. Mice were injected with a lentivirus encoding c-fos:NpHR-EYFP in the BLA and exposed to footshock or nicotine to induce an ensemble of cells in the BLA expressing NpHR-EYFP. Animals were placed in a three-compartment chamber and the CS+ and CS− were infused from opposite ends of the apparatus. Time spent in each compartment was recorded in the presence and absence of optical inhibition. Reproduced with permission from [46].

Neural representations of unconditioned stimuli in the BLA are central to the generation of innate and learned responses, but numerous questions remain to be resolved. Access to neural circuits based upon their physiological response properties may provide insight into how the connectivity and plasticity of these circuits facilitate adaptive behaviour [94,96–99,107,118–120]. Moreover, intersectional approaches that combine activity-dependent targeting of genetic tools with approaches that exploit anatomical or molecular neuronal features may help elucidate circuitry in greater detail. This might ultimately provide an understanding of the mechanisms underlying psychiatric diseases including addiction, post-traumatic stress disorder and anxiety.

7. Caveats and challenges

IEG promoters provide a means to drive the expression of optogenetic, chemogenetic and pharmacogenetic tools in physiologically classified neurons. However, these powerful technologies are not without limitations. For example, some brain regions may not be amenable to these approaches. The transgenic c-fos:lacZ and viral c-fos:ChR2-EYFP-2A-mCherry approaches are only suited to brain areas with low levels of spontaneous activity. The transgenic mice described above demonstrate significant variability in expression across brain areas [34]. Even if the brain area of interest is amenable to the technology, current strategies do not provide access to cells that are inhibited by stimuli, or that are activated in a manner than does not induce IEG expression. Moreover, the requirement to expose an animal to the stimulus of interest in order to induce labelling does not allow one to assess the nature of this circuitry in its naive state. It is therefore impossible to know whether all features of the circuitry existed prior to the inducing event. Alternatively, neural circuitry may acquire its properties by virtue of experience. Current strategies for manipulating neurons based on their physiological response properties may therefore never be able to distinguish between these possibilities. However, accessing neurons based on their physiological response properties might provide insight into how molecular markers map onto physiologically classified populations of neurons. These markers might then provide a means to monitor and manipulate these circuits in their naive state. The development of new transgenic approaches [121], synthetic activity-dependent promoters [122] and novel genetic switches [37] may remove some of these limitations, and thereby help to extend our understanding of the systems-level interactions that mediate behaviour.

We emphasize that even with the development of new molecular approaches, significant challenges will likely remain. Neurons in cortical and subcortical brain structures often exhibit complex physiological response properties. For example, a single neuron can encode different information at different time points. Indeed, neurons in the amygdala and orbitofrontal cortex often respond to multiple conditioned and unconditioned stimuli presented sequentially [83,84,89,123]. Moreover, neurons can exhibit selectivity that is modulated by other task parameters, such as the combined encoding of CS identity and reinforcement contingencies, which has been observed in amygdala and orbitofrontal cortex [88,89]. Neurons that encode information about multiple parameters exhibit ‘mixed selectivity’; mixed selectivity may arise from linear or nonlinear computations that combine information about multiple parameters [124–127]. The application of activity-dependent approaches for manipulating the activity of mixed-selectivity neurons that represent a specific set of parameters faces significant challenges. For example, a given cell may respond to stimulus A when it appears in context 1. Other cells may respond to stimulus A in any context, and yet other cells could respond to any stimulus so long as it appears in context 1. Activity-dependent expression of genetic tools will probably fail to isolate only neurons that respond to stimulus A in context 1, as other cell groups will also be activated when presenting stimulus A in context 1. In general, the prevalence of mixed-selectivity neurons in prefrontal cortex and other brain areas suggests that targeting the manipulation of activity to neurons dedicated to encoding a particular combination of parameters will be difficult. If neurons with these types of complex response properties lack obvious anatomical organization, connectivity, morphology or molecular phenotypes, techniques to test their causal role in complex behaviours may be limited.

Complex cognitive processes, such as abstraction, the construction of cognitive maps and the integration of conditional statements underlying rule-based learning, often require the integration of multiple parameters to generate neural representations of ‘internal’ or ‘abstract’ variables. For example, in the Wisconsin Card Sorting Task, subjects must deduce a rule in effect based on multiple parameters in order to sort cards correctly [128]. This type of rule corresponds to an abstract internal variable. Neurons in prefrontal cortex have been observed to encode rules [129–133], and mixed-selectivity neurons may play a role in rule representation. In particular, nonlinear mixed-selectivity neurons generate high-dimensional neural representations [127]. High-dimensional representations may provide a computationally efficient means of implementing complex cognitive processes that require the integration of information about multiple parameters [127]. Indeed, the dimensionality of a neural representation in prefrontal cortex has been correlated with performance on a demanding cognitive task requiring monkeys to track multiple types of information while applying a rule [127]. If a broad range of higher cognitive processes also relies on high-dimensional neural representations, then the difficulties of targeting neurons that exhibit nonlinear mixed selectivity may impose significant experimental hurdles in testing causal hypotheses regarding complex cognitive behaviour. Despite the triumphs of manipulating neural activity in physiologically classified neurons reviewed here, manipulations of neurons based upon response properties, even in combination with anatomical, morphological and molecular information about the identity of targeted neuronal elements, may prove insufficient for probing the causal role of neural activity in mediating complex cognitive operations.

Authors' contributions

F.G., E.C.S. and C.D.S. wrote the manuscript.

Competing interests

We have no competing interests.

Funding

C.D.S. was supported by NIMH (R01-MH082017) and the Simons Foundation (324162 to Stefano Fusi and C.D.S.).