Using miniscopes and deep learning to compare neurobehavioral representations of psychostimulant and opioid self-administration

Abstract

Concurrent abuse of psychostimulants and opioids represents a growing public health concern. However, preclinical models of substance use disorder often situate psychostimulants and opioids similarly within a unified self-administration procedure. This approach may fail to capture important differences in neurobehavioral activity related to each drug type. A large volume of in vivo literature suggests that, within canonical reward-related brain regions such as the nucleus accumbens (NAc), multiple reward-seeking behaviors may be represented by distinct neural populations. This comparison is often made between drug and natural rewards, however there is evidence for a similar distinction between psychostimulants and opioids. Here, we review the evidence for distinct neurobehavioral characteristics of psychostimulant versus opioid self-administration and consider the utility of two experimental approaches—miniscope calcium imaging and deep learning-assisted behavior analysis—in further exploring this topic.

1.1. Introduction

Abuse of both psychostimulants and opioids represents a significant public health concern [1–3]. These drugs are often abused concurrently, and polysubstance abuse scenarios are of significant clinical interest [4]. Drug self-administration procedures are useful preclinical models of addiction [5]. Animals are trained to perform an operant response—e.g., pressing a lever—to earn an intravenous infusion of drug [6]. A large proportion of self-administration studies have used either psychostimulant (e.g., cocaine, methamphetamine) or opioid (e.g., heroin, morphine) reinforcers (e.g., [6–10]). Some have used a “speedball” (cocaine and heroin mix) reinforcer (e.g., [11,12]), but relatively few have specifically compared psychostimulant and opioid self-administration in a within-subjects design [13–16]. Thus, concurrent abuse of psychostimulants and opioids has received less study than single-drug use.

Despite pronounced differences in the subjective effects of psychostimulants and opioids [17], single-drug self-administration studies contribute to converging literature on drug addiction in general, due to evidence that overlapping circuitry, specifically within the nucleus accumbens (NAc) mediates the reinforcing effects of both drug classes [18] and is necessary for self-administration [19]. NAc has been shown to be similarly involved in operant responding for non-drug rewards such as food, water, and juice [20,21].

At a single-cell level, however, multiple reward-seeking responses are heterogeneously represented within NAc. This functional heterogeneity has largely been characterized in comparisons of drug versus natural rewards [20–24], due to translational interest in examining drug vs. non-drug choice [25,26]. However, similar results have been found in comparisons of cocaine and heroin [16,27,28], suggesting that despite similar overarching involvement of NAc, psychostimulant and opioid self-administration are likely represented by distinct ensembles. This perspective aligns with pharmacological evidence that psychostimulants and opioids produce different means to similar ends in terms of elevating NAc dopamine (DA) [29,30], and neuroadaptations such as the formation of silent synapses [31], as well as behavioral evidence that common effects such as context-indued reinstatement involve overlapping, but not identical circuitry [17,32].

Comparison of both neural activity and behavior during psychostimulant and opioid self-administration is necessary to further characterize their biobehavioral bases. In vivo calcium imaging via miniature fluorescent microscopes (miniscopes) can expand on findings from electrophysiological studies to explore activity from specific cell types and brain circuits and can be combined with deep learning-assisted behavior analysis techniques to characterize the neural correlates of diffuse behaviors. Here, we review the electrophysiological evidence for functional heterogeneity within canonical NAc reward circuitry, as well as parallel evidence for distinct psychostimulant- and opioid-induced effects. We then describe recent advancements in miniscope imaging and deep learning analyses that stand to increase our understanding of the neurobehavioral characteristics of psychostimulant versus opioid self-administration.

2.1. Functional heterogeneity in the nucleus accumbens

2.2. Electrophysiological evidence

The NAc has been implicated in both drug and natural reward-seeking behavior [33]. Lesions of NAc disrupt self-administration of both psychostimulants and opioids [19,34,35] and increased extracellular DA in NAc is associated with operant responding for food [36] and drug [37] as well as with water consumption in thirsty rats [38]. These data have generated interest in NAc as a region in which canonical reward-related processes may become dysregulated and drive addictive drug-seeking [37].

Interest in NAc as a common substrate among reward-seeking behaviors has motivated many single-unit electrophysiology studies examining NAc neuron firing patterns during operant responding for drugs and/or natural rewards [24,33,39–43]. Early investigations supported the notion that responses for drug and natural reward were similarly represented in NAc neuronal activity. Carelli and Deadwyler [39] identified three distinct firing patterns present during both responding for cocaine and responding for water: increased firing immediately prior to the response and either increased or decreased firing immediately following the response, suggesting that cocaine- and water-seeking responses are largely represented by similar patterns of phasic activity in NAc neurons. However, they also observed a fourth pattern unique to cocaine SA, featuring pre-response and post-response activity peaks, suggesting the possibility of a distinct circuit involved in cocaine self-administration.

In a follow-up study, Carelli and colleagues [21] again performed single-unit recording during responding for cocaine or natural rewards, but did so from the same neurons in each session type. The authors observed the types of phasic activity reported by Carelli and Deadwyler [39]. However, of the neurons phasically active during responding for either cocaine or natural reward, fewer than 10% exhibited similar firing patterns in both conditions. In a comparison between natural rewards (food and water), however, 68% of phasically active neurons exhibited similar firing patterns under both conditions. This minimal overlap between neurons active during drug versus natural reward seeking has been reported elsewhere: Bobadilla and colleagues [44] used an activity tagging approach to label NAc core ensembles activated by cocaine seeking versus sucrose seeking and reported a 70% distinction between the two conditions. In the orbitofrontal cortex (OFC), Guillem and colleagues [22,23] observed non-overlapping ensembles during responding for heroin versus saccharin as well as cocaine versus saccharin.

Comparing the neural representation of responding for drug versus natural reward has received attention from researchers wishing to model the human choice between drug and natural reward-seeking behaviors [25,26]. Comparison between different classes of drugs has received less attention, due partly to the popular assumption that drug-seeking behavior in general is mediated by common circuitry despite vastly different subjective effects [17]. However, a study by Chang and colleagues [16] suggests that NAc neural representation of psychostimulant- and opioid-seeking responses may also be minimally overlapping. Researchers trained rats to self-administer both heroin and cocaine (either in separate sessions or in separate periods within single sessions) and recorded activity from NAc and medial prefrontal cortex neurons during self-administration of each drug. Again, a subset of NAc neurons exhibited phasic activity patterns around the time of the response, however, across all training conditions, no more than 25% of these neurons exhibited the same response during both cocaine and heroin self-administration. These results suggested that self-administration of cocaine and heroin is mediated by distinct populations within NAc, despite the assumption that drug self-administration in general is mediated by common circuitry.

2.3. Pharmacodynamics and non-canonical effects

The involvement of distinct NAc neuronal populations in cocaine versus heroin self-administration may be due in part to distinct pharmacological characteristics of each drug [16]. The majority of NAc neurons are GABAergic medium spiny neurons (MSNs) expressing either D1- or D2-family DA receptors [45]. It is well established that psychostimulants and opioids increase extracellular DA in NAc, albeit by different mechanisms. Opioids bind to mu-opioid receptors on GABAergic interneurons in the ventral tegmental area (VTA), disinhibiting principal neurons and inducing downstream DA release in NAc [30], whereas psychostimulants interfere with DA transporters (DAT) to increase DA in the mesolimbic pathway, particularly in NAc and VTA [46,47]. This mechanism further varies among psychostimulants: Cocaine blocks DA reuptake by binding to and inhibiting DAT [48] whereas amphetamines invert DA transport and facilitate DAT-mediated DA efflux [49].

Both psychostimulants and opioids also exert so-called “non-canonical” (i.e., DA-independent) effects on NAc neural activity that may contribute to differential representation in NAc. One such non-canonical effect concerns the formation of silent synapses. Silent synapses are glutamatergic synapses that represent a mechanism of activity-induced plasticity and following their formation, can be either stabilized with additional AMPA receptors or removed via synapse elimination [50]. The mechanism by which silent synapses are generated following repeated drug exposures differs between cocaine and morphine. Graziane and colleagues [51] observed the formation of silent synapses in NAc shell MSNs following five days of either cocaine or morphine injection. However, synapse formation following cocaine involved the insertion of new NMDA receptors, whereas morphine-induced silent synapses were formed from existing synapses via AMPA receptor internalization. Furthermore, the formation of these cocaine- and morphine-induced silent synapses was primarily observed in D1- and D2-expressing MSNs, respectively. After 21 days of withdrawal, cocaine-induced silent synapses were stabilized via insertion of additional AMPA receptors, whereas morphine-induced silent synapses were eliminated. Indeed, another study found that during morphine withdrawal, NAc shell MSNs exhibited markedly decreased spine density for up to 14 days before returning to pre-treatment levels [52]. The longer-term status of unsilenced cocaine-induced synapses in D1-expressing MSNs is unclear, however others have reported that in projections from paraventricular thalamus to NAc, these synapses may not return to basal levels for up to 45 days after cocaine intake [53].

Although these reports do not involve self-administration, they highlight a mechanism by which psychostimulants and opioids may exert distinct effects on NAc neural activity, especially with repeated exposure. In particular, the separate effects of cocaine and morphine on D1- and D2-expressing NAc MSNs may help explain Chang et al.’s [16] observation that cocaine and heroin self-administration involved partially overlapping populations of NAc neurons.

3.1. Behavioral differences between psychostimulant and opioid self-administration

Psychostimulants and opioids induce a range of distinct behavioral effects in rodents. Psychostimulants induce locomotion and rearing at low to medium doses and focused stereotypy at high doses [54–56]. Opioids induce increased running at low doses and hypoactivity, pica, stereotypy, and catatonia at high doses [57–59]. Early drug discrimination studies demonstrated that the interoceptive states produced by different drugs are sufficiently distinct to act as discriminative stimuli for multiple non-drug reward-seeking responses [60]. In other words, animals can distinguish between the effects of different drugs and can use that information to guide specific behaviors. Still, drug self-administration procedures are typically uniform across drug classes, which may fail to capture certain aspects of psychostimulant versus opioid self-administration in humans, particularly “instrumentalization” of self-administration to titrate effects relative to drug availability, dosage, tolerance, and other factors [61,62]. Indeed, altering certain parameters of the drug self-administration procedure can have different effects on the self-administration of different drugs. Below, we highlight two interesting studies in which psychostimulant self-administration and opioid self-administration behavior differed in subtle ways.

First, an experiment by Caprioli and colleagues [13] suggests that cocaine self-administration and heroin self-administration may be differently modulated by the experimental context. Two groups of rats were trained to self-administer cocaine and heroin on alternating days: a Resident group that lived in the operant chambers and a Nonresident group that was transported to the operant chambers from separate home cages each session. Both groups responded more for cocaine throughout training, although this effect was greater in the Nonresident group and in later sessions for both groups. Interestingly, when given the opportunity to choose between cocaine and heroin, Resident rats tended to prefer heroin and Nonresident rats tended to prefer cocaine. These results (see also [14,15], but see [25,26]) suggest that rats’ familiarity with the experimental context had different impacts on rats’ preference for cocaine versus heroin, and more generally imply that cocaine self-administration and heroin self-administration may be differently modulated by context.

Second, an experiment by D’Ottavio et al. [61] suggests that cocaine self-administration and heroin self-administration are differently modulated by the presence of a timeout in the drug self-administration procedure. Many procedures implement a timeout period—typically 20 s—immediately following each drug infusion to promote regular patterns of drug seeking and intake [63]. However, D’Ottavio et al. reported that removal of this timeout from a continuous access self-administration procedure altered patterns of heroin-seeking, promoting a burst-like pattern of self-administration that resulted in higher drug brain concentration relative to a timeout condition. The same was not the case for cocaine seeking. The authors note that the different impact of timeout removal on heroin and cocaine seeking is likely due to the different pharmacokinetics of each drug: the longer half-life of cocaine may produce continued intoxication across timeout periods, whereas the relatively brief (~90 s) half-life of heroin may render heroin self-administration more sensitive to brief periods of drug unavailability.

These data suggest that unified self-administration procedures may obscure subtle but important differences between cocaine and heroin self-administration. They highlight that the self-administration of different types of drugs is likely influenced by factors that are diffuse and difficult to directly observe, such as an animal’s familiarity with the context, and the degree of an animal’s intoxication when responding for additional drug. The extent to which these behavioral differences are related to heterogeneous representation in NAc is unclear. However, NAc neurons have been implicated in the contextual control of both cocaine [64,65] and heroin [66] and more broadly in the integration of hippocampal input into cue- and context-evoked reward seeking [67,68]. A potential explanation for Caprioli et al.’s results, then, might involve segregation of hippocampal inputs to distinct NAc neural populations depending on the nature of the contextual information being conveyed via the hippocampus. Further research targeting activity in this specific pathway is necessary to explore this possibility.

4.1. Miniscopes and deep learning

Decades of in vivo research have indicated that multiple reward-seeking behaviors—including those for psychostimulants and opioids—can be differently represented within canonical reward-related circuitry [16,17]. Additional research has revealed subtle behavioral differences between psychostimulant and opioid self-administration [13,61]. Comparison of psychostimulant and opioid self-administration thus remains of interest for researchers trying to understand how drug-seeking behaviors manifest neurobehaviorally, as well as for researchers trying to develop treatments for polysubstance abuse.

Evidence for heterogeneous representation of psychostimulant and opioid self-administration among individual NAc neurons [16] emphasizes the need for in vivo recording at single-cell resolution. The divergent non-canonical effects of psychostimulants and opioids on D1- and D2-expressing NAc MSNs [31] further suggest that this recording should be compatible with promotor-driven strategies for cell- and circuit-type specificity. Finally, closer examination of the behavioral differences between psychostimulant and opioid self-administration [13,61] requires an approach designed to identify subtle, longitudinal, non-time-locked patterns of behavior.

Single-unit electrophysiology offers high temporal resolution but lacks cell- and circuit-type specificity. Fiber photometry provides this specificity but lacks single-cell resolution. Activity-based markers and inactivation strategies (i.e., fos labelling, Daun02 inactivation [10,69]) offer high spatial resolution and functional specificity but are not compatible with repeated and longitudinal recording approaches. Single-cell calcium imaging offers high spatial resolution, cell- and circuit-type specificity, and modest temporal resolution, however certain calcium imaging approaches—particularly multi-photon approaches—require animals to be head-fixed (e.g., [70,71], but see [72,73]), eliminating a broad spectrum of relevant behaviors.

We suggest that single-photon calcium imaging via miniscopes combined with deep learning approaches to behavior analysis represents an optimal solution to the limitations described above [74–76]. Below, we outline this two-pronged approach to parsing the neurobehavioral signatures of psychostimulant and opioid self-administration in preclinical models.

4.2. Miniscope imaging

Miniscopes can use a variety of fluorescent biosensors to visualize neural activity at sub-second, single-neuron resolution. One common approach uses genetically encoded calcium indicators (GECIs) to record neural activity in vivo: A GECI (e.g., GCaMP) is delivered via viral vector to a brain region of interest and a gradient index (GRIN) lens is implanted in the same region [77], providing optical access to deep brain regions such as NAc, which would not be accessible via single- or multi-photon imaging through a cranial window [78]. The lightweight design of single-photon miniscopes allows neural recording in freely-moving animals with minimal interference to behavior [79].

Miniscopes are increasingly used in operant conditioning and drug self-administration procedures [80–85]. These applications have produced similar results to the electrophysiological studies reviewed above, specifically by identifying distinct representations of multiple reward-seeking responses. For example, Jennings and colleagues [86] identified distinct, but interacting, ensembles in OFC that responded to either social interaction or consumption of a liquid food reward. More recently, our lab [81] identified distinct but overlapping populations of prelimbic cortex neurons active during the performance of either a food-seeking or a cocaine-seeking response.

Miniscopes can offer additional insight into drug self-administration by recording from specific cell types and pathways. Schall and colleagues [87] used a Cre-dependent GECI to record activity from D1 and D2 receptor-expressing NAc MSNs in D1-Cre and D2-Cre mice. Cameron et al. [80] used a similarly promotor-driven approach to record activity in neurons projecting from infralimbic cortex (IL) to NAc; researchers injected a retrogradely transported Cre-recombinase into NAc and a Cre-dependent GCaMP in IL. Thus, miniscope imaging combines the cell type and pathway specificity of fiber photometry [88] with the single-cell resolution of electrophysiology. These advantages will prove useful to researchers trying to parse the contribution of different MSN subtypes in NAc or different mesocorticolimbic pathways involving NAc.

Finally, dual channel miniscopes have recently been developed [89], which will allow recording from two separate biosensors expressing spectrally distinct activity markers. In other words, dual-channel miniscopes allow for simultaneous recording of multiple targets relevant to drug self-administration such as DA receptor-expressing MSNs and extracellular DA in NAc [90], or NAc MSNs and projection terminals from areas such as IL or hippocampus. This dual-channel function can also be used to record activity in one channel and deliver an LED for optogenetic manipulation in the other.

Although miniscopes are increasingly being used in drug self-administration studies, this approach has historically been limited by optical cables connected to miniscopes and infusion lines connected to catheter ports becoming tangled. As a result, simultaneous neural recording and drug self-administration was impossible. Investigators compromised by only recording activity under extinction conditions, treating these sessions as assessments of “drug-seeking” albeit in the absence of drug [80–82]. This approach, however, failed to capture activity related to either ongoing drug self-administration or the actual effects of drugs. Recently, both commercially available and open-source motorized commutators have overcome this limitation [91,92]. For example, Chen et al. [83] were able to record activity from secondary motor cortex during self-administration of cocaine, as opposed to non-reinforced performance of a cocaine-seeking response. These advancements have recently allowed behavioral neuroscientists to leverage the technical advantages of single-cell calcium imaging during drug self-administration and have positioned miniscopes as a novel and viable option for characterizing neural activity during extended periods of psychostimulant and/or opioid intake.

4.3. Deep learning-assisted behavior analyses

Miniscope recording can be augmented by simultaneous synchronized video recording of animal behavior. These videos can be used to train pose estimation models via algorithms such as DeepLabCut [93] and SLEAP [94]. Researchers can then take a variety of approaches to translate that pose estimation data into meaningful information about behavior. At a basic level, pose estimation data can be used to label behaviors that would otherwise require manual scoring (e.g., freezing, rearing, locomotion) as well as task-relevant behaviors that cannot be reliably timestamped by experimental software (e.g., lever approach/investigation). Our lab recently used this approach to characterize neural activity related to 11 such behaviors in mice responding for food and cocaine in alternating sessions, and identified distinct but overlapping neural populations related to each reward type [81].

Alternatively, pose estimation data can be further analyzed with deep learning algorithms such as HUB-DT [95], Keypoint-MoSeq [96], or VAME [97]. The details of these methods have been reviewed elsewhere [75,76]. These approaches typically use pose estimation data—indicating an animal’s posture, orientation, velocity, and location in the operant chamber—to identify behavioral clusters or microstates, which can in turn be assembled into sequences of operant responding. Thus, an operant response can be examined as a continuous and fine-grained sequence, as opposed to a time-locked event such as a lever press.

In the context of drug self-administration, a primarily utility of deep learning-assisted approaches may be the ability to identify drug state-specific behaviors with little to no experimenter input. In a proof-of-concept application of HUB-DT, Lindsay and colleagues [95] injected rats with morphine and recorded their behavior in an open field environment. These recordings were processed with DeepLabCut then analyzed with HUB-DT. When comparing the behavior of these morphine-injected rats with controls, HUB-DT identified several morphine-specific behaviors without explicit input from experimenters. Whether this approach can similarly identify psychostimulant-specific behaviors remains to be seen. However, Lindsay et al.’s result suggests that deep learning-assisted algorithms may be useful to extract spontaneous drug-specific behaviors from large volumes of video data. There are two clear advantages to this approach. First, the need for subjective, error-prone, and labor-intensive behavioral scoring is largely reduced. This is especially useful for drug self-administration procedures, which may involve up to six hours of video recording per session. Second, these methods may be especially useful when integrated with pharmacokinetic/pharmacodynamic (PKPD) modeling software [98]. Various PKPD models combine established parameters regarding drug metabolism with data from self-administration procedures such as subject weight, drug dosage, infusion rate, and infusion time to continuously estimate brain and blood concentrations of the drug being self-administered. Deep learning-assisted approaches—especially those requiring minimal experimenter input (e.g., self-supervised [85] or unsupervised [96] models)—may prove useful in identifying behaviors associated with different periods of drug use, such as the initial “load-up” and later titration of drug level [99].

4.4. Potential applications and limitations

Together, miniscope imaging and deep learning-assisted behavior analyses present a novel method for characterizing cell- and circuit-type-specific neural activity related to diffuse behavioral signatures during ongoing self-administration of one or more drugs. Indeed, continuous recording of both neural and behavioral activity can be carried out in animals implanted with a double-lumen jugular catheter [13] allowing for simultaneous self-administration of psychostimulants and opioids and titration of each according to an individual animal’s preference. To our knowledge, no such experiment has yet been conducted. This approach may help parse neurobehavioral elements of psychostimulant versus opioid self-administration longitudinally, i.e., from initial drug use to prolonged drug use to abstinence and relapse. As noted above, parsing the contributions of D1- and D2-expressing MSNs in NAc to both drug-seeking and drug abstinence is of particular interest. Some recent studies suggest that D1- and D2-expressing MSNs may be involved in distinct temporal components of the reward-seeking sequence [87,100]. Others suggest that each MSN subtype plays a distinct role in honing operant responding by engaging task-relevant “focus” [101] or by suppressing suboptimal alternative actions [102]. In other words, such an approach may be capable of identifying subtle changes in how an animal initiates or experiences the self-administration of psychostimulants and/or opioids, how different cell types within NAc might work in concert to guide and refine this reward-seeking behavior, and whether there are patterns or trends within these data that may predict later propensity for relapse [40,41,103,104].

Although miniscope imaging and deep learning-assisted behavioral analyses offer additional insight into the neurobehavioral mechanisms of psychostimulant and opioid self-administration, they also pose certain limitations. First, calcium imaging offers poorer temporal resolution than electrophysiology. Unlike electrophysiology, which yields high temporal resolution of neural oscillations, calcium sensors themselves may slow calcium dynamics by acting as a buffer. Increase expression of a calcium sensor may in turn lead to more buffering, further altering the appearance of neural activity [105]. The trade-off of using GECIs with low affinity for calcium is a decrease in the ability to detect neural signals, although recent GFP-based GECIs feature improved signal-to-noise ratio and sensor kinetics which have nearly tenfold faster rise times than previous GCaMP iterations. Further, the use of calcium as a proxy for neural activity means that miniscopes do not record neural activity directly. This approach may fail to detect neurons with low activity, and may lead to discrepancies in population-level activity between electrophysiology and miniscope imaging [106]. On the other hand, it can be difficult to reliably distinguish certain types of neurons from others (e.g., DA from non-DA in VTA) via electrophysiological characteristics [107]. Finally, slow-firing neurons (e.g., GABAergic MSNs) may be challenging to detect via either recording method.

Second, caution is warranted when applying deep learning procedures to novel behavioral paradigms. Unsupervised models may be especially prone to overfitting; Kuo et al [76] note that such algorithms are best suited for relatively sustained behaviors separated by brief transitions. On the other hand, supervised approaches require subjective experimenter input and may not be as successful at identifying behavioral clusters that are not already visible to the experimenter. In either case, rigorous validation is recommended for any deep learning-assisted behavior analysis model.

5.1. Concluding remarks

Concurrent abuse of psychostimulants and opioids represents a growing public health concern [3,4]. Preclinical models of drug self-administration have extensively examined each drug type separately, but relatively few studies have compared them directly [14–17]. An important consideration regarding psychostimulant versus opioid self-administration concerns distinct representation within canonical reward-related regions such as NAc: decades of in vivo electrophysiology studies have reported functionally distinct NAc neuronal populations involved in different reward-seeking behaviors, including in cocaine versus heroin self-administration [16,21,39]. Further research is necessary to parse the contributions of different cell types, projection pathways, and other neuromodulatory factors, as well as to more thoroughly characterize behavioral differences between psychostimulant and opioid self-administration. In this mini review, we reviewed evidence for distinct neurobehavioral representation of psychostimulant versus opioid self-administration and discussed the merits of miniscope imaging and deep learning behavior analyses for researchers wishing to explore this comparison further. We suggest that this methodology may be the optimal approach to longitudinally characterizing diffuse behaviors and cell- or circuit-specific neural activity at single-cell resolution in freely moving animals self-administering one or more drugs.