Translational Models of Addiction Phenotypes to Advance Addiction Pharmacotherapy

Abstract

Alcohol and substance use disorders are heterogeneous conditions with limited effective treatment options. While there have been prior attempts to classify addiction subtypes they have not been translated to clinical practice. In an effort to better understand heterogeneity in psychiatric disorders, the National Institute for Mental Health Research Domain Criteria (RDoC) has challenged scientists to think beyond diagnostic symptoms and to consider the underlying features of psychopathology from a neuroscience-based framework. The field of addiction has grappled with this approach by considering several key constructs with the potential to capture RDoC domains. This critical review will focus on the efforts to apply translational models of addiction phenomenology in human clinical samples, including their relative strengths and weaknesses. Opportunities for forward and reverse translation are also discussed. Deep behavioral phenotyping using neuroscience-informed batteries show promise for better understanding of the clinical neuroscience of addiction and advancing precision medicine for alcohol and substance use disorders.

Graphical Abstract

Alcohol and substance use disorders are heterogeneous conditions with limited effective treatment options, and attempts to classify addiction subtypes have not been translated to clinical practice. We consider the underlying features of psychopathology from a neuroscience-based framework and focus on efforts to apply translational models of addiction. Behavioral phenotyping using neuroscience-informed batteries show promise to understand the clinical neuroscience of addiction and precision medicine.

INTRODUCTION

Substance use disorder (SUD) is a medical condition that is characterized by impaired control over substance use despite adverse consequences in psychosocial domains. In the Diagnostic and Statistical Manual of Mental Disorders (DSM-5), SUDs are diagnosed by a patients self-reported feelings and experiences and a clinicians’ understanding of psychiatric phenomenology or behavioral observations. These subjective approaches can lead to diagnostic inconsistencies across patients and clinicians. Additionally, diverse phenomena are captured by a SUD diagnosis. SUD requires any combination of at least two of the 11 diagnostic criteria. This means that studies on SUD (for example, genome-wide association studies) often involve patients with different clinical presentations of SUD. The phenomenological heterogeneity that characterizes the DSM means that a diagnostic category likely encompasses a large number of biologically distinct entities.

SUD heterogeneity has also impacted treatment development. In the case of alcohol use disorder (AUD), there are currently four FDA-approved medications to treat AUD: disulfiram, acamprosate, oral naltrexone, and injectable naltrexone. These medications have only modest efficacy, and given the heterogeneity of AUD, no one medication will work for every single patient suffering from this disorder. The heterogeneity of SUD and other DSM-5 disorders may be a key reason for the withdrawal of pharmaceutical companies from drug development for mental disorders. That is, if the mechanism of action of a given compound is relevant only to about half of the patients for a potential indication (that is, a given diagnostic category), it may be expected that this will result in many failed clinical trials and an inconsistent pattern of clinical trial outcomes.

There have been attempts over the last few decades to target interventions to specific SUD subgroups. One notable example is Project MATCH, which was one of the first large-scale alcohol studies to investigate whether behavioral treatments could be targeted using a personalized treatment approach. In that study, 10 primary matching variables (e.g., perceptions and involvement of alcohol use, cognitive impairment, gender, motivation, severity of psychiatric symptoms, etc.) were studied as moderators of three behavioral treatments—cognitive behavioral therapy, motivational enhancement therapy, and 12-step facilitation¹. Although the three behavioral treatments showed similar effectiveness, results generally failed to show that a particular therapy could be “matched” to particular patients to obtain better treatment outcome. The only effective match was between patients with low psychiatric severity and 12-step facilitation, which resulted in patients having more abstinent days than similar patients treated with cognitive-behavioral therapy. Clearly, some other variables must explain the heterogeneous nature of AUD. Nearly 25 years after the Project MATCH study, the science underlying personalized medicine is more sophisticated. In particular, progress is being made in both neurobiology and pharmacogenetics, which may facilitate the identification of biologically based AUD subtypes and the selection of treatments to target those subtypes^{2, 3}

Despite advances in neuroscience and genetic research during the past two decades, there are still no genetic or other biomarkers that reliably guide the diagnosis of SUD. This makes it difficult to link SUDs to a specific circuit or gene, and in this context, neuroscience and genetic findings have had a limited impact on DSM-5 diagnostic criteria. The fact that neuroscience and genetics have contributed little to the DSM-5 may be due to several reasons. For example, the DSM-5 is central to clinical practice, insurance reimbursement, determinations of disability and service eligibility, among others. Thus, stability in diagnostic categorization is vital for these purposes, and significant changes require substantial validation. Additionally, the DSM classification system has itself impeded progress in the areas of neuroscience relevant to AUD. If to obtain a grant or to publish a paper, one has to select a study population according to a system that is a poor reflection of nature, it is very difficult to advance our understanding of SUD.

In order to address the issue of SUD heterogeneity, there have been calls for the field of psychiatry to move towards a transdiagnostic and neuroscience-based framework to foster development of psychiatric nosology based on pathophysiology rather than clinical presentation. The Research Domain Criteria (RDoC) from the National Institute of Mental Health (NIMH) is one such initiative that is intended to advance the goal of a neuroscience-based research framework for psychiatric diseases⁴. RDoC has been influential in parsing heterogeneity and focusing on behavioral assays that are translatable and biologically informative.

While the application of RDoC to addiction phenotypes has lagged behind other mental health disorders, an Alcohol Addiction RDoC (AARDoC) was proposed as a research framework wherein specific functional domains can be prioritized⁵. The AARDoC posits the following major underlying domains of functioning in AUD: reward, stress and affect regulation, incentive salience, executive function, and social processes. While the framework was proposed in 2015 by the National Institute on Alcohol and Alcoholism (NIAAA), only a few AARDoC domains have received initial validation⁶.

As a compliment to these research frameworks, the Addictions Neuroclinical Assessment (ANA) was proposed by the NIAAA as a clinical framework for the assessment of addictions⁷. The ANA captures information in three of the five AARDoC domains. In its development, the ANA leveraged deep phenotyping with factor analytic methods to construct core neurofunctional domains. The ANA posits three neurofunctional domains that can be leveraged to understand heterogeneity in addiction, incentive salience, negative emotionality, and executive (dys)function^{7, 8}. These domains have been derived across independent alcohol-focused laboratories using a combination of clinical, behavioral, and self-report measures that assess the aforementioned underlying constructs^9–13. Our laboratory recently showed that the ANA can be applied to derive neurofunctional domain in individuals who use methamphetamine¹⁴, providing initial empirical support that the ANA can be applied across a range of substances.

It is possible that the ANA domains can be leveraged to advance precision medicine for AUD using pharmacological and/or behavioral treatments; however, consensus and support for a common battery are needed prior to initiating such experimental studies.

While neuroscience has yet to inform the diagnostic framework of addictive disorders, there are several neuroscience constructs that show promise in advancing pharmacotherapies for addiction (see Figure 1). In an effort to advance the translational neuroscience of addictive disorders, this critical review will focus on the efforts to apply translational models of addiction phenomenology in human clinical samples, including their relative strengths and weaknesses.

FIGURE 1.

Neuroscience-informed constructs applied to understand alcohol and substance use disorders. Neuroscience constructs reviewed include: allostasis, incentive salience, reward prediction error, reward and relief drinking phenotypes, and goal-directed, habit, and compulsive behaviors. In the area of addiction pharmacotherapy, there is a growing body of research that the reward and drinking phenotypes can inform responses to medications.

Neuroscience constructs applied to study alcohol and substance use disorders

Allostasis

The allostatic model of addiction seeks to explain the intricate balance between positive and negative reinforcement in addiction^15–17. This model is informed by the Opponent Process Theory, developed by Solomon and Corbit (1974) to explain how two opposing processes may occur simultaneously and jointly effect motivation¹⁸. In basic terms, it contends that over time, addiction becomes less about positive reinforcement (the activational process, or the a-process) and more about negative reinforcement (the counteradaptive opponent process, or the b-process)¹⁹. This theory seeks to capture the dynamic nature of addiction neurobiology as the brain is continuously adapting to large amounts of alcohol/drug use over extended periods of time, thereby causing a shift in the allostatic set point. Addiction allostasis is defined as the process of maintaining reward function stability through changes in brain reward mechanisms¹⁷. During the reinforcing effects of alcohol intoxication (a-process), there is an increase in GABAergic activity, opioid peptides, and dopamine output in the ventral striatum, which represent the neural substrates of alcohol reward. Conversely, during the counteradaptive opponent process marked by negative affect and withdrawal, there is an increase in corticotropin releasing factor (CRF) activity as well as a decrease in neuropeptide Y (NPY), both of which are key neuromodulators of stress reactivity^{17, 20, 21}. These processes provide the neural basis of reward and negative reinforcement associated with alcohol intoxication and withdrawal, respectively. Over time, the shift in the balance from positive to negative reinforcement is thought to explain what patients describe in their experiences with alcohol or drugs. In other words, patients often describe using drugs/alcohol to feel “normal”, which is consistent with the neuroadaptation in the brain reward circuitry leading to a chronic deviation of the brain reward set point proposed by the allostatic model¹⁵. In this context, it may be useful for clinicians and patients alike, to recognize that from a biological standpoint, chronic and heavy alcohol use causes individuals to drink primarily to alleviate withdrawal and its associated unpleasant affective and physical symptoms.

The neurobiological underpinnings of alcohol withdrawal include changes in the neurochemical systems within the extended amygdala including decreases in neurotransmitter functions subserving the acute reinforcing effects of alcohol (e.g., opioidergic, dopaminergic, GABAergic)¹⁶. An increase in alcohol self-administration can be reliably induced in animal models using a withdrawal state, and such models demonstrated that dopaminergic function is compromised during acute withdrawal²². Animal models have also emphasized the role of dysregulation in the brain stress system, including CRF-mediated processes, to changes in reward function leading to negative reinforcement^{17, 20, 21}. Clinical research has sought to characterize reward and relief processes in clinical samples. In a host of laboratory studies in which individuals receive a standard dose of alcohol, our group found that the positive reinforcing effects of alcohol (i.e., stimulation and positive mood) were associated with alcohol craving in heavy drinkers but not in individuals with alcohol dependence²³. These findings were extended in an independent sample, and it was demonstrated that the reinforcing effects of alcohol were salient determinants of subjective craving in lagged models of subjective response to alcohol and subsequent subjective craving across a range of light drinkers, heavy drinkers, and drinkers with AUD²⁴. In a study combining alcohol administration with alcohol self-administration using a progressive ratio model, we did not observe a relationship between AUD severity and subjective response. AUD severity was associated with greater baseline negative mood, sedation, and craving, but it did not moderate the relationship between subjective response and subsequent self-administration. This finding did not support our prediction based on the allostatic model that higher severity individuals would show self-administration of alcohol driven by negative reinforcement while lower severity individuals would self-administer for positive reinforcement²⁵. A longitudinal study including alcohol administration in the laboratory found support for the early stage phase of the allostatic model. Heavy drinkers exhibiting heightened reward sensitivity and stimulation in response to alcohol were more likely to progress to AUD²⁶. In brief, studies of the subjective rewarding effects of alcohol and negative affect relief have generally provided support for predictions from the allostatic model but only for the early stage of the model. The “dark side of addiction” proposed by the model has not been reliably supported, in large part because experimental studies rely heavily on healthier and younger individuals with more mild AUD presentations^{27, 28}. The disconnect between AUD severity in experimental psychopathology research samples and treatment research samples has been identified as a crucial opportunity for more efficient translation, whereby clinical samples ought to express the pathology severity necessary to fully access the target constructs described in neurobiological models²⁹.

Translation of the allostatic model to clinical samples with AUD have received increased attention. In the treatment domain, it has been argued that the clinical response to naltrexone, an opioid antagonist, may be stronger among individuals who report more positive reinforcement for alcohol (discussed in further detail below under reward/relief drinking)^{30, 31}. This line of inquiry overlapped with efforts to identify genetic markers of susceptibility for alcohol reward and for naltrexone responsivity in the lab^{32, 33} and in the clinic^34–36. While this line of research did not produce the reliable effects required to predict naltrexone response based on the reward-drinking genotype, recent studies using self-report measures of drinking motives have shown that the effect size for treatment with naltrexone is significantly higher among individuals who report drinking for positive reinforcement, compared to individuals who drink to alleviate negative feelings or for normalization^{37, 38}. Collectively, these findings suggest a potential convergence between neurobiologically-informed phenotypes and clinical phenotypes that can effectively improve treatment for AUD in a meaningful fashion. In medication development efforts, it has become increasingly accepted that some medications may work to attenuate the positive reinforcing effects of alcohol (e.g., naltrexone for blocking positive reinforcement) while others may be effective in targeting negative reinforcement mechanisms (e.g., gabapentin, an anti-convulsant GABA modulator, for alleviating protracted withdrawal symptoms and/or mifepristone, a glucocorticoid receptor antagonist, works on the stress system by regulating the amygdala). Alpha-1 blockers like prazosin and doxazosin may help to normalize these stress system changes seen in addictive disorders³⁹. This framework continues to garner support and inform precision medicine efforts.

Incentive Sensitization

Another prominent theory of addiction consists of the incentive sensitization model. The basic tenets of this theory argue that drugs of abuse share the ability to alter brain organization (i.e., produce neuroadaptation) in the brain reward systems rendering the system “sensitized” to drugs and associated stimuli (Robinson & Berridge, 2001). A key contribution of this theory is the dissociation between two aspects of incentive-sensitization, namely “liking” and “wanting.” Specifically, it has been nicely demonstrated that sensitization operates primarily at the subcomponent of reward termed incentive-salience, which is marked by “drug wanting.” While the neural basis of liking is primarily subserved by the endogenous opioid system, the process of wanting has been associated with dopaminergic activity in the brain’s reward circuitry⁴⁰. Sensitization is not simply an inevitable pharmacological consequence of repeated drug use but instead is modulated by environmental factors associated with alcohol and drug intake⁴¹.

The treatment implications of incentive salience are multiple. From a neurobiological standpoint, teaching patients to cope with triggers is akin to training one’s brain to unlearn associations or at a behavioral level, to inhibit a pre-potent (learned) response, such as alcohol use in the presence of a drinking buddy. While learning theory has been influential in the development of highly effective treatments for anxiety disorders, such as exposure-based interventions, similar success is not seen in the case of addiction. Cue exposure treatments for alcohol use disorder have produced mixed results⁴². The lack of strong empirical support for exposure-based treatments for addiction is largely explained by the over-generalizability of the conditioned response. To that end, it is simply not feasible to devise exposure exercises that effectively target all of such triggers. Nevertheless, functional analysis of behavior is commonly used to effectively identify patients’ most salient drug use triggers. Likewise, behavioral techniques for coping with triggers such as avoiding, taking time-outs, and learning refusal skills, represent important components of cognitive behavioral therapy for addiction. What is often lacking from this effective intervention, is the conceptualization of triggers as learned processes that are biologically-based and may evoke the unwanted, yet learned, behavioral response of alcohol or drug use leading to relapse.

Learning theory is particularly useful to understanding the neural underpinnings of incentive salience in addiction. It contends that adaptive responses to various types of functional alteration are displayed not only at the level of single neurons but also at the synapses between neurons, hence the term synaptic plasticity. This phenomenon has been most studied in the form of long-term potentiation (LTP), which is a process of long-lasting facilitation of neurotransmission across neurons when the synapses between them are used repeatedly under certain conditions. These processes are critical to all learning, both adaptive and maladaptive. Recognizing that much like the rewarding properties of alcohol and drugs that operate within the same neurocircuits responsible for normal reward functions in the brain, Pavlovian (or associative) learning during the development of addiction operates through the same signaling pathways subserving all other non-pathological forms of learning⁴³. In clinical terms, patients have learned to form associations between triggers and alcohol or drug consumption through the same biological mechanisms that allow us to associate our favorite restaurant with food. The difference is that that alcohol and drugs of abuse result in unnatural levels of dopamine that may be driving learning to an even greater degree than what would be expected by an unexpected reward (reward prediction errors are discussed in further detail below).

A number of neuroimaging studies have shown that the presentation of alcohol or drug cues, compared to control cues, reliably produces increases in blood flow in brain areas associated with reward (nucleus accumbens, ventral tegmental area [VTA], insula)⁴⁴ and affect regulation (amygdala)⁴⁵. Studies have summarized the neural circuitry reliably involved in cue-reactivity⁴⁶. It has been shown that alcohol cues elicit robust activation of limbic and prefrontal regions, including the ventral striatum, anterior cingulate cortex, and the ventromedial prefrontal cortex in individuals with AUD. Compared to controls, individuals with AUD show greater activation of parietal and temporal regions, including the posterior cingulate, precuneus, and superior temporal gyrus. Cue-elicited activation of the ventral striatum was most frequently correlated with behavioral measures and reduced by treatment^{46, 47}. Importantly, while brain activation is correlated with the subjective experience of craving, captured via self-reports, the correlation is far from perfect and in some cases, not present at all⁴⁸. This suggests that while craving is under conscious awareness, some of it may be subcortical in nature and perhaps inaccessible to patients. That is consistent with patient reports of being on “auto pilot” and having little awareness of their craving levels during a lapse. The use of functional magnetic resonance imaging(fMRI)-based cue reactivity holds promise for screening AUD/SUD treatments and provides a proof-of-mechanism for treatments thought to reduce the incentive salience of alcohol and drugs⁴⁹. Ibudilast, a neuroimmune modulator shows promise as a novel pharmacotherapy for AUD/SUD. A two-week human laboratory study of ibudilast to reduce heavy drinking found that ibudilast, relative to placebo, reduced the odds of heavy drinking across time by 45% and also attenuated alcohol cue-elicited activation in the ventral striatum compared to placebo. Individuals who had attenuated ventral striatal activation and took ibudilast had the fewest number of drinks per drinking day in the week following the scan⁵⁰. Importantly, medications development for addiction have focused on attenuation of craving and a recent meta-analysis has found that drug cues and craving indicators play significant roles in drug use and relapse outcomes in clinical settings⁵¹.

Reward/relief motivated behavior

One approach to classifying heterogeneity in SUD is based on the underlying motivation for substance use, namely primarily using substances for their rewarding effects or primarily using substances for relief. The processes underlying these phenotypes reflect the three-stage addiction cycle⁵². Reward use is hypothesized to be more salient in the binge-intoxication stage, the initial stage of substance use wherein use is motivated by a substance’s positive reinforcing effects. Reward substance use is linked to dopaminergic and opioidergic signaling in the ventral striatum. Relief substance use is more salient in the withdrawal, negative affect stage wherein individual’s use substances to relieve negative emotionality. Relief substance use has been linked to dysregulation in norepinephrine and stress-related signaling in the extended amygdala, as well as dysfunction in glutamate and GABA signaling more broadly. A neuroimaging study from our laboratory aimed to examine the neural correlates of reward/relief behaviors⁵³. Our results indicated that relief/habit drinkers (relief and habit groups were combined in this study due to overlap in clinical characteristics between these groups) showed greater dorsal striatal activation to visual alcohol cues than reward drinkers. However, cue-elicited ventral striatal activation did not differ significantly between groups. Further work identifying the neural correlates of reward and relief/habit drinking present a path towards the refinement of these neuroscience-informed phenotypes with the ultimate goal of informing personalized treatments for AUD.

Our laboratory and others have developed validated scales to identify reward and relief drinkers. Mann, Roos, Hoffmann, Nakovics, Lemenager, Heinz and Witkiewitz³⁷ and Roos, Mann and Witkiewitz⁵⁴ used complex data reduction approaches to identify reward and relief drinkers using items from the Inventory of Drinking Situations (IDS). More recently, Votaw, Mann, Kranzler, Roos, Nakovics and Witkiewitz⁵⁵ developed a 10-item version of the IDS, the Reward and Relief IDS, to identify 4 latent profiles: Low Reward/High Relief, Low Reward/Low Relief, High Reward/High Relief, and High Reward/Low Relief. Our laboratory recently developed the Reward, Relief, and Habit Drinking Scale (RRHDS), a short 4-item self-report measure that can be used to identify drinking phenotypes in the clinic⁵⁶.

There is some evidence that the reward/relief drinking phenotypes predict medication responses to naltrexone, an opioid antagonist, and acamprosate, a modulator of glutamatergic signaling. Several studies have found that primary reward drinkers respond better to naltrexone^{37, 38, 57} compared to placebo, with a notable exception⁵⁴. Roos, Mann and Witkiewitz⁵⁴ showed that primary relief drinkers respond better to acamprosate than placebo, while Mann, Roos, Hoffmann, Nakovics, Lemenager, Heinz and Witkiewitz³⁷ found no medication response to acamprosate in relief drinkers. Using the brief IDS to classify phenotypes, primary high reward drinks responded better to naltrexone and acamprosate versus placebo⁵⁵. To our knowledge, reward and relief phenotypes have not been applied to medication studies for other SUDs.

While identifying reward/relief phenotypes show promise for precision medicine, prospective studies are needed to confirm the reward phenotypes and naltrexone response directly. These studies are feasible given that the RRHDS and the brief IDS can be used to stratify randomization into treatment groups by reward/relief phenotypes. While there is some evidence for increased efficacy of naltrexone in reward users, suvorexant is a dual orexin antagonist that has garnered interest for addictive disorders⁵⁸. Dense orexin projections from the lateral hypothalamus to the VTA provide neurobiological support that orexins may influence responses to rewarding stimuli, including alcohol.⁵⁹ Two ongoing clinical trials will assess suvorexant’s potential as a treatment for AUD (NCT04229095) and comorbid AUD + insomnia (NCT03897062). Ghrelin receptor antagonists may also serve as a biological target for addictive disorders. In human laboratory studies, intravenous ghrelin administration has been shown to increase the urge to drink, increase alcohol self-administration, and modulate brain activity in regions involved in reward processing^{60, 61}. In regard to medications that may be beneficial for relief users, a seven-day human laboratory crossover trial of the neuroimmune modulator, ibudilast, showed that ibudilast improved mood during exposure to alcohol and stress cues and reduced the mood-altering and stimulant effects of alcohol among participants with more severe depressive symptoms⁶². Our laboratory has an ongoing Phase 2 randomized controlled trial (RCT) testing ibudilast for AUD (NCT03594435).

Reward prediction error

Learning through reinforcement is driven by a process through which an individual’s behavior changes based on a pairing of decisions/actions and consequences. This type of learning is thought to be driven by the discrepancy between the predicted, or expected, outcome and the received outcome, termed reward prediction error (RPE)^{63, 64}. RPEs drive behavior towards the obtainment of reward. If an individual receives the predicted reward, they keep the prediction unchanged and engage in the same behavior to obtain the reward; however, if the individual does not receive the predicted reward, they must update their prediction, resulting in a changed behavior to obtain the reward. There are two types of RPEs: positive and negative. Positive prediction errors occur when a reward is better than predicted, whereas negative prediction errors occur when the reward is worse than predicted. Substantial evidence indicates that RPEs are neurally driven by dopaminergic signaling in the midbrain. Activity in midbrain dopamine neurons increases during a positive RPE and decreases during a negative RPE.

As SUDs are often defined by compulsive behavior, or use which occurs despite the experience of negative consequences, there has been substantial translational interest in RPE signaling in the field⁶⁵. Moreover, addictive drugs have direct and indirect effects on dopaminergic neuronal activity, resulting in an increase in dopamine firing which may drive learning through an RPE-like mechanism⁶⁵. A number of substances, including cocaine, amphetamine, alcohol, opioids, and nicotine, all acutely increase dopamine transients in the nucleus accumbens^{66, 67}, providing support for the theory that RPE-like signaling may drive addictive behavior.

Human studies of RPE have largely relied on fMRI paradigms to probe the neural underpinnings of the RPE signal and to examine the relationship between this signal and decision-making behavior. While fMRI task specifics differ between studies, the methodology relies on dynamic changes in response-outcome contingencies to induce RPEs⁶⁸. In a simple version of a RPE task, a participant would be presented with two abstract stimuli. The participant would select between the stimuli and receive positive or negative feedback or reward, typically monetary reward. Initially one stimulus is associated with more frequent experiences of a positive reward (i.e., 80% chance of reward), while the other stimulus is paired with a low likelihood of receiving the positive reward (i.e., 20% chance). After a participant learns these reward contingencies, the contingencies would be unpredictably switched, such that the more frequently rewarded stimulus would become the less frequently rewarded stimulus, resulting in a negative RPE. A number of studies have used these paradigms to investigate RPE in individuals with an SUD. One of the first studies in this area found that chronic nicotine smokers produced a robust RPE signal in the caudate, a region of the dorsal striatum; however, their behavior was not guided by this neural signal⁶⁹. Moreover, when smokers were given nicotine, neural signals guided behavior; thereby indicating that during acute abstinence and withdrawal, neural prediction error signals are intact yet are unable to influence decision-making, mirroring use despite consequences. Similarly, men with AUD displayed an intact striatal RPE signal in the brain; however, these participants did show altered functional connectivity between the striatum and the dorsolateral prefrontal cortex which was associated with impairments in learning⁷⁰. Another study in individuals with AUD used real alcohol rewards, rather than monetary rewards, and found that individuals with AUD had increased positive RPE activity in parietal and occipital regions relative to social drinkers⁷¹. A recent study in adolescents found that AUD symptomology was negatively correlated with optimal decision-making and striatal and prefrontal cortex RPE signal⁷², indicating that these errors in decision making are present in those who may not yet have a severe presentation of the disorder. Findings have been mixed in individuals with a cocaine use disorder. Initial studies found reduced RPE signal in the striatum and orbitofrontal cortex⁷³, and a reduced EEG signal during RPEs⁷⁴. However, another study found that during cocaine deprivation, individuals with a cocaine use disorder, had heightened neural RPE activity in the striatum, which mediated the relationship between chronicity of substance use and the desire to use cocaine⁷⁴. Overall, a recent meta-analysis of 28 studies of RPE studies in substance using populations found that while substance users robustly activate the striatum and insula during RPEs, they show blunted activation in the putamen, inferior frontal gyrus, and right insula, relative to controls⁷⁵. Differences in findings between and across substance-using populations may also be attributed to the satiation state of the population, (i.e., acute abstinence, active use, acute use⁷⁶) and by belief or expectation state^{77, 78}.

While understanding the neural mechanisms underlying RPEs and decision-making in SUDs holds promise for furthering treatment and medication development, few studies have directly used RPE tasks or models to evaluate pharmacological or psychosocial treatments. Instead, results from these studies provide targets for the development of learning-based addiction treatments. One medication that may hold promise to treat decision-making deficits is modafinil. Modafinil is a mild stimulant and cognitive enhancer which acts on several neurotransmitter systems including dopamine. Modafinil has been investigated as a pharmacotherapy for AUD, methamphetamine use disorder, and cocaine use disorder. In individuals with AUD, acute treatment with modafinil improved impulsive decision-making, enhanced frontostriatal connectivity⁷⁹, and increased response inhibition for a subset of compromised participants⁸⁰. In individuals with a methamphetamine use disorder, acute treatment with modafinil improved performance in a learning task and increased activation in frontal brain regions⁸¹. However, the benefits of chronic treatment with modafinil for individuals with AUD and methamphetamine use disorder remain unknown. In individuals with cocaine use disorder, acute or short-term treatment of modafinil reduced activation in the VTA, attenuated self-reported craving⁸², and improved working memory⁸³. However, a RCT of modafinil to treat cocaine use disorder failed to find improvements in clinical or neuropsychological outcomes and did not improve abstinence rates⁸⁴, suggesting that modafinil may not be a useful treatment for cocaine use disorder. Frontal-striatal and insula-striatal pathways have been consistently found to be disrupted in RPE studies^{70, 73, 75}; therefore, targeting these circuits may hold promise to treat SUDs.

Goal-directed choice, habit, and compulsivity

Over the last few decades, three prominent theories of addiction have characterized addiction as either excessive goal-directed choice behaviors, maladaptive habitual behaviors, or compulsive behaviors. Goal-directed behavior is characterized by actions selected based on their resulting consequences⁸⁵. In the context of addiction, the frequency of goal-directed drug taking is determined by the expectation of reward value combined with knowledge regarding the voluntary behaviors necessary to obtain the drug^{86, 87}. While the initial stages of addiction are characterized by goal-directed drug taking behaviors, habit theory suggests that the reinforcing effects of the substance strengthens the association between drug-related stimuli and the drug-seeking response⁸⁸. That is, habit theory suggests that addiction is driven by the strength of the established stimulus-response relationship insofar as the magnitude of the reinforcer is steady. When the reinforcer is re-experienced having a lower or higher value, the stimulus-response relationship and the frequency of drug-taking behavior adjusts accordingly. As such, habitual drug taking behavior is flexible and amenable to change, specifically when the value of the reinforcer shifts. In contrast to habit theory, compulsion theory posits that the stimulus-response relationship controlling substance use cannot be modified by re-experiencing the reinforcer at a different value^{89, 90}. In other words, because compulsive drug taking behavior is driven primarily by the established stimulus-response relationship, drug taking behavior is not impacted by a loss of value in the reinforcer.

Goal-directed decision-making requires the knowledge of the association between actions and their consequences, and are “model-based”, as they rely on a model of the world to determine the best course of action⁹¹. Conversely, habitual decision-making is more rigid and uses past reward associations to make current decisions; referred to “model-free” as it does not require an explicit model of the world for decision making⁹². Several tasks have been developed to investigate goal-directed and habitual decision-making. One of the most prominent and translationally valid tasks is the two-step decision making task developed by Daw and colleagues⁹³. In this sequential two-stage task, participants choose between one of two choices in each stage, leading to a rewarded or non-rewarded outcome of varying probabilities. In the first stage, choices are associated with a transition to a likely (70%) or unlikely (30%) state, where participants make their final selection and receive feedback. In this task, model free, or habitual, decision-making would result in the repetition of a previously rewarded action, even if the transition was unlikely; whereas model-based, or goal-directed, decision-making considers the likelihood of the transition for future decision making. This task has been used to investigate goal-directed and habitual decision-making in rodent models of addiction⁹⁴, individuals with AUD^{95, 96}, and at-risk individuals^{97, 98}.

The outcome-devaluation task is also commonly used in laboratory animals and humans to differentiate between goal-directed versus habit behaviors⁹⁹. In the typical research setting, participants learn that two responses result in two different rewarding outcomes. Of these outcomes is then devalued by allowing participants to consume the outcome to satiety or giving the participant instructions that one of the outcomes is no longer available. After devaluation, the participants are given a choice between the two responses under conditions when the outcomes are not available (i.e., extinction). If responding for the devalued outcome decreases, it suggests that the response is goal-directed and dependent on the current value of the outcome. If responding for the devalued outcome remains steady, this suggests that the response is habitual and driven by the stimulus-response relationship. While the task has been used to provide support for habit theory among individuals who use alcohol¹⁰⁰, cocaine¹⁰¹, and nicotine¹⁰², several studies among clinical samples failed to support habit theory (See Hogarth⁹⁹ for a thorough critique of the habit literature).

To study compulsive behavior, researchers examine actual or hypothetical drug-taking behaviors after punishment or increasing costs. For example, demand tasks examine hypothetical consumption behaviors at increasing costs⁸⁷. The intensity of demand (maximum consumption at low price) is considered to be a relatively pure index of drug value unaffected by costs. Breakpoint represents the price at which drug consumption drops to zero and is thought to be more sensitive to the impact of price costs on the decision to consume. Compulsion theory would be supported if substance use severity was more strongly associated with breakpoint than intensity, suggesting cost insensitivity is more important than drug value¹⁰³. However, meta-analyses and systematic reviews of this literature have found that indices of dependence severity correlate more consistently with measures of intensity than breakpoint, suggesting that dependence is more likely to be driven by greater drug value than cost discounting^104–106. Deficits in reversal learning have been interpreted as evidence for greater cost discounting in addiction. In reversal learning tasks, participants learn that one response has a higher payoff than an alternative choice, before these response−reward contingencies are reversed. Individuals with SUD show deficits in reversal learning despite comparable acquisition of the initial contingencies^107–109. One interpretation is that drug users are less sensitive to the punishment of the incorrect choice, driving persistence of this choice. Taken together, these studies provide minimal support for compulsion theory.

These behavioral tasks have not been widely utilized in the context of treatments for addiction. However, behavioral and pharmacotherapies that reduce drug value may be effective at reducing goal-directed and habitual behaviors (assuming that re-experiencing the drug at a lower value will adjust the stimulus-response relationship). Effective pharmacotherapies that have shown to decrease drug/alcohol demand include naltrexone¹¹⁰ and varenicline¹¹¹. Once “proof of concept” studies have established the validity of these paradigms as predictors of medication response, the tasks can be leveraged in medication development for addiction.

CONCLUSION

Substance and alcohol use disorders are heterogenous conditions, as currently defined in DSM-5 diagnostic criteria. Because the clinical presentation of alcohol and SUDs can vary widely from person to person, effective treatment options are limited, and a one size fits all treatment approach is not feasible. While there have been numerous attempts to classify meaningful subtypes in addiction; they have not translated to clinical practice. Thus, there is a critical need to for an efficient and clinically-useful way to better understand the heterogeneity in alcohol and SUDs and to leverage that knowledge towards targeted interventions. Recent efforts focused on deep behavioral phenotyping using NIAAA’s Addictions Neuroclinical Assessment and/or the National Institute of Drug Abuse’s Phenotypic Battery represent recent examples of neuroscience-informed constructs aimed at elucidating the clinical neuroscience of the disorders. The neuroscience-based constructs identified using deep phenotyping batteries complement core constructs in NIMH’s RDoC and NIAAA’s Alcohol Addiction RDoC and have the potential to extend these research frameworks to the clinic. Once these domains have been independently validated across laboratories, their predictive utility in the human laboratory and clinical trials should be put to the test. These efforts would embed neuroscience-informed phenotyping in medication development and allow for biomarker development as well. There is a critical need to apply the phenomenology to improve treatments and to develop treatment-responsive biomarkers using the underlying biology of core phenotypes (e.g., neural cue-reactivity, magnetic resonance spectroscopy assays of neurometabolites, peripheral inflammation, stress, pharmacogenetics). This is an important direction in the implementation of neuroscience-informed phenotyping. Moreover, there have been recent calls to amplify the neuroscience-informed assessment at the level of clinical outcomes¹¹².

This review addressed several neuroscience-related constructs that are proposed to underly addiction; however, this is not intended to be an exhaustive list, and it is reasonable to assume some phenotypic overlap between constructs. Our research group has tested several of the neuroscience constructs using experimental psychopathology approaches¹¹³. There are several opportunities for forward and reverse translation of neuroscience constructs that are worth mentioning. In the allostatic model of addiction, recruiting more severe samples of patients with AUD may reconcile the differences between the preclinical and human literature. Additionally, reverse translating reward and relief drinking phenotypes in animal models will improve our understanding of the neurodysfunction in addiction. The latter point would be especially useful given the prominent clinical research that reward and relief phenotypes differ in their responses to pharmacological treatments. Additional examples of translational opportunities are provided elsewhere¹¹⁴. The central recommendations from the literature reviewed herein include the following: (a) implementing neuroscience-informed phenotyping as predictors and as outcomes of addiction treatment in contemporary (as compared to post-hoc) samples; (b) testing the clinical utility of the neuroscientific constructs; (c) leveraging clinical neuroscience constructs towards biomarker development; (d) refining biomarkers and behavioral predictors with potential for large scale dissemination and implementation in clinical care settings; and (e) maintaining a healthy level of skepticism regarding the “added benefit” of neuroscience-informed phenotypes and biomarkers beyond what is currently accessible to clinicians.

In summary, neuroscience constructs hold promise for advancing precision medicine for addiction. The field is experiencing a foray into their potential for informing treatment response, with most of the research focused on the reward/relief drinking phenotypes. The deep behavioral phenotyping approaches serve as a starting point to understand heterogeneity in addiction. Once the batteries are fine-tuned and validated, it is purported that they can be leveraged to develop novel interventions that target specific neuroscience-informed mechanisms that underlie addiction. Nevertheless, the “proof is in the pudding.” As such, we contend that a thoughtful translational neuroscience approach should consistently aim for larger effect sizes for addiction treatments, which would result in higher uptake of evidence-based practices and overall population-level improvements in healthcare outcomes for addiction.