An fMRI marker of drug and food craving

Abstract

Craving is a core characteristic of drug addiction and eating disorders. A new study identifies an fMRI-based neural signature of craving that is common to both food and drugs, predicts self-reported craving, distinguishes drug users from non-users, and tracks the efficacy of a cognitive therapy technique to reduce craving.

You are watching TV and a commercial for your favorite junk food comes on. Do you want what you see in the commercial? Do you crave it? How much? Experiences of cue-induced craving like this one are common. A similar process is also thought to underlie substance use and eating disorders, but of much greater potency. Here a particular context, like stress, a person, a neighborhood or any other reminders associated with eating or taking drugs (including drinking or smoking), may elicit an intense or overpowering craving for food or drugs, respectively. We use the same word (craving) to describe these related motivational states; but do they share similar mechanisms in the brain? If we can identify such a mechanism, it could greatly improve our understanding of, and treatments for, eating disorders and drug addiction. In this issue of Nature Neuroscience, Koban et al. identify an fMRI-based Neurobiological Craving Signature (NCS), a novel potential biomarker for these disorders, which suggests a shared neural substrate for craving across multiple drugs and food.

Does the experience and mechanism of craving depend on the specific object of craving, e.g., a certain drug or food? To answer this question, the authors aggregated data from several fMRI experiments designed to measure cue-induced craving, as well as its regulation with a cognitive exercise, across three different groups of drug users (of cigarettes, alcohol, or cocaine) and non-users. Participants viewed images of their drug of choice or palatable foods while imagining the short-term benefits of consuming the item, which elicits craving, or the long-term consequences of use/eating, which decreases craving. Actual craving was assessed by self-report after viewing each image, on a 1–5 scale. To identify a reliable neural pattern that predicts the reported craving levels, the authors applied a common machine learning algorithm, LASSO-PCR, to the fMRI responses to the images. This approach yielded a single pattern of fMRI activity, the NCS, trained to predict craving ratings across food, cigarettes, alcohol, and cocaine in out-of-sample subjects, suggesting these cravings share a common neural basis (Figure 1A).

Figure 1.

Creating and using the Neurobiological Craving Signature (NCS). A) In the scanner, participants were shown images of alcohol, cigarettes, cocaine, or palatable foods and asked to report their craving. A linear model was then trained to predict craving from fMRI responses, resulting in the NCS, a whole-brain pattern of BOLD activations. B) In new, held-out subjects, the NCS significantly predicted craving and identified whether the participant was a drug-user or a non-user.

There is a longstanding debate about the degree of overlap between different forms of craving, which has important implications for the treatment of disorders that feature craving¹. Because a single NCS model worked well across all image types used in this study, Koban et al. provide evidence for at least some common neural substrates across three different drug classes (including the two most commonly used) and food (a primary reinforcer). The authors provide further evidence for a shared neural mechanism by training a model solely to predict drug craving and showing that it also performed well at predicting food craving. Similarly, a model trained on food craving performed above chance at predicting drug craving. From a psychiatric perspective this cross-category prediction suggests that similar treatment approaches could be used when addressing craving in both substance use and eating disorders. Although other important drugs of abuse are yet to be similarly tested, e.g., opiates, these results also suggest that a unified intervention approach could be effectively adapted to regulate craving across different drug classes.

Many have advocated for the development of predictive models and biomarkers based on neuroimaging data as a framework for integrating neuroscience and psychiatry². This computational approach can enhance diagnosis specificity and sensitivity, predict treatment outcomes and guide treatment delivery, identify neural targets for intervention, and facilitate comparisons with animal models². The NCS takes a major step in this direction for drug addiction and eating disorders³. First, the neurophysiology of the NCS is easily interpreted as coordinated increases/decreases in neural activity across different regions of the brain. This pattern can be compared to animal studies of related constructs (e.g., of cue-induced incubation of craving⁴) or targeted as part of clinical interventions in humans (e.g., with brain stimulation or neurofeedback). In addition to providing an interpretable model, two further analyses by Koban et al. highlight the clinical potential and utility of the NCS. Beyond predicting craving (its trained target), the NCS also distinguished (out-of-sample) drug users and non-users based on responses to drug images (Figure 1B). This suggests that the NCS could be an objective marker of drug use and its severity, as remains to be further validated. A useful biomarker for addiction should also predict outcomes, guiding tailored treatments and tracking their efficacy. Koban et al. show that the simple exercise of thinking about future negative consequences, inspired by cognitive behavioral therapy and prior work by the authors⁵, reduced craving ratings in parallel with NCS reductions. This result suggests that the NCS can be used to track treatment efficacy, and perhaps identify, from the get-go, treatment responders. It will be exciting to test these potential uses of the NCS in new, larger and more variable samples (e.g., testing additional drugs of abuse or people with substance use and comorbid disorders).

Neuroimaging has increasingly contributed to the realization that cognitive/motivational states have distributed brain bases⁶, departing from seminal earlier neuropsychological work that mapped specific functions to individual brain regions⁷. The NCS represents an integrated attempt at identifying such a distributed representation of craving in drug addiction and eating disorders. Encouragingly the NCS features large weights in regions that have been identified through decades of neuroimaging research on drug cue-reactivity in people with substance use disorders, including the ventral striatum, insula and the vmPFC⁸. Intriguingly, the NCS also points to several brain regions not commonly implicated in craving, such as the cerebellum and lateral temporal and parietal areas, which have received less attention in the addiction literature. The authors provide additional evidence that NCS activations in such surprising regions are not explained by low level visual features, suggesting these regions are indeed related to craving. An interesting direction for future work will be to further explore the role of these regions and their relationships with the well-studied brain networks in drug addiction and eating disorders.

Craving, an important player in motivated behavior, was recently integrated in the diagnostic criteria of drug addiction⁹. However, the construct of craving is defined on the basis of self-report, which has a myriad of limitations including demand characteristics, the impact of sociodemographic and cultural factors, and the limits of introspection and self-awareness^10,11. Being trained on self-reported craving, it is difficult to disentangle the NCS from brain signals related to this complex mixture of processes underlying self-report in general. Further refinement of the NCS by filtering out or disambiguating contributions of self-report, could therefore yield improved mechanistic understanding and clinical utility for the NCS. One approach would be to compare the NCS with a model trained on self-report of other states that are related yet fundamentally distinct from craving (e.g., drug liking¹²), removing their shared components. Koban et al. do show that different fMRI patterns trained on various ratings of negative affect did not predict self-reported craving, suggesting that these patterns are at least not dominated by self-report of related signals. Another approach is to use psychophysiological and other objective behavioral measures (e.g., pupil dilation, event-related potentials, choice) as supplemental, or validated replacements, for self-reported craving^13,14. The NCS may also be further improved by measuring craving in response to more naturalistic audiovisual stimuli (movies) or other sensory modalities (touch, smell). Further, the predictive performance of the NCS should ultimately be tested quantitatively against other available clinical measures. Does the NCS yield improved clinical outcome predictions over simply using the self-reported measures themselves? Given the difficulty, cost and resources needed for obtaining MRI data relative to simple questionnaires, this is an important litmus test for any fMRI-based biomarker.

As an encouraging parting note, the work by Koban et al. represents a major development in our understanding of the neurophysiology of craving. They provide strong evidence for common neural mechanisms across different objects of craving, helping to resolve a longstanding debate in the field. It will be exciting to see future applications and refinements of the NCS as the field moves towards clinically reliable biomarkers for drug addiction and eating disorders.