See Supplementary Table S3 for PRISMA diagram of the systematic review flowchart.

To conduct a structured review of the role of WM in cognitive control associated with AN and SUD we used the following search inclusions and exclusions as below. Of note, we do not provide a systematic review of WM and cognitive control in healthy controls, given that since the introduction of the WM model by Baddeley and Hitch (1974), there has been an explosion of studies, theories and opinions using various terms synonymous with cognitive control, including cognitive inhibition, affect regulation, self-regulation, top-down control and cognitive–emotion interaction. As such, some contemporary theories of healthy WM are introduced, and thereafter studies are systematically reviewed that measure WM in AN and SUD in line with the opposing extremes of the impulse control model described in Figure 1.

Consulting PubMed, Medline, Science Direct, and manual searches of publication reference lists, we used the following search terms for inclusion in the systematic review of AN and SUD studies: anorexia AND working memory anorexia nervosa AND working memory; anorexia AND working memory AND cognitive control anorexia nervosa AND working memory AND cognitive control; substance use disorder AND working memory; substance use disorder AND working memory AND cognitive control. Of note, we only included AN and SUD (and not, for e.g., ‘eating disorders’ or ‘addiction’) to search for those specific populations we are considering at the extremes of an impulse control spectrum model. Our aim is to systematically review the role of WM in cognitive control in AN and SUD. Exclusion criteria were: systematic reviews/meta-analysis and theoretical/opinion/perspective articles (although some of these articles are referred to in our discussions of the empirical work); articles not written in English; articles measuring other executive functions but not WM; articles where AN/SUD was comorbid/secondary to major mental/neurological disorder (e.g., schizophrenia/cognitive decline associated with seropositive HIV/fetal alcohol syndrome/bipolar disorder); articles within the last decade – January 2010 – present: August 2017 (to include recent neuroscientific literature). As such, we found n = 15 studies that directly measured WM and cognitive control in AN, and n = 93 studies in SUD.

In order to contextualize the findings of the systematic review of studies examining WM and its role in cognitive control in AN and SUD, we first provide an overview of theories and some empirical studies in the healthy human brain.

In an attempt to highlight the mechanisms that may initiate impulse control, involving an interplay between bottom-up, non-conscious processes, and top-down, conscious cognitive processes in the healthy human brain, we turn to a neurocognitive description of the WM model by Baddeley and Hitch (1974) (Figure 2). In this model, a central executive (which is not in itself synonymous with consciousness or cognitive control, but rather an interplay between systems), residing within prefrontal cortex networks presides over what are termed slave systems. These slave systems, according to the model, are tripartite, interacting compartments known as the phonological loop (language), visuospatial scratchpad (visual semantics), and the episodic buffer (short term and episodic memory). These three WM subsystems and their neurobiological substrates will now be considered in more depth.

Firstly, the phonological loop, which may support cognitive ruminations, encompassing the articulatory loop and acoustic store, involves repetitive, conscious mental rehearsal strategies that promote the consolidation of beliefs (e.g., mantras), incorporating a dual network, namely speech production of Broca’s area in the frontal cortex (articulatory loop) and speech comprehension of Wernicke’s area in the temporal cortex (acoustic store). Secondly, interacting with the phonological loop are the dual visual semantic networks, namely the dorsal and ventral visual streams for action and perception (Goodale and Milner, 1992), or the “where” and the “what” pathway (Mishkin and Ungerleider, 1982), respectively. The dorsal “where” stream is primarily sensory, and follows a path from area V1 of the primary visual cortex (which can be activated by non-consciously processed stimuli, see Brooks et al., 2012c) to the parietal cortex, enabling a person to visualize the self-relevance of belief systems in time and space. Of note, it is traditionally the frontoparietal network that activates during WM functional magnetic resonance imaging (fMRI) studies (Rottschy et al., 2012). Conversely, the ventral “what” stream follows a path originating in V1 to the temporal auditory cortex (incorporating Wernicke’s area), with connections to the interoceptive (insular cortex) and episodic memory (hippocampal) regions, giving visual perception a comprehendible, concrete visual form in the mind. Thirdly, the episodic buffer can also be activated by non-conscious stimuli (Brooks et al., 2012c) and interacts with the phonological loop and visual semantic networks via the hippocampal-amygdala network residing close to Wernicke’s area in the medial temporal cortex. The hippocampal-amygdala network is integral to the mesolimbic reward/motivation pathway and interactions with prefrontal cortex, particularly in terms of building saliency and priority maps that influence overt and covert attentional systems (Zelinsky and Bisley, 2015). Poignantly, accumulating neurobiological evidence in animals and humans demonstrates direct prefrontal – hippocampal neural circuitry supporting the top-down modulation of bottom-up processes (Anderson et al., 2016). Furthermore, non-conscious processing of salient visual stimuli originating in V1 and the dopaminergic mesolimbic pathway, may occur up to half a second prior to conscious action tendencies (Libet, 1985), which may shed light on the sequence of activations of these sub-systems of WM in response to stimulation (Figure 3).

Drawing on the Global Workspace Theory (Baars et al., 2013) can help to understand how the central executive of the WM model might preside over these slave systems in order to orchestrate conscious cognitive control of distracting stimuli. The central executive could also be related to the executive control network (ECN) that functions antagonistically to the default mode network (DMN). The ECN is usually related to externally focused goal-related cognitions, whereas the DMN reflects introspection, day-dreaming, and general self-monitoring. In brief, the Global Workspace Theory posits that the updating of conscious perception of fluctuating internal (related to DMN) and external (related to ECN) stimulation (e.g., under conditions of uncertainty) emerges from backstage non-conscious processing. In other words, the conscious stage accommodates, or receives signals from transient actors (of various sensory modalities, e.g., auditory, tactile, and visual) that are guided by a) priority maps and context setting about the self, world, and others (e.g., Zelinsky and Bisley, 2015; Trask et al., 2016) that may over time become unconscious, and (b) by an unconscious audience (self-precepts, automatisms, memories, and language). Recent functional connectivity analyses (e.g., hub detection and network-based statistics) have provided support for the Global Workspace Theory, demonstrating how functional networks in the brain reorganize according to higher cognitive loads, particularly during the n-back WM task (Finc et al., 2017). In particular, it appears that network modularity decreases as cognitive load (on the n-back WM task) increases; specifically, far-reaching network hubs increase but local hubs decrease in conjunction with greater global connectivity with the DMN (Finc et al., 2017). These most recent findings are pertinent in the consideration of the role of WM in cognitive control, suggesting the wider recruitment of backstage, non-conscious neural processes with increasing cognitive demand.

Progressing the Global Workspace Theory, Bayesian Probabilistic Inference, or the Bayesian brain (Nilsson, 1986; Friston, 2012) follows three main principles in terms of how the central executive of the WM system, driven by unconscious backstage processes, may guide decisions under conditions of uncertainty. Firstly, activation of the neural networks within the prefrontal cortex (e.g., anterior cingulate and medial prefrontal) and hippocampus (which supports non-conscious, episodic, and salient memories) likely conform to the likelihood principle. In other words, calculating prediction error and inferring (with generative models) the evaluation of current experiences (interoceptive and exteroceptive) with prior experience (“have I seen it before?”). Secondly, neural systems within the prefrontal cortex (e.g., dorsomedial and orbitofrontal), insular cortex, and basal ganglia (e.g., ventral striatum, hippocampus, and amygdala) process the saliency of stimuli to derive frequentism or in other words, familiarity, by merging prior experience with current beliefs about the stimulus (“did I like it before?”). Finally, Bayesianism confers belief systems, perhaps via activation of the dorsolateral prefrontal cortex (DLPFC), visuospatial and language networks, with frequency of exposure to a certain event in order to update predictions and action tendencies the presence of uncertainty (“how should I respond again now?”).

With the processes of the Bayesian brain and Global Workspace Theory combined, there occurs an epistemic foraging for information (sampling of intrinsic and extrinsic stimuli) until the best prediction about the uncertain future based on prior beliefs while reducing error and free energy, is found (Friston et al., 2015). Relating epistemic foraging to excessive cognitive control, as seen for example in AN, could reflect heightened ruminations, local versus global thinking and attention to detail (Kothari et al., 2013) that support the updating of prior, rigid cognitive models concerning eating, weight, shape, and food (Dell’Osso et al., 2016). Conversely, in studies measuring the ‘jumping to conclusions (JTC)’ bias – indicative of less epistemic foraging and reduced cognitive load – people with AN do not show such a bias (Wittorf et al., 2012; McKenna et al., 2014). Whereas, those with JTC bias have reduced epistemic foraging, collecting less information to arrive at a decision, which is linked to WM deficits and delusional thinking (Garety et al., 2013). As such, people with SUD, who are at risk of delusional and psychotic disorder, usually behave in a habitual, model free manner and ‘jump to conclusions’ (Wittorf et al., 2012; Voon et al., 2015). Considering the healthy brain for now, the transient stage of WM and the conscious illusion of cognitive control, presiding within PFC circuits and interacting with lower order circuits, is consciously experienced as a deliberative, generative inference (particularly if prediction error/free energy is high) that shapes predictions about future events/goals (Schwartenbeck et al., 2015). This is particularly perceived if cognitive load is high, forcing the brain to epistemically forage for a conclusion that solidifies or updates a prediction. Our brains measure the degree to which our predictions about uncertain events are in error, taking more measurements that are held in mind by delay interneurons (see below) for accurate updating of our prior beliefs. This dynamic updating is experienced consciously and is solidified in episodic memory (which ultimately becomes unconscious) to build our internal (sub)optimal cognitive models.

Often studies of WM report insignificance in behavioral performance between healthy control groups and mental disorder populations, but when including brain imaging measures in a study, significant neural differences can often be found (Clark et al., 2017), and for review see below. This might suggest that the dynamic nature of the brain and neuroplasticity mechanisms support the development of compensatory processes that adapt to the unique experiences of the individual and which may, in turn, develop into entrenched, habitual mental disorder if not repeatedly challenged. Repetitive storage, encoding, and retrieval of processed stimuli (“have I seen it before, did I like it before, how should I respond again now?”) within the cortico-parietal-hippocampal circuitry underlies the detection of patterned activity and coincidental events (Basu and Siegelbaum, 2015). Such repetition underlies the dynamic mechanisms of Bayesianism as described above, and may be the process by which WM fosters neuroplasticity and alterations to the subjective experience of cognitive control. This innate neural learning mechanism could be harnessed to improve mental disorder (Lewis, 2015). As a starting point, basal ganglia processes may be at first non-consciously activated via the thalamus/brain stem, V1 and projections onto amygdala and striatal systems that register the saliency of a stimulus (“have I seen it before?”), particularly in terms of Pavlovian conditioning involving neocortex (Chau and Galvez, 2012; LeDoux, 2014). Such non-conscious saliency processes occur backstage, as readiness potentials – at least half a second before they decay (see discussions on meta-neuronal assemblies that register neuronal decay and the link to consciousness: Greenfield, 2017) and are registered by conscious systems in the higher order networks (Libet et al., 1979) (Figure 3). On this basis, subsequent “feeling” occurs (“did I like it before?”) through conscious cognitive processing of neural raw materials (LeDoux, 2014).

The interface between backstage processes and conscious on-stage processes, in other words cognitive–affective interaction, likely begins between the basal ganglia (e.g., amygdala, hippocampus, and striatum), anterior cingulate cortex (ACC) and via direct connectivity to the DLPFC for prediction error detection, reinforcement learning, and Bayesian updating about how to respond (Garrison et al., 2013). The ACC, which is involved in the prediction error detection network (Shen et al., 2015) contributes to the larger mesolimbic pathway, of which the saliency network of the amygdala-hippocampal-insula subsystem is a part and can be activated by non-conscious stimuli (Brooks et al., 2012c; Meneguzzo et al., 2014). Furthermore, it is intriguing to consider that non-conscious episodic memories via the slave systems (e.g., amygdala-hippocampal) can be triggered by stimuli that have been frequently encountered and are therefore salient and specific to the individual, underlying the qualitative nature of adaptive behavior (Desmedt et al., 2015).

The interference effects of non-conscious backstage salient processes on WM in healthy adult participants has been examined by manipulating the presentation of different types of subliminal images (appetitive, aversive, and neutral) while varying the levels of cognitive load (Uher et al., 2014). The rationale for the study was to probe how cognition and emotion interact, in terms of modulating affective arousal and cognitive engagement, since it is proposed by various groups that emotion and cognition compete for limited resources in the brain (Desimone and Duncan, 1995; Lavie, 2000; Pashler et al., 2001; Pessoa, 2009), particularly in the prefrontal-parietal cortex (Jaeggi et al., 2003; Rottschy et al., 2012). Against this background, in this study we showed a scrambled mosaic image on screen that functioned as a backward mask, commonly used in subliminal studies (Figure 4). The backward mask was presented directly after a 20-ms presentation of an aversive (e.g., bloody bodies), appetitive (e.g., high calorie food), or neutral (e.g., utensils) image from the International Affective Picture System (IAPS) (Lang et al., 1996) that were previously rated by independent volunteers in terms of pleasantness, aversion, salience, visual complexity, and recognizably. During the presentation of these subliminal stimuli to a group of healthy adult men and women (mean age: 25 years), we engaged participants in the easier 1-back and more difficult 2-back versions of the N-back WM task (Kirchner, 1958). In conjunction with our hypotheses, we found that when WM load was low, competitive interference between cognitive (e.g., completing the task) and affective processes (e.g., neural responses to subliminal arousing stimuli) on prefrontal cortex attentional systems was high (e.g., there was an increase in errors). However, increasing the cognitive load of the WM task appeared to attenuate the interfering effects of subliminal arousing stimuli (of both positive and negative valence).

Our data complemented findings by others showing that increasing the cognitive load can reduce the interference caused by consciously processed emotional distracters (Van Dillen and Koole, 2009; Van Dillen and Derks, 2012). Our findings were particularly interesting given that a post-experiment forced choice test confirmed that all participants were unaware of the subliminal interfering images during the study. From this we suggested that affective interference does not directly relate to conscious evaluation of the current cognitive activity (e.g., metacognition), and that a top-down regulatory mechanism within the prefrontal cortex appears to exist that is not reliant on conscious metacognition. Thus, it might be that hippocampal mechanisms in response to subliminal stimuli are activated first by the dopaminergic mesolimbic pathway via area V1, such that WM processes (visual semantics and language) are distracted if cognitive load is low (1-back). However, when cognitive load is high (2-back), the central executive might bias visual semantic and language systems in favor of prior goals (e.g., completing the task) such that hippocampal and salient environmental activation does not impinge on WM processes. For a schematic illustration, see Figure 4.

In line with our findings above (Uher et al., 2014), the central executive component of the WM model, particularly involving DLPFC-ACC-insula coupling (Fang et al., 2016) may regulate non-conscious stimulation of the hippocampus to reduce distraction when cognitive engagement is high (e.g., excessive epistemic foraging). Specifically, a rostral (bordering the evaluative OFC) versus caudal (bordering the error detecting ACC) involvement of the DLPFC may ensue according to low versus high cognitive load, respectively (Rottschy et al., 2012), and may be a robust model of cognitive control to consider for excessive cognitive control in AN. Similarly, this may support the notion of WM training to strengthen a person’s ability to evoke WM at a higher load/capacity in order to prevent distraction. However, interactions between lower order non-conscious slave systems and higher order ECNs are also mediated by genetic and epigenetic factors that underlie traits and differing thresholds of cognitive control and WM capacity that we consider next.

In terms of the genetic influences on WM processes, particularly concerning neuroplasticity, it is useful to consider expression within some of the main cellular signaling pathways that are implicated, including dopaminergic (D1–D5 receptors, DAT, and COMT), glutamatergic (BDNF and NMDA) and GABAergic (inhibitory interneuron) systems (Karlsgodt et al., 2011).

The current systematic review has found that to date 16 studies (published between 2010 and 2017) have examined WM performance in AN. In a previous review of all studies examining WM to date (the first study in 2002), it was found that 45% of studies showed patients with AN to have better WM performance than healthy subjects or those who binge; 18% reported worse performance and 37% reported no difference (Brooks, 2016). While the majority of studies in the previous review report superior performance, the heterogeneity of findings is likely due to the mixing of eating disorder subtypes, differing durations of illness and experiments that test different components of WM (e.g., verbal versus visuospatial). For example, varying degrees of restriction versus binge eating concomitant with switching between subtypes of ED over the course of illness can influence WM ability, with those with AN sometimes having better, but also worse WM ability (e.g., Israel et al., 2015; Weider et al., 2015). A longer duration of illness can also influence WM function (Dickson et al., 2008; Pruis et al., 2012; Lao-Kaim et al., 2014), but often WM ability does not correlate with clinical eating disorder measures (e.g., Seed et al., 2002; Fowler et al., 2006; Hatch et al., 2010; Nikendei et al., 2011), suggesting that clinical symptoms during chronic eating disorder are likely transient and secondary to the core cognitive disturbances in those who develop ED that often manifest during adolescence (Lena et al., 2004).

Since the last review in 2016, an additional six studies have been found that examined WM in AN (Phillipou et al., 2015; Biezonski et al., 2016; Giombini et al., 2016; Bentz et al., 2017; Ritschel et al., 2017; Solstrand Dahlberg et al., 2017). Of these six studies, none reported better WM performance in AN compared to HC; some reported normalization of WM function after treatment such as cognitive remediation therapy (CRT) and in-patient treatment (Castro-Fornieles et al., 2010; Giombini et al., 2016), whereas others show that abnormal function remains on recovery (Ritschel et al., 2017). This leaning toward the reporting of worse WM performance in AN may reflect the fact that there may be heightened WM capacity in AN but that this is utilized primarily for the exercising of detailed, rigid, complex cognitions about eating-related concerns. Furthermore, contemporary studies examining AN and WM now seem more often to combine neuropsychological measures of WM with neural measures using brain imaging techniques, which may shed better light on the mechanisms involved in cognitive control. For example, Biezonski et al. (2016) demonstrate thalamo-cortical structural and functional abnormalities in those with AN coincide with deficits on tasks probing WM and cognitive control. Similarly, increased temporoparietal activation in those who have recovered from AN, during a probabilistic learning task suggests a trait marker for excessive cognitive control and inefficiency to learn from unfolding new experience (Ritschel et al., 2017). These findings are suggestive of rigid, increasingly detailed prior beliefs (related to eating and body image) that are supported by excessive WM function that are difficult to alter (e.g., with cognitive therapies alone). Furthermore, larger insular cortex volumes in adolescent females with a new diagnosis of AN are related to significantly slower WM reaction time (Solstrand Dahlberg et al., 2017) and may be reflect proneness of hypersensitivity to distraction (Dickson et al., 2008). Taken together, the recent studies suggest a suboptimal WM performance in those with AN that coincides with hyperactivation of brain networks associated with excessive cognitive control and the perception of saliency (e.g., thalamo-cortical, temporal, parietal and insula). This may support a Bayesian brain notion that those with AN engage in excessive epistemic foraging of intrinsic and extrinsic stimuli to strengthen proof for generative models of future, uncertain goals (pertaining to concerns about eating, shape, weight, and food – “I will only eat a salad at noon”). Such excessive, chronic epistemic foraging may foster neural changes that support habitual, rigid, attention-to-detail style thinking that maintains AN, and that influencing cognitive changes at the neural level (e.g., with WM training) may improve treatment efficacy for AN. It is noteworthy that no studies have yet, to the authors’ knowledge, examined the (epi)genetic underpinnings of WM function and neuroplasticity in AN.

Other research into eating disorders supports the notion of Bayesian brain processes in the development of AN. For example, a diagnosis of restricting AN is linked to perceived self-control and delayed gratification in contrast to binge eaters, supporting a transdiagnostic, spectrum model of impulse/appetite control (Bartholdy et al., 2017). Furthermore, to support the spectrum view of impulse control in eating disorders, AN appears to be protective for the development of SUD (Kaye et al., 2013b). Alongside more prominent clinical symptoms (Treasure and Schmidt, 2013), WM is linked to core deficits, for example that pertain to excessive appetite control by way of repetitive cognitive ruminations. In support of this, a link has recently been demonstrated between AN, negative affect and disease-related ruminations (Seidel et al., 2016). This finding is in line with recent work on the Bayesian brain by Friston et al. (2015; Peters et al., 2017) suggesting that high levels of uncertainty in the environment regarding future events (e.g., high prediction error, free energy/entropy within brain processes) may lead to excessive epistemic foraging and stress. This can be related to negative affect/allostatic overload and to the findings of Seidel et al. (2016), in that if people with AN are excessively ruminating on future goals and attempts to control eating, body shape and weight, they place themselves in a cognitively uncertain situation. This is because, in a world where food and body images are abundant, there is uncertainty about whether appetite suppression and control of the body can indeed be achieved over the longer term. The tension between uncertainty versus excessive epistemic foraging to support prior beliefs would increase allostatic load and result in heightened stress/anxiety, which is commonly comorbid in those with AN. However, it may also be rewarding to experience achievement of appetite control goals in the face of uncertainty/abundance of food/eating/body image cues (O’Hara et al., 2015).

Similarly, it has been demonstrated that an early diagnosis of eating disorder during adolescence is linked to obsessive-compulsivity regarding concerns about shape, weight, and eating (Brooks et al., 2014b). Furthermore, the study also reported that higher obsessive-compulsive scores correlated with slower WM performance on the N-back task. Slower WM performance may be indicative of an interference effect, for example, if the brain is already engaged in ruminations about shape, weight, and eating (triggered non-consciously by images of food). That WM is distracted by affective stimuli in those with AN compared to healthy, and more so as duration of illness progresses, has been corroborated by Dickson et al. (2008). Similarly, another study found that WM performance in adult females with AN is only comprised by non-consciously processed images of food, not other affective subliminal images (Brooks et al., 2012a). Also, the interference effect observed in this study only occurred when neural systems were engaged in a WM task that utilized DLPFC-related processes, and not a conflict monitoring task that utilized ACC-related networks (Brooks et al., 2012a). Taken together, research suggests that WM function may be a feature of excessive rumination and appetitive control in those with AN, and that as the disorder progresses, visual and phonological episodic memories (e.g., images of food and body) become rigidly ingrained in the non-conscious saliency network due to their frequent and repetitive processing, which ultimately influences neuroplasticity processes and renders DLPFC-related neurocircuitry function suboptimal (e.g., biased toward detailed eating disorder cognitions).

Rumination and appetitive control in adolescents and adults at risk for developing AN may at first be deliberative and recreational (e.g., occasional dieting to achieve a visualized weight/body-image goal), but may switch to become ingrained, compulsive and habitual, akin to the neurobiological process of addiction (Everitt, 2014; O’Hara et al., 2015). For example, WM processes rely on the dopaminergic corticolimbic pathway (Goldman-Rakic, 1998), and as such, repetition of phonological and visual images may activate the saliency network, which over time may lead to epigenetic changes that upregulate dopamine receptors in line with altered dopamine levels in the mesolimbic pathway. In neural circuitry terms, deliberative responses appear to activate prefrontal cortex and ventral striatal (nucleus accumbens) networks linked to incentive salience and decisions about wanting and liking, whereas compulsive and habitual responses switch to dorsal striatum (Everitt, 2014; Koob and Volkow, 2016). Repetitive ruminations in those with AN (e.g., “I will only eat salad at noon”), from a WM model perspective, may engender the episodic representation of images evoked by deliberative prefrontal cortex predictive processes, such that internally generated images are eventually furnished with a saliency akin to a concrete object.

Additionally, AN is linked to hypoactivation in the insular cortex – a saliency hub – in response to food images, which is linked to reduced interoception, whereas by comparison the exercising of restraint cognitions to be thin appears more rewarding and addictive (e.g., “nothing tastes as good as skinny feels,” Kaye et al., 2013a). Epigenetic effects in dopaminergic receptor populations within the related basal ganglia network eventually leads to withdrawal, stress, and anxiety (Koob and Volkow, 2016) as more drug consumption – or in the case of AN, restraint cognitions of increasing complexity and detail – need to be exercised, as habituation ensues (Kaye et al., 2015). However, the degree to which WM is exercised according to restraint cognitions is likely transient, determined by current ED subtype and may be translated into neurobiological and epigenetic effects underlying the strength and efficiency of neural networks supporting WM capacity. Considering the potential neurobiological underpinnings of cognitive control and WM in AN, one study observed that while the global brain volume was smaller, the right DLPFC volume negatively correlated with age (between 18 and 45 years) in healthy adults but not in adults with a chronic diagnosis of restricting AN (Brooks et al., 2011a), suggesting that some difference in functionality might save the AN DLPFC from age-related normal atrophy. With this in mind, when weighting for BMI, the right DLPFC volume in RAN positively correlated with restraint cognitions (as measured by the eating disorder examination questionnaire) (Brooks et al., 2011a). While the study was a small pilot, it is intriguing to consider that excessive ruminations over a substantial time period (e.g., chronic eating disorder over approximately 30 years lifespan), measured cross-sectionally in individuals with differing durations of illness, may alter neural plasticity and therefore DLPFC volume.

In this vein, numerous recent studies have linked increased DLPFC function and related regions to WM and disease-related processes in those with AN. For example, a study revealed that increased neural activation during the perception of high calorie visual food stimuli occurs in the bilateral prefrontal cortex (including the DLPFC and ACC) in adolescents with an early diagnosis of eating disorder, as well as adults with chronic AN compared to patients with bulimia nervosa (Brooks et al., 2011c), and that increased DLPFC activation positively correlates with high obsessive-compulsive scores and slower WM performance (Brooks et al., 2014b). Other groups have also linked WM ability in those with restricting AN compared to healthy controls, to greater right DLPFC activation as cognitive load increases (Israel et al., 2015). Moreover, others have suggested that neural activation differences during WM performance in the DLPFC between patients with AN and controls may be distinguished by duration of illness (Lao-Kaim et al., 2014). Also, in a recent review of resting state functional MRI studies of AN, connectivity deficits in corticolimbic networks pertaining to cognitive control were implied (Gaudio et al., 2016). Similarly, a recent study by Boehm et al. (2016) demonstrates that passive (e.g., not engaged in a perception or cognitive task during the scan) functional resting state connectivity was reduced between the DLPFC and frontoparietal network in people who had recovered from AN, suggesting that aberrant WM and cognitive control is a trait marker for the risk of developing an ED. Similarly, increased DLPFC activation while perceiving and responding to images of food (e.g., “how disgusting do you find the images?”) in recovered patients who previously had chronic AN, may not be an indication of ‘good outcome’ (Uher et al., 2003), but rather, a reflection that underlying core cognitive symptoms of aberrant WM are recalcitrant. Taken together, these data suggest that the disruptive influence of non-conscious processes on WM and related attentional bias (Brooks et al., 2011b) increases as AN progresses, which in turn promotes epigenetic changes within the cognitive-affective neural interplay underlying the cognitive control of appetite.

Changes in DLPFC volume as measured by MRI for example, may translate into epigenetic effects and neuroplasticity within prefrontal cortex networks that support WM function. In line with this suggestion, increased volume in the DLPFC may reflect similar findings of increased BDNF SNP rs6265 methylation in ValVal subjects and improved WM (Ursini et al., 2016). Increased volume in the prefrontal cortex could also reflect epigenetic effects within the COMT promoter gene, which is associated with WM function and modulation of prefrontal cortex during WM in humans (Walton et al., 2014). Similarly, prefrontal cortex expression of enzymes required for GABA synthesis and the function of delay interneurons may also be associated with differential structure and function in the DLPFC (Labouesse et al., 2015). However, according to the authors’ knowledge, no studies have yet examined the genetic underpinnings of WM processes in AN and the link to excessive cognitive control.

Against the background of the neurobiology of WM mechanisms associated with AN, the genetic epidemiology of eating disorders has recently been reviewed (Bulik et al., 2016). Firstly, the review highlights that all eating disorders appear to be heritable conditions that are determined by genetic and environmental interaction, which is pertinent to consider here, given that repetitive ruminations about food restriction and environmental stimuli perceived to be relevant (e.g., one’s own shape, weight and eating in comparison to others) may alter cellular brain processes. In this vein, while Bulik et al. (2016) admit that as yet, genome-wide association studies (GWAS) for eating disorders are still underpowered, the top hit (co-morbid with bipolar disorder) for AN is on the SOX2-OT gene on chromosome 3 (Boraska et al., 2014; Liu et al., 2016). The SOX2-OT gene is associated with neurogenesis, and might underlie varying degrees of neuroplasticity and the extent to which neural circuits are rigidly set or can be modified; however, this suggestion is currently speculative. Secondly, the review underlines that a novel method of analysis, namely linkage disequilibrium score regression (LDSR) has demonstrated a strong positive genetic correlation between AN, schizophrenia and bipolar disorder, which suggests that the common variants cumulatively associated with schizophrenia and bipolar risk also increase risk for AN (Bulik-Sullivan et al., 2015) – implicating dopaminergic processes that are associated with WM. Given the strong links between WM and cortical plasticity in those with schizophrenia (Genevsky et al., 2010), the genetic linkage between schizophrenia and AN further implicates WM in pathology.

Another recent review has considered not only the genetic epidemiology of eating disorders but also specific epigenetic links (Yilmaz et al., 2015), which link neuroplasticity and WM processes. Firstly, while anxiety and related disorders (e.g., obsessive-compulsive disorder) are highly comorbid with eating disorders and as such genetic analyses of genes related to serotonergic systems have been extensively studied, Yilmaz et al. (2015) report mixed and underpowered findings. Similarly, while there is some indication that dopamine receptor studies are implicated in levels of binge eating and attention deficit disorder in those with eating disorders, Yilmaz et al. (2015) suggest that replication studies are needed. Furthermore, while leptin receptor and melanocortin genes are associated with weight regulation, Yilmaz et al. (2015) report no significant and consistent linkage to those with eating disorders.

Yilmaz et al. (2015) hint that it is perhaps BDNF that holds the most promise as an epigenetic candidate for eating disorders, given that BDNF is involved in appetite suppression by regulating melanocortin signaling in the hypothalamus (Xu et al., 2003), although no strong links are currently found. Yet, it is interesting to consider, given that the BDNF SNP rs6265, which creates or abolishes a CpG dinucleotide for methylation, is associated with modulation of prefrontal cortex activation in terms of improved WM and reduced methylation in ValMet subjects (Ursini et al., 2016). Thus, while differences in nucleotide sequence between AN and controls may not differ significantly, it is possible that gene expression or methylation patterns, perhaps in BDNF SNP rs6265, may significantly vary in those with eating disorders in relation to measures of WM capacity. Yilmaz et al. (2015) report that due to the link with reward processes, the dopaminergic system has been most extensively studied in terms of epigenetics in eating disorders, which is also pertinent given the role of dopamine in WM (Goldman-Rakic, 1998). Accordingly, there is some indication that those with AN may have increased DAT (SLC6A3; also referred to as DAT) mRNA expression due to hypermethylation of the gene’s promoter region, as well as DRD2 promoter hypermethylation (Frieling et al., 2010), although this is yet to be replicated. And at the time of writing, there has been no convincing links between COMT epigenetic effects and eating disorders. Taken together, given their influence on corticolimbic circuitry, it might be that BDNF and dopaminergic epigenetic mechanisms are most pertinent to the examination of fluctuating WM ability and the link to cognitive control of appetite in those with AN.

At the opposite extreme of the impulse control spectrum model (Figure 1) are those who have reduced impulse control (e.g., binge eaters/SUD, Volkow and Baler, 2015) as proposed by the dual process theory of addiction (Bechara, 2005). With SUD (including stimulant, nicotine, opioid, marijuana, and alcohol use) in mind, we conducted a second systematic review yielding n = 93 studies that have examined WM and cognitive control, between 2010 (e.g., the start of this decade) to the present (August 2017) (Supplementary Table S2). The majority of studies, n = 68 (72%) reported worse WM performance compared to healthy drug-naïve controls or non-drug taking control groups, but it is not clear whether WM deficits are a trait (e.g., cause) or a state (e.g., consequence) of SUD. In attempt to probe the trait versus state effects of SUD, researchers have examined whether WM deficits normalize following a period of abstinence, and in some cases, there is moderate reversal of cognitive deficits following abstinence (Vonmoos et al., 2014). Similarly, in a recent randomized control trial, Bell et al. (2017) reported that abstinence during 3 months of CRT and work therapy for outpatients with SUD was related to improved WM function also present at 6 months follow-up. Furthermore, 3 weeks of marijuana abstinence led to WM improvements in adolescents between the ages of 15–19 years, but attention deficits remained (Hanson et al., 2010). However, withdrawal from heavy substance use (alcohol and drugs) during adolescence may lead to long-lasting, WM deficits that are related to neurotoxicity in later life (Hanson et al., 2011). Nevertheless, across the board, the resounding conclusion from these studies is that WM in SUD is deficient, but that there is some indication that deficits may not be long-lasting and can be improved.

At the opposite end of an impulse-control spectrum, where AN represents excessive cognitive control, conversely those with addiction behaviors, and SUD in particular, represent low cognitive control and impaired self-regulation (Bechara, 2005; Bickel et al., 2011, 2015; Volkow and Baler, 2015; Brooks, 2016; Squeglia and Cservenka, 2017). Given that dopaminergic systems underlying WM (Goldman-Rakic, 1995, 1998) are implicated in addiction (Volkow and Baler, 2015), WM deficits are considered to contribute to core pathology of addiction (e.g., Bechara, 2005). In support of this (but also in support of trait deficits) adolescents with a family history of alcoholism have been shown to be slower on a verbal WM task than those with no family history of alcoholism (Cservenka et al., 2012). And generally, those with alcohol use disorder are usually impaired on the WM executive function domain (Chanraud et al., 2010; Bogg et al., 2012), particularly in a Bayesian brain sense when attempting to make decisions under uncertain conditions (Brevers et al., 2014). Verbal WM is associated with the phonological loop (articulatory loop and acoustic store; Broca and Wernicke’s area, respectively), and so deficits in this – as opposed to the visuospatial – domain may prevent the effective verbal rehearsal of top-down cognitive strategies and recall of future goals (that may underlie excessive ruminations in those with AN, for e.g.). Rehearsing cognitive strategies verbally in mind might be even more difficult for those with SUD in the presence of distractors (McClure and Bickel, 2014). In line with deficits in rehearsing verbal strategies in mind, in those with impulse control disorders, various studies have found that attention deficit symptoms (e.g., impulsivity, “behavioral under-control,” fidgeting) and dysfunctional affective regulation during childhood predict the later development of SUD (Block et al., 1988; Caspi et al., 1996; Màsse and Tremblay, 1997; Moffitt et al., 2011). Moreover, in children with a diagnosis of ADHD there is a higher risk for developing SUD later in life (Molina and Pelham, 2003).

A formal diagnosis of ADHD has been shown to be dissociable in core WM brain regions in adult ADHD patients with and without spatial WM deficits (Mattfeld et al., 2015), suggesting that ADHD symptoms are somewhat necessary but not sufficient to reduce WM ability. This could be due, in part, to the development of compensatory neural mechanisms during adulthood that reduce WM deficits, or that core WM deficits in those with ADHD at risk for SUD aggregate on neural networks underlying the verbal and not spatial WM domain. With this in mind, a recent meta-analysis of adults with SUD showed significant deficits in the verbal WM domain and not others (Baldacchino et al., 2012), which may also reflect evidence that there is a switch from controlled (associated with ventral striatum and prefrontal cortex) to habitual (associated with dorsal striatum and amygdala) drug taking as SUD progresses (Everitt, 2014). In other words, recreational and controlled drug use may be associated with the employment of verbal WM strategies that help to curtail impulsive behaviors, which may be lessened during the switch to habitual drug use and withdrawal. This switch could also hint at epigenetic effects, combined with the evidence that improvements to WM and prefrontal cortex activation have been observed in substance users who have achieved a period of abstinence (Schulte et al., 2014).

Further clues regarding the link between SUD and deficits in WM/cognitive control can be derived from the neuroimaging studies found in our systematic review (n = 31) employing electroencephalography (EEG), diffusion tensor imaging (DTI), single photon emission computed tomography (SPECT), functional and structural magnetic resonance imaging (MRI), including task – and resting state fMRI. Eighteen (58%) of these neuroimaging studies report normal WM performance but with altered, perhaps compensatory – or inefficient – neural processing (Chanraud et al., 2010, 2013; Crego et al., 2010; Jager et al., 2010; Schweinsburg et al., 2010; Sweet et al., 2010; Vollstädt-Klein et al., 2010; Bustamante et al., 2011; Sutherland et al., 2011; Bach et al., 2012; Bogg et al., 2012; Campanella et al., 2013; Charlet et al., 2014; Cousijn et al., 2014a,b; Ma et al., 2014; Loughead et al., 2015; Brooks et al., 2016). Neural and not behavioral differences during WM task performance may relate to the Global Workspace Theory and the recruitment of neuronal networks that activate the DMN (e.g., for self-reflection) when cognitive load is high (Finc et al., 2017). To support the notion that brain imaging data can provide additional insight into the processes of WM than behavioral data alone, a recent study has suggested that compensatory neural mechanisms are at play during variation in cognitive load (Clark et al., 2017).

The neuroimaging studies of WM in SUD have reported that, while WM function appears normal (e.g., no significant difference between cases and controls in behavioral data), there is reduced activation in the PFC network – for example in opiate maintenance patients (Bach et al., 2012). Similarly, reduced right parietal cortex activation in cocaine users (Bustamante et al., 2011), reduced ACC and medial PFC activation in adolescent and adult binge drinkers, respectively, (Crego et al., 2010; Bogg et al., 2012) and reduced functional connectivity in frontostriatal networks in cocaine dependent individuals (Ma et al., 2014). Conversely, other studies have reported increased activation when WM performance is not significantly different between cases and controls. For example, increased bilateral supplementary motor area and PFC/dorsal ACC regions occurs in binge drinkers (Vollstädt-Klein et al., 2010; Campanella et al., 2013; Charlet et al., 2014) and increased PFC activation is observed as cannabis use increases (Jager et al., 2010; Cousijn et al., 2014a). Furthermore, increased PFC activation is shown to remain over 3 years in cannabis users (Cousijn et al., 2014b; Goudriaan, 2014). Similarly, increased bilateral insula, medial superior prefrontal cortices, and right precentral gyrus activation has been observed in marijuana users (Schweinsburg et al., 2010). Increased PFC and posterior cingulate cortex activation during WM tasks may be predictive of relapse – at least for nicotine smokers (Loughead et al., 2015), and in chronic but not acute smokers normal WM performance is associated with increased activation of bilateral PFC, temporal, parietal, and insular cortices (Sweet et al., 2010; Sutherland et al., 2011). Interestingly, one study suggests that chronic alcoholics recruit cerebellar function resources to stimulate PFC networks during WM tasks (Chanraud et al., 2013). Finally, structural brain imaging studies when no significant WM deficit is observed, have reported increased basal ganglia volume using MRI after 4 weeks of cognitive treatment for methamphetamine use (Brooks et al., 2016) and increased frontocerebellar volume in those with alcohol use disorder (Chanraud et al., 2010). Taken together, the neuroimaging studies clearly implicate differential neural activation in brain regions that are typically associated with the WM network, including the frontostriatal, parietal, insula cortices, and cerebellar region. However, it is not yet clear how increased or decreased neural activation or volume in these regions differentiates WM function in those with SUD, with studies suggesting compensatory or inefficient mechanisms across networks to enable normal WM performance.

The remaining n = 11 (35%) brain imaging studies in our systematic review reported worse WM performance in SUD compared to controls and altered neural processes (Bava et al., 2010; Moeller et al., 2010; Smith et al., 2010; Nulsen et al., 2011; Cservenka et al., 2012; Marvel et al., 2012; Fitzpatrick and Crowe, 2013; Ozsoy et al., 2013; Falcone et al., 2014; Claus and Hendershot, 2015; Liang et al., 2016). In conjunction with worse WM performance, structural studies demonstrated reduced fractional anisotropy (FA) in PFC WM networks and higher FA in visual networks that may be linked to cognitive biases in alcohol and marijuana-using adolescents (Bava et al., 2010). Other structural studies showed alcoholic cerebellar degeneration was related to worse WM performance (Fitzpatrick and Crowe, 2013), and reduced hippocampal volumes in adolescents with alcohol use disorder (Ozsoy et al., 2013), which may alter the solidification of episodic memories within the WM network. In terms of brain function, decreased activation in the superior temporal gyrus and worse WM performance was observed in male alcohol users after an acute dose of alcohol (Claus and Hendershot, 2015). Decreased PFC activation and worse WM was observed in youth with a family history of alcoholism (Cservenka et al., 2012). Similarly, abstinent smokers performed slower on a WM task and had decreased DLPFC and ACC activation, but older age appeared to attenuate the effects (Falcone et al., 2014). Cocaine dependent individuals show reduced frontostriatal activation, and increased thalamus activation in relation to treatment response (Moeller et al., 2010). Ecstasy (MDMA) users during an EEG study also showed reduced electrophysiological indices during a WM task (Nulsen et al., 2011). Another study using SPECT measured another type of brain function – dopamine transportation (DAT), and found reduced DAT levels in the striatum of opioid dependent subjects that related to non-perseverative (e.g., omission) errors during a WM task (Liang et al., 2016). On the other hand, increased activation in inferior/superior cerebellum and amygdala in methadone maintenance opioid dependent participants has been observed (Marvel et al., 2012), and increased activation of the middle/superior frontal gyrus and right superior temporal gyrus in young marijuana users during worse WM performance (Smith et al., 2010). Taken together, worse WM performance appears to imply deficits (reduced/aberrant structure and function) in brain areas linked to the WM network, namely the PFC, hippocampus, temporal gyrus – and also the cerebellum.

Some of the neuroimaging studies in our systematic review included pharmacological challenge paradigms that go one step further in identifying the neural mechanisms of WM in SUD. Eleven pharmacological studies were found, including the use of tolcapone (a COMT inhibitor) (Ashare et al., 2013); methadone substitution therapy (an opioid derivative) (Bach et al., 2012; Henry et al., 2012; Marvel et al., 2012; Rapeli et al., 2012; Rass et al., 2015); buprenorphine substitution therapy (an opioid derivative) (Bach et al., 2012; Rapeli et al., 2012); modafinil (a GABA inhibitor that has stimulant properties) (Kalechstein et al., 2010, 2013; Dean et al., 2011; Joos et al., 2013), and rivastigmine (an acetylcholinesterase inhibitor) (Mahoney et al., 2014). The study using tolcapone did not provide strong evidence that reducing the function of COMT significantly alters WM function in abstinent nicotine smokers (Ashare et al., 2013), though that is not to say this approach would not work in other SUDs, e.g., stimulant users, in line with altered dopaminergic function. Studies using methadone or buprenorphine opioid substitution therapy showed no significant influence on WM (Bach et al., 2012; Henry et al., 2012; Marvel et al., 2012; Rapeli et al., 2012; Rass et al., 2015). Conversely, there is accumulating evidence that modafinil may significantly improve attention and WM in those with SUD (Kalechstein et al., 2010, 2013; Dean et al., 2011; Joos et al., 2013), and may be a useful pharmacological partner for other interventions aiming to improve WM, such as WM training. Finally, in the only preliminary study using rivastigmine with cocaine dependent participants to date, there was significant evidence that acetylcholinesterase inhibitors may improve WM deficits (Mahoney et al., 2014). Thus, while most pharmacological agents have not proven to be effective at improving WM deficits in SUD, modafinil and rivastigmine may be beneficial in conjunction with other psychological/cognitive interventions, particularly for those with severe cognitive deficits.

Perhaps one of the most prominent findings arising from the systematic review of WM in SUD is that the PFC – implicating dopaminergic function and the dual process model (Goldman-Rakic, 1995, 1998; Bechara, 2005) – is most altered (both increased and decreased activation). Furthermore, PFC activation may moderate the function of other brain regions implicated in our review, namely the cerebellum, insula, basal ganglia (including hippocampus, amygdala, striatum, ACC), thalamus, temporal gyrus, and parietal cortex, in line with contemporary neural models of addiction (Everitt, 2014). From the viewpoint of dopaminergic dysfunction that pertains to cognitive control deficits, an association implicating PFC volume differences has been reported (Genevsky et al., 2010). Furthermore, the link between psychosis and prolonged SUD, particularly in the use of stimulants is established and highlights the role of prefrontal dopaminergic system dysregulation (Bramness and Rognli, 2016). Moreover, effective treatment for psychosis, which is often observed in those with chronic SUD, targets PFC systems involving the DLPFC (Welch et al., 2011; Li et al., 2015; Kani et al., 2016). Similarly, repetitive transcranial magnetic stimulation of the right DLPFC leads to a significant reduction in craving for substances (Enokibara et al., 2016). Related to this, targeted cognitive training for schizophrenia using auditory (as opposed to visuospatial) WM tasks is most effective at reducing symptoms that might coincide with neuroplasticity changes (Biagianti et al., 2016). Thus, combined these findings suggest that PFC dopamine dysfunction in particular, might underlie variations in cognitive control, particularly with regard to auditory/verbal strategies for future goals and modulation of distracting stimuli that may shed light on the role of WM in SUD.

Various questions remain unanswered as to the neurobiological underpinnings of cognitive control and the link to WM in the healthy brain, and in AN and SUD, for a summary, see Figure 5.

Firstly, it is not yet clear whether unconscious processing is the first step in the sequence of cognitive control within the WM system, whether deliberative metacognitive processing is the key, or if it is a circular relationship between cognitive and affective processes, which are modified over time (Shea et al., 2014). As such, it is not clear how the WM system interacts with the mesolimbic circuitry to forge cognitive control of impulses, although it is likely related to storage, encoding and retrieval of future goals and prior beliefs that are stored in the episodic buffer via the hippocampus. Secondly, it is not yet clear what is the most effective way to measure unconscious processes in the brain, and so far, backward masking (Figure 4), event-related potentials with EEG (Figure 3) and subliminal tasks with fMRI (Brooks et al., 2012c) have been commonly used, but there are other novel experimental methods to consider as described below. Thirdly, given that there is some evidence that the ACC, part of the corticolimbic network, is activated to non-consciously processed stimuli (e.g., Brooks et al., 2012c), it is not yet clear if there are instances where the PFC is activated without conscious awareness/engagement. For example, compulsive repetitive restraint cognitions in those with AN may become unconsciously activated in the PFC, but these may be driven by alterations in bottom-up saliency network processes. Fourthly, it is not yet known whether psychiatric disorder, pertaining to deficits in WM and cognitive control, originates in excessive bottom-up saliency networks, or aberrant top-down conscious metacognitions, or whether both are important. Fifthly, while there is some evidence implicating the dorsal fronto-medial cortex in terms of a conscious veto (“free won’t”) of anticipated behavior (e.g., eating or substance use) (Kühn et al., 2009), which is linked to reduced EEG readiness potentials during voluntary omissions but not rule-based omissions (Misirlisoy and Haggard, 2014), the field remains relatively unexplored. Finally, it is not yet clear whether epigenetic processes can be harnessed to improve cognitive control and far transfer effects of WM training, such as long-term neuroplasticity, clinical improvements and better quality of life. Related to this, it is not known whether WM training can alter regulation of GABAergic and glutamatergic receptors that contribute to cognitive control and reduced delay discounting via PFC activation that may be associated with AN.

Attempts to answer these unyielding questions could consider novel non-invasive experimental approaches, such as, but not limited to, repetitive Transcranial Magnetic Stimulation (rTMS), real-time fMRI (rt-fMRI) and imaging epigenetics. Firstly, a recent meta-analysis demonstrates that rTMS, which is suggested to increase activation in the right DLPFC and has the potential to alter receptor regulation and neurotransmitter function significantly improves craving in those with SUD (Enokibara et al., 2016). Similarly, rTMS of the DLPFC has mixed findings in those with binge eating disorder (Van den Eynde et al., 2010; Gay et al., 2016) but is nevertheless linked to enhanced frontostriatal connectivity and reduced craving (Dunlop et al., 2015). However, rTMS of the DLPFC improves AN symptoms linked to more prudent decision-making (McClelland et al., 2016) and improves WM performance in general (Brunoni and Vanderhasselt, 2014). With these rTMS studies in mind, it is useful to consider that lateralization of function in the DLPFC has been reported, in that the left DLPFC is linked to future control, whereas the right DLPFC is linked to immediate control (Vanderhasselt et al., 2009). Against this background, future studies could examine whether a combination of increasingly difficult WM training and rTMS over the left or right DLPFC could foster long-term changes in the corticolimbic dopaminergic circuitry that translate into clinical and quality of life improvements.

Secondly, rt-fMRI explores biofeedback mechanisms such that participants learn, via conditioning, the association between physiological experience of cognitive control and visual representation of brain activation (Val-Laillet et al., 2015). However, while emerging evidence supports neurofeedback-guided upregulation of hypoactive networks, it appears more difficult to downregulate hyperactive networks (Ihssen et al., 2016). Nevertheless, some recent evidence suggests that motivational neurofeedback can reduce the incentive salience of appetitive food cues, and that this is linked to alterations in the corticolimbic circuitry (Ihssen et al., 2016). Furthermore, Ihssen et al. (2016) conclude that decreased neural responses to salient stimuli appears not to be regulated by top-down processes, but rather arises from subcortical regions related to implicit operant reinforcement of brain activity. Similarly, for SUD (alcohol), rt-fMRI has been shown to be effective in reducing functional connectivity between ACC, insula (limbic saliency network), inferior temporal gyrus (part of the “what” visual pathway, and may indicate reduced cue-induced saliency) and medial frontal gyrus (associated with self-related processing) (Karch et al., 2015). Thus, against this background, rt-fMRI might positively reinforce a participant involved in WM training, by demonstrating how improved WM performance, and thus cognitive control, is linked to neural activation.

Finally, examining epigenetic variations that are related to gene methylation in peripherally assessed DNA (e.g., blood and saliva) and correlating with behavioral and clinical measures associated with brain function is a new field that aims to locate biological mechanisms of risk for psychiatric disorder (Nikolova and Hariri, 2015). It is suggested that the first steps in the field of imaging epigenetics should examining well-established neural circuitry, such as dopaminergic genes and the reward/motivation pathways (Schultz, 2002) and by combining this approach with multi-modal brain imaging techniques (e.g., positron emission tomography, EEG, fMRI) (Nikolova and Hariri, 2015). Furthermore, it is suggested that epigenetic imaging should be guided by the most significant GWAS data, and in terms of AN and SUD the dopaminergic system (BDNF, DAT, and DAD2) and also the SOX2-OT gene associated with neurogenesis are implicated. Specifically, in terms of AN, while most recent GWAS studies do not support the significant data found in single studies, the opioid, leptin, ghrelin, orexin, and serotonin receptor genes may be promising areas of further exploration for epigenetic imaging studies (Gorwood et al., 2016). Specifically, in terms of SUD, there is an urgent need for biomarkers that are associated with chronic addiction to be classified, but presently the dopaminergic reward system appears to be key (Volkow et al., 2015).