Prefrontal cortex activation and stopping performance underlie the beneficial effects of atomoxetine on response inhibition in healthy and cocaine use disorder volunteers

250 Tables: 2 Figures: 3 Supplementary Figures: 4 Supplementary Tables: 3

Neurochemically, response inhibition encompasses dopaminergic and noradrenergic mechanisms operating at distinct cortical and subcortical sites (12,13). Norepinephrine in particular appears to have a preferential contribution to response inhibition in the prefrontal cortex (PFC) (12,14).
Given the prominence of response inhibition difficulties in the conceptualisation of addiction models (7), pharmacological interventions of norepinephrine neurotransmission have been suggested as potentially improving executive inhibitory control in addiction, though its effects in CUD remain to be established (15)(16)(17). Atomoxetine is a well-tolerated selective presynaptic norepinephrine transporter blocker (18)(19)(20) prescribed for Attention Deficit Hyperactivity Disorder (ADHD), which is also characterised by impulsive behavior and poor response control (21). Acute administration of atomoxetine has been found to improve response inhibition in healthy volunteers and in adult ADHD patients (14,22) and longer-term administration in ADHD has been associated with broader attentional control improvements (23). Additionally, atomoxetine was found to upregulate rIFG during stopping in healthy volunteers (24). Further, atomoxetine ameliorated attentional bias to drug-related cues in CUD participants (25).
However, not all studies on atomoxetine have found performance improvements in stopping or concomitant brain correlates (26-29). A possible reason for this may be baseline-dependent individual differences, whereby only those with worse stopping performance benefit from atomoxetine administration. Two studies on atomoxetine administration in older adults with Parkinson's Disease (PD) found that atomoxetine-related improvement in stopping was associated with baseline performance (30,31). In one of these studies, improvement of response inhibition by atomoxetine was also associated with increased rIFG activation (30). Moreover, these studies pointed to atomoxetine enhancing and even restoring abnormal connectivity within the stopping fronto-striatal network (30,31).
In the current study, we investigated the effects of atomoxetine on inhibitory performance and associated brain function in CUD and healthy control participants. We used pharmacological functional magnetic resonance imaging (fMRI) of the stop-signal task (6,32), expecting CUD participants to show performance impairments that could be remediated by atomoxetine. We hypothesized that changes in fronto-striatal regions subserving stopping would underlie any beneficial effects of atomoxetine. We further aimed to identify effects of atomoxetine on effective connectivity of the stopping network. To this end, we built on previous dynamic causal modelling (DCM) of the stop-signal task (9), employing this approach on healthy controls and on CUD participants. We extended the network being investigated to encompass the putamen to better assess the interactions between the prefrontal cortex and basal ganglia.

METHODS AND MATERIALS Participants
A starting sample of twenty-eight healthy participants and twenty-eight individuals who satisfied DSM-IV TR (33) criteria for cocaine dependence, here referred to as cocaine use disorder (CUD) were recruited from drug treatment services by advertisement and word of mouth. Two CUD patients did not complete the task in the scanner and thus were excluded. Healthy control participants were recruited from the Cambridge BioResource volunteer panel (www.cambridgebioresource.org.uk) and had no current or past psychiatric disorders. All participants were screened using the Mini-International Neuropsychiatric Interview (34) and completed the Beck Depression Inventory Version II to record levels of dysphoric mood. We did not assess ADHD but none of our participants had a prior diagnosis of ADHD and had been prescribed stimulants for the treatment of ADHD. Cocaine dependence was verified using the Structured Interview to the DSM-IV (Structured Clinical Interview, SCID (35). Fifteen CUD patients also met the DSM-IV criteria for opioid dependence; 10 of whom were taking methadone or buprenorphine as part of their maintenance therapy. Ten CUD patients also met criteria for tetrahydrocannabinol dependence. Participants were excluded if they 1) had a history of neurological disorder, head or brain injury, history of psychotic disorder, or metabolic disorder, 2) were taking any medication that would interact with atomoxetine such as aripiprazole or bupropion, 3) were pregnant, 4) had MR incompatibilities or 5) had been involved in a clinical trial within the past six months. Urine screens verified recent cocaine use in all CUD patients and were drug negative for all control participants. Additional five CUD patients were excluded from each session due to limited task compliance and race model violations (6), resulting in eight CUD patients precluded from analyses incorporating both sessions (see Supplementary Table 1).

Experimental Procedure and Design
The study followed a randomised, double-blind, placebo-controlled, crossover, balanced design. All participants provided written informed consent, which received ethical approval from National Ethics Committee (12/EE/0519; PI: KD Ersche). Participants received orally either 40mg atomoxetine or a placebo of identical appearance, consistent with previous studies (24,30). At least seven days separated the sessions for each participant, which included a neuropsychological battery and brain imaging (25). Blood samples for plasma were collected 150 minutes after administration (mean 366 ng/mL, standard deviation 200 ng/mL) following established pharmacokinetics immediately after scanning (36,37). Participants underwent structural and functional MRI scanning where they performed the stop-signal task. Generalised linear models (GLM) on SSRT and Go RT were conducted with subject-level random effects (equivalent to mixed-effect Analyses of Variance) to explore the main effects of group (cocaine vs control), drug (atomoxetine vs placebo) and the groupby-drug interaction (nlme and car packages in RStudio v3.4.1). Age and atomoxetine plasma levels were included as covariates. Regression weights with their respective t and p values are reported.
Following the GLM analyses described below, predictors of performance changes due to atomoxetine were explored. ANCOVA models (aov package in RStudio, v3.4.1) were fitted to explain the variability in atomoxetine-dependent changes in SSRT and in Go RT.
MRI Acquisition. MRI data were acquired using a Siemens Trio 3T scanner (Erlangen, Germany).
Stop-Signal Task. On go trials participants were required to respond with left or right key presses to corresponding left or right arrow stimuli (100ms) (6). On stop trials, a stop signal subsequently appeared (an orange upward arrow, 300 ms) and participants had to cancel their planned response.
Left and right arrows were counterbalanced and intermixed, and the delay between go and stop stimuli was adjusted in 50 ms steps from an initial value of 250 ms to achieve 50% successful stopping (38). The task included 48 stop trials and 240 go trials in one block, with stop trials repeating at a later time if participants responded before stop signal onset. Inter-trial-intervals varied randomly between 900 and 1100 ms (39). Participants were instructed to respond as quickly as possible and not to delay responding. Key task outcome measures included mean reaction time (RT) on Go Trials and the SSRT calculated using the integration method with replacement of go omissions (32). Participants who did not meet the assumptions of the race were excluded (6). ΔSSRT was calculated as the difference between SSRT on atomoxetine and SSRT on placebo.

MRI Data Processing and Analyses
Pre-processing. FMRI data processing was carried out using FEAT (FMRI Expert Analysis Tool) Version 6.00, part of FSL (FMRIB's Software Library, www.fmrib.ox.ac.uk/fsl). The first five volumes were discarded to achieve steady-state equilibrium. Registration to structural and standard space images was carried out using FLIRT (46)(47)(48) and FNIRT (49,50). Pre-processing included motion correction using MCFLIRT (Jenkinson 2002); non-brain removal using BET (51); spatial smoothing using a Gaussian kernel of FWHM 5mm and grand-mean intensity normalisation; highpass temporal filtering (100s). First level analysis (52) included four regressors of interest: successful stops, failed stops, successful go and error go responses (all other Go-trials) convolved with doublegamma haemodynamic response function. Temporal derivatives were also included for each of the regressors. Successful stops were contrasted with successful go responses (stopping contrast).
Twenty-four movement parameters were included as covariates of no interest along with a prewhitening step.
GLM Analyses. Two GLMs, one for the placebo and one for the atomoxetine drug condition, used one sample whole brain t-tests to identify significant stop-related group mean activations in the control and cocaine groups using FEAT FLAME1 analysis (53). Conjunction analysis tested for overlap between the groups (easythresh_conj.sh, https://warwick.ac.uk/fac/sci/statistics/staff/academicresearch/nichols/scripts/fsl ). An additional GLM included the difference maps of parameter estimate contrasts for placebo and atomoxetine. Here, one sample t-tests of the atomoxetine versus placebo difference maps were used to evaluate drug effects across subjects and a group-by-drug interaction was tested using independent sample t-tests (CUD vs control) (https://fsl.fmrib.ox.ac.uk/fsl/fslwiki/Randomise/UserGuide).
We also sought to assess whether behavioral effects of atomoxetine were conditional on individual differences in brain activation. To test for associations between drug-related changes in task performance and brain activity, the difference between atomoxetine and placebo in SSRT (ΔSSRT) was added as a covariate to the atomoxetine vs placebo difference GLM. Based on previous findings (24,30), an IFG region of interest (ROI) was applied using the Harvard-Oxford atlas. As significant improvement in Go RT was noted, a parallel yet exploratory analysis was conducted with the difference between drug conditions in Go RT (ΔGoRT) added as a covariate to the go>stop drug difference GLM with primary motor cortex (M1) serving as an ROI (primary motor cortex from the Juelich Atlas). For all analyses, images were thresholded using threshold free cluster enhancement in randomise with 5000 permutations (z>2.3, p<0.05). Demeaned order of drug vs placebo sessions was included as a second-level covariate of no interest in all GLMs. Group mean activations and placeboatomoxetine GLMs were conducted using a whole-brain mask while the ∆SSRT and ∆GoRT analyses were conducted within the IFG and M1 masks, respectively. DCM Analyses. To examine the most likely network identified by group mean GLM maps, and to inspect the directed connectivity in that network, dynamic causal models (DCM; Friston 2003) were built and tested in each group and drug condition in SPM12 (https://www.fil.ion.ucl.ac.uk/spm/software/spm12/). We examined the effective connectivity between well-known nodes of the stopping network that included the IFG, dorsal Anterior Cingulate Cortex (ACC), M1, and STN, building on previous DCM findings (9). We extended the network underlying action initiation and inhibition by adding the putamen, a key node of the indirect corticobasal ganglia pathway (54,55). The addition of the putamen allowed for the assessment of striatal contributions to response inhibition via the indirect pathway (56)

Demographic and Behavioral Results
Demographic, clinical and personality measures are reported in Table 1. The participants were well matched in terms of their sex, education level and pattern of alcohol use as reflected by the AUDIT score. CUD patients were younger, had a lower verbal IQ and significantly higher levels of depressed mood on the BDI-II.

fMRI Results
Whole-brain group mean activations were found in the IFG, dorsal ACC, medial frontal gyrus, parietal and visual areas for the stopping contrast. Conjunction analysis revealed wide-ranging overlap in the above areas activated by both cocaine and control groups in both drug conditions ( Figure 1). On placebo, no significant group differences were observed using a whole brain mask. On atomoxetine, CUD patients showed significantly greater activation than controls in the dorsal ACC (peak MNI coordinates [-6, 16, 52], zmax=3.89, p=0.002). There were no significant drug effects in either group, nor was there an interaction between group and drug.
In sum, both CUD and control participants showed robust and largely consistent activations in the key nodes of the stopping network regardless of drug condition. Thus, associations with drug-dependent performance differences (ΔSSRT) were assessed across the entire sample. At the whole-brain level no results survived threshold-free cluster-enhancement multiple comparison correction. Using the bilateral IFG mask, we identified a robust cluster of right IFG activation that was associated with improved SSRT performance on atomoxetine (Figure 2A, 2B). This finding was in line with our hypothesis that fronto-striatal activation would be associated with atomoxetine-induced stopping improvements.
Analyses for the go>stop contrast revealed group mean activations in the left precentral gyrus (M1), contralateral to the right-handed task response as would be expected (Supplementary Figure 1).
Associations with drug-dependent performance differences (ΔGo RT) showed that within the primary motor cortex (M1) ROI, increased activation in a robust cluster was associated with improved Go RT performance on atomoxetine across all participants ( Figure 2C, 2D).

Predictors of changes in performance induced by atomoxetine
To determine the factors predicting changes in performance following atomoxetine administration, mixed-effect generalized linear models (aov package) assessed the contribution of placebo performance ("baseline"), plasma atomoxetine, and change in rIFG activation on SSRT improvements on atomoxetine. The rIFG region was defined based on the voxelwise fMRI results in which ΔSSRT was included as covariate of interest. The full model explained a significant amount of variance in ΔSSRT in all participants (r 2 = 0.64, F4,41=18.33, p= 1.1e-8).

DCM Results
Overall, DCM connectivity analyses were consistent between the groups and largely replicated previous findings regarding the effective connectivity underlying stopping, though the putamen was not included in previous models (9). Bayesian model selection ( Figure  indicated the same winning model for both groups on placebo (nonlinear C). Additionally, the winning model was the same for both groups on atomoxetine (linear D). On placebo, the winning model included nonlinear modulation of the hyperdirect dACC-STN connection by the IFG ( Figure   3B). In contrast, on atomoxetine this nonlinear modulatory connection was replaced with a fixed connection between the IFG and putamen. Looking at task modulation (the red arrow in Figure 3B), stopping modulated the IFG in controls on placebo, but this changed to the putamen when they were

Individual differences in atomoxetine effects
Importantly, atomoxetine improved stopping performance in a baseline-dependent manner independent from diagnosis, in keeping with findings from a much older cohort of healthy volunteers and PD patients (30,31). Thus, poorer inhibitors benefited the most from atomoxetine compared to placebo. Moreover, this was accompanied by increased rIFG activation such that the greatest improvement in stopping latency was associated with greater upregulation of this region. Higher levels of atomoxetine as detected in the blood were also positively correlated with greater brain activation in the rIFG during successful stopping (24). These results are consistent with the inverted-U modulation by norepinephrine of PFC-mediated cognitive control (58)(59)(60). They are reminiscent of findings found not only with atomoxetine but also with methylphenidate which acts on both the noradrenergic and dopaminergic systems, in accordance with the broader literature of optimal catecholamine levels determining optimal performance (61)(62)(63). This also points to the likely utility of placebo inhibitory performance ("baseline") in predicting subsequent effects of atomoxetine on cognitive control across individuals. Larger studies may be more adequately powered to detect grouplevel benefits to CUD patients driven by those who showed increased rIFG activation and improved stopping performance.
Unexpectedly, atomoxetine also improved response execution compared to placebo. These improvements were found in a baseline-dependent manner, independent from diagnosis. Moreover, faster responding with atomoxetine was positively associated with enhanced activation in primary motor cortex, which is typically activated with response execution (8). Atomoxetine is known to improve attention, but acute administration does not typically promote general response speeding (26,64). Faster response latencies with atomoxetine compared to placebo in the stop-signal task were also reported in ADHD boys (28), suggesting speeding may occur in some situations. The effects on response speeding indicate that atomoxetine facilitated compliance without any concomitant negative effects on stopping, as instructions emphasized to respond as fast as possible and avoid slowing.

Effective connectivity underlying response inhibition and execution
Some of the network connectivity findings are relevant to a general understanding of response inhibition, with others being more specific to its basis in CUD. Present results in controls on placebo provide an important replication and extension of previous DCM findings (9), particularly as such replications are uncommon. Despite some methodological differences, not only was the same winning architecture found, but also there was notable agreement as to the effective connectivity between regions. Specifically, dACC projections to M1 and STN were excitatory, while putamen to the STN and STN to M1 projections were inhibitory (65). Positive modulation of the hyperdirect pathway by rIFG allows for top-down control over response cancellation. Present rIFG-dmPFC connectivity diverges from those previously found (9), possibly due to differences in task instantiation such as their inclusion of no-go trials. Additionally, present results found activation in the dorsal part of the dACC rather than the pre-supplementary motor area, though exact ROI locations appeared spatially adjacent.
A subsequent DCM study by the same authors also did not yield similar results, though it was conducted on a considerably older cohort, suggesting that age may modulate stopping network connectivity (31). This would be consistent with the well-established decline in response inhibition from adulthood to older age, with gradual worsening in stopping efficiency (66,67). The present connectivity analyses also extended the stopping network by incorporating an indirect route from the dmPFC to the STN by way of the putamen, as supported by animal and human research in response execution (68). Atomoxetine selectively altered rIFG modulation of dmPFC to STN connectivity for both control and CUD patients, consistent with the notion that it is a key node by which the drug exerts its influence on response inhibition. Group differences in connectivity were noted, though these did not survive type I error control and were obscured by individual differences. Nevertheless, on the whole the results draw attention to the potential importance of cortico-cortical connectivity in addition to corticostriatal connectivity traditionally associated with impulsivity in CUD (69).

Conclusions and limitations
This study points to baseline-dependent improvements in response inhibition with atomoxetine administration along with concomitant rIFG upregulation in a cohort of CUD patients and in healthy volunteers. CUD patients did not demonstrate impaired stopping or reduced PFC activation compared to controls, contrary to expectations (1,6,70,71). However, CUD patients showed significantly more omissions on Go trials, suggesting a degree of hesitancy manifesting from proactive inhibition (72) being used in this group in view of increased impulsivity (6)(73). Some of the participants were also dependent on opioids and cannabis in addition to cocaine. The stop-signal task is a sensitive measure of response inhibition in stimulant use disorder, with impaired stopping in both clinical and preclinical studies (74)(75)(76). Notably, our participants were active cocaine users, and acute cocaine administration has been shown to improve response inhibition (76,77). Alternatively, lack of case-control differences in response inhibition may be due low of power in our sample or differences between the behavioral and the fMRI versions of the stop signal task as some previous studies of CUD patients also report no significant differences in behavioral performance (78)(79)(80)(81), suggesting that the evidence for SSRT impairments in CUD patients is inconsistent (75). Therefore, atomoxetine might prove more beneficial in drug abstinent CUD patients in recovery, strengthening response inhibition and preventing relapse. Further, it is possible that participants who chose to undertake a lengthy pharmacological study with multiple visits and perform the task adequately in the scanner exhibit good executive control. In accordance with this notion, alcohol use was not increased in this cohort, though they reported high levels of trait impulsivity, chronic and compulsive drug use. The two groups were also not matched on demographic characteristics, though age was added as a covariate. In principle, atomoxetine may alter the hemodynamic response to neural activity obscuring or confounding any differences detected by fMRI though there is some evidence to counteract this (82).
The spatial resolution of fMRI methods restricts precision in regions such as the putamen and STN.
To mitigate this, we followed previous methods where possible (9) and used established anatomical masks. The present study used a dose of 40 mg of atomoxetine which is the standard starting therapeutic dose (14). Whilst greater improvements may have been detected with a larger dose, dosage was guided by safety and tolerability considerations.
The results emphasize the nature of response inhibition functioning as existing along a continuum, with considerable overlap between CUD patients and healthy volunteers in the underlying neural network, determining the overall effects atomoxetine has on its nodes and connectivity. Future studies may explore whether this is the case for other forms of impulsive behaviors such as premature responding also found to be abnormal in stimulant-dependent individuals (73), given the improvements with atomoxetine found in rodent studies (83). The findings also underscore the importance of individual differences within the CUD patients in responding to atomoxetine, as those with worse response control are expected to benefit more from atomoxetine. More generally, the association between rIFG upregulation and successful stopping underscoring the effects of         DCM allows us to estimate generative models of brain connectivity between a set of regions of interest (ROIs), which can then be compared in terms of their posterior probability given the BOLD timeseries data (1). We examined the effective connectivity between well-known nodes of the stopping network that included the inferior frontal gyrus (IFG), dorsal anterior cingulate (ACC), primary motor cortex (M1), and the subthalamic nucleus (STN), building on previous evidence (2)(3)(4)(5)(6)(7)(8).

Acknowledgments and Conflicts of Interest
We extended the network by adding the putamen, a key component of the direct and indirect corticobasal ganglia pathways (9,10) underlying action initiation and response inhibition.
The DCMs allowed us to assess: a) fixed connections between these nodes (DCM.a), b) modulatory effects of the task (successful stop > go contrast) on these connections (DCM.b), c) inputs that drive network activity (all trials, regardless of trial type or outcome) and d) nonlinear modulatory effects of one ROI on connectivity between other ROIs (DCM.d). A set of 33 models guided by a priori hypotheses was compared using Bayesian model selection based on the free-energy bound F, adjusted for model complexity. Further, subject-specific connectivity values can be extracted for the most likely model for each group by drug condition using Bayesian Model Averaging.
Model space included 33 models ( Figure 3A), systematically varying in the location of fixed connections (DCM.a), nonlinear modulatory connections (DCM.d) and location of task modulation effects (DCM.b). Fixed connections in linear models tested for systematic differences in the connectivity between the IFG and putamen (linear models A-F). Linear models A-C aimed to test whether ACC-Putamen-STN pathway could replace the hyperdirect pathway (dACC-STN) in stopping; models D-F tested for the role of the IFG given the presence of the hyperdirect pathway and a parallel pathway from the dACC to the STN via the putamen. In particular, we tested whether the IFG-putamen connection was likely, given the data (model D vs model F) and whether an additional projection from the IFG to STN was likely (model E). Nonlinear models examined the addition of nonlinear modulation of the ACC-STN or the ACC-Putamen-STN pathway by the IFG (models A-C).
Nonlinear models D and E tested whether putamen may be modulating projections from the ACC to M1 or from the STN to M1. Each of the 11 models had three versions, with task demands (successful stop vs go) modulating the IFG, ACC or Putamen in each model. This resulted in 33 models, though three models failed to converge and were excluded. In each model, all ROIs also had an autoinhibitory or an autoexcitatory self-connection and all trials provided driving inputs (DCM.c) to the dACC and the IFG following previous definitions of stopping DCMs (8).
Node activations were calculated for each participant. The first Eigenvariate of the BOLD timeseries was extracted from the 1 st level analysis. All anatomical masks were taken from the Harvard Oxford cortical and subcortical atlases (11)(12)(13)(14). The dorsal ACC and IFG spheres were determined as follows: a search region was first created from the intersection of the significant group mean activation for the successful stop vs go contrast and the anatomical masks of ACC and bilateral IFG (pars opercularis and pars triangularis). This was then masked by a sphere (5 mm radius) of the peak activation from each participant's parameter estimate map for the successful stop vs go contrast. The bilateral motor cortex search region was created from the intersection of the primary motor cortex anatomical mask and significant mean activation for successful go vs stop. Individual-specific spheres (5mm radius) were placed at peak activation for the go vs stop contrast in the search ROI. The putamen sphere (3mm radius) was placed at the subject-specific peak activation in the putamen anatomical region. These connectivity values were subsequently used in subject-level random effects analyses to examine potential effects of group, drug condition and interaction between group and drug condition. To correct for the number of multiple comparisons, we used the Bonferroni method (p<0.004, i.e. 0.05 divided by 12 fixed connections common to all winning models, Figure 3B). Figure S1 summarizes the results of the model selection between thirty DCMs representing competing hypotheses regarding the causal interactions between prefrontal cortical regions and the subcortical response inhibition pathways. The winning models in both groups in both conditions did not differ in terms of fixed connections between ROI pairs (DCM.a). On placebo, but not on atomoxetine, the winning model in both groups included nonlinear modulation of the hyperdirect pathway (ACC to STN) by the IFG (DCM.d). Further, stopping modulated the IFG in the control group in the placebo condition, but the location of stopping modulation changed to the putamen when healthy participants were given atomoxetine. In the Cocaine User Dependent (CUD) group, stopping modulated the putamen activity regardless of drug condition.

DCM Results and Discussion
Having clarified the presence or absence of connections and modulatory effects, we investigated the connectivity strength within the winning model identified by BMS. Estimating connectivity strength of fixed connections (DCM.a) allows us to assess whether a region is providing excitatory or inhibitory inputs to another region, while nonlinear modulations inform us about increase or reduction in connectivity between two regions by a third modulatory ROI.
The analysis revealed striking similarities to the network architecture previously identified in humans and animals. In particular, the hyperdirect pathway (dACC to STN) as well as the nonlinear modulatory influence of the IFG on the hyperdirect pathway were excitatory. This finding is consistent with a role for the IFG in increasing the excitatory connectivity in the hyperdirect pathway when participants successfully stop a motor response and allowing ACC to activate the STN more strongly. The STN exerts inhibitory control over motor cortex (Redgrave et al 2010) and can thus relay the stopping command it received from the ACC to successfully inhibit the initiated response in the motor cortex. Figure S1. Significant activation maps for the [Successful Go>Successful Stop] contrast showcase the motor cortex activation -precentral/postcentral gyrus. Shown is the conjunction between control and CUD group, cluster corrected with the cluster forming threshold of z>2.3 and p<0.05. Figure S2. Relative log evidence for each of the thirty models included in the Bayesian model comparison and selection. Models 1-10 are nonlinear models A, B, C, D, E and linear models A, B, C, D, E, F with task modulation of the IFG; models 11-20 are the nonlinear models A, B, C, D, E and linear models A, B, C, D, E, F with task modulation of the ACC; models 21-30 are nonlinear models A, B, C, D, E and linear models A, B, C, D, E, F with task modulation of the putamen. Three models were excluded since they failed to converge for several subjects: nonlinear model D with IFG modulation, nonlinear model E with ACC modulation and nonlinear model E with putamen modulation. The winning models and are highlighted in red. In the placebo condition, nonlinear model C provided the best fit to data from both control and cocaine groups, although stopping modulated the IFG in the controls and the putamen in the cocaine group. In the atomoxetine condition, linear model D with task modulation of the putamen gathered the most evidence in both control and cocaine groups.

Group differences in Go Omissions
In addition to the results reported in Table 1, we show the distribution of the Go omissions in Figure  S3. Although a robust, significant difference between CUD patients and controls was found (with increased probability of Go Omissions in CUD), the distribution was heavily skewed towards zero (as shown in Figure S3) and we therefore refrain from testing for linear associations between Go Omissions with task-based stop-signal fMRI data. Figure S3. Go omission distribution in healthy controls and CUD patients.