N-acetyl Cysteine attenuates intrinsic functional connectivity, but not neural alcohol cue reactivity, in treatment-seeking individuals with alcohol use disorder

Participants

Thirty-four participants were recruited as part of a larger trial investigating the efficacy of NAC in alcohol use disorder (Morley et al., 2023). Inclusion criteria included AUD according to the DSM-V criteria, a desire to reduce or stop drinking, have consumed at least 21 standard drinks per week or 2 heavy drinking days per week (HDD: ≥5 standard drinks/day for men; ≥4 for women) in the month prior to screening, age 18-70, adequate cognition and English language skills to give valid consent and complete research interviews, and willingness to give written informed consent.

Exclusion criteria were pregnancy or lactation, concurrent use of any psychotropic medication other than antidepressants (provided these were taken at stable doses for at least two months); any other substance dependence other than nicotine; clinically unstable medical (e.g. cancer, end stage liver disease) or psychiatric disorder, (e.g. active psychosis or active suicide risk that precluded trial participation; concurrent use of selenium, vitamin D or other antioxidants; and any alcohol pharmacotherapy used within the previous month. One participant presented significant cognitive dysfunction, and one participant had a significant traumatic brain injury identified during baseline scan and their results were not included in the analyses. 12 participants dropped out of the trial before attending their next scan appointment after the baseline and their data excluded from analyses.

Procedure

Participants underwent a structured interview and medical consultation on day 0 assessing trial eligibility. For the main trial, participants were randomised 1:1 to NAC at 2400 mg/day (2 × 1200 mg capsules twice a day) for 28 days or placebo capsules of identical appearance. Participants completed baseline questionnaires. Imaging sessions were scheduled at baseline prior to first treatment administration (T0) and a follow-up treatment session around 19 days later (SD = 3.73, T1) based on patient availability. Sessions were conducted between 11 am and 4 pm. Participants completed the cue reactivity task approximately 120 minutes after taking their first daily medication capsule.

Assessments

Consumption in the preceding 28 days before T0 scan day was measured using the Timeline followback interview (TLFB) (Sobell and Sobell, 1992) expressed as the number of Australian standard drinks (10 g ethanol) per drinking day (henceforth TLFB units). For T1, participants completed TLFB reporting days since the T0 scan day. The Penn Alcohol Craving Scale (PACS; (Flannery et al., 1999)) measured subjective craving experienced across the week, and the Alcohol Urge Questionnaire (AUQ; (Bohn et al., 1995)) was completed pre-and post scan assessing changes in state craving and urges to drink, with higher scores indicating greater craving, The Alcohol Dependence Scale (ADS) (Skinner and Allen, 1982) measured participants’ AUD severity, with higher scores indicating greater severity and only completed at T0.

fMRI Cue Reactivity

A well-established visual cue reactivity task (Grüsser et al., 2004) was used to measure alcohol cue-elicited brain activity. Stimuli comprised 15 alcohol-related pictures depicting types of alcohol and drinking situations (Grüsser et al., 2004, Wrase et al., 2002), and 30 control images: 15 affective neutral pictures from the International Affective Picture System (Lang et al., 1997) matched for colour and complexity, and 15 scrambled versions of the scrambled alcohol pictures controlling for potential activity related to novelty of neutral images. Images were presented for 6.6 s in blocks of three images of the same type, totalling five blocks per type (alcohol, affective neutral). Stimuli and block order were randomised across subjects, and blocks of the same image type did not follow consecutively. Each condition block was preceded by a fixation cross presented for 10 s modelled as a regressor of no interest. The total task duration was 520 s. Participants were debriefed after test session to address potential continued craving elicited during scanning.

MRI Data Acquisition

MRI data were acquired on a 3-Tesla GE Discovery (GE Healthcare, Milwaukee, Wisconsin, United States) using a 32-channel head coil. A T1-weighted (1-mm^3 voxel resolution) structural scan was acquired for each subject for screening and registration (TR: 7200 ms, TE 2.7 ms, 176 sagittal slices, 1 mm thick, no gap, 256 × 256 × 256 matrix). For BOLD acquisitions, we acquired echoplanar image (EPI) volumes (cue reactivity: n = 183; resting state: n = 200) comprising 39 axial slices in an ascending interleaved fashion with a voxel resolution of 1.88 × 1.88 × 2 mm (TR: 3000 ms, TE 30 ms, FA 90 degrees, FOV 240 mm, matrix 128 × 128, acceleration factor 2, slice gap: 1 mm). Participants’ heads were fixated with foam pads to minimise head movement. For fMRI resting-state acquisition participants were instructed to keep their eyes closed and to focus on nothing in particular.

Image Processing

A detailed overview of imaging processing pipeline is provided in the Supplementary Material, with a summary provided here. Anatomical and functional preprocessing was completed using FMRIPrep ((Esteban et al., 2020, Esteban et al., 2019), RRID:SCR_016216), with the two session T1-weighted structural images corrected for intensity non-uniformity, skull-stripped, and then used for reconstruction of the brain surfaces. Volume-based structural images were segmented into white matter, grey matter, and cerebrospinal fluid, and then spatially normalised into MNI space.

Functional MRI data for the two sessions per subject and per acquisition (cue reactivity, resting-state) were pre-processed with susceptibility distortion correction using a B0-nonuniformity map (i.e., fieldmap) based on two EPI references in opposing phase-encoding directions, and slice-timing correction was completed. Images underwent motion correction, co-registration to structural data, normalisation to MNI space and projection to cortical surface. Functional timeseries were then resampled to FreeSurfer’s (FreeSurfer 6.0.1, surfer.nmr.mgh.harvard.edu) fsaverage space. Cue reactivity data were post-processed within SPM, with functional resampled images smoothed with a Gaussian kernel of 8 mm full-width half maximum (FWHM) to improve sensitivity for group analysis. Resting-state data were further post-processed using eXtensible Connectivity Pipelines (XCP-D) (Ciric et al., 2017, Satterthwaite et al., 2013). Volumes with framewise-displacement (FD) greater than 0.4 were flagged as outliers and excluded. A total of 36 nuisance regressors were regressed out of the BOLD data using the ‘36P’ strategy (Ciric et al., 2017, Satterthwaite et al., 2013). Residual timeseries were then band-pass filtered (0.01-0.08 Hz) and spatially smoothed with a kernel size of 6mm (FWHM). Fisher’s r-to-z transformation was applied to the Pearson correlation coefficients per cell of the resultant matrices of the ROI-to-ROI connectivity matrix using the Schaefer 17-network 400 parcel atlas (Schaefer et al., 2018) for each subject outputted from the XCP-D processing. Of the 400 parcels from the 17 functional networks, 23 parcels contained missing data and were excluded from second-level analyses (See supplementary Material for information), leaving a 377×377 symmetrical functional connectivity matrix per participant per session (46 matrices total).

Statistical analysis

Differences in treatment group sample demographics, clinical and drinking measures, and PACS craving were assessed by two-sample t-tests or Mann-Whitney U tests, where appropriate. We used a linear mixed effects model (LMM) with a random intercept of participant to assess any group differences between pre- and post-scan session AUQ craving compared across T0 and T1 sessions.

fMRI cue reactivity analyses were conducted in SPM12 at two levels. Two conditions were modelled at the first level (subject-specific): alcohol-related (Alcohol) cues; and both control cues combined in a single condition (Control) to control for both novelty and visual complexity of the stimuli, modelled as a box-car function convolved with the canonical haemodynamic response.

Motion correction parameters (six regressors) also modelled at the first level as regressors of no interest. The fixation cross was left as an implicit baseline.

Regions of interest (ROIs) were selected based on areas correlated with alcohol cue reactivity identified by a recent meta-analysis evaluating pharmacotherapy studies in AUD (Zeng et al., 2021). Five ROIs were used: the left and right caudate, the left and right dorsolateral prefrontal cortex (DLPFC), and bilateral ventromedial prefrontal cortex (VMPFC). The caudate was defined here as caudate body from the Harvard-Oxford subcortical probability atlas (http://www.cma.mgh.harvard.edu/fsl_atlas.html). The DLPFC and VMPFC were probabilistic maps defined using the Brainmap database (Fox and Lancaster, 2002) binarized with a threshold of ≥.90. We used the MarsBar toolbox (Brett et al., 2002) to extract the ROI unweighted beta estimates (β) for the baseline and treatment scans which were then exported into R software (Version 4.0.3) for further analyses using LMMs to assess group treatment differences in ROI activity while accounting for other variables. To optimally balance Type I and Type II error associated with multiple comparisons across ROIs, we accounted for the correlation between the dependent variables of the five probabilistic ROI using the Simple Interactive Statistical Analysis Bonferonni tool (http://www.quantitativeskills.com/sisa/calculations/bonfer.htm). We applied a method outlined in Li et al. (2014), whereby the resultant Bonferroni correction is less stringent as it does not assume the variables are obtained from independent subgroups, as would occur in a standard and more conservative Bonferroni correction. The 5 ROI betas had a mean correlation coefficient of r = .066 (number of tests k = 5), resulting in an equivalent corrected alpha of .029 which would be equivalent to a corrected P <.05.

As difference contrast comparisons of Alcohol > Control have low reliability for repeated scans (Bach et al., 2022) the individual contrast images for the Alcohol and Control image conditions were used for second-level models. LMMs were conducted comparing ALC or CON image type across sessions (T0, T1), treatment, and a session-by-treatment interaction. Patient age was added to the models to account for differences in brain activation associated with age and controlling for additional years of harmful drinking seen in older drinkers. Antidepressant use was added as a covariate due to the potential augmentative effects of NAC associated with antidepressant use and mixed evidence regarding improved outcomes when antidepressants and NAC are administered concurrently, possibly due to reductions in oxidative stress (Costa-Campos et al., 2013). Presence of alcohol-related liver disease (ArLD) was added due to its effects on cortical concentrations of GABA (Morley et al., 2020) and potential impacts on cue reactivity related to symptoms (Logge et al., 2019).

ANCOVAs were conducted for T0 and T1 respectively for the ALC > CON contrast to evaluate ROI beta activity at baseline and after treatment and are presented in Supplementary Material. Age was included as a covariate in the analyses for both sessions. The T0 ANCOVA included TLFB drinks per drinking day prior to enrolment, and respective interactions, to evaluate the effect of baseline drinking on cue reactivity. The T1 ANCOVA included TLFB drinks per drinking day as a main effect, and number of treatment days as an additional covariate along with respective interactions to control for the time of T1 scan based on patient availability.

We performed additional exploratory whole brain analyses to examine brain activity not evaluated by the a priori ROIs using the ‘Multivariate and Repeated Measures for Neuroimaging’ (MRM) toolbox (version 1.0, (McFarquhar et al., 2016)) a multivariate and permutation approach which allows advanced statistical modelling of repeated measures mixed effects designs using a multivariate form of the general linear model employed within Matlab. First level contrasts (ALC, CON) were entered into respective second-level full factorial whole-brain general linear models (GLM) comparing NAC and placebo using two sample t-tests to assess group differences in alcohol and control cue reactivity across the two sessions (T0 and T1). Statistical thresholds were set using permutation-based inference, with 5000 permutations conducted to appropriately account for the within and between-subjects’ variance in the model while assessing the betas for the two conditions. Whole-brain analyses were corrected with a family-wise error cluster level inference (pFWEc) used at < 0.05 with a cluster-forming height threshold of p = .001, in accordance with recommendations by Eklund et al. (2016). MNI coordinates are reported in x, y, z dimensions, with regions identified using the SPM Wake Forest University (WFU) Pickatlas toolbox (http://www.fmri.wfubmc.edu/cms/software, version 3.0.5). Whole-brain results are reported in the Supplementary Material.

Resting state data were analysed using CONN toolbox (Whitfield-Gabrieli and Nieto- Castanon, 2012) version 22a (Nieto-Castanon and Whitfield-Gabrieli, 2021). The 46 377×377 ROI- to-ROI z score correlation matrices were entered into a second-level GLM model with within-subject contrast session (T0 < T1 [i.e., -1 1]) evaluating post-treatment effects and between-subjects treatment (NAC > placebo [i.e., 1 -1]) evaluating the effect of NAC. Age, antidepressant use, and ArLD were added as covariates. A parametric multivariate statistics approach was conducted to target the contributions of the individual ROIs. An FDR-corrected ROI-level p-value (MVPA omnibus test) was used with a threshold of PFDR <.05 (Benjamini and Hochberg, 1995) with a connection (height) threshold of p <.01 (uncorrected) to examine the individual connections between the ROIs.

Sample Characteristics and Drinking Outcomes

Table 1 summarises the demographic and clinical characteristics of the sample and drinking outcomes. There were no significant differences between placebo and NAC groups for age, gender, years of education, or unemployment. Both groups demonstrated similar levels of TLFB drinks per drinking day prior to treatment, number of years since alcohol-related problems began, and levels of alcoholic liver disease, tobacco use, and anti-depressant use (p’s > .269). There were no group differences in severity of alcohol dependence measured by ADS (p > 386).

Subjective Craving

Craving was reported both for the previous week using PACS and pre- and post-scan across both sessions with mean AUQ scores presented in Table 1. No differences in PACS subjective craving were seen between groups, between sessions, or interactions (p’s > .731). AUQ scores during sessions showed no differences according to treatment group, pre- versus post-scan, or between sessions, or interactions (p’s > 104) . LMMs of AUQ scores are presented in Supplementary Material. The LMM for T0 assessing baseline and T1 only assessing changes in craving after treatment found no main treatment or pre- versus post-scan differences in AUQ score. However, there a significant main effect of age for the T1 model, with older participants reporting less overall craving during treatment (β = -0.33, p = .038).

Alcohol Cue Reactivity

The contrasts for the separate image types (ALC, CON) are used to evaluate treatment across sessions. The LMM model tables for these ALC and CON contrasts for 5 ROIs are presented in Table 2 and Table 3. The ALC v CON contrast tables at T0, T1, and across sessions are further presented in the Supplementary material for reference.

When evaluating reactivity to ALC cues across T0 and T1 sessions, there were no significant effects seen for treatment group, session, or two-way interaction seen for any of the 5 ROIs. This was similar for CON cues across T0 and T1 sessions, with no significant main effects or interactions seen for the 5 ROIs.

Resting state/intrinsic functional connectivity

There was a significant treatment-by-time interaction seen with reduced functional connectivity patterns observed in the NAC participants compared to placebo after treatment in one ROI seed region, posterior cingulate 9 (MVPA PFDR = .01) in the dorsal attentional (B) network of the Schaefer 400 atlas parcellation. The connectivity target regions included those in the left and right somatosensory networks, left temporal-parietal network, bilateral visual-central and visual-peripheral regions, and salience ventral attentional networks. Fig. 1 presents these connections, with connectivity values reported in Supplementary Material. No evidence of other significant connectivity patterns was seen at either the session or treatment levels, or interaction.

• Download figure
• Open in new tab

Fig 1.

Axial (top-left panel), right sagittal (right panel) and left sagittal (bottom-left panel) views of the significant ROI-to-ROI treatment-by-session interaction. Spheres represent Schaefer-400 parcellation nodes (regions of interest) for the significant connections between the posterior cingulate (dark green), with lines representing connections between nodes that were significant at a nominal, uncorrected p-value threshold of 0.01. Networks are labelled in the figure are color-coded according to Yeo et. al (2011) networks.