Abstract
Alcohol use disorder (AUD) causes complex alterations in the brain that are poorly understood. The heterogeneity of drinking patterns and the high incidence of comorbid factors compromise mechanistic investigations in AUD patients. Here we used male Marchigian Sardinian alcohol-preferring (msP) rats, a well established animal model of chronic alcohol drinking, and a combination of longitudinal resting-state fMRI and manganese-enhanced MRI to provide objective measurements of brain connectivity and activity, respectively. We found that 1 month of chronic alcohol drinking changed the correlation between resting-state networks. The change was not homogeneous, resulting in the reorganization of pairwise interactions and a shift in the equilibrium of functional connections. We identified two fundamentally different forms of network reorganization. First is functional dedifferentiation, which is defined as a regional increase in neuronal activity and overall correlation, with a concomitant decrease in preferential connectivity between specific networks. Through this mechanism, occipital cortical areas lost their specific interaction with sensory-insular cortex, striatal, and sensorimotor networks. Second is functional narrowing, which is defined as an increase in neuronal activity and preferential connectivity between specific brain networks. Functional narrowing strengthened the interaction between striatal and prefrontocortical networks, involving the anterior insular, cingulate, orbitofrontal, prelimbic, and infralimbic cortices. Importantly, these two types of alterations persisted after alcohol discontinuation, suggesting that dedifferentiation and functional narrowing rendered persistent network states. Our results support the idea that chronic alcohol drinking, albeit at moderate intoxicating levels, induces an allostatic change in the brain functional connectivity that propagates into early abstinence.
SIGNIFICANCE STATEMENT Excessive consumption of alcohol is positioned among the top five risk factors for disease and disability. Despite this priority, the transformations that the nervous system undergoes from an alcohol-naive state to a pathologic alcohol drinking are not well understood. In our study, we use an animal model with proven translational validity to study this transformation longitudinally. The results show that shortly after chronic alcohol consumption there is an increase in redundant activity shared by brain structures, and the specific communication shrinks to a set of pathways. This functional dedifferentiation and narrowing are not reversed immediately after alcohol withdrawal but persist during early abstinence. We causally link chronic alcohol drinking with an early and abstinence-persistent retuning of the functional equilibrium of the brain.
Introduction
Excessive consumption of alcohol is positioned among the top five risk factors for disease and disability worldwide and alcohol use disorders (AUDs) have a prevalence of 5.1% in a population aged ≥15 years (World Health Organization, 2018). These alarming numbers worsen in the light of recent studies indicating that there is no safe level of alcohol consumption (Topiwala et al., 2017). Despite this priority and a strongly growing body of literature, we still have a poor understanding of the brain mechanisms underlying this aberrant behavior, especially of those alcohol-induced alterations that are accessible to objective measurements and that may comprise early warning signs for an ongoing disease process (Heilig et al., 2019).
Functional magnetic resonance imaging (fMRI), in particular the measurement of spontaneous or intrinsic neural activity at rest [resting-state fMRI (rs-fMRI)], can be used for examining brain-wide networks (Fox and Raichle, 2007; Cole et al., 2010). With this technique, networks can be generated with nodes defined either by a priori criteria, anatomic or functional, or by data-driven criteria, based on the signal dynamics in the individual imaged points or voxels. In the second modality, the nodes, also called resting-state networks (RSNs), are composed by brain regions that share activity patterns, and one brain region can contribute to more than one RSN. The links connecting pairs of RSNs in a brain-wide network are often defined as the correlation strength between the activity time courses in the nodes. Correlation analysis of blood oxygenation level-dependent (BOLD) signals from rs-fMRIs of AUD patients has shown diverse alteration in various RSNs (Chanraud et al., 2011; Weiland et al., 2014; Müller-Oehring et al., 2015; Dupuy and Chanraud, 2016; Zhu et al., 2016, 2017; Kohno et al., 2017; Vergara et al., 2017). While these studies provide a valuable account of the functional alterations associated with persistent alcohol consumption, establishing causality to the pathophysiology of AUD is challenging. Clinical studies inherently contain various potential confounds such as different patterns of abuse, often including multiple substances, heterogeneity in disease phenotype, somatic and psychiatric comorbidity, smoking, nutrition, and socioeconomic factors, to name a few. Furthermore, the neural alterations occurring during the transition from an alcohol-naive state to a pathologic drinking state are difficult to investigate longitudinally in human subjects.
In contrast, MRIs in laboratory animals provide the tight experimental control necessary for establishing causality. The high translational utility of animal AUD models has been recently demonstrated, using structural and functional MRI, in the identification of a novel pathologic mechanism during the early weeks of abstinence after high voluntary alcohol consumption (De Santis et al., 2019, 2020; Scuppa et al., 2020). However, when investigating functional connectivity (FC) with MRI in experimental rodents, the use of anesthesia needs to be considered (Moreno et al., 2013; Pan et al., 2015). While experiments in rodents (Paasonen et al., 2018) and primates (Vincent et al., 2007) show a comparable organization of resting-state connections in awake versus anesthetized conditions, which mostly recapitulate the underlying known structural connectivity (Barttfeld et al., 2015; Díaz-Parra et al., 2017), complementary measurements of brain activity in awake conditions are desirable. Awake fMRI studies in rodents have been developed (Zhang et al., 2010; Stenroos et al., 2018) but involve habituation periods of intense stress, which will unavoidably interfere with animal models in which the stress is an important factor, as is the case in AUD (Ciccocioppo et al., 2006; Ciccocioppo, 2013).
Other functional imaging modalities, such as manganese-enhanced MRI (MEMRI), permit recordings of brain-wide network activations produced during awake and behaving conditions. The paramagnetic ion manganese (Mn2+) permeates cells through voltage-gated calcium channels and accumulates in depolarized neurons. Since Mn2+ clearance is slow, the pattern of activity-dependent Mn2+ accumulation produced in awake conditions can then be measured as increased signal intensity in T1-weighted images (Silva et al., 2004; Eschenko et al., 2010a). This technique has been successfully used to identifying neural circuits associated with specific behaviors (Eschenko et al., 2010a; Gildish et al., 2012; Bangasser et al., 2013; McGuire et al., 2013; Laine et al., 2017; Yang and Li, 2020), including cocaine addiction (Cannella et al., 2020) and excessive alcohol drinking in rats (Dudek et al., 2015, 2016; Pallarés et al., 2021).
Here we used a multimodal MRI approach combining anesthetized BOLD-contrast fMRI and awake MEMRI. We thus complemented functional connectivity with regional neuronal activity analysis to provide new insight in the early remodeling of brain networks initiated by chronic alcohol drinking. We used the Marchigian Sardinian alcohol-preferring (msP) rat line (Ciccocioppo, 2013), a well established model that has demonstrated strong translational value (De Santis et al., 2019, 2020). This rat line consumes pharmacologically relevant blood alcohol levels inducing reward, antidepressant and anxiolytic actions (Ciccocioppo et al., 2006; Borruto et al., 2021), matching some of the characteristic alcohol effects in AUD patients.
Following 1 month of voluntary alcohol drinking, we identified widespread changes in functional connectivity patterns and increased neuronal activations in prefrontocortical and striatal regions. Interestingly, two fundamentally different forms of network reorganization emerged according to the pattern of connectivity induced by alcohol drinking. These changes defined a new alcohol-driven equilibrium state that does not return to baseline on alcohol withdrawal but propagates into early abstinence.
Materials and Methods
Animals and experimental design.
A total of 72 male msP rats (age, 4 months; weight, 370–480 g) were imported from the breeding facility at University of Camerino, Italy (Ciccocioppo et al., 2006; Ciccocioppo, 2013). Two cohorts (n = 36 rats) were used for longitudinal fMRI studies, and another two cohorts (n = 36 rats) were used for MEMRI. Part of the MEMRI data were used in a recent report comparing two alcohol-preferring rat strains (Cohort 3 below; Pallarés et al., 2021). For the experiments, the rats were individually housed in transparent polycarbonate cages with bedding, a wood stick, and nesting material, with ad libitum access to food and water, regulated temperature (21 ± 1°C), and relative humidity (55 ± 10%), and on a regular 12 h light/dark cycle (lights on at 8:00 A.M.). All experiments were approved by the Animal Care and Use Committee of our institution (Consejo Superior de Investigaciones Científicas) and comply with the Spanish (law 32/2007) and European regulations (EU directive 86/609, EU decree 2001–486, and EU recommendation 2007/526/EC).
Two rs-fMRI longitudinal studies were conducted. Animals of Cohort 1 (n = 18) had free access to alcohol in a two-bottle free-choice paradigm (two 250 ml drinking bottles: one filled with water, the other containing 10% ethanol in water, changed and measured twice per week) and were imaged at the following three time points: (1) before alcohol access [naive condition, time point 1 (TP1)]; (2) after 30 d of alcohol drinking (alcohol condition, TP2); and (3) after 1 week without alcohol access (abstinent condition, TP3). For the last scan, only nine randomly selected animals were used. Cohort 2 (n = 18) was imaged twice in the alcohol-naive condition with a 30 d interval (TP1 and TP2) and was used as a control to account for any age-related differences in brain resting state FC (Fig. 1A).
Cohort 3 for the MEMRI experiment was assessed in a cross-sectional design, with the alcohol-drinking group (n = 14) having access to water and alcohol, as described above, and the control group (n = 14) having access only to water during the same 30 d (see Fig. 7A). In this cohort, MnCl2 was administered using an osmotic minipump (model 2001, ALZET) delivering 200 µl of MnCl2 (1 µl/h) during a 7 d infusion period. MnCl2 was prepared as an isotonic solution dissolved in Tris-buffered saline, pH 7.4. The total dose of infused MnCl2 was 80 mg/kg. Before the surgery, the pumps were primed overnight in a 37°C saline solution. The pumps were implanted subcutaneously on the dorsum, slightly caudal to the scapulae. Animals were imaged immediately following the 7 d MnCl2 infusion.
In Cohort 4, we investigated the influence of the manganese administration time window on the obtained activity maps. In the previous MEMRI experiment, the obtained activation maps represented the integrated manganese accumulation over the last drinking week. However, in the rs-fMRI experiments, we measured functional connectivity on the last day of alcohol drinking. To compare both imaging time windows, we performed one additional MEMRI experiment with an acute injection of 80 mg/kg, i.p., MnCl2 (in Tris-buffered saline), pH 7.4, before both the alcohol-drinking phase (TP1) and the final 24 h drinking session (TP2), and then imaged manganese accumulation 24 h later. In this within-subjects design (n = 8), functional MEMRI maps represent brain activity on the last day of alcohol drinking (see Fig. 6A).
MRI acquisition protocol.
The experiments were conducted in a horizontal 7 tesla MRI scanner containing a 30-cm-diameter bore (Biospec 70/30v, Bruker Medical) and a 675 mT/m actively shielded gradient coil (model BGA 12-S, Bruker) of 11.4 cm inner diameter. To achieve the highest possible signal-to-noise ratio (SNR), a 1H rat brain receive-only phase array coil with integrated combiner and preamplifier, no tune/no match was used together with the actively detuned transmit-only resonator (BrukerBioSpin MRI). MRI images were acquired and minimally preprocessed with a Hewlett-Packard console running Paravision 5.1 software (Bruker Medical) on a Linux platform. The msP rats were placed in an MRI-compatible stereotaxic device with adjustable ear bars and bite bars on the magnet bed. To maintain the vascular reactivity triggered by neuronal activation, the temperature of the rats was preserved (37 ± 0.5°C) using a water blanket connected to a temperature-regulated water bath (SAHARA S5P Heated Bath Circulators, Thermo Fisher Scientific). Furthermore, vital constants were monitored (MouseOx, Starr Life Sciences) to make sure that the following optimal values were fulfilled: heart rate (300 ± 50 beats/min), oxygen saturation (>95%), and breathing rate (90 ± 10 breaths/min; Pacheco-Torres et al., 2018; Pérez-Cervera et al., 2018).
rs-fMRI data acquisition.
The msP rats were anesthetized with 1% isoflurane in oxygen (0.8–1 L/min). rs-fMRI acquisition was performed using a gradient echo-EPI sequence with the following parameters: field of view (FOV) = 25 × 25 mm; slice thickness = 1 mm; 15 coronal slices; five 3 min runs; matrix size = 96 × 96 voxels; flip angle = 60°; echo time (TE) = 15 ms; and repetition time (TR) = 2000 ms. To facilitate the registration of these functional images to a standard space, we also collected anatomic T2-weighted MRI images, using rapid acquisition relaxation enhanced sequence (RARE) and applying the following acquisition parameters: FOV = 25 × 25 mm; slice thickness = 1 mm; 15 coronal slices; matrix size = 192 × 192 voxels; RARE factor = 8; effective TE = 56 ms; and TR = 2000 ms.
MEMRI data acquisition.
T1-weighted images were acquired on the anesthetized rats (1% isoflurane in oxygen, 0.8–1 L/min) using a multislice multiecho pulse sequence (TR = 300 ms; TE = 14 ms; averages = 8; FOV = 32 × 32 mm; and 9 horizontal slices of 0.5 mm thickness with matrix size = 256 × 256 voxels, resulting in 0.125 × 0.125 × 0.5 mm3 voxel resolution).
rs-fMRI image preprocessing.
rs-fMRI data were preprocessed within runs using FSL 5.07 tools (Oxford Center for Functional MRI of the Brain, Oxford, UK; Smith et al., 2004; Jenkinson et al., 2012) and MATLAB 2014a (MathWorks). First, the Bruker images were converted to NIfTI (Neuroimaging Informatics Technology Initiative) data format, and the voxel resolution was scaled up by 10 to mimic human brain size and apply correct spatial transformations (Kalthoff et al., 2011; Pan et al., 2015). Afterward, motion correction was applied with respect to the middle volume (Jenkinson et al., 2002) and the brain was segmented (Smith, 2002). Next, the transformation matrix to register the functional images to a rat brain T2-weighted MRI template (Schwarz et al., 2006) was calculated as the concatenation of two matrices: (1) rigid matrix to coregister the functional images onto the anatomic image; and (2) affine matrix to register the anatomic images to the standard template. The next step was noise reduction, including a 4 mm full-width at half-maximum (FWHM) Gaussian smoothing (Smith and Brady, 1997). Subsequently, global 4D mean-based intensity normalization was applied and the variance tied to six motion parameters (rotations and translations along the three principal axes) was regressed out. This step was followed by a bandpass temporal filtering (nonlinear high-pass filter, σ = 50 s; Gaussian linear low-pass filter, σ = 2 s) to retain frequencies in the 0.01–0.1 Hz range, the frequency band with biological relevance for rodents under isoflurane anesthesia (Pan et al., 2013). Finally, the transformation matrix was applied to the functional images, which became normalized to a standard space (Schwarz et al., 2006).
Next, to obtain run-specific time series of each RSN, the following two steps were performed: group probabilistic independent component analysis (PICA) and spatial regression.
Group probabilistic independent component analysis.
PICA (Beckmann et al., 2005) was applied to the data (individual runs) in naive and alcohol conditions in Cohort 1, using MELODIC (Multivariate Exploratory Linear Decomposition into Independent Components) version 3.14, part of FSL (FMRIB Software Library; Smith et al., 2004; Jenkinson et al., 2012). The aim of this data-driven approach is to find group linear-mixed, independent and non-Gaussian sources, leading to spatial maps with minimal spatial redundancy, each of them having voxels that share a unique time course (Smith et al., 2014). Before this analysis, each brain-extracted image was voxel-wise preprocessed, including demeaning and variance normalization. Then, these data were temporally concatenated across control and alcohol conditions, whitened, and projected into a 17-dimensional subspace using principal component analysis. Seventeen dimensions were the best balance between the fusion and separation of brain regions of the same RSN. The whitened observations were decomposed into sets of vectors that describe signal variation across the temporal domain (time courses), the run domain, and the spatial domain (maps) by optimizing for non-Gaussian spatial source distributions using a fixed-point iteration technique (Hyvärinen, 1997). Then, the estimated component maps were divided by the SD of the residual noise and threshold to place equal loss on false positives and false negatives, by fitting a mixture model to the histogram of intensity values (Beckmann and Smith, 2004). Group-PICA parceled the functional data into 17 independent components (ICs)—spatial maps and time courses—some reflecting spatial artifacts, blood vessels, or CSF (Fig. 1C), and others corresponding to well characterized RSNs (Fig. 1B; Becerra et al., 2011).
Spatial regression.
To obtain run-specific time courses for the ICs identified by group-PICA, we used the first step of dual-regression (i.e., spatial regression). The whole set of group-PICA spatial maps was spatially regressed onto the rs-fMRI data, obtaining run-specific β-coefficients that characterize the temporal dynamics for each spatial map—time courses for each group-PICA map—and condition (Beckmann et al., 2009; Smith et al., 2014).
Direct and indirect rs-FC.
To investigate the interactions between RSN in the brain-wide network, we followed an approach based on multiple multivariate linear regressions (mlrs), combining full and partial models (López-Madrona et al., 2020). For each RSN or node (Noded), we have defined an mlr, where the target node (Noded) is defined by a linear combination of the other nodes, or factors, plus a residual, as follows:
This model gives the total variance of Noded that can be explained by all the other nodes (variance of
Using different subsets of this model, it is possible to identify the variance that is explained by each factor or combination of them. For example, if only one factor is included in the model:
We would obtain the total variance that can be explained by Node1, independent of the contributions of other nodes. It can be compared with the full correlation between Noded and Node1. We call this variance var1full.
To obtain the variance that is explained solely by Node1, which cannot be explained by any other factor, we remove Node1 from the full mlr, as follows:
The difference between the varfull, minus the variance explained with all factors except Node1, would indicate the specific contribution of Node1 to Noded. We call this variance var1par. This can be compared with the partial correlation between Noded and Node1.
The next question would be, in case that var1par is lower, which of the other nodes are explaining the same variance of Node1? We can address the variance shared by two nodes with a submodel including them. Continuing with the example, if we are interested in the variance of Noded that is explained by Node1 but that can be also explained by Node2, the first step is to compute an mlr with both factors, as follows:
We call var12full the variance that is explained by Node1 and Node2. Then, we compute the submodel including only the second factor, obtaining var2full. In a similar manner as before, the variance that is explained by Node1 that cannot be accounted by Node2 is obtained as the difference between var12full minus var2full. We call this variance var1par2. This value should be equal or higher than var1par, and lower or equal than var1full. Furthermore, the difference between var1full minus var1par2 would be the proportion of variance explained by Node1, which is also explained by Node2. In other words, this can be interpreted as the indirect information between Noded and Node1, which is mediated by Node2. Note that this indirect pathway may also be shared by other nodes in the network. However, we did not further partialize the model as it may result in negligible connectivity values.
We first used Cohorts 1 and 2 to assess whether changes in RSN connectivity were solely driven by alcohol drinking or had an age-related contribution. To this end, we first computed the difference in the global connectivity (average of all full connectivity values) between the two time points (1 month interval) in each cohort, and then compared the difference using a two-tailed Wilcoxon rank-sum test. We then tested in Cohort 1 the longitudinal transformations induced by alcohol drinking. Finally, using Cohort 1, we further tested whether 1 week of abstinence after 1 month of alcohol exposure induced any further alteration or homeostatic recovery of connectivity. To compare network interactions between conditions, we first concatenated the time series per subject and group, and computed the regressions in the concatenated data. After that, we averaged the results across subjects and computed the difference between conditions. For the statistical analysis, we followed a surrogate approach. We computed 1000 surrogates by randomly switching the runs of each subject between experimental conditions. The result of each surrogate in each condition was given by the average across subjects. Finally, we computed the difference between both conditions. These surrogate distributions can be approximated to normal. The p-values associated with each result were determined as the cumulative surrogate distribution evaluated at that point (i.e., difference between conditions). We corrected for multiple comparisons by multiplying the resultant p-values by the number of links analyzed (21 pairs).
Estimating rs-FC with regularized partial correlation.
We further used a more restrictive method to measure partial correlation between RSNs, which is especially suited for time series with a low number of samples, as in the case of fMRI. We used FSLNets toolbox version 0.6, part of FSL. After demeaning each time course, the time courses from the discarded ICs were regressed out of the time courses of the RSNs (Griffanti et al., 2014), and the rest of the analysis was focused only on the cleaned time courses of the RSNs. We next computed the L2-regularized partial correlation (ridge regression with ρ = 0.01; Smith et al., 2011, 2015) between each pair of time courses, for each run, with the goal of solely focusing on direct network connections. Then, the partial correlation values were converted into Fisher's z-transformed values considering the temporal smoothness of the data, and the network matrices were averaged across five runs to obtain one network matrix per subject.
For the statistical analysis, the BOLD signal at each voxel was modeled with the general linear model (Friston et al., 1994), applied using permutation-based nonparametric testing, correcting for multiple comparisons. The p-values were family wise error rate corrected for multiple comparisons and contrasts (Winkler et al., 2014, 2016) in an analysis subjected to 5000 permutations. The difference in L2-regularized partial correlation values of each pair of networks between the two time points (1 month interval) were calculated for the drinking (Cohort 1) and control (Cohort 2) groups, and were compared using a two-way ANOVA followed by a Bonferroni's multiple-comparison test. In addition, longitudinal comparisons between naive and alcohol-exposed states were performed with a paired t test in Cohort 1. Finally, the effect of abstinence was further tested in Cohort 1 with a paired t test between-network rs-FC of the nine msP rats in alcohol-exposed versus abstinence conditions.
MEMRI image preprocessing.
All images were preprocessed with custom-developed MATLAB functions (version R2014a; MathWorks) and FSL 5.07 tools (Oxford Center for Functional MRI of the Brain; Smith et al., 2004; Jenkinson et al., 2012). First, the MEMRI images were converted to NIfTI format, scaled up by a factor of 10 and brain extracted (Smith, 2002). Then, these T1-weighted images were coregistered and further registered to a stereotaxic rat brain MRI template (Schwarz et al., 2006) by a 12-parameter affine transformation (Jenkinson et al., 2002). This template is coregistered to a digitized atlas (Paxinos and Watson, 2007), which enables atlas-based generation of region of interest (ROI) masks for detailed anatomic analysis. The resulting images were smoothed to improve SNR using a Gaussian kernel with 4 × 4 × 4 mm3 FWHM. Afterward, the percentage of voxels that showed a statistically significant increase in signal intensity per ROI was calculated.
MEMRI data analysis.
For identifying the brain regions in which activation in the alcohol-drinking group differed from that in the control group, voxel-wise independent t tests were performed in SPM8 (Wellcome Trust Center for Neuroimaging, Institute of Neurology, University College London, London, UK). An initial arbitrary significance threshold was set to p < 0.01 (uncorrected) for individual voxels, and then a cluster size threshold was applied to correct for multiple comparisons at p < 0.05. The cluster size threshold value was determined using Monte Carlo simulation. More restrictive statistical thresholds were also tested (see Results). Finally, the voxels with significant differences in activity across control and alcohol conditions were grouped into brain regions corresponding to the RSN found with fMRI, and the percentage of activated voxels within each one was calculated.
Results
Resting-state networks
During the 1 month of alcohol access (Fig. 1A), animals had an average alcohol consumption of 5.8 ± 1.4 g/kg/d. To assess changes in resting-state connectivity induced by this level of alcohol consumption, we first used group-PICA to define data-driven and robust components from the rs-fMRI data in the longitudinal imaging experiment (Fig. 1B). We found seven distinct RSNs that we named according to the dominant brain regions represented in the network. In some cases (see below), the identified networks resembled those found in humans and rodents under operational/cognitive definitions, such as the salience, the reward, or the default-mode networks. We found a striatal network (StrN), composed of the caudate putamen, nucleus accumbens (NAc), and the ventral pallidum, also invading a portion of the dorsal hippocampus (HC). The sensory-cortex network (SCN), which included mainly posterior insular cortex, primary somatosensory cortex (S1), and secondary somatosensory cortex (S2), and invaded some primary motor cortex regions, with S1 being the region with the highest IC values. The motor-retrosplenial network (MRN), formed by primary and secondary motor cortices and retrosplenial cortices, invading small portions of S1 and S2. The medial prefrontal-cingulate network (mPCN) is principally composed by the anterior cingulate cortex and the prelimbic cortex, and contributed by the infralimbic cortex, lateral orbital cortex, and the anterior insula (AI). The parietal-association network (PAN) is centered in the parietal cortex (Pa) and extends into some regions of S1 and S2. The occipital network (OccN), primarily centered on the primary visual cortex (V1) and secondary visual cortex, spreading into the HC, S1, and laterodorsal thalamic nucleus. And the temporal-cortex network (TCN), which included the temporal association cortex, primary and secondary auditory cortices, and part of V1.
Network connectivity
We first investigated FC in the brain-wide network of RSNs using multiple multivariate linear regression (see Materials and Methods), differentiating among global, specific (partialized), and indirect interactions. Global interaction after 1 month, accounting for the full relation among all pairs of RSNs (comparable to full correlation), was found to be significantly higher in alcohol-drinking compared with control (water-drinking) animals (two-tailed Wilcoxon rank-sum test, p = 0.0031; Fig. 2A). While all RSNs showed an overall increase in connectivity in the alcohol condition, the magnitude of this change was heterogeneously distributed among individual networks (Fig. 2B).
The heterogeneity of the effects of alcohol on RSN connectivity was even more evident at the level of individual links (Fig. 3A). For instance, the full connectivity of the StrN was significantly increased only with the mPCN and MRN, but not with the other networks. With these results, we speculated that the specific information shared by the RSNs might have changed after chronic alcohol drinking. We therefore computed direct interactions in the network, as the partial relation between two RSNs that cannot be explained by any other RSNs (comparable to partial correlation; Fig. 3B). Complementarily, we also computed the information between two RSNs that is also contained in a third RSN (indirect relation; Fig. 3C). Interestingly, while most partial interactions were stable across conditions, we found that the balance of specific pair-wise interactions between some RSNs was dramatically changed in the alcohol group (Fig. 3B). Furthermore, contrary to the consistent increase in full relations, specific pair-wise interactions could increase or decrease. Significant increases were found between the StrN and mPCN (p < 0.025 based on surrogate measures; see Materials and Methods) and SCN-PAN (p < 0.025), and decreases between the StrN-OccN (p < 0.025) and MRN-PAN (p < 0.05). Once again, these differences were not found in the control group (Cohort 2), which showed no significant differences in partialized connectivity between the two time points spaced by 1 month, but for a small but significant increase in the StrN–SCN coupling (Fig. 4). Consistent with an increase in global FC (Fig. 2) not matched by a comparable increase in direct interactions, the indirect FC was also increased by alcohol drinking (two-tailed Wilcoxon matched-pairs signed-rank test, p < 0.0001; Fig. 3C). Particularly important was the increase among the MRN, SCN, and PAN networks, suggesting a functional clustering driven by alcohol consumption.
These analyses demonstrated that 1 month of alcohol drinking induces a global increase in functional coupling between RSNs, mainly explained by an increase in global versus specific interactions, thus increasing the redundancy of the information shared in the network. Importantly though, the analysis further showed a redistribution in the equilibrium of functional connections, resulting in a change in the proportion of specific pair-wise interactions between defined RSNs. It has been suggested, however, that when the number of observations per network node is limited, as often occurs in fMRI experiments, an efficient method to estimate the complete set of direct interactions is to compute a partial correlation with a Lasso regularization (Smith et al., 2011). Therefore, we next set out to validate our findings with L2 regularized partial correlation computed at the two time points, TP1 and TP2, in water-drinking controls and alcohol-exposed animals, and testing the differences (TP2 – TP1) between both groups. We found strong statistical interaction between RSNs and group (two-way ANOVAL F(20,714) = 5.9, p < 0.0001). Post hoc Bonferroni's multiple-comparisons test identified a significant decrease in connectivity between the StrN and OccN (t(714) = 3.1, p = 0.046), and increased connectivity between the StrN and mPCN (t(714) = 7.1, p < 0.0001) and the PAN and SCN (t(714) = 3.2, p = 0.03; Fig. 5). These results support the findings using multivariate regression models.
We next investigated whether the FC alterations persisted or recovered after the cessation of alcohol drinking. In a group of animals, we interrupted the access to alcohol for 1 week, acquiring a new imaging dataset after this period and computing full, partial, and indirect relations in the network of RSNs (Fig. 6). We found no statistically significant differences in the abstinence condition versus alcohol drinking. Similarly, L2 regularized partial correlation between RSNs during early abstinence showed no significant differences with respect to the alcohol-drinking condition (all corrected p-values > 0.05). However, since these analyses were performed in a subgroup of animals, the reduction in the sample size in the abstinence group has to be considered when interpreting this result. Overall, these results suggest the idea of an alcohol-driven reorganization of between-networks FC that sets a new equilibrium of specific interactions that persists in the initial stage of abstinence.
Brain activity measured by MEMRI is increased by alcohol drinking in awake rats
To directly assess regional brain activity in awake rats, we used two different protocols of MEMRI. In the first protocol, MnCl2 was administered within a time window of 7 d using osmotic minipumps and produced no detectable effects on the behavior of the animal, including motor activity, and did not interfere with the amount of daily alcohol intake (4.4 ± 0.6 vs 4.4 ± 0.5 g/kg/d in the 7 d before vs after MnCl2 infusion, respectively; two-tailed paired t test, t(13) = 0.21 p > 0.05), as previously shown (Eschenko et al., 2010a,b; Dudek et al., 2015, 2016; Pallarés et al., 2021). In the second protocol, the same dose of MnCl2 (80 mg/kg, i.p.), administered acutely, produced an observable decrease in motor activity (Eschenko et al., 2010a,b) and a 31.5% reduction in alcohol intake (from 4.6 ± 0.7 to 2.8 ± 0.7 g/kg) over the 24 h post-MnCl2 administration period (two-tailed paired t test, t(7) = 2.61 p = 0.04). Daily consumption in the three groups exposed to alcohol, the rs-fMRI Cohort 1 and the two MEMRI experiments (before MnCl2 administration), was undistinguishable (ANOVA: F(2,37) = 0.47, p = 0.63).
Importantly, both manganese regime administrations produced comparable results. Statistical parametric maps for both MEMRI experiments are presented in Figure 7 and demonstrate significant (cluster size corrected, p < 0.05) increases of T1 signal intensity in the alcohol-drinking rats compared with the control conditions. More conservative cluster-size corrections yielded no statistically significant results. However, the replication of the MEMRI results in two independent experiments, and with brain-wide statistics, demonstrated the robustness of the findings. Additional experiments increasing the MnCl2 administration dose to increase the signal-to-noise ratio were not considered because of the associated toxic effects (see above). No decreases in T1 signal were found to be statistically significant. An increase in T1 intensity reflects the accumulation of Mn2+ into brain tissue and suggests increased brain activation (Fig. 7). Activity increases were largely confined to prefrontal cortical regions, mostly the orbitofrontal, anterior insular, prelimbic, and infralimbic regions, and cingulate cortices (corresponding to the mPCN in rs-fMRI). Less extensive but significant activations were also found in primary and secondary motor regions and retrosplenial cortex (MRN) and S1 (SCN). Subcortically, the caudate–putamen and ventral striatum (StrN) showed statistically significant signal increases. Olfactory areas such as the olfactory bulb, anterior olfactory nucleus, and piriform cortex were also identified. The time course of the MEMRI experiment with acute manganese administration better matched the time course of the fMRI experiment, in which FC is measured in the last 24 h of alcohol drinking but had behavioral side effects. On the other hand, the administration with osmotic minipumps results in the integration of activity over the last 7 d of the alcohol-drinking period but did not interfere with the behavior of the animals. Both together, yielding highly comparable activation maps, produced a robust result showing activations in four of seven RSNs, with the strongest activations found in the mPCN followed by StrN and MRN (Fig. 7D). Neuronal activity as measured by MEMRI did not provide evidence in support of brain regions downregulating their activity after 1 month of chronic alcohol drinking.
Discussion
We show that a relatively brief period of chronic excessive alcohol drinking is enough to induce alterations in rs-fMRI connectivity and MEMRI brain activity. These alterations were age independent and supported by findings both in anesthetized and awake conditions, respectively. By combining both imaging techniques, we show a convergent increase in activity and altered connectivity in prefrontal cortical and ventral striatal structures. In the cortex, the AI, prelimbic, infralimbic, and cingulate cortices accumulate the strongest evidence of alteration driven by alcohol drinking. Quantifying the longitudinal transformations from an alcohol-naive state to a chronic drinking situation, we highlight two important adaptations. First, there is an early increase in redundant activity shared by all RSNs that decreases the specific information content in the system. We called this effect network dedifferentiation, as the activity becomes more similar across all network nodes. Second, the increase in correlated activity is heterogeneous, causing the rearrangement of pair-wise interactions. Reminiscent of a system that finds its equilibrium away from the original homeostatic state, the new highly correlated equilibrium allocates specific interactions between pairs of RSNs differently. We called this effect network narrowing, since the new equilibrium preserves most of the pair-wise interactions but retunes a specific set of connections, potentiating some and depressing others. These alterations persisted after drinking cessation, further suggesting an alcohol-induced allostatic load in the brain (Koob and Moal, 1997; Koob, 2003; McEwen, 2019) that does not recover during the first week of alcohol withdrawal.
The strength and novelty of our findings rely on the following three fundamental aspects of the study design: (1) the use of an animal model with demonstrated translational validity (Ciccocioppo et al., 2006; Hansson et al., 2006; Hirth et al., 2016; De Santis et al., 2019, 2020); (2) the use of two convergent imaging modalities that reinforce each other and allow the investigation of neuronal activity and connectivity, noninvasively and in the entire brain; and (3) the within-subject longitudinal design, possible in animal models, which is optimal for identifying the transformations during the trajectory from a naive state to a chronic alcohol-drinking state. In human studies, cross-sectional designs are the standard, with the control group used to contrast brain changes in AUD patients commonly composed by social drinkers. Furthermore, by using the animal model we avoid the comorbidities and heterogeneous drinking habits normally associated with AUD patients, allowing causal inference on the effects of alcohol.
Brain functional connectivity
The RSNs found in our work are consistent with those found in other rodent fMRI studies (Hutchison et al., 2010; Becerra et al., 2011; Lu et al., 2012; Tsai et al., 2020) and are common to RSNs found in humans (Smith et al., 2009). The first observation was a predominant, although heterogeneous, increase in the interactions among all RSNs at an early stage of chronic alcohol drinking. This observation, together with the increase in brain activation found with MEMRI, points to an initial hyperactive and hyperconnected state induced early by chronic alcohol drinking. These results resonate with some rs-fMRI findings in long-term AUD patients (Zhu et al., 2017) and contrast with others (Vergara et al., 2017). However, finding a translational counterpart for our observations in the literature might be difficult because they were obtained from a longitudinal experiment, which is not feasible in AUD patients, and were focused on brain-wide interactions in an early stage of chronic alcohol drinking, which is not commonly investigated in clinal studies. Therefore, this initial hyperactive and hyperconnected state might have been overlooked in clinical studies. Supporting this view, a study in alcohol-dependent and abstinent mice using the expression of the immediate early gene c-fos as a readout of neuronal activity, found widespread increases in correlated activation producing a decrease in brain network modularity (Kimbrough et al., 2020). Decreased modularity can be equated here with network dedifferentiation, which translates into less compartmentalized communication channels. Furthermore, as mentioned before, clinical studies suffer from various potential confounds, including somatic and psychiatric comorbidities and different patterns of abuse, likely contributing to the large variability of the results found in fMRI studies. For instance, a study with 188 subjects (including drinking, smoking, and drinking-and-smoking subjects) that found network alterations suggesting network hypoconnectivity, also showed that the concurrent use of nicotine and alcohol affects brain connectivity, requiring careful consideration of interaction effects (Vergara et al., 2017). Another recent study investigating network interactions globally in AUD patients, rather than seed-based studies or studies with a priori selection of ROI, has shown network alterations consistent with hyperconnectivity, with participation of the salience, default mode, executive, and reward networks (Zhu et al., 2017). Overall, these results highlight the importance of translational studies involving well controlled animal models to establish causal relations and investigate underlying mechanisms in AUD.
For the pair-wise alcohol-induced and abstinence-persistent changes in FC identified in our study, we can also find translational counterparts in separate human studies. The increase in the specific interaction between mPCN and StrN confirms the finding of increased resting-state FC between the anterior cingulate cortex and the striatum found in a seed-based analysis in 27 abstinent alcoholic subjects (16 ± 12.8 weeks of abstinence) versus 26 age-matched control subjects (Müller-Oehring et al., 2015). Similarly, the shift in maximal FC strength of the orbitofrontal cortex toward the ventral striatum, was found in another resting-state study comparing 66 AUD patients with 40 healthy control subjects (Gerchen et al., 2019). Peak strength connectivity in this study correlated with alcohol dependency and the urge to consume alcohol. Using between-RSNs connectivity as features in a machine learning platform, a recent study showed that pair-wise connectivity between the reward network (largely involving our StrN) and the executive control and default mode networks (with key structures represented in the mPCN) conveyed the maximal information to classify AUD patients (n = 46 vs 46 healthy control subjects; Zhu et al., 2018). Finally, the anterior insular cortex, an important region of the salience network and contained in our mPCN, was highlighted in several AUD fMRI studies, with a postulated role in the dysregulation of interoceptive awareness in drinkers (Orban et al., 2013; Vergara et al., 2017; Zhu et al., 2017; Halcomb et al., 2019; Bordier et al., 2022; for review, see Sommer et al., 2022). Strengthened connectivity within an amygdalar–striatal network, compatible with the increased mPCN–StrN-specific interaction, was found in a cohort of 25 alcohol-dependent patients and 26 healthy control subjects, with a salience network correlating with impulsivity measured in the delay discount task (Zhu et al., 2017).
The decrease in FC between OccN and SCN matches the visuomotor hypoconnectivity found in social drinkers after acute alcohol consumption (Luchtmann et al., 2013) and, in AUD patients, connectivity between the occipital gyrus and the supplementary motor area negatively correlated with AUDIT (Alcohol Use Disorders Identification Test) values (Fede et al., 2019). The finding in social drinkers confirmed that the visuomotor pathway is affected early during alcohol drinking, and further indicated that alterations in these networks may persist even during early abstinence. Furthermore, a recent study in a cohort of adults with problem drinking patterns (n = 59) showed that connectivity with visual, visual association, and sensorimotor regions was important for predicting AUD severity (Fede et al., 2019). Another rs-fMRI study with 43 volunteers diagnosed with AUD and 26 healthy control subjects using the striatum as the seed region (Kohno et al., 2017) found increased FC with the anterior cingulate cortex in the AUD group and decreased connectivity with the lateral occipital cortex, equivalent to the increased StrN–mPCN and decreased StrN–OccN connectivity, respectively, found in our present study. These studies pointed to relapse vulnerability because of craving dysregulation reflected in the increased connectivity in striatolimbic regions and decreased corticostriatal connectivity. Our preclinical study supports this interpretation and further indicates that these changes can be directly linked to alcohol, starting soon after chronic excessive drinking begins, and propagate into early abstinence.
Brain activity
MEMRI allowed us to assess the effects of chronic alcohol consumption on global brain activity in freely behaving animals. MEMRI activations were mostly localized in the striatum and prefrontocortical structures. Activity in the ventral striatum occurs during reinforced learning and has long been associated with the rewarding effects of drugs of abuse, including alcohol (Koob et al., 1998). Comparable evidence supporting the striatal activation in different rat models of alcohol consumption have been reported using 2-deoxyglucose uptake (Porrino et al., 1998) or c-Fos immunostaining (Bachtell et al., 1999; Hansson et al., 2007; Vilpoux et al., 2009), and alcohol consumption was shown to evoke dopamine release in the accumbens of sP rats, the parental strain of msPs (De Montis et al., 2004). In humans, ventral striatal activation has been consistently found to be associated with the presentation of alcoholic beverages or related sensory cues (Myrick et al., 2004; Filbey et al., 2008; Gilman et al., 2008; Vollstädt-Klein et al., 2020).
Brain regions in the basic motivational circuit have also been consistently found in imaging studies with AUD patients, including medial and orbital regions of the prefrontal cortex (Spanagel, 2009) and insular cortex (Naqvi and Bechara, 2010), which were also found in our MEMRI maps. Activation of the insular cortex and ventral striatum in response to alcohol cues further correlated with drinking levels in humans (Ihssen et al., 2011). Activations in both the prelimbic and infralimbic subdivisions of the medial prefrontal cortex were found in a postdependent rat model of alcohol addiction (Meinhardt and Sommer, 2015), likely associated with the alcohol-dependent phenotype. It is important to note that both structures are part of the mPCN identified in our rs-fMRI study, which showed enhanced direct interaction with the StrN; that is, both structures and the NAc are more active (MEMRI) and functionally coupled (rs-fMRI) after chronic alcohol drinking, with a likely effect on the reward system (Haber and Knutson, 2010). In msP rats, with an innate anxiety-like behavior (Ciccocioppo et al., 2006; Borruto et al., 2021), the infralimbic connection may contribute to the anxiolytic effects of alcohol, whereas the medial prefrontal cortex activation could mediate alcohol-seeking behavior (Peters et al., 2009; Meinhardt et al., 2013; Pfarr et al., 2015, 2018). We have previously hypothesized that msP rats are a phenocopy of a postdependent alcohol condition in which alcohol is consumed as an attempt to self-medicate from high background levels of stress (Hansson et al., 2006; Björk et al., 2010; Herman et al., 2013; Hirth et al., 2016). In this context, the combined rs-fMRI and MEMRI results may be interpreted as an allostatic network reorganization induced by alcohol to compensate a preexisting dysfunction that leads to the chronification of alcohol consumption.
Study limitations
Some limitations of the present study need to be considered. First, for the reasons stated in the introduction, fMRI experiments in awake rat AUD models present important drawbacks, and, thus, the RSN activity was extracted from lightly anesthetized animals. While previous studies showed robust RSNs in anesthetized versus awake conditions (Barttfeld et al., 2015; Díaz-Parra et al., 2017; Paasonen et al., 2018), important network dynamics could have been masked. This limitation was avoided in the MEMRI study, in which manganese accumulation occurs in awake animals, although it reflects neuronal activity rather than connectivity. On the other hand, the sensitivity of the MEMRI technique to map brain activity relies on manganese ion concentration in the extracellular space, but increasing it over certain limits may result in toxic effects (Eschenko et al., 2010a,b). To overcome this limitation and the trade-off of sensitivity and toxicity, we used low doses of MnCl2 and two different regimes of administration, as previously described (Eschenko et al., 2010a). The obtained results were highly robust across protocols. However, subtler alterations in brain activity may have gone undetected by these protocols. An important consideration is also the complexity of AUD, resulting from several genetic, environmental, and personality factors, and therefore is hardly represented by a single animal model. The msP line, in addition to showing a preference for alcohol, presents high stress sensitivity and a propensity to negative affect (Borruto et al., 2021), and therefore, it more likely resembles a subpopulation of AUD patients that drink to alleviate negative mood and for stress-relieving purposes. More generally, the 1 month period of drinking in rats is not expected to mimic all the effects of alcohol consumption in long-term AUD patients, but more likely the early transformations induced by chronic drinking. The translational validity of the model, however, has been validated in several recent studies (De Santis et al., 2019, 2020). Finally, the present study was performed in males. However, important sex differences in alcohol exposure have been found (Agabio et al., 2017; Fama, 2020) that warrant further work comparing RSN connectivity and brain activity in males and females.
Concluding remarks
To conclude, we combined rs-fMRI and MEMRI in msP rats engaged in chronic moderate drinking, to investigate brain adaptations induced by alcohol drinking and withdrawal. Alcohol drinking led to a global but heterogeneous increase in BOLD signal correlation and a concomitant increase in activity in frontal and prefrontal cortical regions and striatal areas. This new network configuration relays communication through indirect pathways, favoring global over local information processing. We hypothesize that increased neuronal excitability is responsible for both enhanced neuronal activation and functional coupling in RSNs. We further hypothesize that the increased redundancy in network interactions (dedifferentiation) produced by the alcohol-driven hyperconnected state compromises cognitive function. Because of its heterogeneous nature, the global increase in FC produces a second brain adaptation that involves the readjustment of a specific set of connections between pairs of RSNs, remodeling specific pathways and narrowing communication channels. Importantly, these alterations persisted after alcohol discontinuation, during abstinence, suggesting that the brain network reaches a new functional equilibrium after a relatively brief period of chronic alcohol consumption that does not recover spontaneously, at least during the early stage of abstinence. Whether this new state contributes to the vulnerability to relapse needs to be investigated. Initial studies involving chemogenetic manipulation of distinct network nodes (Haaranen et al., 2020) suggest that those manipulations involve highly complex networks and that their behavioral outcomes are difficult to predict. Future research using animal models is needed to understand this transition to develop therapies aiming to recover the network homeostasis.
Footnotes
This research was supported by funding from the European Union Horizon 2020 research and innovation program under Grant 668863 (SyBil-AA); the ERA-Net NEURON framework with grants given to the project “Translational Neuroimaging in Alcoholism” (TRANSALC) by the Academy of Finland (Grant TRANSALC 01EW1112), the Bundesministerium für Bildung und Forschung (Grant FKZ01EW1112), and the Spanish Ministerio de Ciencia e Innovación (Grant PIM2010ERN-00679). S.C. and D.M. were funded by the Spanish Ministerio de Economía y Competitividad (MINECO) cofinanced by the European Regional Development Fund (ERDF) under Grants BFU2015-64380-C2-1-R and BFU2015-64380-C2-2-R, and the Generalitat Valenciana Government through the Prometeo Program (Grant PROMETEO/2019/015). S.C. was further supported by the Spanish Agency of Research (AEI) and ERDF under Grant PGC2018-101055-B-I00 and the Spanish Ministerio de Sanidad, Servicios Sociales e Igualdad (Grant 2017I065). Ú.P.-R. was supported by Grant FPU13/03537 and research stay internship EST15/00516 from the Spanish Ministerio de Educación, Cultura y Deporte (MECD). Further funding was obtained from the Spanish State Research Agency through the Severo Ochoa Program for Centres of Excellence in R&D (Grant SEV- 2017-0723). W.H.S. was further supported by the Deutsche Forschungsgemeinschaft (Center Grant TRR265-B08). We thank Eugene Duff and Stephen Smith for comments on a previous version of this manuscript. We also thank Begoña Fernández for technical assistance.
The authors declare no competing financial interests.
- Correspondence should be addressed to David Moratal at dmoratal{at}eln.upv.es or Santiago Canals at scanals{at}umh.es