CREB‐BDNF pathway influences alcohol cue‐elicited activation in drinkers

Abstract

Alcohol use disorder (AUD) is suggested to have polygenic risk factors and also exhibits neurological complications, strongly encouraging a translational study to explore the associations between aggregates of genetic variants and brain function alterations related to alcohol use. In this study, we used a semiblind multivariate approach, parallel independent component analysis with multiple references (pICA‐MR) to investigate relationships of genome‐wide single nucleotide polymorphisms with alcohol cue‐elicited brain activations in 315 heavy drinkers, where pICA‐MR assesses multiple reference genes for their architecture and functional influences on neurobiological conditions. The genetic component derived from the cAMP‐response element‐binding protein and ‐brain derived neurotrophic factor (CREB‐BDNF) pathway reference was significantly associated (r = −0.38, P = 3.98 × 10⁻¹²) with an imaging component reflecting hyperactivation in precuneus, superior parietal lobule, and posterior cingulate for drinkers with more severe alcohol dependence symptoms. The highlighted brain regions participate in many cognitive processes and have been robustly implicated in craving‐related studies. The genetic factor highlighted the CREB and BDNF references, as well as other genes including GRM5, GRM7, GRID1, GRIN2A, PRKCA, and PRKCB. Ingenuity Pathway Analysis indicated that the genetic component was enriched in synaptic plasticity, GABA, and protein kinase A signaling. Collectively, our findings suggest that genetic variations in various neural plasticity and signaling pathways partially explain the variance of precuneus reactivity to alcohol cues which appears to be associated with AUD severity. Hum Brain Mapp 36:3007–3019, 2015. © 2015 Wiley Periodicals, Inc.

INTRODUCTION

Alcohol use disorders (AUD) present a substantial health and economic issue. The lifetime prevalence is estimated to be 17.8% for alcohol abuse and 12.5% for alcohol dependence [Hasin et al., 2007]. Genetic factors have been shown to affect liability to AUD, with the heritability estimated to be 40–60% while the remainder of variance is likely attributable to environmental factors [Mayfield et al., 2008; Prescott and Kendler, 1999]. Great efforts have been made toward unraveling the genetic etiology of AUD. Candidate gene and unbiased genome‐wide association studies (GWAS) provided evidence for a number of susceptibility variants, highlighting genes involved in various neural signaling pathways, including dopaminergic [Conner et al., 2005], glutamatergic [Schumann et al., 2008], and GABAergic [Bierut et al., 2010] systems. Genes encoding alcohol dehydrogenase (ADH) enzymes playing a key role in alcohol metabolism are also implicated in AUD vulnerability [Luo et al., 2007].

Despite the growing knowledge on susceptibility loci contributing to individual differences in drinking behavior, the genetic findings generally have modest effect sizes at best [Heath et al., 2011; Kapoor et al., 2013]. For instance, in a large GWAS of alcohol dependence where thousands of participants were included for investigations [Bierut et al., 2010], no single nucleotide polymorphism (SNP) could pass the genome‐wide significance threshold of 5.00 × 10⁻⁸ [Stranger et al., 2011]. Instead the highlighted 15 SNPs yielded suggestive associations with P < 10⁻⁵, yet none of them replicated findings of a previous GWAS [Treutlein et al., 2010]. This is a common challenge in complex trait mapping. Indeed, like many other complex disorders, AUD is suggested to have polygenic risk factors [Johnson et al., 2006; Salvatore et al., 2014], such that the underlying genetic architecture involves many variants with modest individual effect sizes, which may interact to confer the liability. Understanding the mechanism of genetic effects becomes even more complicated due to phenotypic heterogeneity where genetic variants can exert influences on various phenotypes through different biological mechanisms [Wong and Schumann, 2008].

In this work, we used a novel semiblind multivariate approach, parallel independent component analysis with multiple references (pICA‐MR) [Chen et al., 2014], to investigate the genetic basis underlying brain function related to AUD. Specifically, functional magnetic resonance imaging (fMRI) data were collected from participants exposed to a well‐established alcohol cue paradigm [Claus et al., 2011; Filbey et al., 2008]. The method, pICA‐MR is an extension of pICA‐R [Chen et al., 2013] to assess multiple genetic references in one single analysis. PICA‐MR extracts independent components from the imaging and genetic modalities separately to assess aggregate effects of multiple variables, posing a promising model for polygenicity. It also enhances intermodality associations, providing a translational framework for exploring genetic underpinnings of neuronal functions, which might ultimately lead to clinical manifestations of the disorder. In addition, genetic references (i.e., multiple sets of SNPs) are incorporated to pinpoint components of particular attribute and pICA‐MR delineates the references' architecture and functional influence in a data‐driven manner. This method is expected to help characterize mechanisms of interest in high‐dimensional complex data.

MATERIALS AND METHODS

Participants

A total of 326 participants enrolled in the study to investigate genetic and neurobiological traits related to AUD [Claus et al., 2011]. The University of New Mexico Human Research Review Committee approved the study. All the participants were recruited from the greater Albuquerque metropolitan region and provided written informed consent. The inclusion criterion was based on alcohol consumption, requiring participants to drink at least five times in the past month with at least five (for men) or four (for women) drinks per drinking occasion. The exclusion criteria included a history of severe alcohol withdrawal, brain‐related medical problems, or symptoms of psychosis. In addition, participants were required to be sober during the data collection, with a breath alcohol concentration of 0.00. Sixty‐one and eighty percentage of participants met Diagnostic and Statistical Manual of Mental Disorders Fourth Edition (DSM‐IV) criteria for current alcohol dependence and alcohol abuse, while 70 and 93% met lifetime criteria for alcohol dependence and alcohol abuse. After preprocessing, 315 participants (220 males and 95 females) for which good quality fMRI and SNP data were collected, were included in the analysis. Table 1 provides the demographic information.

Table 1.

Demographic information of participants

Number of participants
	Male (220)		Female (95)
Caucasian	99		43
African American	4		2
Asian	2		0
Latino	54		28
Native American	13		3
Mixed	47		19
Unreported	1		0
Age of participants
	Male		Female
Mean ± SD	31.74 ± 9.43		32.52 ± 10.58
Range	21–56		21–55

Data Collection and Preprocessing

Behavioral assessment

Participants completed a variety of questionnaires at a baseline assessment appointment, including the Alcohol Dependence Scale (ADS) [Skinner and Horn, 1984], the Alcohol Use Disorder Identification Test (AUDIT) [Babor et al., 2001], and the Impaired Control Scale (ICS) for alcohol [Heather et al., 1998]. Symptom counts for current/lifetime alcohol dependence/abuse were tallied from the Structured Clinical Interview for DSM Disorders. Years of regular drinking, age at first drink, and age that regular drinking began were derived from a drinking history questionnaire. Depression was assessed using the Beck Depression Inventory. We excluded relatively incomplete measures where data were missing for more than 25 participants. Finally, a total of 23 behavioral measures were investigated for associations with identified imaging and genetic components, as listed in Table 2. The missing ratio was no greater than 4/315. It should be noted that most of these behavioral measures showed significant associations with age, except for AgeFirstDrink, EStress‐tot, and Stress‐tot.

Table 2.

Alcohol dependence assessment

Assessment	Sub‐category	Description	Score range	Score mean ± SD
ADS	ADS‐con	Loss of behavior control	0–18	7.87 ± 4.16
	ADS‐obs	Obsessive drinking style	0–5	1.07 ± 1.33
	ADS‐per	Psychoperceptual withdrawal	0–13	2.00 ± 2.34
	ADS‐phy	Psychophysical withdrawal	0–10	2.63 ± 2.16
	ADS‐tot	Total ADS	1–43	13.57 ± 8.24
AUDIT	AUDIT‐tot	Total AUDIT score	4–39	19.02 ± 7.75
	AUDIT‐consump	Alcohol consumption total	0–12	9.21 ± 2.06
	AUDIT‐dep	Alcohol dependence total	0–12	3.78 ± 3.04
	AUDIT‐probs	Alcohol problems total	0–16	5.96 ± 3.99
ICS	ICS‐tot	Total ICS	0–96	45.65 ± 21.24
	ICS‐ac	Attempted control	0–20	8.13 ± 4.90
	ICS‐fc	Failed control	0–40	19.12 ± 9.64
	ICS‐pc	Perceived control	0–40	18.40 ± 9.67
Alcohol symptom count	PA‐count	Past alcohol abuse symptom count	0–4	2.01 ± 1.24
	CA‐count	Current alcohol abuse symptom count	0–4	1.66 ± 1.19
	PD‐count	Past alcohol dependence symptom count	0–7	3.46 ± 2.21
	CD‐count	Current alcohol dependence symptom count	0–7	3.50 ± 2.44
Drinking history	AgeDrink	Probable age that regular drinking first occurred	11–39.75	18.96 ± 4.40
	YearsDrink	Probable number of years of regular drinking	1–41	13.01 ± 9.17
	AgeFirstDrink	Probable age of first drink	4–28	14.39 ± 2.88
Stress	EStress‐tot	Total early stress for ages before 19	0–12	3.04 ± 2.61
	Stress‐tot	Early stress total (all ages reported)	0–12	3.46 ± 2.77
	BDI‐tot	Total Beck Depression Inventory	0–47	11.65 ± 9.81

Functional MRI

Brain activation data were collected from 317 participants during an alcohol craving task as described in [Claus et al., 2011; Filbey et al., 2008], see Supporting Information Figure S1 for the schematic of a single taste cue trial. Participants were exposed to 1 ml of alcoholic (individual preferred) or juice (litchi) beverages pseudorandomly presented during the MRI scans. Each taste cue trial sequentially consisted of a 2‐s “Ready” prompt, a 24‐s taste cue presentation, and a 16‐s washout period. During the cue presentation, participants tasted the presented beverage (second 1–10 and 12–22) and then swallowed (second 10–12 and 22–24). No stimuli were presented during the washout and participants viewed the word “Rest”. Two 9‐min runs were conducted for each participant, with a single run spanning 12 trials, six for each tastant. A 3T Siemens Trio was used for the data collection. The echo‐planar gradient‐echo pulse sequence was configured as follows: TR = 2 s, TE = 29 ms, flip angle = 75°, voxel size = 3.75 × 3.75 × 4.55 mm³. The collected fMRI data were preprocessed with Statistical Parametric Mapping 5 (SPM5, http://www.fil.ion.ucl.ac.uk/spm). Standard motion correction was performed and images were normalized to the Montreal Neurological Institute template [Jenkinson et al., 2002] and resliced to 3 × 3 × 3 mm³. An 8‐mm full‐width half‐maximum Gaussian kernel was used for spatial smoothing. Finally, alcohol versus juice contrast images were extracted from a total of 54,937 voxels with no missing values and used for subsequent association analyses. One participant outlier was excluded due to a low correlation (>3‐SD) with the mean activation across all the participants.

SNP data

Saliva samples were collected from 324 participants for DNA extraction. Genotyping for all participants was performed at the Mind Research Network using the Illumina Infinium Human 1M‐Duo assay spanning 1,199,187 SNP loci. BeadStudio was used to make the final genotype calls. A series of standard quality control procedures were then performed with PLINK [Purcell et al., 2007]. Specifically, SNPs and participants were first examined for a genotyping rate threshold of 95%; SNPs were excluded if they deviated from Hardy–Weinberg Equilibrium with a threshold of 10⁻⁶ or failed to be missing at random with a threshold of 10⁻¹⁰; Two participants were excluded due to high heterozygosity (3‐SD greater than the mean); Another two participants were excluded due to relatedness with identity‐by‐descent values greater than 0.1875; Minor allele frequency threshold was set to 0.05. After the quality control, discrete numbers were assigned to the categorical genotypes: 0 for no minor allele, 1 for one minor allele, and 2 for two minor alleles. Subsequently, we replaced the missing genotypes using high linkage disequilibrium (LD) loci if available (correlation > 0.80), excluded 18,809 SNPs with a missing ratio greater than 1%, and replaced the rest missing genotypes with the major alleles of individual loci. The resulting 717,129 autosomal SNPs were then used for population stratification correction by principal component analysis (PCA), and three principal components (PC1, 2, and 4) differed significantly among ethnic groups (P = 9.85 × 10⁻⁷⁹, 3.23 × 10⁻⁸⁶ and 3.21 × 10⁻⁵⁵, respectively) while exhibiting no significant associations with drinking behavior measures. Then the data reconstructed from these three ethnicity‐related PCs were subtracted from the original data to eliminate the influence of population structure. In the corrected SNP data, a Q‐Q plot for P‐values of association with the AUDIT‐tot score tested against a uniform distribution showed no clear indication of population structure (Supporting Information Fig. S2).

Association Analysis

The fMRI contrast images were analyzed in conjunction with the SNP data using pICA‐MR [Chen et al., 2014], which extends the pICA [Liu et al., 2009] and pICA‐R [Chen et al., 2013] (Fusion ICA Toolbox, http://mialab.mrn.org/software/fit) approach to accommodate multiple genetic references. Figure 1 shows the flowchart of pICA‐MR. As a multivariate approach, pICA‐MR decomposes the two datasets, X₁ and X₂, into linear combinations of underlying components separately and in parallel, as depicted in Eq. (1). S, A, and W denote the component, mixing, and unmixing matrices, respectively. The subscript d runs from 1 to 2, denoting two data modalities.

$${\mathrm{X}}_d={\mathrm{A}}_d{\mathrm{S}}_d\to {\mathrm{S}}_d={\mathrm{W}}_d{\mathrm{X}}_d,{ \mathrm{A}}_d={\mathrm{W}}_d^{-1}, d = 1, 2$$ (1)

Figure 1.

A flowchart of pICA‐MR.

The data decomposition primarily builds on infomax ICA [Amari, 1998; Bell and Sejnowski, 1995], where independent sets of covarying variables are organized into different components. The mathematical model is described in Eqs. (2) and (3). F ₁ is the objective function of infomax for the imaging modality (modality 1), where independence among components is optimized by maximizing the entropy (H). f _y(Y) is the probability density function of the sigmoid function Y, which is chosen based on the form of the expected underlying source distribution [Lee et al., 1999]. E is the expected value and W₀ is the bias vector. In contrast to F₁, the objective function for the genetic modality (modality 2) F₂, is modified based on Infomax so that components are not only independent but closely resemble the reference matrix r. Each row of r represents a reference vector, which is of the same dimension as the SNP data with nonzero elements representing a group of reference loci likely contributing in a coordinated manner. PICA‐MR then calculates the Euclidean distances between each of the reference vectors r_i (ith row of r) and all the components (only using nonzero reference loci ${\stackrel{\mathrm{\sim }}{\mathrm{r}}}_{\mathrm{i}}$, a subvector of ${\mathrm{r}}_{\mathrm{i}}$) to determine the closest one as the corresponding constrained component, whose distance ( ${‖\left|{\stackrel{\sim }{\mathrm{S}}}_{{2,\mathrm{k}}_i}\right|‐{\stackrel{\sim }{\mathrm{r}}}_i‖}_2^2$) is further minimized. ${\stackrel{\mathrm{\sim }}{\mathrm{S}}}_{{2,\mathrm{k}}_{\mathrm{i}}}$ denotes a subvector of the r_i‐constrained component ${\mathrm{S}}_{{2,\mathrm{k}}_{\mathrm{i}}}$ (the k_i ^th row of S₂) and is computed by ${\mathrm{W}}_{{2,\mathrm{k}}_{\mathrm{i}}}$(the k_i^th row of W₂) multiplying ${\stackrel{\mathrm{\sim }}{\mathrm{X}}}_2$ (a submatrix of X₂). ǁ·ǁ₂ represents the L₂‐norm Euclidian distance, and λ is the weight parameter. Note that pICA‐MR incorporates the reference constraint in a way similar to pICA‐R, where the distance between the component and reference vector is optimized for the reference loci only. This is to enable a semiblind decomposition which is partially constrained to highlight the presumed causal loci in the resulting components while still allowing the remaining loci to show their own importance driven by the data. As in pICA‐R, the weight parameter λ is adaptively adjusted to avoid false positives due to the reference constraint dominating over the independence metric. In the case when λ is completely de‐emphasized, pICA‐MR converges to regular pICA. Conversely, in contrast to pICA‐R dealing with one single reference (vector), in pICA‐MR, multiple reference vectors can constrain the same or different components, reflecting functional architecture in a data‐driven manner. Finally, the intermodality association function F₃ maximizes the correlations computed over the columns of the loading matrices A₁ and A₂.

$$\begin{array}{c}{\mathrm{Y}}_d=\frac{1}{1+e^{-{\mathrm{U}}_d}},{\mathrm{U}}_d={\mathrm{W}}_{\mathrm{d}}{\mathrm{X}}_d+{\mathrm{W}}_{d0}\hfill \\ F_1=\text{max}\left\{\mathrm{H}\left({\mathrm{Y}}_1\right)\right\}=\text{max}\left\{-\mathrm{E}\left[\text{ln} f_{{\mathrm{y}}_1}\left({\mathrm{Y}}_1\right)\right]\right\}\hfill \\ F_2=\text{max}\left\{\lambda \mathrm{H}\left({\mathrm{Y}}_2\right)+\left(1-\lambda \right)\left[-{\sum }_{i=1}^{\mathrm{I}}{\text{dist}}^2\left({\stackrel{\sim }{\text{r}}}_i,\left|{\stackrel{\sim }{\text{S}}}_{2,{\mathrm{k}}_i}\right|\right)\right]\right\}\hfill \\ =\text{max}\left\{\lambda \left(-\mathrm{E}\left[\text{ln} f_{{\mathrm{y}}_2}\left({\mathrm{Y}}_2\right)\right]\right)+\left(1-\lambda \right)\left(-{\sum }_{i=1}^{\mathrm{I}}{\left|\right|\left|{\mathrm{W}}_{2,{\mathrm{k}}_i}{\stackrel{\sim }{\text{X}}}_2\right|-{\stackrel{\sim }{\text{r}}}_i\left|\right|}_2^2\right)\right\} \hfill \end{array}$$ (2)

$$\begin{array}{c}{F}_3=\text{max}\left\{{\sum }_{i,j}{\text{Corr}}^2\left({\mathrm{A}}_{1,i},{\mathrm{A}}_{2,j}\right)\right\}\hfill \\ =\text{max}\left\{{\sum }_{i,j}\frac{{\text{Cov}}^2\left({\mathrm{A}}_{1,i},{\mathrm{A}}_{2,j}\right)}{\text{Var}\left({\mathrm{A}}_{1,i}\right)\text{Var}\left({\mathrm{A}}_{2,j}\right)}\right\}\hfill \end{array}$$ (3)

The three objective functions are solved through iteratively updating the unmixing matrices W₁ and W₂ using gradient optimization. The updating rules are similar to [Chen et al., 2013]. Cautions have been exercised to avoid overfitting and false positive findings. First, a threshold of correlation is imposed such that only component pairs exhibiting above‐threshold correlations are further optimized. In this work, we used a conservative threshold of 0.25 for the sample size of 315, requiring that the component pair showed a significant correlation before the optimization. In addition, an adaptive learning is incorporated through monitoring the components' independence. The optimization of correlation and distance will be depressed if it compromises the metric of independence [Chen et al., 2013; Liu et al., 2009]. Simulations suggest that pICA‐MR effectively identifies similarity among multiple references and the detection power is comparable to pICA‐R and hence significantly improved compared to blind methods [Chen et al., 2014]. It should be noted that currently pICA‐MR is configured to apply references only to the genetic modality, as the fMRI modality is less challenging and in general, the regular ICA is competent. However, the reference constraint can be extended to both modalities if needed.

In this work, the number of fMRI components was estimated by minimum description length (MDL) [Rissanen, 1978] on uncorrelated voxels, yielding the most efficient representation of the original data. MDL proves an effective metric for neuroimaging data as signals of interest in general carry much more variance than noises. However, it is less applicable to SNP data, as most genetic factors account for small amounts of variance, except for those related to population structure. Therefore, we developed an approach to estimate the component number based on components' consistency for SNP data, such that the most stable decomposition could be obtained [Chen et al., 2012].

Six sets of strongly hypothesized genes were selected based on previously reported associations with AUD, as listed in Table 3. Each set consisted of a few genes which might affect brain function in a coordinated manner, and was tested separately. CHRNA3, CHRNA5, and CHRNB4 are located closely in 15q24‐25 and are among the most replicated findings for smoking behaviors [Bierut, 2010; Caporaso et al., 2009] while also implicated in risk of alcohol dependence [Wang et al., 2009]. Cytogenetic band 4p12 hosting GABA receptor genes were shown to present moderate odds ratios for alcohol dependence in a GWAS [Bierut et al., 2010] and independent contributions from individual receptors have been suggested [Enoch et al., 2009]. Another cluster of GABA receptor genes resides in 5q34. Both mouse and human studies provided evidence for their modulatory roles in alcohol dependence [Radel et al., 2005]. ADH is a primary enzyme involved in alcohol metabolism and variants in the gene cluster ADH1A‐ADH1B‐ADH1C have been found associated with risk of alcoholism [Edenberg and Foroud, 2006; Luo et al., 2007]. The opioidergic system is considered as mediating drug‐induced feelings and playing an important role in substance rewarding properties [Gianoulakis, 2001]. Related polymorphisms have been identified as associated with AUD [Filbey et al., 2008b; Zhang et al., 2008]. cAMP‐response element‐binding protein (CREB) is a key transcriptional factor for neuronal growth and regulates the expression of brain derived neurotrophic factor (BDNF). These two genes have been found to interact in a variety of brain regions and play a critical role in addiction [Carlezon et al., 2005; Crews et al., 2007; Pandey, 2003]. A reference matrix was generated for each gene set with each row representing a group of reference SNPs derived from a single gene based on LD structure (r ² > 0.2, [Ripke et al., 2011]), as described in [Chen et al., 2013]. Specifically, in this work, most of the reference genes spanned tens of SNPs forming a single LD block in our data, which were directly used to generate the reference matrix. One exception was CREB5 hosting 228 SNPs, for which multiple LD blocks were identified and represented in multiple reference matrices. The associations of all the reference SNPs with four major behavioral measures (ADS‐tot, AUDIT‐tot, ICS‐tot, and BDI‐tot) were provided in Supporting Information Table S1.

Table 3.

Tested genetic references

Reference	Genes	Literature
15q24‐25	CHRNA3, CHRNA5, CHRNB4	Bierut, 2010; Caporaso et al., 2009; Wang et al., 2009
4p12	GABRA4, GABRA2, GABRG1, GABRB1	Bierut et al., 2010; Enoch et al., 2009
5q34	GABRB2, GABRA6, GABRA1, GABRG2	Radel et al., 2005
4q23	ADH1A, ADH1B, ADH1C	Edenberg et al., 2006; Luo et al., 2007
Opioid system	OPRM1, OPRK1, OPRD1	Filbey et al., 2008; Zhang et al., 2008
CREB‐BDNF	CREB1, CREB5, BDNF	Carlezon et al., 2005; Crews et al., 2007; Pandey, 2003

To assess the fidelity of the identified association, we applied 10‐fold validation and permutation test. In the 10‐fold validation, the participants were divided into 10 subsets. Then in each of the 10 runs, nine subsets were included and pICA‐MR was applied to the 90% of the original data to extract components and evaluate the correlations. This was to confirm that the observed SNP‐fMRI association was consistently presented by all the participants, instead of being raised by specific subsets. In the 1,000‐run permutation test, all the participants were included and permuted in each run and sent for pICA‐MR analysis. Then based on the top associated component pair of each run, we calculated the tail probability to evaluate the significance level of the identified SNP‐fMRI association. This permutation test investigated the possibility of the identified SNP‐fMRI association occurring in randomly rearranged participants. It also helped evaluate overfitting which may manifest as producing inflated correlations from permuted data.

To understand the functional influences, the identified SNP and fMRI components were further investigated for associations with all the behavioral measures listed in Table 2. Multiple regression analysis was conducted to control for sex and population structure. Particularly, the covariates of population structure included both the self‐reported race/ethnicity (Table 1) and the top three PCs extracted from the corrected SNP data [Ripke et al., 2011]. False discovery rate (FDR) was applied to account for multiple tests given the moderate to high correlations among most of the behavioral measures.

RESULTS

The fMRI (315 × 54,937) and SNP (315 × 717,129) data were analyzed with pICA‐MR. The number of components was estimated to be 15 for fMRI and 11 for SNP, respectively. Among the tested gene sets, CREB‐BDNF elicited a significant SNP‐fMRI correlation of −0.38 (P = 3.98 × 10⁻¹²), passing the Bonferroni threshold of 5.05 × 10⁻⁵ which corrected for the six tested genetic references and the combinations of all extracted components (15 sMRI × 11 SNP) in each test. The three genes BDNF, CREB1, and CREB5 were identified to constrain the same SNP component. Supporting Information Table S2 summarizes the recruited reference SNPs, which consisted of all the genotyped loci in BDNF (15 SNPs) and CREB1 (20 SNPs), and an LD block spanning 20 SNPs in CREB5. After regressing out controlling variables (age, sex, population structure), the SNP‐fMRI association remained significant, exhibiting a correlation of −0.35 (P = 2.19 × 10⁻¹⁰), as shown in Figure 2a. The identified SNP‐fMRI pair exhibited stable correlations in the 10‐fold validation, ranging from 0.23 to 0.33 with a median of 0.27. In the 1,000‐run permutation, only one permuted sample exhibited a SNP‐fMRI association stronger than that observed in the original data, yielding a significant P‐value of 0.001 for our finding. Other tested gene sets presented insignificant or marginally significant SNP‐fMRI associations, as summarized in Supporting Information Table S3.

Figure 2.

Scatter plots of: (a) the fMRI and SNP loadings; (b) the fMRI loading and CD‐count. [Color figure can be viewed in the online issue, which is available at http://wileyonlinelibrary.com.]

The identified fMRI component, with FDR control, was significantly associated with a number of behavioral measures, including CD‐count, ICS‐fc, ICS‐tot, and PD‐count, but not YearsDrink. The most significant association was observed from CD‐count, exhibiting a correlation of 0.24 (P = 1.73 × 10⁻⁵) after regressing out sex and population structure, as shown in Figure 2b. In addition, the correlation remained significant (r = 0.18, P = 1.31 × 10⁻³) after age was further regressed out. This, together with the fact that no significant association was observed between the fMRI component and YearsDrink, suggested that the variance in activation reflected more current dependence symptoms rather than cumulative alcohol use. The fMRI loadings also showed significant associations with ICS‐fc, ICS‐tot, and PD‐count (correlations of 0.24, 0.23, and 0.21, respectively), all of which highly correlated with CD‐count (correlations of 0.68, 0.72, and 0.82, respectively). Due to this collinearity, the individual effect could not be disentangled and we chose to focus on the most significantly associated symptom CD‐count in the following discussion. Figure 3 shows the spatial map of the identified fMRI component thresholded at |Z| > 2. Table 4 summarizes the Talairach atlas labels [Lancaster et al., 2000] of the mapped brain network, including precuneus, superior, and inferior parietal lobules, as well as posterior cingulate cortex (PCC). These highlighted brain regions remained highly consistent with a |Z| threshold ranging from 2 to 3.

Figure 3.

Spatial map of brain network for the identified fMRI component (|Z| > 2). [Color figure can be viewed in the online issue, which is available at http://wileyonlinelibrary.com.]

Table 4.

Talairach labels of identified brain regions (|Z| > 2)

Brain region	Brodmann area	L/R volume (cm3)	L/R random effects, max Z (x,y,z)
Precuneus	7, 19, 39, 31	16.8/14.1	9.18(0,−58,61)/9.41(3,−58,64)
Superior Parietal Lobule	7, 5	8.9/7.6	8.55(−3,−67,56)/8.72(6,−64,58)
Postcentral Gyrus	7, 5, 3, 2, 40, 1	5.3/4.5	8.06(0,−46,66)/9.03(3,−52,66)
Inferior Parietal Lobule	40, 7, 39	3.5/3.5	4.28(−39,−49,61)/5.11(39,−52,58)
Cuneus	19, 18, 7, 30	2.7/3.7	4.53(0,‐82,40)/4.59(27,−83,37)
Paracentral Lobule	5, 4, 6, 7	2.9/1.8	7.28(0,−46,63)/5.71(3,−37,68)
Posterior Cingulate	29, 30, 23	1.6/1.0	3.25(−6,−41,5)/2.99(6,−41,5)

With FDR control, the identified SNP component did not exhibit any significant association with the behavioral measures. However, a trend of correlation was observed with AUDIT‐dep (r = −0.14, P = 1.14 × 10⁻², uncorrected). Figure 4 shows a Manhattan plot of weights of loci for the identified SNP component, where the z‐score threshold of 3.13 is marked to present the selection of top 2,020 SNPs (for details, see Supporting Information Fig. S3). Supporting Information Table S4 provides a summary of all the selected top contributing SNPs, including SNP position, hosting gene, and z‐scored component weight. One thousand and nineteen out of the 2,020 top contributing SNPs were mapped to 457 unique genes, which were used for Ingenuity Pathway Analysis (IPA: Ingenuity® Systems, http://www.ingenuity.com). IPA revealed a number of enriched canonical pathways, including synaptic long term depression (LTD, 1.70 × 10⁻⁵) and potentiation (LTP, 5.89 × 10⁻³), CREB Signaling in Neurons (6.31 × 10⁻⁴), protein kinase A (PKA) signaling (1.26 × 10⁻²), as well as GABA receptor signaling (2.24 × 10⁻²), as summarized in Table 5. IPA also indicated a significant enrichment of psychiatric disorders in our finding, including bipolar disorder (7.56 × 10⁻⁴), schizophrenia (5.50 × 10⁻³), and major depression (4.37 × 10⁻²). The identified genes were also significantly overrepresented in neuritogenesis (2.81 × 10⁻⁴) and other developmental functions (see Supporting Information Table S5 for a complete summary of pathway analyses). As there is no golden rule to select top contributing SNPs, we also tested top 1,000–5,000 SNPs and similar pathway analyses results were obtained (Supporting Information Table S6).

Figure 4.

Manhattan plot for the identified SNP component. The black line represents the z‐score threshold of 3.13. [Color figure can be viewed in the online issue, which is available at http://wileyonlinelibrary.com.]

Table 5.

Pathway analyses

IPA canonical pathway	Genes	P‐value
Synaptic long term depression	GNA14,ITPR1,GRM5,GRM7,GNAI3,GRID1,PLB1, RYR3,LYN,GNAT2,PPP2R1B,PRKCB,PRKCA	1.70E‐05
CREB signaling in neurons	GRM5,GRM7,GNAI3,GRIN2A,GRID1,GNAT2, PIK3CD,GNA14,ITPR1,CREB5,PRKCA,PRKCB	6.31E‐04
Axonal guidance signaling	LRRC4C,ITSN1,KALRN,RAC1,GNA14,ROBO1, ADAMTS2,GNAI3,SRGAP3,NTRK3,DCC,ADAM19,RTN4, ADAM23,GNAT2,PIK3CD,WNT5B,PRKCA,PRKCB	5.13E‐03
Synaptic long term potentiation	GRM5,GRM7,GRIN2A,GNA14,ITPR1,CREB5,PRKCA, PRKCB	5.89E‐03
Protein kinase A signaling	PTPN7,PTPRD,PTPN3,ITPR1,NFKB1,CREB5,PDE1C, GNAI3,HHAT,ADD3,RYR3,DCC,PTPRS,PDE8B, PRKCB,PRKCA	1.26E‐02
GABA receptor signaling	GABRG3,GABRR3,AP2M1,AP1B1	2.24E‐02

To further confirm the genetic influence on brain function alterations, we performed a regression analysis between the fMRI and SNP loadings while controlling for age, sex, population structure as well as associated behavioral measures of CD‐count, ICS‐fc, ICS‐tot, and PD‐count. The SNP component still showed a significant regression effect (P = 9.18 × 10⁻¹⁰) on the fMRI component.

DISCUSSION

In this work, pICA‐MR was used to identify SNP‐fMRI associations under the guidance of genetic references. Compared to pICA‐R which deals with reference SNPs from a single gene, pICA‐MR has been extended to accommodate a set of genes which might contribute in a coordinated manner. Simulations showed reliable convergence for all the tested scenarios [Chen et al., 2014] and recommended that pICA‐MR reference selection should be based on a relatively strong hypothesis which deserves further investigation in the following aspects. First, pICA‐MR examines reference SNPs in a multivariate framework for potential aggregate effects. Second, pICA‐MR investigates how the references constrain the same or different genetic components and their further associations with imaging components. This may help further our understanding of genetic influence across various disrupted neural mechanisms involved in a complex disorder. Last but not least, pICA‐MR performs a semi‐blind decomposition where the genetic references serve as guidance and other variables are allowed to show their own importance driven by the data. In this way, we expect to obtain a more complete picture of the underlying genetic architecture. Specifically, in this work, we tested six strongly hypothesized gene sets consistently implicated in alcohol studies. While there exists a large library of alcohol related genes, they will be left for future investigations.

The fMRI loadings exhibited a positive correlation with CD‐count, together with positive activations of the component, indicating that participants experiencing more severe alcohol dependence symptoms had higher regional activations when exposed to the taste of alcohol. The brain network comprised precuneus, superior and inferior parietal lobules, and PCC, as shown in Figure 3 and listed in Table 4. The brain network covered the midline, which might be attributable to spatial smoothing using an 8‐mm Gaussian kernel. Precuneus belongs to associative cortices and is known to be involved in a wide range of highly integrated tasks, including episodic memory retrieval, self‐referential processes, and consciousness [Cavanna and Trimble, 2006]. Its involvement in a variety of processes builds on its anatomical wide‐spread connections with both adjacent areas such as superior parietal lobule and PCC, and frontal lobes including prefrontal cortex and anterior cingulate cortex [Cavanna and Trimble, 2006]. Although not generally targeted for AUD, precuneus and parietal regions have been robustly implicated in craving studies, where activation elicited by drug‐related cues has been found associated with severity of dependence [Claus et al., 2011; Liu et al., 2013; Park et al., 2007; Tapert et al., 2004]. The drug‐cue‐elicited activation in precuneus is possibly a reflection of its recruitment in episodic memory retrieval, as the triggering of craving can be considered as a conditioned response where the recollection of past experience as episodic memories serves as a conditioned cue [Robbins et al., 2008]. PCC is a key player in the salience network [Sutherland et al., 2012] and frequently implicated in the processing of drug‐related stimuli [Tapert et al., 2004; Wrase et al., 2007]. Its functional alteration has proven to underlie the concurrent use of alcohol and tobacco [Liu et al., in press]. Of particular interest, as shown in a meta‐analysis on fMRI studies of alcohol cue reactivity, brain activation in precuneus and PCC, instead of the mesolimbic system, is selectively enhanced and most effectively differentiates cases from controls [Schacht et al., 2013]. Overall, the identified brain network echoes considerable similar findings and deserves more attention to elucidate the neuropathology of addiction.

The associated genetic component elicited by the three genes, BDNF, CREB1, and CREB5, negatively correlated with the fMRI component, indicating that participants carrying lower loadings on the SNP component presented higher brain activation in the identified precuneus and parietal regions. Pathway analyses delineated a complex genetic architecture emphasizing neural plasticity and signaling pathways based on the 457 genes highlighted in the component. A meta‐analysis of mouse models identified 3,800 genes differentially expressed between models of high and low amounts of alcohol consumption [Mulligan et al., 2006], suggesting that thousands of genes are involved in the pathology of AUD. Our findings appeared to be in line with the animal study and further suggested that the CREB‐BDNF pathway likely serves as a hub of the polygenetic effect related to AUD, as illustrated below in several associated canonical pathways.

CREB functions as a transcription factor and is well known for its role in neural plasticity and long‐term memory [Carlezon et al., 2005]. BDNF is a CREB regulated gene and active in synaptic plasticity [Bramham and Messaoudi, 2005]. Together with several glutamatergic genes (GRM5, GRM7, GRID1, and GRIN2A) and protein kinase C genes (PRKCA and PRKCB), they signify the CREB signaling, synaptic LTD, and synaptic LTP pathways. Synaptic LTP and LTD are two forms of synaptic plasticity which enhances or weakens, respectively, the synchronized stimulations between neurons, thus allowing the refinement of neuronal circuits underlying learning and memory [Malenka and Bear, 2004]. It is commonly recognized that synaptic plasticity plays an important role in the development of addiction, through which use of drug progresses from impulsive to compulsive behavior [Kauer and Malenka, 2007]. A meta‐analysis further indicated that a genetic component might affect addiction through regulating synaptic plasticity [Li et al., 2008], where LTD and LTP are among the top enriched pathways for the 396 addiction‐related genes implicated in two or more independent studies. Particularly, one SNP in BDNF (rs6265_A or Val66Met, “A” represents the minor allele) has been identified as predicting relapse in alcohol dependence, where minor allele carriers showed decreased vulnerability to relapse [Wojnar et al., 2009]. This is consistent with our finding where the same SNP exhibited a positive weight, indicating that minor allele carriers presented lower brain activation which was associated with less severe alcohol dependence symptoms. The SNP in GRIN2A (rs4628972_G) and three SNPs in PRKCA (rs17688881_C, rs721429_A and rs7217618_C) presented negative weights, indicating that the participants carrying more minor alleles showed higher brain activations and more severe alcohol dependence symptoms. The opposite was observed for the rest (rs1000061_G in GRM5, rs1353832_C in GRM7, rs1863824_C in GRID1, rs8077110_T in PRKCA, and rs880824_A in PRKCB) which all presented positive weights.

As a major inhibitory neurotransmitter in the central nervous system, GABAergic signaling is implicated in addiction [Bierut et al., 2010]. It has been reported that chronic cocaine uses decrease GABAergic synapse function, such that LTP induction is not effectively suppressed at excitatory synapses [Liu et al., 2005]. Polymorphisms in GABA receptor genes are consistently identified as susceptibility loci to alcohol dependence [Bierut et al., 2010], including a number of SNPs in GABRG3 [Dick et al., 2004]. In our finding, all the three identified SNPs in GABRG3 (rs12439549_G, rs4438262_G, and rs3922613_G) and two SNPs (rs1874864_G and rs7638369_T) in GABRR3 contributed with positive weights, while the rest (rs1844934_T, rs1688378_A, and rs1492054_C) exhibited negative weights. Another pathway, cAMP‐PKA pathway, is noteworthy here, which is a primary signaling cascade modulating numerous cellular events in neurons, including synaptic plasticity [Waltereit and Weller, 2003]. It is documented that activation of the cAMP‐PKA signaling leads to increased activity of CREB [Ron and Jurd, 2005], and genetic mutation reducing cAMP‐PKA signaling results in increased sensitivity to the sedative effects of ethanol [Wand et al., 2001]. The detail of how each SNP in this pathway contributes to the genetic component can be found in Supporting Information Table S4.

It is noted that the identified genetic component was not specific to AUD. The aforementioned canonical pathways and neurodevelopmental functions such as neuritogenesis and axonal guidance (Supporting Information Table S5) are part of much broader neural processes. IPA also implicated that the associated genes were overrepresented in other neuropsychiatric diseases, including bipolar disorder, schizophrenia, and major depression. This suggests a genetic basis for the comorbidity among these disorders, which has been the focus of recent large‐scale genetic studies [Johnson et al., 2009; Lee et al., 2013]. Overall, the genetic component delineates a relatively general substrate whose disruptions may interact with other factors to trigger various diseases including AUD, which is in accord with the fact that the identified brain network also participates in various tasks not specific to those related to AUD. Future studies that combine genetics and neuroimaging across a heterogeneous sample of psychiatric disorders will be the best approach for understanding the degree to which there is overlap in genetic contributions or brain circuits among comorbid conditions.

In summary, using a novel semiblind multivariate approach, we demonstrate gene interactions within the polygenic model of AUD selected by the CREB‐BDNF gene set. The extracted SNP component exhibits enrichment in neural plasticity and signaling pathways. The captured brain network, highlighting the precuneus, superior parietal, and PCC regions, is related to alcohol dependence symptom. The significant SNP‐fMRI association indicates that the genetic factor affecting neural plasticity may influence precuneus, superior parietal, and PCC responses to alcohol cues. We speculate one likely mechanism is through learning and memory function executed in these brain regions that is implicated in addiction [Courtney et al., 2014; Robbins et al., 2008] and modulated by neural plasticity.

Supporting information

Supporting Information

Supporting Information Table 1.

Supporting Information Table 4.