Polygenic risk score-derived subcortical connectivity mediates attention-deficit hyperactivity disorder diagnosis

STRUCTURED ABSTRACT

Background

ADHD has substantial heritability and a recent large scale investigation has identified common genome-wide significant loci associated with increased risk for ADHD. Along the same lines, many studies using non-invasive neuroimaging have identified differences in brain functional connectivity in children with ADHD. Here, we attempted to bridge these literatures to identify differences in functional connectivity associated with common genetic risk for ADHD using polygenic risk score (PRS).

Methods

We computed ADHD polygenic risk scores (PRS) for each participant in our sample (n=315, children aged 7 to 13, 196 with ADHD and 119 unaffected comparison children) using ADHD data from the Psychiatric Genetics Consortium as a discovery set. Magnetic resonance imaging was used to evaluate resting state functional connectivity of targeted sub-cortical structures.

Results

The functional connectivity between two region pairs demonstrated a significant correlation to the PRS: right caudate-parietal cortex, and nucleus accumbens-occipital cortex. Connectivity between these areas, in addition to being correlated with the PRS, was correlated with ADHD status. The connection between the caudate and the parietal region acted as a statistical suppressor, such that when it was included in a path model, the association between the PRS and the ADHD status was enhanced.

Conclusions

Results suggest that functional connectivity to certain subcortical brain regions are directly altered by genetic variants and certain cortico-subcortical connections may modulate ADHD-related genetic effects.

INTRODUCTION

Children diagnosed with ADHD have poor lifelong outcomes and their disorder incurs substantial societal costs despite available treatments, which are only partially effective (1). Thus, understanding ADHD’s etiology and pathophysiology is a public health priority. To do so requires clarifying how genetic and environmental risks affect the brain and behavior. Through the past decade, researchers identified many putative genetic and brain imaging markers tied to ADHD symptoms and diagnoses. However, the relationship between genetic variation, functional brain organization, and ADHD remains unknown.

ADHD has substantial heritability of liability of diagnosis (2, 3) although molecular data relying on common gene variants (single nucleotide polymorphisms or SNPs) reveal a lower but still substantial SNP heritability than twin studies (4). A large collaborative study identified a dozen common genome-wide significant loci (4); rare genetic variants (<1% of the population) have also been reported (3). While individual common genetic markers confer a very small amount of risk, cumulative genetic risk can be measured using polygenic risk scores, which sum the effects of many common variants into a cumulative genetic effect (5–9). We recently used this method to show that the effects of common variants on ADHD is mediated by alterations in working memory and arousal (10).

Polygenic risk scores for psychiatric disorders have been informative in the sense that they provide an individual probabilistic susceptibility to disease. They have the advantage of predicting an individual’s risk for ADHD based on molecular information and can be compared to other risk indices such as family history, contributing to eventual progress on biomarker algorithms and personalized approaches to care. Because these risk scores are associated with various ADHD phenotype indicators (10,11), we sought to extend this effort to discover mediators of genetic effect in the brain for ADHD.

Resting-state functional connectivity provides a logical avenue for investigating how PRS influence the developing brain. ADHD symptoms are substantially heritable (2); furthermore, the connectivity within and between brain networks are equally not all equally heritable, with networks such as the dorsal and ventral attention, and default mode demonstrating substantial heritability(11, 12). In a recent report with 24 ADHD and 52 nuclear extended families, Sudre and colleagues reported on brain regions that had heritable functional connectivity within regions such as the default mode network (DMN), the ventral attentional network, and the frontoparietal network (13). ADHD symptoms such as impulsivity/hyperactivity and inattention are previously associated with decreased functional connectivity within the DMN (14).

To the best of our knowledge, it is still unknown whether a higher PRS influences communication between brain regions (i.e. functional connectivity) or whether these altered functional connections mediate the relationship between PRS and ADHD status. In the current report, we use resting state functional connectivity MRI and PRS in a large sample of children with and without ADHD to examine genetic contributions to connectivity in ADHD. Given the recent evidence for the ability of PRSs to predict various behavioral disorders, such as autism spectrum disorder, schizophrenia, bipolar disorder, and ADHD, (5), we hypothesized the PRS would be associated with ADHD-related brain connectivity.

Children with ADHD display marked differences from their typically-developing peers in behaviors such as emotional regulation, working memory, reward processing, and regulation of motor control (15, 16). These symptoms have provided the direction for investigations into the neurobiological origins of the disorder, and several brain networks have been previously shown to exhibit atypical function(13, 14, 17). We focused our study of functional connectivity on circuits involving subcortical regions that have been previously implicated structurally in ADHD, that we had previously associated with aspects of ADHD phenotype, and thought to be involved in both cognitive control ( the left and right caudate (18, 19)) and in emotional regulation (left and right nucleus accumbens (17, 18, 20, 21), and left and right amygdala (22)(18)). We hypothesized that PRS is associated with altered connectivity to these nuclei. We selected the nucleus accumbens for its strong involvement with directing attention to towards motivationally-relevant goals by actively promoting the likelihood of selection, and vigor to obtain these rewards, avoidance of aversive consequences, novel stimuli exploration, and because we previously showed its involvement in individual differences in reward function in ADHD (see (23), for review). Additionally, we selected the caudate nucleus for its contribution toward activation/inhibition of action schemas, and goal-directed actions(24), and observed volumetric differences among children with ADHD. We examined differences in connectivity the amygdala due in part to the observed emotional dysregulation and processing rewarding environmental stimuli and its centrality in our prior study of individual variation in emotion regulation in ADHD (22).

The fundamental concept behind the polygenic risk score is that it provides a cumulative indicator of genetic risk based on the differential allelic frequency between cases and control populations in a disorder where many genes appear to be responsible for the phenotype. It has become a primary tool for tracking routes of common genetic liability in complex disease. We therefore aimed to use this genetic measure of ADHD risk to test the genetic basis of connectivity to regions that 1) demonstrate differential gene expression between cases and controls and 2) have been shown previously in the literature to have aberrant connectivity compared to typically developing children. Preliminary results employed Functional Mapping and Annotation of Genome-Wide Association Studies (FUMA GWAS), a web-based platform that performs gene-set based association analysis. The platform calculates which differentially expressed genes (DEG – that is sets of genes which are more or less expressed in a specific tissue type, see supplementary figure S3) demonstrate tissue specificity of prioritized genes within these regions (see supplementary methods).

METHODS

Participants

656 participants were genotyped (424 ADHD, 232 Typically-developing (TD)). Of these, 514 participants were determined based on genotype to be of European ancestry and were used for the genetic analysis (see Relatedness and Population Stratification in supplementary methods). Of these, 315 (119 TD and 196 ADHD) participants had at least 4 minutes of useable resting-state-functional magnetic resonance imaging (rs-fMRI) (ages mean = 10.394 and std. dev. = 1.552, n = 109 female) free of known medical or neurological conditions (self-report), characterized by our stringent research diagnostic process to satisfy ADHD criteria according to the DSM-IV and DSM-5 (APA 1994; 2013) (See table 1 and Recruitment and Clinical Assessment below). All methods were approved by the Institutional Review Board at Oregon Health and Science University. Because age and sex might affect how certain regions will be functionally connected, we conducted a 2 × 2 (sex by diagnosis (ADHD vs. TD)) ANOVA with age as the dependent variable. There were no differences in age between sexes or diagnosis groups (sex: F[1,358] = 0.011, p = 0.916; diagnosis: F[1,358] = 1.126, p=0.289) or interaction between these variables (sex*diagnosis: F[1,358] = 0.352, p=0.553).

Table 1:

Participant demographics by group.

	European Ancestry-only intake group(n=514)		European Ancestry -only(n=315) with at least 4 minutes of low motion rs-fMRI data
	Control	ADHD(all)	Control	ADHD(all)
N	177	337	119	196
% male	54%	72%	54%	72%
Age at intake	9.4(1.5)	9.5(1.5)	10.13(1.58)	10.56(1.47)

Recruitment and Clinical Assessment

The sample was recruited from the community using a case-finding procedure to identify ADHD cases and non-ADHD controls from among community volunteers. A parent/legal guardian provided written informed consent, and children provided written informed assent. Volunteers responded to a mass mailing to all families with children in the target age range in the metropolitan area. After initial screening, a detailed clinical research evaluation was conducted using standardized, well-normed rating scales from parent and teacher, parent semi-structured clinical interview conducted by reliable and trained clinical interviewers, child intellectual testing, and clinical observation as reported, previously (10). Best estimate research diagnoses and final diagnostic eligibility were established by a team of two experienced clinicians (a child psychiatrist and a child psychologist) who independently arrived at the diagnosis and conferenced any disagreements. They achieved an acceptable concordance rate for identifying qualified ADHD cases (k>.7).

Working memory

In addition to studying ADHD as a diagnosis and a dimension, we examined working memory as an outcome. We did this because working memory is frequently cited as a possible mediator of ADHD effects via neural alterations (25–27), and because we recently found in this sample that it statistically mediated polygenic effects (10). To evaluate working memory (technically, “working memory storage” as we did not administer dual tasks tests), we administered three widely used measures: digit span backward(28), spatial span backward (29), and N-back(30). The methods are standard and described in detail in the online supplement.

Genotyping

Methods for Genotyping have been reported previously (10) and are further described in the supplementary methods. Briefly, participants were genotyped using the PSYCHChip v1.1. Candidate SNPs that were not initially genotyped were imputed with the 1000 Genomes reference panel (see supplementary methods). SNPs with association p ≤ 0.5 from the GWAS were included for the calculation of polygenic risk scores (though other thresholds were also tested which demonstrated a highly similar pattern of results; see Supplementary Table S7). Polygenic risk scores were calculated for each participant based on PGC GWAS for ADHD (N= 20,183 ADHD cases and 35,191 controls, see supplementary methods). The final number of SNPs used for the score is 193,692.

FUMA Analysis of Differential gene expression

Using summary statistics from the ADHD GWAS, we determined differential gene expression in various tissues. FUMA GWAS was used to map SNPs to genes, and using several expression repositories, map gene expression to tissues (31). Differentially expressed gene sets are determined be performing a two-sided t-test per each tissue versus all other types. By default genes that have a Bonferroni corrected p-value <0.05 and an absolute log fold change > 0.58 are selected as DEG. For the signed DEG, the direction of expression is taken into account(31). For full details of differential gene expression, see supplementary methods. These regions that demonstrate differential gene expression (along with the current literature), were then used were then used to select seeds used in seed-based correlation.

Neuroimaging

Scan Details

Structural and functional data were acquired using a 3T Siemens Trio Tim equipped with a 12-channel head coil. One T1 weighted structural image was acquired per participant (TR = 2300ms, TE = 3.58ms, FOV = 256 × 256, orientation = saggital, voxel size=1mm × 1 mm × 1.1mm slice thickness). Resting state data was acquired in 3, 5 minutes blocks using blood oxygen level-dependent (BOLD) contrast (TR = 2500ms, TE = 30ms, flip angle = 90°, FOV = 240mm, in-plane resolution = 3.8mm × 3.8mm with a slice thickness of 3.8 mm). Participants were instructed to lie still and fixate on a cross-hair at the center of their visual field. Structural T1s and dense time series for each participant was processed using modified pipelines from the Human Connectome Project (HCP) (32). Time courses were then motion censored to minimize the effects of head motion (see supplemental methods).

Parcellations and functional networks

For bold time series, signal intensity was calculated as the average signal within the ROIs(33)(32). We used the subcortical segmentations derived from FreeSurfer to define subcortical structures (Fischl et al 2002). After HCP processing, 6 candidate regions of interest (ROIs) (3 left and 3 right) were used as seed regions, and the time series of resting-state scans were used to calculate a Pearson’s correlation against all other 91,282 greyordinates (see Figure 1). The 3 candidate sites were selected based on previous subcortical differences in children with ADHD (18, 34, 35). Hoogman and colleagues (18) examined the volumes of the thalamus, putamen, hippocampus and pallidum, however to preserve power, and due to the multiple sub nuclei of the thalamus, and pallidum, we focused specifically on 3 sites. The candidate sites that were most theoretically relevant were: the left and right amygdala, the left and right nucleus accumbens, and caudate nucleus, and the left and right sides were examined separately for a total of 6 tests.

Figure 1:

Method of Analysis. For more details, see Methods section. A) Dense time series data was generated from each participant using the Freesurfer processing pipeline(32). Outlier frames or frames with FD>0.2mm were removed (B vs C). D) Using the subcortical parcellations from Freesurfer, we extracted regions of interest that correspond with the right and left amygdala, nucleus accumbens, and caudate. We used these regions as seed to correlate with all other regions from the dense times series. This resulted in a vector of correlation values between the region of interest and all other greyordinates. Scalars were generated for each participant (shown as S1-S5as an example), and then correlated the polygenic risk score at each greyordinate (E). Within the scatterplot, each participant is represented by one circle.

Data Analysis

A linear regression was performed at each greyordinate between the connectivity between the seed ROI and that greyordinate, and the PRS using FSL PALM (Permutation Analysis of Linear Models; Winkler et al. 2014). That is to say, we first calculated seed-greyordinate connectivity, then we correlated the connectivity to the PRS (where individual PRS is the constant across correlations, see Figure 1). We decided to combine the ADHD and TD children into a single cohort, due to the observed a continuous distribution of PRS. Correlation coefficients from the regression were converted to Z-scores, which were then thresholded at 1.96 to determine putative clusters. These clusters were contiguous brain regions where the Fisher Z > 1.96 (i.e. these represent the ~5% of PRS-connectivity correlations). To correct for multiple comparisons, cluster sizes were then permutation-based tested with 10,000 permutations, to test where the cluster extent was significant, (α was set at −log p= 1.3). Each seed based correlation was tested independently, but permutated in the same manner. This method has been used previously, (36, 37) and is described further in the supplementary methods. The correlation between time series for the seed, and the time series for the cluster was then calculated for each subject for use in the mediation analysis.

A path diagram was used to map one possible causal model (acknowledging the cross sectional data; see Bentler (38). In that context, statistical mediation analysis was conducted in Mplus (Version 8, Los Angeles, CA: Muthén & Muthén) by using diagnosis (ADHD/TD) as the outcome variable in a multivariate logistic regression, with PRS and connectivity as predictors. Sex and age were used as covariates in the diagnosis, and cluster connectivity, however only sex was used as covariate for only PRS, as we would not expect age to influence PRS. Direct effects of PRS on diagnosis and indirect effect of PRS through connectivity to diagnosis were measured. The multivariate model included age as a covariate for seed-cluster connectivity and diagnosis, and sex as covariate for the PRS, seed-cluster connectivity, and diagnosis. In addition to ADHD diagnosis as a categorical dependent variable, we examined ADHD as a dimensional latent variable by parent report and by teacher report, as well as a working memory latent variable selected from a cognitive battery (see supplementary methods and Figure S1). For the parent latent variable, indicators were the inattention and hyperactivity subscale scores on the ADHD Rating Scale (ADHD-RS), the Kiddie Schedule for Affective Disorders (K-SAD), the Conners’ Parent Rating Scale-3 (CPRS-3), and the Strengths and Difficulties Questionnaire (SDQ). For the teacher rating scale, it was built from the ADHD-RS, the SPRS-3 and the SDQ. All regression beta weights reported in the results section are standardized.

RESULTS

When we examined the results of the differential gene expression in specific tissue types, we found that several brain regions demonstrated both significant upregulated and down regulated DEG (Figure 1). Regions of the brain that demonstrated significantly up-regulated differential gene expression were: the amygdala (Number of genes = 1847, p = 9.305 × 10⁻⁴, Bonferroni corrected p = 0.04932), the caudate (Number of genes = 2100, p = 7.634 × 10⁻⁴, Bonferroni corrected p = 0.04046), hippocampus (Number of genes = 1952, p = 7.3363 × 10⁻⁴, Bonferroni corrected p = 0.0388), and the nucleus accumbens (Number of genes = 2248, p = 7.817 × 10⁻⁴, Bonferroni corrected p = 0.04143).

We therefore examined functional connectivity for 6 circuits involving subcortical regions: the left and right caudate, left and right nucleus accumbens, and left and right amygdala (Table 2).

Table 2:

Seed-based linear correlation analysis by region.

Region	Left Caudate nucleus	Right Caudate Nucleus	Left Nucleus accumbens	Right Nucleus Accumbens	Left Amygdala	Right Amygdala
Fisher Z range	−3.62 – 3.78	−3.61 –3.49	−4.18–3.72	−4.08–4.27	−3.31 –3.9	−3.59–3.56
Clusters above threshold	13	18	24	15	13	13
Clusters survived permutation	1	0	0	2	0	0
Maximum cluster Fisher Z family-wise error - log p-value	1.39	0.58	0.728	3.15	0.299	0.518
Cluster p-value	0.041	0.263	0.187	7.08 × 10−4	0.502	0.303

Results of linear regression between left or right caudate nucleus connectivity and PRS (homogenous/heterogeneous)

The PRS was associated significantly with connectivity between the left caudate nucleus and a cluster within the intraparietal sulcus (z= −1.96 to −3.78, (z(10,000 permutations, n=315), family-wise error p= 0.041). The statistical map in the lower row of Figure 3 (right columns) shows a Pearson’s correlation (Fisher Z transformed) of the PRS to the left caudate nucleus connectivity at each greyordinate. The PRS was also significantly correlated with connectivity between the left caudate and a cluster in the right parietal cortex (r = 0.026 ± 0.162 S.D., see Figure 4). For the right caudate nucleus, we did not observe any significant association between its connectivity and the PRS after correction.

Figure 3.

Correlation between connectivity from subcortical seed regions and polygenic risk score (PRS). (Upper rows) Fisher z values based on correlation are shown for the left and right nucleus accumbens and caudate nucleus. Fisher z values were then thresholded at 1.96 to create putative clusters. (Lower rows) Permutation testing was performed on clusters, and then only clusters surviving p > .05 were deemed significant. When the left caudate was used as a seed region, one cluster of the parietal lobe showed a significant correlation with PRS. Additionally, when the right nucleus accumbens was used as a seed, the connectivity to the occipital cortex was significantly correlated to PRS. No other seed region’s connectivity demonstrated significant correlation to PRS. Hotter colors indicate greater Fisher z values, and cooler colors represent smaller Fisher z values. Yellow regions indicate significant clusters.

Figure 4.

Correlation of connectivity between the seed regions and clusters correlated with polygenic risk score (PRS). The clusters found in Figure 2 were used to make a region of interest (see Methods and Materials). Solid line indicates linear regression (left caudate–parietal cluster: R² = .041; right nucleus accumbens–occipital cluster: R² = −.021). Pearson’s correlation between left caudate–parietal region pair and PRS = −.202. Pearson’s correlation between right nucleus accumbens and occipital cortices pair = .144. Each circle represents 1 participant. Filled circles represent typically developing children, and open circles represent children with attention-deficit/hyperactivity disorder. nuc. accum, nucleus accumbens.

Results of linear regression between left or right nucleus accumbens connectivity and PRS

When we examined the statistical map of the right nucleus accumbens connectivity, we found a regions of the occipital lobe near the calcarine sulcus showed significant correlation with the PRS (Figure 3, upper left images) (Fischer Z values of cluster =1.96–3.15. After permutation testing, two cluster were significant (left cortex: family-wise error p-value= 5.00 × 10⁻⁴, right cortex: family-wise error p = 1.58 × 10⁻³). The regions of significance for the left and right were combined to form one cluster(see Figure 3) which was then used in the mediation analysis (see below). The mean correlation for the right nucleus accumbens to the cluster was −0.0998 ±0.167 S.D. (see figure 4). We did not observe any significant correlation between the PRS and the connectivity of different clusters correlated to the left nucleus accumbens after correction, though we did observed a similar pattern of correlation observed in the right nucleus accumbens.

Results of linear regression between amygdala connectivity and PRS

We did not observe connectivity between brain regions and either the left or right amygdala that had a significant correlation to the PRS (see supplement for details).

Caudate-parietal cluster Mediation Analyses

We conducted a mediation analysis to determine the extent to which the correlation between the seed region and the clusters mediated the relationship between the PRS and ADHD diagnosis (Figure 5A). Connectivity between the caudate and parietal cluster mediated the correlation between PRS and diagnosis. As expected, the PRS statistically predicted ADHD diagnosis (total PRS to diagnosis effect β= 0.153(0.073 S.E.), p=0.038) (supplementary tableS1). As expected, the model showed that connectivity to the caudate-cluster was significantly correlated with the PRS (β=−0.467(0.152), p=0.002)), such that subjects with higher PRS had a less connectivity between these regions. That is to say, the higher the ADHD genetic risk, the more the correlation between the 2 regions decreases. Importantly, increased caudate-cluster connectivity was associated with an ADHD diagnosis (β=0.507(0.160), p=0.002). The relationship between the PRS and ADHD diagnosis was thus suppressed by connectivity between the left caudate nucleus and the right parietal cortex, with 59.5% of the direct effect of the PGS on the diagnosis counteracted by connectivity of the cluster to the caudate nucleus (Figure 5B). As a result, in the mediation model (see Figure 5 Panel A), the association of the PRS with ADHD was significantly stronger with the suppressor-mediator in the model. Although sex of the participant was covaried in that model, we conducted a sensitivity analysis in a subset with matched proportions of males and females (Controls: n=119 (46% female) ADHD: n=118 (46% female; see Supplementary Results Table S2). The effect sizes were similar and in the same direction (although the loss of sample size rendered the total indirect effect non-significant (p=.152), the size of the indirect effect remained similar (β=−.185, vs. β=−0.237 in the full model).

Figure 5.

Mediation model of significant clusters. (A) Full statistical structural equation model for caudate-parietal mediation. See Methods and Materials for full details of model. (B) Mediation model indicates that 60% of the effect of polygenic risk score (PRS) on diagnosis is mediated by how the caudate communicates with the parietal cortex. (C) Statistical model with connectivity between the nucleus accumbens–occipital area as the mediator. (D) Mediation model demonstrates that while connectivity to these regions was significantly correlated with PRS, it does not appear to mediate the relationship between PRS and the diagnosis. For all subfigures, dark lines indicate a significant effect in the model, and dashed lines indicate a nonsignificant effect. Yellow regions are a visual reference for the region pair in the model. For visual clarity, β weights for the principal components are not shown. For clarity, sex, age, and principle components (and medication status; see Supplemental Results) are omitted in the graphical representation for panels (B) and (C). DX, diagnosis; Nuc. Accum., nucleus accumbens.

Furthermore, we conducted similar path-model based mediation analysis on the significantly correlated clusters found in Figure 3 with working memory as the outcome variable. Results demonstrate that caudate-parietal connectivity did not predict working memory (β=−0.069(0.064), p=0.279); however, as already shown earlier, the PRS did predict working memory (β=−0.194(0.060), p=0.001), which we had shown previously (10). See Supplementary table S1 for full model results.

Lastly, we ran a mediation analysis with parent reported symptoms of ADHD as a latent variable. As expected, the left caudate-right parietal connectivity was significantly correlated with the PRS (β=−0.195 (0.045), p<0.001), but was unrelated to the parent-reported symptoms (β=−0.027(0.060), p = 0.655). However, parent reported symptoms were related to PRS (β=0.138(0.059), p=0.020). This relationship between PRS and parent reported symptoms is consistent with what has been previously reported in the literature(10). Lastly, we conducted a mediation analysis on teacher-reported symptoms of ADHD as a latent variable using left caudate-parietal region pair. While PRS were marginally correlated with teacher ratings latent variable of ADHD (β=0.122(0.062), p =0.051), we observed that left caudate-right parietal connectivity was unrelated to the teacher ratings (symptoms (β=0.77(0.060), p = 0.200). Effect sizes were nearly identical when previous history of current psychiatric mediation (see Supplementary Table S3) or previous history of medication use (Supplementary Table S4) was added was included in the model.

Nucleus accumbens-occipital cortex mediation analyses

When we conducted the same mediation analysis with the connectivity between the right nucleus accumbens and the cluster in occipital cortex (right nucleus accumbens to bilateral occipital cortices (clusters from each occipital cortex were combined to form 1 cluster, Figure 3), as already noted the right nucleus accumbens-occipital cortex connectivity was associated with a greater PRS (β=0.270(0.117), p=0.021), and the connectivity was associated with diagnosis (β=0.582(0.146), p<0.001). The PRS became unrelated to diagnosis suggesting mediation (β=−0.005(0.120), p=0.969, p=0.0860, Figure 5D). Additional outcome variables did not demonstrate significant mediation (see Supplementary results).

DISCUSSION

This was to our knowledge the first study to investigate the association of ADHD polygenic risk with functional connectivity in a targeted set of subcortical brain regions. Findings did confirm PRS association to ADHD-associated brain circuits, although with some caution due to the direction of effects that we discuss below. Taken together, results suggest that ADHD genetic load involves for part of its effect particular brain circuits (communication between these region pairs).

With regard to the caudate findings, we note that the caudate participates in widely distributed frontal-subcortical neural loops, and subserves multiple functions, including behavioral selection, regulation of motor output, and cognitive and emotional processing. Relatively few studies have investigated functional connectivity to the caudate with respect to ADHD. The caudate nuclei appear to be key regions of altered activation in children with ADHD during tasks that require cognitive control(39, 40). Abnormal activation has been observed in the caudate and ventral striatum in individuals with ADHD compared to controls(41, 42). Lastly, brain activity within the superior parietal cortex has been shown to be correlated with the availability of dopamine within the caudate(43). This indicates that lower dopaminergic function in children with ADHD has the potential to affect prominent connectivity hubs(41). As well, substantial literature has explored structural variation in the caudate in relation to ADHD, suggesting it is importance in disease etiology (18, 44–46).

The connectivity between to the right parietal cortex as a pathway for genetic influence is potentially very interesting. Meta-analyses have attempted to specify its role based on anatomical function in parietal cortex, with more lateral regions encoding saccadic eye movements, regions near the medial inferior parietal sulcus, encoding reach, and anterior parietal sulcus encoding, grasping (47), timing(48–50). These regions have been previously shown to be involved in interference inhibition, and attention allocation(19, 51, 52), and are regions shown to be activated by stimulant medication (53, 54). Previous work has also shown that children with ADHD demonstrate lower short- and long- range connectivity to/from the superior parietal cortex(41). Connectivity between the parietal cortex and the orbitofrontal cortex has been shown to be lower in children with ADHD(41). Reward-associated regions (striatum and anterior cingulate cortex) appear to have higher connectivity to orbitofrontal cortex(41), however we did not observe that the connectivity of a subset of these regions (caudate and nucleus accumbens) were associated with the PRS. The functional asymmetry between parietal cortices may explain why we observed genetic effects only for the left caudate seed. This would indicate that ADHD-related SNPs are involved with the functions localized to right parietal region. Because both the parietal cortex and caudate are heavily involved in motor control, and appear to be related to ADHD diagnosis, it may be reasonable to assume that these SNPs convey some risk for the regulation of motor behavior in ADHD. However, emerging evidence suggests that severity of hyperactivity in adults with ADHD was correlated with connectivity between the caudate and an auditory/sensorimotor independent component (55). Other metrics such as degree centrality specifically within the caudate has also been shown to be abnormal in children with ADHD when compared to typically developing children (56).

In our present experiment, we were unable to show a significant correlation between the connectivity to either the left or the right amygdala and the PRS, which was somewhat unexpected given previous findings that demonstrated reduced amygdala volume (18) and clear clinical demarcations in negative emotionality, which has been shown to be associated with reduced amygdala-insula connectivity among children with ADHD (22). The absence of a significant genetic correlation with amygdala connectivity may be due in part to the heterogeneity of children with ADHD (57, 58) or may suggest that emotional disruption involving the amygdala is not as strongly associated with common genetic influences as some other aspects of the phenotype. Karalunas and colleagues(22) noted that ADHD-related emotional patterns were related to functional connectivity between the amygdala and the posterior cingulate, and amygdala and the anterior insula(22).

Interestingly, we found that PRS seems to predict connectivity between occipital cortex and nucleus accumbens, suggesting a possible link between how reward systems communicate with the very early visual processing. This is tangentially supported by a recent study that found differences in grey matter volume(46) in visual cortex in patients with ADHD. One advantage of using resting-state functional connectivity to examine differences between ADHD and controls, compared to volumetric measurements is that one does not have to normalize for intracranial volume, which appears to grow differentially among affected and non-affected siblings(18).

By virtue of the way that polygenic risk scores are computed(5), we would expect that higher risk scores significantly predict ADHD in our model. The association of connectivity with the PRS is an important clue to pathophysiology. However, it is important to realize that the PRS only explains a small portion of the variance in either ADHD or connectivity; other factors are also shaping these phenotypes. (i.e., regulated by a complicated pattern of SNPs which is not captured by the PRS, is epigenetically modified, or most likely is a complicated pattern across these factors). We were also able to demonstrate that the PRS was positively associated with connectivity between the right nucleus accumbens and occipital cortex, yet negatively associated with connectivity between the left caudate and the parietal cluster (Figure 4). This suggests that ADHD risk alleles are involved in development of functional connections between the nucleus accumbens and occipital cortex, while at the same time they may inhibit development of functional connections between left caudate and parietal clusters. In other words, although we cannot isolate biological functions in the polygenic context here, ADHD risk-associated SNPs might be involved in fostering communication between reward and visual processing and also prevent communication between motor regulation and sensorimotor transformations.

Interpretation of mediation Results

In our mediation model, we provide the first report that the communication between parietal cortex and the caudate regions predicts ADHD diagnosis. Furthermore, in these cross-sectional data the diagnostic value of the PRS appears to be statistically mediated by this connection such that the connectivity from the caudate to parietal cortex suppresses the direct genetic effect on diagnosis relationship. When this effect is included in the model the association between the PRS and ADHD is revealed to be considerably stronger. However, further identification of such effects could help clarify ways to increase the sensitivity of polygenic scores. While age cannot account for our findings due to our statistical controls, we note that subcortical connectivity and volumes change with age and development (connectivity:(59), volumes: (18)). These results may not generalize to older or younger samples of ADHD. Interestingly, we found that despite caudate-parietal connectivity being significantly positively associated with diagnosis, this connectivity suppresses the direct effect of the PRS on diagnosis. That is to say increased genetic risk correlates with ADHD diagnosis, but this relationship is reduced the aforementioned connectivity.

Suppression mediation effects are rare in behavioral research although well described in the methodological literature (60, 61). Certain patterns of correlation, including the situation here in which opposite signs appear for some of the paths, results in the strength of the relationship between the PRS and diagnosis being increased when the suppressor mediator is in the model. While initially puzzling, these results imply that this connection likely conveys resiliency against the direct genetic effects of ADHD, perhaps via adaptive or homeostatic response to the genetic risk effect. This relationship, coupled with the marginally significant mediation of genetic PRS effect by the nucleus accumbens-occipital cortex circuit, together suggest that these models have the potential for improvement in larger samples with power to model relationships that are more complex. In such models, it should be possible to consider simultaneously several brain regions where the 1) communication patterns are associated with ADHD and 2) are genetically influenced. This is consistent with several theories that suggest that the disorder is multisystemic, with multiple parallel deficits, each contributing to a different cognitive domain (50, 51, 62, 63).

We examined connectivity from parcellated seed regions to all other greyordinates. This approach neglects to examine aberrant local circuitry within each of these candidate regions. von Rhien and colleagues have demonstrated that differences in global striatal structure did not exist and adolescents with ADHD and healthy controls, yet differences were observed in connectivity between anterior and posterior segments of the putamen. These results suggest that for structures where local computational differences between anterior/posterior/ or rostral/caudal functions are segregated, differential connections arising from subregions and intraregional communication must be accounted for rather than simply observing cortico-subcortical connectivity differences.

Finally, our current study examined seed-based connectivity using PRS. However, these hypothesis-driven candidate seed regions are most certainly not a comprehensive list of possible seed regions. Indeed, our mediation model could potentially be improved with a curated subset of the many possibly connections that correlate with the PRS. We limited our analysis to these specific regions due to their support in the literature and from our gene expression findings, while limiting the number of possible seeds to conserve statistical power. We hope that the present results, which indicate both a genetic and neural contribution toward ADHD diagnosis, will catalyze the field to further investigate the genetic bases brain connectivity aspects of ADHD. GWAS has started to identify candidate SNPs that convey significant risk for ADHD, and it is crucial that we identify the neural consequences of those risk alleles.

Figure 2.

Tissues that demonstrate differential gene expression. Bars indicate Bonferroni corrected p value of the differential gene expression for various tissue types. (Top row) Tissue types that demonstrate significant gene upregulation. (Middle row) Genes that are significantly downregulated in each tissue type. (Bottom row) Brain regions that are shown to be either upregulated or downregulated (as shown by a 2-sided t test). For a full list of genes that were used, see Supplemental Figure S3. Significant differences in gene expression between participants with attention-deficit/hyperactivity disorder and control participants are shown in red. DEG, differentially expressed genes.