Inhibition‐directed multimodal imaging fusion patterns in adults with ADHD and its potential underlying “gene‐brain‐cognition” relationship

Abstract Aims Inhibition deficits have been suggested to be a core cognitive impairment in attention‐deficit/hyperactivity disorder (ADHD). Exploring imaging patterns and the potential genetic components associated with inhibition deficits would definitely promote our understanding of the neuropathological mechanism of ADHD. This study aims to investigate the multimodal imaging fusion features related to inhibition deficits in adults with ADHD (aADHD) and to make an exploratory analysis of the role of inhibition‐related gene, NOS1, on those brain alterations. Methods Specifically, multisite canonical correlation analysis with reference plus joint independent component analysis (MCCAR + jICA) was conducted to identify the joint co‐varying gray matter volume (GMV) and the functional connectivity (FC) features related to inhibition in 69 aADHD and 44 healthy controls. Then, mediation analysis was employed to detect the relationship among inhibition‐related imaging features, NOS1 ex1f‐VNTR genotypes, and inhibition. Results Inhibition‐directed multimodal imaging fusion patterns of aADHD were reduced GMV and FC in inhibition network and increased GMV and FC in default mode network. The results showed a significant indirect effect of NOS1 ex1f‐VNTR on inhibition via FC component [effect size = −0.54 (SE = 0.29), 95% CI = −1.16 to −0.01]. In addition, the results indicated a significant indirect effect of GMV on the inhibition via FC component [effect size = 0.43 (SE = 0.23), 95% CI = 0.12 to 1.00]. Conclusion The findings suggested that reduced GMV and FC in inhibition network and increased GMV and FC in default mode network were jointly responsible for inhibition deficits in aADHD. Both the NOS1 ex1f‐VNTR genotypes and GMV might influence the inhibition through the mediation effect of the aforementioned FC (NOS1/GMV→FC→Inhibition).


| INTRODUC TI ON
Attention-deficit/hyperactivity disorder (ADHD), characterized by age-inappropriate inattention, hyperactivity, and impulsivity, is a childhood-onset neurodevelopment disorder which could persist into adulthood. 1 Inhibition deficits, the core cognitive impairments in ADHD, 2 are closed to the core symptoms of ADHD 3 and might be responsible for secondary deficits in other executive functions. 4 Thus, understanding the mechanism of inhibition would definitely promote our understanding and interpreting the pathogenesis of ADHD.
A core network of inhibition was identified in healthy subjects most prominently in the inferior frontal gyrus, superior frontal gyrus, the supplementary/pre-supplementary motor area, basal ganglia, and temporal/parietal areas. 5,6 More recently, studies reported conflicting activation in the cerebellum, cingulate, mesial frontal, and parietotemporal areas in inhibition process of adults with ADHD (aADHD). [7][8][9] Compelling evidence has confirmed that inhibition impairments involved both brain structural (eg, gray matter volume, GMV) and functional (eg, functional connectivity, FC) dysfunctions in aADHD. 10 The information provided by the single imaging modality is limited, while multi-mode could provide much more rich information. The main goal in integrating/fusing studies is to take advantage of the relative strengths of each modality to provide results synergistically for a more comprehensive understanding of brain activity. The joint analysis of multiple brain imaging measures could capture more views and covariations of brain modalities, which was believed to be essential for understanding brain networks and their relationship with human cognition and behavior. However, most studies focused either single modality or a multimodal comparison after individual analysis in each mode, in which cases, the cross-information among multiple modalities was often missed. In addition, compared with unsupervised fusion, supervised multimodal fusion takes advantage of the prior knowledge, for example, clinical symptoms or cognitive functions, to guide the fusion analysis and can discover more precisely and robustly goal-directed imaging features. 11 There have been studies using supervised multimodal fusion approach to look for multimodal biomarkers of schizophrenia 11 and depression, 12 but not yet for ADHD.
Inhibition was potentially inheritable characteristics in ADHD. 13,14 Recently, the NOS1 gene, encoding neuronal nitric oxide synthase (nNOS), has been suggested to be a candidate gene for aADHD. 15 The nNOS-derived nitric oxide (NO) acts as a second messenger downstream of the N-methyl-D-aspartate (NMDA) receptor and interacts with the monoaminergic systems. 16 Numerous evidence has supported the crucial role of the NOS1-nNOS-NO pathway in brain development and functions, and in the etiology of neuropsychiatric disorders including ADHD. 17 In the existing literature, a functional variant, NOS1 ex1f-VNTR, has been indicated to be associated with the hyperactive-impulsive behavior in aADHD. 18 Imaging genetic exploration showed that the risk short allele (S) of NOS1 ex1f-VNTR was associated with abnormal activation of the dorsolateral prefrontal cortex and inferior frontal cortex during the response inhibition tasks. 19 This functional variant might affect the activation of anterior cingulated cortex or ventral striatum, and finally lead to the impulsivity-related behavior. 20,21 In addition to the brain function, the NOS1 ex1f-VNTR also potentially influenced the brain structure, such as the white matter microstructure. 22 However, no efforts to explore the genetic effects of NOS1 ex1f-VNTR on the multimodal imaging patterns has been reported, much less the inhibition-guided fusion analyses.
In our present study, we firstly aim to identify the co-varying multimodal MRI patterns in aADHD under the guidance of inhibition function, which has been suggested to be a core impairment of ADHD; Secondly, the potential genetic effects of NOS1 ex1f-VNTR on inhibition and its related brain alterations were investigated to build a potential "gene-brain-behavior" relationship.

| Participants
A total of 119 subjects were recruited in the study. Four aADHD and two controls were excluded because they failed to finish the MRI scan. by a senior psychiatrist for diagnosis and potential comorbidity. ADHD

Conclusion:
The findings suggested that reduced GMV and FC in inhibition network and increased GMV and FC in default mode network were jointly responsible for inhibition deficits in aADHD. Both the NOS1 ex1f-VNTR genotypes and GMV might influence the inhibition through the mediation effect of the aforementioned FC (NOS1/ GMV →FC→ Inhibition).

K E Y W O R D S
adults with attention-deficit/hyperactivity disorder, inhibition, multimodal MRI fusion, NOS1 ex1f-VNTR diagnosis was further confirmed using the Conner's Adult ADHD Diagnostic Interview. 24 Patients who were taking drugs (2 with OROS-MPH and 2 with Atomoxetine) were asked to undergo a washout period of at least 48 h before the MRI scan. This study was approved by the Research Ethics Review Board of Peking University Sixth Hospital.
Written informed consent was obtained from all participants.

| MRI data preprocessing and genotyping
The T1-weighted structural MRI (sMRI) and resting-state functional MRI (fMRI) data were analyzed by the Data Processing Assistant for Resting-State fMRI 27 (DPARSF, http://rfmri.org/DPARSF) according to standard preprocessing procedures. The GMV was analyzed by voxel based morphometry (VBM) and sMRI data were re-oriented, segmented into white matter, gray matter, and cerebrospinal fluid using The brief flowchart of the supervised data fusion strategy. This model simultaneously maximizes the inter-modality covariation and correlations of joint independent component and the prior information (the reference). MCCAR, multisite canonical correlation analysis with reference; jICA, joint independent component, IC_ref, joint independent component related to the reference a unified segmentation approach, 28 normalized into the Montreal Neurological Institute (MNI) space, modulated, and smoothed using a 8 mm Gaussian kernel. fMRI data preprocessing included removing first ten volumes, slice timing correction, head motion correction, spatial normalization to the MNI template, resampling to 3 × 3 × 3 mm 3 , smoothing using a 4 mm Gaussian kernel, temporal band-pass filtering (0.01 to 0.1 Hz), nuisance signal regression (including 6 head motion parameters, white matter, cerebrospinal fluid, and global signals), and head motion scrubbing. 29 The FC was analyzed by calculating Pearson correlation coefficients between pairs of node time courses within the whole brain and normalizing to z value using Fisher's transformation.
The peripheral blood genotyping of the NOS1 ex1f-VNTR were performed according to the methods described in Rief et al. 30 We first converted the genotypes to short (16-27 "GT" repeats, S) and long (28-36 "GT" repeats, L) alleles followed the methods described in Weber et al., 18 then counted three common repeats with relative high frequency at single allele level, including 25-repeat (25R), 26-repeat (26R), and 30-repeat (30R). For example, three genotypes were defined according to the allele accounts of 25R: 25R + /25R + indicated the carriers with two 25R, while 25R + /25R − as carriers with one 25R, and 25R − /25R − as carriers without 25R.

| Fusion with reference
Multisite canonical correlation analysis with reference plus joint independent component analysis (MCCAR + jICA) is a supervised, goal-directed model that enables a priori information (eg, cognitive function and symptoms) as a reference to guide multimodal data fusion. This model can maximize the inter-modality covariation and precisely identify the co-varying multimodal component closely related to reference information, which may not be detected by a blind (unsupervised) multimodal fusion approach. In this study, the GMV feature and FC feature were jointly analyzed by "MCCAR + jICA" in which "inhibition" was set as the reference to guide the joint decomposition of this two MRI features. For the code and more information about this method, please refer to the Fusion ICA Toolbox (FIT, https://www.nitrc.org/proje cts/fit) and the methodology paper. 11 Thirty components were estimated according to an improved minimum description length criterion. 31 Two-sample t-tests were further performed on mixing coefficients of each component for each modality. We aimed to investigate the joint independent component (IC), which is significantly correlated with inhibition, co-varying among modalities and group discriminative. The brief flowchart of the supervised data fusion strategy could be found in Figure 1.

| Analyses of "gene (NOS1 ex1f-VNTR)cognition (inhibition)" and "gene (NOS1 ex1f-VNTR)brain (structural/functional patterns)" relationship
The analyses of covariance (ANCOVA) was performed to explore the potential genetic effects of NOS1 ex1f-VNTR on inhibition and its related brain structural/functional alterations, while genotypes, ADHD diagnosis and genotype*ADHD diagnosis interaction were entered into the model with gender and age as covariates. When the main genotypic effect and/or genotype*diagnosis interaction effects existed, genetic association was then analyzed in case and control groups separately.

| Mediation analysis
Once a joint GMV-FC fusion pattern was identified to be group discriminative as well as associated with inhibition performance, and inhibition exhibited significant difference in different NOS1 ex1f-VNTR genotypes, mediation analysis was further conducted to better delineate the gene-sMRI/fMRI-inhibition pathways.
PROCESS macro (http://wwwaf hayes com/publi c/process) in SPSS v.23 was used to conduct these analyses. Bootstrapping was used to estimate the confidence interval (CI) of sampling distributions and the indirect effects of mediators. In this study, bootstrapping sampled from the data 5000 times and provided 95% CI of indirect effects. If the 95% CI does not contain zero, it suggests that the mediating effect is significant at p < 0.05 level. In all mediation analysis, age and gender were controlled as covariates.

| Demographic information and Clinical characteristics
Detailed information of the participants recruited in the current study was shown in Table 1. Twelve patients had a history of major depressive disorder and four patients had a history of anxiety disorder. One patient had ever taken osmotic release system methylphenidate (OROS-MPH) for two months in 2013. Four patients were taking drugs (2 with OROS-MPH and 2 with Atomoxetine). Age, gender, and FIQ were not significantly different between aADHD and HC. As expected, aADHD were with higher core symptoms than HC, including inattentive, hyperactive/impulsive, and total symptoms.

| Inhibition-guided multimodal co-varying imaging patterns and their association with ADHD core symptoms and inhibition-related cognitive functions
The spatial maps of the joint component (denoted as IC_ref, with the same IC order in both structural and functional modalities) were transformed into Z scores, visualized at |Z| > 3 as in Figure 2A  In addition to "inhibition" function, we also conducted correlation analyses for ADHD core symptoms and other multiple inhibitionrelated cognitive domains such as emotional control, shifting, and self-monitoring which constituted the BRI of BRIEF_A to further identify the reliability of the findings to some extent. As shown in Figure 2D, IC_ref loadings in FC were significantly linked with inattentive (r = 0.24, p = 0.020 † ), hyperactive/impulsive (r = 0.26, p = 0.012 † ), and total scores (r = 0.29, p = 0.004 † ) after controlling for age, gender, and diagnosis. Increased IC_ref loadings in GMV were significantly correlated with higher inattentive (r = 0.21, p = 0.043) and marginally significantly correlated with total scores (r = 0.18, p = 0.080) after controlling for age, gender, and diagnosis.
Both GMV loadings and FC loadings showed significantly positive correlation with self-monitoring and BRI in ADHD patients after controlling for age, gender, and diagnosis. Moreover, GMV loadings were significantly correlated with emotional control (more details see Table S2).

| Genetic effects of NOS1 ex1f-VNTR genotypes on inhibition and its related brain structural/functional patterns
Marginally significant interaction of NOS1 ex1f-VNTR genotypes (based on the allele counts of 25R) and ADHD diagnosis was found for inhibition function (F = 2.82, p = 0.067). When analyzing in ADHD and controls separately, significant association was only found in aADHD (F = 3.76, p = 0.032). As shown in Figure 3A, post hoc analysis showed that "Inhibit" score in 25R + /25R + carriers was significant lower than that in 25R − /25R − [(14.63 ± 1.02) versus   Figure 3B). No significant and/or marginally significant association was found for GMV.

| Mediation effects
Based on the above findings of "gene-cognition," "gene-brain," and "brain-cognition" relationship, the following mediation analyses

| DISCUSS ION
This study provided several new insights into the relationships between inhibition performance, related brain structural/functional alterations, and NOS1 ex1f-VNTR genotypes in aADHD. First, we attempt to reveal how abnormally structural and functional neuroimaging features jointly contribute to inhibition impairment in aADHD.
To the best of our knowledge, this is the first study to explore the inhibition-guided multimodal brain imaging fusion in aADHD.
Second, we explored how genetic variants and brain structural/ functional alterations worked relatedly to cause inhibition deficit by constructing a potential "gene-brain-cognition" pathway. Our findings indicated that both the inhibition network and the DMN were involved in inhibition deficits for aADHD. In addition, the FC might mediate the relationship of the functional genetic variants, NOS1 ex1f-VNTR and inhibition, as well as the relationship of the GMV and inhibition in aADHD.
Recently, more and more studies were designed to explore brain morphometric and functional abnormalities and alterations in aADHD.

F I G U R E 2
The identified joint independent component. The joint independent component (A) indicated significant group difference (B) and association with inhibition function (C) and ADHD core symptoms (D). IC_ref_GMV, the reference guided joint independent component_ gray matter volume; IC_ref_FC, the reference guided joint independent component_functional connectivity; aADHD, adults with attentiondeficit/hyperactivity disorder; HC, healthy controls

F I G U R E 3
The relationship of NOS1 ex1f-VNTR (25R), inhibition, and inhibition-directed FC in aADHD. Genetic effect of 25R was found on inhibition (A) and inhibition-directed FC pattern (B), and the FC-mediated NOS1 ex1f-VNTR (25R) and inhibition in aADHD (C). NOS1 ex1f-VNTR, nitric oxide synthase1 variable number of tandem repeats in exon 1f; aADHD, adults with attention-deficit/ hyperactivity disorder, FC: functional connectivity aADHD-related morphometric alterations and abnormal activation emerged in large-scale networks, predominantly located within the frontoparietal, visual, dorsal attention, and default mode networks and so on. [32][33][34][35] In analyses limited to inhibition tasks, aADHD-related abnormal activation included several frontal regions bilaterally, parietooccipital cortex, the right superior temporal gyrus, the left inferior occipital gyrus, and the right thalamus. 32 Due to the limitation of analytical methods, there were a few researches focusing on structural anomalies related to inhibition and the results were inconsistent. 10 In the study, a supervised, goal-directed fusion model was used to explore the underlying structure-function fusion pathogenesis of inhibition in aADHD. "Inhibit" score was used as a specific prior reference to guide the fusion in this study. Results showed that the gradual addition of GMV and FC variables could improve the interpretation level of inhibition variability (more details see the supplementary results and Table S3), indicating multi-modalities provided more enriched imaging information than unimodal pattern. In addition, it was able to extract more sensitive structure-function dysfunctions associated with the inhibition itself than unsupervised fusion approaches. 11,36,37 Specifically, the modality-common co-  38 Anterior cingulate activation was considered to be error-related in a conflict situation, or even a more general role as a conflict detector, independent of whether an error occurred. 39 The parietal lobes are thought to reflect reorientation of attention and the maintenance of task sets during inhibition process. 6,40 The lateral cerebellum projects to the motor area through the thalamus and to the prefrontal cortex, which may contribute to both the preparation and inhibition of movement. 9,41 In addition to abnormal alternations of above regions within the inhibition network, failed suppression of irrelevant network (the DMN) was key for inhibition impairments. 13,42,43 Behavioral inhibition depends on the integrity of the inhibition network and the DMN.
Not only reduced GMV and FC within the inhibition network, but also increased GMV and FC within the DMN were involved in inhibition deficits in aADHD. This implicated the abnormalities observed in ADHD patients affected not only isolated brain structure regions but also the functional connectivity between these brain regions, thus providing a structure-function fusion neural mechanism for inhibition deficits in aADHD. A central assumption of neuroscience is that brain structure could predict and/or is linked to brain function. 44 Accordingly, we have also conducted mediation analyses for the potential "GMV →FC→ Inhibition" model. As expected, the GMV of F I G U R E 4 The relationship of inhibition, inhibition-directed GMV, and inhibition-directed FC in aADHD. GMV could affect inhibition not only directly, but also through the mediation effect of the FC in aADHD. aADHD, adults with attention-deficit/hyperactivity disorder; GMV, gray matter volume; FC, functional connectivity the abovementioned regions could play a role not only directly but also through the mediation of FC in inhibition impairment. As far as we know, this is the first attempt to guide multimodal imaging data fusion with cognitive scores as a reference to seek potential multimodal neural markers of inhibition deficits in ADHD.
In our present study, we have found a marginal genetic association of NOS1 ex1f-VNTR with impaired inhibition function in aADHD, which was consistent with previous report in some extent. 45 However, we should note that the potentially risk allele indicated in our present study was the "25R," but not the short allele (S) or the "21R" reported previously in the Caucasian populations. 18 One potential reason might be the ethnic difference. When comparing the allelic distribution, we could found that the most common alleles in Chinese Han subjects were 25R, 26R, and 30R, while that in Caucasian subjects were 20R, 25R, and 31R. 18 Similarly, the negative association for the short allele (S) in our study may be due to the potentially different definition of S or L allele in Chinese Han subjects which needs future functional analyses. Further imaging genetic analyses indicated that this gene-cognition association was potentially mediated by the brain resting-state functional alteration in inhibition and default mode networks. In a previous report, significant decreased activity in dorsolateral prefrontal cortex and inferior frontal cortex was found for risk short allele carriers of NOS1 ex1f-VNTR during inhibition process. 19 Reif et al. speculated that the hypoactivation in prefrontal cortex in risk allele carriers, leading to compromised inhibition processes, might be due to decreased NO production as a consequence of reduced expression of NOS1 gene. 46 Besides dorsolateral prefrontal cortex and inferior frontal cortex, previous studies reported abnormal anterior cingulate cortex, parietal cortex and basal ganglia activation in risk NOS1 allele carriers, 20,21,30,46 which were all the central regions for inhibition. Consistent with these results, our results showed that NOS1 ex1f-VNTR affected inhibition through the mediation of FC of the inhibition network and DMN.
There are some limitations in current study. Firstly, the sample size in genetic subgroup is relatively small, which might lead to marginal findings in genetic association analyses. In addition, we did not perform analyses for males and females separately, but just set as covariate. A larger sample size may lead to more robust results and help us to illustrate the potential gender-specific relationship in the future. Secondly, fusion modal included only two modalities of GMV and FC in this study. Inhibition may also be affected by white matter integrity, local spontaneous brain activity, or cerebral blood flow. The fusion of these modalities would provide more informative knowledge of inhibition and could be studied in the future.

| CON CLUS IONS
Our results showed that reduced GMV and FC in the inhibition network and increased GMV and FC in the DMN were jointly responsible for inhibition deficits in aADHD. More importantly, both the functional NOS1 ex1f-VNTR variant and the GMV might affect inhibition through the mediation of the abnormal FC of these brain regions, identified a potential gene-brain-cognition pathway for inhibition in aADHD. This multimodal fusion method may be helpful in searching for meaningful biological markers and cross-information links in multiple neuroimaging modalities. Integration of multimodal brain imaging and individual-level genetic data delineates a new way for elucidating the pathophysiology and provides potential biomarkers of inhibition impairments.

ACK N OWLED G EM ENT
We are glad to thank all the subjects who participate in this research.

CO N FLI C T O F I NTE R E S T
All authors reported no potential conflicts of interest.

DATA AVA I L A B I L I T Y S TAT E M E N T
The raw data required to reproduce these findings cannot be shared at this time as the data also forms part of an ongoing study.