Investigating microstructure of white matter tracts as candidate endophenotypes of Social Anxiety Disorder - Findings from the Leiden Family Lab study on Social Anxiety Disorder (LFLSAD).

BACKGROUND: Social anxiety disorder (SAD) is a mental illness with a complex, partially genetic background. Differences in characteristics of white matter (WM) microstructure have been reported in patients with SAD compared to healthy controls. Also, WM characteristics are moderately to highly heritable. Endophenotypes are measurable characteristics on the road from genotype to phenotype, putatively reflective of genetically based disease mechanisms. In search of candidate endophenotypes of SAD we used a unique sample of SAD patients and their family members of two generations to explore microstructure of WM tracts as candidate endophenotypes. We focused on two endophenotype criteria: co-segregation with social anxiety within the families, and heritability.
METHODS: Participants (n = 94 from 8 families genetically vulnerable for SAD) took part in the Leiden Family Lab Study on Social Anxiety Disorder (LFLSAD). We employed tract-based spatial statistics to examine structural WM characteristics, being fractional anisotropy (FA), axial diffusivity (AD), mean diffusivity (MD) and radial diffusivity (RD), in three a-priori defined tracts of interest: uncinate fasciculus (UF), superior longitudinal fasciculus (SLF) and inferior longitudinal fasciculus (ILF). Associations with social anxiety symptoms and heritability were estimated.
RESULTS: Increased FA in the left and right SLF co-segregated with symptoms of social anxiety. These findings were coupled with decreased RD and MD. All characteristics of WM microstructure were estimated to be at least moderately heritable.
CONCLUSION: These findings suggest that alterations in WM microstructure in the SLF could be candidate endophenotypes of SAD, as they co-segregated within families genetically vulnerable for SAD and are heritable. These findings further elucidate the genetic susceptibility to SAD and improve our understanding of the overall etiology.

Introduction

Social anxiety disorder (SAD) is a highly prevalent mental illness with a typical onset during late childhood and early adolescence (Bas-Hoogendam et al., 2020a). Patients with SAD experience an excessive fear of negative evaluation in social situations, which are either avoided or endured with intense fear or anxiety (American Psychiatric Association, 2013). SAD often has a chronic course and high comorbidity rates with other mental disorders (Beesdo-Baum et al., 2012, Blanco et al., 2011, Fehm et al., 2005, Grant et al., 2005, Koyuncu et al., 2019, Steinert et al., 2013). Unfortunately, the wide range of cognitive behavioral and pharmacological therapies available proves to be insufficient in 30–40% of patients (Stein & Stein, 2008). To improve current therapies and develop new interventions, the etiology of SAD should be further elucidated (Brühl et al., 2014, Fox and Kalin, 2014).

In the past decades, research into the neurobiological background of mental disorders has increased and evolved, with a special interest in neuroimaging of the brain. Indeed, several mental disorders, including SAD, have been associated with altered brain functioning (Brühl et al., 2014, Jenkins et al., 2016). For example, previous research in patients with SAD has reported abnormalities in the activity of the anxiety circuitry and its regulatory networks, such as hyperactivation of the amygdala and insula in relation to negative emotional stimuli (Etkin & Wager, 2007) and decreased functional coupling between the amygdala and regulatory prefrontal cortical regions such as the orbitofrontal cortex (Hahn et al., 2011). In addition, altered resting state functional connectivity has been reported, such as decreased connectivity between the amygdala and frontal regions such as the medial, dorsolateral and ventrolateral prefrontal cortex (PFC) and between regions implicated in the default mode network like the medial PFC and the lateral parietal cortex (reviewed in MacNamara et al. (2016)).

A recent neurofunctional model on SAD suggested disturbed emotion regulation networks in patients with SAD, with an imbalance between hyperactive parietal and medial occipital brain regions and fear circuitry on the one hand, and impairments in regulatory control networks in frontal areas on the other hand. This imbalance might be partly attributed to impaired communication between these areas due to decreased structural connectivity, such as abnormal microstructure of several white matter (WM) tracts (Brühl et al., 2014).

A method to study characteristics of WM microstructure of the brain is diffusion tensor imaging (DTI), which in turn can be utilized to model four commonly used scalar measures of diffusivity: fractional anisotropy (FA), axial diffusivity (AD), radial diffusivity (RD) and mean diffusivity (MD). Differences in microstructure of several WM tracts, such as the uncinate fasciculus (UF), superior longitudinal fasciculus (SLF) and inferior longitudinal fasciculus (ILF), have been reported frequently in Magnetic Resonance Imaging (MRI) studies in SAD. For an overview of current literature on DTI studies in SAD, see Supplemental Table 1a and 1b, summarizing the work by Baur et al., 2011, Baur et al., 2013, Jenkins et al., 2016, Liao et al., 2011, Phan et al., 2009, Qiu et al., 2014, Tükel et al., 2017. The UF connects the prefrontal cortex with the anteromedial temporal lobe and is thought to be, among others, involved in social emotional processing (Von Der Heide et al., 2013). In voxelwise analyses, decreased FA of the left and right UF has been previously reported in SAD patients compared to healthy controls (Baur et al., 2011, Phan et al., 2009, Qiu et al., 2014) and one study replicated these results using tractography analyses (Baur et al., 2013). The SLF is roughly divided in three subparts (I, II and III) and is the major tract connecting parietal cortices to prefrontal areas (Schmahmann et al., 2007). It is thought to subserve visuo-spatial attention and language functionality (Schmahmann & Pandya, 2006), and in voxelwise studies decreased FA of the left SLF has been reported in patients with SAD (Baur et al., 2011, Tükel et al., 2017). Interestingly, this finding was replicated in a voxelwise meta-analysis of FA (Jenkins et al., 2016). The ILF originates from medial and lateral anterior temporal areas and terminates in the occipital lobe. This tract is suggested to play a role in facial recognition (Catani et al., 2003, Maller et al., 2019, Panesar et al., 2018) and one voxelwise study reported decreased FA in patients with SAD compared to healthy controls (Tükel et al., 2017). Interestingly, when using a tract-of-interest (TOI) approach, an increase in average FA of the bilateral UF and right ILF has been reported after 10 weeks of cognitive behavioral group therapy (Steiger et al., 2017) and higher values of a combined tractography measure of AD, MD, RD and FA of the right ILF has been found to predict better clinical response to cognitive behavioral therapy in individual and group settings (Whitfield-Gabrieli et al., 2016). Also, an inverse voxelwise relation between FA in a putative amygdala-prefrontal pathway and trait anxiety has been reported (Kim & Whalen, 2009; Kim et al., 2016). Interestingly, the three WM tracts mentioned above are also thought to be involved in several resting-state functional brain regions and networks in which altered connectivity has been reported in SAD patients, such as regions involved in the DMN, the ventral attention network and the fronto-parietal network, (MacNamara et al., 2016, Peer et al., 2017, Sylvester et al., 2012). These results suggest that abnormal WM microstructure, together with other underlying neurobiological processes, could be linked to SAD.

It is thought that the overall vulnerability to develop SAD is based on complex interactions between genetic (dis)advantages and liabilities, epigenetic factors and environmental factors (Bas-Hoogendam et al., 2020a, Spence and Rapee, 2016, Wong and Rapee, 2016). For example, studies using twins and families indicated that genetic and non-shared environmental factors explained roughly equally most of the individual differences in SAD (Scaini et al., 2014) and a recent genome wide association study (GWAS) analysis confirmed a heritable basis of SAD (Stein et al., 2017). To study the genetic vulnerability to SAD more closely, an endophenotype approach could be used. Endophenotypes are defined as biological or psychological markers of a disorder, which are thought to be in the causal chain between genetic contributions to a disorder and diagnosable symptoms of psychopathology (Gottesman and Gould, 2003, Lenzenweger, 2013). The endophenotype approach assumes that underlying measurable components of a mental illness are heritable and present before the development of clinical symptoms and include, for example, neurobiological changes in brain structure and function. It is thought that endophenotypes are theoretically capable of providing greater statistical power to localize and identify disease related biomarkers than affection status alone (Glahn et al., 2019). To be considered an endophenotype, a candidate psychoneurobiological characteristic has to be associated with the disorder (criterion 1), state-independent (criterion 2), heritable (criterion 3) and co-segregate with the disorder within families of probands whilst already present in a preclinical state (criterion 4) (Gottesman and Gould, 2003, Lenzenweger, 2013, Miller and Rockstroh, 2013, Roffman, 2019).

Several neurobiological candidate endophenotypes for SAD have been proposed, such as WM and grey matter (GM) characteristics, amygdala and prefrontal brain function and alterations in functional connectivity networks (Bas-Hoogendam et al., 2016). The Leiden Family Lab study on Social Anxiety Disorder (LFLSAD) is to our knowledge the first comprehensive two-generation family neuroimaging study on SAD and has been designed specifically to examine the heritability and first part of the co-segregation criteria of candidate endophenotypes of SAD (Bas-Hoogendam et al., 2018a). Previous results of this study suggest that several characteristics of brain structure and function, like cortical and subcortical GM characteristics, increased and prolonged amygdala activation, and increased brain activity whilst processing unintentional social norm violations, could be endophenotypes of SAD (Bas-Hoogendam et al., 2018b, Bas-Hoogendam et al., 2019b, Bas-Hoogendam et al., 2019c, Bas-Hoogendam et al., 2020b, Bas‐Hoogendam et al., 2019a). However, it still remains to be elucidated whether characteristics of WM microstructure could be candidate endophenotypes of SAD. Results of previous studies suggest that this could indeed be the case, as they were found to be associated with the disease in case-control studies (endophenotype criterion 1) and are at least moderately (h2 ≥ 0.2) to highly (h2 ≥ 0.6) heritable as shown by multiple studies in healthy twins (endophenotype criterion 3) (Budisavljevic et al., 2016, Kochunov et al., 2015). However, to our knowledge the criterion of co-segregation within families of probands (criterion 4, first element) has not been examined in previous work.

Using DTI data from the LFLSAD, the present study investigated characteristics of WM microstructure as candidate endophenotypes of SAD, with a focus on: 1) co-segregation within families genetically enriched for SAD and 2) estimation of heritability of WM characteristics. We employed tract-based spatial statistics (TBSS; (Smith et al., 2006)) to examine the association between measures of WM microstructure and measures of social anxiety. Our main parameter of interest was FA, while we used other parameters, being AD, MD and RD, to gain more insight into underlying WM microstructure (cf. Aghajani et al., 2014, van der Werff et al., 2014). Based on previous research we expected to find a negative association between the level of social anxiety symptoms and FA and AD in the UF, SLF and ILF, coupled with a positive association between clinical symptoms and RD and MD. Furthermore, we expected estimates of all WM parameters to be at least moderately heritable. We employed a-priori defined TOI analyses in these three tracts using two types of analyses: a voxelwise analysis and an analysis of the averaged values of WM parameters over the whole TOI. In addition, we performed an exploratory voxelwise analysis of the whole WM skeleton to investigate WM microstructure outside the a-priori defined regions.

Methods and materials

Participants

The LFLSAD is a multiplex (families were selected based on a minimum of two (sub)clinical SAD cases within one nuclear family) and multigenerational (multiple nuclear families encompassing two generations from the same family participated) family study on SAD (total sample: n = 132, from nine families, MRI participants n = 113), designed to investigate candidate neurobiological endophenotypes of SAD. A design like this is especially powerful to investigate environmental and genetic influences on SAD-related characteristics (Bas-Hoogendam et al., 2018a). The background, objectives and methods as well as clinical characteristics of the sample and an a priori power analysis are described in more detail elsewhere (Bas-Hoogendam et al., 2018a). Preregistration of the study is available on . The LFLSAD sample consists of families selected on presence of a primary diagnosis of SAD in a parent (25 – 55 years old; ‘proband’), with a child who met the criteria for clinical or subclinical SAD (living at home, 8 – 21 years old; ‘proband’s SA child’). Furthermore, the partner of the proband, other children of this nuclear family (≥8 years of age), siblings of the proband and their partners and children (≥8 years of age) were invited to participate. Thus, the sample consisted of two generations of family members: the generation of the proband (generation 1) and the generation of the proband’s SA-child (generation 2; see Fig. 1). Exclusion criteria for the proband or proband’s SA child were comorbidity other than internalizing disorders; other family members were included regardless of the presence of psychopathology. Exclusion criteria for all participants for the MRI experiment were general MRI contraindications, such as metal implants or pregnancy.

Fig. 1

Family structure in the LFLSAD. Example of a family within the Leiden Family Lab study on Social Anxiety Disorder. Families were included based on the combination of a parent with social anxiety disorder (SAD; “proband”: depicted in red) and a proband's child with SAD (red) or (sub)clinical SAD (orange). In addition, family members of two generations were invited, independent from the presence of SAD within these family members (no SAD: light blue; did not participate: gray). Grandparents (Generation 0; white) were not invited for participation. This family is slightly modified to guarantee anonymity; however, the number of family members and the frequency of (sub)clinical SAD are depicted truthfully. Squares and circles represent men and women, respectively. This figure is a reprint of the figure published in (Bas-Hoogendam et al., 2018a). (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

MRI data from one family (n = 3 family members) had to be excluded from the present analysis as the proband from this family was not able to participate in the MRI experiment due to an MRI contraindication. DTI data from two subjects (9.4 y and 18.5 y) were not available due to an early stop of data acquisition during the MRI experiment, which consisted of several structural and functional scans as described elsewhere (total duration of the MRI protocol: 54 min 47 s) (Bas-Hoogendam et al., 2018a). Thus, data from n = 108 participants was available for initial DTI analysis.

Ethics

The LFLSAD study was approved by the Medical Ethical Committee of the Leiden University Medical Center (P12.061). All participants provided informed consent according to the Declaration of Helsinki; both parents signed the informed consent form for their children, while children between 12 and 18 years of age signed the form themselves as well.

Phenotyping

Confinement of diagnosis

o determine the presence of DSM-IV diagnoses, with special attention to SAD, the Mini-International Neuropsychiatric Interview (MINI)-Plus (version 5.0.0) or MINI-Kid (version 6.0) was used by experienced clinicians and was voice-recorded for review (Bauhuis et al., 2013, Sheehan et al., 1998, Sheehan et al., 2010, van Vliet and de Beurs, 2007). Diagnosis of clinical SAD was determined using the DSM-IV-TR criteria for the generalized subtype of SAD, but a clinician verified that DSM-5 criteria were also met to establish the diagnosis. Subclinical SAD was diagnosed when DSM-5 criteria were met, but important areas of functioning were not impaired (criterion G) (American Psychiatric Association, 2013).

Questionnaires

All participants completed self-report questionnaires regarding anxiety-related symptoms. If applicable, age-appropriate questionnaires were used. Among others, we measured social anxiety symptoms, using the Liebowitz Social Anxiety Scale (LSAS-SR) for adults (≥18 years of age) (FRESCO et al., 2001, Mennin et al., 2002) and the Social Anxiety Scale for Adolescents (SAS-A) for younger participants (<18 years of age) (La Greca & Lopez, 1998), the intensity of fear of negative evaluation using the revised Brief Fear of Negative Evaluation (BFNE – II scale) (Carleton et al., 2006, Leary, 1983) and the level of trait anxiety using the State-Trait Anxiety Inventory (STAI) (Spielberger et al., 1970). Furthermore, depressive symptoms were evaluated by the self-report Beck Depression Inventory (BDI-II) for adults (Beck et al., 1996, Van der Does, 2002) or the Children’s Depression Inventory (CDI) for adolescents (Kovacs, 1985). To analyze the scores of the age-appropriate self-report questionnaires, z-scores were computed for the level of social anxiety symptoms and depressive symptoms as described previously (Bas-Hoogendam et al., 2018a).

Incidental missing values on the self-report questionnaires were replaced by the mean value of the completed items. Data on the BFNE-II was missing for one participant. Differences in scores on self-report questionnaires between participants with (sub)clinical SAD and non-affected relatives were assessed by fitting regression models in R (R Core Team, 2019). Within these models, outcomes from self-report questionnaires (levels of social anxiety symptoms (z-score; z-SA), intensity of fear of negative evaluation (FNE), depressive symptoms (z-score) and trait anxiety) were modelled as dependent variables and (sub)clinical SAD as independent variable. Genetic correlations between family members were accounted for by including random effects in the models, and gender and age (centered) were added as covariates. The Bonferroni method was used to correct p-values for multiple comparisons (4 tests, corrected p-value = 0.0125).

MRI data acquisition

Scanning was performed at Leiden University Medical Center using a 3.0 T Philips Achieva MRI scanner (Philips Medical Systems, Best, The Netherlands), equipped with a 32-channel Sensitivity Encoding (SENSE) head coil. The MRI session consisted of several structural and functional scans, as described elsewhere (total duration of the MRI protocol: 54 min 47 s) (Bas-Hoogendam et al., 2018a). Of interest for the present study are two diffusion weighted imaging (DWI) scans with the following characteristics: repetition time (TR) 7316 ms, echo time (TE) 69 ms, field of view (FOV) 240 × 240 × 150 mm, acquisition matrix 128 × 128 with 75 slices, slice thickness 2 mm, voxel size 1.9 × 2.4 mm. DWI scans (b = 1000 s/mm2) were acquired in 30 directions with two additional non-DWI scans (b = 0 s/mm2) serving as reference scans. The two DWI scans were collected with reversed phase-encode blips, resulting in pairs of images with distortions going in opposite directions (anterior-posterior and posterior-anterior). Scan duration for each scan was about 4 min 30 s. Furthermore, one high-resolution T1-weighted structural scan was acquired with the following characteristics: 140 slices, resolution 0.875 mm × 0.875 mm × 1.2 mm, FOV = 224 mm × 168 mm × 177.333 mm, TR = 9.8 ms, TE = 4.59 ms, flip angle = 8°. All structural MRI scans were inspected by a neuroradiologist. No clinically relevant characteristics were reported in any of the participants.

Data analysis

DTI processing

DTI models diffusivity of water molecules across the brain, using tensors. These tensors consist of three main eigenvalues (λ1, λ2 and λ3) which in turn can be used to calculate the four most commonly used characteristics of WM microstructure: fractional anisotropy (FA), axial diffusivity (AD), radial diffusivity (RD), and mean diffusivity (MD) (Alexander et al., 2007). FA provides a relative difference between the largest eigenvalue as compared to the others, reflecting the tendency of water molecules to diffuse in one direction as opposed to all others and could therefore be described as a general indicator for WM microstructure (e.g. myelin thickness, membrane integrity) (Alexander et al., 2007, Hasan et al., 2004). AD is defined as the first eigenvalue (λ1) and reflects water diffusion along the principal direction of the fiber, displaying fiber bundle coherence and axonal integrity (Budde et al., 2009). RD is defined as the average of the second and third eigenvalue (λ2 and λ3) and reflects water diffusion perpendicular to the principal direction of the fiber, thus being more indicative of the level of myelinization (Song et al., 2005). MD is defined as the average of the three eigenvalues and hence reflects average water diffusion in all directions within a fiber, thus putatively reflective of a degree of myelination (Horsfield & Jones, 2002). In general, decreased FA is coupled with decreased AD and / or increased RD and MD and vice versa (Alexander et al., 2007, Kochunov et al., 2007).

Image pre-processing and analyses were performed using the Oxford Centre for Functional Magnetic Resonance Imaging of the Brain (FMRIB) Software Library (FSL) (Smith et al., 2004). The susceptibility-induced off-resonance field from the two pairs of DWI images was estimated using a method similar to that described in (Andersson et al., 2003) and the two images were combined into a single corrected one. Afterwards, the Brain Extraction Tool (BET) was used to remove non-brain tissue from the non-diffusion images (Smith, 2002). Image distortion and motion artefacts induced by eddy currents or inter-volume head motions were corrected (Andersson & Sotiropoulos, 2016) and image quality was statistically evaluated afterwards (Bastiani et al., 2019). No outliers were detected. Individual FA images and primary (λ1), secondary (λ2) and tertiary (λ3) eigenvalues were created by fitting a tensor model to the raw diffusion data using FMRIB’s Diffusion Toolbox (FDT) (Behrens et al., 2003). Individual maps of other diffusivity measures were calculated out of eigenvalues, defining AD as λ1, RD as λ23 = (λ2 + λ3)/2 and MD as λ123 = (λ1 + λ2 + λ3)/3. The individual vector and raw FA images were visually and statistically evaluated for alignment on WM tracts according to standardized protocols, designed to facilitate harmonized image analysis across multiple sites (http://enigma.ini.usc.edu/protocols/dti-protocols/).

Then, a study-specific custom FA template was created. All subjects' FA data were slightly eroded and aligned into a common space using the nonlinear registration tool FNIRT (Andersson et al., 2007a, Andersson et al., 2007b), which uses a b-spline representation of the registration warp field (Rueckert et al., 1999). Afterwards, a mean FA image and distance map to the masked template were created. Individual FA and non-FA (AD, MD and RD) images were then projected onto the template. Subsequently, quality control was performed twofold: we visually inspected the registered images for misalignment onto the skeleton, and individual projection distances of the extracted skeletons onto the template were calculated to detect outliers (defined as individual projection distance to the template exceeding the threshold of 3.8 mm), which could represent bad alignment to the template (Acheson et al., 2017). All images were well aligned, and no outliers were detected.

Investigation of WM candidate endophenotypes

We focused on two endophenotype criteria: (i) co-segregation of the candidate endophenotypes with the disorder within families (the first element of criterion 4) and (ii) heritability estimation (h2) of characteristics of WM microstructure (criterion 3). Co-segregation of WM characteristics within families was examined by exploring the relationship between parameters of WM microstructure and two dimensional measures of social anxiety (z-SA and FNE) using three a priori defined WM tracts based on previous literature: the SLF, ILF and UF (see Supplemental Fig. 1 for TOIs, for an overview of current literature, see Supplemental Table 1a and 1b). To investigate WM microstructure outside the a-priori defined regions, we also examined the association between WM parameters and a diagnosis of (sub)clinical SAD (see Supplemental section).

Selection of tracts

Binary unilateral masks of the SLF, UF and ILF were created using the probabilistic Johns Hopkins University (JHU) white-matter tractography atlas (Mori et al., 2005) provided by FSL, thresholded at a conventional 20% (Smith et al., 2006). Using the mean FA skeleton, each tract was unilaterally masked to include only voxels comprised in both the tract and the skeleton. This confines the statistical analysis to voxels from the center of the tract, thereby minimizing anatomic intersubject variability, deviations in registration and partial volume effects (Westlye et al., 2009).

Co-segregation

To examine co-segregation of WM characteristics, voxelwise and average TOI analyses were conducted. We report findings uncorrected for the number of tracts because of the use of a priori defined TOIs, which are possibly also functionally related (Burkhouse et al., 2017) and because of the innovative and more explorative nature of the present study (to the best of our knowledge, this is the first comprehensive family study on SAD). To investigate WM microstructure outside the a-priori defined regions, an exploratory analysis was conducted to investigate voxelwise associations between FA in the WM skeleton and the level of self-reported symptoms of SAD was conducted using NINGA (see Supplemental section for more details and results).

For the voxelwise and average TOI analyses, sensitivity analyses were conducted to control for (i) psychopathology other than SAD and (ii) severity of depressive symptoms. Therefore, all participants with past and/or present (comorbid) psychopathology other than SAD were excluded (sensitivity analysis 1); or the z-score of the level of depressive symptoms was added as a covariate in the analyses (sensitivity analysis 2). Details and results of these analyses are included in the Supplemental section.

Voxelwise TOI analysis

First, we performed voxelwise analyses within each of the TOIs to examine subtle localized differences which could disappear in average TOI analyses due to the size of the TOI. In these voxelwise TOI analyses, FA was used as principal outcome measure as this is a general indicator of WM microstructure. Additional WM parameters, being AD, MD and RD, were examined for significant clusters only, to provide complementary information about WM microstructure and to aid interpretation of FA changes. The voxelwise TOI analyses were conducted by performing multiple nonparametric regression analyses using the NeuroImaging Nonparametric Genetic Analysis (NINGA) toolbox (Ganjgahi et al., 2015, Ganjgahi et al., 2018) as methods previously used in analyses of fMRI data of the LFLSAD could, due to the specific structure of TBSS data and assumptions regarding the random field theory, not account for family-wise errors (FWE). NINGA implements linear mixed effect for covariate inference in presence of family relatedness using an approximate non-iterative random effect estimator based on restricted maximum likelihood function. It uses permutation test to provide essential spatial statistics inference tools for uncorrected and family-wise error (FWE) corrected p-values. Individual levels of self-reported social anxiety (z-SA and FNE) were modelled as independent variables and voxelwise FA values as the dependent variable. Covariate inference was incorporated in the model to account for nuisance kinship, gender and age. We used Threshold-Free Cluster Enhancement (TFCE) statistics to define significant clusters, and permutation testing to provide FWE corrected p-values at a conventional threshold of α = 0.05.

To gain more insight in the direction of FA in clusters displaying significant associations with SA-symptoms, RD, MD and AD were examined by extracting and binarizing the significant cluster from the previous analysis. This mask was then used to comprise only the relevant voxels in individual skeletons of AD, MD and RD. Next, individual levels of self-reported social anxiety symptoms (z-SA and FNE) were modelled as independent variables and values of AD, MD or RD in the relevant cluster as the dependent variable. Covariate inference was incorporated to account for nuisance kinship, gender and age.

Average TOI analysis

Second, we conducted conventional TOI analyses using average individual values of WM parameters over the whole tract to allow comparison with previous literature (“average TOI analyses”) and to ensure continuity with previous analyses performed on data from the LFLSAD (Bas-Hoogendam et al., 2018b). That is, following methods previously described (Bas-Hoogendam et al., 2018b, Bas-Hoogendam et al., 2019b, Bas-Hoogendam et al., 2020b, Bas‐Hoogendam et al., 2019a), associations between individual average values of WM parameters (FA, MD, RD and AD over the whole tract) per tract and clinical symptoms of SAD (z-SA and FNE) were examined by performing multiple regression analyses using linear mixed models in R (R Core Team, 2019). Average values of FA, AD, MD and RD were extracted for each individual unilaterally per tract. Per TOI, mean WM parameters were modelled as dependent variables and the outcomes of self-report questionnaires as independent variables. Correlations between family members were accounted for by including random effects in the models (Tissier et al., 2017). Both age (centered) and gender (centered) were included as covariates. As most of the dependent variables were non-normally distributed, the robustness of the linear mixed model used was confirmed by checking the distribution of the residuals of the phenotypes with the Shapiro-Wilk normality test and visual inspection, which showed all residuals followed an approximate normal distribution.

Heritability

Next, general heritability (h2) of the WM microstructure characteristics was estimated, using methods previously used in analyses of the LFLSAD sample to ensure consistency (Bas-Hoogendam et al., 2018b, Bas-Hoogendam et al., 2019b, Bas-Hoogendam et al., 2019c, Bas-Hoogendam et al., 2020b, Bas‐Hoogendam et al., 2019a). This method estimates heritability by jointly modelling SAD status and the individual average values of FA, AD, MD, and RD within all TOIs in a multivariate-mixed probit model, by which the familial relationship and ascertainment of the families (based on SAD in the proband and (sub)clinical SAD in the proband’s SA-child) were taken into account (Tissier et al., 2017). To adjust for age and gender, these variables were included as covariates (both centered) in the marginal regression models. Variance of the random effects was determined using maximum‐likelihood estimates; subsequently, heritability was estimated (Tissier et al., 2017).

Results

Sample characteristics

Sample characteristics are summarized in Table 1. Participants with (sub)clinical SAD reported significantly higher levels of social anxiety (self-reported social anxiety symptoms (z-SA) and FNE), depressive symptoms and trait anxiety compared to their non-SAD relatives but did not differ with respect to gender-distribution, generation, age or IQ. For a more elaborate description of this sample, including diagnostic information and details on quality checking and data availability, we refer to the Supplemental section and previous publications on the LFLSAD in general (Bas-Hoogendam et al., 2018a) and the MRI sample in particular (Bas-Hoogendam et al., 2018b, Bas-Hoogendam et al., 2019b, Bas-Hoogendam et al., 2020b, Bas‐Hoogendam et al., 2019a).

Table 1

Characteristics of participants with and without (sub)clinical SAD.

	(Sub)clinical SAD (n = 31)	No SAD (n = 57)	Statistical analysis
Demographics
Male / Female (n)	13 / 18	28 / 29	χ2 = 0.42, p = 0.52
Generation 1 / Generation 2 (n)	19 / 12	27 / 30	χ2 = 1.56, p = 0.21
Age in years (mean ± SD); range	33.7 ± 15.5 (9.2–59.6)	32.9 ± 14.8 (9.6–61.5)	β ± SE = 0.8 ± 3.3, p = 0.80
Estimated IQ (mean ± SD)	102.2 ± 12.2	105.5 ± 10.8	β ± SE = -3.1 ± 2.4, p = 0.21
Diagnostic information (n)
Clinical SAD	15	0	χ2 = 33.3, p < 0.001**
Self-report measures (mean ± SD)
Social anxiety symptoms (z-score)	2.4 ± 3.3	0.7 ± 1.3	β ± SE = 1.9 ± 0.5, p < 0.001**
FNE	23.0 ± 12.4	12.0 ± 7.6	β ± SE = 10.6 ± 2.2, p < 0.001**
Depressive symptoms (z-score)	0.05 ± 0.9	−0.6 ± 0.6	β ± SE = 0.6 ± 0.2, p = 0.001**
Trait anxiety	38.0 ± 9.8	33.2 ± 8.6	β ± SE = 5.0 ± 2.0, p = 0.01*

SAD: social anxiety disorder; FNE: fear of negative evaluation; STAI: state-trait anxiety inventory; SD: standard deviation.

Sample for dimensional analysis: n = 94 for z-SA, n = 93 for FNE. Data on the presence of subclinical SAD were, due to technical reasons, lost for six family members (remaining sample for categorical analysis: n = 88). ** significant at Bonferroni corrected p-value of 0.0125; * significant at uncorrected p-value of 0.05.

Co-segregration

Three bilateral TOIs were examined in voxelwise and average TOI analyses to explore the association between WM microstructure and clinical symptoms, as measured by self-reported levels of social anxiety (z-SA) and intensity of FNE. Both analyses revealed that higher levels of z-SA and FNE were significantly associated with higher FA values and lower MD and RD values in the left and right SLF. These significant findings will be discussed more in depth in the following paragraphs. We did not find any significant correlations between levels of social anxiety or FNE and WM microstructure in the bilateral UF or ILF.

Voxelwise TOI analysis

Significant associations are summarized in Table 2 and illustrated in Fig. 2. A cluster in the left SLF was significantly positively associated with levels of social anxiety and FNE (β = 0.147, p = 0.006 and β = 0.039, p = 0.002 resp.). In addition, a cluster in the right SLF was significantly positively associated with levels of FNE (β = 0.027, p = 0.04). In both clusters, these findings were coupled with significant negative associations between the level of clinical symptoms and MD and RD. AD was not significantly associated with clinical symptoms.

Table 2

Significant associations between measures of SA and voxelwise TOI analyses of fractional anisotropy in the SLF; post-hoc analyses of additional WM parameters within the clusters.

Clinical measure	Side	WM parameter	Voxels (mm3)	Peak MNI coordinates			β	p
				x	y	z
z-SA	L	FA	207	–32	−39	29	0.147	0.006
		AD					0.071	0.37
		MD					−0.012	0.03
		RD					−0.076	< 0.001
FNE	L	FA	178	–33	−38	29	0.039	0.002
		AD					0.017	0.49
		MD					−0.004	0.01
		RD					−0.021	< 0.001
FNE	R	FA	51	34	−30	32	0.027	0.04
		AD					0.013	0.50
		MD					−0.023	< 0.001
		RD					−0.023	< 0.001

Threshold-free cluster enhancement (TFCE) and family-wise error (FWE) corrected at p-values < 0.05. ß-values and p-values represent the outcome of the analyses on mean values of white matter integrity over all voxels. z-SA: social anxiety (z-score); FNE: fear of negative evaluation; FA: fractional anisotropy; AD: axial diffusivity; MD: mean diffusivity; RD: radial diffusivity; L: left; R: right.

Fig. 2

Significant clusters from voxelwise TOI analyses of fractional anisotropy in the superior longitudinal fasciculus. Sagittal, coronal and axial sections of the WM skeleton (blue), with subregions of the superior longitudinal fasciculus (SLF) showing significant associations of fractional anisotropy (FA) with levels of A) social anxiety (z-score) and B) fear of negative evaluation within families genetically enriched for social anxiety disorder (SAD) (p < 0.05, threshold-free cluster enhancement (TFCE) and family-wise error (FWE) corrected (yellow/orange)). The color bar indicates p-values. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

Average TOI analysis

For all TOIs and all four parameters of WM microstructure, associations with clinical symptoms can be found in Table 3. In accordance with findings from the voxelwise TOI analyses, mean FA in the left SLF demonstrated a significant positive association with levels of self-reported social anxiety (β = 0.002, p = 0.03) and a near-significant association with levels of FNE (β = 3.67E-04, p = 0.06). In addition, mean FA in the right SLF showed a significant positive association with levels of intensity of FNE (β = 3.58E-04, p = 0.04). This finding was coupled with a significant negative association with mean RD values (β = −4.39E-07, p = 0.04) and a marginally significant, negative association with mean MD values (β = −3.15E-07, p = 0.09).

Table 3

Associations between measures of SA and average values of white matter parameters in tracts of interest.

TOI	Side	WM parameter	Effect of social anxiety(z-score)			Effect of FNE			Heritability estimate
			β	SE	p	β	SE	p	h2	SE
ILF	L	FA	0.001	0.001	0.58	3.44E-04	2.18E-04	0.11	0.30	0.02
		AD	−4.46E-07	1.98E-06	0.82	−1.37E-07	4.21E-07	0.74	0.99	0.48
		MD	−9.58E-07	1.52E-06	0.53	−3.38E-07	3.25E-07	0.30	0.90	0.13
		RD	−1.20E-06	1.45E-06	0.41	−4.17E-07	3.11E-07	0.18	0.66	0.34
	R	FA	0.001	0.001	0.28	1.89E-04	1.77E-04	0.29	0.41	0.09
		AD	−3.46E-07	1.29E-06	0.79	−1.15E-07	2.61E-07	0.66	0.87	0.07
		MD	−1.09E-06	9.70E-07	0.26	−2.57E-07	1.95E-07	0.19	0.79	0.05
		RD	−1.35E-06	1.05E-06	0.20	−2.66E-07	2.19E-07	0.23	0.57	0.00
SLF	L	FA	0.002	0.001	0.03	3.67E-04	1.92E-04	0.06	0.41	0.05
		AD	1.15E-06	1.33E-06	0.39	2.76E-08	2.75E-07	0.92	0.73	0.21
		MD	−7.33E-07	1.13E-06	0.51	−3.05E-07	2.34E-07	0.19	0.70	0.01
		RD	−1.68E-06	1.21E-06	0.17	−4.57E-07	2.55E-07	0.07	0.59	0.03
	R	FA	0.001	0.001	0.07	3.58E-04	1.76E-04	0.04	0.31	0.02
		AD	1.21E-07	1.14E-06	0.92	2.03E-08	2.26E-07	0.93	0.84	0.05
		MD	−1.02E-06	9.14E-07	0.27	−3.15E-07	1.84E-07	0.09	0.78	0.05
		RD	−1.56E-06	1.04E-06	0.13	−4.39E-07	2.16E-07	0.04	0.58	0.04
UF	L	FA	0	0.001	0.79	2.68E-04	3.19E-04	0.40	0.37	0.05
		AD	1.16E-06	2.14E-06	0.59	−1.43E-07	4.67E-07	0.76	0.53	0.00
		MD	2.66E-07	1.56E-06	0.86	−2.84E-07	3.38E-07	0.40	0.51	0.01
		RD	−5.01E-08	1.68E-06	0.98	−3.58E-07	3.59E-07	0.32	0.51	0.10
	R	FA	0	0.001	0.67	1.69E-04	2.48E-04	0.50	0.61	0.02
		AD	1.05E-06	1.74E-06	0.55	2.69E-07	3.77E-07	0.48	0.26	0.01
		MD	−1.53E-07	9.36E-07	0.87	−6.65E-09	1.91E-07	0.97	0.73	0.01
		RD	−4.98E-07	1.11E-06	0.65	−1.12E-07	2.27E-07	0.62	0.72	0.14

TOI: tract of interest; WM: white matter; ILF: inferior longitudinal fasciculus; SLF: superior longitudinal fasciculus; UF: uncinate fasciculus; L: left; R: right; FA: fractional anisotropy; AD: axial diffusivity; MD: mean diffusivity; RD: radial diffusivity; FNE: fear of negative evaluation; SE: standard error; h2: heritability estimate.

Heritability

Per TOI, heritability of every WM parameter was estimated over its average value. All results displayed at least moderate heritability (0.2 – 0.4), extending to very high heritability (0.9 – 1.0) for some WM parameters (Kendler & Prescott, 2006). Results are summarized in Table 3.

Discussion

In the present study we investigated whether characteristics of WM microstructure could be candidate endophenotypes of SAD. To our knowledge, this is the first comprehensive family study on WM microstructure in SAD, which enabled us to specifically examine co-segregation of WM characteristics in families of probands, as affected and non-affected family members have participated in this study (Bas-Hoogendam et al., 2018a). As recently stated by Glahn et al. (2019), a multiplex, multigenerational family design like the LFLSAD is particularly powerful to investigate candidate endophenotypes as “[..] Reduced environmental variation among family members can reduce noise, improving statistical power to observe genotype-phenotype associations. [..] Designs that require multiple affected individuals in a family may result in a more severe phenotypic profile and a different underlying genetic architecture as compared to simplex families. [..] Family selection also impacts the distribution of phenotypes among unaffected family members, with members of multiplex families generally having greater endophenotype impairment than simplex family members.”

In the present work, we focused on two endophenotype criteria, namely co-segregation of WM characteristics with social anxiety within participating families and estimation of heritability of these WM characteristics. Voxelwise and average tracts of interest (TOI) analyses were used to examine associations between measures of self-reported social anxiety and WM characteristics in the UF, SLF and ILF. For all three TOIs, heritability of WM characteristics was estimated.

Our analyses revealed that increased FA in the left and right SLF co-segregated with social anxiety within families enriched for SAD. These findings were coupled with decreased MD and RD and were consistent across both TOI analyses. The voxelwise results suggest that significant clusters are located in the SLF II. Furthermore, and in line with previous literature, all WM characteristics were estimated to be at least moderately heritable (Kochunov et al., 2015, Kochunov et al., 2016).

The SLF II is the major part of the SLF and is mostly concerned with visuospatial attention and processing. Structurally it connects the caudal part of the inferior parietal lobule (IPL) and intraparietal sulcus with the posterior part of the prefrontal cortices. Functionally, the SLF II is thought to connect the parietal part of the ventral attention network with the prefrontal component of the dorsal attention network and is involved in the default mode network (DMN) (Alves et al., 2019, Barbeau et al., 2020, Makris et al., 2005, Parr and Friston, 2017, Schmahmann and Pandya, 2006, Schmahmann et al., 2007, Thiebaut de Schotten et al., 2011, Vossel et al., 2014).

Interestingly, previous studies that reported decreased FA and increased RD in the SLF described clusters in a different subpart of the SLF, namely SLF III, in patients with SAD compared to healthy controls (see Supplementary Table 1a and 1b; Baur et al., 2011, Qiu et al., 2014, Tükel et al., 2017. These findings were replicated in a meta-analysis using these three studies (Jenkins et al., 2016). It should be noted that the SLF III differs from the SLF II as it extends from the supramarginal gyrus to the ventral premotor regions and is thought, among others, to be involved in language processing (Makris et al., 2005, Schmahmann and Pandya, 2006); Schmahmann et al. (2007). In addition, the studies mentioned above had a different study design: they examined WM in SAD patients versus healthy controls in a case-control design, whilst we investigated WM microstructure in families genetically enriched for SAD using a unique family study design. To the best of our knowledge, only one other study examining WM in SAD, using a case-control design and tractography analyses, reported increased FA (Liao et al., 2011). In this study, the FA-increase was located in fibers passing through the genu of the corpus callosum.

Current literature suggests that increased FA coupled with decreased MD and RD and unchanged AD could imply dense axonal packaging or increased myelination (Alexander et al., 2007, Alexander et al., 2011, Feldman et al., 2010, Jones et al., 2013). It should be noted that increased FA does not necessarily mean a better connection of the WM tract involved; instead, this could be suggestive of a compensatory mechanism, due to reduced crossing WM fibers or a more coherent alignment of fibers in the SLF II (Haber et al., 2020, Jones et al., 2013, Thomason and Thompson, 2011).

Our findings might seem contradictory to the neurofunctional model of the socially anxious brain described by Brühl et al. (2014), which proposed that decreased structural connectivity of, among others, the SLF could contribute to decoupling of hyperactive parietal and medial occipital brain regions from other networks involved in emotion regulation such as amygdala, limbic, salience and ventral attention networks. However, this subset of the model was based on the three DTI studies described above (Baur et al., 2011, Qiu et al., 2014, Tükel et al., 2017) and might yet be partly conceptual.

As recently reviewed by Bas-Hoogendam and Westenberg (2020), neuroimaging studies in SAD have reported multimodal changes in the brain. For example, changes in functional connectivity of the DMN have been reported. The DMN is thought to be involved in, among others, social referencing (Kim and Yoon, 2018, MacNamara et al., 2016). In addition, cortical thickness and surface area of the IPL are positively associated with SAD (Bas-Hoogendam et al., 2018b). Also, heightened activity in the medial temporal gyrus, superior temporal gyrus, and superior temporal sulcus during unintentional social norm processing is associated with SA (Bas-Hoogendam et al., 2019b) and a recent fMRI study by Kim et al. (2018) reported heightened processing and prolonged attention during social threats in the IPL and the supramarginal gyrus. As the SLF II connects the IPL to the middle and superior frontal gyri, these results are relevant for the interpretation of our present findings.

We propose a few hypotheses for our findings, but realize that the current literature is not yet clear about the differentiation of subsets of the SLF and AF, and their involvement in SAD. First, the SLF II might be involved in the DMN as it connects different hubs of this network, and changes in the DMN have been reported in patients with SAD (MacNamara et al., 2016). Second, the SLF II might be involved in abnormal visual processing as this tract is associated with posterior temporal parts and the attention network, and visual biases have been reported in SAD. Finally, FA is known to be decreased when strong fibers in multiple directions are present (e.g. in a crossing fiber area). If myelination of one fiber bundle in one of these directions is decreased, FA could be increased. Therefore, as our significant findings are in an area with a high amount of crossing fibers, we cautiously propose that increased social anxiety could be associated with lower myelination in one fiber bundle in a crossing fiber area.

Limitations and recommendations for future studies

Although this study is the first comprehensive two-generation family neuroimaging study on SAD, thus enabling investigation of WM microstructure as potential endophenotype of SAD, the findings of this study should be interpreted in light of its limitations. First, some participants have mental comorbid disorders, which thus might have influenced our results. However, as comorbidity is high in the clinical population of SAD patients (Blanco et al., 2011, Fehm et al., 2005, Grant et al., 2005, Koyuncu et al., 2019), we deem our cohort as a representative sample. In addition, results of sensitivity analyses, in which we excluded participants with comorbid mental disorders, were in line with the main results. In addition, we would like to mention that although we used a continuous scale to analyze social anxiety rather than a dichotomous SAD versus non-SAD approach, we recognize our findings might not be SAD specific as changes in the SLF have also been reported in other mental emotional disorders, for example in a transdiagnostic meta-analysis of emotional disorders (Jenkins et al., 2016). We therefore recommend future studies to consider transdiagnostic approaches, for example using the Research Domain Criteria framework (Owen, 2014).

Second, microstructure of WM tracts has been examined both in voxelwise and average TOI analyses to allow for interpretation in the light of the current literature whilst also being able to detect subtle changes in WM tracts. As we did not use tractography to analyze the DTI data, a method which is able to trace anatomical connections of WM between several brain regions and thus examine crossing fibers and different subsets of the SLF, we cannot investigate the idea that increased FA is a result of decreased myelination in one fiber bundle in a crossing fiber area. In addition, no post-hoc correction for the number of TOIs (n = 3, bilateral) has been applied thus results should be interpreted carefully. Also, analyses were done in the WM skeleton (and in TOIs), limiting it to the core of major WM tracts. As a result, more peripheral WM parameters in smaller tracts were not investigated. Therefore, future analyses should consider including tractography to investigate crossing fibers and distinguish between the different subsets of the SLF and their involvement in SAD. Also, we have only corrected linearly for age whilst the development of WM is known to peak at 28 years, followed by a slow decline (Kochunov et al., 2012). In addition, the genetic data for GWAS analyses are not yet available, thus our findings cannot be linked to genetic variation yet. Finally, due to the lack of healthy control families without SAD, we could not examine the second part of the fourth endophenotype criterium (nonaffected family members show other levels of the endophenotype compared to the general population). As endophenotypes are thought to be present before the development of clinical symptoms, they should thus be measurable in unaffected family members and differ from healthy control families. This could not be investigated in the present study. In addition, as a cross-sectional design has been used, we could not examine the second endophenotype criterium (state-independency). Therefore, to further examine its potential role as an endophenotype, we advise future studies to consider including control families to compare changes in the SLF in nonaffected family members to SLF microstructure in healthy control families, and to consider a longitudinal design to study trait-stability.

Conclusion

The findings of the present work confirmed our hypothesis that altered white matter microstructure could be a candidate endophenotype of SAD. However, contradictory to our hypothesis of decreased FA, we found that increased FA in the SLF II co-segregated with SA within families genetically enriched for SAD. This was coupled with decreased RD and MD. Furthermore, all white matter characteristics were estimated to be at least moderately heritable, thus supporting the heritability criterion for endophenotypes. These findings might further elucidate the genetic susceptibility to SAD and improve our understanding of the overall etiology.

Funding sources

The Leiden Family Lab study on Social Anxiety Disorder and Janna Marie Bas-Hoogendam were funded by Leiden University Research Profile “Health, Prevention and the Human Life Cycle”. These funding sources had no involvement in writing this article nor in the decision to submit this work for publication.

CRediT authorship contribution statement

Eline F. Roelofs: Methodology, Software, Formal analysis, Data curation, Writing - original draft, Visualization. Janna Marie Bas-Hoogendam: Conceptualization, Data curation, Investigation, Methodology, Formal analysis, Software, Project administration, Supervision, Writing - review & editing. Hanneke van Ewijk: Supervision, Writing - review & editing. Habib Ganjgahi: Software, Writing - review & editing. Steven J.A. van der Werff: Methodology, Writing - review & editing. Marjolein E.A. Barendse: Methodology, Investigation, Writing - review & editing. P. Michiel Westenberg: Conceptualization, Funding acquisition, Investigation, Methodology, Resources, Writing - review & editing. Robert R.J.M Vermeiren: Supervision, Writing - review & editing. Nic J.A. van der Wee: Conceptualization, Funding acquisition, Investigation, Methodology, Resources, Supervision, Writing - review & editing.

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.