Extended‐amygdala intrinsic functional connectivity networks: A population study

Abstract Pre‐clinical and human neuroimaging research implicates the extended‐amygdala (ExtA) (including the bed nucleus of the stria terminalis [BST] and central nucleus of the amygdala [CeA]) in networks mediating negative emotional states associated with stress and substance‐use behaviours. The extent to which individual ExtA structures form a functionally integrated unit is controversial. We utilised a large sample (n > 1,000 healthy young adult humans) to compare the intrinsic functional connectivity networks (ICNs) of the BST and CeA using task‐free functional magnetic resonance imaging (fMRI) data from the Human Connectome Project. We assessed whether inter‐individual differences within these ICNs were related to two principal components representing negative disposition and alcohol use. Building on recent primate evidence, we tested whether within BST‐CeA intrinsic functional connectivity (iFC) was heritable and further examined co‐heritability with our principal components. We demonstrate the BST and CeA to have discrete, but largely overlapping ICNs similar to previous findings. We found no evidence that within BST—CeA iFC was heritable; however, post hoc analyses found significant BST iFC heritability with the broader superficial and centromedial amygdala regions. There were no significant correlations or co‐heritability associations with our principal components either across the ICNs or for specific BST‐Amygdala iFC. Possible differences in phenotype associations across task‐free, task‐based, and clinical fMRI are discussed, along with suggestions for more causal investigative paradigms that make use of the now well‐established ExtA ICNs.

Part of the interest in the ExtA stems from its anatomic location.
With structural connections to areas including sensory, mnemonic, affective, and regulatory processing regions, the ExtA is strategically placed to coordinate activities in multiple "limbic lobe" areas for the development of behavioural responses through its output channels (Avery et al., 2016;Fox et al., 2015;Heimer & Van Hoesen, 2006). As such, and in particular because of its direct outputs to the hypothalamic pituitary adrenal axis, it has been implicated in multiple behaviours linked to the processing of threat, stressors, and negative emotional states Giardino et al., 2018;Lebow & Chen, 2016).
That the ExtA is a key component within a stress-related network further implicates it as an area of interest for substance-use behaviours (Avery et al., 2016;Erikson, Wei, & Walker, 2018;Stamatakis et al., 2014;Volkow et al., 2016). Specifically, the ExtA is thought to be important in the dysphoric state associated with drug withdrawal and stress-induced relapse and has been associated with cellular changes following alcohol use (Avery et al., 2016;Ch'ng, Fu, Brown, McDougall, & Lawrence, 2018;Erikson et al., 2018;Roberto et al., 2020;Stamatakis et al., 2014;Volkow et al., 2016). Association of the ExtA with both alcohol and anxiety is especially interesting given the high comorbidity between the two, with anxiety often precipitating alcohol use and being a hallmark of withdrawal (Gilpin et al., 2015). Experimental evidence for involvement in fear, anxiety, stress, and substance-use derives from a multitude of lesion, optogenetic, and neural tracing studies in animals and, more recently, human neuroimaging studies (for reviews, see Ahrens et al., 2018;Avery et al., 2016;Ch'ng et al., 2018;Goode, Acca, & Maren, 2020;Lebow & Chen, 2016).
Despite some agreement regarding the ExtA ICNs (overlapping connections to medial prefrontal, hippocampal, wider amygdala, and thalamic regions), because of data acquisition, processing differences (such as brain coverage and choice of mask), and repeated use of the same samples, the convergence between studies can be hard to assess (Table 1). Thus, our first aim was to establish the ICNs of the BST and CeA in a large (n= > 1,000) independent dataset-the Young Adults Human Connectome Project (HCP). A major strength of this approach is our use of the HCP data. The HCP contains high-quality imaging data, with most participants having undergone an hour of tf-fMRI (Glasser et al., , 2016. Scan lengths longer than 10 min are important as studies have highlighted the negative effects of short scan times on the stability of brain function estimates (Birn et al., 2013;Elliott et al., 2020). There is presently some debate as to whether the ExtA acts mostly as a unified structure, or whether its components represent separate systems underlying different processes, in particular with regard to fear versus anxiety processing or in the tracking of threat imminence Goode et al., 2019Goode et al., , 2020Hur et al., 2020;Tillman et al., 2018;Walker, Miles, & Davis, 2009). Therefore, we utilised this sample to examine the degree of overlap between the ICNs of the CeA and BST; giving an indirect indication as to the similarity of their functions (Gorka et al., 2018;Oler et al., 2012;Tillman et al., 2018;Torrisi et al., 2015Torrisi et al., , 2019Weis et al., 2019).
While phenotypes such as anxiety, fear, depression, and substance use are often studied as if they were separate constructs, they are frequently highly comorbid and demonstrate an overlap of symptoms (Hur, Stockbridge, Fox, & Shackman, 2019;Plana-Ripoll et al., 2019). Recent work has suggested that these phenotypes can be represented by broader overarching constructs, conceptualised as "dispositional negativity" or simply "negative affect" (Hur et al., 2019;Krueger et al., 2018;Shackman et al., 2018;Waszczuk et al., 2020).
Genetic correlation studies have lent credence to this hypothesis, demonstrating that many phenotypically similar traits such as anxiety and depression also share a large proportion of underlying genetic risk factors (Allegrini et al., 2020;Hur et al., 2019;Waszczuk et al., 2020).
Human and non-human primate neuroimaging work suggests that dispositional negativity traits are associated with networks that include the ExtA, with a particular focus on the central amygdala (Hur et al., 2019). Consequently, to expand on this previous work, we placed self-report questionnaire measures examining phenotypes of interest (anxiety, depression, fear, and alcohol use) into a principal Here, we addressed this issue by using a large population-level sample containing multiple measures of relevant phenotypes.
Because psychological traits are underpinned by the brain, understanding whether psychological traits and brain function share underlying genetic influences can be useful for identifying where research may be able to detect biological mechanisms contributing to both.
Despite its apparent importance in a range of psychopathology-linked behaviours, to our knowledge only one study to date has examined genetic co-variance of psychopathology-associated traits with ExtA iFC. This study used a pedigree of rhesus monkeys to demonstrate that iFC between the CeA and an area consistent with the BST was co-heritable with anxious temperament (pgr = 0.87) .
While heritability estimates do not alone provide information about the nature of shared genetic mechanisms (Turkheimer, 2016), this result suggests that ExtA iFC and anxiety-related traits may be influenced by common genetic factors.
Therefore, we used the kinship structure of the HCP data to estimate within BST-CeA iFC heritability and co-heritability with our principal components. Thus, we aimed to extend the non-human primate finding of Fox et al. to humans by demonstrating that within BST-CeA iFC is both heritable and co-heritable with anxiety-related traits . Previous evidence has also reported significant BST iFC to other amygdala sub-nuclei in humans (Hofmann & Straube, 2019). Hence, we further ran a post hoc analysis to assess the heritability and co-heritability (with the principle components) of BST iFC to the centromedial, basolateral, and superficial amygdala regions.

| Principal component analysis
In this study, phenotypes of interest were those related to anxiety, depression, fear, and substance use. There are multiple instruments in this dataset measuring each of these constructs and these phenotypes are frequently highly correlated. Therefore, we performed PCA and reduced data dimensionality by extracting the minimum number of latent components that summarise the maximum amount of information contained in the original measures. The questionnaire measures outlined in the next section were joined into a single dataset and were tested for sampling adequacy using a Kaiser-Meyer-Olkin (KMO) test (Dziuban and Shirkey, 1974)

| Questionnaire selection
The questionnaires used were administered to each participant by the HCP team and all measures were selected from the NIH toolbox, a well-validated set of metrics for quick assessment of cognitive, emotional, sensory and motor functions (Weintraub et al., 2013). Items were selected if they measured anxiety, stress, fear, or substance use.
Where individual items were not provided, we used the relevant questionnaire subscales (Table 2). For the substance use metrics, we only included measures of alcohol use, as self-reported smoking and "harder" drug use rates were low (<20% for tobacco use, <8% ever used cocaine). In total, nine measures were selected (

| HCP pre-processing
We used the minimally processed tf-fMRI 3 T dataset, described elsewhere . Scripts to run the pipeline are freely available online at https://github.com/Washington-University/HCPpipelines. Briefly, the pipeline applies gradient distortion correction to account for spatial distortions, followed by volume realignment to compensate for subject motion, coregistration of the fMRI data to the structural image, non-linear registration to MNI space, intensity normalisation to a mean of 10,000, bias field removal, and masking of data with a brain mask. Structured noise was  (2019)). This pipeline was optimised for the HCP dataset and had the aim of maximising the reduction of structured noise components, such as those caused by subject motion, while retaining spatially specific bold signal components (i.e., ICNs) (Glasser et al., 2016). This was reportedly achieved with better than 99% accuracy (Glasser et al., 2016;Griffanti et al., 2014). To reduce the effects of signal drop-out (Schwaferts, 2017) Both seeds were thresholded at 25% before use (Tillman et al., 2018) ( Figures 1 and 2).
The 3 T 2 mm BST mask was generated by a manual segmentation process undertaken on 10 healthy individuals using a scanning sequence that provided high grey matter/white matter/CSF contrast (Theiss et al., 2017) (Figures 1 and 2). The protocol was found to have high reliability among raters (Dice similarity coefficient ≥ 0.85).

| Whole-brain seed-based correlation analysis
Seed-based correlation iFC analysis provides a measure of temporal coherence between a seed-region's blood-oxygenation-leveldependent (BOLD) activation over time and that of the target regions.
Temporal coherence in tf-fMRI data is used to infer iFC (Battistella et al., 2020;Suárez et al., 2020;Thomas Yeo et al., 2011). To run the analysis we used the ciftify_seed_corr tool downloaded from https:// edickie.github.io/ciftify/#/ (Dickie et al., 2019), which was in turn adapted from the HCP minimal processing pipeline . This works by first extracting a mean time-series of the seed-region. This time-series is then correlated with the mean timeseries of the target regions, producing a Fisher's r correlation map.
These correlation coefficients are then converted to normally distributed z-scores using a Fisher r-z transform (Fisher, 1915). This produces a z-map for each participant that represents the strength of the correlation of activity for each target region and the seed-region. We used a whole-brain voxel-wise approach, meaning that our target regions were every 2 mm voxel in the brain.
2.5 | fMRI statistical analysis 2.5.1 | Permutation-based one-sample t tests Following the creation of a single z-map for each participant, all of these images were visually inspected. Twenty-three participants had images removed from further analysis due to having either sections of the signal missing or for having z-score distributions containing too many values within the outer or inner tail distributions (assessed via fslstats -r-R and histogram plots). The remaining 1,071 participants had their images merged across all participants to create a 4D image using the fslmerge tool (Jenkinson et al., 2012). Permutation-based one-sample t tests were then run to see which voxels had activity that was significantly correlated with the seed-regions across all participants. This was done using FSL's PALM command line tool (Winkler et al., 2016).
For the quantification of the whole brain ICNs, we wanted the results to be generalisable to the wider population, thus we were not interested in the influence of family effects across the whole network.
Therefore, because our sample was made up of siblings, it was important to account for relatedness such that model estimations were not inflated. PALM permits a kinship matrix that details the family structures within the population. PALM shuffles the data within and between blocks according to this family structure, avoiding relatedness The bed nucleus of the stria terminalis (BST) (blue) and central nucleus of the amygdala (CeA) (red) seeds The bed nucleus of the stria terminalis (BST) seed (blue), coronal section confounding the results. The kinship file was generated with the HCP2Blocks MATLAB script provided online at https://brainder.org/ 2016/08/01/three-hcp-utilities (Winkler, Webster, Vidaurre, Nichols, & Smith, 2015).
PALM has several optional commands. We used threshold-free cluster-enhancement (TFCE) and Gamma approximation. Briefly, TFCE enhances cluster-like structures in the data without having to define somewhat arbitrary cluster thresholds beforehand (Smith & Nichols, 2009). Gamma approximation is an option used to speed up the analysis by running a smaller number of permutations, computing empirically the moments of the permutation distribution and then fitting a gamma distribution (Winkler et al., 2016). The number of permutations used was 1,000.

| Post hoc thresholding of PALM output images
Given the large sample size, the vast majority of voxels in the brain were statistically significantly correlated to our seed-regions after family wise error rate correction. To reveal meaningful connections and to reduce noise, we further thresholded the images post hoc using the t-statistic. This was done by visually inspecting the output images and choosing a t-score that met the criteria of delineating meaningful anatomical structures in the brain, while keeping the maximum amount of signal (Tillman et al., 2018). The t-threshold we used for both seed-images was 9. Using the -saveglm option from PALM, we saw that this equated to a minimum Cohen's d value of 0.275 (Winkler et al., 2016). While we are confident this was an appropriate threshold, given the somewhat arbitrary nature of this method, thresholded and un-thresholded output images have been uploaded to NeuroVault for inspection at https://identifiers.org/neurovault. collection:8076.

| Analysing shared and unique BST and CeA networks
To assess the shared ICNs between the BST and CeA, we used a minimum conjunction (Boolean "AND") to combine the t-thresholded PALM output images of each seed (Nichols, Brett, Andersson, Wager, & Poline, 2005;Tillman et al., 2018). This created a new image displaying the areas of ICNs that overlapped between the two ExtA regions.
To assess the unique BST and CeA networks, we performed a single group paired difference t test using the method outlined on the FSL GLM website (https://fsl.fmrib.ox.ac.uk/fsl/fslwiki/GLM#Single-Group_Paired_Difference_.28Paired_T-Test.29). Briefly, to get the unique BST ICN, we subtracted each participants BST z-score image from their CeA z-score image and then ran a one-sample permutation t test on this difference map. This was repeated for the CeA network (CeA-BST z maps, followed by a one-sample t test). A mask was used to restrict analysis to the regions that were found to be connected to one or both seeds in the original one-sample t tests, thus avoiding the need to interpret differences in regions not significantly connected to the seeds (Tillman et al., 2018).

| Region identification
Connected regions were identified using a mixture of the Oxford cortical/sub-cortical atlas and the Juelich Histological Atlas, both provided with FSL (Jenkinson et al., 2012). For iFC to basal ganglia structures and the hypothalamus, we used a collection of masks provided online at Neurovault (https://identifiers.org/neurovault. collection:3145) (Pauli, Nili, & Tyszka, 2018).

| Intrinsic connectivity networks and principal component association tests
Following the one-sample t tests for each seed region, we then created a mask of the t-thresholded significantly connected regions. This mask was then applied to the 4D image of participants connectivity zmaps to select only the thresholded connected voxels for association testing with our PC's and for gender effects. We used the PALM command-line tool, with TFCE, Gamma-approximation, and event blocks to control for family relatedness (see Section 2.5.1). As well as the standard correction for multiple comparisons within each image, PALM further allows for correction across different contrasts with the -corr-con option (Winkler et al., 2016). This option was used along with the -demean function, which mean-centres the variables, and the -cmcx function, which allows for synchronised permutations accounting for repeated elements in the design matrix. Three tests were run in total on each seed-image, one each for the two principal components and one for gender (male, female 2.7 | Within BST-amygdala heritability, coheritability, and phenotype association analysis   4 and 5, right). There was further shared iFC with pre-and post-central gyri, extending bilaterally to primary motor and sensory regions, and shared connectivity with the angular gyrus/superior lateral occipital cortex. There were no significant voxels directly within either the BST or CeA masks, suggesting that the two regions were not co-activated at rest. There was however a bilaterally BSTconnected amygdala cluster directly adjacent (within a single voxel) to the CeA mask ( Figure 6). The bilateral SLEA region connecting the BST and CeA also demonstrated overlapping connectivity (Figure 7).

| BST > CeA connectivity
The BST had more extensive iFC with the occipital lobe, in particular within the superior occipital cortex, the intracalcarine cortex, and at the occipital pole (Figures 4 and 8, left). There was also greater BST iFC with the posterior and anterior cingulate gyrus, posterior thalamus, precuneus cortex, left and right caudate, globus pallidus, lateral superior frontal gyrus, paracingulate gyrus, and ventral tegmental area (Figures 4 and 8, left).

| CeA > BST connectivity
The

| PCA results
The selected questionnaire items ( Note: Significantly connected clusters to the BST following the one-sample permutation test. Images were thresholded at t= > 9 before clusters were identified. Brain regions were listed if they had >50% chance of being within a cluster. Max t is the maximum t-stat located within a cluster. X, Y, and Z columns represent the location of the centre of gravity for the cluster. positively on measures capturing negative disposition, such as anxiety, depression and perceived stress, and was therefore named the "negative disposition" component. The second component had significant loadings from alcohol measures and was therefore labelled the "alcohol use" component. See Table 5 and Figure 9 for a breakdown of the PCA results.

| Intrinsic functional connectivity networks and principal components
The PALM corr-con analysis provided no evidence that the negative disposition or alcohol use components were significantly associated with increased or decreased iFC across the ExtA ICNs in our sample.
Gender was also not associated with the BST or CeA ICNs after correction for multiple comparisons.

| Univariate heritability analysis
Twin-based heritability analysis of within BST-CeA iFC found no evidence for heritability ( Table 6). Analysis of within BST-centromedial iFC found that this connection was significantly heritable at H2r = 0.15 (Table 6). BST-superficial iFC had a heritability estimate of H2r = 0.14, but was marginally outside the bounds of statistical Note: Significantly connected clusters with the CeA following the one-sample permutation test. Images were thresholded at t= > 9 before clusters were identified. Brain regions were listed if they had = > 50% chance of being within a cluster. Max t is the maximum t-stat located within a cluster. X, Y, and Z columns represent the location of the centre of gravity for the cluster. Hemi indicates the hemisphere in which the cluster resides where B = bilateral, R = right, and L = Left. For ease of interpretation clusters shown are those with a minimum of 10 connected voxels. Abbreviation: CeA, central nucleus of the amygdala.
significance after FDR correction (Table 6). BST-basolateral iFC showed no evidence of significant heritability (

F I G U R E 4
The bed nucleus of the stria terminalis (BST) and central nucleus of the amygdala (CeA) share a common intrinsic functional connectivity pattern, in particular with pre-frontal cortex, amygdala, hippocampus, superior temporal sulcus, insula, and precuneus. They also share connectivity with areas of the motor and sensory cortex

| Bivariate heritability analysis
Co-heritability analysis did not reveal any significant phenotypic, environmental, or genetic correlations with either of the principal components for any of the amygdala sub-regions (see supplementary material for bivariate SOLARIUS outputs).

| Summary of findings
Using a large young adult human sample, we revealed distinct, but overlapping, ExtA ICNs that are largely consistent with findings from smaller previous human neuroimaging studies (Avery et al., 2014;Gorka et al., 2018;Oler et al., 2012Oler et al., , 2017Tillman et al., 2018;Torrisi et al., 2015;Weis et al., 2019). Genetic analysis of within BST-CeA iFC provided no evidence for a heritable connection.
However, post hoc analysis of amygdala sub-regions revealed evidence for small heritability estimates for BST-centromedial and superficial regions. PCA reduced scores on nine questionnaire measures of anxiety, fear, depression, and substance use to two components, which we interpret as "negative disposition" and "alcohol use." Contrary to our hypotheses, we report no evidence for associations of these phenotypes across the ExtA ICNs. We also found no evidence that specific BST iFC to any of the tested amygdala regions were co-heritable or otherwise correlated with either of the components.
F I G U R E 5 Axial section demonstrating shared connectivity of bed nucleus of the stria terminalis (BST) and central nucleus of the amygdala (CeA) with the hippocampus, insular, temporal gyri, frontal orbital and medial prefrontal cortex. The CeA has more extensive connectivity generally with each of these regions and of note displays unique connectivity along amygdalo-hippocampal regions F I G U R E 6 The bed nucleus of the stria terminalis (BST)-correlated amygdala cluster (red) and central nucleus of the amygdala (CeA) seed (blue)

| Intrinsic connectivity networks of the ExtA
Our shared ICN results are in broad agreement with the previous literature, specifically demonstrating overlapping connections within a now widely reported ExtA ICN that includes the mPFC, bilateral hippocampus, insular regions, wider amygdala areas, and the precuneus (Avery et al., 2014;Gorka et al., 2018;Pedersen et al., 2020;Tillman et al., 2018;Torrisi et al., 2015Torrisi et al., , 2019Weis et al., 2019). We report shared iFC to lateral temporal regions, including the superior and middle temporal gyri and the temporal poles, again largely consistent with previous human iFC results. Whist amygdala structural connections to lateral temporal regions are well characterised (Folloni et al., 2019;Janak & Tye, 2015;Klingler & Gloor, 1960), this is not the case for the BST and it has been suggested that BST-temporal pole connectivity may even be unique to humans (Avery et al., 2014). We demonstrate shared iFC to areas of the sensory/motor cortex, auditory regions, and to lateral occipital areas, something also reported by Tillman et al. (2018). This largely cortical sensory-motor connectivity is consistent with the suggestion that the ExtA serves as an integrator of sensory information, which can then prepare the motor and endocrine systems to act according to the emotional salience and threat- Negatively correlated variables are positioned on opposite sides of the plot. Variables that are away from the centre are well represented by that component. Here, it is shown that we can neatly cluster two separate components, representing negative disposition (PC1) or alcohol use (PC2).
Bottom right: The screen plot displays the amount of variance explained by each component. The first two components capture 62% of the total variance of the original questionnaire measures. See Table 2 for a description of questionnaire measures only measure correlated BOLD activity, without taking into account more elaborate models that assess causality, we are not permitted to make inferences regarding the direction of connectivity (Rogers, Morgan, Newton, & Gore, 2007). However, an extensive body of work on the amygdala suggests that many of the CeA connections are mediated through the basolateral amygdala to the CeA, which in turn serves primarily as an output to basal forebrain structures (Janak & Tye, 2015). The picture is complex, however, and many studies have also shown direct structural connections with the CeA region, for example, from agranular and dysgranular regions of the insular in Macaques and from the ventral hippocampus in mice (Stefanacci & Amaral, 2002;Xu et al., 2016).
Given pre-clinical and human imaging results demonstrating structural and functional connectivity between the CeA and BST (Avery et al., 2014;Davis, Walker, Miles, & Grillon, 2010;Fox et al., 2018;Gorka et al., 2018;Hofmann & Straube, 2019;Martin et al., 1991;Oler et al., 2017;Torrisi et al., 2015), we expected to find evidence of strong iFC between our BST and CeA masks, however this was not quite the case. After thresholding, we did not find evidence of CeA iFC with the BST, although we did find a bilateral BSTfunctionally connected region directly adjacent to the original CeA mask ( Figure 6). Given the small size of the structures, many studies refer to "areas consistent with" the BST and CeA . These discrepancies can likely be explained by the difficultly of accurately delineating the amygdala sub-regions using MRI and/or the noisy nature of tf-fMRI data (Kedo et al., 2018;Sylvester et al., 2020).
Our results revealed minimal connectivity to the thalamus. Given thalamic connectivity is widely reported in structural and functional studies in both pre-clinical and human studies (Fox et al., 2015;Lebow & Chen, 2016), it seems likely that this may be due to a difference in data acquisition or pre-processing. Although speculative, the discrepancy could perhaps be explained by signal drop-out, something that has been shown to affect FC estimates of the thalamus in the HCP data (Schwaferts, 2017).
In general, though, our findings are highly consistent with the smaller previous studies, and in particular are similar to those of  (Pedersen et al., 2020;Torrisi et al., 2019).

| Heritability and co-heritability of within BST-amygdala iFC
Contrary to recent primate evidence , we do not report evidence of a heritable functional connection between the BST and CeA. A post hoc analysis did reveal evidence for a small magnitude of heritability between the BST and the centromedial and superficial amygdala regions; however, there was no evidence of iFC coheritability with either of the principal components (negative disposition, alcohol use).
Although brain morphology and development are reliably heritable (Jansen, Mous, White, Posthuma, & Polderman, 2015), this is not necessarily the case for iFC where heritability estimates can frequently be zero (Elliott et al., 2018;Jansen et al., 2015). In an analysis of the  (Elliott et al., 2018). The reasons for low iFC heritability estimates are not well understood but could reflect either comparatively noisy signal or simply the greater context-dependent variability inherent within fluctuating connections (Cabeza, Stanley, & Moscovitch, 2018). This makes the Fox primate finding of high heritability (.45) all the more interesting, although the usefulness of comparing the strength of heritability estimates across samples is limited as they are highly influenced by their particular environment; something compounded by comparing across species (Turkheimer, 2016). The fact that we found a heritable connection with the centromedial and superficial amygdala, and not specifically the CeA as was reported in Fox et al., may again reflect difficulties in locating small anatomical regions within the amygdala. With this in mind, our finding of H2r results of .14, while smaller than the non-human primate evidence, is not zero and is broadly in line with other estimates of the heritability of iFC findings in humans (Elliott et al., 2018). Further examination in other human samples could perhaps assess whether individualised task-based, naturalistic fMRI, behaviourally defined (rather than self-reported) negative disposition phenotypes, and/or the use of clinical groups influences the heritability estimates of ExtA iFC (Finn et al., 2017). Larger twinsamples with 7 T MRI data and rich phenotyping would also help to resolve issues around the delineation of amygdala sub-region boundaries while allowing for co-heritability analysis, which is after all of primary interest given the suggestion of shared genetic mechanisms.

| Principal components and ExtA iFC
Our first principal component grouped together questionnaire items that represented aspects of negative disposition (stress, fear, anxiety, depression), supporting previous work (Hur et al., 2019;Krueger et al., 2018;Shackman et al., 2018;Waszczuk et al., 2020 Pleil et al., 2016;Roberto et al., 2020;Volkow et al., 2016), there is very little investigation of the ExtA and alcohol use in humans; with most work tending to focus on the amygdala proper (Hur et al., 2018;Lebow & Chen, 2016). One study that did specifically examine ExtA iFC found that under the influence of alcohol, BST and CeA reactivity to emotional faces was dampened (Hur et al., 2018). Although we did not find evidence of a self-report alcohol-use association in our sample, given the importance of understanding alcohol use behaviours and the strength of evidence from the animal literature, ExtA neuroimaging work on the effects of alcohol in humans should remain a priority. Getting participants to drink alcohol (Hur et al., 2018), utilising heavy drinkers, or making use of task-based fMRI (Finn et al., 2017) could be a more fruitful approach for identifying ExtAalcohol associations.
Our estimates of negative disposition and alcohol use heritability were broadly in line, if not slightly smaller, than similar human studies (Han & Adolphs, 2020;Kranzler et al., 2019;Swan, Carmelli, Rosenman, Fabsitz, & Christian, 1990;Zheng, Plomin, & von Stumm, 2016). As mentioned above, however (Section 4.3), direct comparison of the strength of heritability estimates across samples is of limited value, and as such should not be over-interpreted (Turkheimer, 2016). The covariates sex and age 2 were statistically significantly associated with alcohol use and negative disposition, respectively. Age 2 explained only a tiny amount of variance, and so interpretation is limited in this case. The finding that being male is associated with a small increase in alcohol use scores, however, is in line with recent findings of US samples .

| Limitations
Our study has some limitations. First, our analyses were conducted using 3 T MRI data. Although imaging at this field strength has been found to accurately capture small regions such as the BST (Theiss et al., 2017), higher resolution, and individualised anatomical parcellations, would enable better characterisation of ExtA iFC networks. Additionally, it is the case that even the small BST structure is made up of further sub-nuclei that may have distinct functions, a point that is difficult to address using human MRI Kim et al., 2013). Second, as is the case with all seed-based correlation analyses, the interpretation of the results is correlational only and mechanistic inferences including the directionality of the connections cannot be inferred (Mohanty et al., 2020;Pearlson, 2017). Third, although we aimed to be consistent with similar tf-fMRI HCP studies (Hofmann & Straube, 2019), our choice to favour some pre-processing techniques over others, such as global signal regression, could have impacted our findings (Glasser et al., 2016;Murphy & Fox, 2017). This is unfortunately a limitation upon all fMRI studies until a consensus approach on pre-processing steps can be reached (Murphy & Fox, 2017). Finally, our questionnaire measures were all self-report, which can sometimes affect the accuracy of the phenotyping (Rosenman, Tennekoon, & Hill, 2011). This may be a particular problem for self-reported drinking behaviour as previous studies have shown heavy-drinking to be underreported (Northcote & Livingston, 2011).

| Conclusions and future directions
We used a large sample of high quality tf-fMRI data to assess the ICNs of the two key ExtA nodes. Our ICN findings largely replicated previous tf-fMRI mapping work, implicating the nodes in mostly overlapping ICNs that includes iFC with medial pre-frontal, hippocampal, wider amygdala, lateral temporal, and precuneus regions. Although for our analysis we intended to establish the ExtA ICNs unencumbered by family relatedness, so as to enable inferences to the wider population, future work could intentionally explore how family relatedness influences the networks. This would allow for heritability and co-heritability analysis across the entire ICNs, instead of a priori selected regions. We report for the first time in humans that within BST-centromedial and superficial amygdala iFC is heritable. We did not replicate the recent non-human primate finding  of BST-CeA iFC co-heritability with an anxietyrelated phenotype. We found no evidence for network associations with negative disposition or alcohol use principal components. Recent work has suggested that self-report trait effects may not be associated with the same neural networks as those identified under task-based conditions and in clinical groups. Future work should explore further these differences by using a combination of self-report, task-based measures, and clinical groups (e.g., Porta-Casteràs et al., 2020). Given that this tf-fMRI network appears to be reliably delineated across healthy samples, researchers should move towards more causal approaches to probe its function. As it has been shown that the ExtA has many functional and structural cortical connections, one approach could be to use brain stimulation techniques to alter the ExtA network via a cortical node to see whether this impacts on related functions. This type of analysis has already been used effectively to probe other subcortical-cortical networks, for example, those involving memory and the hippocampus (e.g., Warren, Hermiller, Nilakantan, & Voss, 2019).

CONFLICT OF INTEREST
The authors declare no conflict of interest.

DATA AVAILABILITY STATEMENT
All primary data used in this study is available on request from the HCP, which can be found at https://www.humanconnectome.org/ study/hcp-young-adult. Output images from this study have been uploaded to NeuroVault at https://identifiers.org/neurovault. collection:8076. The code used to generate the data can be made available upon request to the lead author (berrysc@cardiff.ac.uk).