An afferent subcortical white matter pathway to the amygdala facilitates fear recognition

Our ability to rapidly detect threats is thought to be subserved by a subcortical pathway that quickly conveys visual information to the amygdala. This neural shortcut has been demonstrated in animals but has rarely been shown in the human brain. Importantly, it remains unclear whether such a pathway might influence neural activity and behaviour. We conducted a multimodal neuroimaging study of 622 participants from the Human Connectome Project. We applied probabilistic tractography to diffusion-weighted images, reconstructing a subcortical pathway to the amygdala from the superior colliculus via the pulvinar. We then computationally modelled the flow of neural activity during a face-viewing task and found strong evidence for a functionally-afferent subcortical pathway. Critically, individuals with greater fibre density in this pathway also had stronger dynamic coupling and enhanced fearful face recognition. Our findings provide converging evidence for the recruitment of an afferent subcortical route to the amygdala in the human brain that facilitates fear recognition.


Introduction
Decades ago, rodent research uncovered a subcortical pathway to the amygdala that rapidly transmits auditory signals of threat even when the auditory cortex is destroyed 1 . Since this discovery, researchers have sought an equivalent visual pathway that might explain how it is that people with a lesioned primary visual cortex can still respond to affective visual stimuli that they cannot consciously see 2 . The superior colliculus, pulvinar, and amygdala have been identified as nodes of a human subcortical route to the amygdala that bypasses the cortex 3 . These subcortical areas consistently coactivate in cortically blind patients 4,5 (as well as in healthy adults 3, 6, 7 ) when they view emotional stimuli, such as angry or fearful faces. Magnetoencephalography studies using computational modelling have investigated whether the activation of these subcortical nodes is causally related. These studies have consistently found evidence for a forward connection between the pulvinar and amygdala [8][9][10] . The dynamic causal relationship between the superior colliculus and the pulvinar, however, remains unexplored in the human brain 11 . The pulvinar also has several functional and cytoarchitectural subregions 12 and it is unclear how these connect to the superior colliculus and the amygdala and what roles these subregions may play in mediating transmission along the subcortical route 13,14 . As such, the hypothesis that the subcortical route rapidly transfers information from the retina to the amygdala without interference has been heavily criticised 14,15 . Furthermore, the pulvinar is highly connected with a widespread network of cortical regions that may contribute to transmission along the subcortical route 16,17 . Hence, it remains unknown whether the functional activation of the human superior colliculus, pulvinar and amygdala during affective processing bears any relation to an underlying structural pathway 14 .
Recent animal research has revealed several potential direct subcortical pathways that have a causal relationship with fearful behaviour in response to visual threats [18][19][20] . In the absence of relevant post-mortem human research, however, our anatomical knowledge of the human subcortical route to the amygdala can only be derived from tractography of diffusion-weighted images (DWI). Tamietto et al. (2012) examined DWIs from a blindsight patient whose left primary visual cortex was destroyed. The white matter structure of the subcortical route was estimated for the patient and for six healthy, age-matched controls. Critically, the subcortical route had greater fractional anisotropy in the patient's damaged hemisphere, suggesting a neuroplastic increase in myelination to compensate for the disrupted cortical pathways 21 . In a similar study, Rafal et al. (2015) used tractography to investigate the subcortical route in 20 healthy humans and 8 macaques. The subcortical route was reconstructed in both hemispheres for 19 of the 20 human participants and 7 of the 8 macaques 22 . Notably, this sample of human participants was recently expanded and reexamined. Individuals with greater fractional anisotropy along the subcortical route also had a stronger bias toward threat when making saccades to scenes 13 .
Diffusion tractography may grant insight into the strength of anatomical connectivity between regions but it cannot reveal the direction of information transfer, nor can it be used as direct evidence alone for the anatomical existence of a neural pathway. The anatomical presence and the direction-specific neural flow of emotional visual information along the subcortical route has never been concurrently investigated in humans to definitively show that the subcortical route is a direct, afferent pathway specifically associated with fear 14,15 . Such a finding would have important implications for the very foundation of visual threat perception, given this pathway's potential for rapid information transfer 8,23 and unconscious processing 2 . Here, we aimed to comprehensively investigate this putative amygdala pathway in a large sample of over 600 healthy human adults from the Human Connectome Project (HCP) dataset using a multimodal imaging approach to encompass structure, function and behaviour. First, we used DWI to reconstruct the subcortical route from the superior colliculus to the amygdala, via the pulvinar, and estimated its fibre density on an unprecedentedly large sample. We then examined whether this putative pathway targeted the same pulvinar functional subdivisions. Next, we modelled the direction-specific flow of neural activity for faces, testing whether a functional subcortical route is recruited to transmit information toward the amygdala. Finally, we asked whether the fibre density of the subcortical route predicts both fearful face recognition as well as the strength of dynamic coupling between the superior colliculus, pulvinar, and amygdala.

Reconstructing the subcortical route using tractography
The first step in our investigation was to evaluate the evidence for an anatomical subcortical route to the amygdala in the healthy human brain. We exploited high-quality neuroimaging data from a large sample of 622 participants made available by the HCP 24 . We then reconstructed the white matter structure of the subcortical route using two complementary tractography methods for cross-validation. We began with global tractography, a Bayesian approach to reconstructing whole-brain fibre configurations that best explain DWI data (see Online Methods for details). We discovered that the superior colliculus (SC) was connected to the pulvinar (fibre counts for PUL; left: M = 13.23, SD = 5.56, right: M = 13.00, SD = 5.59, minimum of 2 fibres per participant). The pulvinar and the amygdala were also connected (fibre counts for left: M = 5.33, SD = 2.79, and right: M = 6.75, SD = 2.90), with most participants having at least one connecting fibre (zero fibres for left PUL-AMG for 8 participantsonly 1.28% of total sample). Thus, this relatively conservative method of fibre reconstruction (as it takes into account the entire brain) can reliably detect evidence for a subcortical route across a large sample of participants. We also found a hemispheric lateralisation, such that there were more reconstructed fibres for the right than the left PUL-AMG pathway (F(1,617) = 39.865, p = 5.206 x 10 -10 , ηp 2 = .061; t(617) = -9.785, p = 4.070 x 10 -21 ).
To assess the validity of our findings we used a second tractography method, namely "local" probabilistic streamline tractography, as used by Rafal et al. (2015) to reconstruct the subcortical route to the amygdala 22 . We generated streamlines between our regions of interest (ROIs) and found that the superior colliculus connected to the pulvinar (streamline counts for left: M = 1403.32, SD = 417.16, right: M = 111.59, SD = 358.60, minimum of six streamlines per participant) and the pulvinar connected to the amygdala (left: M = 575.42, SD = 203.03, right: M = 575.42, SD = 248.85, minimum 66 streamlines per participant). To evaluate whether these streamlines counts were reconstructed significantly above chance, we compared the numbers with those produced by a null distribution algorithm, which essentially generated streamlines at random 25 . We found that the number of streamlines was significantly different from chance for each connection, as determined by a series of paired t-tests (see Fig. 1), suggesting that the DWI data produced meaningful streamlines between our ROIs.
We employed a recently developed method, SIFT2, which estimates the apparent fibre density of the streamlines connecting two regions of interest. This method more accurately represents the true underlying white matter structure 26 . The apparent fibre density of the streamlines generated using local tractography followed the same pattern as the global tractography fibre counts, such that there was greater fibre density for the SC-PUL connection (left: M = 6.25, SEM = 0.09, right: M = 5.34, SEM = 0.07) than the PUL-AMG connection (left: M = 3.98, SEM = 0.05, right: M = 4.94, SEM = 0.06, F(1,608) = 257.135, p = 1.603 x 10 -48 , ηp 2 = .297). Fibre density was also greater on the right than the left for the PUL-AMG connection (t(608) = -18.205, p = 1.960 x 10 -59 ) but, additionally, we also observed greater fibre density for the left than right SC-PUL connection (t(608) = 10.749, p = 8.600 x 10 -25, F(1,608) = 401.417, p = 6.002 x 10 -69 , ηp 2 = .398). Taken together, our tractography analyses provide strong convergent evidence for a subcortical white matter pathway to the amygdala in the human brain. The streamline count of the local tractography is shown (green-yellow for SC-PUL, orange-pink for PUL-AMG) compared with the streamline count generated by a null distribution (grey). * = pFDR < .05

A direct route through the right inferior pulvinar
We have replicated evidence for a subcortical white matter route to the amygdala from the superior colliculus via the pulvinar using multiple methods for cross-validation in an unprecedented sample over 30 times larger than previous studies 13,21,22 . However, it remains unknown whether this functions as a continuous pathway 14 . This is critical because the pulvinar is highly connected with the visual, temporal, parietal, and frontal cortices 16 and thus may simply mediate communication between these areas and the amygdala, rather than enable a direct route for rapid fear processing. To investigate this, we asked whether the two halves of the subcortical route (i.e. SC-PUL and PUL-AMG) connected to the same subregion of the pulvinar. This would serve as the fastest and thus most direct path from the superior colliculus to the amygdala. To answer this question, we investigated whether the connectivity strength to and from the same pulvinar subregion (from the superior colliculus and to the amygdala, respectively) were positively correlated, which would suggest the concurrent use of that white matter pathway.
We used a parcellated pulvinar mask generated by a previous meta-analysis of functional co-activation 12 that maps closely onto the pulvinar's cytoarchitecture 27 . We then repeated the initial tractography analysis but instead of using one pulvinar region of interest (i.e., target) we used five pulvinar subregions: inferior, superior, anterior, medial, and lateral (see Fig. 2A). We then correlated the fibre density measures (from both local and global tractography) between each pulvinar subregion-specific SC-PUL and PUL-AMG pathway (see Online Methods for details of principal components analysis which was used to extract the unique variance per subregion-specific pathway), resulting in a total of 100 Pearson's correlations. This revealed several significant relationships between subregion-specific pathways, many of which were between pathways that connected to different pulvinar subregions (see Fig. 2C, off-diagonal matrix entries). Critically, however, the right inferior pulvinar was the only subregion to show a significant positive relationship with both the superior colliculus and amygdala for both global and local tractography methods for the same pulvinar subregion (global: r = .137, pbonf = .030, local: r = .188, pbonf = 2.982 x 10 -6 , see Fig.  2C diagonal matrix entries). No other relationships were consistent and significant across both methods for the same pulvinar subregion. The inferior pulvinar also had the strongest fibre density for connections to the amygdala and was also one of the strongest subregions to connect to the superior colliculus. Thus, while there were many connections that correlated across pulvinar subregions, these relationships were not consistent across tractography methods. Therefore, the mutual connections with the inferior pulvinar observed for both tractography methods provided the strongest evidence for a direct pathway from the superior colliculus to the amygdala. The thickness of the lines indicates connectivity strength, relative to connectivity between other subregions and that region. B) Global track counts and local summed weights for the PUL-AMG (pink box) and SC-PUL (green box) connections. The five subregions are colour-coded as per the 3D renders in A). C) Asterisks indicate significant correlations (pFDR < .05) between the subregion-specific connection strength of the SC-PUL (columns) and PUL-AMG (rows) pathways. Diagonal matrix entries correspond to the correlation of white matter pathways linking superior colliculus and the amygdala via the same pulvinar subregion, suggesting a continuous pathway from the superior colliculus to the amygdala. Off-diagonal matrix entries correspond to pathways linking superior colliculus and the amygdala via different pulvinar subregions. Red indicates the strength of positive correlations and black indicates the strength of negative correlations, where the magnitude of Pearson's r is indicated by the colour bar. The key finding is highlighted by two black boxes, showing that the connections between the inferior pulvinar and the superior colliculus and amygdala were significantly positively correlated for both global and local tractography.

Greater fibre density predicts better fearful face perception
We wanted to translate our work in humans to animal research that has demonstrated clear relationships between the anatomical presence of a subcortical route and fearful behaviour [18][19][20] . To this end, we examined data from the Penn Emotion Recognition task, which was implemented in the HCP battery of behavioural tests. Note that this task was different from the behavioural task used in the fMRI experiment. In this task, participants were serially presented with 40 faces that were either happy, sad, angry, fearful, or neutral. We then investigated the association between these scores (see Fig. A) with the fibre density of the subcortical route. We chose not to include happy or neutral expressions in our analysis because the data were substantially negatively skewed (skewness for: happy = -5.821; neutral = -2.053; angry = -0.719; fearful = -1.188; sad = -1.090). Thus, we entered the global and local fibre density measures for the SC-PUL and PUL-AMG pathways into a multivariate regression with recognition accuracy scores for fearful, angry, and sad faces as covariates (we excluded 25 outliers, 4.02% of sample, with a Cook's distance of more than 3 SDs from the mean). Overall, there was a significant multivariate relationship between tractography and recognition accuracy for fearful faces (F(8, 585) = 2.100, p = .034, Wilk's Λ = .972, np 2 = .028; see Fig. B). This was driven predominantly by fibre density of the left (F(1,592) = 12.213, p = .001, np 2 = .020) and right (F(1,592) = 8.915, p = .003, np 2 = .015) PUL-AMG connections' local fibre density. These results suggest that the fibre density of the subcortical route is associated with fearful face recognition more so than with other negative (sad) or threatening (angry) emotional expressions.

Figure 3. Relationship between behavioural performance and tractography. A)
Histograms of scores (out of eight) for correctly recognising the emotional expression of angry, fearful, and sad faces from the Penn Recognition Test. Normal distribution function is plotted. B) The relationships between recognition of angry (purple), fearful (red), and sad (blue) faces (xaxes) and the left (top row) and right (bottom row) local fibre density of the PUL-AMG connection (y-axes) produced by a multivariate regression. Fearful face recognition accuracy was significantly related to local fibre density for the left and right PUL-AMG connections.

Subcortical and cortical BOLD signal to emotional faces
We used dynamic causal modelling to infer the dynamic (or effective) connectivity between each node of the subcortical route and determine the directionality of the functional interactions occurring along the anatomical pathway mapping described above. First, it was necessary to establish any differences in functional activation within these nodes. To do this, we used the "Emotion" task from the HCP battery of fMRI experiments, in which participants performed a matching task using images of faces or shapes. We contrasted activation in face versus shape blocks and reviewed the results at the whole-brain group level across all 622 participants, p < .05 FWE (see Fig. 4). This revealed a network of significant BOLD clusters spread across occipital, temporal, frontal, parietal, and subcortical areas, replicating previous work with this dataset 28 . Critically, the most significant cluster included the left and right amygdala as well as the left and right fusiform gyri (FG) and inferior occipital gyri (IOG). These latter two regions are key nodes in the cortical visual processing stream for faces, which may feed information forward to the amygdala 2 . We used the SPM Anatomy Toolbox to confirm the anatomical positions of our functional activations. In the absence of an anatomical template for the superior colliculus and the pulvinar, we masked the map of statistically significant voxels (p < .05, FWE) with our a priori manual anatomically-defined superior colliculi mask and functionally-defined pulvinar masks from Barron et al. (2015; see Online Methods for ROI generation). This revealed significant voxels in each area (proportion of significant voxels within each mask: left SC = 65.08%, right SC = 73.55%, left PUL = 36.13%, right PUL = 51.49%). Therefore, the "Emotion" HCP task established functional activation in the three subcortical nodes of interest, as well as two nodes of a potential cortical pathway to the amygdala for conveying information about emotional faces.

A forwards-only subcortical route is engaged during face processing
After observing significant BOLD signal in regions along the subcortical route as well as in other visual cortical areas, we next asked whether these regions were dynamically connected. We designed a space of testable models that mapped onto the functional hypothesis of a subcortical route to the amygdala that operates alongside a cortical visual pathway and is modulated by faces. Due to the presence of the IOG and FG in the whole-brain correct fMRI activation and their known roles in face processing 29 , we defined several plausible functional cortical connections to the amygdala. These consisted of reciprocal pathways between IOG and FG, FG and amygdala, IOG and amygdala, as well as pulvinar and IOG (see Fig. 5). We named models containing both a cortical and subcortical route to the amygdala as "Dual" models, whereas models in which the subcortical route was absent were named "Cortical". Our model space also included different sources of visual input, namely to the superior colliculus, pulvinar, or both, given that the pulvinar also receives direct retinal input 30 as well as input via the superior colliculus31. This gave us six families of models: 1) Cortical with SC input, 2) Cortical with PUL input, 3) Cortical with SC and PUL input, 4) Dual with SC input, 5) Dual with PUL input, and 6) Dual with SC and PUL input. We considered all possible combinations of forwards and reciprocal (forwards and backwards) cortical and subcortical connections, giving us a comprehensive model space of 102 models.
Of the available 622 participants, we conducted dynamic causal modelling on a subset of 237 participants who had sufficient above-threshold activation in all ROIs (these were defined by the subcortical masks used thus far and by spheres surrounding the coordinate of peak group BOLD signal in the IOG and FG; see Online Methods for more details). We conducted Bayesian model selection on the model space (grouped by families) to estimate how well the models explained the data, taking into account model complexity. We used the random effects implementation to account for potential individual differences in the recruitment of a subcortical pathway for viewing faces32. The winning family was the "Dual with SC and PUL input" family (expected probability = 67.34%, exceedance probability = 100%) and the winning model across the entire model space was also within this family (expected probability = 21.24%, exceedance probability = 98.01%, protected exceedance probability = 98.18%; see Fig. 5). This model included reciprocal cortical connections between IOG and FG and between FG and the amygdala. It also included a forwards-only subcortical route from the superior colliculus to the pulvinar to the amygdala, with input to both the superior colliculus and the pulvinar. The Bayesian Omnibus Risk score was P =1.78 x 10 -124 , indicating a very small chance that the winning model was indistinguishable from all models tested 33 . We replicated this finding (same winning family and winning model) on a subsample consisting of only the unrelated (i.e. non-sibling) participants within this group (49 participants; see Supplementary materials). The winning model revealed that the functional network that best explained the fMRI activity in our sample of 237 participants included visual inputs to the superior colliculus and pulvinar, forward connections from superior colliculus to the amygdala via the pulvinar, and recurrent interactions between IOG and FG, as well as between FG and amygdala. To extrapolate this finding to the general population and assess the consistency of dynamic coupling at each individual connection, we performed inferential statistics (t-tests) on the parameter estimates of each connection within the winning model (i.e. connectivity strength, in their natural space). We looked at the connectivity modulation parameters that represent the change in connection strength caused by the effect of faces. We removed extreme outliers (>3 SDs from mean) participants from each connection (M = 5.25, range = 3 to 8 participants excluded) and found that all connectivity modulations were significant except for the backward connection from left and right FG to left IOG (see Fig. 6). These results suggest that the modulation of these connections by faces was consistently strong and so we can infer that a subcortical route for processing faces is likely present within the general population.

Greater fibre density relates to stronger effective connectivity
Our findings from tractography, fMRI, and dynamic causal modelling provide convergent evidence for a subcortical route to the amygdala in humans. The final question we set out to answer was whether this converging evidence was correlated, such that participants with stronger structural connectivity also had stronger effective connectivity. In other words, we asked whether the structural connectivity along the subcortical amygdala route enables functional interactions amongst the nodes that lie within it. We computed eight correlations to examine the relationship between each parameter estimate and the corresponding global fibre count and local summed weights per connection (left and right SC-PUL and PUL-AMG). After removing outliers from each pairing according to Cook's Distance (threshold = 3 SDs from the mean), we discovered that participants with more global fibres also had greater modulatory activity for the right PUL-AMG connection (r = .200, pbonf = .023, N = 220; see Fig. 7). We successfully replicated this finding within a subsample of unrelated participants (49 participants; see Supplementary Materials). Thus, our study is the first to successfully harmonise functional and structural information about the subcortical route to the amygdala.

Discussion
The elusive subcortical route to the amygdala has posed a unique challenge in studies of the human brain, due to its depth and its rapid, short-lived activity. Evidence has accumulated over recent years across many studies using various neuroimaging modalities showing that this pathway may underlie primitive threatrelated behaviour. These studies, however, often take a unimodal approach on typically small samples, making it difficult to relate the specific structural connections between superior colliculus, pulvinar, and amygdala to observed functional brain activity and behavioural output. Our study, which used a large sample of participants from the HCP, unequivocally supports the existence of a subcortical route to the amygdala via the pulvinar in the healthy human adult brain that facilitates dynamic coupling between these regions and enhances fear recognition. We reconstructed the subcortical route to the amygdala using sophisticated tractography methods and found that the white matter fibre density significantly predicted individuals' ability to recognise fearful faces. We then computationally modelled the functional neural networks along this structurally connected network that were engaged while people viewed emotional faces. We found that it was far more likely for the network to include a subcortical visual route to the amygdala than a cortical route alone. Finally, we revealed converging evidence from structural and effectivity connectivity, such that the fibre density of the right pulvinar to amygdala pathway was positively correlated with the strength of the dynamic coupling (i.e. effective connectivity) between these regions. This study marks the first time that structural and effective connectivity have been concurrently investigated in the one, large sample to address the controversy on the existence and functional role of the putative subcortical route to the amygdala. Tractography of diffusion images is susceptible to both false positives and false negatives and thus is seldom used in isolation to determine the existence of particular neuroanatomical pathways 34 . We established the validity of our tractographically-reconstructed subcortical route by directly relating our measures of fibre density to both behaviour and effective connectivity, as well as by using two different tractography methods. Had the fibre density measures been simply due to noise, we would not have expected these theoretically relevant relationships with fearful face processing to emerge within this large sample of individuals. Our findings reconcile the long-debated question of whether a pathway from the superior colliculus to the pulvinar to the amygdala is anatomically present and functionally involved in processing visual threat information in humans 14,15,35 .
Another disputed feature of the subcortical route is its continuity, given discrepancies in the animal literature between which pulvinar subregions connect to the superior colliculus and the amygdala 15 . We provide new evidence that the right inferior pulvinar may mediate direct communication between the superior colliculus and the amygdala in the human brain, building upon previous animal research 31,[36][37][38] . Moreover, our results show that only the connection from the pulvinar to the amygdala is specifically related to fearful face processing. This suggests a central role for the pulvinar in processing emotional faces, as has been found by previous studies in the macaque 11,39 . We did not exhaustively explore the extent to which the cortex contributes information to the pulvinar-amygdala connection. The winning effective connectivity model, however, did not include cortical connections between the pulvinar and the inferior occipital gyrus or fusiform gyrus. Hence, it is unlikely that these cortical visual areas contribute to the information transmitted along the subcortical route. It remains to be shown how interactions between the pulvinar and other cortical areas, such as the inferotemporal cortex 17 , may influence activity along the pulvinar-amygdala connection.
Our findings open avenues for future studies on how this subcortical pathway might influence threat-related behaviour. Recent research has revealed that the superior colliculus, pulvinar, and amygdala show greater BOLD signal when individuals with autism fixate on the eyes of a face, more so than controls 40 . Future studies might therefore explore how the fibre density of this pathway differs in autistic populations. Furthermore, the propensity for the subcortical route to increase pre-attentive 41 or unconscious 2 processing of threatening stimuli might play a significant and relatively unexplored role in anxiety disorders, such as social anxiety and post-traumatic stress disorder.
In conclusion, our study settles a long-held debate over the existence and function of a subcortical route to the amygdala in the human brain. Our multimodal neuroimaging approach, leveraged by computational modelling, provides convergent evidence for a fundamental and conserved pulvinar-amygdala pathway that is specifically involved in fear. We demonstrate that the white matter tracts that form the subcortical structural pathway from the pulvinar to the amygdala enables functional, dynamic interactions amongst the regions that lie within it. Critically, we show that structural connectivity between the pulvinar and the amygdala leads to better recognition of fearful expressions in human faces.

Participants
We used the data from the publicly available Human Connectome Project (HCP) S900 release, collected between 2012 and 2016 24 . The S900 release contained 730 young adults with complete MRI and dMRI data, as well as fMRI data for the emotion task 42 . Of these, we excluded 95 people due to positive drug/alcohol tests and an additional 13 for abnormal colour vision. This resulted in a final sample of 622 participants aged between 22 and 36 years (M = 28.81, SD = 3.68 years), 259 of whom were male and 363 female, with 569 right-handed and 53 left-handed. Within our sample, 495 participants were related to one or more other participants (328 families in total). This included 53 pairs of monozygotic twins, 50 pairs of dizygotic twins, and 289 participants with one or more non-twin siblings in the sample. The remaining 127 participants were unrelated. We acknowledged that the many siblings in the HCP sample might spuriously decrease the variance in our neural measures (due to the structural and functional similarity between siblings, for example) and thus influence our statistics. Because of this, we replicated some of the analyses from the full sample on the subsample of unrelated participants (see Supplementary Materials).

dMRI Processing dMRI Acquisition
The HCP scanned participants in sessions over two days using a custom-made Siemens 3T "Connectome Skyra" (Siemens, Erlangen, Germany) with a 32-channel head coil, located at Washington University, St Louis, USA. They collected two separate T1-weighted high-resolution MPRAGE images (voxel size = 0.7mm isotropic, field of view = 224mm, matrix = 320, 256 saggital slices, TR = 2,400ms, TE = 2.14ms, TI = 1,000ms, flip angle = 8°, bandwidth = 210Hz per pixel, echo spacing = 7.6ms). We only used the first T1 image of the two sessions in our analysis. The HCP collected multi-band multi-shell diffusion-weighted images in a single session also using the Connectome Skyra

Global Intensity Normalisation
First, we corrected for low-frequency B1 field inhomogeneities in the dMRI volumes. We then conducted global intensity normalisation across each participant's corrected diffusion-weighted image so that we could later perform quantitative analyses of fibre density (i.e. apparent fibre density 45 ). This step normalises the median white matter b=0 intensity (i.e. non-diffusion-weighted image) across participants so that the proportion of one tissue type within a voxel does not influence the diffusion-weighted signal in another. Given our large sample size, we selected a subset of 62 participants (approximately 10% of the sample) to create a representative fractional anisotropy (FA) population template and white matter mask. We then used the population template and white matter mask to normalise the white matter intensity of all 622 participants' dMRI volumes.

Response Function Estimation
We segmented each participant's T1 image into five tissue types (cortical grey matter, subcortical grey matter, white matter, CSF, and pathological tissue) using the Freesurfer parcellation image provided by the HCP. We then estimated response functions (i.e. the signal expected for a voxel containing a single, coherently-oriented fibre bundle) for grey matter, white matter, and CSF using the Multi-Shell Multi-Tissue (MSMT) Constrained Spherical Deconvolution (CSD) algorithm 46 . After completing this step for all participants, we averaged their response functions to produce representative response functions per tissue type. We then conducted MSMT CSD on each participant again using the group averaged response functions, producing individual multi-tissue fibre orientation distributions (FODs).

fMRI Processing fMRI Acquisition
As with the dMRI data, the HCP acquired whole-brain gradient-echo echo planar imaging (EPI) data using the Connectome Skyra (TR = 720ms, TE = 33.1ms, flip angle = 52°, bandwidth = 2,290 Hz/Px, in-plane field of view = 208 x 180mm, 72 slices at 2mm thick, voxel size = 2mm isotropic, echo spacing = 0.58ms) with a multiband factor of eight. They collected data in a one-hour session (either on the same day as the dMRI or one day before/after) along with two or three other functional tasks in the HCP battery. For the emotion task, there were two runs, one with right-to-left phase encoding and the other with left-to-right phase encoding, each with 176 frames at a duration of 2 minutes and 16 seconds.

fMRI Task
The HCP developed the "emotion task" from the paradigm presented by Hariri, Mattay, Tessitore, Kolachana, Fera, Goldman, Egan and Weinberger 47 . Participants were presented with three visual stimuli at a time (one image at the top and two at the bottom) using E-Prime 48 . Participants were then instructed to make a button press indicating which of the two images at the bottom (left or right) matched the image at the top. Images were either a face (angry or fearful) or a shape (circle, horizontal oval, or vertical oval). The stimuli were presented on screen for 2,000ms separated by a 1,000ms inter-stimulus interval. A cue was presented at the beginning of each block to indicate the block type (i.e. "shape" or "face"), where each block contained either 6 faces trials or 6 shapes trials. Finally, a fixation cross was presented for eight seconds at the end of each of each run. The last block of each run only contained the first three trials due to a technical error that occurred early in HCP data collection. As the first block was always a shape block, our analysis was conducted on three shape blocks and 2.5 face blocks.

fMRI Preprocessing
We used the minimally preprocessed fMRI data provided by the HCP corrected for gradient distortion, motion, and field map-based EPI distortion. The HCP intensity normalised the data and spatially transformed it to MNI152 space using FSL (see Glasser et al. 2013 for full details on preprocessing pipeline). We further increased the signal-to-noise ratio of the fMRI data in SPM12 (SPM12, www.fil.ion.ucl.ac.uk/spm) by applying spatial smoothing using a 4mm Gaussian kernel 49 .

Regions of Interest
We chose the superior colliculus, pulvinar, and amygdala as our ROIs We created masks of these ROIs in standard MNI space using FSL. For the amygdala (AMG) mask, we used the probabilistic Harvard-Oxford Subcortical atlas at a threshold of at least 50% probability. For the pulvinar (PUL), we were interested in the structure as a whole, as well as its subregions. To do this, we used the parcellated pulvinar mask generated by Baron et al. (2015), who isolated five distinct pulvinar clusters based on functional coactivation profiles in fMRI data from 29,597 participants across 7,772 experiments 12 . For the pulvinar as a whole ROI, we merged the five clusters together and used FSL to manually fill any holes in the resultant mask. Finally, we manually created masks for the left and right superior colliculi (SC) in the absence of an atlas-based mask by drawing the boundaries of the superior colliculus over the MNI152 single participant T1 template with reference to an anatomical atlas 50 and filling the centre. We then used FSL to warp these masks into native diffusion space for each participant's tractography analysis. All our ROIs in MNI space are freely available online from the Open Science Framework: doi:10.17605/OSF.IO/KBPWM.

dMRI analysis
In this study, we implemented two tractography methods that use different approaches to white matter reconstruction for cross-method validation. We first used the multi-tissue model of global tractography. This method takes a Bayesian approach to reconstructing a full-brain fibre configuration using a generative signal model to best explain the underlying data. It is less sensitive to noise that may accumulate for longer distance tracts in other "local" tractography methods throughout their stepwise approach 51,52 . Hence, for comparison, we computed probabilistic tractography between our regions of interest 53 . This method also uses a Bayesian approach to account for one or more distributions of fibre orientations within each voxel, thus incorporating uncertainty into the model 54 . To acquire a biologically-accurate measure of apparent fibre density 45 along the resultant streamlines, we used the Spherical-Deconvolution Informed Filtering of Tractograms version 2 (SIFT2) method to weight each streamline by a cross-sectional area multiplier directly related to the underlying data 26,45 .

Global tractography
Global tractography is a data-driven Bayesian approach to estimating the whole-brain fibre configuration that best explains the underlying diffusion-weighted images. As opposed to local streamline tracking, global tractography accounts for the spatial continuity of fibres and thus is better able to discriminate crossing and fanning fibre geometries 51 . Furthermore, because the simultaneously-reconstructed fibre configurations are optimised with respect to the data at hand, the density of the final tractogram quantitatively represents the apparent fibre density (AFD; i.e. the proportion of space occupied by white matter fibres 45 ).
We conducted global tractography on the global-intensity-normalised DWI volumes for each participant using the group-averaged multi-tissue response functions. After 250 million iterations to optimise a full brain reconstruction, we filtered the tractogram using the ROI masks described above to isolate fibres that terminated in 1) both the superior colliculus and pulvinar masks, and 2) both the pulvinar and amygdala masks. We also used the masks for the five individual functionally-defined pulvinar subregions to isolate the subregion-specific fibres connecting to the superior colliculus and to the amygdala. Before any statistical analysis, we removed four participants with extreme values more than 3 SDs from the mean of all connections, as these likely indicated artefactual imaging data (final N = 618). These outliers were also removed from the pulvinar subregion-specific analyses.

Local tractography
As a less conservative approach than global tractography, we also conducted local probabilistic tractography using our ROIs as seeding and terminating regions. We used the iFOD2 algorithm, iteratively planting a seed point 25,000 times (or until at least 10,000 tracks had been selected and written) in each voxel of the seeding ROI 44 . We applied the anatomically-constrained variation of this technique, whereby each participant's five-tissue-type segmented T1 image provided biologically-realistic priors for streamline generation, reducing the likelihood of false positives 55 . We edited the final streamlines so that only those that terminated in our ROIs remained.
Using these methods, we traced streamlines between the two halves of the subcortical route (i.e. SC to PUL, PUL to AMG). We then reversed the seeding location (i.e. PUL to SC, AMG to PUL) and based all statistics on the average between the forwards and backwards seeding directions to reduce any influence of possible asymmetries in seed ROI volume. We applied SIFT2 to these streamlines to enable us to quantitatively assess the connectivity. SIFT2 makes this possible by weighting the streamlines by a cross-sectional multiplier such that the sum of these weighting factors better represents the underlying white matter fibre density 55 . Before computing any statistics, we removed 13 participants whose average summed weights across each connection were more than 3 SDs from the mean (final N = 609). These outliers were also absent from the pulvinars subregion-specific analyses.

Pulvinar subregion analysis
We were interested in which of the subregion-specific pathways between the superior colliculus and the pulvinar correlated with those between the amygdala and pulvinar. We constructed correlation matrices with 5 subregion SC-PUL variables on one axis and 5 subregion PUL-AMG variables on the other. This was done per hemisphere and per local/global tractography method (see Fig. 2). The SC-PUL variables and PUL-AMG variables were highly intercorrelated within each set, meaning that any relationship between an SC-PUL subregion and PUL-AMG subregion would be strongly influenced by other SC-PUL subregions. To isolate the unshared variance of each subregion-specific pathway, we performed principal components analysis (using Varimax rotation orthogonalise the variables) to specifically extract five components. Bartlett's test of sphericity was significant for each analysis and the minimum Kaiser-Meyer-Olkin measure of sampling adequacy was at least .696 for global tractography and .687 for local tractography.

General linear modelling
Using the spatially smoothed fMRI data, we convolved the onset of each Face and Shape block with a canonical hemodynamic response function (HRF) using SPM12. We closely modelled this first-level general linear model (GLM) analysis on the work by Hillebrandt et al., (2014), such that we did not slice time correct the multiband data due to the fast TR. We partitioned the GLM into sessions (left-to-right and right-to-left encoding) and we included head motion as a regressor. We generated statistical parametric maps (SPMs) of the expected BOLD signal for faces minus shapes and shapes minus faces.
We then entered the faces minus shapes contrast into a second-level analysis (a one-sample t-test) across all participants. After examining the estimated BOLD signal to Faces at the whole-brain level (p < .05, familywise error corrected using Random Field Theory 11 ), we applied the superior colliculus, pulvinar, and amygdala a priori defined masks to more specifically estimate functional activation in these anatomicallydefined areas.

Dynamic causal modelling
We implemented Dynamic Causal Modelling (DCM) to infer the causal direction of information flow between neural regions using a biophysically informed generative model 56 . First, we examined the map of significant activation produced by the fMRI analysis of the "Emotion" HCP task. Based on this and our a priori hypotheses, we defined the left and right superior colliculus, pulvinar, amygdala, inferior occipital gyrus (IOG), and fusiform gyrus (FG) as our ROIs. For the two gyri, we used MNI coordinates of the most significant peak from the group level analysis (left IOG: -22 -92 -10, right IOG: 28 -90 -8, left FG: -38 -50 -20, right FG: 40 -52 -18). We then placed spheres with a radius of 12mm around these four coordinates to search for the participant-specific local maxima within each participant's session-specific SPM for the faces minus shapes contrast (adjusted for the t effects of interest, p < .05 uncorrected). Note that for the purposes of extracting the fMRI data for the DCM nodes, one does not need corrected p-values 49 . Next, we defined the ROIs by a 6mm radius sphere around the participant-and session-specific local maxima. For the subcortical areas of interest, we defined the initial search radius by the anatomically defined ROI masks (as described above) instead of significant peaks from the group analysis to confine our search within subcortical grey matter.
We used a "two-state" DCM model, which accounts for both excitatory and inhibitory neural populations 49,57 . Our model space was dictated by our specific, theory-driven hypotheses about subcortical and cortical visual pathways to the amygdala, as well as by the significant regions of the BOLD signal observed at the group level in our GLM analysis. Both face and shape blocks contributed to input parameters within each model. All endogenous and intrinsic connections in each model were modulated by the effect of faces over shapes.
To specify a DCM, each participant needed to have above-threshold activation (at p < .05, uncorrected) within each ROI across both scanning sessions. This was the case for 237 out of the 622 participants. The ROIs with the lowest number of above-threshold participants were the left and right superior colliculi (69 and 58 participants, respectively), followed by the left and right pulvinar (4 and 3 participants, respectively). This may be due to the bilateral superior colliculi's relatively smaller volume as well as lower statistical power (its mean t-statistic was approximately 10.57 compared with 33.35 for IOG and 30.14 for FG). Critically, the 237 participants with above-threshold activity in all ROIs did not differ significantly from the other 385 participants in the global or local tractography results, performance on the Penn Emotion Recognition task, volume of the thalamus/amygdala/fusiform area/lateral occipital area (as computed by Freesurfer), reaction time/accuracy to faces/shapes during the fMRI task, age, or gender. Therefore, using the information available to us, we had no evidence to assume that our DCM sample was biased by any confounding variable.