Skip to main content
NCBI home page
NIHPA Author Manuscripts logoLink to NIHPA Author Manuscripts
. Author manuscript; available in PMC: 2022 Aug 4.
Published in final edited form as: Neuron. 2021 Jun 23;109(15):2404–2412.e5. doi: 10.1016/j.neuron.2021.06.001

A human colliculus-pulvinar-amygdala pathway encodes negative emotion

Philip A Kragel 1,2,3,#,*,†, Marta Čeko 1, Jordan Theriault 4, Danlei Chen 4, Ajay B Satpute 4,5, Lawrence L Wald 5, Martin A Lindquist 6, Lisa Feldman Barrett 4,5, Tor D Wager 1,7,*
PMCID: PMC8349850  NIHMSID: NIHMS1717997  PMID: 34166604

Summary:

Animals must rapidly respond to threats to survive. In rodents, threat-related signals are processed through a subcortical pathway from the superior colliculus to the amygdala, a putative ‘low road’ to affective behavior. This pathway has not been well characterized in humans. We developed a novel pathway-identification framework, which uses pattern recognition to identify connected neural populations and optimize measurement of inter-region connectivity. We first verified that the model identifies known thalamocortical pathways with high sensitivity and specificity in 7T (N = 56) and 3T (N = 48) fMRI experiments. Then, we identified a human functional superior colliculus-pulvinar-amygdala pathway. Activity in this pathway encodes the intensity of normative emotional responses to negative images and sounds, but not pleasant images or painful stimuli. These results provide a functional description of a human ‘low road’ pathway selective for negative exteroceptive events and demonstrate a promising method for characterizing human functional brain pathways.

eToc Blurb:

Kragel et al. implement a new functional connectivity technique that estimates inter-region connections using local multivariate patterns. They identify a superior colliculus-pulvinar-amygdala pathway at 7T and 3T that responds to aversive images and sounds, but not painful stimuli. This method provides a general framework for identifying and characterizing functional pathways.

Rapidly detecting threats in the environment is crucial for survival. In rodents, behavioral responses to visual danger signals are mediated by neural pathways from the superior colliculus to the pulvinar (Zhou et al., 2017) and pulvinar to the amygdala (Wei et al., 2015). This pathway, sometimes referred to as the “low road” to amygdala-mediated threat (de Gelder et al., 2011; LeDoux, 1998; Pessoa and Adolphs, 2010), enables reflexive behavioral responses (e.g., orienting) via collicular projections (Schneider, 1969), and interfaces with mechanisms in the amygdala that shape current and future defensive behaviors (Tamietto and de Gelder, 2010). The target of this pathway, neurons in the basolateral amygdala (Wei et al., 2015), comprises one of many interdigitated neural populations involved in pain-related affect (Corder et al., 2019), anxiety (Tye et al., 2011), reward and punishment (Burgos-Robles et al., 2017), among other behaviors (Janak and Tye, 2015). In humans, this pathway is thought to mediate unconscious processing of affective visual stimuli (Morris et al., 1999; Tamietto and de Gelder, 2010; Vuilleumier et al., 2003; Whalen et al., 2004). However, the pathways identified in animal research reflect communication among specific neural ensembles (Janak and Tye, 2015; Kyriazi et al., 2018) that are not clearly resolved with human neuroimaging. Though structural connections consistent with the collico-pulvinar-amygdala pathway have been identified (Abivardi and Bach, 2017; McFadyen et al. 2019; Rafal et al., 2015; Tamietto et al., 2012), its precise location and functions in humans requires further study. In particular, it has not been functionally distinguished from other amygdalar circuits, and there is as of yet no defined measure of functional connectivity in this pathway. More precise identification of the pathway would allow more precise characterization of its relationships with human affect and behavior.

The adoption of pattern recognition techniques has moved fMRI closer to the level of neural representation in humans (Kragel et al., 2018; Norman et al., 2006; Poldrack and Farah, 2015), potentially even uncovering information coded by cortical columns (Haynes and Rees, 2005; Kamitani and Tong, 2005), but the mainstay of connectivity research in humans examines relationships between coarse anatomical regions, not neural populations. As a result, there are ongoing debates about the colliculus-pulvinar-amygdala pathway, including whether it exists in primates (Pessoa and Adolphs, 2010). If so, it is unclear whether it selectively responds to emotional events (as opposed to attentionally demanding or behaviorally relevant events more generally), and whether it relates to human subjective emotional experience (LeDoux, 2014). We addressed these questions by extending pattern recognition approaches in fMRI (Haxby et al., 2001; Kriegeskorte et al., 2006; Poldrack and Farah, 2015) to modeling multivariate brain pathways in humans. We sought to identify a human colliculus-pulvinar-amygdala pathway and characterize its relationships with human emotion and pain.

Recent advances in connectivity models using multivariate fMRI patterns have the potential to identify human pathways connecting specific neural populations with increased precision (Anzellotti and Coutanche, 2018; Anzellotti et al., 2017; Coutanche and Thompson-Schill, 2013; Woo et al., 2014). Conventional approaches focus on the connectivity among anatomical regions or large functional parcels (Friston, 1994), which averages over many off-target signals. Multivariate pattern analysis can help by identifying signals that are more sensitive to particular subsets of neural populations (Haxby et al., 2001; Kamitani and Tong, 2005). For example, distinct multivariate patterns within the same brain region can exhibit unique profiles of connectivity and differential associations across varieties of affect (Anzellotti and Coutanche, 2018; Anzellotti et al., 2017; Coutanche and Thompson-Schill, 2013; Woo et al., 2014).

The approach we developed–Multivariate Pathway Identification (MPathI)–was designed specifically to model connections between distinct populations located in different brain regions, as is frequently demonstrated in contemporary animal studies (Deisseroth, 2011). MPathI accomplishes this goal using an extension of Partial Least Squares to optimize the covariance between two multivariate patterns (latent sources reflective of neural populations) and estimate the strength of functional correlations between them. Training the model to identify a population-level (group) pattern and testing on independent participants (Kragel et al., 2018; Woo et al., 2017) stabilizes model weights, prioritizes generalizability, and provides unbiased estimates of (a) connectivity measured by the pathway model (b) prediction of subjective emotion, and (c) differences from standard connectivity measures. Unlike similar measures (e.g., canonical correlation (Hardoon et al., 2007)), optimizing covariance prioritizes identification of larger, more robust signals in each region (see STAR Methods). In addition to testing whether particular pathways can be identified in humans with the assistance of pattern recognition algorithms, MPathI and related techniques provide optimized measures of pathways that can be tested for relationships with behavior, psychopathology, and treatments.

We first evaluated whether MPathI can be used to identify thalamocortical pathways with high sensitivity and specificity. This is important because fMRI-based measures of functional correlations may be driven by signals other than direct connections (e.g., by global signals or indirect connections). To this end, we acquired high-resolution (1.1 mm isotropic) functional MRI at 7T with whole-brain coverage (N = 56; see STAR Methods). We estimated the strength of connectivity in monosynaptic pathways from lateral geniculate nucleus (LGN) to primary visual cortex (V1) and from medial geniculate nucleus (MGN) to primary auditory cortex (A1). Pathway models produced multivoxel patterns in each pair of regions that maximally covary (i.e., LGNV1-V1LGN and MGNA1-A1MGN). We compared the strength of these connections against crossed (off-target) connections (i.e., LGNV1-V1MGN and MGNA1-A1Lgn), which served as negative controls. Connectivity in target pathways (i.e., LGNV1-V1MGN and MGNA1-A1Lgn) was strong and positive (r = 0.3891 ± .0167 sem, p < 0.0001), whereas “crossed” connections were nonsignificant (r = 0.0359 ± .0205 sem). The direct comparison was highly significant (Δr = .353 ± .024 sem, t55 = 13.83, p < .0001, see Fig. 1B). Connectivity strength discriminated between target and off-target pathways with 85% ± 2.4% accuracy (sensitivity = 95%, 95% CI = [90% 98%]; specificity = 76% [68% 84%]; AUROC = .92).

Figure 1. Multivariate Pathway Identification (MPathI) discriminates direct from indirect thalamo-cortical sensory pathways.

Figure 1.

Open in a new tab

(A) Visual (LGN-V1, shown in red) and auditory (MGN-A1, shown in green) regions of interest used to evaluate the sensitivity and specificity of MPathI. (B) Axial views of patterns in the primary auditory cortex (A1), medial geniculate nucleus (MGN), lateral geniculate nucleus (LGN), and primary visual cortex (V1) that exhibit maximal functional connectivity with one another. Patterns were identified using Partial Least Squares to maximize connectivity between MGN and A1 (i.e., patterns MGNA1 and A1MGN) in one model and LGN and V1 in another (i.e., patterns LGNV1 and V1LGN). Unthresholded patterns that exhibit maximal connectivity are shown for display purposes. Cool colors correspond to negative model weights whereas positive colors indicate positive weights. (C) Correlations for direct (LGNV1-V1LGN and MGNA1-A1MGN) and indirect (LGNV1-V1MGN and MGNA1-A1LGN) pathways fit on preprocessed BOLD time-series from Expt. 1 (N = 56). Each circle corresponds to one participant, solid black lines indicate mean, light gray regions indicate one standard deviation, and dark gray areas indicate 95% confidence intervals. ***p < .0001

These findings were replicated in a separate fMRI study at conventional 3T field strength (Experiment 2, N = 48), with some loss of precision but good sensitivity and specificity. Experiment 2 differed in other important ways as well: it included stimulation with affective visual, auditory, mechanical, and thermal stimuli, and the MPathI model was estimated at the population (group) level, allowing correlation between pathway response and affective experience (see below and Methods). MPathI demonstrated positive cross-validated connectivity estimates for target pathways (r = 0.547 ± .031 sem, p < 0.0001), that were stronger than estimates for off-target connections (r = 0.492 ± .027 sem; Δr = .055 ± .016 sem, t47 = 3.86, p < .0001; accuracy = 76.0%; AUROC = .783). Together, these observations suggest that MPathI is sensitive to connections in known thalamocortical pathways and shows regional specificity across nearby regions of the thalamus. Such specificity would not be expected if pathway connectivity estimates were driven by global signals shared across multiple brain regions.

Given evidence that MPathI can accurately identify known pathways, we next examined whether it could be used to detect a superior colliculus-pulvinar-amygdala pathway in humans at 7T. This pathway is quiescent at rest in non-human animals (Ramcharan et al., 2005; Zhou et al., 2018), making it difficult to isolate in the absence of behavior (e.g., during resting-state fMRI). To induce relevant activity, we scanned participants during tasks involving visual manipulation of negative emotion and painful mechanical stimulation (see STAR Methods). Despite evidence of anatomical connectivity in humans (Abivardi and Bach, 2017; Rafal et al., 2015) the colliculus and amygdala connect to potentially different pulvinar populations (Pessoa and Adolphs, 2010) making it unclear whether the pathway is conserved across species, and if it is, how best to optimize a measure that captures its connectivity. To accommodate the possibility of both direct and multi-synaptic pathways through the pulvinar (Pessoa and Adolphs, 2010), we fit separate models for the superior colliculus-pulvinar and pulvinar-amygdala segments of the pathway (Fig. 2B). In Experiment 1, MPathI revealed patterns of activity in the superior colliculus optimized to covary with activity in the pulvinar (which we denote SCPulv), and patterns of activity in the pulvinar and amygdala optimized to covary with one another (i.e., PulvAmy and AmyPulv; Fig. 2B). Both the pulvinar-amygdala model (r = .7071 ± .0179 sem, t55 = 39.588,p < .0001; Fig. 2C) and superior colliculus-pulvinar model (r = .7652 ± .0128 sem, t55 = 59.683, p < .0001) revealed strong connectivity. Connectivity was substantially stronger and more consistent across individuals than estimates from standard approaches based on region-average signals (pulvinar-amygdala: correlation difference Δr = .402 ± .032 sem, t55 = 12.766, p < .0001; superior colliculus-pulvinar: Δr = .328 ±.026 sem, t55 = 12.618,p < .0001). These observations demonstrate that our approach captures information conveyed by fine-scale patterns of fMRI activity that are not captured by standard connectivity measures.

Figure 2. A colliculus–pulvinar–amygdala pathway selective for negative emotion in humans.

Figure 2.

Open in a new tab

(A) Hypothesized pathway and anatomical regions of interest rendered in MNI space. (B) Axial views of patterns in the amygdala (Amy, top; peak t55 = 3.36, p = .0014), pulvinar (Pulv, middle; peak t55 = 4.00, p = .0001), and superior colliculus (SC, bottom; peak t55 = 3.32, p = .0016) that exhibit maximal functional connectivity with one another (target regions are denoted in subscript). Patterns that maximize connectivity between the pulvinar and amygdala were identified in one model and the superior colliculus and pulvinar in another. Unthresholded patterns that exhibit maximal connectivity are shown for display purposes. Cool colors correspond to negative model weights whereas positive colors indicate positive weights. (C) Pathway models are more sensitive than conventional measures based on the mean signal of each region. Plots depict functional correlations estimated with multivariate pathway models (yellow circles) and mean signal in each region (gray circles). Each point corresponds to a single subject. The left panels depict data from Expt. 1 (N = 56) and the right panels Expt. 2 (N = 48). Solid lines depict mean connectivity estimates, and boxes indicate standard deviation and 95% confidence intervals. (D) Activation of the PulvAmy pattern mediates the relationship between SCPulv activity and negative emotion for images (top row) and sounds (bottom row). Scatter plots show the relationship between SCPulv activity and PulvAmy activity (path a, left) and between PulvAmy activity and Negative Emotion (path b, right). Lines show least squares fit between variables for each subject (N = 48). *p < .05, **p < .01, ***p < .001

In animal models, the superior colliculus-pulvinar-amygdala pathway is selectively activated by threat cues, particularly visual cues (Wei et al. 2015). In Experiment 2 (N = 48), we evaluated whether human MPathI connectivity estimates are consistent with this pattern of selectivity. We scanned participants at 3T during a multi-modal affect induction task designed to test the selectivity of affective responses. The task included five different types of stimuli assessed across two separate scanning sessions: Negative images, negative sounds, mechanical pain, thermal pain, and positive images, with 4 intensity levels per stimulus type (20 event types). We estimated BOLD responses to each stimulus presentation (6 trials × 4 intensity levels × 5 stimulus types × 48 participants, totaling 5,760 single-trial images). We fit pathway models using data aggregated across multiple subjects and evaluated their performance using cross-validation in independent subjects with 48-fold cross-validation. As in Experiment 1, both the pulvinar-amygdala model (r = .683 ± .017 sem, t47 = 25.83, p < .0001) and superior colliculus-pulvinar model (r = .859 ± .010 sem, t47 = 35.50, p < .0001) revealed strong connectivity (Fig. 2C), at levels greater than those estimated using standard approaches based on region-average signals (pulvinar-amygdala: Δr = .151 ± .021 sem, t47 = 8.22, p < .0001; superior colliculus-pulvinar: Δr = .177 ±.012 sem, t47 = 16.97, p < .0001).

We next tested the colliculus-pulvinar-amygdala pathway’s sensitivity and selectivity for negative emotional stimuli, compared with salient painful and positive emotional stimuli. The colliculus-pulvinar-amygdala pathway in rodents conveys information about threats, whereas reward and pain signals reach the amygdala through different pathways, e.g., via other thalamic nuclei or parabrachial projections (Doron and LeDoux, 1999; Han et al., 2015; Palmiter, 2018; Tye et al., 2008). Therefore, we expected the pulvinar-amygdala pathway to selectively track the aversiveness of negative images (and possibly sounds), but not painful heat, painful pressure, or positive images. To test this prediction, we regressed PulvAmy responses on normative rankings of aversiveness and pleasantness (see STAR Methods for details). PulvAmy responses increased linearly with the subjective aversiveness of pictures (β= 0.0011 ± 0.00026 sem, t47 = 4.363, p < .0001) and sounds (β= 0.0019 ± 0.00026 sem, t47 = 4.363, p < .0001), showing that the pathway tracks negative emotion in humans across multiple sensory modalities. However, PulvAmy responses showed little relationship with mechanical pain (β= 0.0005 ± 0.00024 sem, t47 = 1.91, p = .0622), thermal pain (β= 0.00025 ± 0.00029 sem, t47 = .852, p = .399), or pleasantness of positive pictures (β = −0.00026 ± 0.00028 sem, t47 = −0.904, p = .371). PulvAmy activation was more sensitive to the aversiveness of images and sounds than other stimulus types (Δβ= 0.0013 ± 0.00024 sem, t47 = 5.600, p < .0001). Similar results were found for colliculus and amygdala patterns (See Supplemental Table S1); however, interestingly, effect sizes were smallest in the amygdala, perhaps because of its functional heterogeneity (Janak and Tye, 2015; Kyriazi et al., 2018) or local fMRI signal inhomogeneity. These results establish that a pathway model optimized for colliculus-pulvinar-amygdala connectivity (without reference to behavior) predicts normative human judgements of aversiveness and that this pathway is sensitive and specific to exteroceptive processing of threatening stimuli as in non-human research (Wei et al., 2015).

Human affective responses may be mediated by both direct and indirect pathways from the superior colliculus to the amygdala. The existence of a direct colliculus-pulvinar-amygdala pathway is contentious because threat-related collicular inputs target the inferior pulvinar, but pulvinar-amygdala projections originate in the medial pulvinar (Pessoa and Adolphs, 2010). This region exhibits strong functional correlations with higher-level visual areas (Arcaro et al., 2018), suggesting cortical involvement in threat processing, rather than a direct subcortical pathway. However, other findings show that projections from inferior layers of superior colliculus to the medial pulvinar and on to the amygdala mediate “fear-like” freezing in rodents (Wei et al., 2015). To evaluate evidence for a direct (vs. indirect) pathway, we performed a multi-level mediation analysis (Atlas et al., 2010), testing if fMRI activity in PulvAmy (the pulvinar pattern optimized for amygdala connectivity) formally mediates the relationship between SCPulv activity and normative human judgments of aversiveness. Significant mediation implies that SCpulv-pulvAmy connectivity covaries with the pulvAmy to emotion intensity connection and is most consistent with a direct pathway. We observed significant mediation effects (Fig. 2D) for both aversive pictures (path a × b = 3.720 ± 1.566 sem, z = 2.356, p = .018) and sounds (path a × b = 16.379 ± 3.579 sem, z = 3.580, p < .0001). A similar mediation analysis using online measures of self-report revealed similar effects for negative sounds (path a × b = 2.00 ±.450 sem, z = 3.722, p <.0001), which varied primarily in terms of basic acoustic features, but not images (path a × b = .2344 ± .2529 sem, z = .8543, p = .3930), suggesting that this pathway may not be sensitive to evaluative processes that either vary over time or between individuals. These results identify a human subcortical pathway involved in aversiveness and are consistent with direct projections from superior colliculus to pulvinar and on to amygdala, though they do not preclude the existence of other pathways as well.

Standard connectivity approaches did not reveal significant mediation of aversiveness. Mediation using the mean signal in each region produced much smaller, non-significant effects, both for pictures (path a × b = 0.0029 ± 0.0171 sem, z = 0.1728, p = 0.875, 95% CI of difference = [−2.738 −13.964]) and sounds (path a × b = 0.0186 ± 0.0283 sem, z = 0.711, p = 0.518, 95% CI of difference = [−13.890 -31.820]). Thus, standard connectivity approaches are not sufficient to identify subcortical pathways mediating emotional experience.

The pathway we identified likely does not operate in isolation. Recent neuroscientific accounts (Pessoa and Adolphs, 2010) and studies of the amygdala’s role in coordinating widely distributed activity (Gründemann et al., 2019; Stringer et al., 2019) suggest that the colliculus-pulvinar-amygdala pathway may be part of a broader cortical-subcortical network for generating and coordinating threat responses, and multiple cortico-amygdala pathways could play a similar role. To compare cortical and thalamic pathways to the amygdala, we used MPathI to identify optimized pathways connecting the amygdala with each of a series of local regions spanning the thalamus (Krauth et al., 2010) and cortex (Glasser et al., 2016). We performed principal component analysis (PCA) on the amygdala-related cortical and thalamic pattern responses (“pathways” to amygdala below) to identify commonalities in amygdala connectivity across multiple pathways (Fig. 2) and situate the pulvinar-amygdala pathway among them.

This analysis revealed that activity in the pulvinar-amygdala pathway was functionally similar to pathways connecting the amygdala with multiple portions of ‘visual’, ‘somatomotor’, ‘frontoparietal’, ‘attention’, and ‘limbic’ networks (Yeo et al., 2011). One group (Component 1, Fig. 3A) included the “low road” pulvinar-amygdala pathway, pathways from the insula and mid-cingulate cortex to amygdala, and others. It was related to multiple large-scale networks (Fig. 3B), and was the only component associated with ‘somatomotor’, ‘attention’, and ‘frontoparietal’ networks. This group selectively responded to aversive images and sounds, but not painful or positive stimuli (Fig. 3C; Table S2). A second group of pathways (Component 2) included parts of the traditional visual cortical “high road” to the amygdala, including amygdala connections with the lateral geniculate nucleus and visual cortex (predominantly the ventral stream), and was associated with ‘visual’ and ‘limbic’ cortical networks (Fig. 3B). These pathways selectively responded to the aversiveness of negative pictures, but not other stimuli. A third group (Component 3) included amygdala connections with anterior temporal lobe and OFC, and loaded predominantly on the ‘limbic’ cortical network (Fig. 3B). These pathways were not associated with emotional intensity in this study. They are known to have dense connectivity with the amygdala and have been implicated in mnemonic (Olson et al., 2007) and associative (Shenhav et al., 2013) processing of affective information, including contextual influences on emotion that were not manipulated here. Together, these findings identify three different profiles of amygdala activity, each correlated with distinct large-scale systems. The colliculus-pulvinar-amygdala pathway was part of a broader system anatomically and functionally consistent with a “low road” to visually and aurally induced threat, whereas other systems were consistent with “high road” visual cortical pathways to amygdala and conceptual “top down” influences on amygdala.

Figure 3. Distributed pathways to the amygdala are sensitive to negative stimuli.

Figure 3.

Open in a new tab

(A) Decomposition of cortical and thalamic signals that covary with activity in the amygdala, rendered onto the brain. pathway models were estimated between the amygdala and cortical areas identified from a multimodal parcellation (Glasser et al., 2016), and an anatomical atlas of thalamus and adjacent structures (17 regions). Principal component analysis reduces 197 estimates of activation that covary with the amygdala to three orthogonal dimensions. Loadings of each region in this three-dimensional functional space are conveyed using an RGB colorspace (red = component 1, green = component 2, blue = component 3). (B) Spatial similarity of principal components and resting-state networks from Yeo et al. (Yeo et al., 2011. The width of each line indicates the degree of correlation between each component and binary maps of large-scale cortical networks. The first principal component is correlated with ‘somatomotor’, ‘frontoparietal’, and ‘attention’ networks; the second component loads on ‘visual’ and ‘limbic’ networks; and the third component solely loads predominantly on the ‘limbic’ network. (C) Relationship between activation of each component and the intensity of affective stimuli. Scatterplots show the relationship between activity of each component and the intensity of affective stimuli. Lines show least squares fit between variables for each subject (N = 48). (D) Scatterplot depicts the mean and standard error of model coefficients for each region in the three-dimensional space, highlighting gradients of activity across regions (see Table S3 for full details). The pulvinar region (Pulvamy) of focus here is in the top right breakout, with among the highest loadings on Component 1 and near-zero loadings on Components 2 and 3.

The pattern is consistent with “multiple roads” accounts (Pessoa and Adolphs, 2010), but also consistent with the importance of the collicular-pulvinar-amygdala pathway for unconscious affective responses to visual stimuli (Tamietto and de Gelder, 2010). The amygdala appears to be important for some types of negative affect but not others, and other pathways may be more relevant for affective responses in other modalities (taste, smell, or somatosensory). In addition, the relationship with conscious affective and emotional experience requires further study. Lesion studies indicate that the amygdala is not always necessary for emotional experience (Anderson and Phelps, 2002; Feinstein et al., 2013). We found the colliculo-pulvinar pathway tracked normative affect responses across aversive sounds and images, demonstrating multi-modal properties and extending previous work showing pulvinar and amygdala coactivation in response to visual threat cues (Liddell et al., 2005; Morris et al., 1999; Vuilleumier et al., 2003). Pathway activity also tracked reported affective experience, though it did so more strongly for sounds than images. Notably, the sounds used here are perceptually simpler and more uniform than the emotional images, which may be evaluated variably across individuals and require more conceptual processing. Recent theories (Barrett, 2017; LeDoux and Brown, 2017) posit that transformation and re-representation of sensory signals is a defining feature of emotional experience, and consciousness more generally (Dehaene et al., 2017). In light of these views, and the many pathways that convey emotion-related information (Gothard, 2020), the sensory signals carried by the colliculus-pulvinar-amygdala pathway likely do not explain the full complexity of emotional experience in humans, but contribute as one of multiple stages in a distributed information processing system.

In sum, this study provided a more precise functional identification and characterization of the colliculus-pulvinar-amygdala pathway than has previously been available in humans. The pathway exhibited a strong and selective relationship with the aversiveness of auditory and visual stimuli. This and other thalamocortical pathways were most clearly identified using ultra high-field fMRI (7T; also see Wang et al., 2020), demonstrating a benefit of high-resolution imaging that may provide increasing advantages as analytic tools develop to capitalize on it. However, the multivariate pattern approach inherent in MPathI helped to identify signals clearly related to specific functional pathways even at 3T. MPathI measures substantially outperformed conventional region average-based connectivity and discriminated amygdala contributions from the colliculo-pulvinar pathway from at least two other components contributing to amygdala activity. These findings support the promise of MPathI and related techniques as important tools for “next-generation” human brain connectivity (Anzellotti and Coutanche, 2018; Anzellotti et al., 2017; Basti et al., 2020; Woo et al., 2017). The ability to identify such pathways in humans addresses a crucial gap between rapidly emerging animal research on neural pathways and the assessment of functional correlations with human neuroimaging. It provides a framework for investigating the functional sensitivity and specificity of brain pathways across task paradigms and species, and provides a step towards understanding how activity in multiple pathways jointly relates to subjective affective experience.

STAR Methods

Resource Availability

Lead Contact.

Further information and requests for resources should be directed to and will be fulfilled by the Lead Contact, Philip Kragel (pkragel@emory.edu).

Materials availability.

This study did not generate new unique reagents.

Data and code availability.

fMRI data are available at https://osf.io/werk2/. MATLAB code for analyses is available at: https://github.com/canlab.

Experimental Model and Subject Details

7T fMRI Study (Expt 1)

This study included 56 participants (Mage = 26.46 years, SD = 5.87 years, 27 female). All recruited participants were between the ages of 18 and 40 years, were right-handed, had normal or corrected to normal vision, were not pregnant, were fluent English speakers, had no known neurological or psychiatric illnesses, and were recruited from the greater Boston area. Participants were excluded from the study if they were claustrophobic or had any metal implants that could cause harm during scanning. All participants provided written informed consent and study procedures were completed as approved by the Partners’ Healthcare Institutional Review Board.

3T fMRI Study (Expt 2)

This study included 48 adult participants (mean ± SD age: 25.1 ± 7.1; 21 female, 27 male; 7 left-handed; 40 white and 8 non-white (1 Hispanic, 5 Asian, 1 Black, and 1 American Indian)). All participants were healthy, with normal or corrected to normal vision and normal hearing, and with no history of psychiatric, physiological or pain disorders and neurological conditions, no current pain symptoms, and no MRI contraindications. Eligibility was assessed with a general health questionnaire, a pain safety screening form, and an MRI safety screening form. Participants were recruited from the Boulder/Denver Metro Area. The institutional review board of the University of Colorado Boulder approved the study, and all participants provided written informed consent.

Method Details

7T fMRI Study (Expt 1)

Experimental Paradigm

Participants completed a probabilistic avoidance learning task during fMRI. In this task (Roy et al., 2014), participants learned to associate different visual cues (circles or triangles) with the aversiveness of stimuli (i.e., mechanical stimulation to the bed of the thumb (N = 31) or unpleasant visual images from the IAPS (Lang et al., 1997) (N = 25). Mechanical stimuli were delivered at non-painful (3 kg/cm2) or painful (5 kg/cm2) pressure levels. Visual stimuli were randomly selected from negative (normative valence = 2.93 ± .67 SD) and neutral (normative valence = 5.52 ± .57 SD) IAPS images. 24 trials of each type were presented in each of five runs (120 trials total). Participants we presented with both visual cues on every trial, followed by a decision phase, a jittered interstimulus-interval, and probabilistic reinforcement based on their decision. Reinforcement rates were predetermined based on a random walk between 20% and 80%, with outcomes determined randomly on every trial.

MRI data acquisition and preprocessing

Gradient-echo echo-planar imaging BOLD-fMRI was performed on a 7 tesla Siemens MRI scanner. Functional images were acquired using GRAPPA-EPI sequence: echo time = 28 ms, repetition time = 2.34 s, flip angle = 75°, number of slices = 123, slice orientation = transversal (axial), anterior to posterior phase encoding, voxel size = 1.1 mm isotropic, gap between slices = 0 mm, field of view = 205 × 205 mm2, GRAPPA acceleration factor = 3; echo spacing = 0.82 ms, bandwidth = 1414 Hz per pixel, partial Fourier in the phase encode direction: 7/8. A custom-built 32-channel radiofrequency coil head array was used for reception. Radiofrequency transmission was provided by a detunable band-pass birdcage coil. Structural images were acquired using a T1-weighted EPI sequence: echo time = 22 ms, repetition time = 8.52 s, flip angle = 90°, number of slices = 126, slice orientation = transversal (axial), voxel size = 1.1 mm isotropic, gap between slices = 0 mm, field of view = 205 × 205 mm2, GRAPPA acceleration factor = 3; echo spacing = 0.82 ms, bandwidth =1414 Hz per pixel, partial Fourier in the phase encode direction: 6/8. This sequence was selected so that functional and structural data would have similar spatial distortions to facilitate coregistration and subsequent normalization of data. Results included in this manuscript come from preprocessing performed using FMRIPREP (Esteban et al., 2019), a Nipype (Gorgolewski et al., 2011) based tool. Spatial normalization to the ICBM 152 Nonlinear Asymmetrical template version 2009c was performed through nonlinear registration with the antsRegistration tool of ANTs v2.1.0 (Avants et al., 2008; Gorgolewski et al., 2011), using brain-extracted versions of both T1w volume and template. Brain tissue segmentation of cerebrospinal fluid (CSF), white matter (WM) and gray matter (GM) was performed on the brain-extracted T1w using fast (Zhang et al., 2001) (FSL v5.0.9). Functional data was slice time corrected using 3dTshift from AFNI v16.2.07 (Cox, 1996) and motion corrected using mcflirt (FSL v5.0.9). This was followed by co-registration to the corresponding T1w using boundary-based registration (Greve and Fischl, 2009) with six degrees of freedom, using flirt (FSL). Motion correcting transformations, BOLD-to-T1w transformation and T1w-to-template (MNI) warp were concatenated and applied in a single step using ants ApplyTransforms (ANTs v2.1.0) using Lanczos interpolation.

3T fMRI Study (Expt. 2)

Experimental Paradigm

Participants received five different types of stimulation (mechanical pain, thermal pain, aversive auditory, aversive visual, and pleasant visual), each at four stimulus intensities. 24 stimuli of each type (6 per intensity level) were presented over six fMRI runs in random order, with different stimulus types intermixed within runs. Following stimulation on each trial, participants made behavioral ratings of their subjective experience. Participants were instructed to answer the question ‘How much do you want to avoid this experience in the future?’. Ratings were made with a visual analog rating scale, with anchors ‘Not at all’ and ‘Most’ displayed at the ends of the scale.

Stimuli

Visual stimulation was administered on the MRI screen and included normed pictures from the International Affective Picture System (Lang et al., 1997), see Table S4. We created four ‘stimulus intensity levels’ by selecting seven images per intensity level in a two-step process: preliminary selection based on normed valence ratings (averaged across male and female raters) from the IAPS database; and final selection based on ratings by 10 lab members (5 male, 5 female) in response to the question “How aversive is this image? 1-100”. The chosen images included photographs of animals (7), bodily illness and injury (12), and industrial and human waste (9). Each picture was presented for 10 sec.

Pressure stimulation was administered using an in-house pressure pain device. This MRI-safe device provides dynamically controlled pressure stimulation using LabView software (National Instruments, Austin, TX). Four pressure levels were applied to an applicator placed on the left thumbnail for 10 sec each (Level 1: 4 kg/cm2, Level 2: 5 kg/cm2, Level 3: 6 kg/cm2; Level 4: 7 kg/cm2).

Thermal stimulation was administered using an ATS Pathway System (Medoc Ltd., Haifa, Israel) with a 16 × 16 mm Peltier-based contact thermode. Four stimulus intensity levels were delivered to the thenar eminence of the left hand (Level 1: 45°C, Level 2: 46°C, Level 3: 47°C, Level 4: 48°C). Each thermal stimulus lasted 10 seconds (1.5 sec ramp-up, 1.5 sec ramp-down, 7 sec at target temperature).

Auditory stimulation was administered using MRI-compatible headphones. We used the sound of a knife scraping on a bottle, which is a reliable aversive auditory stimulus (Kumar et al., 2008). Four stimulus intensity levels were created by scaling the amplitude of a single audio file (Level 1: Lv4 −8 dB, Level 2: Lv4 −4 dB, Level 3: Lv4 −1 dB, Level 4: original amplitude). Auditory stimuli lasted 10 seconds each.

Quantification and Statistical Analysis

MRI data acquisition and preprocessing

Whole-brain fMRI data were acquired on a 3T Siemens MAGNETOM Prisma Fit MRI scanner at the Intermountain Neuroimaging Consortium facility at the University of Colorado, Boulder. Structural images were acquired using high-resolution T1 spoiled gradient recall images (SPGR) for anatomical localization and warping to standard MNI space. Functional images were acquired with a multiband EPI sequence (TR = 460 ms, TE = 27.2 ms, field of view = 220 mm, multiband acceleration factor = 8, flip angle = 44°, 64 × 64 image matrix, 2.7 mm isotropic voxels, 56 interleaved slices, phase encoding posterior > anterior). Six runs of 7.17 mins duration (934 total measurements) were acquired. Stimulus presentation and behavioral data acquisition were controlled using the psychophysics toolbox (Brainard, 1997) for MATLAB (The MathWorks, Inc., Natick, MA).

fMRI data were preprocessed using an automated pipeline implemented by the Mind Research Network, Albuquerque, NM. The preprocessing steps included: distortion correction using FSL’s top-up tool (https://fsl.fmrib.ox.ac.uk/fsl/), realignment (affine alignment of first EPI volume (reference image) to T1, followed by affine alignment of all EPI volumes to the reference image and estimation of the motion parameter file (sepi_vr_motion.1D, AFNI, https://afni.nimh.nih.gov/), spatial normalization of the T1 image (T1 normalization to MNI space (nonlinear transform), normalization of EPI image to MNI space (3dNWarpApply, AFNI, https://afni.nimh.nih.gov/), interpolation to 2 mm isotropic voxels (for better alignment with templates in standard MNI152 space to facilitate prospective testing) and smoothing with a 6 mm FWHM kernel (SPM 8, https://www.fil.ion.ucl.ac.uk/spm/software/spm8/).

Prior to first level analysis, we removed the first four volumes to allow for image intensity stabilization. We also identified image-wise outliers by computing both the mean and the standard deviation (across voxels) of intensity values for each image for all slices to remove intermittent gradient and severe motion-related artifacts.

fMRI data analysis

Data were analyzed using SPM12 (http://www.fil.ion.ucl.ac.uk/spm) and custom MATLAB code available from the authors’ website (http://github.com/canlab/CanlabCore). First-level general linear model (GLM) analyses were conducted in SPM12. The six runs were concatenated for each subject. Boxcar regressors, convolved with the canonical hemodynamic response function, were constructed to model periods for the 10-second stimulation and 4-7 second rating periods. The fixation cross was used as an implicit baseline. A high-pass filter of 0.008 Hz was applied. Nuisance variables included: regressors coding for each run (intercept for each run); linear drift across time within each run; the six estimated head movement parameters (x, y, z, roll, pitch, and yaw), their mean-centered squares, their derivatives, and squared derivative for each run (total 24 columns); and motion outliers (spikes) identified in the previous step. A single-trial model was used to uniquely estimate the response to every stimulus to assess functional connectivity.

Multivariate pathway identification

Functional connectivity between the amygdala, pulvinar, and superior colliculus was estimated using Partial Least Squares (PLS) regression, which identifies latent multivariate patterns that maximize the covariance between two blocks of data (Wold et al., 2001). This approach for estimating the covariance between two putatively connected neural populations, which we call multivariate pathway identification (MPathI, see Fig. S1), is an extension of multivariate methods that have recently been proposed to estimate functional connectivity. These methods identify how patterns of activity in one region relate to patterns in another region, and produce estimates of connectivity based on multivariate dependence (Anzellotti and Coutanche, 2018; Anzellotti et al., 2017; Basti et al., 2020; Coutanche and Thompson-Schill, 2013). We note that MPathI is not an estimation procedure, but a framework for studying connectivity. A family of multivariate methods could be used for estimation, including PLS, canonical correlation, or other multivariate regression approaches. We were motivated to develop a new approach (that utilizes PLS for estimation) because it is uniquely aligned with our goal of identifying functionally connected neural populations, based on the following considerations: 1) it estimates the spatial distribution of underlying neural populations, unlike representational similarity and distance-based methods (Basti et al. 2020; Geerligs et al., 2016) that more flexibly search for statistical dependence; 2) it is constrained to estimate connectivity using robust, reproducible signals by reducing model complexity, which is a problem for related methods such as canonical correlation analysis (Le Floch et al., 2012; Smith et al., 2015); 3) it does not aim to explain all of the variance in each region, as is done when estimating multivariate pattern dependence (Anzellotti et al., 2017), but characterizes signals that covary between regions; and 4) it is a linear estimation procedure, which increases interpretability. These properties allow PLS to constrain the flexibility of connectivity models and allow them to be validated against what is known about neural pathways in animal models.

We used SIMPLS as developed by deJong (1993) to identify latent scores (T and U) and loadings (P and Q) that maximize the covariance between the centered variables X0 and Y0.

  1. Compute the cross-product S = X0’Y0

  2. Compute the singular value decomposition of S

  3. Get weights: r = first left singular vector

  4. Compute X scores: t = X0r

  5. Compute X loadings: p = X0’t/(t’t)

  6. Compute Y loadings: c = Y0’t

  7. Compute Y scores: u = Y0c

    With two additional steps to finding regression coefficients that map activity in one block onto the latent scores of the second block (where A+ is the Moore-Penrose pseudoinverse of matrix A):

  8. Compute Z = X+u

  9. Compute V = Y+t

    These coefficients can be used to make predictions on out-of-sample data, e.g., in cross-validation or prospective tests.

In Expt. 1 preprocessed BOLD time-series data were used as inputs for pathway models, and separate connectivity models were fit for each subject. To assess generalizability across subjects and stimulus types in Expt. 2, PLS models were fit using single-trial estimates of BOLD responses (Gazzaley et al., 2004) to aversive thermal, mechanical, auditory, and visual stimuli, in addition to a set of pleasant visual stimuli which were used as a control. These data were concatenated across subjects (5,760 total trials, 120 per subject). For the pulvinar–amygdala model, the predictor (X) block of variables included all 208 voxels in an anatomically defined mask of the pulvinar and the outcome (Y) block included all 240 voxels in the amygdala (Amunts et al., 2005, see Fig. S2 for depiction of anatomical ROIs). The superior colliculus–pulvinar model included responses in all voxels in a hand-drawn mask of the superior colliculus (41 voxels) as the predictor block and pulvinar responses as the outcome.

This set of regions is particularly well suited to pattern-based estimation of connectivity because each region has a fine-grained structure with unique profiles of connectivity. The superior colliculus is composed of multiple layers, with superficial layers predominantly receiving visual inputs and deeper layers integrating multisensory information and coordinating orienting behavior (Tardif et al., 2005; May, 2006). Both the pulvinar and amygdala are composed of multiple subnuclei, which have distinct profiles of subcortical and cortical connectivity (Arcaro et al., 2018; Barron et al., 2015; Chang and Yu, 2018; Gattass et al., 2017; Guedj and Vuilleumier, 2020; Hrybouski et al., 2016). Pattern-based analysis of fMRI activity can capture information coded in neural substrates exceeding the resolution of single voxels, as it samples population activity distributed across voxels, acting as a complex spatiotemporal filter (Kriegeskorte et al., 2010). Because anatomy varies across individuals, and population activity may be blurred due to hemodynamic filtering, our masks were not set at a conservative threshold (i.e., including any voxels that could plausibly be identified as a specific region), anticipating that MPathI would identify that the key regions within each region that covary with one another.

Each PLS model was specified to identify a pair of latent variables that maximally covary with one another. The patterns of activity in each region that predict latent activity in the other region were estimated using least-norm regression. Inference on these patterns was made using bootstrap resampling (5,000 iterations, treating each subject as a block and randomly sampling subjects with replacement) with normal approximation for inference. Leave-one-subject-out cross-validation was performed to estimate the strength of functional connections (using the Pearson correlation between actual and estimated brain activity as the measure of interest). Inference on cross-validated estimates of functional connectivity were performed using block permutation tests (1,000 iterations) in which the order of trials in the outcome block were scrambled independently for each subject.

Sensitivity analysis

The sensitivity of the brain pathway models was evaluated by comparing the correlation of multivariate model estimates (e.g., correlations of latent activity common to the pulvinar and amygdala) to that of the mean signals in the same regions. We tested for differences in correlation coefficients by converting them to z-scores using the Fisher transform and performing a t-test of differences between pathway models and the mean signal.

Control analyses

To evaluate whether the MPathI approach can classify pathways with known monosynaptic connections from those that do not, we constructed a set of connectivity models between thalamic nuclei and primary sensory cortex. These included one pathway model from the lateral geniculate nucleus (LGN, from the Morel Atlas (Krauth et al., 2010)) to primary visual cortex (V1, from a multimodal parcellation of the cortex (Glasser et al., 2016)) and another from the medial geniculate nucleus (MGN, from the Morel atlas) to primary auditory cortex (A1 from (Glasser et al., 2016)). We also estimated the connectivity of “crossed” connections (i.e., from LGN to V1MGN and from MGN to A1LGN) based on the activity of latent patterns in each region. We used these four estimates to classify established pathways (i.e., LGN-V1 and MGN-V1) from indirect pathways. Inference on the strength of individual pathways was performed using one-sample t-tests on Fisher transformed correlation coefficients.

Mediation analysis

A pathway from the pulvinar to the amygdala is thought to mediate emotional responses by rapidly conveying information about the environment from the superior colliculus. To test this hypothesis, we performed a series of mediation analyses identifying three statistical paths to characterize the effects of the activity of different brain pathways on normative aversiveness: 1) path a characterizes the effect of the SCpulv pathway on PulvAmy activity; 2) path b reflects the relationship between PulvAmy activity and normative ratings of aversiveness, and 3) path ab reflects pulvinar activity formally mediating the link between superior colliculus activity and normative aversiveness, reducing the strength of the direct path c ’ between these variables. We performed mediation analyses to assess relationships between brain activity and normative differences in aversiveness because they are more closely linked to the function of the colliculus-pulvinar-amygdala pathway, i.e., to rapidly evaluate threatening sensory cues. Compared to trial-by-trial self-reports, normative ratings are more directly linked to the nature of stimuli, less influenced by decision variables, including social and contextual factors, and are less influenced by individual differences in introspection and self-report, which are likely mediated by other brain systems including insular cortices and the default network (Chang et al., 2013; Gu et al., 2013; Kleckner et al., 2017). We tested the generalizability of this pathway using mediation analysis both for negative images and sounds and assessed specificity by performing the same mediation analysis on responses to painful stimuli (thermal and mechanical) and positive images. We additionally evaluated whether this pathway is sensitive to moment-to-moment differences in affective experience, as opposed to the normative intensity of negative valence, by performing the same analysis with on-line ratings.

Cortical and thalamic pathway estimation

To evaluate whether activity in the colliculus-pulvinar-amygdala pathway is distinct from cortical activity during emotion processing, we estimated pathway models between the amygdala and parcels of the Glasser atlas combined across hemispheres (Glasser et al., 2016) and anatomically defined regions of the thalamus (Krauth et al., 2010). This produced a series of beta estimates for 197 regions. These estimates were concatenated into a single 5,760 (trial) by 197 (region) matrix that was subjected to principal component analysis (PCA). Bootstrap resampling (5,000 samples, with block resampling of subjects to keep all images from a subject together) was performed to estimate the standard errors of coefficients from PCA.

The activity of each component was estimated by multiplying the PCA coefficients and activation estimates for all pathways. The association between each component and subjective aversiveness (and pleasantness) ratings was estimated using the procedures described in the sensitivity analysis section.

Supplementary Material

1
Table S3.

Supplemental Material Titles:

Mean and Bootstrap Standard Error of Coefficients from Principal Component Analysis of Brain Pathway Activation, Related to Figure 3.

Key Resources Table

REAGENT or RESOURCE SOURCE IDENTIFIER
Deposited Data
Brain models and maps This Paper https://identifiers.org/neurovault.collection:9982
Processed fMRI data This Paper https://osf.io/werk2
Software and Algorithms
SPM8 Wellcome Trust Centre for Neuroimaging https://www.fil.ion.ucl.ac.uk/spm/software/spm8/
SPM12 Wellcome Trust Centre for Neuroimaging https://www.fil.ion.ucl.ac.uk/spm/software/spm12/
fMRIPrep Poldrack Lab https://www.fmriprep.org/
CANLab Core Tools Wager Lab https://github.com/canlab/CanlabCore/
Open in a new tab

Highlights:

  • Multivariate fMRI models identify the colliculus-pulvinar-amygdala pathway in humans

  • The pathway responds to negative visual and auditory, but not painful stimuli

  • The pathway is associated with broad patterns of cortico-amygdalar connectivity

Acknowledgments:

We thank Daniel Ott for assistance with data collection and Mickela Heilicher for literature review and thoughtful discussions about this project.

Funding:

This work was supported by the National Institutes of Health grants R01 MH076136, DA046064, MH116026, and EB026549 and National Cancer Institute grant U01 CA193632.

Footnotes

Publisher's Disclaimer: This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.

Declaration of Interests: None of the authors of this manuscript have a financial interest related to this work.

References:

  1. Abivardi A, and Bach DR (2017). Deconstructing white matter connectivity of human amygdala nuclei with thalamus and cortex subdivisions in vivo. Hum. Brain Mapp 38, 3927–3940. [DOI] [PMC free article] [PubMed] [Google Scholar]
  2. Amunts K, Kedo O, Kindler M, Pieperhoff P, Mohlberg H, Shah NJ, Habel U, Schneider F, and Zilles K (2005). Cytoarchitectonic mapping of the human amygdala, hippocampal region and entorhinal cortex: intersubject variability and probability maps. Anat. Embryol 210, 343–352. [DOI] [PubMed] [Google Scholar]
  3. Anderson AK, and Phelps EA (2002). Is the human amygdala critical for the subjective experience of emotion? Evidence of intact dispositional affect in patients with amygdala lesions. J. Cogn. Neurosci 14, 709–720. [DOI] [PubMed] [Google Scholar]
  4. Anzellotti S, and Coutanche MN (2018). Beyond Functional Connectivity: Investigating Networks of Multivariate Representations. Trends Cogn. Sci 22, 258–269. [DOI] [PubMed] [Google Scholar]
  5. Anzellotti S, Caramazza A, and Saxe R (2017). Multivariate pattern dependence. PLoS Comput. Biol 13, e1005799. [DOI] [PMC free article] [PubMed] [Google Scholar]
  6. Arcaro MJ, Pinsk MA, Chen J, and Kastner S (2018). Organizing principles of pulvino-cortical functional coupling in humans. Nat. Commun 9, 5382. [DOI] [PMC free article] [PubMed] [Google Scholar]
  7. Atlas LY, Atlas LY, Bolger N, Lindquist MA, and Wager TD (2010). Brain Mediators of Predictive Cue Effects on Perceived Pain. Journal of Neuroscience 30, 12964–12977. [DOI] [PMC free article] [PubMed] [Google Scholar]
  8. Avants BB, Epstein CL, Grossman M, and Gee JC (2008). Symmetric diffeomorphic image registration with cross-correlation: evaluating automated labeling of elderly and neurodegenerative brain. Med. Image Anal 12, 26–41. [DOI] [PMC free article] [PubMed] [Google Scholar]
  9. Barrett LF (2017). The theory of constructed emotion: an active inference account of interoception and categorization. Soc. Cogn. Affect. Neurosci 12, 1833. [DOI] [PMC free article] [PubMed] [Google Scholar]
  10. Barron DS, Eickhoff SB, Clos M, and Fox PT (2015). Human pulvinar functional organization and connectivity. Hum. Brain Mapp 36, 2417–2431. [DOI] [PMC free article] [PubMed] [Google Scholar]
  11. Basti A, Nili H, Hauk O, Marzetti L, and Henson RN (2020). Multi-dimensional connectivity: a conceptual and mathematical review. NeuroImage 221, 117179. [DOI] [PubMed] [Google Scholar]
  12. Basti A, Mur M, Kriegeskorte N, Pizzella V, Marzetti L, and Hauk O Analysing linear multivariate pattern transformations in neuroimaging data. [DOI] [PMC free article] [PubMed] [Google Scholar]
  13. Brainard DH (1997). The Psychophysics Toolbox. Spat. Vis 10, 433–436. [PubMed] [Google Scholar]
  14. Burgos-Robles A, Kimchi EY, Izadmehr EM, Porzenheim MJ, Ramos-Guasp WA, Nieh EH, Felix-Ortiz AC, Namburi P, Leppla CA, Presbrey KN, et al. (2017). Amygdala inputs to prefrontal cortex guide behavior amid conflicting cues of reward and punishment. Nat. Neurosci 20, 824–835. [DOI] [PMC free article] [PubMed] [Google Scholar]
  15. Chang J, and Yu R (2018). Alternations in functional connectivity of amygdalar subregions under acute social stress. Neurobiol Stress 9, 264–270. [DOI] [PMC free article] [PubMed] [Google Scholar]
  16. Chang LJ, Yarkoni T, Khaw MW, and Sanfey AG (2013). Decoding the role of the insula in human cognition: functional parcellation and large-scale reverse inference. Cereb. Cortex 23, 739–749. [DOI] [PMC free article] [PubMed] [Google Scholar]
  17. Corder G, Ahanonu B, Grewe BF, Wang D, Schnitzer MJ, and Scherrer G (2019). An amygdalar neural ensemble that encodes the unpleasantness of pain. Science 363, 276–281. [DOI] [PMC free article] [PubMed] [Google Scholar]
  18. Coutanche MN, and Thompson-Schill SL (2013). Informational connectivity: identifying synchronized discriminability of multi-voxel patterns across the brain. Front. Hum. Neurosci 7, 15. [DOI] [PMC free article] [PubMed] [Google Scholar]
  19. Cox RW (1996). AFNI: software for analysis and visualization of functional magnetic resonance neuroimages. Comput. Biomed. Res 29, 162–173. [DOI] [PubMed] [Google Scholar]
  20. Dehaene S, Lau H, and Kouider S (2017). What is consciousness, and could machines have it? Science 358, 486–492. [DOI] [PubMed] [Google Scholar]
  21. Deisseroth K (2011). Optogenetics. Nat. Methods 8, 26–29. [DOI] [PMC free article] [PubMed] [Google Scholar]
  22. deJong. (1993). SIMPLS: An alternative approach to partial least squares regression. Chemometrics Intellig. Lab. Syst 18, 251–263. [Google Scholar]
  23. Doron NN, and Ledoux JE (1999). Organization of projections to the lateral amygdala from auditory and visual areas of the thalamus in the rat. J. Comp. Neurol 412, 383–409. [PubMed] [Google Scholar]
  24. Esteban O, Markiewicz CJ, Blair RW, Moodie CA, Isik AI, Erramuzpe A, Kent JD, Goncalves M, DuPre E, Snyder M, et al. (2019). fMRIPrep: a robust preprocessing pipeline for functional MRI. Nat. Methods 16, 111–116. [DOI] [PMC free article] [PubMed] [Google Scholar]
  25. Feinstein JS, Buzza C, Hurlemann R, Follmer RL, Dahdaleh NS, Coryell WH, Welsh MJ, Tranel D, and Wemmie JA (2013). Fear and panic in humans with bilateral amygdala damage. Nat. Neurosci 16, 270–272. [DOI] [PMC free article] [PubMed] [Google Scholar]
  26. Friston KJ (1994). Functional and effective connectivity in neuroimaging: A synthesis. Hum. Brain Mapp 2, 56–78. [Google Scholar]
  27. Gattass R, Soares JGM, and Lima B (2017). The Pulvinar Thalamic Nucleus of Non-Human Primates: Architectonic and Functional Subdivisions (Springer; ). [Google Scholar]
  28. Gazzaley A, Rissman J, and D’esposito M (2004). Functional connectivity during working memory maintenance. Cognitive, Affective, & Behavioral Neuroscience 4, 580–599. [DOI] [PubMed] [Google Scholar]
  29. Geerligs L, Cam-Can, and Henson RN (2016). Functional connectivity and structural covariance between regions of interest can be measured more accurately using multivariate distance correlation. Neuroimage 135, 16–31. [DOI] [PMC free article] [PubMed] [Google Scholar]
  30. de Gelder B, van Honk J, and Tamietto M (2011). Emotion in the brain: of low roads, high roads and roads less travelled. Nat. Rev. Neurosci 12, 425; author reply 425. [DOI] [PubMed] [Google Scholar]
  31. Glasser MF, Coalson TS, Robinson EC, Hacker CD, Harwell J, Yacoub E, Ugurbil K, Andersson J, Beckmann CF, Jenkinson M, et al. (2016). A multi-modal parcellation of human cerebral cortex. Nature 536, 171–178. [DOI] [PMC free article] [PubMed] [Google Scholar]
  32. Gorgolewski K, Burns CD, Madison C, Clark D, Halchenko YO, Waskom ML, and Ghosh SS (2011). Nipype: a flexible, lightweight and extensible neuroimaging data processing framework in python. Front. Neuroinform. 5, 13. [DOI] [PMC free article] [PubMed] [Google Scholar]
  33. Gothard KM (2020). Multidimensional processing in the amygdala. Nat. Rev. Neurosci [DOI] [PMC free article] [PubMed] [Google Scholar]
  34. Greve DN, and Fischl B (2009). Accurate and robust brain image alignment using boundary-based registration. Neuroimage 48, 63–72. [DOI] [PMC free article] [PubMed] [Google Scholar]
  35. Gründemann J, Bitterman Y, Lu T, Krabbe S, Grewe BF, Schnitzer MJ, and Lüthi A (2019). Amygdala ensembles encode behavioral states. Science 364. [DOI] [PubMed] [Google Scholar]
  36. Gu X, Hof PR, Friston KJ, and Fan J (2013). Anterior insular cortex and emotional awareness. J. Comp. Neurol 521, 3371–3388. [DOI] [PMC free article] [PubMed] [Google Scholar]
  37. Guedj C, and Vuilleumier P (2020). Functional connectivity fingerprints of the human pulvinar: Decoding its role in cognition. Neuroimage 221, 117162. [DOI] [PubMed] [Google Scholar]
  38. Han S, Soleiman MT, Soden ME, Zweifel LS, and Palmiter RD (2015). Elucidating an Affective Pain Circuit that Creates a Threat Memory. Cell 162, 363–374. [DOI] [PMC free article] [PubMed] [Google Scholar]
  39. Hardoon DR, Mourão-Miranda J, Brammer M, and Shawe-Taylor J (2007). Unsupervised analysis of fMRI data using kernel canonical correlation. Neuroimage 37, 1250–1259. [DOI] [PubMed] [Google Scholar]
  40. Haxby JV, Gobbini MI, Furey ML, Ishai A, Schouten JL, and Pietrini P (2001). Distributed and overlapping representations of faces and objects in ventral temporal cortex. Science 293, 2425–2430. [DOI] [PubMed] [Google Scholar]
  41. Haynes J-D, and Rees G (2005). Predicting the orientation of invisible stimuli from activity in human primary visual cortex. Nat. Neurosci 8, 686–691. [DOI] [PubMed] [Google Scholar]
  42. Hrybouski S, Aghamohammadi-Sereshki A, Madan CR, Shafer AT, Baron CA, Seres P, Beaulieu C, Olsen F, and Malykhin NV (2016). Amygdala subnuclei response and connectivity during emotional processing. Neuroimage 133, 98–110. [DOI] [PubMed] [Google Scholar]
  43. Janak PH, and Tye KM (2015). From circuits to behaviour in the amygdala. Nature 517, 284–292. [DOI] [PMC free article] [PubMed] [Google Scholar]
  44. Kamitani Y, and Tong F (2005). Decoding the visual and subjective contents of the human brain. Nat. Neurosci 8, 679–685. [DOI] [PMC free article] [PubMed] [Google Scholar]
  45. Keuken MC, Bazin PL, Crown L, Hootsmans J, Laufer A, Müller-Axt C, Sier R, van der Putten EJ, Schäfer A, Turner R, & Forstmann BU (2014). Quantifying inter-individual anatomical variability in the subcortex using 7T structural MRI. NeuroImage, 94, 40–46. [DOI] [PubMed] [Google Scholar]
  46. Kleckner IR, Zhang J, Touroutoglou A, Chanes L, Xia C, Kyle Simmons W, Quigley KS, Dickerson BC, and Barrett LF (2017). Evidence for a large-scale brain system supporting allostasis and interoception in humans. Nature Human Behaviour 1. [DOI] [PMC free article] [PubMed] [Google Scholar]
  47. Kragel PA, Koban L, Barrett LF, and Wager TD (2018). Representation, Pattern Information, and Brain Signatures: From Neurons to Neuroimaging. Neuron 99, 257–273. [DOI] [PMC free article] [PubMed] [Google Scholar]
  48. Krauth A, Blanc R, Poveda A, Jeanmonod D, Morel A, and Székely G (2010). A mean three-dimensional atlas of the human thalamus: generation from multiple histological data. Neuroimage 49, 2053–2062. [DOI] [PubMed] [Google Scholar]
  49. Kriegeskorte N, Goebel R, and Bandettini P (2006). Information-based functional brain mapping. Proc. Natl. Acad. Sci. U. S. A 103, 3863–3868. [DOI] [PMC free article] [PubMed] [Google Scholar]
  50. Kriegeskorte N, Cusack R, and Bandettini P (2010). How does an fMRI voxel sample the neuronal activity pattern: compact-kernel or complex spatiotemporal filter? Neuroimage 49, 1965–1976. [DOI] [PMC free article] [PubMed] [Google Scholar]
  51. Kumar S, Forster HM, Bailey P, and Griffiths TD (2008). Mapping unpleasantness of sounds to their auditory representation. J. Acoust. Soc. Am 124, 3810–3817. [DOI] [PubMed] [Google Scholar]
  52. Kyriazi P, Headley DB, and Pare D (2018). Multi-dimensional Coding by Basolateral Amygdala Neurons. Neuron 99, 1315–1328.e5. [DOI] [PMC free article] [PubMed] [Google Scholar]
  53. Lang PJ, Bradley MM, Cuthbert BN, and Others (1997). International affective picture system (IAPS): Technical manual and affective ratings. NIMH Center for the Study of Emotion and Attention 1, 39–58. [Google Scholar]
  54. Ledoux J (1998). The Emotional Brain: The Mysterious Underpinnings of Emotional Life (Simon and Schuster; ). [Google Scholar]
  55. LeDoux JE (2014). Coming to terms with fear. Proc. Natl. Acad. Sci. U. S. A 111, 2871–2878. [DOI] [PMC free article] [PubMed] [Google Scholar]
  56. LeDoux JE, and Brown R (2017). A higher-order theory of emotional consciousness. Proc. Natl. Acad. Sci. U. S. A 114, E2016–E2025. [DOI] [PMC free article] [PubMed] [Google Scholar]
  57. Le Floch E, Guillemot V, Frouin V, Pinel P, Lalanne C, Trinchera L, Tenenhaus A, Moreno A, Zilbovicius M, Bourgeron T, et al. (2012). Significant correlation between a set of genetic polymorphisms and a functional brain network revealed by feature selection and sparse Partial Least Squares. Neuroimage 63, 11–24. [DOI] [PubMed] [Google Scholar]
  58. Liddell BJ, Brown KJ, Kemp AH, Barton MJ, Das P, Peduto A, Gordon E, and Williams LM (2005). A direct brainstem–amygdala–cortical “alarm” system for subliminal signals of fear. NeuroImage 24, 235–243. [DOI] [PubMed] [Google Scholar]
  59. May PJ (2006). The mammalian superior colliculus: laminar structure and connections. Prog. Brain Res. 151, 321–378. [DOI] [PubMed] [Google Scholar]
  60. McFadyen J, Mattingley JB, & Garrido MI (2019). An afferent white matter pathway from the pulvinar to the amygdala facilitates fear recognition. elife, 8, e40766. [DOI] [PMC free article] [PubMed] [Google Scholar]
  61. Morris JS, Öhman A, and Dolan RJ (1999). A subcortical pathway to the right amygdala mediating “unseen” fear. Proc. Natl. Acad. Sci. U. S. A 96, 1680–1685. [DOI] [PMC free article] [PubMed] [Google Scholar]
  62. Norman KA, Polyn SM, Detre GJ, and Haxby JV (2006). Beyond mind-reading: multi-voxel pattern analysis of fMRI data. Trends Cogn. Sci 10, 424–430. [DOI] [PubMed] [Google Scholar]
  63. Olson IR, Plotzker A, and Ezzyat Y (2007). The Enigmatic temporal pole: a review of findings on social and emotional processing. Brain 130, 1718–1731. [DOI] [PubMed] [Google Scholar]
  64. Palmiter RD (2018). The Parabrachial Nucleus: CGRP Neurons Function as a General Alarm. Trends Neurosci. 41, 280–293. [DOI] [PMC free article] [PubMed] [Google Scholar]
  65. Pessoa L, and Adolphs R (2010). Emotion processing and the amygdala: from a “low road” to “many roads” of evaluating biological significance. Nat. Rev. Neurosci 11, 773–783. [DOI] [PMC free article] [PubMed] [Google Scholar]
  66. Poldrack RA, and Farah MJ (2015). Progress and challenges in probing the human brain. Nature 526, 371–379. [DOI] [PubMed] [Google Scholar]
  67. Rafal RD, Koller K, Bultitude JH, Mullins P, Ward R, Mitchell AS, and Bell AH (2015). Connectivity between the superior colliculus and the amygdala in humans and macaque monkeys: virtual dissection with probabilistic DTI tractography. J. Neurophysiol 114, 1947–1962. [DOI] [PMC free article] [PubMed] [Google Scholar]
  68. Ramcharan EJ, Gnadt JW, and Sherman SM (2005). Higher-order thalamic relays burst more than first-order relays. Proc. Natl. Acad. Sci. U. S. A 102, 12236–12241. [DOI] [PMC free article] [PubMed] [Google Scholar]
  69. Roy M, Shohamy D, Daw N, Jepma M, Wimmer GE, and Wager TD (2014). Representation of aversive prediction errors in the human periaqueductal gray. Nat. Neurosci 17, 1607–1612. [DOI] [PMC free article] [PubMed] [Google Scholar]
  70. Schneider GE (1969). Two visual systems. Science 163, 895–902. [DOI] [PubMed] [Google Scholar]
  71. Shenhav A, Barrett LF, and Bar M (2013). Affective value and associative processing share a cortical substrate. Cogn. Affect. Behav. Neurosci 13, 46–59. [DOI] [PMC free article] [PubMed] [Google Scholar]
  72. Smith SM, Nichols TE, Vidaurre D, Winkler AM, Behrens TEJ, Glasser MF, Ugurbil K, Barch DM, Van Essen DC, and Miller KL (2015). A positive-negative mode of population covariation links brain connectivity, demographics and behavior. Nat. Neurosci 18, 1565–1567. [DOI] [PMC free article] [PubMed] [Google Scholar]
  73. Stringer C, Pachitariu M, Steinmetz N, Reddy CB, Carandini M, and Harris KD (2019). Spontaneous behaviors drive multidimensional, brainwide activity. Science 364, 255. [DOI] [PMC free article] [PubMed] [Google Scholar]
  74. Tamietto M, and de Gelder B (2010). Neural bases of the non-conscious perception of emotional signals. Nat. Rev. Neurosci 11, 697–709. [DOI] [PubMed] [Google Scholar]
  75. Tamietto M, Pullens P, de Gelder B, Weiskrantz L, and Goebel R (2012). Subcortical connections to human amygdala and changes following destruction of the visual cortex. Curr. Biol 22, 1449–1455. [DOI] [PubMed] [Google Scholar]
  76. Tardif E, Delacuisine B, Probst A, and Clarke S (2005). Intrinsic connectivity of human superior colliculus. Exp. Brain Res 166, 316–324. [DOI] [PubMed] [Google Scholar]
  77. Tye KM, Stuber GD, de Ridder B, Bonci A, and Janak PH (2008). Rapid strengthening of thalamo-amygdala synapses mediates cue–reward learning. Nature 453, 1253–1257. [DOI] [PMC free article] [PubMed] [Google Scholar]
  78. Tye KM, Prakash R, Kim S-Y, Fenno LE, Grosenick L, Zarabi H, Thompson KR, Gradinaru V, Ramakrishnan C, and Deisseroth K (2011). Amygdala circuitry mediating reversible and bidirectional control of anxiety. Nature 471, 358–362. [DOI] [PMC free article] [PubMed] [Google Scholar]
  79. Vuilleumier P, Armony JL, Driver J, and Dolan RJ (2003). Distinct spatial frequency sensitivities for processing faces and emotional expressions. Nat. Neurosci 6, 624–631. [DOI] [PubMed] [Google Scholar]
  80. Wang YC, Bianciardi M, Chanes L, & Satpute AB (2020). Ultra high field fMRI of human superior colliculi activity during affective visual processing. Scientific reports, 10(1), 1–7. [DOI] [PMC free article] [PubMed] [Google Scholar]
  81. Wei P, Liu N, Zhang Z, Liu X, Tang Y, He X, Wu B, Zhou Z, Liu Y, Li J, et al. (2015). Processing of visually evoked innate fear by a non-canonical thalamic pathway. Nat. Commun 6, 6756. [DOI] [PMC free article] [PubMed] [Google Scholar]
  82. Whalen PJ, Kagan J, Cook RG, Davis FC, Kim H, Polis S, McLaren DG, Somerville LH, McLean AA, Maxwell JS, et al. (2004). Human amygdala responsivity to masked fearful eye whites. Science 306, 2061. [DOI] [PubMed] [Google Scholar]
  83. Wold S, SjÖstrÖm M, and Eriksson L (2001). PLS-regression: a basic tool of chemometrics. Chemometrics Intellig. Lab. Syst 58, 109–130. [Google Scholar]
  84. Woo C-W, Koban L, Kross E, Lindquist MA, Banich MT, Ruzic L, Andrews-Hanna JR, and Wager TD (2014). Separate neural representations for physical pain and social rejection. Nat. Commun 5, 5380. [DOI] [PMC free article] [PubMed] [Google Scholar]
  85. Woo C-W, Chang LJ, Lindquist MA, and Wager TD (2017). Building better biomarkers: brain models in translational neuroimaging. Nat. Neurosci 20, 365–377. [DOI] [PMC free article] [PubMed] [Google Scholar]
  86. Yeo BTT, Krienen FM, Sepulcre J, Sabuncu MR, Lashkari D, Hollinshead M, Roffman JL, Smoller JW, Zöllei L, Polimeni JR, et al. (2011). The organization of the human cerebral cortex estimated by intrinsic functional connectivity. J. Neurophysiol 106, 1125–1165. [DOI] [PMC free article] [PubMed] [Google Scholar]
  87. Zhang Y, Brady M, and Smith S (2001). Segmentation of brain MR images through a hidden Markov random field model and the expectation-maximization algorithm. IEEE Trans. Med. Imaging 20, 45–57. [DOI] [PubMed] [Google Scholar]
  88. Zhou N, Masterson SP, Damron JK, Guido W, and Bickford ME (2018). The Mouse Pulvinar Nucleus Links the Lateral Extrastriate Cortex, Striatum, and Amygdala. J. Neurosci. 38, 347–362. [DOI] [PMC free article] [PubMed] [Google Scholar]
  89. Zhou NA, Maire PS, Masterson SP, and Bickford ME (2017). The mouse pulvinar nucleus: Organization of the tectorecipient zones. Vis. Neurosci. 34, E011. [DOI] [PMC free article] [PubMed] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

1
NIHMS1717997-supplement-1.pdf (383KB, pdf)
Table S3.

Supplemental Material Titles:

Mean and Bootstrap Standard Error of Coefficients from Principal Component Analysis of Brain Pathway Activation, Related to Figure 3.

NIHMS1717997-supplement-2.xlsx (19.4KB, xlsx)

Data Availability Statement

fMRI data are available at https://osf.io/werk2/. MATLAB code for analyses is available at: https://github.com/canlab.

RESOURCES