Orienting to fear under transient focal disruption of the human amygdala

Abstract Responding to threat is under strong survival pressure, promoting the evolution of systems highly optimized for the task. Though the amygdala is implicated in ‘detecting’ threat, its role in the action that immediately follows—‘orienting’—remains unclear. Critical to mounting a targeted response, such early action requires speed, accuracy, and resilience optimally achieved through conserved, parsimonious, dedicated systems, insured against neural loss by a parallelized functional organization. These characteristics tend to conceal the underlying substrate not only from correlative methods but also from focal disruption over time scales long enough for compensatory adaptation to take place. In a study of six patients with intracranial electrodes temporarily implanted for the clinical evaluation of focal epilepsy, we investigated gaze orienting to fear during focal, transient, unilateral direct electrical disruption of the amygdala. We showed that the amygdala is necessary for rapid gaze shifts towards faces presented in the contralateral hemifield regardless of their emotional expression, establishing its functional lateralization. Behaviourally dissociating the location of presented fear from the direction of the response, we implicated the amygdala not only in detecting contralateral faces, but also in automatically orienting specifically towards fearful ones. This salience-specific role was demonstrated within a drift-diffusion model of action to manifest as an orientation bias towards the location of potential threat. Pixel-wise analysis of target facial morphology revealed scleral exposure as its primary driver, and induced gamma oscillations—obtained from intracranial local field potentials—as its time-locked electrophysiological correlate. The amygdala is here reconceptualized as a functionally lateralized instrument of early action, reconciling previous conflicting accounts confined to detection, and revealing a neural organisation analogous to the superior colliculus, with which it is phylogenetically kin. Greater clarity on its role has the potential to guide therapeutic resection, still frequently complicated by impairments of cognition and behaviour related to threat, and inform novel focal stimulation techniques for the management of neuropsychiatric conditions.


S2 Supplementary
We compiled a set of 89 pairs of images, each pair consisting of a fearful and neutral facial expression posed by the same actor. Images were taken from three databases: the Karolinska Directed Emotional Faces (http://www.emotionlab. se/resources/kdef) 1

, the Warsaw Set of Emotional Facial Expression Pictures
(http://www.emotional-face.org) 2 , and the Radboud Faces Database (http://www.socsci. ru.nl:8180/RaFD2/RaFD) 3 . Each image was processed to minimise variation in spatial and contrast characteristics irrelevant to the expressed emotion. Contrast was linearly normalised in the full 8-bit greyscale range. To align the images, Face++ facial landmark algorithm (https://www.faceplusplus.com) was used to identify 83 facial landmarks per image, from which a rigid transform was computed using coherent point drift registration implemented in Matlab (CPD v2.1 4 ). The landmarks were also used to crop each face to the hair and jaw lines, excluding ears, and present it on a black background.
For the pixel-wise analyses, the same images were non-rigidly aligned (to improve pixel-level registration) using a regularised non-linear transform also implemented in CPD. Images were smoothed with a 4 pixel FWHM Gaussian kernel prior to analysis in SPM.

Data acquisition
Pre-operatively, whole-brain T1-weighted magnetic resonance imaging of 0.94x0.94x1.1mm resolution was acquired on a 3T General Electric Excite HDx scanner (General Electric, Milwaukee, WI, USA) using an eightchannel array head coil for reception and the body coil for transmission, with standard imaging gradients (maximum strength 40 mT/m and slew rate 150 T/m/s). Post-electrode implantation, the participants underwent an uncontrasted whole-head CT scan of resolution 0.43x0.43x1.2mm. (SOMATOM Definition 128slice, Siemens Healthcare GmbH, Erlangen, Federal Republic of Germany).

Analysis
We sought to precisely determine the location of the electrode contacts and therefore the electrical disruption sites within the amygdala of each patient in standard stereotactic space. This required two forms of image registration: a native space, within-subject registration of the post-implantation CT to the pre-implantation MRI, and a MNI space, template registration of the MRI, applying the derived parameters so as to secondarily warp the natively registered CT into the same space. We used a previously developed and validated non-linear CT-MRI registration algorithm developed to optimally account for electrode artefact 5 . The procedure is described briefly below.
CT PRE-PROCESSING: a rigid body coregistration to the standard SPM12 tissue probability map was performed based on normalised mutual information with adjustment from a Procrustes analysis weighted by the white and grey matter compartments. This placed the scan in rigid register with the MNI template space.

S4
So as to focus subsequent processing on tissue-relevant contrast, an identical copy of the scan was windowed so as to zero all voxels outwith 0-100 Hounsfield units. This was then filtered with an Oracle-based 3D discrete cosine transform filter 6 to enhance tissue contrast. All subsequent operations were performed on this image, and the final transformation was replicated on the original image at the end. MR PRE-PROCESSING: As for the CT, a rigid body coregistration to the standard SPM12 tissue probability map was performed based on normalised mutual information with adjustment from a Procrustes analysis weighted by the white and grey matter compartments. This placed the scan in rigid register with the MNI template space, and also with the CT. The scan was then resliced using 4th degree b-spline interpolation to procedure, with default parameters, was used to generate segmented compartments in native space for each of the standard 6 tissue classes, as well as a set of parameters for non-linear transformation into MNI space of this and any other image in register with it.

NON-LINEAR REGISTRATION OF CT TO MR
normalisation procedure on the windowed, filtered CT scan, but instead of using the standard MNI space template tissue compartments as tissue prior probability maps we used the individual MR-derived tissue compartments in native space. The resultant transformation is therefore not into standard stereotactic space but rather the native space of the T1 (which the CT already shares), adjusted to introduce some conformity with the native MR tissue segmentation. Other than removing the affine registration step and any bias routine involves explicitly modelling anomalous signal, this adjustment was robust to the artefact created by the metal contacts. The deformation field describing the transformation was applied both to the CT and its corresponding landmark image.

NORMALISATION (TRANSFORMATION INTO MNI SPACE):
The deformation field estimated from the MR scan segmentation and normalisation was applied to the grey and white matter compartments of the MR image, transforming them into in standard MNI space. The same deformation field was used to transform the final MR-registered CT into the same space. All transformed images were resliced to 1mm isotropic voxels. The output of MR-derived grey and white matter images, and CT-derived electrode locations were consequently visualised in ParaView (http://www.paraview.org). The location of the amygdala, derived from a 0.1 threshold of the probabilistic maps of amygdala subnuclei derived from the SPM Anatomy Toolbox 7,8 was also superimposed.   Choice and a joint model of both. For each of the three models, the parameters of the response distribution (D) are given. Response latencies were modelled as a shifted lognormal distribution whose parameters reflect the mean ( ) and standard deviation (σ) of the log-transformed latency, relative to an additional shift parameter (δ, the time of the earliest possible response). The binary choice of orientation target, was modelled with a Bernoulli response distribution with a rate parameter θ representing the degree of Fearful face preference. The drift-diffusion model has four critical parameters: a (the decision threshold) the distance between response boundaries representing response caution; z (the bias) the competition starting point which determines if there S7 is a prior preference towards one of the outcomes; v (the drift-rate) which determines the rate of evidence accumulation towards the outcomes (e.g. easier choices accumulate evidence to one outcome more rapidly) and t (the non-decision time). In every model, each parameter is modelled either by an intercept , or a linear predictor vector j of length J multiplied by J predictors generated from the factorial interaction coding of the population-level factors of interest listed (X). Factorial interactions are specified by a colon (e.g. INSTRUCTION:DISRUPTION specifies the interaction between these two terms) Additionally, all parameters are modelled by vector η s of length S multiplied by S subject-specific dummy variables (Z) to account for betweensubject variance. Link functions are specified if they were needed to bound parameters to positive numbers (Log) or values between 0 and 1 (Logit). DISRUPTION is modelled as a factor with three levels (None, Ipsilateral to fearful face, Contralateral to fearful face) except where highlighted with a star (*DISRUPTION), in which case it is modelled as a factor with two levels (Present, None). INSTRUCTION has the levels Yield and Oppose. EMOTION has the levels Fearful face and Neutral face. The HEMIFIELD in which the fearful face was presented has the levels Left and Right. Prior distributions are represented as follows: a normal distribution is represented as normal(mean, standard deviation), a student t distribution is specified as student_t(degrees of freedom, mean, standard deviation) and a positive only student t distribution is specified as halfstudent_t(degrees of freedom, mean, standard deviation).

Robust Prior Selection
We followed a consistent conservative approach employing weakly informative prior distributions over model parameters as described in Supplementary Table 2 9 implausible and are liable to cause degeneracies in non-linear models. Student t distributions were preferred over Normal distributions were possible owing to their improved behaviour with the heavy tailed data characteristic of behaviour 10 . Priors over continuous unbounded population-level parameters of interest were generally modelled with a Student t distribution (df=5, mean=0, sd=5). Subject-level intercepts were modelled hierarchically with normal priors whose mean and standard deviation were learned from the data (partial pooling) by placing Student t (df=3, mean=0, sd=2.5) and Half-Student t (df=3, mean=0, sd=2.5) priors over the Normal mean and standard deviations respectively. The shift parameter in the latency model and nondecision time in the drift-diffusion model are both linked to, and cannot be higher than, the minimum response latency. We therefore applied tighter constraints to these parameters with a normal prior (mean=-3, sd=1).
When transformed this results in a prior predictive distribution with 95% credibility interval between 7-350ms, which covers the reasonable range of values for these parameters.

Posterior Estimation
Calculating the Bayesian posterior for such models can be analytically intractable and therefore posterior parameter estimates were generated using a Markov Chain Monte Carlo (MCMC       )). Taken together, these results suggest that the orientation towards Dark crosses may be errors (early and less frequent). Although this analysis has reduced power as compared to the Faces task due to fewer participants and greater imbalance between salient and nonsalient images (salient images were strongly preferred), there was no similarity between the major findings in the Faces task and the low-level Crosses tasks. As such, it is likely that mechanistic inferences derived from the Faces task are face-specific, and cannot be generalised to detection of other salient objects. CI = Credibility Interval.        x LATENCY revealed that scleral information was associated with orientation latency was associated with orientation latency as a function of disruption (see Supplementary Table 11). The F-statistic map for this contrast is displayed above overlaid onto a mean face. Presented F-statistics are thresholded at p<0.001 uncorrected for visualisation. Post hoc t-tests within each disruption condition are presented in Figure 4.
Foveated minus neglected face in the Yield condition