1. Neuroscience
Download icon

Social aversive generalization learning sharpens the tuning of visuocortical neurons to facial identity cues

  1. Yannik Stegmann  Is a corresponding author
  2. Lea Ahrens
  3. Paul Pauli
  4. Andreas Keil
  5. Matthias J Wieser
  1. Department of Psychology (Biological Psychology, Clinical Psychology, and Psychotherapy), University of Würzburg, Germany
  2. Center for Mental Health, Medical Faculty, University of Würzburg, Germany
  3. Center for the Study of Emotion and Attention, University of Florida, United States
  4. Department of Psychology, Education, and Child Studies, Erasmus University Rotterdam, Netherlands
Research Article
  • Cited 0
  • Views 543
  • Annotations
Cite this article as: eLife 2020;9:e55204 doi: 10.7554/eLife.55204

Abstract

Defensive system activation promotes heightened perception of threat signals, and excessive attention to threat signals has been discussed as a contributory factor in the etiology of anxiety disorders. However, a mechanistic account of attentional modulation during fear-relevant processes, especially during fear generalization remains elusive. To test the hypothesis that social fear generalization prompts sharpened tuning in the visuocortical representation of social threat cues, 67 healthy participants underwent differential fear conditioning, followed by a generalization test in which participants viewed faces varying in similarity with the threat-associated face. We found that generalization of social threat sharpens visuocortical tuning of social threat cues, whereas ratings of fearfulness showed generalization, linearly decreasing with decreasing similarity to the threat-associated face. Moreover, individuals who reported greater anxiety in social situations also showed heightened sharpened tuning of visuocortical neurons to facial identity cues, indicating the behavioral relevance of visuocortical tuning during generalization learning.

Introduction

Selectively perceiving and differentially responding to cues associated with threat versus safety is a fundamental function of the vertebrate brain. The dysregulation of these functions is at the core of many psychiatric problems. Over the past decade, basic and applied researchers interested in mental health have focused on the contribution of dysfunctional associative learning mechanisms to the etiology of anxiety disorders (Dymond et al., 2015). Given its intuitive relation with exaggerated fear and anxiety, the process of overgeneralization—showing threat responses to safety cues that resemble threat-associated stimuli—has been of particular interest. However, previous clinical and translational work has yielded contradictory findings. While some authors observed overgeneralization in patients with anxiety disorders (Kaczkurkin et al., 2017; Lissek et al., 2014b; Lissek et al., 2010), others did not (Ahrens et al., 2016; Tinoco-González et al., 2015). This lack of convergent findings may be due to the fact that different physiological systems respond differently to varying similarity with a fear stimulus.

When individuals are in a state of fear, defensive mechanisms are activated with the goal of engaging in adaptive action, for example in fighting or escaping the threat. This defensive engagement is indexed by somato-visceral measures, such as fear-potentiated startle, skin conductance, and cardiovascular responses (Boecker and Pauli, 2019; Bradley and Lang, 2000). These measures have been considered in previous fear generalization experiments (Ahrens et al., 2016; Lissek et al., 2010; Torrents-Rodas et al., 2013), with mixed results regarding the nature and variability of fear generalization across a range of cues varying in similarity with a threat cue (Ahrens et al., 2016). In addition to preparing autonomic and motor efferent systems for action, defensive mobilization also includes heightened sensory processing, that is, perception and attention to threat (Robinson et al., 2013). In line with this notion, a substantial body of research has shown that stimuli predicting threat are attended more than neutral cues, and that heightened attention toward threatening stimuli is pronounced in patients with anxiety disorders (Bar-Haim et al., 2007). At the same time, several studies focused on the importance of perceptual discriminability of threatening stimuli (Struyf et al., 2017; Zaman et al., 2019). As a consequence, excessive attention to threatening stimuli has been discussed as causal or contributory in the etiology of anxiety disorders (Clark and Wells, 1995; Rapee and Heimberg, 1997). Direct neurophysiological evidence of heightened attention to threat in clinical disorders is scarce, however, and findings of both heightened and diminished attention have been reported. The absence of a mechanistic account and lack of direct unequivocal evidence of attention dysfunction may be a result of the using indirect measures of attention to threat. The present study uses electrophysiological measurements from visual cortex to test mechanistic hypotheses derived from the structure and function of the human visual system.

Direct visuocortical responses to a specific stimulus may be quantified with the steady-state visually evoked potential (ssVEP, Müller et al., 1998). The ssVEP is an oscillatory neuronal response to stimuli that are periodically modulated in luminance. Heightened ssVEP amplitudes mark increased visuocortical activation and can be used to index attentional processes. For example, attended features (Müller et al., 2006) and selective spatial attention (Müller et al., 1998) facilitate visuocortical activation compared to unattended features and locations. SsVEPs are also sensitive to emotional processes and show increased amplitudes for emotional compared to neutral stimuli (Keil et al., 2003; Kemp et al., 2002; McTeague et al., 2011). Therefore, they provide a promising method for testing hypotheses regarding changing perception and attention as participants undergo fear generalization learning (Wieser et al., 2016). The amplitude of the ssVEP differentiates threat from safety signals, being selectively heightened for conditioned threat cues (reviewed in Miskovic and Keil, 2012). Building on these findings, McTeague et al., 2015 utilized ssVEPs to study population-level tuning of orientation-selective neurons in the primary visual cortex during fear generalization. In the neuroscience literature, tuning functions are often used to to describe differences in sensitivity of a response (neural or behavioral) along a physical feature gradient. For example, orientation tuning functions denote how neurons in the retinotopic visual cortex selectively respond to specific orientations (see Figure 1). At the population level, especially with scalp record fields, information regarding the tuning of individual neurons is obscure. Changes in the preferential tuning of population-level responses along a feature gradient can however be assessed with suitable research designs: To examine the extent to which aversive learning affects the population-level orientation tuning reflected in ssVEPs, McTeague et al., 2015 used high-contrast grating stimuli differing in orientation as conditioned stimuli. During pre-conditioning, the ssVEP amplitude evoked by each orientation was the same, resulting in a flat tuning curve. During acquisition, only the grating stimulus in the middle of this stimulus continuum was paired with an aversive noise (i.e. the CS+). After several conditioning trials, the CS+ prompted enhanced visuocortical engagement, accompanied by a suppression of the grating orientations with highest similarity to the CS+. This tuning pattern, which contrasts with the gradually decreasing generalization gradient observed in self-report and somato-visceral indices of learning, suggests lateral inhibitory interactions among orientation-selective neuronal populations in the visual cortex.

Different tuning functions during aversive learning.

The flat grey line depicts relative (behavioral or neural) responses along a physical feature gradient during pre-conditioning. The black curve shows a possible tuning function for lateral inhibition after aversive conditioning as seen in orientation-selective neuronal populations in the visual cortex (McTeague et al., 2015). Relative responses are enhanced for the specific feature associated with the aversive event and supressed for the most similar features. In contrast, the purple curve depicts a gradually decreasing generalization gradient observed in self-report and somato-visceral indices of generalization learning.

In the present study, we tested the hypothesis that sensory systems, when presented with a similarity gradient around a social threat stimulus, undergo changes to sharpen their tuning properties toward the relevant feature. Paralleling work on orientation-selective neurons discussed above, we expected amplification of visuocortical responses to the threat-associated face and a selective suppression of responses to the face most similar to the threat-associated face, reflecting inhibitory interactions between neuronal populations that represent facial features. This hypothesis is grounded in work showing substantial evidence for single-unit and population level (LFPs, fMRI) tuning in face-specific areas in the human and primate brain (Freiwald and Tsao, 2010; Freiwald et al., 2009; Gilaie-Dotan and Malach, 2007; Leopold et al., 2006; Loffler et al., 2005). These studies have demonstrated that there are neurons and neuronal populations in face-sensitive cortical areas, like the occipital face area (OFA) and the fusiform face area (FFA), which show gradual responses to varying facial identify, often referred to as ‘tuning’ to facial identities (Chang and Tsao, 2017). Here, we examine the malleability of population-level tuning as observers learn to associate one identity along a gradient of morphs with an aversive outcome. To further establish the behavioral relevance of visuocortical tuning, we also examine the extent to which such sharpened visuocortical tuning is associated with interindividual differences in social anxiety.

Results

Habituation

Steady-state visually evoked potentials (ssVEPs)

To induce ssVEPs, two different facial stimuli (CS+ and CS-) were presented with a flickering frequency of 12 Hz (see Figure 2). After converting the electrocortical signal to current source density (CSD) estimates and transforming it into the frequency-domain, the signal-to-noise ratio (SNR) was calculated by dividing the power of the driving frequency by the mean of the spectral power at six adjacent frequency bins, leaving out the two immediate neighbors. The resulting SNRs were pooled across eight sensor locations over the occipital pole for statistical analyses. During habituation, the linear mixed model analysis including the within-factor CS-type (two levels: CS+ vs CS-) and social anxiety as a continuous between-variable revealed a significant main effect for social anxiety, indicating that higher social anxiety was associated with higher ssVEP-SNRs in general and a non-significant trend for CS-type (CS+: M = 3.61, SD = 1.91; CS-: M = 3.39 SD=1.77). There was no CS-type x social anxiety interaction (see Table 1).

Experimental design.

Habituation, acquisition and generalization phase are illustrated. Stimuli were randomly presented for 3 s during each of the three parts of the experiment. The US consisted of a 1500 ms presentation of the respective CS+ face displaying a fearful expression, which was accompanied by a 95 dB shrill female scream. The assignment of face to CS+/CS- was counterbalanced between participants.

Table 1
Results of the linear mixed model analyses during habituation.
ssVEP-SNRs:
CS-TypeF(1,65)=3.18p=0.079=0.047CI = [.000,. 187]
Social anxietyt(65)=2.18p=0.033β = 0.48SE = 0.21
CS-Type x Social anxietyF(1,65)=0.69p=0.408=0.011CI = [.000,. 112]
Valence:
CS-TypeF(1,65)=4.07p=0.048=0.059CI = [.001,. 207]
Social anxietyt(65)=1.34p=0.186β = −0.15SE = 0.12
CS-Type x Social anxietyF(1,65)=1.16p=0.285=0.018CI = [.000,. 130]
Arousal:
CS-TypeF(1,65)=0.34p=0.562=0.005CI = [.000,. 096]
Social anxietyt(65)=1.24p=0.221β = 0.21SE = 0.17
CS-Type x Social anxietyF(1,65)=2.18p=0.145=0.032CI = [.000,. 162]

Valence and arousal ratings

The linear mixed model analysis of valence ratings revealed neither a significant main effect of social anxiety nor a CS-type x social anxiety interaction (see Table 1). However, subjects rated the CS- as slightly more unpleasant than the CS+, (CS+: M = 4.60, SD = 1.22; CS-: M = 4.99 SD=1.25). With regard to arousal ratings, there was no significant effect for CS-type (CS+: M = 3.69, SD = 1.61; CS-: M = 3.58 SD=1.58), social anxiety, or CS-type x social anxiety interaction. This result points out that none of the two faces elicited more arousal at the beginning of the experiment.

Acquisition

SsVEPs

After the CS+ had been paired with the US (see Figure 2), linear mixed model analyses for ssVEPs yielded a main effect of CS-type and a small effect of social anxiety, demonstrating that subjects reacted with higher amplitudes to the CS+ (M = 3.69, SD = 1.80) compared to the CS- (M = 3.33, SD = 1.56, see Figure 3a) and that higher social anxiety was associated with higher amplitudes in general (see Table 2). There was no significant CS-type x social anxiety interaction.

Boxplots and means (diamonds) of the (a) 12 Hz ssVEP signal-to-noise ratios (SNR) during the acquisition phase and mean US expectancy (b), arousal (c), and unpleasantness ratings (d) after acquisition.
Table 2
Results of the linear mixed model analyses during acquisition learning.
ssVEP-SNRs:
CS-TypeF(1,65)=5.50p=0.022=0.078CI = [.003,. 235]
Social anxietyt(65)=2.00p=0.050β = 0.37SE = 0.19
CS-Type x Social anxietyF(1,65)=1.57p=0.168=0.013CI = [.000,. 118]
Valence:
CS-TypeF(1,65)=49.51p<0.001=0.432CI = [.271,. 587]
Social anxietyt(65)=0.80p=0.424β = 0.09SE = 0.12
CS-Type x Social anxietyF(1,65)=0.00p=0.982=0.000CI = [.000,. 075]
Arousal:
CS-TypeF(1,65)=91.13p<0.001=0.584CI = [.447,. 705]
Social anxietyt(65)=0.05p=0.959β = 0.01SE = 0.16
CS-Type x Social anxietyF(1,65)=1.29p=0.259=0.020CI = [.000,. 135]
US expectancy:
CS-TypeF(1,65)=323.15p<0.001=0.833CI = [.771,. 884]
Social anxietyt(65)=0.17p<0.867β = −0.30SE = 1.77
CS-Type x Social anxietyF(1,65)=0.15p=0.736=0.002CI = [.000,. 083]

Valence and arousal ratings

After acquisition, there was a significant main effect of CS-type for both valence and arousal ratings (see Table 2), demonstrating that subjects rated the CS+ as more unpleasant (M = 6.42, SD = 1.41) and more arousing (M = 6.51, SD = 1.63) compared to the CS- (unpleasantness: M = 4.73, SD = 1.34; arousal: M = 3.69, SD = 1.94, see Figure 3c and d). Main effects of social anxiety and CS-type x social anxiety interactions were not significant.

US expectancy rating

The analysis detected a main effect of CS-type (see Table 2), with higher US expectancy ratings for the CS+ (M = 72.09, SD = 21.71) compared to the CS- (M = 8.66, SD = 18.74), underlining that the experimental manipulation was effective (see Figure 3b). There was no effect of social anxiety or CS-type x social anxiety interaction.

Generalization

Steady-state visually evoked potentials

During the generalization test, four additional faces (GS1 – 4) were presented by morphing the two faces of the previous phases together in 20% steps, with the GS1 being the most similar to the CS+ and the GS4 being the most similar to the CS- (see Figure 2). The CS-type (six levels: CS+ vs GS1 vs GS2 vs GS3 vs GS4 vs CS-) x social anxiety linear mixed model provided a main effect of CS-type and a marginal effect of social anxiety (see Figures 4 and 5a and Table 3). Post-hoc contrasts indicated significant differences between GS4 vs CS-, t(325) = 2.94, p=0.003, GS3 vs CS-, t(325) = 3.24, p=0.001, and CS+ vs CS-, t(325) = 2.17, p=0.031, but not between GS2 vs CS-, t(325) = 1.75, p=0.081, and GS1 vs CS-, t(325) = 0.82, p=0.411. However, the omnibus-test revealed no stimulus type x social anxiety interaction.

Mean scalp topographies of the 12 Hz ssVEP signal-to-noise ratios to the conditions during the generalization test.
Generalization test: (a) Mean 12 Hz signal-to-noise ratios (SNR) ±SEM of the ssVEP during the generalization test.

(b) Mean US-expectancy, (c) arousal and, (d) unpleasantness ratings ± SEM after generalization. Prediction intervals are shown for the mean ±1 SD of the SPAI covariate to illustrate the influence of social anxiety.

Table 3
Results of the linear mixed model analyses during generalization learning.
ssVEP-SNRs:
CS-TypeF(5,325)=3.39p=0.009=0.045CI = [.020,. 111]
Social anxietyt(65)=1.94,p=0.056β = 0.40SE = 0.21
CS-Type x Social anxietyF(5,325)=1.57p=0.167=0.024 CI = [.009,. 080
Valence:
CS-TypeF(5,325)=35.83p<0.001=0.355CI = [.286,. 436]
Social anxietyt(65)=0.17p<0.867β = 0.13SE = 0.12
CS-Type x Social anxietyF(5,325)=1.83p=0.107=0.027CI = [.011,. 086]
Arousal:
CS-TypeF(5,325)=66.80p<0.001=0.507CI = [.443,. 574]
Social anxietyt(65)=1.12,p=0.267β = −0.12SE = 0.18
CS-Type x Social anxietyF(5,325)=0.71p=0.618=0.011CI = [.005,. 060]
US expectancy:
CS-TypeF(5,325)=57.30p<0.001=0.469CI = [.402,. 540]
Social anxietyt(65)=1.37p=0.174β = −2.79SE = 2.03
CS-Type x Social anxietyF(5,325)=0.97p=0.435=0.015CI = [.006,. 067]

To follow up on the frequentist statistics and to seek converging evidence from a Bayesian perspective, we compared the fit of a lateral inhibition pattern to a quadratic and linear trend, using weighted predictors in Bayesian linear models (see Figure 6). To this end, we constructed weight vectors that reflected the predictions of the alternative hypotheses for the experimental conditions. This approach allowed us to leverage the specific model predictions under the lateral inhibition, generalization, and linear hypotheses, and to quantify the fit between the empirical data and model predictions in one test across all conditions. The lateral inhibition pattern was expressed as the difference of two Gaussians (weights: +2,–2, +0.5, +1, +0.5,–2 for CS+, GS1, GS2, GS3, GS4, CS-), paralleling previous studies on visuocortical tuning (Antov et al., 2020; McTeague et al., 2015). The quadratic (weights: +2.5334, +1.0934,–0.0267, −0.8267,–1.3067, −1.4667) and linear (weights: +2.5, +1.5, +0.5,–0.5, −1.5,–2.5) trend were modeled after the analyses of the linear and quadratic component of the generalization gradient, which are commonly employed in the fear generalization literature (Ahrens et al., 2016; Lissek et al., 2014a; Lissek et al., 2014b). In a first step, each Bayesian model was compared to the ‘random intercept only’ model (null model, 0), before transitive Bayes factors were calculated to obtain the relative evidence of one model over another. These transitive Bayes factors allow a direct comparison between the competing models. For a summary of resulting Bayes factors for each candidate main effect, interaction and predictor weight model see Table 4. The main effect of weighted CS-type received support for the lateral inhibition pattern only, BFLa1/0 = 13.26, but not for the quadratic, BFQ1/0 = 0.14, or linear trend, BFLi1/0 = 0.13. Further including a main effect of social anxiety did not lead to substantially increased support for the lateral inhibition pattern, BFLa2/La1 = 1.38, the quadratic, BFQ2/Q1 = 1.08, or linear trend, BFLi2/Li1 = 1.36. The strongest evidence could be found for the full interaction model (SNR ~CS type + social anxiety + CS-type x social anxiety) and the lateral inhibition pattern, BFLa3/0 = 57.32, which substantially extends the support of the main effects model, BFLa3/La2 = 3.13, suggesting that the accentuation of the lateral inhibition pattern increased with higher social anxiety. By contrast, the full interaction models for the Quadratic, BFQ3/0 = 0.11, and Linear trend, BFLi3/0 = 0.38, still yielded less evidence than the null model (see Figure 6). Importantly, by comparing the main and interaction effect models of each contrast against the same random intercept model, it was possible to derive the relative evidence of the lateral inhibition model over the quadratic and linear trend (see Table 4, last two columns). Results for the full interaction model demonstrated that the lateral inhibition model is 858 times more likely than the quadratic trend model and 502 times more likely than the linear trend model.

Bayesian model fit: Topographical distributions of the Bayes Factor for comparing the main effect model ‘SNR ~ CS-type’ to the null model for each contrast.

Weights used for the contrasts are displayed at the bottom row. Natural log-transformed BFs are illustrated, so that positive values display support for the full effect model while negative values display support for the null model.

Table 4
Summary of the Bayesian linear model analysis.
BFM/0ModelLateral inhibitionQuadratic trendLinear trendInhibition vs quadraticInhibition vs
Linear
M1:SNR ~ CS-type13.260.140.13104.1797.62
M2:SNR ~ CStype + SA18.330.150.17105.49125.16
M3:SNR ~ CS type + SA + CS-type x SA57.320.110.38858.52502.184
  1. Bayes factors of main and interaction effect models (M1 – M3) compared to the ‘random intercept only’ model (Null model) for the lateral inhibition pattern (weights: +2,–2, +0.5, +1, +0.5,–2 for CS+, GS1, GS2, GS3, GS4, CS-), quadratic trend (weights: +2.5334, +1.0934,–0.0267, −0.8267,–1.3067, −1.4667) and linear trend (weights: +2.5, +1.5, +0.5,–0.5, −1.5,–2.5). The last two columns display direct model comparisons between the lateral inhibition pattern to the quadratic or linear trend by dividing respective BFs for each main and interaction effect model. SA, social anxiety.

To follow up on the interaction of CS-type and social anxiety and to demonstrate the effect of social anxiety on the accentuation of the lateral inhibition pattern, we performed an additional regression analysis. In a first step, we calculated a visuocortical tuning index on subject level, which is the scalar product of the weights of the lateral inhibition model (2, –2, 0.5, 1, 0.5, –2) and the respective individual ssVEP responses. This visuocortical tuning index increases with a stronger accentuation of the lateral inhibition pattern and decreases if individual responses deviate from the pattern. For comparison, we calculated similar indices for the quadratic and linear trend. In a second step, we analyzed a linear regression model with social anxiety as predictor and the ssVEP indices as depended variables (see Figure 7). The regression analysis revealed a moderate, positive correlation between social anxiety and the visuocortical tuning index, r(65) = .288, p = 0.018, BF1/0 = 3.65, confirming that the accentuation of the lateral inhibition pattern increased with higher social anxiety, while the weak correlations between social anxiety and the index for the quadratic, r(65) = .146, p = 0.238, BF2/0 = 0.52, or linear trend, r(65) = .137, p = 0.268, BF3/0 = 0.49, missed significance and did not receive support from Bayesian-analyses.

Comparison of the correlations between social anxiety and the different contrast models.

Only the visuocortical tuning index as a parameter for the accentuation of the lateral inhibition pattern significantly increased with social anxiety. Marginal density plots display the distributions of the social anxiety scores and the ssVEP indices.

Valence and arousal ratings

The linear mixed models of the ratings yielded a significant main effect of CS-type for both valence and arousal ratings (see Table 3). Following the procedure of the ssVEP analysis, follow-up contrasts were calculated. With regard to arousal (see Figure 5c), subjects differentiated among the CS- and the CS+ plus three GS: CS- versus CS+, t(325) = 14.93, p<0.001, CS- versus GS1, t(325) = 10.28, p<0.001, CS- versus GS2, t(325) = 5.24, p<0.001, and CS- versus GS3, t(325) = 3.12, p=0.002, but not CS- versus GS4, t(325) = 1.06, p=0.290. For valence ratings (see Figure 4d), results showed differences between CS- and CS+, t(325) = 10.95, p<0.001, CS- and GS1, t(325) = 7.14, p<0.001, CS- and GS2, t(325) = 3.41, p<0.001, and CS- and GS3, t(325) = 2.06, p=0.040, but not between CS- and GS4, t(325) = .64, p=0.526. These results suggest that all subjects transferred their fear response reflected in both valence and arousal ratings from the CS+ to three GSs. In addition, there were no main effects of social anxiety nor CS-type x social anxiety interactions.

US expectancy ratings

US expectancy analysis revealed a main effect of CS-type (see Figure 5b and Table 3). Post-hoc contrasts yielded significant effects for the comparison of CS- versus CS+ t(325) = 13.73, p<0.001; GS1: t(325) = 8.88, p<0.001; and GS2: t(325) = 3.84, p<0.001. The differences between CS- and GS3: t(325) = 1.69, p=0.091, and CS- and GS4, t(325) = 1.14, p=0.254, were not significant. As found in valence and arousal ratings, subjects showed an enhanced tendency to generalize their conditioned fear reaction, indicated by the fact that they expected the GS1 and GS2 to be followed by the US, although they had never been paired with the US. The main effect of social anxiety and the CS-type x social anxiety interaction were not significant.

Discussion

The goal of the present study was to test the hypothesis that aversive generalization learning prompts sharpened representations of facial identity, reflecting inhibitory interactions between neuronal populations that represent facial features associated with threat versus safety. Second, we leveraged interindividual differences in social anxiety to examine whether the sharpened visuocortical tuning to facial identity is heightened in those characterized by higher social anxiety. For this purpose, steady-state visually evoked potentials (ssVEPs) as well as valence, arousal and US expectancy ratings were recorded in a fear conditioning and generalization paradigm with social cues.

Crucial for the later generalization test, successful fear conditioning was indexed in both ssVEP amplitude and arousal, valence and US expectancy ratings during acquisition, and social anxiety was not associated with stronger discrimination between conditioned stimuli or generally increased ratings. Thus, the CS+ during acquisition elicited increased visuocortical responses reflecting enhanced sensory engagement (Stegmann et al., 2019a; Wieser et al., 2014c). However, during habituation and acquisition, subjects with higher social anxiety showed overall amplified visuocortical responses to the face stimuli, which is in line with the notion of generally heightened sensitivity to facial expressions in social anxiety disorder (McTeague et al., 2018) and has been observed previously in patients with SAD using ssVEPs (McTeague et al., 2011) and in high socially anxious individuals using ERP components (Mühlberger et al., 2009) as markers of early perceptual processing.

During fear generalization, ratings of arousal, valence and US expectancy monotonically diminished with decreasing similarity to the CS+, indicating that subjects transferred their fear response from the threat signaling face to similar faces, although those had not been associated with the aversive outcome. These results corroborate previous studies on conditioned generalization, which demonstrated gradual, monotonic, generalization effects for subjective ratings (Lissek et al., 2008; McTeague et al., 2015; Stegmann et al., 2019b), somato-visceral measures, such as fear-potentiated startle (Lissek et al., 2008), skin conductance responses (Stegmann et al., 2019b; Torrents-Rodas et al., 2013) and heart rate (Ahrens et al., 2016) and electrocortical responses, that is, late positive potentials (Nelson et al., 2015).

Importantly, the visuocortical responses did not show a monotonic generalization gradient, but instead displayed a response pattern consistent with sharpening of the threat face representation. The amplitude of the ssVEPs – in contrast to the ratings – did not gradually diminish with increasing distance from the CS+. Instead, ssVEP amplitude for the CS+ was increased, but followed by an immediate reduction for the closest GS (GS1), and a slow increase in response to the remaining GSs, before it was again reduced for the learned safety cue (CS-). This observation is consistent with the response pattern of orientation sensitive neurons in the visual cortex during a fear generalization paradigm with enhanced visuocortical engagement to the CS+ and a suppression of the grating orientations with the highest similarity to the CS+ (McTeague et al., 2015). In line, the Bayesian analyses demonstrated that the observed ssVEP generalization data more likely corresponds to the lateral inhibition model compared to the null model, which received more support from the Bayesian analyses than the quadratic or linear trend model.

Together, these findings suggest that there is a dissociation in the aversive generalization patterns of sensory compared to efferent and autonomic systems, which is consistent with the idea that fear generalization is an active and multifaceted process that integrates a wide variety of signals to organize adaptive fear responses (Onat and Büchel, 2015). A large body of work has shown that visuocortical engagement with specific stimulus features varies with the motivational relevance of these features (Bradley et al., 2012). The present results are in line with this notion, but also suggest that these adaptive sensory changes observed during learning differ from the efferent system’s responses. To date, there are not many studies that are directly targeting the functional differences between sensory and efferent systems. Using steady-state visual evoked fields (the magnetoencephalographic counterpart of ssVEPs), two studies could demonstrate that threat-induced sensory changes in low-level visual areas occur independently of the conscious awareness of the CS-US contingencies (Moratti and Keil, 2009; Yuan et al., 2018), suggesting that visuocortical responses are to some extent independent of higher level cognitive processes. In contrast, it has been suggested, that threat-induced changes in the sensory system represent short-term plasticity in the early visual cortex, which might be induced by projections from subcortical structures like the amygdala or thalamus (Miskovic and Keil, 2012). This could be one reason for the divergence between visuocortical and subjective responses, as verbal reports reflect a more cognitive component of the defensive response (LeDoux and Pine, 2016). On the other hand, McTeague et al., 2015 also found a dissociation between the visuocortical and the fear-potentiated startle responses. The startle reflex, as an index of the autonomic component of the efferent system, is also assumed to be directly modulated by amygdala projections (Davis, 2006). In this case, the discrepancies between sensory and efferent response patterns might be mediated by different subregions of the amygdala, which modulate the sensory and efferent system according to their supposed functions in threat detection and defensive responding. Taking an evolutionary perspective, it is most adaptive for the organism to enhance sensory specificity in the visual cortex to distinguish the motivational information-providing stimulus from others sensory signals as reflected in the lateral inhibition model. On the other hand, the ‘efferent’ readiness to respond to a potential threat is generalized, as reflected in a monotonically decreasing generalization gradient, because a false alarm is less costly than a - potentially fatal - miss. This evolutionary interpretation assumes that the constantly changing and diverse environment in which humans find themselves demands and thus favors plastic physiological mechanisms, which may differ between systems in order to optimize functioning (Miskovic and Keil, 2012).

The present study also has implications regarding the neural mechanisms mediating aversive generalization learning, and regarding the neocortical changes underlying learning and memory more broadly: McTeague et al., 2015, using a generalization gradient of oriented gratings, explained their finding of visuocortical sharpening as a consequence of lateral inhibitory interactions among orientation-selective neuronal populations in the primary visual cortex. In this case, signals from frontoparietal attention networks may selectively facilitate CS+ representations in visual cortex, prompting local inhibitory interactions between adjacent cortical units. This process is thought to prompt suppression of the features represented by the most spatially proximal populations. In fact, ongoing computational modeling efforts in our laboratories explain these and other generalization data best by assuming that top-down signals take the shape of a generalization gradient (paralleling behavioral and autonomic data), and it is the organization of visual cortex that turns this gradient into a sharpening pattern through lateral inhibition. As such, sharpening would not be inherited by anterior areas but would result from the organization and geometry of visual cortical areas. This interpretation is further supported by research demonstrating that the orientation-tuning functions of visual neurons may be shaped by short-term plastic processes (Dragoi et al., 2000).

We hypothesized that in a situation in which facial identity predicts threat versus safety, a similar mechanism may operate in the cortical tissue that is specific not to orientation but to the features encoding facial identity. There are two loci where such tuning may occur: First, as aversive learning proceeds, projections from anterior structures may increasingly target lower-level representations of individual facial features in retinotopic areas, thus prompting inhibitory interactions between individual physical facial features such as orientation or spatial frequency, which differ across the morphing gradient. This is consistent with findings and models in perceptual learning (e.g. reverse hierarchy theory; Ahissar and Hochstein, 2004) and would prompt a mid-occipital topography of sharpening effects in the present study.

Second, inhibitory interactions may occur between similar faces along a gradient of morphs, in face-sensitive cortical areas. This alternative hypothesis is in line with recent evidence suggesting that neuronal populations in face-sensitive cortical patches encode identity by combining their population tuning to sets of high-order shape and appearance dimensions (e.g. Chang and Tsao, 2017). Neural populations that are sensitive to such features may be located in the early visual cortex or at later stages of the face-processing pathway, for example the occipital face area (OFA) or the fusiform face area (FFA) (Duchaine and Yovel, 2015). The present study suggests that the amplification of these selective neuronal cell ensembles to a given threat face prompts lateral inhibitory interactions among neurons that are selective to slightly different facial features. Our results support the hypothesis that such interactions take place in extra-striate, higher order visuocortical areas, as we exploited the assets of high-density EEG recording and CSD-transformation leading to a precise topographical distribution of the ssVEP signal. These signal topographies reveal indeed right-lateralized activation peaks for the signal-to-noise ratios of the ssVEP signal (Figure 4) as well as for the model fit of the lateral inhibition pattern (Figure 6). Thus, our results are compatible with the idea of a right-hemispheric dominance of face perception (Rossion, 2014) and suggest an involvement of the right OFA or FFA in sharpening face discrimination on the basis of lateral inhibition.

Furthermore, Bayesian model and correlational analyses revealed that visuocortical system engagement is associated with self-reported symptoms of social anxiety. This result adds to a growing body of literature on attentional biases in social anxiety, including evidence from reaction-time tasks, (Bantin et al., 2016), eye-tracking (Wieser et al., 2009), event-related potentials (ERPs, Mühlberger et al., 2009; Wieser et al., 2010) and ssVEPs (McTeague et al., 2018; McTeague et al., 2011). Here, we provide further evidence that cortical engagement in response to threat-associated phobic-relevant stimuli is dimensionally related to social anxiety. It is important to mention, however, that we - replicating Ahrens et al., 2016 – found no association between strength of social anxiety and (over)generalization of conditioned fear in terms of a higher fear responses (efferent system) to the generalization stimuli. Instead, we observed social anxiety to be associated with a more pronounced lateral inhibition pattern in visuocortical responses to generalization stimuli. Please note that the effect of social anxiety on visuocortical responses was only evident in the Bayesian-analysis but not in the corresponding ANOVA. This discrepancy results from the differences regarding statistical power between omnibus- and contrast-analysis (Furr and Rosenthal, 2003). The ANOVA tests for any differences, while the contrast-analysis only tests for deviations from the specified pattern. This notion is substantiated by the finding that social anxiety was not associated with visuocortical responses for the quadratic or linear trend model, but only for the lateral inhibition model. We conclude that this latter finding is indicative for the functional relevance of the lateral inhibition model in visuocortical tuning during generalization learning. Given the healthy subjects of this study, however, with those showing a psychiatric disorder being excluded, future studies should examine diagnosed and treatment-seeking SAD patients to substantiate initial findings and to draw conclusions on how sensory generalization versus sharpening contributes to the etiology or maintenance of social anxiety disorder or other psychopathologies.

One important limitation to our findings is that there was no substantial increase in differential ssVEP-SNRs (CS+ vs CS-) from habituation to acquisition, t(66) = 0.83, p = 0.409, d = 0.10, CI95 = [−0.14, 0.34]. A reason for the lack of such a difference may be related to the trendwise differences between CS+ and CS- at baseline, which often indicate different responses to one of the two faces. However, this seems unlikely in our study, because the mapping of the two faces to the CS+ and CS- was fully counterbalanced between subjects. In addition, in the differential conditioning literature, cross-phase comparisons are typically avoided because they confound any conditioning effects with time-dependent effects such as adaptation, in which both the CS+ and CS- evoked responses decline over time, from habituation to acquisition and finally into the extinction phase. This has been well established in meta-analyses (e.g. Fullana et al., 2016), and is true for a wide range of dependent measures, including startle, skin conductance, fMRI, and ssVEPs where this pattern of findings has been noted and systematically addressed several times (e.g. Keil et al., 2013; Moratti and Keil, 2005). Thus, fear conditioning studies interested in temporal changes previously utilized analysis on a trial-by-trial level in order to avoid confoundation with these types of adaptation processes (e.g. Sjouwerman et al., 2016; Weike et al., 2007; Wieser et al., 2014c).

Post-hoc analyses comparing pre- and post-conditioning: although a lack of statistically robust changes between the experimental phases would not affect our interpretation of the main results regarding the generalization phase, we addressed this potential concern in a post-hoc analysis quantifying the amount of conditioning-related effects above and beyond the initial difference in habituation by means of parametric bootstrapping in combination with Bayesian analyses (Efron, 2012).

To this end, we computed a distribution for the CS+ versus CS- difference in ssVEP amplitude separately for habituation (to be used as prior distribution) and acqusition (posterior distribution) based on 100,000 bootstrapped group mean differences. Odds for conditioning effects to occur were estimated from these two distributions (the odds of the CS+-CS- difference to be positive), and the BF of interest was given as the ratio of posterior (acquisition) over prior (habituation) odds. The error of this process was estimated by repeating the above process 100 times and measuring the standard error of the resulting mean. This BF corresponds to the change in confidence that the ssVEP for the CS+ is greater than ssVEP for the CS- in acquisition (the posterior), relative to the prior distribution, which was estimated from the CS+ versus CS- difference during habituation. This analysis yielded a BF (posterior odds over prior odds) of 4.77, error = 1.3%, suggesting that the acquisition data provided substantial evidence for the notion that fear conditioning selectively amplifies the CS+ evoked ssVEP above and beyond the differences present during habituation.

In conclusion, our study extends current notions of generalization learning, by demonstrating the involvement of inhibitory interactions among feature-specific neurons in the visuocortical system during fear generalization to facial stimuli. We found that the accentuation of the lateral inhibition pattern increased with the severity of social anxiety. Future research may examine stability of the lateral inhibition response pattern during extinction as well as the role of perceptual sharpening in fear extinction learning.

Materials and methods

Subjects

Subjects were 67 undergraduate students (age: M = 24.10, SD = 6.33; 48 female) with normal or corrected vision and without past or present psychiatric diagnosis or family history of epilepsy (self-report), who were paid or received course credit for participation. The sample size was based on the effect sizes of previous aversive conditioning studies using ssVEPs (McTeague et al., 2015; Miskovic and Keil, 2013) and adapted for covariate-analyses. Subjects completed the Social Phobia and Anxiety Inventory (SPAI, German version, Beidel et al., 1989) as a self-report measurement of social anxiety (M = 67.70, SD = 19.74, Min = 32.48, Max = 126.60). Prior to participation, written informed consent was obtained from each participant. The study was approved by the ethics review board of the Medical Faculty of the University of Würzburg (87/13).

Stimuli and apparatus

Request a detailed protocol

Conditioned stimuli (CS) consisted of two pictures of female actresses with a neutral facial expression taken from the NimStim Set of Facial Expressions (Tottenham et al., 2009). Pictures were adjusted for luminance and brightness, converted to gray-scale and presented using Presentation (Neurobehavioral Systems, Inc, Albany, CA). One of the actresses was randomly selected as threat cue for each participant (CS+) while the other served as safety signal (CS-). Pictures were shown on a gray background on a 17-inch monitor (resolution = 1,280×1,024 pixel) in a flickering mode at a frequency of 12 Hz in order to elicit ssVEPs. Face-specific areas are often targeted with relatively slow driving frequencies (Baldauf and Desimone, 2014), but face-specific processing has also been isolated from paradigms with faster frequencies (Campagnoli et al., 2019; Gruss et al., 2012; Wieser et al., 2014a; Wieser et al., 2014c).

The US consisted of the respective CS+ face displaying a fearful expression and a simultaneously presented 95 dB shrill female scream of the IADS (Bradley and Lang, 1999). Four generalization stimuli (GS) were created by morphing the two faces in 20% steps using a face-morphing software (Squirlz Morph; Xiberpix, Solihull, UK). The GS most similar to the CS+ is referred to as GS1 and the GS most similar to the CS- as GS4 (see Figure 2).

Design and procedure

Request a detailed protocol

The experiment comprised three blocks (habituation, acquisition, generalization). Habituation and acquisition consisted of 30 trials (two faces, each presented 15 times), while there were 90 trials in the generalization phase (six faces, each presented 15 times), resulting in 150 trials in total. After completing the questionnaires, EEG electrodes were applied to participants, who were seated in a noise-reduced, darkened room one meter distant to the screen. In the habituation phase of the experiment, faces were presented for 3000 ms without reinforcement. During acquisition, one of the faces (CS+) was paired in 12 of 15 trials (80% reinforcement) with the US, which lasted 1500 ms and was presented at the offset of the CS+ with a sound volume of 95 dB by Labtech speakers (Labtech International Ltd., Ringmer, East Sussex, GB) and a Kenwood KA-3010-Amplifier (Kenwood Electronics, Heusenstamm, GER). Subjects were not informed of any specific relation among the CSs and the US prior to the experiment and the assignment of faces to CS-conditions was counterbalanced across subjects. Generalization consisted of the CS+, CS-, and four GS, each presented 15 times (90 trials). While CS- and GS were never reinforced, 6 of the 15 CS+ were still followed by the US to prevent early extinction (40% reinforcement, Lissek et al., 2008) (see Figure 2). The presentation order of the faces within each block was pseudo-randomized such that no more than two of the same faces could occur in a row. After each trial, a gray screen with a fixation cross was presented. Inter-trial intervals differed between 2000 and 2500 ms. At the end of each phase, subjects rated the valence (ranging from 1 - ‘very pleasant’ to 9 - ‘very unpleasant’) and arousal (ranging from 1 – ‘very calm’ to 9 – ‘very arousing’) of the faces using a computer-based version of the Self-Assessment Manikin Scale (Bradley and Lang, 1994). Moreover, subjects were asked to rate US expectancy after acquisition and generalization from 0% to 100% as a response to the question ‘What is the likelihood that the currently presented face is followed by a scream?’ to measure successful learning of the CS-US association.

EEG recording and analysis

Request a detailed protocol

Electrocortical activity was measured via 129 electrodes using an Electrical Geodesics (EGI, Eugene, OR) high-density EEG System referenced to Cz, recorded with a sampling rate of 250 Hz and online bandpass filtered with 0.1 and 100 Hz and a 50 Hz notch filter. The threshold of impedances was kept below 50 kΩ as recommended for the Electrical Geodesics high-impedance amplifiers. Offline, EEG analyses were implemented using the software EMEGS (Electro Magnetic Encephalography) for Matlab (Peyk et al., 2011). First, epochs of 600 ms pre-stimulus and 3000 ms post-stimulus onset were extracted, and data were filtered with a low-pass filter of 40 Hz. In a second step, artifact rejection was conducted according to the SCADS procedure (Junghöfer et al., 2000). This way, outlying channels could be identified and interpolated from the full channel set and artifact-contaminated trials could be excluded from the analyses. Trials were rejected when more than 20 channels out of 129 were outliers as per the statistical parameters used for artifact identification. In a next step, remaining trials were averaged in the time domain for each subject according to the different experimental conditions. To increase spatial resolution of the EEG signal, we then calculated the current source densities (CSD) of the time-averaged data. The CSD transformation offers the advantage of reducing the negative impact of volume conduction and thereby effectively minimizing unwanted topographical variability between subjects by quantifying source densities (Junghöfer et al., 1997; Kayser and Tenke, 2015). The CSD approach relies on the spatial Laplacian (the second spatial derivative) of the scalp potential to estimate the potential distribution at the cortical surface. Here, we used the CSD algorithm described by Junghöfer et al., 1997, with λ = 0.2, which is well suited for dense-array EEG montages and has been used in previous studies investigating ssVEP responses to facial stimuli (McTeague et al., 2011) or during fear generalization (McTeague et al., 2015). The CSD time series values were then transformed into the frequency domain using a Fast-Fourier-Transformation on a time interval between 500 and 3000 ms after stimulus-onset. The first 500 ms after stimulus onset were omitted due to initial non-stationary components of the ssVEP. In a next step, we obtained the signal-to-noise ratio (SNR) for the driving frequency of 12 Hz by dividing the power of the driving frequency by the mean of the spectral power at six adjacent frequency bins, leaving out the two immediate neighbors. The SNR is a unitless measure that accounts for both the evoked signal and the random noise in the data and has recently been used in other ssVEP paradigms as well (Barry-Anwar et al., 2018; Boylan et al., 2019).

For statistical analysis, the ssVEP activity was pooled across sensor Oz and seven neighbouring electrodes (EGI sensors 70, 71, 72, 74, 75, 76, 82, 83; Wieser et al., 2014a; Wieser and Keil, 2014b).

Statistical analysis

Request a detailed protocol

Stimulus differences in ssVEP amplitudes as well as valence and arousal ratings during habituation and acquisition were analyzed with linear mixed models with the within-subject factor CS-type (2: CS+, CS-) and mean-centered SPAI scores as covariates. Both main effects and the interaction of CS-type and SPAI were entered as fixed effects. Subjects were entered as random intercepts to the model. For the generalization phase, the same linear mixed model was analyzed, though the factor CS-type now included six levels (CS+, GS1, GS2, GS3, GS4, CS-). Follow-up tests for significant effects of CS-type were analyzed using simple contrasts with the CS- as reference level (Lissek et al., 2008). US expectancy ratings underwent the same analyses except for the habituation phase, because US expectancy ratings just started after the acquisition phase. Significance was evaluated using the Kenward-Rogers approximation for degrees of freedom (Kenward and Roger, 1997; Luke, 2017). Alpha was set at p<0.05 (two-tailed). Linear mixed models were conducted in the R software environment (version 3.5.0.; R Development Core Team, 2020), using the package ‘lme4’ (version 1.1–20; Bates et al., 2015). Standardized effect sizes and confidence intervals for the discrete factors of the linear mixed models were calculated as partial-, using the package ‘r2glmm’ (version 0.1.2; Jaeger et al., 2017).

To compare the lateral inhibition pattern to a quadratic and linear fit, Bayesian linear models were used. For this analysis, a pre-specified weight vector for each contrast entered the linear regression as predictors. The lateral inhibition pattern was expressed as the difference of two Gaussians (weights: +2,–2, +0.5, +1, +0.5,–2 for CS+, GS1, GS2, GS3, GS4, CS-). Note, that in contrast to the study of McTeague et al., 2015, generalization learning occurred along a continuum from CS+ to CS-, which is why the weight vector was adapted to only one half of the previously used model. For the quadratic and linear trends, the following weights were used, respectively: quadratic (weights: +2.5334, +1.0934,–0.0267, −0.8267,–1.3067, −1.4667) and linear (weights: +2.5, +1.5, +0.5,–0.5, −1.5,–2.5), paralleling previous work on fear generalization (Ahrens et al., 2016; Lissek et al., 2014a; Lissek et al., 2014b). The linear model analysis is mathematically insensitive to any linear transformation of the contrast vectors, which is why every vector was centered around zero to facilitate model comparisons. In each model, centered SPAI scores were entered as an additional predictor variable and subjects were entered as random intercepts to the model. Transitive Bayes factors (BFs) were then calculated for each candidate main effect, interaction and predictor weight model. Interpretation of Bayes factors follows guidelines developed by Jeffreys (1961). Bayesian analyses were conducted in R, using the package ‘BayesFactor’ (version 0.9) and default JZS-priors (Rouder et al., 2012). To follow up on the hierarchical model selection, individual visuocortical tuning indices were calculated for the lateral inhibition pattern. This index was defined as the scalar product of the weights of the lateral inhibition pattern and the subjects’ individual ssVEP-response to the corresponding stimulus, so that higher/lower visuocortical tuning indices indicate stronger/weaker accentuation of the lateral inhibition pattern. Similar indices were calculated for the quadratic and linear trend. In a next step, these indices were correlated with the SPAI scores. For the frequentist analysis, Pearson’s r was calculated, and the alpha level was set to .05, while the population correlation parameter ρ was estimated for a two-sided Bayes factor analysis (Ly et al., 2016).

References

  1. 1
  2. 2
  3. 3
  4. 4
  5. 5
  6. 6
  7. 7
  8. 8
  9. 9
  10. 10
  11. 11
  12. 12
  13. 13
  14. 14
    The International Affective Digitized Sounds (IADS): Stimuli Instruction Manual and Affective Ratings
    1. MM Bradley
    2. PJ Lang
    (1999)
    NIMH Center for the Study of Emotion and Attention.
  15. 15
    Emotion and Motivation
    1. MM Bradley
    2. PJ Lang
    (2000)
    In: JT Cacioppo, LG Tassinary, G Berntson, editors. Handbook of Psychophysiology. Cambridge University Press. pp. 602–642.
  16. 16
  17. 17
  18. 18
    A cognitive model of social phobia
    1. DM Clark
    2. A Wells
    (1995)
    In: RG Heimberg, editors. Social Phobia: Diagnosis, Assessment, and Treatment. New York: Guilford Press. pp. 69–93.
  19. 19
  20. 20
  21. 21
  22. 22
  23. 23
  24. 24
  25. 25
  26. 26
  27. 27
  28. 28
  29. 29
  30. 30
  31. 31
  32. 32
  33. 33
  34. 34
  35. 35
  36. 36
  37. 37
  38. 38
  39. 39
  40. 40
  41. 41
  42. 42
  43. 43
  44. 44
  45. 45
  46. 46
  47. 47
  48. 48
  49. 49
  50. 50
  51. 51
  52. 52
  53. 53
  54. 54
  55. 55
  56. 56
  57. 57
  58. 58
  59. 59
  60. 60
  61. 61
    R: A language and environment for statistical computing
    1. R Development Core Team
    (2020)
    R Foundation for Statistical Computing, Vienna, Austria.
  62. 62
  63. 63
  64. 64
  65. 65
  66. 66
  67. 67
  68. 68
  69. 69
  70. 70
  71. 71
  72. 72
  73. 73
  74. 74
  75. 75
  76. 76
  77. 77
  78. 78
  79. 79
  80. 80
  81. 81

Decision letter

  1. Richard B Ivry
    Senior Editor; University of California, Berkeley, United States
  2. Jonas Obleser
    Reviewing Editor; University of Lübeck, Germany
  3. Christian Keitel
    Reviewer; University of Stirling, United Kingdom
  4. Ulrike M Krämer
    Reviewer

In the interests of transparency, eLife publishes the most substantive revision requests and the accompanying author responses.

Acceptance summary:

Stegmann et al. address an important and interesting question, namely whether social fear conditioning leads to changes in neural processing of fear-conditioned faces versus other faces that are more or less similar to the acquired one. They investigate whether sharpened visuo-cortical tuning, a phenomenon usually reported for low-level visual features such as orientation, also affects the neural processing of social threat cues. The study uses a relatively simple and straightforward fear-conditioning paradigm and an non-invasive measure of brain electric activity (so-called steady-state visual evoked potentials, SSVEPs) to measure visual-cortical processing as well as a series of self-report measures. The results suggest sharpened tuning, correlated with interindividual differences in social anxiety. The reviewers and the Reviewing Editor collectively found considerable merit in the approach and findings and agreed that a case in terms of a lateral inhibition-like situation can be made, also in the light of earlier literature.

Decision letter after peer review:

Thank you for submitting your article "Social aversive generalization learning sharpens the tuning of visuocortical neurons to facial identity cues" for consideration by eLife. Your article has been reviewed by three peer reviewers, and the evaluation has been overseen by Jonas Obleser, the Reviewing Editor, and Richard Ivry as the Senior Editor. The following individuals involved in review of your submission have agreed to reveal their identity: Christian Keitel (Reviewer #1); Ulrike M. Krämer, PhD (Reviewer #3).

The reviewers and editors have discussed the reviews with one another, and the Reviewing Editor has drafted this decision to help you prepare a revised submission.

Summary:

The reviewers and the Reviewing Editor collectively found considerable merit in your approach and findings. We agreed that a case in terms of a lateral inhibition-like situation can be made, also in the light of earlier literature. However, there was clear consensus that the manuscript in its current form is not convincing yet and will require (1) refined statistical evidence being put forward, and (2) more detailed grounding in the principles of visual cortical processing and its modulation by (conditioned) fear, that is, a stronger framework behind the authors' assumptions and interpretations.

See below for further guidance to preparing a revised version.

Depending on revision and accompanying rebuttal letter, we will consult all or some of the present reviewers again.

Essential revisions:

– The degree to which the current data do show a sharpening of cortical tuning is contentious, at least. As such, title and Abstract may be misleading. SSVEPs are used, which measure population responses in visual cortex broadly (it is unclear whether the signals here are even stemming from face-specific areas – most likely not, as previous studies have shown that slower frequencies (<12Hz) are more effective in driving higher regions in visual cortex, see Baldauf and Desimone, 2014), and the "tuning" is inferred from a Bayesian model comparison that is rather indirect.

– While the study is trying to understand mechanisms of fear-conditioning, it lacks an account of how lateral inhibition in the processing of fear-conditioned faces would be implemented in the brain. The comparison to orientation tuning is challenging. Do the authors assume a face similarity map in visual cortex where similar faces are represented nearby each other, and so processing of similar faces can be attenuated through lateral inhibition? Would this happen via an attentional mechanisms, or through other connections? None of this seemed specified, and we struggled to understand of how such "tuning" would occur in fearful face representations.

– Related, the authors do not really come up with a suggestion how to link the sharpened tuning of the visuo-cortical responses on the one hand and the generalization of the subjective fear experience on the other hand. It is only stated that fear generalization is an "active and multifaceted process". As this is a central aspect of this study and pertains to the functional relevance of the observed enhanced tuning of the SSVEP response, one would like to see a bit more discussion of this discrepancy.

– For the generalization phase, the GS1-4 faces were treated as a single factor, which did not seem right to the reviewers. These should be separated into 4 different factors, given the authors' assumption that these 4 faces should show different modulations (i.e., lateral inhibition model and also linear decrease model).

– The support for lateral inhibition model only comes from a comparison of a series of different models to a random intercept model. This is weak evidence in favor of a lateral inhibition model. Before doing the modeling, it should first be tested whether the data for the GS1-4 faces differs reliably between each other (see last point), which is doubtful looking at the figure.

Second, a fair modeling approach would be to compare the models against each other (linear vs. lateral inhibition), and not take each model and compare it to the random intercept model, and then pick the model that shows the highest BF. Transitive BFs might be utilized here, accordingly.

– Each of the tested models does not appear to be well justified based on previous literature or any mechanistic accounts of fear conditioning, and it remains a bit unclear why these models were chosen.

– A potentially non-negligible difference (p = 0.07) between CS+ and CS- at baseline was noted, with the same directionality as the significant effect after fear condition. This seems problematic given that the difference remains relatively small after (and during) conditioning (though normalized effect sizes are not reported). We find such a small change extremely difficult to interpret. To be convinced that this is a real change in SNR through conditioning, we would want to see 1) statistics showing a reliable increase of the SSVEP response (so pre- and post training comparisons) which are currency lacking, and 2) effect sizes (at best with CIs for these, J.O.).

– The statistical results are in many places interpreted in a seeming all-or-none fashion, which does not seem appropriate. Effects with p-values of e.g. 0.05 are interpreted as true, while effects with p-values of e.g., 0.056 are interpreted as absent. Again, effect sizes should be reported, and a non-dichotomising interpretation of effects is encouraged.

– Related, the ANOVA results for the generalization show, if we understand correctly, a main effect for the CS-type but no main effect or interaction with social anxiety. The Bayesian linear models does find support for a full interaction model including a main effect and interaction with social anxiety. This discrepancy is not really resolved or discussed so far.

– Reviewers were not unequivocally convinced that the Current source density transforms are useful or adequate for SSVEPs. Please make sure to justify their use better (potentially also demonstrating their effect on overall conclusions/results, e.g. in a supplement). The authors should motivate their choice more strongly and report parameters λ/m.

– The authors have added an additional analysis that relates a measure of visuo-cortical tuning (i.e. the individual sharpening of response profiles after fear conditioning) to a measure of social anxiety. Indeed, they find a moderate positive correlation, strengthening their point that the individual sensitivity to threat cues can affect visuo-cortical tuning. Computing individual cortical tuning as weighted sum of SSVEP and model weights seems elegant. However, the reader is somewhat left at the authors' mercy to confide in the sensitivity of this measure. We strongly suggest to repeat this analysis with the weights of the Quadratic and Linear trend models and compare them with the lateral inhibition situation to show that there is no correlation in these cases.

– The manuscript is generally well-written and easy to read. The Results section however suffers from heavy in-text reporting of statistics. Reading flow will be improved by reporting stats in tables instead. Also consider the use of supplementary tables.

[Editors' note: further revisions were suggested prior to acceptance, as described below.]

Thank you for resubmitting your work entitled "Social aversive generalization learning sharpens the tuning of visuocortical neurons to facial identity cues" for further consideration by eLife. Your revised article has been evaluated by Richard Ivry (Senior Editor) and a Reviewing Editor.

The manuscript has been improved considerably.

Your work addresses an important and interesting question, namely whether social fear conditioning leads to changes in neural processing of fear-conditioned faces versus other faces that are more or less similar to the acquired one.

Reviewers and myself collectively found considerable merit in the approach and findings. We agreed that a case in terms of a lateral inhibition-like situation can be made, also in the light of earlier literature. The authors did a compelling job in carefully addressing most and certainly all major questions from us reviewers and editors.

We do believe there remain two issues that require further attention.

We anticipate that the reviewing and senior editor should be able to act on the revision without consulting the reviewers.

Requested revisions:

1) One reviewer alerted us that one issue raised in the initial round of reviews was not addressed, the potentially quite important pre-post comparison for SSVEP amplitude (i.e., pre-and post-training). In our view, it would be an important analysis to ask if there is an effect over time. We think proper reporting (and discussion) of such a pre-post comparison is important for making inferences that these results point to a change in neural tuning due to habituation.

2) The other remaining issue if of lesser concern:

You have not done a direct comparison of the models to each other. Please consider reporting these comparisons. The reviewer provides the following guidance: "While I follow that there is some benefit in comparing all models against a random intercept model, I would still want to see whether the differences between the GS1-4 faces differs reliably from each other (follow-up paired t-test)."

https://doi.org/10.7554/eLife.55204.sa1

Author response

Essential revisions:

– The degree to which the current data do show a sharpening of cortical tuning is contentious, at least. As such, title and Abstract may be misleading. SSVEPs are used, which measure population responses in visual cortex broadly (it is unclear whether the signals here are even stemming from face-specific areas – most likely not, as previous studies have shown that slower frequencies (<12Hz) are more effective in driving higher regions in visual cortex, see Baldauf and Desimone, 2014), and the "tuning" is inferred from a Bayesian model comparison that is rather indirect.

Defining what we mean by “tuning” is indeed crucial to avoid misconceptions. In the revised version, we give now an explicit definition of population-level tuning early on in the Introduction and explain the rationale for studying population-level tuning. Specifically, we base our conceptualization on conventions in vision neuroscience and psychophysics, where tuning functions are used to describe differences in sensitivity of a response (neural or behavioral) along a physical feature gradient. As the reviewers mention, probably best known is research conducted on orientation tuning of individual neurons in retinotopic visual cortex, meaning that visual neurons selectively respond to specific orientations as noted by the reviewers. There is also substantial evidence for single-unit and—importantly—population level (LFPs, fMRI) tuning in face-specific areas stemming from research in primates (e.g. Leopold et al., 2006; Freiwald et al., 2009; Freiwald and Tsao, 2010) and humans (e.g. Gilaie-Dotan et al. 2006; or Loffler et al., 2005). These studies have demonstrated that there are neurons and neuronal populations in face-sensitive cortical areas (patches in macaques), like the OFA and the FFA, which show gradual responses to varying facial identify, often referred to as “tuning” to facial identities. In this growing literature, authors often use the present approach, i.e. fitting a-priori tuning profiles to neural response data across a gradient of morphed faces (e.g., Koyano et al., 2019).

We agree with the reviewers that ssVEPs likely reflect population responses across different cortical areas. Changes of the tuning of a population response would then imply that the population response changes its profile across a feature gradient of interest. Sharpened tuning, as operationally defined in this sense then, is intended not to denote an absolute quality of the cortical tissue but the process in which the selective population response is altered across the gradient, in the present case by responding less to features that closely resemble a reference stimulus, or an attended stimulus. A shorter version of the above narrative is now included in the manuscript. This concern is also linked to concerns raised regarding the neural mechanism that would prompt population-level sharpening, addressed in the context of the next comment. We added the following section and a figure to illustrate different examples of tuning functions to our Introduction:

“The amplitude of the ssVEP differentiates threat from safety signals, being selectively heightened for conditioned threat cues (reviewed in Miskovic and Keil, 2012). […] Here, we examine the malleability of population-level tuning as observers learn to associate one identity along a gradient of morphs with an aversive outcome.”

Finally, to adress the concerns regarding slower frequencies being more effective in eliciting signals in face-specific areas, we added a paragraph with references to the Materials and methods section, stating that even though face-specific areas are often targeted with relatively slow driving frequencies (Baldauf and Desimone, 2014), face-specific processing has also been isolated from paradigms with faster frequencies, when combined with appropriate experimental manipulations (Campagnoli et al., 2019; Gruss, Wieser, Schweinberger and Keil, 2012; Wieser, Flaisch and Pauli, 2014; Wieser et al., 2014)

– While the study is trying to understand mechanisms of fear-conditioning, it lacks an account of how lateral inhibition in the processing of fear-conditioned faces would be implemented in the brain. The comparison to orientation tuning is challenging. Do the authors assume a face similarity map in visual cortex where similar faces are represented nearby each other, and so processing of similar faces can be attenuated through lateral inhibition? Would this happen via an attentional mechanisms, or through other connections? None of this seemed specified, and we struggled to understand of how such "tuning" would occur in fearful face representations.

We expanded on the hypothetical neural mechanisms in the revised version of the manuscript. In brief, based on the animal studies cited above, and based on the extant work on neuroplastic changes for tuning to low level features, we considered two hypotheses that rely on the same computational principle: As stated in McTeague et al., 2015 in the case of orientation tuning, signals from frontoparietal attention networks may selectively faciliate CS+ representations in visual cortex, prompting local inhibitory interactions between adjacent cortical units (here: orientation columns). This process is thought to prompt suppression of the features represented by the most spatially proximal populations. We have developed a computational model (currently under review) that explains these and other data best by assuming that top-down signals take the shape of a generalization gradient (paralleling behavioral and autonomic data), and it is the organization of visual cortex that turns this gradient into a sharpening pattern through lateral inhibition. As such, sharpening would not be inherited by anterior areas but would result from the organization of visual cortical areas.

In the case of facial identity gradients, as alluded to by the reviewers, there are two loci where such tuning may occur, each of which being supported by published findigs:

First, as aversive learning proceeds, projections from anterior structures may increasingly target lower-level representations of individual facial features in retinotopic areas, thus prompting inhibitory interactions between individual physical facial features such as orientation or spatial frequency, which differ across the morphing gradient. This is consistent with findings and models in perceptual learning (e.g. reverse hierarchy theory; Ahissar and Hochstein, 2004) and would prompt a mid-occipital topography of sharpening effects in the present study.

Second, inhibitory interactions may occur between similar faces along a gradient of morphs, in face-sensitive cortical areas. This alternative hypothesis is in line with recent evidence suggesting that neuronal populations in face sensitive cortical patches encode identity by combining their population tuning to sets of high-order shape and appearance dimensions (e.g., Chang and Tsao, 2017). In the present study, this hypoythesis would be supported by a right-occipital topography of sharpened tuning.

The full discussion of the neural mechanisms now reads as follows:

“McTeague et al., 2015, using a generalization gradient of oriented gratings, explained their finding of visuocortical sharpening as a consequence of lateral inhibitory interactions among orientation-selective neuronal populations in the primary visual cortex. […] Thus, our results are compatible with the idea of a right-hemispheric dominance of face perception (Rossion, 2014) and suggest an involvement of the right OFC or FFA in sharpening face discrimination on the basis of lateral inhibition.”

– Related, the authors do not really come up with a suggestion how to link the sharpened tuning of the visuo-cortical responses on the one hand and the generalization of the subjective fear experience on the other hand. It is only stated that fear generalization is an "active and multifaceted process". As this is a central aspect of this study and pertains to the functional relevance of the observed enhanced tuning of the SSVEP response, one would like to see a bit more discussion of this discrepancy.

This is indeed a central aspect of our study. We hope that the account above assisted in making explicit the hypothesis that generalization occurs widely in brain and behavior, as established in decades of research, and that sensory areas turn generalization-tuned top-down signals into local sharpening because of their organizational principle of lateral inhibition. This hypothesis is now made more explit in the manuscript. In addition, we extended the paragraph, where we thoroughly discuss the discrepancies between sensory and efferent systems from an evolutionary perspective, by a discussion on the underlying mechanisms of the different output systems based on the existing literature. It reads as follow:

“A large body of work has shown that visuocortical engagement with specific stimulus features varies with the motivational relevance of these features (Bradley, Keil and Lang, 2012). […] This evolutionary interpretation assumes that the constantly changing and diverse environment in which humans find themselves demands and thus favors plastic physiological mechanisms, which may differ between systems in order to optimize functioning (Miskovic and Keil, 2012).”

– For the generalization phase, the GS1-4 faces were treated as a single factor, which did not seem right to the reviewers. These should be separated into 4 different factors, given the authors' assumption that these 4 faces should show different modulations (i.e., lateral inhibition model and also linear decrease model).

Thank you, for pointing out this potential source of confusion. We completely agree that the different generalization stimuli represent separate levels of the analysis, and had entered them accordingly in the previous version. However, to avoid further confusion, we changed the description of the factor on page x from “CS-type (6: CS+ vs GS1-4 vs CS-)“ to “CS-type (6 levels: CS+ vs GS1 vs GS2 vs GS3 vs GS4 vs CS-)”.

– The support for lateral inhibition model only comes from a comparison of a series of different models to a random intercept model. This is weak evidence in favor of a lateral inhibition model. Before doing the modeling, it should first be tested whether the data for the GS1-4 faces differs reliably between each other (see last point), which is doubtful looking at the figure.

Second, a fair modeling approach would be to compare the models against each other (linear vs. lateral inhibition), and not take each model and compare it to the random intercept model, and then pick the model that shows the highest BF. Transitive BFs might be utilized here, accordingly.

Again, we agree that comparisons to the random intercept (Null model), would have resulted in weak evidence, only. However, by comparing every model against the same random intercept model, the resulting BFs share a common denominator. Thus, as mentioned by the reviewers, this allows us to use transitive BFs, which we used in Table 4 (and in the text), and which we now emphasize in the revised version throughout. This is an important prerequisite for consecutive model comparisons. Here, we not only compared models in which we hierarchically included the main and interaction effects of CS-type and social anxiety for each contrast model by dividing the respective BFs, but we also compared the contrast models against each other. The results of these comparisons can be seen in the last two columns of Table 4, showing that the lateral inhibition model is 858 times more likely than the quadratic trend model and 502 times more likely than the linear trend model. To clarify this, we added the following paragraph to the Results section:

“Importantly, by comparing the main and interaction effect models of each contrast against the same random intercept model, it was possible to derive the relative evidence of the lateral inhibition model over the quadratic and linear trend (see Table 4, last two columns). Results for the full interaction model demonstrated that the lateral inhibition model is 858 times more likely than the quadratic trend model and 502 times more likely than the linear trend model.”

– Each of the tested models does not appear to be well justified based on previous literature or any mechanistic accounts of fear conditioning, and it remains a bit unclear why these models were chosen.

We expanded on the background of the models in the Introduction and Discussion. In our study, we compared three different models; the lateral inhibition model, a linear and a quadratic trend model. We decided to use the linear and quadratic trend as comparisons since they are frequently employed in the fear generalization literature, where they are used to indicate an overgeneralization of fear (e.g. Lissek, 2014; Lissek, 2014; Ahrens, 2016). In contrast to the trend analyses in these studies, we explicitly expressed these trends as numeric values, which is mathematically equivalent to the implicit approach. The contrast vectors are also insensitive to any linear transformations, as long as the linear and quadratic properties stay intact. Our main model, the lateral inhibition model was based on previous studies on visuocortical tuning (McTeague, 2015; and Antov et al., 2020), which expressed the lateral inhibition model as a difference of two Gaussian distributions. However, as we have stated in our Materials and methods section, we adapted the lateral inhibition model to only one half of the previously used model, as in our study, generalization learning occurred along a continuum from the threat stimulus (CS+) to the safety stimulus (CS-) and not along a continuum with the CS+ is in the middle. To facilitate model comparisons, the sum of each contrast vectors was rescaled to equal zero. We included the following description with references to the Results section:

“The lateral inhibition pattern was expressed as the difference of two Gaussians (weights: +2, -2, +0.5, +1, +0.5, -2 for CS+, GS1, GS2, GS3, GS4, CS-), paralleling previous studies on visuocortical tuning (Antov et al., 2020; McTeague et al., 2015). The quadratic (weights: +2.5334, +1.0934, -0.0267, -0.8267, -1.3067, -1.4667) and linear (weights: +2.5, +1.5, +0.5, -0.5, -1.5, -2.5) trend were modelled after the analyses of the linear and quadratic component of the generalization gradient, which are commonly employed in the fear generalization literature (Ahrens et al., 2016; Lissek et al., 2014; Lissek et al., 2014).”

And updated the Materials and methods section to:

“To compare the lateral inhibition pattern to a quadratic and linear fit, Bayesian linear models were used. […] The linear model analysis is mathematically insensitive to any linear transformation of the contrast vectors, which is why every vector was centered around zero to facilitate model comparisons.“

– A potentially non-negligible difference (p = 0.07) between CS+ and CS- at baseline was noted, with the same directionality as the significant effect after fear condition. This seems problematic given that the difference remains relatively small after (and during) conditioning (though normalized effect sizes are not reported). We find such a small change extremely difficult to interpret. To be convinced that this is a real change in SNR through conditioning, we would want to see 1) statistics showing a reliable increase of the SSVEP response (so pre- and post training comparisons) which are currency lacking, and 2) effect sizes (at best with CIs for these, J.O.).

You address an important point, which we indeed did not consider earlier. There seem to be two concerns here:

First, do ssVEP amplitudes change from habituation to acquisition? To test for pre and post-training differences, we directly compared the difference between CS+ and CS- in ssVEP-SNRs during habituation versus acquisition as well as during habituation versus generalization with simple t-tests. However, the t-test for CS-differences in habituation compared to acquisition, t(66) = 0.83, p = .409, d = 0.10, CI = [-0.14; 0.34], as well was the t-test for habituation compared to generalization, t(66) = 0.52, p = .606, d = 0.06, CI = [-0.18; 0.30] showed no significant differences. This is to a certain extent a limitation to our study, which we now address in our Discussion. Yet, in the differential conditioning literature, cross-phase comparisons are typically avoided because they confound any conditioning effects with time-dependent effects such as adaptation, in which both the CS+ and CS- evoked responses decline over time, from habituation to acquisition and finally into the extinction phase. This has been well established in meta-analyses (e.g., Fullana et al., 2016), and is true for a wide range of dependent measures, including startle, skin conductance, fMRI, and ssVEPs where this pattern of findings has been noted and systematically addressed several times (e.g., Moratti et al., 2005; Keil et al. 2013). Thus, fear conditioning studies interested in temporal changes previously utilized analysis on a trial-by-trial level in order to avoid confoundation with these types of adaptation processes (e.g. Sjouwerman, Niehaus and Lonsdorf, 2015; Weike, Schupp and Hamm, 2007; Wieser, Miskovic, Rausch and Keil, 2014).

Second, is there a benchmark conditioning effect, given that there are visible differences in the same direction during habituation? To address this, we conducted a post-hoc analysis quantifying the amount of conditioning-related effects above and beyond the initial difference in habituation by means of parametric bootstrapping in combination with Bayesian analyses (Efron, 2012). We computed a distribution for the CS+ versus CS- difference in ssVEP amplitude separately for habituation (to be used as prior distribution) and acqusition (posterior distribution) based on 100000 bootstrapped group mean differences. Odds for conditioning effects to occur were estimated from these two distributions (i.e., the odds of the CS+-CS- difference to be positive), and the BF of interest is given as the ratio of posterior (acquisition) over prior (habituation) odds. The error of this process was estimated by repeating the above process 100 times and measuring the standard error of the resulting mean. This BF corresponds to the change in confidence that the ssVEP for the CS+ is greater than ssVEP for the CS- in acquisition (the posterior), relative to the prior distribution, which was estimated from the CS+ versus CS- difference during habituation. This analysis yielded a BF (posterior odds over prior odds) of 4.77, error=1.3%, suggesting that the acquisition data provided substantial evidence for the notion that fear conditioning selectively amplifies the ssVEP, above and beyond the differences present during habituation. This is now mentioned in the Discussion. Finally, when using face CSs, differences during habituation often indicate different responses to one of the two faces, an explanation which seems unlikely in our study, since the mapping of the two faces to the CS+ and CS- was fully counterbalanced between subjects.

In conclusion, the present data provide evidence for the notion that conditioning selectively heightens the ssVEP evoked by the CS+. This being said, we also emphasize in the manuscript that a lack of significant changes between the experimental phases, which were kept short to allow ample generalization learning, would not affect our interpretation of the main results regarding the generalization phase. We added the following qualifications to our Discussion:

“One important limitation to our findings is that there was no substantial increase in differential ssVEP-SNRs (CS+ vs CS-) from habituation to acquisition, t(66) = 0.83, p = .409, d = 0.10, CI95 = [-0.14, 0.34]. […] However, even though there were reliable differences between CS+ and CS- during acquisition and generalization, future studies may replicate our findings using a different set of stimuli to rule out the influence of baseline differences.”

– The statistical results are in many places interpreted in a seeming all-or-none fashion, which does not seem appropriate. Effects with p-values of e.g. 0.05 are interpreted as true, while effects with p-values of e.g., 0.056 are interpreted as absent. Again, effect sizes should be reported, and a non-dichotomising interpretation of effects is encouraged.

We agree on the importance of effect sizes and their confidence intervals, which is why we already reported R²-values and their 95%-Cis as effect sizes for the discrete factors (CS-type) and betas and their SE as effect sizes for the continuous factors (social anxiety) in the linear mixed model analyses (but please also note this discussion on the issues of effect sizes in linear mixed model analyses: https://stats.stackexchange.com/questions/95054/how-to-get-an-overall-p-value-and-effect-size-for-a-categorical-factor-in-a-mi). To further discourage a dichotomizing interpretation of effects, we rephrased the Results section carefully.

– Related, the ANOVA results for the generalization show, if we understand correctly, a main effect for the CS-type but no main effect or interaction with social anxiety. The Bayesian linear models does find support for a full interaction model including a main effect and interaction with social anxiety. This discrepancy is not really resolved or discussed so far.

This is true and, importantly, the effect of social anxiety received only support for the lateral inhibition model. This discrepancy is not a conflict between the ANOVA and Bayesian-analysis but results from the differences between a non-directional omnibus-test (which the ANOVA is) and the explicit test of specific contrasts (which we did with Bayesian-analyses). While the omnibus-analysis tests the influence of social anxiety in any direction, the contrast analyses tests if social anxiety is associated with responses following the specified pattern only, which strongly increases the statistical power. In line with this, it is important to mention, that social anxiety was not associated with the quadratic or linear trend model. To clarify this matter, we added following to the Discussion:

“Please note that the effect of social anxiety on visuocortical responses was only evident in the Bayesian-analysis but not in the corresponding ANOVA. This discrepancy results from the differences regarding statistical power between omnibus- and contrast-analysis (Furr and Rosenthal, 2003). The ANOVA tests for any differences, while the contrast-analysis only tests for deviations from the specified pattern. This notion is substantiated by the finding that social anxiety was not associated with visuocortical responses for the quadratic or linear trend model, but only for the lateral inhibition model.”

– Reviewers were not unequivocally convinced that the Current source density transforms are useful or adequate for SSVEPs. Please make sure to justify their use better (potentially also demonstrating their effect on overall conclusions/results, e.g. in a supplement). The authors should motivate their choice more strongly and report parameters λ/m.

We have added a paragraph highlighing the benefits of the CSD transformation for reducing unwanted topographical variability between subjects by quantifying source densities. We cite methodological papers and empirical papers using this technique for that purpose. We also report now that we used a λ = .2 for all analyses. The paragraph in the Materials and methods section reads as follows:

“To increase spatial resolution of the EEG signal, we then calculated the current source densities (CSD) of the time-averaged data. The CSD transformation offers the advantage of reducing the negative impact of volume conduction and thereby effectively minimizing unwanted topographical variability between subjects by quantifying source densities (Junghöfer et al., 1997; Kayser and Tenke, 2015). The CSD approach relies on the spatial Laplacian (the second spatial derivative) of the scalp potential to estimate the potential distribution at the cortical surface. Here we used the CSD algorithm described by Junghöfer et al., 1997, with λ = .2, which is well suited for dense-array EEG montages and has been used in previous studies investigating ssVEP responses to facial stimuli (McTeague et al., 2011) or during fear generalization (McTeague et al., 2015).”

– The authors have added an additional analysis that relates a measure of visuo-cortical tuning (i.e. the individual sharpening of response profiles after fear conditioning) to a measure of social anxiety. Indeed, they find a moderate positive correlation, strengthening their point that the individual sensitivity to threat cues can affect visuo-cortical tuning. Computing individual cortical tuning as weighted sum of SSVEP and model weights seems elegant. However, the reader is somewhat left at the authors' mercy to confide in the sensitivity of this measure. We strongly suggest to repeat this analysis with the weights of the Quadratic and Linear trend models and compare them with the lateral inhibition situation to show that there is no correlation in these cases.

We would like to thank the reviewers for this suggestion. We now have calculated the scalar product for the quadratic and linear trend models and added the corresponding regression analyses to the Results section. We also updated the figure accordingly. The text now reads as follows:

“In a first step, we calculated a visuocortical tuning index on subject level, which is the scalar product of the weights of the lateral inhibition model (2, -2, 0.5, 1, 0.5, -2) and the respective individual ssVEP responses. […] The regression analysis revealed a moderate, positive correlation between social anxiety and the visuocortical tuning index, r(65) = .288, p = .018, BF1/0 = 3.65, confirming that the accentuation of the lateral inhibition pattern increased with higher social anxiety, while the weak correlations between social anxiety and the index for the quadratic, r(65) = .146, p = .238, BF2/0 = 0.52, or linear trend, r(65) = .137, p = .268, BF3/0 = 0.49, did not reach the frequentist significance threshold and did not receive support from Bayesian analyses.”

– The manuscript is generally well-written and easy to read. The Results section however suffers from heavy in-text reporting of statistics. Reading flow will be improved by reporting stats in tables instead. Also consider the use of supplementary tables.

We agree and have moved the statistics of all omnibus tests from the text to tables. We accordingly revised the Results section.

[Editors' note: further revisions were suggested prior to acceptance, as described below.]

Requested revisions:

1) One reviewer alerted us that one issue raised in the initial round of reviews was not addressed, the potentially quite important pre-post comparison for SSVEP amplitude (i.e., pre-and post-training). In our view, it would be an important analysis to ask if there is an effect over time. We think proper reporting (and discussion) of such a pre-post comparison is important for making inferences that these results point to a change in neural tuning due to habituation.

Thank you for bringing this issue to our attention. We can only assume that we failed to explain and appropriately highlight the pre-post comparisons that we included in the revised version (t-tests and a directed Bayesian bootstrapped test that directly tests the conditioning effects on SSVEP amplitude pre- versus post-conditioning). To avoid confusion, in the last response letter, we reported and discussed pre-post training comparisons analyzed with t-tests (including effect sizes and CIs) and conducted an additional bootstrap-based Bayesian test to examine conditioning-related changes above and beyond the initial difference in habituation. As a reminder, the BF for a pre-post conditioning effect (CS+ minus CS-) was 4.7, but some of the direct amplitude comparisons of the same conditions from pre to post, did not reach the level of NHST significance, with small effect sizes, prompting us to add two paragraphs of discussing this as a limitation. To make this comparison more obvious and make sure that readers will find this important information in the manuscript, we added a subheading, and added language that highlights these tests.

2) The other remaining issue if of lesser concern:

You have not done a direct comparison of the models to each other. Please consider reporting these comparisons. The reviewer provides the following guidance: "While I follow that there is some benefit in comparing all models against a random intercept model, I would still want to see whether the differences between the GS1-4 faces differs reliably from each other (follow-up paired t-test)."

This comment seems to raise two different issues. First, how does the lateral inhibition model perform against the quadratic or linear trend model? This being an important point, we already had addressed this question in the original manuscript by not only comparing each model against the random intercept model, but we also directly compared the different models to each other. We also elaborated on this approach in the initial round of revisions (see Comment #5). Again, we can only assume that this information is not obvious enough in its presentation, so we amplified its visibility by adding additional sentences explaining and highlighting them.

Second, we indeed did not report follow-up contrasts for ssVEP amplitudes, since we addressed the follow-up analysis with the more powerful, and in our view more elegant and theory-driven, model comparisons. In response to the reviewer’s concern and, to keep reporting of the ssVEP results consistent with the results for ratings, we now report post-hoc contrasts (t-tests, as requested) for the linear mixed model analysis of ssVEP amplitudes. In line with the fear generalization literature, all contrasts were referenced to the conditioned safety cue (Lissek et al., 2008; Lissek et al., 2010). Please note that we are not able to calculate standardized effect sizes for contrasts in linear mixed model analysis. We added the following paragraph to the ssVEP Results section:

“Post-hoc contrasts indicated significant differences between GS4 vs CS-, t(325)=2.94, p=.003, GS3 vs CS-, t(325)=3.24, p=.001, and CS+ vs CS-, t(325)=2.17, p=.031, but not between GS2 vs CS-, t(325)=1.75, p=.081, and GS1 vs CS-, t(325)=0.82, p=.411.”

https://doi.org/10.7554/eLife.55204.sa2

Article and author information

Author details

  1. Yannik Stegmann

    Department of Psychology (Biological Psychology, Clinical Psychology, and Psychotherapy), University of Würzburg, Würzburg, Germany
    Contribution
    Data curation, Software, Formal analysis, Visualization, Writing - original draft, Writing - review and editing
    For correspondence
    yannik.stegmann@uni-wuerzburg.de
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0002-0933-8492
  2. Lea Ahrens

    Department of Psychology (Biological Psychology, Clinical Psychology, and Psychotherapy), University of Würzburg, Würzburg, Germany
    Contribution
    Conceptualization, Data curation, Investigation, Methodology
    Competing interests
    No competing interests declared
  3. Paul Pauli

    1. Department of Psychology (Biological Psychology, Clinical Psychology, and Psychotherapy), University of Würzburg, Würzburg, Germany
    2. Center for Mental Health, Medical Faculty, University of Würzburg, Würzburg, Germany
    Contribution
    Conceptualization, Supervision, Funding acquisition, Visualization, Methodology, Project administration, Writing - review and editing
    Competing interests
    No competing interests declared
  4. Andreas Keil

    Center for the Study of Emotion and Attention, University of Florida, Gainesville, United States
    Contribution
    Data curation, Software, Supervision, Visualization, Methodology, Writing - review and editing
    Competing interests
    No competing interests declared
  5. Matthias J Wieser

    1. Department of Psychology (Biological Psychology, Clinical Psychology, and Psychotherapy), University of Würzburg, Würzburg, Germany
    2. Department of Psychology, Education, and Child Studies, Erasmus University Rotterdam, Rotterdam, Netherlands
    Contribution
    Conceptualization, Resources, Data curation, Software, Formal analysis, Supervision, Funding acquisition, Methodology, Project administration, Writing - review and editing
    Competing interests
    No competing interests declared

Funding

Deutsche Forschungsgemeinschaft (44541416 - TRR-58)

  • Paul Pauli
  • Matthias J Wieser

The funders had no role in study design, data collection and interpretation, or the decision to submit the work for publication.

Acknowledgements

This work was supported by the German Research Foundation – project number 44541416 - TRR-58, projects B05 of the 2nd funding period and B01 of the 3rd funding period

Ethics

Human subjects: Prior to participation, written informed consent was obtained from each participant. The study was approved by the ethics review board of the University of Würzburg (87/13).

Senior Editor

  1. Richard B Ivry, University of California, Berkeley, United States

Reviewing Editor

  1. Jonas Obleser, University of Lübeck, Germany

Reviewers

  1. Christian Keitel, University of Stirling, United Kingdom
  2. Ulrike M Krämer

Publication history

  1. Received: January 16, 2020
  2. Accepted: June 8, 2020
  3. Accepted Manuscript published: June 9, 2020 (version 1)
  4. Version of Record published: June 23, 2020 (version 2)

Copyright

© 2020, Stegmann et al.

This article is distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use and redistribution provided that the original author and source are credited.

Metrics

  • 543
    Page views
  • 100
    Downloads
  • 0
    Citations

Article citation count generated by polling the highest count across the following sources: Crossref, PubMed Central, Scopus.

Download links

A two-part list of links to download the article, or parts of the article, in various formats.

Downloads (link to download the article as PDF)

Download citations (links to download the citations from this article in formats compatible with various reference manager tools)

Open citations (links to open the citations from this article in various online reference manager services)

Further reading

    1. Computational and Systems Biology
    2. Neuroscience
    Chen Chen et al.
    Research Article

    While animals track or search for targets, sensory organs make small unexplained movements on top of the primary task-related motions. While multiple theories for these movements exist—in that they support infotaxis, gain adaptation, spectral whitening, and high-pass filtering—predicted trajectories show poor fit to measured trajectories. We propose a new theory for these movements called energy-constrained proportional betting, where the probability of moving to a location is proportional to an expectation of how informative it will be balanced against the movement’s predicted energetic cost. Trajectories generated in this way show good agreement with measured trajectories of fish tracking an object using electrosense, a mammal and an insect localizing an odor source, and a moth tracking a flower using vision. Our theory unifies the metabolic cost of motion with information theory. It predicts sense organ movements in animals and can prescribe sensor motion for robots to enhance performance.

    1. Ecology
    2. Neuroscience
    Felix JH Hol et al.
    Tools and Resources

    Female mosquitoes need a blood meal to reproduce, and in obtaining this essential nutrient they transmit deadly pathogens. Although crucial for the spread of mosquito-borne diseases, blood feeding remains poorly understood due to technological limitations. Indeed, studies often expose human subjects to assess biting behavior. Here, we present the biteOscope, a device that attracts mosquitoes to a host mimic which they bite to obtain an artificial blood meal. The host mimic is transparent, allowing high-resolution imaging of the feeding mosquito. Using machine learning we extract detailed behavioral statistics describing the locomotion, pose, biting, and feeding dynamics of Aedes aegypti, Aedes albopictus, Anopheles stephensi, and Anopheles coluzzii. In addition to characterizing behavioral patterns, we discover that the common insect repellent DEET repels Anopheles coluzzii upon contact with their legs. The biteOscope provides a new perspective on mosquito blood feeding, enabling the high-throughput quantitative characterization of this lethal behavior.