Autism spectrum disorder (ASD) is a broad group of disorders, characterized by impairments in communication and social awareness, and by repetitive stereotyped behaviors (American Psychiatric Association, 2013). In addition to these well-known impairments, there are some differences in perceptual functioning, particularly in the visual domain (Dakin and Frith, 2005; Kellerman et al., 2005; Pellicano et al., 2005; Dale and Salt, 2008; McCleery et al., 2009; Simmons et al., 2009). These perceptual differences extend to the neurotypical population (Almeida et al., 2010; Sutherland and Crewther, 2010; Jackson et al., 2013) for individuals with high scores on the Autism Spectrum Quotient (AQ) personality scale (Baron-Cohen et al., 2001). This suggests that the underlying physiology of ASD may be distributed throughout a broader autistic phenotype, which contributes to visual differences in these individuals (Bailey et al., 1995; Braddick et al., 2003).

Detection of potentially threatening stimuli, such as fearful or angry faces involves a distributed network of cortical and subcortical regions (LeDoux, 1998; Johnson, 2005; Kragel et al., 2016). Several researchers have proposed that rapid processing of threat occurs via a direct, subcortical route from the pulvinar and superior colliculus to the amygdala (LeDoux, 1998; Morris et al., 1998a; Öhman, 2005). Direct evidence of a subcortical (retino-collicular-pulvinar) route to the amygdala has come from animal studies (LeDoux, 1998), whereas functional evidence in humans has been inferred based on evidence of emotional face processing in cortically blind participants (Morris et al., 2001), unconscious processing of emotional stimuli (Morris et al., 1998a), and magnetoencephalographic (MEG) responses to fearful stimuli in the amygdala and subcortical structures (Streit et al., 2003). Clear anatomical evidence from in vivo DTI tractography conducted in humans demonstrates connections between the amygdala and superior colliculus (Rafal et al., 2015); and amygdala and pulvinar (Tamietto et al., 2012). In support of this model, connectivity analyses have shown that fearful stimuli engage a distributed network of brain sites including the bilateral fusiform gyrus, dorsal/anterior precuneus, amygdala, hippocampus, and parahippocampal regions (Kragel and LaBar, 2015).

In order to compare cortical and subcortical routes for affective information processing, several researchers have taken advantage of the tuning properties of cells in the subcortical structures (Vuilleumier et al., 2003; Öhman, 2005; Vlamings et al., 2009). Cells in the input layers of the superior colliculus receive input from the magnocellular pathway (Schiller et al., 1979). Magnocellular neurons have relatively large receptive fields and respond preferentially to low spatial frequency (LSF) input, compared with parvocellular neurons which have smaller receptive fields and respond preferentially to high spatial frequency (HSF) input (Livingstone and Hubel, 1988). Processing of coarse information contained in the LPFs precedes processing of more detailed information contained in the HSFs, with magnocellular input reaching V1 25–31 ms earlier than parvocellular input (Klistorner et al., 1997; Bullier, 2001; Bar et al., 2006; Sutherland and Crewther, 2010; Crewther et al., 2016). This temporal precedence has been termed the ‘magnocellular advantage’ (Laycock et al., 2007).

In support of a magnocellular route for fearful face processing, Vuilleumier et al. (2003) found enhanced BOLD responses in the superior colliculus and amygdala structures for LSF, but not HSF, fearful expressions. However, the amygdala also receives input from the anterior inferotemporal cortex (Aggleton, 1993), so it is not ‘blind’ to HSF fearful face representations. In fact, the amygdala is likely to have numerous connections with subcortical and cortical structures, which process and integrate both LSF and HSF, to form a ‘whole’ fearful face percept. Evidence for multiple pathways comes from connectivity analyses demonstrating that fearful stimuli engage a distributed network of brain sites including the bilateral fusiform gyrus, dorsal/anterior precuneus, amygdala, hippocampus, and parahippocampal regions (Kragel and LaBar, 2015). In addition, evidence for multiple pathways is presented by one fMRI study demonstrating that during unconscious processing using binocular rivalry and backward masking, when visual input must be processed via a subcortical route, the amygdala has limited capacity to differentiate between facial emotions (Morris et al., 1998b; Williams et al., 2004).

Sets of hybrid face stimuli have been used in numerous studies in order to investigate the relative contributions of LSF and HSF information to various aspects of emotional face processing, while controlling the natural variation of spatial frequencies within the images (Pourtois et al., 2005; Laeng et al., 2010, 2013; Prete et al., 2014, 2015a,b,c, 2016). LSF information generally includes coarse, global visual features whereas HSF information includes finer, more detailed features (De Valois and De Valois, 1988). Hybrid face stimuli are created from different combinations of the LSF and HSF filtered components of neutral and emotional faces. By equating the luminance and contrast of LSF and HSF within one image, only the change in emotional expression of the face remains (Pourtois et al., 2005). Hybrid faces have been used to assess conscious report of fearful face perception, where HSF fearful and neutral expressions are more rapidly discriminated than LSF expressions (Stein et al., 2014). However, behavioral report does not necessarily reflect the rapid, neural processing that occurs prior to conscious perception of a fearful face, which can be investigated with electroencephalogram (EEG). For consistency, throughout this paper we will refer to hybrids created from LSF and HSF, fearful (F) and neutral (N) face stimuli as follows: F_LSF–F_HSF, F_LSF–N_HSF, N_LSF–F_HSF, and N_LSF–N_HSF.

To elucidate the time course of rapid fear perception, electrophysiological studies have primarily focused on the P100 and N170 waveforms. The visual P100 is a fast response, typically evoked between 90 and 140 ms, that is maximal over lateral–occipital–parietal sites, and appears to originate from striate and extrastriate neural generators (Clark and Hillyard, 1996; Mangun et al., 1997; Allison et al., 1999). Cueing studies suggest the P100 amplitude is an index of the attentional gain mechanisms that suppress responses to irrelevant stimuli (Hillyard and Anllo-Vento, 1998). In addition, the P100 is sensitive to affect, with greater amplitude responses to fearful face presentation, compared to other emotions including neutral, happy, angry, sad, disgust, and surprise faces (Pizzagalli et al., 1999; Batty and Taylor, 2003; Pourtois et al., 2004; Magnée et al., 2008; Feng et al., 2009; Forscher and Li, 2012; Meaux et al., 2013; Smith et al., 2013; Zhang et al., 2013). The N170 is negative peak that occurs approximately 170 ms post-stimulus, originating from a network of face/object processing regions including the fusiform gyrus, superior temporal sulcus and inferior, middle and superior temporal gyri (Henson et al., 2003). N170 amplitude is sensitive to configural processing of faces, as evidenced by higher amplitude response to upright over inverted faces (Bentin et al., 1996). Affective input also modulates these early waveforms, with greater amplitude P100 and N170 responses to fearful or angry expressions (Batty and Taylor, 2003; Pegna et al., 2008).

Several studies have investigated whether fearful face modulation of visual ERPs relies on LSF (i.e., magnocellular) input. The P100 amplitude is enhanced for F_LSF–N_HSF hybrid faces compared to N_LSF–N_HSF hybrid faces in the right hemisphere, but not for N_LSF–F_HSF hybrids (Pourtois et al., 2005); however, this effect is only observed when the images have been equated for luminance and contrast (Vlamings et al., 2009). In addition, Vlamings et al. (2009) found shorter P100 latency for F_LSF–N_HSF compared to N_LSF–F_HSF hybrids, which trended toward significance in the fearful expression condition. Findings regarding the effects of emotion and spatial frequency on the N170 amplitude have been mixed. Some studies reported no effect of emotion on the N170 response (Holmes et al., 2005; Pourtois et al., 2005), however, Vlamings et al. (2009) reported N170 enhancement for F_LSF–N_HSF but not N_LSF–F_HSF faces, regardless of whether the stimuli were equated for luminance and contrast.

Early electrophysiological responses to fearful faces vary across the autistic spectrum, indicating there are individual differences in the neural pathways that produce these responses. In autistic individuals, the P100 tends to be delayed, reduced in amplitude and less lateralized than in neurotypical groups (Bailey et al., 2005; Wong et al., 2008; McCleery et al., 2009; Luo et al., 2010; Batty et al., 2011; Fujita et al., 2013; Tye et al., 2013; Wagner et al., 2013; Key and Corbett, 2014; Bonnard-Couton et al., 2015; Lassalle and Itier, 2015; Anzures et al., 2016). In an investigation of face processing in the broader autistic phenotype, Stavropoulos et al. (2016) found that P100 and N170 latencies tended to be slower and decreased for people with high AQ scores, compared to people with low AQ scores. They found that non-consciously perceived emotional faces elicited enhanced neural responses regardless of AQ score, yet they did not use hybrid faces so it is unclear whether the two groups utilized LSF information in the same way. De Jong et al. (2008) investigated the effects of gaze cueing on ERPs and found that for neurotypical observers, gaze cueing had a stronger effect on ERPs for LSF filtered faces; whereas for autistic observers, cueing effects were stronger for HSF filtered faces. These findings indicate that the P100 and N170 responses to fearful emotional faces may be particularly affected in individuals with high autistic tendency. Therefore, we aimed to compare the effects of LSF and HSF fearful expressions on ERP responses for high and low AQ groups.

Several visual studies provide evidence for magnocellular/dorsal stream abnormalities in ASD (Pellicano et al., 2005; Milne et al., 2006; McCleery et al., 2009) and in the broader autistic phenotype (Almeida et al., 2010; Sutherland and Crewther, 2010; Jackson et al., 2013; Thompson et al., 2015; Stavropoulos et al., 2016). Other studies have indicated that the relationship between magnocellular function and ASD is more complicated (Bertone et al., 2003; Del Viva et al., 2006). The current view is that ASD involves complex interactions between multiple visual pathways, rather than a specific magnocellular/dorsal stream impairment (Simmons et al., 2009; Thye et al., 2017). Despite the nature of magnocellular impairment remaining unresolved, there are differences in the rapid output of magnocellular neurons to V1 in high autistic tendency (Sutherland and Crewther, 2010; Jackson et al., 2013). Moreover, as discussed above, magnocellular neurons have a preference for LSF input, resulting in rapid V1 activation. Neural efficiency of magnocellular neurons can be studied through Weiner kernel analysis of multifocal visual evoked potentials (mfVEP) (Sutherland and Crewther, 2010; Jackson et al., 2013). The first slice of the second order, non-linear VEP kernel (K2.1) measures non-linearity in neural recovery (referenced one video frame back). Studies of the contrast response function indicate that the N60 K2.1 waveform, and early components (with the same latency) of the K2.2 waveform (second slice- referenced two frames back) are of magnocellular origin (Baseler and Sutter, 1997). Higher amplitudes of these waveforms indicate a greater degree of inefficiency in neural recovery by magnocellular neurons. The amplitude of the N60 negativity in the first slice of the second order kernel (K2.1) is elevated in high AQ observers, indicative of poor magnocellular recovery rate (Jackson et al., 2013). Hence, Jackson et al.’s (2013) findings are consistent with the lack of efficiency of the magnocellular pathway in ASD.

Based on prior research by Vlamings et al. (2009) we hypothesized that the low AQ group would exhibit significantly higher amplitude and shorter latency P100 and N170 responses to F_HSF–F_LSF hybrids compared to N_HSF–N_LSF hybrids, and to F_LSF–N_HSF compared to F_HSF–N_LSF hybrids. Based on the findings by Stavropoulos et al. (2016), it was hypothesized that the P100 and N170 would be lower for the high AQ group than for the low AQ group. Furthermore, we predicted that F_LSF–N_HSF expressions would have a greater influence on response amplitudes in the low AQ group than in the high AQ group. We aimed to use non-linear VEP to assess magnocellular recovery, as indexed by the amplitude of the K2.1 waveform. It was hypothesized, the K2.1 response amplitude would be higher in the high AQ group, which is indicative of inefficient recovery rate of neurons within the magnocellular system.

Grand mean ERP topographies and waveforms for the low and high AQ groups are presented in Figure 3. The P100 topographies for the low and high AQ groups (Figures 3A,C, respectively) show there was a bilateral, occipital positivity at 115 ms latency for each of the hybrid face conditions. The N170 topographies for the low and high AQ groups (Figures 3D,G, respectively) show there was a right lateralized negativity, with strongest activations at electrode sites P8 and PO8, for each of the hybrid face conditions. Means and standard deviations for the low and high AQ groups are presented in Figure 4, with comparisons for P100 amplitude and latency, N170 amplitude and latency, and behavioral performance. As explained in the analyses section, we defined ERP amplitude as the area under the waveform (peak or trough latency ± 30 ms); however, it should be noted that we observed a similar pattern of results when we compared the waveform areas or the peak voltages across hybrid face conditions and AQ groups.

Here, we conducted two investigations in groups with low and high autistic tendency recruited from within the neurotypical population. Firstly, we asked whether fearful expression affects the P100 and N170 ERP responses in the same way for low and high AQ groups. As expected, fearful expression modulated P100 amplitude in the low AQ group, but not in the high AQ group. In the low AQ group, P100 amplitudes were significantly greater for F_LSF–F_HSF faces than for N_LSF–N_HSF faces. Interestingly, while P100 amplitude was significantly greater with F_LSF–F_HSF than N_LSF–F_HSF faces_, it was not significantly different with F_LSF–F_HSF and F_LSF–N_HSF faces. This suggests that the effect of fearful expression on producing greater P100 amplitude is mostly via fearful expression being carried by LSF. By contrast, in the high AQ group there was an overall reduction in P100 and N170 amplitude and there was no effect of fearful expression on ERP amplitudes. P100 and N170 latencies were faster for the F_LSF–F_HSF hybrids than for the other conditions, but there were no significant latency differences between the two groups. We used Wiener kernel analysis of the visual evoked potential to investigate whether differences between the low and high AQ groups could reflect differences in magnocellular processing. The K2.1 non-linear VEP response amplitude was higher in the high versus low AQ group, which suggests inefficient neural recovery within the magnocellular stream (Sutherland and Crewther, 2010; Jackson et al., 2013).

We demonstrated that for people with low levels of autistic personality traits, fearful emotion tends to affect the P100 amplitude but not the N170 amplitude. Differences in the effects of our hybrid emotional stimuli on the P100 and N170 can be interpreted in terms of what we know about these early potentials. The literature suggests that P100 amplitude modulation reflects both rapid orienting to salient information (Hillyard and Anllo-Vento, 1998) and rapid identification of threat-related input, such as fearful facial expression (Batty and Taylor, 2003; Pourtois et al., 2004; Magnée et al., 2008; Feng et al., 2009; Luo et al., 2010; Forscher and Li, 2012; Meaux et al., 2013; Zhang et al., 2013). Consistent with this literature, our findings suggest that differences in facial emotion can be processed as early as the P100, but only in groups with low levels of autistic tendency. The N170 amplitude is sensitive to configural processing of faces, yet there are mixed findings as to whether it is influenced by facial emotion (Holmes et al., 2005; Pourtois et al., 2005). Vlamings et al. (2009) found increased N170 amplitude responses to fearful hybrids compared to neutral hybrids, yet in our experiment, these differences were not observed for either AQ group. On the contrary, we did observe some latency differences for the whole sample, with faster N170 responses for F_LSF–F_HSF hybrids than for the other hybrid conditions. The slower N170 responses for the hybrids with mixed expressions (i.e., F_LSF–N_HSF and N_LSF–F_HSF) might reflect disruptions in configural processing of these stimuli (Bentin et al., 1996).

Consistent with Stavropoulos et al. (2016), the P100 and N170 amplitudes were lower in the high AQ group than in the low AQ group, yet contrary to their findings, we did not observe any clear P100 or N170 latency differences between the two AQ groups. There are several differences between our experiments that may explain these differences in results. First of all, we used hybrid face stimuli, whereas they used unaltered (i.e., broadband spatial frequency) face stimuli. Secondly, they observed the greatest between-groups differences when the stimuli were presented non-consciously (16 ms presentation time), whereas our stimuli were presented for 500 ms. Finally, their high AQ sample had very low amplitude/high noise ERP results. This may imply that their high AQ sample had greater face processing impairments than our high AQ sample. Given that we recorded clear P100 and N170 responses from both groups, we were able to make meaningful comparisons for the effects of facial emotion on ERP responses in the low and high AQ neurotypical groups.

Differences in the effects of facial emotion on P100 responses for the low and high AQ groups may reflect underlying differences in magnocellular function. In the low AQ group, P100 amplitudes were greater for F_LSF–F_HSF than N_LSF–F_HSF faces_, yet there was no difference in P100 amplitude for F_LSF–F_HSF and F_LSF–N_HSF faces. This is consistent with evidence from Vlamings et al. (2009) that LSF input contributes to rapid detection of fearful expression. Recent studies have demonstrated magnocellular projections from the pulvinar to the orbitofrontal cortex, which allow for rapid feedback to bias visual processing toward behaviorally relevant stimuli (Bar et al., 2006; Kveraga et al., 2007). Given the spatial frequency preferences of the magnocellular and parvocellular pathways (Livingstone and Hubel, 1988), our results are consistent with evidence that the effects of fearful expression on early visual processing are likely to be carried by the magnocellular pathway. The K2.1 non-linear VEP response amplitude was higher in the high AQ group, which suggests inefficient neural recovery within the magnocellular stream (Sutherland and Crewther, 2010; Jackson et al., 2013).

In addition, fearful expression did not tend to affect P100 amplitudes in the high AQ group in any hybrid condition, compared to the low AQ group, which demonstrated sensitivity to hybrid conditions. This suggests that even within the neurotypical population, very early processing of fear-related input varies for individuals with different levels of autistic personality traits. While the exact nature of magnocellular impairment in high autistic tendency individuals remains unresolved (Simmons et al., 2009; Thye et al., 2017); the differences in early ERP responses to fearful faces that we observed for low and high AQ groups could reflect a magnocellular difference in processing LSF facial emotion.

Our results for the non-linear VEP analysis are consistent with magnocellular differences between the low and high AQ groups. Previous investigations of the contrast-response functions for non-linear VEP components (Klistorner et al., 1997; Jackson et al., 2013) indicate that the K2.1 and early components of the K2.2 waveform are of magnocellular origin, whereas the later K2.2 components are of parvocellular origin. Our results are consistent with previous findings that magnocellularly driven VEP non-linearities tend to be greater in groups with high AQ, but parvocellularly driven VEP non-linearities tend to be similar for both groups (Jackson et al., 2013). This implies that the magnocellular pathway recovers less efficiently from rapid stimulation in individuals with high AQ than in individuals with low AQ. Our results suggest that magnocellular projections, which would normally enable rapid detection of threatening stimuli, are less efficient within the broader autistic phenotype. This may contribute to explaining why the high AQ group produced reduced P100 ERP responses to all hybrid face conditions.

The results of our behavioral task are seemingly at odds with our ERP results. Both high and low AQ groups performed similarly in detecting fearful expression in the F_LSF–F_HSF, F_LSF–N_HSF, N_LSF–F_HSF, and N_LSF–N_HSF hybrid faces. More interestingly, our results indicated that both AQ groups relied upon HSF information to recognize fearful expressions. This HSF preference, however, was elicited in a task that required a 500 ms wait time before response. The conscious detection and report of fearful expression within a hybrid face, as described in prior studies of conscious perception, has demonstrated reliance on HSF facial information (Williams et al., 2004). However, these behavioral measures are not sensitive to very early visual processing stages that precede conscious awareness. We demonstrated differences in ERP responses to fearful and neutral face stimuli as early as 100 ms post-presentation in the low AQ group, but not in the high AQ group. Our non-linear VEP experiment found no significant between-group differences in the parvocellularly driven K2.2 N95 or P150 waveforms, which suggests the differences in the fearful face P100 are magnocellularly driven.

In conclusion, for observers with low AQ, we found that fearful expression enhances rapid electrophysiological responses to faces. Our results suggest that this enhancement is likely to occur via rapidly processed, magnocellular input. For observers with high AQ, we did not find any difference in ERP responses to fearful and neutral faces. Consistent with previous studies (Jackson et al., 2013), our high AQ group showed abnormal temporal processing in the magnocellular pathway. These results support the notion that autism involves differences in processing of LSF information. Our results suggest that magnocellular projections, which would normally enable rapid detection of threatening stimuli, are not utilized efficiently for those with higher autistic tendency. Hence, visual processing differences may underlie some of the socio-cognitive aspects of autism.

AB contributed to this manuscript, including creating the first experimental design, conducting literature review, participant testing sessions, EEG and behavioral analysis and the preparation of this manuscript. LH contributed this manuscript by creating the second experimental design, programming analysis, EEG analysis of second experiment, creating manuscript figures, manuscript section for second experiment and manuscript editing. TF-B contributed by assisting in research participant testing sessions. DC contributed by supervising second experiment design and manuscript editing.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.