1. Human Biology and Medicine
  2. Neuroscience
Download icon

Default mode-visual network hypoconnectivity in an autism subtype with pronounced social visual engagement difficulties

  1. Michael V Lombardo  Is a corresponding author
  2. Lisa Eyler
  3. Adrienne Moore
  4. Michael Datko
  5. Cynthia Carter Barnes
  6. Debra Cha
  7. Eric Courchesne
  8. Karen Pierce  Is a corresponding author
  1. Center for Neuroscience and Cognitive Systems @UniTn, Istituto Italiano di Tecnologia, Italy
  2. University of Cambridge, United Kingdom
  3. University of California, San Diego, United States
  4. VISN 22 Mental Illness Research, Education, and Clinical Center, VA San Diego Healthcare System, United States
Research Article
  • Cited 2
  • Views 855
  • Annotations
Cite this article as: eLife 2019;8:e47427 doi: 10.7554/eLife.47427

Abstract

Social visual engagement difficulties are hallmark early signs of autism (ASD) and are easily quantified using eye tracking methods. However, it is unclear how these difficulties are linked to atypical early functional brain organization in ASD. With resting state fMRI data in a large sample of ASD toddlers and other non-ASD comparison groups, we find ASD-related functional hypoconnnectivity between ‘social brain’ circuitry such as the default mode network (DMN) and visual and attention networks. An eye tracking-identified ASD subtype with pronounced early social visual engagement difficulties (GeoPref ASD) is characterized by marked DMN-occipito-temporal cortex (OTC) hypoconnectivity. Increased DMN-OTC hypoconnectivity is also related to increased severity of social-communication difficulties, but only in GeoPref ASD. Early and pronounced social-visual circuit hypoconnectivity is a key underlying neurobiological feature describing GeoPref ASD and may be critical for future social-communicative development and represent new treatment targets for early intervention in these individuals.

eLife digest

Many parents of children with autism spectrum disorder (ASD) spot the first signs when their child is still a toddler, by noticing that their child is less interested than other toddlers in people and in social play. These early differences in behavior can have long-term implications for brain development. The brains of toddlers with little interest in social stimuli will receive less social input than those of other toddlers. This will make it even harder for the brain to develop the circuits required to support social skills.

But even among children with ASD, there are large differences in children's interest in the social world. One way of measuring these differences is to track eye movements. Lombardo et al. presented toddlers with and without ASD with images of moving colorful geometric shapes next to videos of dancing children. The majority of toddlers, including most of those with ASD, spent more time looking at the children than the shapes. But about 20% of the toddlers with ASD spent most of their time looking at the shapes. These toddlers also had the most severe social symptoms.

To find out why, Lombardo et al. measured the toddlers' brain activity while they slept. During sleep, or when at rest, the brain shows stereotyped patterns of activity. Groups of brain regions that work together – such as those involved in vision – fire in synchrony. Lombardo et al. found that toddlers who preferred looking at shapes over people showed different patterns of brain activity while asleep compared to other children. In the toddlers who preferred shapes, brain networks involved in social skills were less likely to coordinate their activity with networks that support vision and attention.

These findings suggest there may be multiple subtypes of ASD, with different symptoms resulting from different patterns of brain activity. At present, all children who receive a diagnosis of ASD receive much the same behavioral therapy. But in the future, studies of brain networks could allow children to receive more specific diagnoses. This could in turn lead to more effective and personalized treatments.

Introduction

Social visual engagement difficulties, defined as lack of preference for social stimuli often combined with a strong preference for and attention towards non-social stimuli (Dawson et al., 1998; Dawson et al., 2004; Klin et al., 2015; Chawarska et al., 2013; Falck-Ytter et al., 2013; Klin et al., 2009; Nakano et al., 2010; Shic et al., 2011; von Hofsten et al., 2009; Pierce et al., 2011a; Pierce et al., 2016a; Jones and Klin, 2013; Falck-Ytter et al., 2018), are key early developmental features of autism spectrum disorders (ASD). These difficulties are central in early ASD screening and diagnostic tools (Lord et al., 2000; Pierce et al., 2011b; Robins et al., 2001). A child’s preferences and attention early in life can potentially have large impact on future development and outcome (Klin et al., 2015). Reduced social visual engagement behaviors actively select and/or neglect specific types of information from the environment as input to the developing brain. A continual stream of atypical non-social input due to reduced social visual engagement in ASD may have detrimental impact on experience-expectant and experience-dependent processes (Greenough et al., 1987; Holtmaat and Svoboda, 2009; Huttenlocher, 2002) that help to sculpt functional specialization and development in the social brain (Klin et al., 2015; Klin et al., 2003; Mundy et al., 2009; Johnson et al., 2015; Johnson, 2017). These features are also highly relevant for early intervention. Many early interventions that show success for some individuals (Bacon et al., 2014; Dawson et al., 2010; Kasari et al., 2006; Pickles et al., 2016) hinge critically upon the idea of changing this attribute of early ASD development. The hope is that early intervention will increase engagement between the child and the social world and enable experience-dependent neuroplasticity to divert a child towards more typical developmental trajectories (Dawson, 2008).

Eye tracking studies of children and adults with ASD have been central in quantifying deficits in social visual engagement (Klin et al., 2009; Nakano et al., 2010; Shic et al., 2011; von Hofsten et al., 2009; Pierce et al., 2011a; Pierce et al., 2016a; Jones and Klin, 2013; Falck-Ytter et al., 2018). Nonetheless, considerable heterogeneity exists across ASD individuals and the early-age neural bases explaining such features and their developmental variability are not well understood (Chita-Tegmark, 2016; Guillon et al., 2014). Disentangling early-age heterogeneity is the foundation for making significant progress towards the goals of stratified psychiatry and precision medicine (Collins and Varmus, 2015; Kapur et al., 2012; Lai et al., 2013; Lombardo et al., 2019) for ASD and is therefore one of the biggest challenges in the field. Attempts to identify strong neural underpinnings behind ASD are hindered by mixing potentially different ASD subtypes with different biology (Lombardo et al., 2015; Lombardo et al., 2018a). Furthermore, the clinical relevance of parsing heterogeneity into discrete subtypes is also highly salient - not all individuals take the same developmental path or have similar outcomes (Lord et al., 2015) and individuals may vary considerably in responsiveness to early intervention (Bacon et al., 2014).

We recently discovered (Pierce et al., 2011a) and then replicated (Pierce et al., 2016a) the finding of one such clinically relevant subtype with marked lack of early social visual engagement. This subtype can be identified and the deficit objectively quantified with a novel eye tracking preferential looking paradigm, the GeoPref Test (Figure 1A). The paradigm displays dynamic non-social colorful geometric patterns side-by-side with social images of happy children in motion. A subset of toddlers comprising around 20% of the ASD population, spend less than 30% of task time looking at social displays and instead prefer to attend to geometric patterns 70–100% of the task time. This degree of marked lack of social visual engagement is seldom seen in non-ASD comparison groups, displaying 98% specificity (Pierce et al., 2016a). Thus, this unique subtype of ASD toddlers, referred to as ‘GeoPref ASD’, displays a specific and extreme lack of preference for socially compelling stimuli. GeoPref ASD was originally identified in early screening population-based samples that have high generalizability across the ASD spectrum (Pierce et al., 2011a). In the largest eye tracking study of ASD to date (Pierce et al., 2016a), we replicated in a large independent sample (n = 334) the same ASD subtype using the original cutoffs derived in our first discovery on this topic (Pierce et al., 2011a). GeoPref ASD toddlers are also more severe on a variety of other clinical behavioral measures, indicating that this subtype is highly clinically relevant beyond social visual engagement (Pierce et al., 2016a). The GeoPref Test has high test-retest reliability (Pierce et al., 2016a) and is simple, fast, and easy to implement, making it a robust behavioral assay for identification of a clinically highly relevant early ASD subtype.

Figure 1 with 1 supplement see all
Identification of the GeoPref ASD subtype and behavioral differentiation in ASD symptoms, verbal, non-verbal, and adaptive behavior domains.

Panel A shows examples of the stimuli used in the GeoPref eye tracking test as well as example fixations from a GeoPref ASD individual (pink), and a nonGeo ASD individual (blue). The red dots superimposed on the stimulus show visual fixations and the size of the red dots indicate fixation duration. Panel B shows a scatter-boxplot of eye tracking data on the GeoPref test for subjects who also had rsfMRI data available (GeoPref ASD, n = 16, pink; nonGeo ASD, n = 62, blue; language/developmental delay, LD/DD, n = 15 yellow; typically developing siblings of ASD individuals, TD ASDSib, n = 16 green; typically developing toddlers, n = 55, turquoise). The middle line of the boxplot represents the median. The box boundaries represent the interquartile range (IQR; Q1 = 25th percentile, Q3 = 75th percentile), while the whiskers indicate the a distance of 1.5*IQR. Percentage of time fixating on the geometric visual stimulus is plotted on the y-axis and group membership is plotted on the x-axis. The cutoff threshold of 69% is noted as the dashed line. GeoPref ASD toddlers (pink) fall above the cutoff, while all other ASD toddlers (nonGeo ASD; blue) fall below the cutoff. Panel C depicts individual and group-level developmental trajectories for longitudinal data from GeoPref ASD (n = 60, pink) or nonGeo ASD (n = 62, blue) on ADOS, Mullen Early Scales of Learning, and Vineland Adaptive Behavior subscales. All measures show a significant main effect of subtype passing FDR q < 0.05. Mullen Receptive Language and Visual Reception subscales additionally show significant (FDR q < 0.05) age*subtype interactions, indicative of different developmental trajectories between the subtypes. The image of a child shown in panel A is taken from a commercially available video (Yoga Kids 3; Gaiam, Boulder, Colorado, http://www.gaiam.com, created by Marsha Wenig, http://yogakids.com) and re-produced here with permission.

© 2003 Gaiam Americas, Inc. All Rights Reserved. Figure 1A is taken from a commercially available video (Yoga Kids 3; Gaiam, Boulder, Colorado, http://www.gaiam.com, created by Marsha Wenig, http://yogakids.com) and re-produced here with permission.

In the current work, we aimed to identify how intrinsic functional connectivity between neural circuits, as measured by resting state fMRI (rsfMRI), is affected in early ASD development and whether heterogeneity in early social visual engagement is a key factor explaining such connectivity differences. Several networks were examined that have high relevance for early social visual engagement. Primary visual cortex, visual association cortices, and networks involved in attentional or salience processing are of high relevance, given their importance in hierarchical processing of features from social visual stimuli (Haxby et al., 2001; Kriegeskorte et al., 2008; Uddin et al., 2013; Mottron et al., 2006; Felleman and Van Essen, 1991; Yang et al., 2015). Subcortical areas such as the amygdala and ventral striatum are also of relevance given theories about social motivation as a key driver of social engagement difficulties in autism (Chevallier et al., 2012; Mosconi et al., 2009; Elison et al., 2013). A large-scale network, the default mode network, is also key since this network is one of the primary networks of the ‘social brain’ involved in high-level social-cognitive and social-communicative processing (e.g., mentalizing, joint attention) (Lombardo et al., 2010a; Redcay et al., 2013; Van Overwalle, 2009; Schurz et al., 2014; Nummenmaa and Calder, 2009; Eggebrecht et al., 2017; Alcalá-López et al., 2018; Redcay and Schilbach, 2019; Schilbach et al., 2008). We analyzed rsfMRI data in one of the largest and youngest samples of ASD to date and compare ASD to several non-ASD comparison groups – typically developing (TD) toddlers, toddlers with language or general developmental delay (LD/DD) and TD toddlers with an older ASD sibling (TD ASDSib). We also examined how connectivity differences may be better modeled by taking into account heterogeneity in early social visual engagement. Based on prior work theorizing altered connectivity between high-level social-cognitive networks in frontal cortex and posterior networks involved in sensation, perception and attention (Courchesne and Pierce, 2005), we hypothesized that there may be atypical and heterogeneous functional connectivity between higher-level social brain networks such as the default mode network (DMN) and posterior lower-level networks that are integral for visual perception and attention. We further hypothesized that GeoPref ASD toddlers may display the most extreme effects on functional connectivity. Finally, to test the utility of our ASD subtype for predicting individual differences in social-communication symptomatology, we predicted that robust relationships between functional connectivity and social-communication behavior would be apparent only within the GeoPref ASD subtype.

Results

Behavioral and developmental characteristics of the GeoPref ASD subtype

Our primary aim in this work was to test if early ASD subtypes defined on the basis of distinctions in early social engagement behavior would also differ at the level of macroscale neural circuit organization as measured with rsfMRI. However, because little prior work exists characterizing these subtypes, we first set out to detail how these subtypes are characterized both behaviorally and developmentally over the first 4 years of life. Using longitudinal data from 12 to 48 months of life, we assessed how the GeoPref and nonGeo ASD subtypes differ with respect to clinical behavioral trajectories. Every measure across the ADOS, Mullen Early Scales of Learning and the Vineland Adaptive Behavior subscales showed a significant effect of group, except for the Vineland Motor subscale. This indicates that there is an on-average difference between GeoPref and nonGeo ASD in this early developmental period manifesting with the GeoPref ASD subtype being more severely affected. In addition, the Mullen Receptive Language and Visual Reception subscales also showed evidence of an age*subtype interaction, which manifests as GeoPref ASD showing a much steeper downwards trajectory of severity compared to nonGeo ASD. These steeper downward trajectories in GeoPref ASD should not be interpreted as this subtype losing skills over development. Rather, since the scores analyzed here are normed T-scores, the correct interpretation is that as individuals get older, this subtype may be falling behind age-appropriate norms at a faster rate than nonGeo ASD. No other measure showed evidence of this kind of age*subtype interaction (Figure 1C; Figure 1—figure supplement 1 and Supplementary file 1). These insights generally extend prior work (Pierce et al., 2011a; Pierce et al., 2016a) by showing that the GeoPref ASD subtype distinction is indeed a subtype distinction with clinical importance extending beyond the domain of early social visual engagement. These results underscore that GeoPref ASD individuals are more severely affected across the ASD symptom dyad, as well as other language, cognitive, motor, and adaptive behavior domains. These data also suggest that receptive language and non-verbal cognitive skills may be particular domains where GeoPref ASD individuals may take on different developmental paths and progressively fall further behind typically-developing age norms compared to nonGeo ASD peers.

Functional hypoconnectivity between the default mode network and visual or attention networks in ASD

To assess ASD differences in functional connectivity we modeled the rsfMRI data with two approaches – 1) an unstratified case-control model approach and 2) a stratified subtype model approach. These two modeling approaches can be empirically compared to discern whether subtyping by the GeoPref ASD or nonGeo ASD subtypes is an important distinction for explaining macroscale functional neural circuit organization. Using an unstratified case-control model comparing all ASD toddlers (n = 109) to toddlers from multiple non-ASD comparison groups (TD, TD ASDSib, LD/DD; n = 55, n = 16, and n = 15 respectively), we identify three component pairs passing multiple comparisons correction at FDR q < 0.05. These effects feature reduced default mode network (DMN; IC10) connectivity in ASD with primary visual cortex (PVC; IC05: F(3, 189)=6.62, p=2.79e-4, partial η2 = 0.099, Cohen’s d > 0.52), visual association circuitry located in occipito-temporal cortex (OTC; IC02: F(3, 189)=5.98, p=6.41e-4, partial η2 = 0.088, Cohen’s d > 0.51) and the dorsal attention network (DAN; IC09: F(3, 189)=5.93, p=6.87e-4, partial η2 = 0.087, Cohen’s d > 0.40) (Figure 2). All other comparisons between non-ASD comparison groups showed relatively small effect size differences, indicating that the ASD group is the primary reason for these larger reductions in connectivity. See Supplementary file 2 for a full list of all statistics for these case-control comparisons. This modeling approach suggests that without taking into account heterogeneity in social engagement, ASD as a whole shows specific reductions in DMN connectivity with visual and attention circuits compared to TD, but also compared to LD/DD toddlers and TD siblings of ASD individuals.

Figure 2 with 3 supplements see all
Functional hypoconnectivity between DMN, visual and attention networks in ASD.

The left column shows surface renderings of ICA components of visual association areas in occipito-temporal cortex (OTC; IC02), primary visual cortex (PVC; IC05), the dorsal attention network (DAN; IC09) and the default mode network (DMN; IC10). The middle column shows scatter-boxplots for DMN-OTC, DMN-PVC, and DMN-DAN connectivity across GeoPref ASD (pink), nonGeo ASD (blue), ASD with no eye tracking data (ASD no ET; magenta), LD/DD (yellow), TD ASDSib (green), and TD (turquoise). Standardized effect sizes (Cohen’s d) are reported in the plots for comparisons of all ASD individuals combined, compared to the other non-ASD comparison groups. The middle line of the boxplot represents the median. The box boundaries represent the interquartile range (IQR; Q1 = 25th percentile, Q3 = 75th percentile), while the whiskers indicate the a distance of 1.5*IQR. The right column uses heatmaps to show standardized effect sizes (Cohen’s d) for all pairwise comparisons of groups with rsfMRI and eye tracking data used in the subtype model. Note that effect sizes are depicted as absolute values, and the directionality of the effects can be seen in the scatter-boxplots. Cells outlined in thick black lines are specific comparisons that survive FDR q < 0.05. Cells outlined in thinner black lines are comparisons that survive FDR q < 0.1.

Marked functional hypoconnectivity between default mode and visual association circuitry in GeoPref ASD

Using the eye tracking subtypes from the GeoPref Test, we next recomputed stratified subtype models to test whether subtyping by social engagement heterogeneity provides a better explanation of differences in macroscale intrinsic functional neural circuit organization. DMN-OTC and DMN-PVC appear again in these stratified models as effects passing FDR q < 0.05, despite the fact that n = 31 ASD subjects from the original unstratified case-control model had to be dropped in this analysis since they lacked any eye tracking data to make the GeoPref or nonGeo ASD distinction. Descriptive standardized effect sizes for all pairwise comparisons of ASD subtypes and non-ASD comparison groups are shown in Figure 2 and pairwise comparisons surviving FDR q < 0.05 correction are highlighted in Figure 2 with black outlines around specific cells. See Supplementary file 3 for statistics from pairwise group comparisons.

We then used the Akaike Information Criteria (AIC) to evaluate whether a stratified subtype model was better than the traditional unstratified case-control model. Models that result in a difference in AIC values (e.g., ΔAIC) less than or equal to 2, indicate that both models under comparison are largely similar to each other. However, models with ΔAIC ≥4 tend to show less support for the comparison model relative to the best model (i.e. the model with the lowest AIC) (Burnham and Anderson, 2004). Here we find that the stratified subtype model is the best performing model for DMN-OTC (IC10-IC02) connectivity, since it produces a lower AIC compared to the unstratified case-control model (subtype model AIC = −51.32; case-control model AIC = −47.60). The ΔAIC for DMN-OTC is 3.72, which is closest to the rule of 4 ≤ ΔAIC ≤7 suggested by Burnham and Anderson (2004) indicating substantially less support for the case-control model being a good contender against the subtype model. However, both DMN-PVC and DMN-DAN comparisons showed that the stratified and unstratified models were largely similar to each other, with difference in AIC values consistently below 2 (DMN-PVC: case-control model AIC = −89.54, subtype model AIC = −87.72; ΔAIC = 1.82; DMN-DAN: case-control model AIC = 19.20, subtype model AIC = 20.29; ΔAIC = 1.09). In addition to ΔAIC, we also used 5-fold cross validation to compute mean absolute percentage error (MAPE). In this context, better models produce lower MAPE estimates. For DMN-OTC, MAPE was much lower for the subtype model (MAPE = 125.5452) compared to the case-control model (MAPE = 135.2241), indicating the subtype model reduced the absolute percentage of error by around 9.67%. However, for DMN-PVC and DMN-DAN, MAPE was lower for the case-control model (DMN-PVC subtype MAPE = 152.8099, case-control MAPE = 152.1738; DMN-DAN subtype MAPE = 156.0619, case-control MAPE = 152.8081), but in both cases the reduction in MAPE was <1%.

The subtype model can also be compared against models that ignore categorical subtype or diagnostic labels and instead model variability on continuous measures. Here we compared the subtype model to a transdiagnostic model using the continuous measure of fixation on the geometric stimulus from the GeoPref test as the primary predictor. The subtype model was unanimously a better explanatory model than a transdiagnostic model for each of the three component pairs, with ΔAIC >5 (DMN-OTC: subtype AIC = −51.32, transdiagnostic AIC = −46.13; ΔAIC = 5.19 DMN-PVC: subtype AIC = −87.72, transdiagnostic AIC = −74.63; ΔAIC = 13.09; DMN-DAN: subtype AIC = 20.29, transdiagnostic AIC = 25.42; ΔAIC = 5.13) (Figure 2—figure supplement 1).

These results generally suggest that heterogeneity in early social engagement is important for some, but not all, functional connectivity differences present in early ASD toddlers. This evidence supports the idea that in toddlerhood, ASD is characterized by functional hypoconnectivity between the DMN and visual circuits in PVC and OTC as well as attention circuitry (DAN). However, for the subtype of individuals with the most pronounced social engagement difficulties (GeoPref ASD), it is DMN-OTC connectivity that may be most important in distinguishing them from their other ASD counterparts.

Subtype-specific association between DMN-OTC connectivity and social-communication difficulties

We next tested whether the two ASD subtypes would differ in the relationship between functional connectivity and social-communication difficulties. This test is critical for examining the question of whether the two ASD subtypes should indeed be considered as discrete types of ASD with different underlying neurobiological mechanisms relevant for the behavioral phenotype. Here we find evidence for subtype-specific associations for DMN-OTC connectivity. A strong negative correlation between DMN-OTC connectivity and ADOS social affect total is present in the GeoPref ASD subtype (r = −0.78, p=0.001, 95% CI = [−0.99,–0.35]), indicating that reductions in DMN-OTC connectivity are associated with more enhanced social-communication difficulties. No such relationship exists for nonGeo ASD (r = 0.06, p=0.64, 95% CI = [−0.20, 0.33]) (Figure 3). Correlation strength also significantly differed between the subtypes (z = 3.71, p=0.0001). In contrast, DMN-PVC showed no evidence of relationships in either subtype (GeoPref ASD r = 0.04, p=0.88, 95% CI = [−0.81, 0.75]; nonGeo ASD r = 0.22, p=0.12, 95% CI = [−0.10, 0.49]) or for differences in the strength of the relationship between subtypes (z = 0.59, p=0.55). DMN-DAN showed no evidence of a strong relationship in GeoPref ASD (r = −0.33, p=0.25, 95% CI = [−0.83, 0.58]). However, there was a significant, although weak positive relationship in nonGeo ASD (r = 0.29, p=0.04, 95% CI = [0.05, 0.49]). The differences in the strength of the relationship between subtypes was also weak, though significant (z = 2.14, p=0.03) (Figure 3).

Connectivity-social-communication relationships.

This figure shows functional connectivity-social-communication relationships for DMN-OTC (Panel A), DMN-PVC (Panel B), and DMN-DAN (Panel C). The relationship for GeoPref ASD is shown in pink, while nonGeo ASD is shown in blue. Both ADOS social affect and connectivity scores shown in the plot are covariate adjusted scores (taking into account age) from the robust regression model evaluating the relationship.

Discussion

Although social visual engagement difficulties are central early developmental features of ASD with high clinical and translational relevance, the neural bases behind these features are not well understood. Here we find robust evidence in early ASD development for atypical intrinsic functional brain organization involving circuitry relevant to social visual engagement. Central to our findings is the importance of early developing social brain circuitry, the default mode network (DMN) (Lombardo et al., 2010a; Kennedy and Adolphs, 2012; Kennedy and Courchesne, 2008a; Kennedy et al., 2006; Padmanabhan et al., 2017; Lombardo et al., 2011; Lombardo et al., 2010b), and its interactions with more basic neural circuitry involved in visual or attention processes. The DMN has been extensively studied in older ASD patients and has been found to be atypical in multiple studies using resting state and task-based social paradigms (Padmanabhan et al., 2017; Lombardo et al., 2010b; Di Martino et al., 2014; Kennedy and Courchesne, 2008b; Hong et al., 2019; Lai et al., 2019; Holiga et al., 2019). However, a novel addition to this literature is our work here showing that in very early development there is functional hypoconnectivity between DMN and primary visual cortex (PVC) as well as visual association circuitry in occipito-temporal cortex (OTC). Here we find that at this point in early development normative groups (e.g., TD) show robust non-zero DMN-OTC connectivity. In contrast, in older typically-developing samples, similar types of robust functional connectivity relationships between these networks are not extensively reported in literature, as DMN is typically highly segregated and embedded at an opposite pole within functional connectivity gradients from sensory cortices (Smith et al., 2015; Margulies et al., 2016). These effects along with DMN-dorsal attention network (DAN) hypoconnectivity show evidence for a case-control difference compared to TD and other non-ASD comparison toddler age groups, and these effects are prominent across both ASD cohorts with and without eye tracking data. This finding suggests that in ASD within the first 4 years of life the developing ‘social brain’ default mode network is on-average functionally disconnected from visual cortices and the dorsal attention network. This finding is important with respect to why social engagement difficulties are a hallmark symptom in early ASD development and shared across most ASD toddlers.

In contrast to generalized case-control differences, we also observed key evidence that early heterogeneity in social visual engagement is important for explaining functional hypoconnectivity between DMN and OTC. The GeoPref ASD subtype is a primary factor behind DMN-OTC functional hypoconnectivity in ASD and this subtype distinction is a better explanatory model than a traditional case-control model or other transdiagnostic models. While GeoPref ASD is behaviorally more severe on a range of different domains, this functional connectivity difference is specific to GeoPref ASD. Other non-ASD toddlers with general developmental or language delay do not show such effects. Furthermore, the magnitude of DMN-OTC functional hypoconnectivity specific to GeoPref ASD was strongly associated with individual-level variability in social-communication difficulties measured on an ASD ‘gold standard’ diagnostic instrument (i.e. ADOS). While DMN-PVC and DMN-DAN connectivity reductions are also prominent in GeoPref ASD toddlers, it could be DMN-OTC connectivity that has the largest implications for the different clinical presentation specific to this subtype. Early functional hypoconnectivity between DMN-OTC circuitry could have important differential impact for this subtype with regards to further developmental functional specialization of the social brain for social cognitive and social-communicative functions. In contrast, the less affected DMN-OTC circuitry in nonGeo ASD individuals could be an indication of the plasticity in this circuitry as a result of relatively more enhanced social experience in this subtype. Given that many early interventions heavily focus on enhancing early social experience (Bacon et al., 2014; Dawson et al., 2010; Kasari et al., 2006; Pickles et al., 2016; Dawson, 2008), DMN-OTC connectivity may be a key target for better understanding why individuals heavily vary in response to early intervention treatment.

It is unknown what neurobiological processes underlie these differences in DMN connectivity with visual and attention circuitry. Recent evidence from two twin studies suggests that variability in early social visual engagement is underpinned by genetic factors potentially linked to heritable common genetic variants (Constantino et al., 2017; Kennedy et al., 2017). How such potential genetic starting points lead to further pathophysiology up to the behavioral phenotype is still a mystery. However, it is clear from recent work that neural pathophysiology in ASD likely begins at the earliest prenatal periods of development (Courchesne et al., 2011; Stoner et al., 2014; Parikshak et al., 2013; Willsey et al., 2013; Marchetto et al., 2017; Parikshak et al., 2016; Courchesne et al., 2019). This prenatal pathophysiology is diverse but includes increased cell proliferation, aberrant cell migration, and downregulated early axonogenesis during early prenatal periods, whereas starting in later prenatal development and continuing throughout the first years of life there is also downregulation of processes typically important for synapse development (Courchesne et al., 2011; Stoner et al., 2014; Parikshak et al., 2013; Willsey et al., 2013; Marchetto et al., 2017; Parikshak et al., 2016; Courchesne et al., 2019). Microglia activation and upregulation of immune/inflammation and protein synthesis (i.e. translation initiation) processes are prominent in ASD throughout life and could have further potential impact on cell signaling relevant to neuronal, glial, and synaptic development (Parikshak et al., 2016; Lombardo et al., 2017; Lombardo et al., 2018b; Pramparo et al., 2015a; Gupta et al., 2014; Voineagu et al., 2011; Chow et al., 2012; Morgan et al., 2010; Takano, 2015). Collectively, postmortem, genomic and genetic, early-age neuroimaging, animal model and patient induced pluripotent stem cell (iPSC) model studies indicate prenatal and postnatal ASD cortex displays an abnormal overabundance of small and possibly poorly differentiated neurons with undergrown axons (Solso et al., 2016), focal laminar and migration defects, neuroinflammation and aberrant synapse formation and function (Courchesne et al., 2019). It has long been theorized that such early abnormalities could cause disconnection of higher-order frontal social and communication networks from lower-level posterior perceptual networks and thereby impair attention to and integration of relevant social, emotional and communicative events during development in ASD (Courchesne and Pierce, 2005). While these early developmental abnormalities are numerous and diverse, a key goal for future work will be to determine which processes, or collection of processes working together, lead to the early emergence of ASD and different social engagement subtypes. One compelling hypothesis is that prenatal pathophysiology could cause a cascade in early development leading towards the ultimate initiation of atypical social engagement behaviors. Aggregation of common genetic risk associated with atypical social visual engagement (Constantino et al., 2017; Kennedy et al., 2017) may be relevant to atypical prenatal development processes (Courchesne et al., 2019). These atypical prenatal processes could trigger early behavioral adaptation responses that manifest as different ways to explore and sample the social environment in early development (Johnson et al., 2015; Johnson, 2017). Different social visual engagement behavior could then lead to different experience-dependent brain development in different ASD toddlers. The end result of this chain could be the emergence of behavioral and neural subtypes but also with some shared aspects of neural abnormalities that are specific to ASD. It will be crucial for future work to examine what set of genetic risk variants and early prenatal pathophysiology may underpin variation in early social visual engagement in ASD and how such mechanisms may lead to different experience-dependent development and neural circuitry. Work utilizing iPSC modeling to recapitulate some of the early prenatal neurobiological processes occurring in such individuals seems highly appropriate for beginning to answer such questions (Marchetto et al., 2017).

The current work may also be important for explaining how and why early developmental functional specialization of social brain circuitry occurs in typical development, and how it may be atypical in some individuals with ASD. Reduced early social engagement behaviors that are adaptations to atypical prenatal starting points could potentially lead to the construction and long-term maintenance of atypical early developmental environmental niches for an individual (Johnson et al., 2015; Johnson, 2017). An environmental niche constructed and maintained from largely sampling the non-social instead of the social world around an infant could have a large impact on how postnatal neural circuits are formed. It is well known that neural circuits are sculpted by experience, particularly in the early years of life when biological processes such as synaptogenesis, axon expansion, and cell-growth are normally at their peak (Greenough et al., 1987; Holtmaat and Svoboda, 2009; Huttenlocher, 2002; Kang et al., 2011). Selective biases to the type of information sampled from the environment in GeoPref ASD may provide the wrong type of input to facilitate social cognitive and social-communicative functional specialization of circuits such as the DMN. Circuits such as the DMN undergo protracted developmental periods of interactive specialization (Johnson, 2011), whereby specialization is achieved through interactions between circuits. Our work here suggests that functional interaction between visual and social brain circuits may be crucial to early social behavior in this ASD subtype, and may have further impact for aiding the developing specialization of function within social brain circuitry.

In order to put this work into context, there are a variety of strengths, caveats, and limitations to underscore. Methodologically, the findings are bolstered by being one of the largest sample sizes to date of rsfMRI data in very young ASD toddlers. The early age range of our sample stands out with respect to most of the existing rsfMRI evidence in the literature, since a large majority of studies examine much older individuals. An interpretational caveat with studies of older individuals is that it is unclear which effects are due to core pathophysiology relevant to ASD and which emerged at much earlier developmental time-points versus effects that could be explained as adaptation or other compensatory mechanisms that occur later in life and are not necessarily linked to core early neurodevelopmental mechanisms of relevance to ASD. While adaptation and compensatory effects are still likely even at very early ages, the advantage of the early age range in this sample (relative to studies on older individuals) is the enhanced ability to interpret the results with regards to possible core mechanisms that are at work at ages when the earliest behavioral symptoms relevant to ASD diagnosis manifest.

The findings are also bolstered by a methodological strength of making comparisons against other non-ASD comparison groups, such as LD/DD and TD ASDSib. Such non-ASD and non-TD comparison groups are relatively rare compared to most neuroimaging studies in the current literature. Comparisons between such groups allow for added inferences about specificity of effects to ASD or its subtypes and also help to show that these effects are not simply due to being clinically more severe across a range of domains. The sample ascertainment strategy based on early population-screening is another considerable strength that should result in higher generalizability compared with other studies that study a smaller subsection of the population, such as baby sibling studies. In fact, data generated using this unique sample ascertainment strategy revealed that functional connectivity within the examined networks including between the DMN and OTC did not significantly differ between unaffected siblings and typically developing toddlers. Interpretation of this finding warrants further investigation but could suggest that the development of early connectivity between social and visual attention brain networks could be key to whether or not an infant eventually manifests symptoms of ASD.

Finally, our study goes beyond most work using case-control designs and identifies heterogeneity in a relatively infrequent subtype that comprises around 20% of the early ASD population. In this study, the sample size of the GeoPref ASD subtype was relatively small (n = 16). While small sample size can be problematic for reasons of suboptimal coverage of the ASD spectrum in case-control comparisons (Lombardo et al., 2019), our design here confronts this issue head first by parsing some aspect of the heterogeneity in ASD, rather than simply being a small sample size case-control comparison. However, because of the imbalance in sample sizes between GeoPref and nonGeo ASD subtypes, future work is necessary where the sample sizes are more balanced. This will require much effort to enrich samples with the GeoPref ASD subtype, given that it is present in around 20% of the early diagnosed ASD population. Second, while small sample size can be problematic for statistical power reasons, we have shown what are the minimal effect sizes at such sample sizes for adequate statistical power (Figure 2—figure supplement 2), and our estimated effect sizes are well above such minimal effects sizes. The brain-behavior correlation estimates are also supplemented by the reporting of bootstrap confidence intervals. Since small sample sizes can result in inflated estimates of correlation, the bootstrap confidence intervals allow for reporting of the distribution of sample correlation estimates that could have been observed.

Finally, it is important to underscore that the GeoPref ASD subtype is based on one relatively quick and simple early eye tracking test. This test picks up about 20% of all early diagnosed ASD individuals as within this subtype. It could be that integration of a larger battery of eye tracking measures and tests may help enhance sensitivity and detect a larger percentage of individuals that may be part of this subtype (Moore et al., 2018). Furthermore, use of eye tracking measures that go beyond static stimuli and use interaction-based paradigms (Redcay and Schilbach, 2019; Schilbach et al., 2013) may further enhance sensitivity. All of these refinements could potentially aid in the identification of early biological bases behind different ASD subtypes.

In conclusion, we identified that functional hypoconnectivity between ‘social brain’ circuitry, the DMN, and low-level visual networks is highly important in the early development of ASD. While DMN-PVC functional connectivity is reduced on-average across ASD toddlers, DMN-OTC functional connectivity is heavily reduced in the GeoPref ASD subtype. Individual-level variation in DMN-OTC functional hypoconnectivity is associated with the degree of social-communication difficulties, but only within the GeoPref ASD subtype. This subtype can be identified with high levels of precision via a simple eye tracking test of social or non-social visual preferences in very early development (Pierce et al., 2011a). Early social-visual functional hypoconnectivity is a key underlying neurobiological feature describing GeoPref ASD and may be critical for future social-communicative development. Thus, we theorize that a neurobiologically driven bias (particularly from prenatal development) leading towards neglect for the social world in early development, if left untreated, would significantly diminish opportunities for social learning and for bootstrapping experience-dependent change within developing neural circuits. This, we suggest, may limit developmental functional specialization within the social brain and lead towards more permanent long-term social, communicative, and cognitive difficulties. Therefore, it is an important goal to hone in on early-age molecular biomarkers that personalize risk, since very early identification may have larger clinical and translational benefits under such contexts. Early detection and interventions that successfully ameliorate early atypical functional connectivity between social brain and visual circuitry might improve social development and outcome for ASD individuals. Despite the narrow view of some institutions (USPSTF) (Siu et al., 2016), early risk detection is an absolute necessity (Pierce et al., 2016b), as is research to devise effective early interventions tailored to specific ASD individuals.

Materials and methods

Participants

This study was approved by the Institutional Review Board at University of California, San Diego (UCSD Human Research Protection Program protocols 091539, 081722, or 110049). Parents provided written informed consent according to the Declaration of Helsinki and were paid for their participation. Identical to the approach used in our earlier studies (Pierce et al., 2011a; Pierce et al., 2016a; Lombardo et al., 2015; Pramparo et al., 2015a; Pramparo et al., 2015b), toddlers were recruited through two mechanisms: community referrals (e.g., website) or a general population-based screening method called the 1 Year Well-Baby Check-Up Approach (Pierce et al., 2011b) that allowed for the prospective study of ASD beginning at 12 months based on a toddler’s failure of the CSBS-DP Infant-Toddler Checklist (Wetherby and Prizant, 2002; Wetherby et al., 2008). All toddlers were tracked from an intake assessment age of 12–24 months and followed roughly every 12 months until 3–4 years of age. All toddlers, including normal control subjects, participated in a series of tests collected longitudinally across all visits, including the Autism Diagnostic Observation Schedule (ADOS; Module T, 1, or 2) (Lord et al., 2000), the Mullen Scales of Early Learning (Mullen, 1995), and the Vineland Adaptive Behavior Scales (Sparrow et al., 1984). All testing occurred at the University of California, San Diego Autism Center of Excellence (ACE).

A total of n = 195 toddlers aged 12 to 48 months were scanned with rsfMRI during natural sleep for the current study. Of the total n = 195, n = 109 were ASD toddlers and this ASD group could be further split into n = 16 GeoPref ASD toddlers (11 male, five female), n = 62 nonGeo ASD toddlers (49 male, 13 female), and a further n = 31 ASD toddlers (27 male, four female) that could not be stratified in either GeoPref or nonGeo ASD subtypes because they lacked eye tracking data needed for such stratification. The GeoPref ASD subtype is defined as toddlers who spent 69% or more of their time fixated on the dynamic geometric stimulus in the GeoPref eye tracking test (Pierce et al., 2011a; Pierce et al., 2016a). The nonGeo ASD subtype represents individuals that spent less than 69% of time fixated on the dynamic geometric stimulus. The 69% threshold for the GeoPref test has been shown in past studies to replicably isolate the GeoPref ASD subtype and maximize specificity with respect to many other contrast groups (Pierce et al., 2011a; Pierce et al., 2016a). Several toddlers from multiple non-ASD groups were examined as contrast groups to the ASD sample. These toddlers consisted of n = 55 typically-developing control toddlers (TD, 37 male, 18 female), n = 15 with language or globally developmental delay (LD/DD, 10 male, five female), and n = 16 TD toddlers who were younger siblings of children already diagnosed with ASD (TD ASDSib, eight male, eight female). See Table 1 for characteristics such as age at scanning and age at eye tracking for all groups. Amongst the total n = 195 toddlers with rsfMRI data, an ANOVA on age at scanning found no group-differences across the groups (F(5, 189)=0.89, p=0.48). A chi-square analysis across the total n = 195 identified a subtle trend-level difference in the proportion of males and females distributed across groups, with more even proportions of males and females in non-ASD comparison groups than the ASD groups (χ2(5)=9.88, p=0.07). In all subsequent connectivity analyses, we statistically controlled for sex and age at scanning as covariates.

Table 1
Descriptive statistics across all groups for age at rsfMRI scan, sex, and head motion measurements.
N
(M, F)
Mean age at eye
tracking in
months (SD)
Mean age at rsfMRI
scan in months (SD)
Age range at
rsfMRI scan in
months
Mean framewise
displacement (SD)
ASD no ET (rsfMRI data only; no eye tracking data)31 (27,4)-29.69 (8.88)13.21–43.630.09 (0.08)
GeoPref ASD16 (11,5)28.37 (7.77)29.92 (8.71)14.16–43.790.06 (0.02)
nonGeo ASD62 (49,13)26.30 (8.35)29.37 (8.35)12.35–44.050.09 (0.12)
LD/DD15 (10,5)19.36 (4.15)25.12 (7.97)13.37–39.750.10 (0.05)
TD ASDSib16 (8,8)19.79 (6.20)26.74 (9.38)12.52–44.090.08 (0.04)
TD55 (37,18)23.07 (9.07)29.61 (10.14)13.17–47.930.10 (0.11)

Because the n = 16 within the GeoPref ASD subtype is a relatively small sample size, we ran a statistical power analysis simulation to identify the minimum effect size needed to detect reject the null hypothesis with an alpha of 0.05% and 80% power at the current sample sizes (e.g., n = 16 GeoPref ASD vs n = 55 TD). In this simulation, we generated data from two populations (hypothetically GeoPref ASD and TD), with the minimal effect size difference between the populations that achieves 80% power at these sample sizes. The population size of each group was set to n = 10,000,000. We then ran 100,000 simulated experiments whereby we randomly sampled without replacement n = 16 (e.g., GeoPref ASD) and n = 55 (e.g., TD) from each population, and then computed effect size and ran a hypothesis test (e.g., independent samples t-test) with the alpha set to 0.05. In Figure 2—figure supplement 2, we illustrate how variable sample effect size estimates can be given this context of n = 16 vs n = 55 (e.g., sample effect size estimates range from d < 0 to d > 1.5). Power is empirically shown here (see red histogram in Figure 2—figure supplement 2) since exactly 80,000 of the 100,000 (80%) experiments rejected the null hypothesis at alpha = 0.05. The minimum effect size to achieve 80% power with these sample sizes is d = 0.80751.

Eye tracking paradigm

Request a detailed protocol

Eye-tracking data from this study has already been reported in two prior studies (Pierce et al., 2011a; Pierce et al., 2016a). Briefly recapping the GeoPref eye tracking test, toddlers were seated on their parent’s lap 60 cm in front of the Tobii T120 eye tracking monitor and were presented a movie consisting of dynamic geometric or social stimuli on either side of the screen. The dynamic geometric stimulus was produced from recordings of animated screen saver programs. The dynamic social stimulus was produced from a series of short sequences of children doing yoga, which included images of children moving in a dramatic manner (e. g., waving arms and appearing as if dancing). These clips were used with permission from the commercially available video Yoga Kids 3 (2003 Gaiam Americas, Inc; https://www.gaiam.com) and specific permissions were given to use the clips for research purposes in published scholarly work. Audio information was discarded. The final stimulus was composed of 2 rectangular areas of interest (AOIs) horizontally distributed containing either the geometric or social stimulus and were changed in a simultaneous, time-linked fashion. The side (left/right) of presentation of geometric or social stimuli were randomly assigned across subject and diagnosis and percent fixation within each AOI calculated. The final movie contained a total of 28 scenes with single-scene duration varying from 2 to 4 s for a total presentation time of 60 s at 24 frames per second.

Longitudinal clinical behavioral data

Request a detailed protocol

To examine behavioral trajectories on other clinical measures (i.e. ADOS, Mullen, Vineland) we utilized the largest sample of GeoPref ASD toddlers available to boost power beyond the smaller sample that also had rsfMRI data available. An additional 44 GeoPref ASD toddlers who had valid eye tracking and clinical data (but not rsfMRI) were included with the n = 16 GeoPref ASD toddlers with eye tracking and rsfMRI data. Thus, a total of n = 60 GeoPref ASD and n = 62 nonGeo ASD toddlers were analyzed for differences in clinical behavioral trajectories, as illustrated in Figure 1C and Figure 1—figure supplement 1. We used linear mixed-effect modeling analyses (modeling random slopes and intercepts) to model within-individual trajectories and group-level trajectories. These analyses were implemented on z-scored data with the lme function contained within the nlme library in R. The dependent variables were either ADOS, Mullen, or Vineland subscale scores, while the independent variables modeled were always age, group, and the age*group interaction. A comparison of linear versus quadratic models indicated that a linear model provided a better fit (linear model, AIC = −158.09; quadratic model, AIC = −148.64) and thus, the linear model was used.

fMRI data acquisition

Request a detailed protocol

Imaging data were collected on a 1.5 Tesla General Electric MRI scanner during natural sleep at night; no sedation was used. High-resolution T1-weighted anatomical scans were collected for warping fMRI data into standard atlas space. Blood oxygenation level-dependent (BOLD) signal was measured across the whole brain with echoplanar imaging (TE = 30 ms, TR = 2500 ms, flip angle = 90°, bandwidth = 70 kHz, field of view = 25.6 cm, in-plane resolution = 4×4 mm, slice thickness = 4 mm, 31 slices). The resting state session lasted 6 min and 25 s resulting in 154 total whole brain volumes. Within the preprocessing we discarded the first four volumes to allow for T2-stabilization effects, leaving a total of 150 volumes for the final analysis.

fMRI data preprocessing

Request a detailed protocol

Preprocessing of the resting state data was split into two components; core preprocessing and denoising. Core preprocessing was implemented with AFNI (http://afni.nimh.nih.gov/) using the tool speedypp.py (http://bit.ly/23u2vZp) (Kundu et al., 2012). This core preprocessing pipeline included the following steps: (i) slice acquisition correction using heptic (7th order) Lagrange polynomial interpolation; (ii) rigid-body head movement correction to the first frame of data, using quintic (5th order) polynomial interpolation to estimate the realignment parameters (3 displacements and three rotations); (iii) obliquity transform to the structural image; (iv) affine co-registration to the skull-stripped structural image using a gray matter mask; (v) nonlinear warping to MNI space (MNI152 template) with AFNI 3dQwarp; (v) spatial smoothing (6 mm FWHM); and (vi) a within-run intensity normalization to a whole-brain median of 1000. Core preprocessing was followed by denoising steps to further remove motion-related and other artifacts. Denoising steps included: (vii) wavelet time series despiking (‘wavelet denoising’); (viii) confound signal regression including the six motion parameters estimated in (ii), their first order temporal derivatives, and ventricular cerebrospinal fluid (CSF) signal (referred to as 13-parameter regression). The wavelet denoising method has been shown to mitigate substantial spatial and temporal heterogeneity in motion-related artifact that manifests linearly or non-linearly and can do so without the need for data scrubbing (Patel et al., 2014). Wavelet denoising is implemented with the Brain Wavelet toolbox (http://www.brainwavelet.org). The 13-parameter regression of motion and CSF signals was achieved using AFNI 3dBandpass with the –ort argument. To further characterize and describe motion and its impact on the data, we computed framewise displacement and DVARS (Power et al., 2012). Examples of how denoising impacts high and low motion subjects can be found in the Figure 2—figure supplement 3. Between-group comparisons showed that all groups were similar with respect to head motion as measured by mean framewise displacement (FD) (F(5,189) = 0.67, p=0.64) with all groups showing on-average less than 0.12 mm motion (see Table 1). Furthermore, mean DVARS measurements were similar across all groups before (F(5,189) = 0.28, p=0.92) and after denoising (F(5,189) = 0.57, p=0.72). Both of these results indicate that motion does not asymmetrically affect one group more than the others.

Connectivity analyses

Request a detailed protocol

To assess functional connectivity between neural circuits we utilized the unsupervised data-driven method of independent component analysis (ICA) to conduct a group-ICA and then used dual regression to back-project spatial maps and individual time series for each component and subject. Both group-ICA and dual regression were implemented with FSL’s MELODIC and Dual Regression tools (www.fmrib.ox.ac.uk/fsl). For group-ICA, the dimensionality estimate was fixed to a pre-specified dimensionality of 30 components, as in most cases with low-dimensional ICA, the number of meaningful components can be anywhere from 10 to 30 (Smith et al., 2013). Higher-dimensional solutions were not sought, although they may be important in future work, particularly with regard to further fractionating known networks like the DMN into subsystems (Kernbach et al., 2018). Given our a priori hypotheses regarding specific processes such as social cognition, attention, salience, visual perceptive, affective, and reward processes, we examined corresponding known networks that are involved in such processes – namely, the default mode, dorsal attention, salience, several visual networks, and amygdala/striatum-centered networks.

Time courses for each subject and each component were used to model between-component connectivity. This was achieved by constructing a partial correlation matrix using Tikhonov-regularization (i.e. ridge regression, rho = 1) as implemented within the nets_netmats.m function in the FSLNets MATLAB toolbox (https://fsl.fmrib.ox.ac.uk/fsl/fslwiki/FSLNets). The aim of utilizing partial correlations was to estimate direct connection strengths in a more accurate manner than can be achieved with full correlations, which allow more for indirect connections to influence connectivity strength (Smith et al., 2013; Marrelec et al., 2006; Smith et al., 2011). Partial correlations were then converted into Z-statistics using Fisher’s transformation for further statistical analyses and the lower diagonal of each subject’s partial correlation matrix was extracted.

General linear models (GLM) were utilized to test for between-group differences in partial correlations – 1) an unstratified case-control model and 2) a stratified subtype model. Both GLM models were implemented with the lm function in R. In these GLMs the partial correlation for a particular component pair was the dependent variable. For independent variables, in all GLMs we used sex and age at scanning as covariates of no interest. The primary independent variable of interest was the group variable. Under the case-control model, group was composed of ASD and non-ASD labels. ASD was the sole label for all ASD subjects and no stratification by eye tracking data was done. The controls in this model were the other non-ASD comparison groups – TD, LD/DD and TD ASDSib. These groups were treated as other separate groups in the model rather than being collapsed into one ‘control’ group. In the stratified subtype model, group utilized the ASD subtype labels identified through the eye tracking data – either the GeoPref and nonGeo ASD subtypes. Group labels for each non-ASD comparison group were also used in this group variable, as was the case for the case-control model. These models were computed for each component pair analyzed and results were thresholded at FDR q < 0.05 for multiple comparisons correction. If any component pairs survived FDR correction, these components were further followed-up with analyses on each pairwise group-comparison and with FDR correction at q < 0.05. These pairwise group-comparisons were analyses using non-parametric permutation tests (10,000 permutations), in order to estimate p-values for each pairwise comparison in a manner that is robust to distributional assumptions. Permutation p-values were then used to compute FDR and only comparisons passing FDR q < 0.05 were deemed significant. Effect sizes from all pairwise group comparisons were estimated as standardized effect size (Cohen’s d) using the cohen.d function in the effsize library in R.

To compare the case-control model to the subtype model, we utilized the Akaike Information Criteria (AIC) statistic. AIC was chosen as the model comparison criterion since it is optimal, compared to other criteria such as BIC, under contexts similar to ASD where the true reality is one of complexity or ‘tapering effects’ (Burnham and Anderson, 2004). In the context of comparing models, typically the model with the lowest AIC value is considered the best model. However, to facilitate interpretation regarding whether there was strong support for either of the models being compared, we computed the difference in AIC, whereby ΔAICi = AICi AICmin. AICmin is the AIC for the model with the lowest AIC value and AICi is the AIC for the i-th model being compared. Burham and Anderson (Burnham and Anderson, 2004) note that ΔAICi ≤ 2 indicates that the AICi model has strong support for being equally as good as the AICmin model. When 4 ≤ ΔAICi ≤ 7, this indicates considerably less support for the AICi model compared to the AICmin model. To further support model selection, we also implemented 5-fold cross validation to compute mean absolute percentage error (MAPE) on held-out unseen data. MAPE is computed as MAPE = mean(abs((Ai – Pi)/Ai)*100), where Ai is actual test data point i, Pi is predicted test data point i, and abs refers to the absolute value. The model with the lowest MAPE is the best model to choose, and we report these results alongside ΔAICi. MAPE is computed for each cross validation fold, and then the average is reported across folds. The difference in MAPE values also allows for interpretability with regards to how much of a reduction in percentage error is gained by the better model.

We also used AIC and ΔAICi to compare the subtype model to a continuous transdiagnostic model. These models utilized all data from all groups with both fMRI and eye tracking data. The continuous transdiagnostic model utilized the continuous measure of percent fixation on the geometric stimulus and without any variable indicating group membership.

Connectivity-Behavior relationships

Request a detailed protocol

To test for relationships between connectivity and social-communication difficulties on the ADOS, we computed robust regression partial correlations between connectivity and ADOS social affect total scores, while covarying for age at scanning. These correlations were computed using the robust regression MATLAB toolbox (https://github.com/canlab/RobustToolbox; Wager et al., 2005). To test for difference between-groups in the strength of such correlations we used the paired.r function in the psych R library to compute Z-statistics and p-values. We also performed bootstrapping to compute 95% confidence intervals around sample correlation estimates, using 100,000 bootstrap resamples. This analysis was done to allow for reporting of the distribution of sample correlation estimates that could have been observed.

Data and code availability

Request a detailed protocol

Tidy data and analysis code are available at https://github.com/mvlombardo/geoprefrsfmri (Lombardo, 2019; copy archived at https://github.com/elifesciences-publications/geoprefrsfmri). 

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
  15. 15
  16. 16
  17. 17
  18. 18
  19. 19
  20. 20
  21. 21
  22. 22
  23. 23
  24. 24
  25. 25
  26. 26
  27. 27
  28. 28
  29. 29
  30. 30
    Neural Plasticity: The Effects of Environment on the Development of Cerebral Cortex
    1. PR Huttenlocher
    (2002)
    Cambridge: Harvard University Press.
  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
    The enactive mind, or from actions to cognition: lessons from autism
    1. A Klin
    2. W Jones
    3. R Schultz
    4. F Volkmar
    (2003)
    Philosophical Transactions of the Royal Society of London. Series B: Biological Sciences 358:345–360.
    https://doi.org/10.1098/rstb.2002.1202
  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
    Developmental trajectories as autism phenotypes
    1. C Lord
    2. S Bishop
    3. D Anderson
    (2015)
    American Journal of Medical Genetics Part C: Seminars in Medical Genetics 169:198–208.
    https://doi.org/10.1002/ajmg.c.31440
  62. 62
  63. 63
  64. 64
  65. 65
  66. 66
  67. 67
  68. 68
  69. 69
    Mullen Scales of Early Learning
    1. EM Mullen
    (1995)
    Circle Pine: American Guidance Service, Inc.
  70. 70
  71. 71
  72. 72
  73. 73
    The default mode network in autism
    1. A Padmanabhan
    2. CJ Lynch
    3. M Schaer
    4. V Menon
    (2017)
    Biological Psychiatry: Cognitive Neuroscience and Neuroimaging 2:476–486.
    https://doi.org/10.1016/j.bpsc.2017.04.004
  74. 74
  75. 75
  76. 76
  77. 77
  78. 78
  79. 79
  80. 80
  81. 81
  82. 82
  83. 83
  84. 84
  85. 85
  86. 86
  87. 87
  88. 88
  89. 89
  90. 90
  91. 91
  92. 92
  93. 93
  94. 94
  95. 95
  96. 96
  97. 97
    Vineland Scales of Adaptive Behavior: Interview Edition, Survey Form Manual
    1. S Sparrow
    2. D Balla
    3. D Cicchetti
    (1984)
    Circle Pines: American Guidance Service.
  98. 98
  99. 99
  100. 100
  101. 101
  102. 102
  103. 103
  104. 104
  105. 105
  106. 106
    Communication and Symbolic Behavior Scales Developmental Profile, First Normed Edition
    1. A Wetherby
    2. B Prizant
    (2002)
    Baltimore: Paul H. Brookes.
  107. 107
  108. 108

Decision letter

  1. Christian Büchel
    Senior and Reviewing Editor; University Medical Center Hamburg-Eppendorf, Germany
  2. Leo Schilbach
    Reviewer
  3. Danilo Bzdok
    Reviewer; RWTH Aachen, Germany

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

Acceptance summary:

This study reports important eye tracking and resting state fMRI results in a large cohort of toddlers with autism spectrum disorder (ASD). The authors used eye tracking to identify an ASD subtype characterized by early social visual engagement difficulties (GeoPref ASD) and can show that this is related to marked default mode network (DMN)-occipito-temporal cortex (OTC) hypoconnectivity. This DMN-OTC hypoconnectivity is also related to increased severity of social-communication difficulties in GeoPref ASD. This early and pronounced social-visual circuit hypoconnectivity could prove critical for future social-communicative development and could also lead to novel early interventions in these individuals.

Decision letter after peer review:

[Editors’ note: the authors were asked to provide a plan for revisions before the editors issued a final decision. What follows is the editors’ letter requesting such plan.]

Thank you for sending your article entitled "Reduced default mode-visual network functional connectivity in autistic toddlers with social engagement difficulties" for peer review at eLife. Your article is being evaluated by Christian Büchel as the Senior Editor, a Reviewing Editor, and three reviewers.

Given the list of essential revisions, possibly including new experiments, the editors and reviewers invite you to respond with an action plan and timetable for the completion of the additional work. We plan to share your responses with the reviewers and then issue a binding recommendation.

As you will see from the reviews below the main issues raised by the reviewers are the following:

1) Somewhat arbitrary choice of ADOS scale items.

This seems to be a somewhat arbitrary collection of items from the ADOS which needs to be clarified and motivated.

2) Problematic statistics for the comparison of the GeoPref ASD group (n=16) and the relatively large nonGeo ASD group (n=62).

The authors are encouraged to provide further evidence e.g. by using a non-parametric approach (e.g. bootstrapping).

3) Problematic model comparison using AIC.

The comparison of the stratified subtype model and the traditional unstratified case-control model with regards to DMN-OTC is an important finding. AIC as used can be problematic and the authors are encouraged to provide additional evidence for model superiority e.g. using BIC.

Reviewer #1:

The study by Lombardo et al. builds on their previous behavioral research which convincingly identified a subgroup of children with ASD who show a strong preference for dynamic, non-social geometric visual stimuli relative to videos of children in motion. Here, they examine whether there are specific neural signatures for this subgroup of children, referred to as "GeoPref ASD," relative to children with ASD who show a preference for social stimuli as well as several groups of control children. The authors report that the GeoPref ASD group has: (1) reduced social engagement, measured with item scores from the ADOS associated with primarily visual social functions, (2) reduced connectivity between occipital-temporal cortex (OTC) and the default mode network (DMN), and (3) subtype-specific associations for DMN-OTC connectivity.

This study represents an exciting and important direction for this line of research, and the imaging component of this work has many strengths: the extent of resting data collected in toddlers with ASD is impressive, and the brain networks approach is sophisticated and well-conceived. Unfortunately, there are a few critical issues that preclude a more favorable review here. First, there is concern with the social engagement subscale of the ADOS that the authors have used. On its surface, this seems to be a somewhat arbitrary collection of items from the ADOS which apparently has never been used before. For example, it is unclear why other items from the ADOS were not included in the author's social engagement subscale, such as Shared Enjoyment in Interaction and Requesting. Given the centrality of the social engagement measure for all of the brain and behavioral analyses, its arbitrariness and unknown clinical value represent substantial limitations of this work. Furthermore, it is misleading to state that "the GeoPref ASD subtype distinction is indeed a distinction that generalizes beyond eye tracking tests and shows differentiation in early social engagement difficulties, as measured by gold standard symptom based diagnostic measures" (subsection “Behavioral and developmental characteristics of the GeoPref ASD subtype”). While the author's social engagement measure is certainly derived from the gold standard symptom based diagnostic measure (i.e. the ADOS), it currently has no known clinical value. It would have been preferable if the authors had used a more established subscale from the ADOS, such as the social affect subscale, or another measure (e.g. SRS) as the primary behavioral measure here.

A second issue is related to the group comparison performed between the GeoPref ASD and the nonGeo ASD groups for the DMN-OTC connectivity data (i.e., Supplementary file 3), which is a critical one for the hypotheses detailed in this work. As noted by the authors, the sample size of the GeoPref ASD group is small (n=16), and the methods are not clear about what statistical method was used for this comparison. Given the distribution of DMN-OTC connectivity scores in Figure 2, which shows a high degree of overlap in DMN-OTC connectivity between the small GeoPref ASD group and the relatively large nonGeo ASD group (n=62), the p = 0.01 group difference result is surprising, and the statistic used to compute this comparison is a critical omission. I imagine this would not survive non-parametric testing in the event the n=16 group is not normally distributed, which the authors do not report.

A third issue is found in the Results where the authors compare the stratified subtype model and the traditional unstratified case-control model with regards to DMN-OTC, which is a critical result for this work. The authors used the Akaike Information Criteria (AIC) statistic to evaluate these two models and, according to the literature, "the individual AIC values are not interpretable as they contain arbitrary constants and are much affected by sample size" (Burnham and Anderson, 2004). Therefore ΔAIC between models needs to be computed. The authors did not report ΔAIC results. Moreover, when they are computed the key result for the DMN-OTC connectivity between stratified subtype model and the traditional unstratified case-control model yields ΔAIC = 3.72. Since Burnham and Anderson state that ΔAIC between 4-7 do not show strong support, it appears that these AIC results are not particularly compelling. It is possible that I am missing something here however it should be noted that the AIC methods are very brief and lack citations.

A fourth issue is related to a strongly worded statement regarding the behavioral data that I am not convinced is warranted given the results. The authors state that "Developmentally, the social engagement and other developmental domains show trajectories indicative of poorer development in GeoPref ASD" (subsection “Behavioral and developmental characteristics of the GeoPref ASD subtype”). This statement about trajectories would suggest that there is an [Age x Subgroup] interaction in the ANOVA results, however in the same paragraph it is stated that there is "no strong evidence of differences in slope of developmental trajectories between the two subtypes" (i.e., no [Age x Subgroup] interaction, Supplementary file 1). If there is no interaction, no claims regarding differential trajectories can be made.

Reviewer #2:

In this most interesting paper the authors describe their efforts of detecting an ASD subgroup that displays a specific and extreme lack of preference for socially compelling stimuli as measured by an eye tracking preferential looking paradigm (GeoPref). In previous studies the authors have demonstrated the utility of this task to separate subgroups within the ASD group, but also in comparison to non-ASD control groups with an impressively high specificity.

In the study here described the authors go on to relate the behavioral read out of their task to functional connectivity as assessed by resting state fMRI in toddlers with ASD and other control groups. Importantly, the authors also report efforts which were undertaken to investigate the clinical characteristics of the GeoPref ASD subgroup. Here, it was found that the GeoPref subgroup, indeed, shows more pronounced social impairments as measured by a selected group of ADOS items that the authors claim taps into social engagement difficulties. fMRI results demonstrate the ASD group as a whole exhibits connectivity reduction of the so-called default mode network (DMN) with visual and attention-related brain areas. The authors than included GeoPref for further stratification and report that this improves the findings for DMN-visual cortex connectivity. Finally, they also used a transdiagnostic model which used fixation data as the primary predictor and found that it does not outperform the subtype model.

Taken together, this is an exciting paper that describes advances in the characterization of a subgroup with ASD at the neural level.

1) How were the ADOS items chosen? What is the empirical evidence that makes it plausible to choose these particular items and not others?

2) Did the authors also compare subgroup-specific differences for the communication and social interaction scales of the ADOS? Does the GeoPref subgroup have higher ADOS scores overall?

3) Did the authors also explore the relationship of the GeoPref findings with parental report?

4) Recently, it has been suggested (Schilbach, 2019) that psychiatry needs to move towards observer-independent characterizations of social interaction behavior (e.g. via full body motion tracking) to provide a more comprehensive and clinically relevant analysis. Could this approach have helped the here described study? Could interaction-based phenotyping have helped the transdiagnostic analysis, because fixation data is too limited?

5) What is the clinical relevance of the additional fMRI results? It's great to see that the behavior differences are reflected at the level of the brain and it helps to provide a better characterization of the subgroup, but will this be relevant in the clinic?

6) In the comparison of ASD subtypes surrounding Figure 3 it might be helpful to include the data from TD toddlers to indicated how similar they are to the less affected ASD group.

7) Are the connectivity findings consistent with recent results from Dukart et al. (2019)?

8) The discussion of DMN findings in relation to social cognition and social interaction appears somewhat limited. Many other papers have addressed the relationship of physiological baseline of the brain and the psychological baseline of thinking about others and its relevance for different psychiatric disorders (e.g. Schilbach et al., 2008, Mars et al., 2012, Schilbach et al., 2015, Spreng, 2015).

9) There are recent findings for a subspecialization of different DMN nodes (Kernbach et al., 2018), which raises the question whether other targeted analyses could have been useful to further corroborate the current findings.

10) In my personal opinion the Discussion could be shortened, because it includes a relatively long coverage of putative genetic mechanisms and ideas for future research (iPSC), which are all very interesting, but not closely connected to the findings of the study.

Reviewer #3:

Lombardo, Pierce and colleagues provide cognitive experiments with eye tracking that build on a recently discovered eye-tacking based autism "subtype" – an eye tracking preferential looking paradigm. The authors thus replicate and extend this potential subgroup marker in a large independent sample (n=334). This work may contribute exciting new evidence on dysregulation between social visual engagement and the long-standing interest in higher social cognition deficits, like perspective taking, probably subserve by the default mode network at the earliest periods of development. Although this reviewer is favorable of the work, several shortcomings should be addressed for publication in eLife.

- Materials and methods section: The implementation of which linear model package was used is explained. However, several aspects of the non-imaging data analysis appear to be missing. Please include more details, including whether or not data columns have been z-scored, which dependent and independent variables are been fed into the model based on what rational.

A similar observation is made by this reviewer about the linear model applied to the partial correlation lower-triangle matrices – it remains largely unclear how the 'stratified' vs. 'unstratified' analysis models were set up. Please specify the exact input variables, output variables, including transformations and potential regularization schemes. Given the high number of component-component connectivity strengths, it is relevant to the reader how linear models were exactly estimated in this scenario of high dimensionality. Finally, how many separate models were estimated in each of these scenarios?

In short, mostly mentioning that the 'lm' function was used is insufficient, given that a large majority of quantitative models fit in the behavioral and biological sciences are some form of linear or generalized linear model.

- Discussion section: Please provide reflection and weighing of the results and findings given the uneven division of subject split into n=16 GeoPref ASD toddlers (11 male, 5 female) and n=62 nonGeo ASD. 16 may appear small, compared to the 195 overall participants.

- Potentially selective citations: other authors have built a body of work on gaze cognition and the brain, such as Schilbach et al., Vogeley et al., and other. The interpretation and introduction may profit from a more balanced relation of the current investigation to existing work.

- Materials and methods section: There may be a slight misunderstanding behind the meaning of bootstrapping as expressed in "to give ranges around the sample correlation estimates": the added value of running a bootstrap analysis in the present context relies on inference on the distribution of outcomes in participant samples that one could have observed.

- Discussion section: The conclusion of "default mode network is functionally disconnected with visual cortices and dorsal attention network on average in ASD" may be overstretched or at least imprecise. Some readers may take this as meaning that DMN and visual cortex are not connected at all in autism, while the authors found a statistically different strengths of functional connectivity between the DMN visual-related components.

[Editors’ note: formal revisions were requested, following approval of the authors’ plan of action.]

Thank you for providing a revision plan for your article titled “Reduced default mode-visual network functional connectivity in autistic toddlers with social engagement difficulties”. We are pleased to inform you that the editors and reviewers have approved your revision plan and look forward to receiving your revised article when ready. After assessing your response, the reviewers had some additional queries, copied below for reference. In your final "response to the reviewers" document, please also address these comments.

Additional essential revisions:

1) I feel that only established measures of social communication should be reported for individuals with ASD: using standard instruments is critical for interpretable clinical research. Therefore, I would greatly prefer if they did not report their novel measure of Social Engagement. Plus, I would like to verify that their "ADOS CoSo Total," an abbreviation that I am not familiar with, is the same as the Social Affect subscore of the ADOS, which is an established measure of social communication for individuals with ASD.

2) Given the small n (n=16) and the fact that Figure 2 appears to show outliers (see whiskers on box plots), a non-parametric 2-sample t-test (Mann-Whitney) seems appropriate.

3) I am confused by the author's response about interpretation of AIC. This is from the Burnham and Anderson (2004) paper they cite:

"The larger the Δi, the less plausible is fitted model i as being the best approximating model in the candidate set....Some simple rules of thumb are often useful in assessing the relative merits of models in the set: Models having Δi ≤ 2 have substantial support (evidence), those in which 4 ≤ Δi ≤7 have considerably less support, and models having Δi > 10 have essentially no support. These rough guidelines have similar counterparts in the Bayesian literature (Raftery 1996)."

My understanding is that this means that the ΔAICi reported by the authors (3.7) shows "considerably less support" that the fitted model is the best approximating model. I am not an expert in AIC so maybe I am missing something here.

[Editors’ note: this article was then rejected after discussions between the reviewers, but the authors were invited to resubmit after an appeal against the decision.]

Thank you for re-submitting your work entitled "Default mode-visual network hypoconnectivity in an autism subtype with pronounced social visual engagement difficulties" for consideration by eLife. Your article has been reviewed by a Senior Editor, a Reviewing Editor, and three reviewers.

Our decision has been reached after consultation between the reviewers. Based on these discussions and the individual reviews below, we regret to inform you that your work will not be considered further for publication in eLife.

As you can see from the individual comments all reviewers found the data-set to be very interesting and unique. However, the revisions did not convince two of the three reviewers as you can see from their comments. Robustness of findings is of great importance and especially reviewer #1 continues to be concerned about the statistical procedures, which resonates with the comments made by reviewer #3.

Reviewer #1:

While I appreciate that the authors removed the social engagement subscale, there are several critical items from my previous review of this work that I do not feel were sufficiently addressed in this resubmission:

1) It is not clear to me why a non-parametric 2-sample t-test (Mann-Whitney) test, the standard statistical measure in the context of a small n and outliers, was not performed on the critical GeoPref vs. nonGeo ASD group comparisons. This test was specifically requested in my previous review. If this is a robust brain signature for GeoPref ASD, this result should be significant using a Mann-Whitney measure as well as the reported bootstrapping results.

2) I remain concerned about the ΔAIC results. Based on Burnham and Anderson's heuristic, I would expect that strong support for the subtype model would be accompanied by a larger ΔAIC. The reported ΔAIC result (3.7) was not within the range for "considerably less support" (i.e., ΔAIC between 4-7) and is certainly well below the ΔAICs associated with "essentially no support." I would suggest that the editor consults with an expert on AIC to help resolve this issue.

3) I remain concerned about the weak theoretical framing surrounding the DMN-OTC results, a point that I highlighted in my initial review of this work. The DMN is typically considered a task-negative network and understanding its link with task-positive OTC in neurotypical individuals or individuals with ASD is not mentioned by the authors. This is critical given the importance of the DMN-OTC results for this paper.

Reviewer #2:

The authors have satisfactorily addressed all, but one of my previous comments:

I still feel that the discussion of putative genetic mechanisms is interesting, but not adequate in light of the results reported.

Reviewer #3:

This reviewer thanks the authors for their efforts on this first revision of the manuscript. Several of my core concerns of the previous version of the manuscript revolved around details and inconsistencies in the statistical modeling.

As one example from the revised manuscript:

"Both GLM models were implemented with the lm function in R. In these GLMs the partial correlation for a particular component pair was the dependent variable. […] The primary dependent variable of interest was the group variable."

Here, the authors first mention the connectivity strengths as the dependent variable first, and then instead mention the group variable. This, and other remaining problems (explanation of bootstrapping etc.) make me doubt that the results and conclusion stand on solid ground – which I have already expressed in detail regarding the previous version of the manuscript.

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

Thank you for resubmitting your work entitled "Default mode-visual network hypoconnectivity in an autism subtype with pronounced social visual engagement difficulties" for further consideration by eLife. Your appeal has been evaluated by Christian Büchel (Senior Editor) and two statistical Editors.

I am sorry that this took so long, but after consulting with one statistical editor, 2 of 3 reviewers remained skeptical and were still against publication of your paper and I therefore had to get a second statistical opinion.

In essence, both statistical editors agree with your line of reasoning concerning sampling procedures and the use of ΔAIC. Therefore, I am very happy to tell you that we are willing to consider publishing your paper. However, during the statistical review of your paper, it has been flagged that the paper often suggests causal relationships indicated by using the verb "drive" (including the abstract). However, the data do not allow this interpretation, because you are simply observing a correlation. This needs to be changed before acceptance.

https://doi.org/10.7554/eLife.47427.sa1

Author response

[Editors’ note: what follows is the authors’ plan to address the revisions.]

As you will see from the reviews below the main issues raised by the reviewers are the following:

1) Somewhat arbitrary choice of ADOS scale items.

This seems to be a somewhat arbitrary collection of items from the ADOS which needs to be clarified and motivated.

We can see from the detailed reviewer comments what motivates this first issue here regarding the choice of ADOS scale items that go into the social engagement index we have used. In summary, our response and action plan for this issue is that we can default back to simply using the ADOS social-communication total for the connectivity-behavior relationship analysis. Reviewer 1 had suggested this for the analysis of connectivity-behavioral relationships, and we don’t have any arguments against this. We would still like to use the social engagement index as an independent behavioral validation of differences between the eye tracking defined subtypes, since we feel it is an important point to show that the subtypes are also differentiated on a symptom-based index derived from a gold standard diagnostic measure like the ADOS. In terms of adding items to this index, we can also integrate the two items that reviewer 1 had suggested (shared enjoyment in interaction and requesting). Finally, in a revision, we can immediately add more details about the justifications for selecting these particular items for the social engagement index. In terms of a timetable for implementing these changes, we have already implemented them, and can insert them into a revision immediately.

Below you will find more detailed justifications that specifically answer this first point about what was the rationale for selecting those specific items. The quick answer is that we selected them based on face validity around the construct of social visual engagement. Below you can also find the statistical results for the connectivity-behavior relationships when we used a revised version of the social engagement index (incorporating shared enjoyment in interaction and requesting) and when the ADOS social-communication total was used.

– Justification for selection of specific items for the social engagement index

In the manuscript, we selected items from the ADOS for the social engagement index on the basis of their face validity in being relevant to the construct of social visual engagement. Social visual engagement was the construct under study, given the emphasis on the eye tracking defined subtypes. We had originally called it a “social orienting index”, but changed the label to “social engagement” due to other studies in the literature opting for this label rather than social orienting. Since not all items are purely visual in nature and may require other aspects of behavior (e.g., communication), we were being conservative in referring to this index simply as a “social engagement” index, rather than only being something pertaining purely to visual social engagement. However, our primary intention was for this to mainly be something pertaining to early social visual engagement behaviors.

Items such as eye contact and integration of eye gaze in social overtures are clear items relevant to visually attending to social world. Pointing, initiating and responding to joint attention, and showing are all relevant with regards to the manipulation of attention of self or others and/or the joint sharing of attention/focus between self and other. Response to name was the final key item that was chosen because it signifies attention to the social world by attending to auditory stimuli and then using such stimuli to behave in a way relevant to the stimulus (i.e. turn and look at the person calling their name). Of the total set of items on the ADOS, these were the select items that fulfilled our view of what we think is most relevant to “social visual engagement”, defined as attending and focusing on the social world around the child.

Other items that the reviewers had suggested such as shared enjoyment in interaction and requesting might be relevant to social engagement, but are less relevant to social visual engagement per se. To be conservative in the items we chose, the lack of visual engagement in these items is the primary reason why we did not include them. In particular, shared enjoyment in interaction also adds in an affective component (e.g., pleasure). While this is quite an important item, it is not as relevant to the way we had defined social visual engagement as attending and focusing on the social world around the child. Such a definition does not define what type of affect the children must show towards others. Requesting was also not included as it is more of a communicative item involving getting another person to do something for the child. The biggest distinction we can think of between requesting and what we would consider as “social visual engagement” is the difference between requesting and joint attention. Autistic children can show requesting behavior but without the key aspect of social intent to engage and focus on others or attempting to direct the other’s attention for a non-imperative purpose.

Thus, while these were our initial motivations, we can agree with the editor and reviewers that such a subset of items may not necessarily be independent from other “social” items in the sense of being statistically independent. In this sense, one could say our selection of items was arbitrary from a statistical standpoint, since it was not informed on any statistical basis, but rather was based on the face validity of important behaviors in the ADOS relevant to social visual engagement. Therefore, a much broader look at the social and communicative total from the ADOS could be examined. We did not initially examine this with regards to connectivity-behavior relationships in the first instance because our intention with the social engagement index was to first use it as a real-world symptom-based validation of the eye tracking defined subtypes. Second, we believed that an a priori index specific to the domain of social visual engagement was the best way to examine connectivity-behavior relationships. We have no issue with simply defaulting to reporting relationships between the ADOS social-communication total and connectivity. While there is a loss of specificity in the inferences being made, an analysis of relationship between connectivity and ADOS social-communication total would allow us to see if connectivity has an association with the broader symptom domain of which diagnosis is particularly reliant on.

– Results when a revised social engagement index or ADOS social-communication total is used in connectivity-behavior relationships

Below is a table summarizing analyses looking at connectivity-behavior relationships. The column “Prior version” shows the statistics from the last submission. The column “Revised version” shows the ADOS social engagement index when including the items of shared enjoyment in interaction and requesting. The final column “ADOS CoSo Total” shows statistics when the total ADOS score for all social-communicative algorithm items is used. These results show that DMN-OTC connectivity-social communication relationships for GeoPref ASD are just as strong (if not stronger), and thus the correlation here is not necessarily restricted solely to items with relevance to social visual engagement. Since there is no a priori reason to suspect on a statistical basis that the selected items are independent of other items on the ADOS, and because the ADOS algorithm holds a much stronger clinical foothold in past research, we can agree with the editor and reviewers that probably the most principled thing to do is to proceed using the ADOS social-communication algorithm total in these analyses.

Prior versionRevised versionADOS CoSo Total
GeoPref ASD
IC10-IC02 (DMN-OTC)
r = -0.61, p = 0.025,
CI = [-0.97, -0.23]
r = -0.67, p = 0.009,
CI = [-0.95, -0.32]
r = -0.78, p = 0.001,
CI = [-0.99, -0.35]
nonGeo ASD
IC10-IC02 (DMN-OTC)
r = 0.03, p = 0.80,
CI = [-0.26, 0.33]
r = 0.07, p = 0.59,
CI = [-0.22, 0.36]
r = 0.06, p = 0.64,
CI = [-0.20, 0.33]
GeoPref ASD
IC10-IC05 (DMN-PVC)
r = -0.10, p = 0.74,
CI = [-0.95, 0.67]
r = -0.05, p = 0.85,
CI = [-0.88, 0.72]
r = 0.04, p = 0.88,
CI = [-0.81, 0.75]
nonGeo ASD
IC10-IC05 (DMN-PVC)
r = 0.13, p = 0.32,
CI = [-0.17, 0.39]
r = 0.11, p = 0.42,
CI = [-0.21, 0.39]
r = 0.22, p = 0.12,
CI = [-0.10, 0.49]
GeoPref ASD
IC10-IC09 (DMN-DAN)
r = -0.09, p = 0.75,
CI = [-0.94, 0.57]
r = -0.26, p = 0.41,
CI = [-0.95, 0.44]
r = -0.33, p = 0.25,
CI = [-0.83, 0.58]
nonGeo ASD
IC10-IC09 (DMN-DAN)
r = 0.25, p = 0.07,
CI = [-0.02, 0.46]
r = 0.29, p = 0.035,
CI = [0.09, 0.49]
r = 0.29, p = 0.045,
CI = [0.05, 0.49]

2) Problematic statistics for the comparison of the GeoPref ASD group (n=16) and the relatively large nonGeo ASD group (n=62). The authors are encouraged to provide further evidence e.g. by using a non-parametric approach (e.g. bootstrapping)

We have now used a non-parametric approach to hypothesis testing (e.g., permutation test). After 10,000 permutations, we can assign a p-value to each hypothesis test by comparing the actual t-statistic under the true group labels with the null distribution of t-statistics recomputed when the group labels are randomly shuffled. This approach results in the same inferences as the prior approach with parametric hypothesis tests. See the tables presented below.

Because the results are effectively the same, we leave it to the decision of the editor and reviewers whether you would like us to use the parametric or non-parametric results. In terms of a timetable, these revisions to the analyses have already been coded up and computed and can be included in a revision immediately.

IC10-IC02 (DMN-OTC)

Comparisont-stat (parametric)p-value (parametric)p-value (non-parametric)FDR q (parametric)FDR q (non-parametric)
GeoPref ASD vs.nonGeo ASD *-2.65440.01310.01560.02620.0312
GeoPref ASD vs.LD/DD *-2.75160.01090.01060.02620.0265
GeoPref ASD vs.TypSibASD *-3.63080.00100.00170.00520.0085
GeoPref ASD vs.TD *-4.66310.00010.00040.00090.0040
nonGeo ASD vs.LD/DD-1.13290.27140.26650.38770.3807
nonGeo ASD vs.TypSibASD-1.72130.09640.09390.16070.1565
nonGeo ASD vs.TD *-2.70460.00790.00820.02620.0265
LD/DD vs.TypSibASD-0.10250.91920.91210.91920.9121
LD/DD vs.TD-0.27180.78880.78770.91920.9121
TypSibASD vs.TD-0.21820.82900.82380.91920.9121

IC10-IC05 (DMN-PVC)

Comparisont-stat (parametric)p-value (parametric)p-value (non-parametric)FDR q (parametric)FDR q (non-parametric)
GeoPref ASD vs.nonGeo ASD-0.54700.58940.59600.65490.6623
GeoPref ASD vs.LD/DD *-2.67990.01380.01180.04610.0345
GeoPref ASD vs.TypSibASD *-2.28020.03000.02840.05610.0487
GeoPref ASD vs.TD *-2.64390.01350.01210.04610.0345
nonGeo ASD vs.LD/DD *-2.60460.01890.01380.04730.0345
nonGeo ASD vs.TypSibASD *-2.26360.03360.02920.05610.0487
nonGeo ASD vs.TD *-3.06790.00270.00250.02710.0250
LD/DD vs.TypSibASD1.04240.30810.31020.38510.3877
LD/DD vs.TD1.31880.20470.20020.29240.2860
TypSibASD vs.TD0.25470.80110.80590.80110.8059

IC10-IC09 (DMN-DAN)

Comparisont-stat (parametric)p-value (parametric)p-value (non-parametric)FDR q (parametric)FDR q (non-parametric)
GeoPref ASD vs.nonGeo ASD-1.03270.31280.31480.39100.3935
GeoPref ASD vs.LD/DD-2.32830.02710.02820.09040.0940
GeoPref ASD vs.TypSibASD-1.76280.08870.08680.17750.1736
GeoPref ASD vs.TD *-2.85170.00850.00900.04240.0450
nonGeo ASD vs.LD/DD-1.93400.06710.06310.16770.1577
nonGeo ASD vs.TypSibASD-1.19090.24380.24040.39100.3935
nonGeo ASD vs.TD *-2.86840.00500.00550.04240.0450
LD/DD vs.TypSibASD0.86340.39570.40160.43960.4462
LD/DD vs.TD0.05350.95780.95680.95780.9568
TypSibASD vs.TD-1.09190.28280.28280.39100.3935

3) Problematic model comparison using AIC.

The comparison of the stratified subtype model and the traditional unstratified case-control model with regards to DMN-OTC is an important finding. AIC as used can be problematic and the authors are encouraged to provide additional evidence for model superiority e.g. using BIC.

To summarize our response on this issue, we argue that AIC is the best measure of model selection criteria in this specific context, particularly with respect to BIC. We also would like to thank the reviewer for pointing out that we should have gone beyond simply evaluating AIC based on which model produced the lowest AIC. Although the winner under that rule was the subtype model, ΔAIC is further evidence we should report. However, contrary to the ΔAIC measure showing that “the results are not that compelling”, this ΔAIC measure actually shows the opposite – there is indeed evidence to support that the case-control model is not as compelling a model as the best performing subtype model. As such, we do not see the need to use other metrics of model selection criteria (e.g., BIC). As action plan, we think it is appropriate to add to the manuscript the reviewer’s suggestion of the ΔAIC measure. This action can be implemented in a revision immediately. Below is a much more detailed response on each of these points.

– Using AIC, not BIC, in the case of complex “tapering effects”

We have used the AIC as a criterion for selecting whether the best model is a case-control or subtype model. We have not utilized other criteria like BIC because as stated by Burnham and Anderson (2004), BIC is best when one suspects that the true model of reality is one characterized by a few big effects, rather than complex effects (what Burnham and Anderson (2004) call “tapering effects”). Indeed, one philosophical problem about BIC’s use in practice is this assumption that the true model of reality be one that is not complex, since it is hardly ever the case that reality is not complex, at least with respect to phenotypes like autism. Burnham and Anderson (2004) exemplify this difference in assumptions about the nature of the true model by contrasting Figure 1 vs. 2 in their paper. BIC works best under the situation of their Figure 2 where the ground truth is that there are only a few large effects, and all other effects are negligible. AIC works best when the ground truth is likely indicative of complexity (or “tapering effects”) as shown in their Figure 1. Illustrating this with the exact words of Burnham and Anderson (2004):

“It is known (from numerous papers and simulations in the literature) that in the tapering-effects context (Figure 1), AIC performs better than BIC. If this is the context of one’s real data analysis, then AIC should be used” (Burnham and Anderson, 2004, pg. 285).

Given our current understanding of autism, the reality is likely one of complexity. In fact, the very reason for studying heterogeneity is because the effects underlying the label of autism are likely complex. There is no reason to suspect the underlying truth (i.e. full model) for explaining features of autism will be a simplistic model with few very large effects (see Happe, Ronald, & Plomin, 2006). For this reason, AIC is preferred as the best measure of model selection criterion in this situation, especially when compared to BIC.

The other issue to illustrate for why BIC should not be utilized is because the computation behind BIC can be considered as mathematically identical to AIC, but with a primary difference in the penalty parameter (k) for model complexity. For AIC, k = 2, whereas for BIC k = log(n) and is typically much larger than k for AIC. Burnham and Anderson (2004) describe this difference as basically an argument that reduces down to the priors on the models (and they illustrate in their paper that AIC is just as Bayesian as BIC). BIC has a prior that the best model should be more simplistic with few large effects, hence a larger penalty parameter for complexity, whereas AIC has the prior that the model should likely be complex, and hence there is a smaller penalty parameter for complexity of the model. It makes little logical sense for us to use AIC and then compute BIC again, which is simply another computation of AIC with a different penalty parameter that implicitly assumes the true model should be simple with few large effects. For all the reasons mentioned above, this is why we prefer not to confuse the situation by reporting both model selection criterion. Had we needed to provide the BIC for the DMN-OTC models, it would have shown that the subtype model produces a lower BIC than the case-control model for DMN-OTC (subtype BIC = -26.53; case-control BIC = -25.90). As Burnham and Anderson (2004) note, the best model to select for BIC is the one producing the lowest BIC value. Hence, even here with BIC, the overall result is the same as AIC in showing that the subtype model is better than the case-control model. However, we do not see how this is necessarily helpful and actually only introduces more confusion, as conceptually there are issues with using BIC in our context when we are fairly certain the situation in autism is contrary to the expectations of BIC.

– ΔAIC supports the subtype model is better than the case-control model

Finally, we also would like to point out that contrary to the reviewer comment about ΔAIC, this evidence actually supports the idea that the subtype model is indeed a better model and that there is considerably less support for the idea that the case-control model is as good. Burnham and Anderson (2004) outline that ΔAIC is computed as follows: ΔAICi = AICi – AICmin. AICmin is the AIC for the model with the lowest AIC – that is, the subtype model for DMN-OTC. AICi is then the AIC for the case-control model. ΔAICi is the metric used to make an inference about whether there is evidence supporting that the case-control model (AICi) is as good as the subtype model (AICmin). The ΔAICi = 3.72. Burnham and Anderson (2004) give this heuristic advice for interpreting ΔAICi:

“Models having Δi ≤ 2 have substantial support (evidence), those in which 4 ≤ Δi ≤ 7 have considerably less support, and models having Δi > 10 have essentially no support”(Burnham and Anderson, 2004, pg. 271).

In other words, a model compared to the model with the lowest AIC (i.e. the best model) that results in 4 ≤ Δi ≤ 7 has considerably less support than the best model with the lowest AIC. This means there is “considerably less support” for the case-control model (AICi) compared to the subtype model (AICmin). This contradicts the statement in review 1, it appears that these AIC results are not particularly compelling. The opposite is actually the case – the evidence here suggests that there is considerably less support for the case-control model being as good as the best model (AICmin), the subtype model. Therefore, AIC is the best model selection criterion for us to use in this context, and the evidence is clear that the subtype model is a better model than the case-control model for the DMN-OTC comparison.

[Editors’ notes: the authors’ response after being formally invited to submit a revised submission follows.]

As you will see from the reviews below the main issues raised by the reviewers are the following:

1) Somewhat arbitrary choice of ADOS scale items.

This seems to be a somewhat arbitrary collection of items from the ADOS which needs to be clarified and motivated.

Per the request of the reviewers, in the revision we have now used the ADOS social affect total and have dropped the usage of an ADOS social engagement index.

2) Problematic statistics for the comparison of the GeoPref ASD group (n=16) and the relatively large nonGeo ASD group (n=62). The authors are encouraged to provide further evidence e.g. by using a non-parametric approach (e.g. bootstrapping)

We have now used a non-parametric approach to hypothesis testing (e.g., permutation test). After 10,000 permutations, we can assign a p-value to each hypothesis test by comparing the actual t-statistic under the true group labels with the null distribution of t-statistics recomputed when the group labels are randomly shuffled. This approach results in the same inferences as the prior approach with parametric hypothesis tests. See the tables presented in our revision plan. As suggested by the reviewers we utilize these non-parametric results in the revised manuscript.

3) Problematic model comparison using AIC.

The comparison of the stratified subtype model and the traditional unstratified case-control model with regards to DMN-OTC is an important finding. AIC as used can be problematic and the authors are encouraged to provide additional evidence for model superiority e.g. using BIC.

Regarding whether to use AIC or BIC, Burnham and Anderson (2004) noted:

“It is known (from numerous papers and simulations in the literature) that in the tapering-effects context (Figure 1), AIC performs better than BIC. If this is the context of one’s real data analysis, then AIC should be used”.

Burnham and Anderson (2004) use the terminology of “tapering effects” to mean a context with high complexity. In these situations, AIC is preferred over BIC. BIC is preferred in situations where the true reality is one where a few large effects dominate. Given our current understanding of autism, the reality is likely one of complexity. In fact, the very reason for studying heterogeneity is because the effects underlying the label of autism are likely complex. There is no reason to suspect the underlying truth (i.e. full model) for explaining features of autism will be a simplistic model with few very large effects (see Happe, Ronald and Plomin, 2006). For this reason, AIC is preferred as the best measure of model selection criterion in this situation, especially when compared to BIC.

The other issue to illustrate for why BIC should not be utilized is because the computation behind BIC is mathematically identical to AIC, but with a primary difference in the penalty parameter (k) for model complexity. For AIC, k = 2, whereas for BIC k = log(n) and is typically much larger than k for AIC. Burnham and Anderson (2004) describe this difference as basically an argument that reduces down to the priors on the models (and they illustrate in their paper that AIC is just as Bayesian as BIC). BIC has a prior that the best model should be more simplistic with a few large effects, hence a larger penalty parameter for complexity, whereas AIC has the prior that the model should likely be complex, and hence there is a smaller penalty parameter for complexity of the model. It makes little logical sense for us to use AIC and then compute BIC again, which is simply another computation of AIC with a different penalty parameter that implicitly assumes the true model should be simple with a few large effects.

Finally, we also would like to point out that contrary to the reviewer comment about ΔAIC, there is not compelling evidence that the case-control model is as good as the subtype model. Burnham and Anderson (2004) outline that ΔAIC is computed as follows: ΔAICi = AICi – AICmin. AICmin is the AIC for the model with the lowest AIC – that is, the subtype model for DMN-OTC. AICi is then the AIC for the case-control model. ΔAICi is the metric used to make an inference about whether there is evidence supporting that the case-control model (AICi) is as good as the subtype model (AICmin). The ΔAICi = 3.72. Burnham and Anderson (2004) give this heuristic advice for interpreting ΔAICi:

“Models having Δi ≤ 2 have substantial support (evidence), those in which 4 ≤ Δi ≤ 7 have considerably less support, and models having Δi > 10 have essentially no support.”

In other words, a model compared to the model with the lowest AIC (i.e. the best model) that results in 4 ≤ Δi ≤ 7 has considerably less support than the best model with the lowest AIC. The observed ΔAICi is closest to the above situation, and this would mean that there is “considerably less support” for the case-control model (AICi) compared to the subtype model (AICmin). This contradicts the statement in review 1, “it appears that these AIC results are not particularly compelling”. The opposite is actually the case – the evidence here suggests that there is considerably less support for the case-control model being as good as the best model (AICmin), the subtype model.

1) I feel that only established measures of social communication should be reported for individuals with ASD: using standard instruments is critical for interpretable clinical research. Therefore, I would greatly prefer if they did not report their novel measure of Social Engagement. Plus, I would like to verify that their "ADOS CoSo Total," an abbreviation that I am not familiar with, is the same as the Social Affect subscore of the ADOS, which is an established measure of social communication for individuals with ASD.

We now utilize the ADOS Social Affect score and have dropped the social engagement index. Our abbreviation of ADOS CoSo Total is indeed the same as the Social Affect subscore.

2) Given the small n (n=16) and the fact that Figure 2 appears to show outliers (see whiskers on box plots), a non-parametric 2-sample t-test (Mann-Whitney) seems appropriate.

As noted above, we have now utilized the non-parametric permutation tests as the primary tests of the specific between-group comparisons.

3) I am confused by the author's response about interpretation of AIC. This is from the Burnham and Anderson (2004) paper they cite:

"The larger the Δi, the less plausible is fitted model i as being the best approximating model in the candidate set....Some simple rules of thumb are often useful in assessing the relative merits of models in the set: Models having Δi ≤ 2 have substantial support (evidence), those in which 4 ≤ Δi ≤ 7 have considerably less support, and models having Δi > 10 have essentially no support. These rough guidelines have similar counterparts in the Bayesian literature (Raftery 1996)."

My understanding is that this means that the ΔAICi reported by the authors (3.7) shows "considerably less support" that the fitted model is the best approximating model. I am not an expert in AIC so maybe I am missing something here.

ΔAICi is a metric indicating whether model i shows compelling evidence for being a model as good as the model with the lowest AIC (e.g., AICmin). In this study, AICmin is the subtype model. Thus, ΔAICi indicates whether the case-control model is as good of a model as the subtype. The heuristic suggested by Burnham and Anderson is that 4 ≤ ΔAICi ≤ 7 indicates that there is considerably less support for model i (e.g., the case-control model) being as good as the best model (e.g., subtype model). Our ΔAICi = 3.72, and this is closest to the aforementioned heuristic suggested by Burnham and Anderson (2004). This ΔAICi would not fit for other heuristics (i.e. ΔAICi ≤ 2) whereby model i (the case-control model) has substantial support for being as good as the best model.

Reviewer #1:

[…] Unfortunately, there are a few critical issues that preclude a more favorable review here. First, there is concern with the social engagement subscale of the ADOS that the authors have used. On its surface, this seems to be a somewhat arbitrary collection of items from the ADOS which apparently has never been used before. For example, it is unclear why other items from the ADOS were not included in the author's social engagement subscale, such as Shared Enjoyment in Interaction and Requesting. Given the centrality of the social engagement measure for all of the brain and behavioral analyses, its arbitrariness and unknown clinical value represent substantial limitations of this work. Furthermore, it is misleading to state that "the GeoPref ASD subtype distinction is indeed a distinction that generalizes beyond eye tracking tests and shows differentiation in early social engagement difficulties, as measured by gold standard symptom based diagnostic measures" (subsection “Behavioral and developmental characteristics of the GeoPref ASD subtype”). While the author's social engagement measure is certainly derived from the gold standard symptom based diagnostic measure (i.e., the ADOS), it currently has no known clinical value. It would have been preferable if the authors had used a more established subscale from the ADOS, such as the social affect subscale, or another measure (e.g., SRS) as the primary behavioral measure here.

Per the request of the reviewers, in the revision we have now defaulted back to simply using the ADOS social affect total instead of the social engagement index.

A second issue is related to the group comparison performed between the GeoPref ASD and the nonGeo ASD groups for the DMN-OTC connectivity data (i.e., Supplementary file 3), which is a critical one for the hypotheses detailed in this work. As noted by the authors, the sample size of the GeoPref ASD group is small (n=16), and the methods are not clear about what statistical method was used for this comparison. Given the distribution of DMN-OTC connectivity scores in Figure 2, which shows a high degree of overlap in DMN-OTC connectivity between the small GeoPref ASD group and the relatively large nonGeo ASD group (n=62), the p = 0.01 group difference result is surprising, and the statistic used to compute this comparison is a critical omission. I imagine this would not survive non-parametric testing in the event the n=16 group is not normally distributed, which the authors do not report.

We have now used a non-parametric approach to hypothesis testing (e.g., permutation test). After 10,000 permutations, we can assign a p-value to each hypothesis test by comparing the actual t-statistic under the true group labels with the null distribution of t-statistics recomputed when the group labels are randomly shuffled. This approach results in the same inferences as the prior approach with parametric hypothesis tests. See the tables presented in the revision plan. As suggested by the reviewers we utilize these non-parametric results for the revised manuscript.

A third issue is found in the Results where the authors compare the stratified subtype model and the traditional unstratified case-control model with regards to DMN-OTC, which is a critical result for this work. The authors used the Akaike Information Criteria (AIC) statistic to evaluate these two models and, according to the literature, "the individual AIC values are not interpretable as they contain arbitrary constants and are much affected by sample size" (Burnham and Anderson, 2004). Therefore ΔAIC between models needs to be computed. The authors did not report ΔAIC results. Moreover, when they are computed the key result for the DMN-OTC connectivity between stratified subtype model and the traditional unstratified case-control model yields ΔAIC = 3.72. Since Burnham and Anderson state that ΔAIC between 4-7 do not show strong support, it appears that these AIC results are not particularly compelling. It is possible that I am missing something here however it should be noted that the AIC methods are very brief and lack citations.

We have now reported ΔAIC in the revised manuscript. However, this result is contrary to the reviewer suggests in this comment. This evidence supports the idea there is not compelling evidence that the case-control model is as good as the subtype model. Burnham and Anderson (2004) outline that ΔAIC is computed as follows: ΔAICi = AICi – AICmin. AICmin is the AIC for the model with the lowest AIC – that is, the subtype model for DMN-OTC. AICi is then the AIC for the case-control model. ΔAICi is the metric used to make an inference about whether there is evidence supporting that the case-control model (AICi) is as good as the subtype model (AICmin). The ΔAICi = 3.72. Burnham and Anderson (2004) give this heuristic advice for interpreting ΔAICi:

“Models having Δi ≤ 2 have substantial support (evidence), those in which 4 ≤ Δi ≤ 7 have considerably less support, and models having Δi > 10 have essentially no support.”

In other words, a model compared to the model with the lowest AIC (i.e. the best model) that results in 4 ≤ Δi ≤ 7 has considerably less support than the best model with the lowest AIC. The observed ΔAICi is closest to the above situation, and this would mean that there is “considerably less support” for the case-control model (AICi) compared to the subtype model (AICmin). This contradicts the statement in review 1, “it appears that these AIC results are not particularly compelling”. The opposite is actually the case – the evidence here suggests that there is considerably less support for the case-control model being as good as the best model (AICmin), the subtype model.

A fourth issue is related to a strongly worded statement regarding the behavioral data that I am not convinced is warranted given the results. The authors state that "Developmentally, the social engagement and other developmental domains show trajectories indicative of poorer development in GeoPref ASD" (subsection “Behavioral and developmental characteristics of the GeoPref ASD subtype”). This statement about trajectories would suggest that there is an [Age x Subgroup] interaction in the ANOVA results, however in the same paragraph it is stated that there is "no strong evidence of differences in slope of developmental trajectories between the two subtypes" (i.e., no [Age x Subgroup] interaction, Supplementary file 1). If there is no interaction, no claims regarding differential trajectories can be made.

We apologize for the confusion in how we have phrased this statement. We intended to state that the GeoPref ASD group is poorer across social engagement and other developmental domains across the age range studied (and as evidenced by the general group effect). We have reworded this part of the Results section in the revision to avoid this confusion.

Reviewer #2:

[…]

1) How were the ADOS items chosen? What is the empirical evidence that makes it plausible to choose these particular items and not others?

Per the request of the reviewers, in the revision we have now defaulted back to simply using the ADOS social affect total instead of the social engagement index.

2) Did the authors also compare subgroup-specific differences for the communication and social interaction scales of the ADOS? Does the GeoPref subgroup have higher ADOS scores overall?

These analyses were already reported in Figure 1C and Supplementary file 1 and Figure 1—figure supplement 1. The GeoPref ASD subtype did show overall higher levels of severity on the ADOS social affect and the repetitive restricted behaviors total, and there was no difference in the slopes of the trajectories between groups.

3) Did the authors also explore the relationship of the GeoPref findings with parental report?

We did not explore this relationship as parental report measures (e.g., the ADI-R) were not available.

4) Recently, it has been suggested (Schilbach, 2019) that psychiatry needs to move towards observer-independent characterizations of social interaction behavior (e.g. via full body motion tracking) to provide a more comprehensive and clinically relevant analysis. Could this approach have helped the here described study? Could interaction-based phenotyping have helped the transdiagnostic analysis, because fixation data is too limited?

Interaction-based methods would be great to study in the future. We have inserted comments about this in the revision.

5) What is the clinical relevance of the additional fMRI results? It's great to see that the behavior differences are reflected at the level of the brain and it helps to provide a better characterization of the subgroup, but will this be relevant in the clinic?

The fMRI data helps us to better answer hypotheses about whether phenotypically differentiated ASD subtypes are also different on-average at a neural level. We do not have evidence from classifiers or individualized prediction models that utilize rsfMRI data that could answer any questions relevant to real-world applications in clinical settings.

6) In the comparison of ASD subtypes surrounding Figure 3 it might be helpful to include the data from TD toddlers to indicated how similar they are to the less affected ASD group.

ADOS data in TD toddlers shows floor effects, since the ADOS is not sensitive for picking up typically-developing variability and is more tailored to assessing atypical social-communication and RRB behaviors in suspected ASD cases. Thus, because of the floor effects, the data cannot be analyzed in a similar way as those shown with ASD children in Figure 3.

7) Are the connectivity findings consistent with recent results from Dukart et al. (2019)?

Unfortunately, it is not possible to easily compare the current results with those of the recent paper by Holiga et al., 2019. First, the age ranges are completely different, with our dataset examining much younger toddlers with ASD, while Holiga et al. is restricted to examining individuals age 6 years or older. Second, Holiga et al. uses a different metric connectivity – weighted degree centrality (WDC). WDC is incompatible with the ICA between-network method we have used in this paper. With WDC as utilized by Holiga et al., one computes the sum of correlation coefficients between a seed voxel and all other voxels above some arbitrary pre-specified threshold (e.g., r > 0.25). This metric is much more global and less specific than connectivity between specific components or networks, as we have measured in the current study.

8) The discussion of DMN findings in relation to social cognition and social interaction appears somewhat limited. Many other papers have addressed the relationship of physiological baseline of the brain and the psychological baseline of thinking about others and its relevance for different psychiatric disorders (e.g. Schilbach et al., 2008, Mars et al., 2012, Schilbach et al., 2015, Spreng, 2015).

Yes, the DMN is an important network for social cognition, and much of our own past work has demonstrated this and how it may be affected in autism (e.g., Lombardo et al., 2010). We did not expand more on this in the Discussion section for reasons of length of the entire Discussion section. However, we have added more references specific to the topic of DMN regions and their role in autism.

9) There are recent findings for a subspecialization of different DMN nodes (Kernbach et al., 2018), which raises the question whether other targeted analyses could have been useful to further corroborate the current findings.

Certainly, there are ways of splitting the DMN into further subsystems. However, since we were using a data-driven method like ICA, we used the ICA-identified components for further downstream analysis. In the future it may be possible to have methods like ICA or others to search for more granular decompositions of networks and to split the DMN into subsystems. This is beyond the scope of the current investigation, but could be an important topic for future work, and we have specifically noted this in the revision.

10) In my personal opinion the Discussion could be shortened, because it includes a relatively long coverage of putative genetic mechanisms and ideas for future research (iPSC), which are all very interesting, but not closely connected to the findings of the study.

The paragraph discussing more molecular biological mechanisms is something we would like to retain in the Discussion, as it attempts to discuss what may be underlying these neural circuitry differences in ASD subtypes, and possible ways to move forward in future research to get a better grasp on what underlies these differences in living patients.

Reviewer #3:

[…]

- Materials and methods section: The implementation of which linear model package was used is explained. However, several aspects of the non-imaging data analysis appear to be missing. Please include more details, including whether or not data columns have been z-scored, which dependent and independent variables are been fed into the model based on what rational.

We have now included z-scoring of the data columns in the revision.

In these longitudinal analyses the only independent variables are age and group and the dependent variable for any particular model is whatever developmental measure we are investigating (e.g., Mullen subscales, Vineland subscales, ADOS total scores). We have made this much more explicit in the revision.

For those that are interested, the code for this and all other analyses are laid out in our GitHub repo for this paper: https://github.com/mvlombardo/geoprefrsfmri.

A similar observation is made by this reviewer about the linear model applied to the partial correlation lower-triangle matrices – it remains largely unclear how the 'stratified' vs. 'unstratified' analysis models were set up. Please specify the exact input variables, output variables, including transformations and potential regularization schemes. Given the high number of component-component connectivity strengths, it is relevant to the reader how linear models were exactly estimated in this scenario of high dimensionality. Finally, how many separate models were estimated in each of these scenarios?

The partial correlations matrices were extracted per each individual subject using the MATLAB script nets_netmats.m from the FSLNets MATLAB toolbox (https://fsl.fmrib.ox.ac.uk/fsl/fslwiki/FSLNets) with it set to run ridge regression (Tikhonov-regularization) with the rho parameter set to 1, as we have done in other independent studies (e.g., Lombardo et al., 2019). The input to this function is simply a [time, component] matrix, and the output is a [component, component] partial correlation matrix, where in each cell of that matrix is the partial correlation between that seed component and a target component, partialing out the effects of all other components, using ridge regression for the regularization.

Only one model was implemented. We did not implement separate models. This extraction of partial correlations has nothing to do with the stratification analyses that were done at the group-level, as all of these extractions of partial correlations are done within each subject and all of this occurs before any group-level modeling, which we do in R with the lm function for running a general linear model.

In the revision, we have stated explicitly the input and output variables for the general linear models examining case-control or subtype models. Lombardo et al. (2019).

In short, mostly mentioning that the 'lm' function was used is insufficient, given that a large majority of quantitative models fit in the behavioral and biological sciences are some form of linear or generalized linear model.

The group-level analysis in R was implemented using the lm function in R. The type of linear model lm can run depends on the formula it is given. Here it is a general linear model (GLM), and we have now provided more information in the revised manuscript so that this is clear. The primary independent variable of interest is a factor variable indicating case or control status or a factor variable indicating the specific group labels (e.g., ASD subtype labels), and there are additional covariates of no interest like scan age and sex. After estimating the general linear model, the model object can be passed to the anova function in R in order to extract F-stats and p-values.

- Discussion section: Please provide reflection and weighing of the results and findings given the uneven division of subject split into n=16 GeoPref ASD toddlers (11 male, 5 female) and n=62 nonGeo ASD. 16 may appear small, compared to the 195 overall participants.

In the revision we have discussed more about the uneven divisions between GeoPref and nonGeo ASD.

- Potentially selective citations: other authors have built a body of work on gaze cognition and the brain, such as Schilbach et al., Vogeley et al., and other. The interpretation and introduction may profit from a more balanced relation of the current investigation to existing work.

We have cited this work in the revision.

- Materials and methods section: There may be a slight misunderstanding behind the meaning of bootstrapping as expressed in "to give ranges around the sample correlation estimates": the added value of running a bootstrap analysis in the present context relies on inference on the distribution of outcomes in participant samples that one could have observed.

We have reworded this in the revision.

- Discussion section: The conclusion of "default mode network is functionally disconnected with visual cortices and dorsal attention network on average in ASD." may be overstretched or at least imprecise. Some readers may take this as meaning that DMN and visual cortex are not connected at all in autism, while the authors found a statistically different strengths of functional connectivity between the DMN visual-related components.

We have removed these types of statements in the revision.

[Editors’ note: the author responses to the second round of peer review follow.]

As you can see from the individual comments all reviewers found the data-set to be very interesting and unique. However, the revisions did not convince two of the three reviewers as you can see from their comments. Robustness of findings is of great importance and especially reviewer #1 continues to be concerned about the statistical procedures, which resonates with the comments made by reviewer #3.

Reviewer #1:

While I appreciate that the authors removed the social engagement subscale, there are several critical items from my previous review of this work that I do not feel were sufficiently addressed in this resubmission:

1) It is not clear to me why a non-parametric 2-sample t-test (Mann-Whitney) test, the standard statistical measure in the context of a small n and outliers, was not performed on the critical GeoPref vs. nonGeo ASD group comparisons. This test was specifically requested in my previous review. If this is a robust brain signature for GeoPref ASD, this result should be significant using a Mann-Whitney measure as well as the reported bootstrapping results.

We apologize for the confusion, but it was not clear to us that reviewer 1 was asking for the Mann-Whitney U test as the one and only test of non-parametric inference that would be valid.

The Mann-Whitney U test is a relatively antiquated method for running non-parametric hypothesis testing. It was developed during a time before the advent of fast computers to implement more exact computationally intensive resampling techniques. Most modern statisticians would recommend resampling methods like the permutation test rather than these older tests. This is what we have done.

Please see the full tables below that show the results of Mann-Whitney U tests alongside the original parametric tests and the permutation tests from the last revision. You can see that the Mann-Whitney U results show the same effects as the original parametric hypothesis tests we had initially reported, as well as the non-parametric permutation test results we included in the last revision.

Below are tables that show the results of Mann-Whitney U tests alongside the original parametric tests and the permutation tests from the last revision. In our revised manuscript we do not further include Mann-Whitney U tests, since they do not provide any additional useful information over and above the parametic and non-parametric permutation tests that were included in the previous revision.

IC10-IC02 (DMN-OTC)

Comparisont-stat (parametric)p-value (parameteric)p-value (perm)p-value (Mann-Whitney)FDR q (parametric)FDR q (perm)FDR q (Mann-Whitney)
GeoPref ASD vs.nonGeo ASD *-2.65440.01310.01560.01130.02620.03120.0304
GeoPref ASD vs.LD/DD *-2.75160.01090.01060.01520.02620.02650.0304
GeoPref ASD vs.TypSibASD *-3.63080.00100.00170.00100.00520.00850.0052
GeoPref ASD vs.TD *-4.66310.00010.00040.000090.00090.00400.0009
nonGeo ASD vs.LD/DD-1.13290.27140.26650.27140.38770.38070.3878
nonGeo ASD vs.TypSibASD-1.72130.09640.09390.14250.16070.15650.2375
nonGeo ASD vs.TD *-2.70460.00790.00820.01360.02620.02650.0304
LD/DD vs.TypSibASD-0.10250.91920.91210.95340.91920.91210.9534
LD/DD vs.TD-0.27180.78880.78770.85230.91920.91210.9470
TypSibASD vs.TD-0.21820.82900.82380.71530.91920.91210.8941

IC10-IC05 (DMN-PVC)

Comparisont-stat (parametric)p-value (parameteric)p-value (perm)p-value (Mann-Whitney)FDR q (parametric)FDR q (perm)FDR q (Mann-Whitney)
GeoPref ASD vs.nonGeo ASD-0.54700.58940.59600.94570.65490.66230.9835
GeoPref ASD vs.LD/DD *-2.67990.01380.01180.00820.04610.03450.0273
GeoPref ASD vs.TypSibASD-2.28020.03000.02840.03510.05610.04870.0599
GeoPref ASD vs.TD *-2.64390.01350.01210.01450.04610.03450.0364
nonGeo ASD vs.LD/DD *-2.60460.01890.01380.00190.04730.03450.0106
nonGeo ASD vs.TypSibASD-2.26360.03360.02920.03590.05610.04870.0599
nonGeo ASD vs.TD *-3.06790.00270.00250.00210.02710.02500.0106
LD/DD vs.TypSibASD1.04240.30810.31020.16270.38510.38770.2034
LD/DD vs.TD1.31880.20470.20020.04980.29240.28600.0712
TypSibASD vs.TD0.25470.80110.80590.98350.80110.80590.9835

IC10-IC09 (DMN-DAN)

Comparisont-stat (parametric)p-value (parameteric)p-value (perm)p-value (Mann-Whitney)FDR q (parametric)FDR q (perm)FDR q (Mann-Whitney)
GeoPref ASD vs.nonGeo ASD-1.03270.31280.31480.19170.39100.39350.3195
GeoPref ASD vs.LD/DD-2.32830.02710.02820.01700.09040.09400.0569
GeoPref ASD vs.TypSibASD-1.76280.08870.08680.06140.17750.17360.1536
GeoPref ASD vs.TD *-2.85170.00850.00900.00340.04240.04500.0345
nonGeo ASD vs.LD/DD-1.93400.06710.06310.11210.16770.15770.2243
nonGeo ASD vs.TypSibASD-1.19090.24380.24040.33750.39100.39350.4219
nonGeo ASD vs.TD *-2.86840.00500.00550.00910.04240.04500.0455
LD/DD vs.TypSibASD0.86340.39570.40160.62590.43960.44620.6955
LD/DD vs.TD0.05350.95780.95680.97710.95780.95680.9771
TypSibASD vs.TD-1.09190.28280.28280.32510.39100.39350.4219

2) I remain concerned about the ΔAIC results: Based on Burnham and Anderson's heuristic, I would expect that strong support for the subtype model would be accompanied by a larger ΔAIC. The reported ΔAIC result (3.7) was not within the range for "considerably less support" (i.e., ΔAIC between 4-7) and is certainly well below the ΔAICs associated with "essentially no support." I would suggest that the editor consults with an expert on AIC to help resolve this issue.

Because reviewer #1 remains uncertain about the AIC result, our response is that rather than continuing to discuss AIC, we can show other standard acceptable ways to compare models. One way is to simply compute the models on a subset of data (e.g., training data) and make predictions on held-out unseen data (e.g., test data), as is done with cross-validation techniques. Here we use 5-fold cross-validation to train the models on an 80% subset of the data, and then make predictions on the left-out 20% that it has not seen. The predictions that are made on this new held-out unseen data are then compared to the actual held-out unseen data and we can compute the prediction error for each model (e.g., mean squared prediction error, MSPE). In this context, it is simple to evaluate which model is best – the model that produces the lowest MSPE is the better model. We show that the subtype mode indeed produces lower MSPE than the case-control model for the DMN-OTC comparison (see second table below). Thus, we show with simple cross-validation that the subtype model for DMN-OTC makes better predictions on new unseen data than the case-control model.

Cross-validated (5-fold) mean squared prediction error (MSPE) for the subtype and case-control models.

DMN-OTCDMN-PVCDMN-DAN
Subtype Model0.04190.03280.0632
Case-Control Model0.04300.03260.0634

We can also compute mean absolute percentage error (MAPE) as perhaps a more interpretable quantity than MSPE. MAPE is computed as the mean absolute value of percentage error and is computed as MAPE = mean(abs((Ai – Pi)/Ai)*100), where Ai is actual test data point i, Pi is predicted test data point i, and abs refers to the absolute value. The model with the lowest MAPE is the best model to choose. MAPE is computed for each cross validation fold, and then the average is reported across folds. The difference in MAPE values also allows for interpretability with regards to how much of a reduction in percentage error is gained by the better model.

Cross-validated (5-fold) mean absolute percentage error (MAPE) for the subtype and case-control models.

DMN-OTCDMN-PVCDMN-DAN
Subtype Model125.5452152.8099156.0619
Case-Control Model135.2241152.1738152.8081

For DMN-OTC, MAPE was much lower for the subtype model (MAPE = 125.5452) compared to the case-control model (MAPE = 135.2241), indicating the subtype model reduced the absolute percentage of error by around 9.67%. However, for DMN-PVC and DMN-DAN, MAPE was lower for the case-control model (DMN-PVC subtype MAPE = 152.8099, case-control MAPE = 152.1738; DMN-DAN subtype MAPE = 156.0619, case-control MAPE = 152.8081), but in both cases the reduction in MAPE was <1%. In a revision we report cross validated MAPE to supplement the analyses showing ΔAIC.

3) I remain concerned about the weak theoretical framing surrounding the DMN-OTC results, a point that I highlighted in my initial review of this work. The DMN is typically considered a task-negative network, and understanding its link with task-positive OTC in neurotypical individuals or individuals with ASD is not mentioned by the authors. This is critical given the importance of the DMN-OTC results for this paper.

Reviewer #1 was concerned about the theoretical framing surrounding the DMN-OTC. Reviewer #1’s suggestion was that we introduce altogether new ideas that we did not originally have or justify at the beginning of this work, as part of the Introduction.

In our previous revision, we did our best to expand on our justifications and references in the Introduction, since the reviewer here is referring to previous comments about further expanding on ideas we had mentioned in the Introduction. However, it is clear from the reviewer’s comments here, that what they want is for us to address something they have in mind regarding task-negative versus task-positive networks. This was not part of our initial justifications. Going into detail like this that uses what is known about the results in an Introduction is an example of HARKing – Hypothesizing After Results are Known (Kerr, 1988). We therefore do not feel it is appropriate to cite this literature in the Introduction and, in the revised manuscript, we have not followed the reviewer’s suggestion here on this matter.

Reviewer #3:

This reviewer thanks the authors for their efforts on this first revision of the manuscript. Several of my core concerns of the previous version of the manuscript revolved around details and inconsistencies in the statistical modeling.

As one example from the revised manuscript:

"Both GLM models were implemented with the lm function in R. In these GLMs the partial correlation for a particular component pair was the dependent variable.. […] The primary dependent variable of interest was the group variable."

Here, the authors first mention the connectivity strengths as the dependent variable first, and then instead mention the group variable… This, and other remaining problems (explanation of bootstrapping etc.) make me doubt that the results and conclusion stand on solid ground – which I have already expressed in detail regarding the previous version of the manuscript.

This reviewer’s concern arose because we made a typo in our revision where we accidentally wrote “dependent” instead of “independent”.

Specifically, we wrote, “The primary dependent variable of interest was the group variable.” It should be, “The primary independent variable of interest was the group variable.”

This typo lead the reviewer to have concerns about the statistical modeling, but our modeling is solid and all the code for the analyses is all available on GitHub (https://github.com/mvlombardo/geoprefrsfmri) and was provided in the last revision. It is clear in the code how we did the analysis, including that “The primary independent variable of interest was the group variable.”

We are sorry for the typo and in a revised manuscript we have corrected this typo and have further made sure that all language referring to the analyses are clear.

Regarding the way bootstrapping is explained, our bootstrapping analysis helps us to generate a sampling distribution around the test statistic in question. The 95% CI bounds on this bootstrapped sampling distribution are reported, to give the reader an idea of how variable the test statistic could be with respect to its bootstrapped sampling distribution. This is consistent with what we have written in the revision.

https://doi.org/10.7554/eLife.47427.sa2

Article and author information

Author details

  1. Michael V Lombardo

    1. Laboratory for Autism and Neurodevelopmental Disorders, Center for Neuroscience and Cognitive Systems @UniTn, Istituto Italiano di Tecnologia, Rovereto, Italy
    2. Autism Research Centre, Department of Psychiatry, University of Cambridge, Cambridge, United Kingdom
    Contribution
    Conceptualization, Data curation, Software, Formal analysis, Funding acquisition, Visualization, Methodology, Writing—original draft, Writing—review and editing
    For correspondence
    mvlombardo@gmail.com
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0001-6780-8619
  2. Lisa Eyler

    1. Department of Psychiatry, University of California, San Diego, San Diego, United States
    2. VISN 22 Mental Illness Research, Education, and Clinical Center, VA San Diego Healthcare System, San Diego, United States
    Contribution
    Investigation, Methodology, Writing—review and editing, Data collection
    Competing interests
    No competing interests declared
  3. Adrienne Moore

    Autism Center of Excellence, Department of Neuroscience, University of California, San Diego, San Diego, United States
    Contribution
    Writing—review and editing, Data collection
    Competing interests
    No competing interests declared
  4. Michael Datko

    Autism Center of Excellence, Department of Neuroscience, University of California, San Diego, San Diego, United States
    Contribution
    Methodology, Writing—review and editing
    Competing interests
    No competing interests declared
  5. Cynthia Carter Barnes

    Autism Center of Excellence, Department of Neuroscience, University of California, San Diego, San Diego, United States
    Contribution
    Writing—review and editing, Data collection
    Competing interests
    No competing interests declared
  6. Debra Cha

    Autism Center of Excellence, Department of Neuroscience, University of California, San Diego, San Diego, United States
    Contribution
    Writing—review and editing, Data collection
    Competing interests
    No competing interests declared
  7. Eric Courchesne

    Autism Center of Excellence, Department of Neuroscience, University of California, San Diego, San Diego, United States
    Contribution
    Conceptualization, Supervision, Funding acquisition, Methodology, Writing—original draft, Project administration, Writing—review and editing, Data collection
    Competing interests
    No competing interests declared
  8. Karen Pierce

    Autism Center of Excellence, Department of Neuroscience, University of California, San Diego, San Diego, United States
    Contribution
    Conceptualization, Supervision, Funding acquisition, Methodology, Writing—original draft, Project administration, Writing—review and editing, Data collection
    For correspondence
    kpierce@ucsd.edu
    Competing interests
    An invention disclosure form was filed by KP with the University of California, San Diego, on March 5, 2010.

Funding

H2020 European Research Council (755816)

  • Michael V Lombardo

National Institute of Mental Health (R01-MH080134)

  • Eric Courchesne
  • Karen Pierce

National Institute of Mental Health (P50-MH081755)

  • Eric Courchesne
  • Karen Pierce

National Institute on Deafness and Other Communication Disorders (R01-DC016385)

  • Michael V Lombardo
  • Eric Courchesne
  • Karen Pierce

CDMRP (AR130409)

  • Eric Courchesne

Jesus College, University of Cambridge (Fellowship)

  • Michael V Lombardo

British Academy (Fellowship)

  • Michael V Lombardo

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

Acknowledgements

This research was supported by grants NIMH R01-MH080134 (EC, KP), NFAR grant (KP), NIMH Autism Center of Excellence grant P50-MH081755 (EC, KP), NIDCD R01-DC016385 (EC, KP), CDMRP AR130409 (EC), and an ERC Starting Grant (ERC-2017-STG; 755816) to MVL and fellowships from Jesus College, Cambridge and the British Academy to MVL. We thank Richard Znamirowski, Clelia Ahrens-Barbeau, Stephanie Solso, Kathleen Campbell, Maisi Mayo, and Julia Young for help with data collection, Stuart Spendlove and Melanie Weinfeld for assistance with clinical characterization of subjects.

Ethics

Human subjects: This study was approved by the Institutional Review Board at University of California, San Diego (UCSD Human Research Protection Program protocols 091539, 081722, or 110049). Parents provided written informed consent according to the Declaration of Helsinki and were paid for their participation.

Senior and Reviewing Editor

  1. Christian Büchel, University Medical Center Hamburg-Eppendorf, Germany

Reviewers

  1. Leo Schilbach
  2. Danilo Bzdok, RWTH Aachen, Germany

Publication history

  1. Received: April 4, 2019
  2. Accepted: November 8, 2019
  3. Version of Record published: December 17, 2019 (version 1)

Copyright

© 2019, Lombardo 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

  • 855
    Page views
  • 99
    Downloads
  • 2
    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. Human Biology and Medicine
    Abidemi Junaid et al.
    Research Article

    TNFa signaling in the vascular endothelium elicits multiple inflammatory responses that drive vascular destabilization and leakage. Bioactive lipids are main drivers of these processes. In vitro mechanistic studies of bioactive lipids have been largely based on two-dimensional endothelial cell cultures that, due to lack of laminar flow and the growth of the cells on non-compliant stiff substrates, often display a pro-inflammatory phenotype. This complicates the assessment of inflammatory processes. Three-dimensional microvessels-on-a-chip models provide a unique opportunity to generate endothelial microvessels in a more physiological environment. Using an optimized targeted liquid chromatography-tandem mass spectrometry measurements of a panel of pro- and anti-inflammatory bioactive lipids, we measure the profile changes upon administration of TNFa. We demonstrate that bioactive lipid profiles can be readily detected from three-dimensional microvessels-on-a-chip and display a more dynamic, less inflammatory response to TNFa, that resembles more the human situation, compared to classical two-dimensional endothelial cell cultures.

    1. Human Biology and Medicine
    2. Neuroscience
    Charlotte M Pretzsch, Dorothea L Floris
    Insight

    A metric called the Hurst exponent could be a useful biomarker for studies exploring brain differences between men and women with autism spectrum disorder.