A DATA DRIVEN APPROACH REVEALS THAT ANOMALOUS MOTOR SYSTEM CONNECTIVITY IS ASSOCIATED WITH THE SEVERITY OF CORE AUTISM SYMPTOMS

Participants

The Johns Hopkins Medical Institutional Review Board approved this experiment. Oral assent and written informed consent were obtained from all participants and their legal guardians.

Four hundred and nineteen 8- to 12-year-old children participated at the Kennedy Krieger Institute (Table 1): 105 children with ASD (age=10.4 ± 1.3 years) and 314 typically developing (TD) children (age=10.3 ± 1.2 years), with the groups balanced on age and handedness assessed using the Edinburgh Handedness Inventory (Oldfield, 1971). Intellectual ability was assessed using the General Ability Index (GAI) derived from Weschlers Intelligence Scale for Children, fourth or fifth editions (WISC-IV, -V).

Diagnosis

Diagnosis Children with ASD met DSM-IV or −5 criteria for autism spectrum disorder (ASD) (American Psychiatric Association, 2013). Diagnoses were confirmed using the Autism Diagnostic Observation Schedule-Generic (Lord et al., 2000), the Autism Diagnostic Interview-Revised (Lord, Rutter, & Le Couteur, 1994), or both and review by a child neurologist with more than 30 years of experience with ASD diagnosis (SHM). Children with ASD were excluded if they had identifiable causes of autism or known neurological disorders, including epilepsy, based on parent responses during an initial phone screening. Children were excluded from the TD group if they had a first-degree relative with ASD or if parent responses to the Diagnostic Interview for Children and Adolescents-IV (Welner, Reich, Herjanic, Jung, & Amado, 1987) or Kiddie Schedule for Affective Disorders (KSADS) (Chambers et al., 1985), depending on the date of enrollment, revealed a history of a developmental or psychiatric disorder, with the exception of simple phobias. Social deficits were quantified using the Social Responsiveness Scale (SRS) or SRS-2 questionnaire (Constantino et al., 2003) and restrictive and repetitive behaviors (RRBs) were quantified using the total score from the Repetitive Behavior Scale-Revised (RBS-R) (Bodfish, Symons, Parker, & Lewis, 2000).

Medication

Over half of children in the ASD group (n = 61) were actively taking psychoactive medications including stimulants (n = 29), selective serotonin reuptake inhibitors (SSRI) (n = 21), clonidine (n = 3), risperidone (n = 4), and lithium (n = 1). Those taking stimulant medications were asked to withhold their stimulant medications the day of and the day prior to participating in the study to avoid effects of stimulants on cognitive, behavioral, and motor measures. All other medications were continued as prescribed.

Motor Assessments

fMRI Acquisition and Preprocessing

All participants completed a mock scanning session to habituate to the MRI environment. Scans were acquired on a Phillips 3T scanner (Achieva; Philips Healthcare, Best, The Netherlands) using a single-shot, partially parallel, gradient-recalled echo planar sequence with sensitivity encoding (repetition time [TR]/echo time = 2500/30 ms, flip angle = 70°, sensitivity encoding acceleration factor of 2, 3-mm axial slices with no slice gap, in-plane resolution of 3.05 × 3.15 mm [84 × 81 voxels]), and either an 8-channel (ASD: n = 62; TD: n = 165) or a 32-channel (ASD: n = 43; TD: n = 149) head coil. An ascending slice order was used, and the first 10 seconds were discarded at the time of acquisition to allow for magnetization stabilization. Children were instructed to relax and focus on a crosshair while remaining as still as possible. Preprocessing was completed using SPM12 (Wellcome Trust Centre for Neuroimaging, London, United Kingdom) and custom code written in MATLAB (The Mathworks, Inc., Natick, Massachusetts). Resting-state-fMRI scans were slice-time adjusted using the slice acquired at the middle of the TR as a reference, and rigid body realignment parameters were estimated to adjust for motion. The volume collected in the middle of the scan was spatially normalized using the Montreal Neurological Institute (MNI) EPI template (Calhoun et al., 2017). The estimated rigid body and nonlinear spatial transformations were applied to the functional data, producing 2-mm isotropic voxels. Linear trends were removed, the data were spatially smoothed using a Gaussian filter (6-mm full width at half maximum kernel), and voxel time series were variance normalized. Prior to imaging analysis, all scans were quality checked by two trained research assistants. Each image was checked to ensure full coverage of the cerebellum. No data included in subsequent analyses contained between-volume translational movements larger than 3 mm or rotational movements larger than 3 degrees. Mean framewise displacement (FD) was calculated using the realignment estimates (Power, Barnes, Snyder, Schlaggar, & Petersen, 2012). Several volumes were discarded from the beginning/end of four scans to uphold these criteria.

Group Independent Component Analysis (ICA)

Spatially Constrained Group ICA

Data were decomposed into temporally coherent networks using the Group ICA of fMRI Toolbox (GIFT v3.0b: https://trendscenter.org/software/gift/; Medical Image Analysis Lab, Albuquerque, New Mexico) (Calhoun, Adali, Pearlson, & Pekar, 2001; Erhardt et al., 2011). Spatial independent component analysis (ICA) estimates maximally independent spatial source maps (SMs) in fMRI data from their linearly mixed fMRI signal time courses (TCs). First, to estimate model order we used an information theoretic approach (Li, Adalı, & Calhoun, 2007) while restraining the algorithm to voxels in our motor system mask. We chose the number of independent components (ICs) for the group to be the maximum dimension estimate across participants (model order = 48) (Nebel et al., 2016).

Prior to ICA, each participant’s preprocessed data were reduced to 96 principle components (PCs) using principal component analysis (PCA) to decrease computational demands, denoise the fMRI signal, and improve back-reconstruction of participant-specific SMs and TCs (Erhardt et al., 2011). Participant-specific PCs were temporally concatenated and a second PCA reduced the data to 48 PCs using multi-power iteration (Rachakonda, Silva, Liu, & Calhoun, 2016). ICA was repeated on these 48 group-level PCs 100 times using the Infomax algorithm (Bell & Sejnowski, 1995) with randomized initial conditions in GIFT (Himberg, Hyvärinen, & Esposito, 2004). ICs were clustered across iterations, modes of each cluster were identified, and aggregate SMs were taken from the iteration with the maximum number of cluster modes (Himberg et al., 2004). Participant-specific SMs and TCs were then generated from the aggregate IC decomposition using a method based on PCA compression and projection (V.D. Calhoun et al., 2001).

Manual Inspection of Independent Components

Highly stable ICA decomposition (cluster quality index > 0.9) was observed for all but three components that were subsequently removed from further analysis. The remaining 45 ICs were manually inspected and hand-classified as signal or noise using spatial map and spectral plot features (Griffanti et al., 2017). Spectral features including fractional amplitude of low-frequency fluctuations (fALFF), ratio of spectral power below 0.1 Hz to spectral power between 0.15 and 0.25 Hz (Allen et al., 2011), and dynamic range, the difference between peak and minimum spectral power, are shown for each IC in Figure 2. Ten components were classified as noise resulting in 35 ICs classified as signal. The remaining 35 motor ICs and the MNI coordinates of the voxel contributing most strongly to each are shown in Figure 3.

Statistical Analyses

Within-Subnetwork Connectivity

A significant group difference was observed after FDR correction for only one SM (IC-24) that spatially corresponded with the right posterolateral cerebellum (Figure 4A), such that children with ASD showed weaker connectivity within a region spanning right Crus I and right Crus II (Figure 4D). From this point forward IC-24 will be referred to as right posterior cerebellum as the spatial extent of the cluster spans right lobule VI, Crus I, Crus II, and lobule VIIb (Figure 4C) (peak MNI: x = 32, y = −76, z = −38) (Diedrichsen et al., 2009). The spatial extent of the significant cluster from the right posterior cerebellum component is shown in Figure 4D and overlaid on the SUIT cerebellar atlas revealing that it spanned both right Crus I and right Crus II. Peak between-group differences (ASD < TD) in right Crus I within-network connectivity were observed at MNI coordinates [x = 24, y = −74, z = −38; t-statistic = −6.85] with the centroid of the cluster at MNI coordinates [x = 28, y = −70, z = −38; t-statistic = −4.51]. Mean z-score residuals from voxels within the t-map thresholded (t > μ + 4σ) right posterior cerebellum subnetwork for each child after regressing nuisance covariates (age, sex, MRI coil type, mean FD) are shown in Figure 4B.

Between-Subnetwork Connectivity

Connectivity between the cerebellar right posterior cerebellum component (IC-24) and all non-cerebellar motor system subnetworks (n = 25) was examined to identify group differences in the intrinsic synchrony between connectivity pairs. Using a MANCOVA, we observed a significant effect of diagnosis on functional connectivity between right posterior cerebellum (IC-24) and the non-cerebellar motor system components (n = 25) (F_(25,369) = 1.762, p = 0.014, partial η² = 0.107). Post-hocs suggest this effect was driven by group differences in right posterior cerebellum functional connectivity with four other motor components. As compared to TD children, children with ASD showed weaker right posterior cerebellum connectivity with left inferior parietal lobule (IPL) (p = 0.02, partial η² = 0.013; TD > ASD); bilateral dorsolateral premotor (DL-PM) (p = 0.02, partial η² = 0.013; TD > ASD); supplementary motor area (SMA) (p = 0.02, partial η² = 0.013; TD > ASD) and stronger connectivity with bilateral dorsomedial M1 (DM-M1) (p = 0.01, partial η² = 0.010; ASD > TD) (Figure 5).

Brain and Behavior Relationships

Brain and Motor Function Relationships

ASD-associated Disruption of Right Posterior Cerebellum Within-Network Connectivity

In examining within-network connectivity of the motor system components, we found a region within the right Crus I/II of children with ASD that contributed significantly less to the component as compared to TD children. Extant evidence suggests the Crus I/II region we identified (peak MNI: 24, −74, −38) is involved in higher-level cognitive and complex motor tasks as compared to simple motor tasks (Baumann et al., 2007; Ilg et al., 2013; Salmi et al., 2010) and that, at rest, this region is most functionally connected with the default mode network (Buckner et al., 2011). Further, the peak coordinates of group differences in the right Crus I/II cluster has been shown to be reciprocally connected to the mentalizing network, including the right/left temporoparietal junction (TPJ) and dorsolateral prefrontal cortex, and involved with processes important for social-cognition (Bernard et al., 2014; Van Overwalle & Mariën, 2016). The specificity of abnormal connectivity lateralized to the right posterior cerebellum component appears to support several structural MRI findings of specific gray matter deficits in right Crus I/II of children with ASD (D’Mello et al., 2015; Rojas et al., 2006; Stoodley, 2014; Wilson, Tregellas, Hagerman, Rogers, & Rojas, 2009; Yang et al., 2016) that are associated with both social-communication skills (D’Mello et al., 2015; Riva et al., 2013) and repetitive and stereotyped behaviors (D’Mello et al., 2015; Rojas et al., 2006). Although we did not examine the extent to which structural deficits within the right Crus I/II might contribute to decreased within-component connectivity, previous research suggests the specificity of pathological structural MRI findings are reflected in intra-network connectivity of ICA- derived intrinsic subsystems (Manoliu et al., 2014). Future studies should examine the extent to which right posterior cerebellum within-network connectivity disruptions observed in this study are associated with structural MRI findings in a large sample of children with ASD.

Anomalous Right Posterior Cerebellum Connectivity with the Motor System

The posterolateral regions of the cerebellum, including Crus I/II, contribute to higher-level social-cognitive (Stoodley & Schmahmann, 2009; Van Overwalle & Mariën, 2016) and visuomotor tracking/adaptation tasks (Neely, Coombes, Planetta, & Vaillancourt, 2013; Tzvi, Koeth, Karabanov, Siebner, & Krämer, 2020; Vaillancourt et al., 2006; Vaillancourt, Thulborn, & Corcos, 2003; Z. Wang et al., 2019), but not to simple pure motor tasks (Buckner et al., 2011; Salmi et al., 2010; Stoodley & Schmahmann, 2009). Viral-tracing and neuroimaging findings show the posterior cerebellum is reciprocally connected with multiple subnetworks of the motor system including the basal ganglia, primary motor (M1), premotor cortex, and inferior parietal lobule (IPL) (Bostan, Dum, & Strick, 2013; Buckner et al., 2011). In the current study, children with ASD, as compared to TD children, showed anomalous functional connectivity between the right posterior cerebellum and several other motor system subnetworks, specifically left IPL, dorsolateral premotor, dorsomedial M1, and the SMA – all regions involved visuomotor tasks (Neely et al., 2013; Vaillancourt et al., 2006, 2003; Z. Wang et al., 2019). Given the functional relevance of these cortical regions, our findings suggest disrupted functional connectivity involving the right posterior cerebellum may contribute to motor anticipatory, planning, execution, and learning deficits observed in children with ASD (Cook, Blakemore, & Press, 2013; Dziuk et al., 2007; Marko et al., 2015; Mosconi et al., 2015; Schmitz, Martineau, Barthélémy, & Assaiante, 2003). However, in the current study multiple regression analysis revealed no right posterior cerebellum-motor system connectivity pairs were significantly associated with performance on motor imitation and ball catching (Ament et al., 2015; Craig et al., 2018; MacNeil & Mostofsky, 2012). Our findings of no association of brain connectivity measures with motor abilities may result from poor discriminative ability of these motor tasks to detect small inter-individual differences between children. For example, our assessment of motor imitation is video coded with each gesture scored as correct or incorrect and ball catching performance is quantified by recording successful and non-successful catches. Assessment of visual-motor integration abilities using more objective, quantifiable, and technologically advanced techniques have been shown to improve discriminative ability of motor imitation assessments (Tunçgenç et al., 2020). Future studies should aim to examine brain-connectivity associations with visual-motor skills, such as motor imitation, in children with ASD using assessments that are more objective, quantifiable, and demonstrate improved discriminative ability over traditional standardized motor assessments.

Evidence from task-based fMRI studies using visuomotor imitation and tracking paradigms suggest that the right posterior cerebellum contributes to visual-motor processing in concert with M1, premotor, basal ganglia, and IPL regions (Mühlau et al., 2005; Neely et al., 2013; Okamoto et al., 2018; Unruh et al., 2019; Vaillancourt et al., 2006, 2003; Z. Wang et al., 2019). Further, direct cerebellum Crus I recording in animals show that Crus I stores internal representations that allow prediction of the sensory consequences of internally generated motor commands and prediction of trajectories of dynamic visual stimuli within the environment (Cerminara et al., 2009; Liu et al., 2003; Proville et al., 2014). The contribution of the Crus I region of the cerebellum in guiding sensorimotor behavior (Proville et al., 2014; Spraker et al., 2012; Vaillancourt et al., 2006) appears to be linked to formation of internal representations of predictable dynamic visual stimuli (Cerminara et al., 2009) and limb movement (Liu et al., 2003). Cerminara et al. (2009), showed that Purkinje cells in Crus I respond to predictable dynamic visual stimuli and continue to respond even when vision of the target is occluded, providing evidence of an internal representation of external target motion (Cerminara et al., 2009). This role of Crus I is likely mediated through known connections with the IPL (Buckner et al., 2011; Clower et al., 2001; Glickstein, 2000; Moulton et al., 2017; Stoodley et al., 2018) and autism-associated abnormalities in right Crus I connectivity with left IPL connectivity may thereby explain ASD-specific findings of impaired motor-tracking of dynamic visual stimuli (Lidstone et al., 2020), impaired ball-catching (Ament et al., 2015; Landa et al., 2016; Whyatt & Craig, 2013; Whyatt & Craig, 2012) and reduced sensitivity to visual feedback during motor learning (Haswell et al., 2009; Izawa et al., 2012; Marko et al., 2015). Visual internal representations of arm movement are also stored in Crus I (Liu et al., 2003), the absence or inaccuracy of which may limit the ability of children with ASD to predict the sensory outcomes of limb movement from visual feedback (Glazebrook et al., 2009; Schmitz et al., 2003; Zheng et al., 2019) and to form internal action models using visual-error feedback (Marko et al., 2015). In children with ASD, right Crus I dysfunction is also implicated in motor behaviors crucial for social-communication such as motor imitation (Jack, Englander, & Morris, 2011; Jack, Keifer, & Pelphrey, 2017; Jack & Morris, 2014) and social eye contact (Laidi et al., 2017; Skefos et al., 2014). Collectively, these findings suggest that internal representations of external dynamic trajectories of objects in the environment and visually guided internally generated limb movement are either inaccurate or absent in children with ASD.

Right Posterior Cerebellar Connectivity and the ASD Behavioral Phenotype