Atypical brain asymmetry in autism – a candidate for clinically meaningful stratification

Background Autism Spectrum Disorder (henceforth ‘autism’) is a highly heterogeneous neurodevelopmental condition with few effective treatments for core and associated features. To make progress we need to both identify and validate neural markers that help to parse heterogeneity to tailor therapies to specific neurobiological profiles. Atypical hemispheric lateralization is a stable feature across studies in autism, however its potential of lateralization as a neural stratification marker has not been widely examined. Methods In order to dissect heterogeneity in lateralization in autism, we used the large EU-AIMS Longitudinal European Autism Project dataset comprising 352 individuals with autism and 233 neurotypical (NT) controls as well as a replication dataset from ABIDE (513 autism, 691 NT) using a promising approach that moves beyond mean-group comparisons. We derived grey matter voxelwise laterality values for each subject and modelled individual deviations from the normative pattern of brain laterality across age using normative modeling. Results Results showed that individuals with autism had highly individualized patterns of both extreme right- and leftward deviations, particularly in language-, motor- and visuospatial regions, associated with symptom severity. Language delay (LD) explained most variance in extreme rightward patterns, whereas core autism symptom severity explained most variance in extreme leftward patterns. Follow-up analyses showed that a stepwise pattern emerged with individuals with autism with LD showing more pronounced rightward deviations than autism individuals without LD. Conclusion Our analyses corroborate the need for novel (dimensional) approaches to delineate the heterogeneous neuroanatomy in autism, and indicate atypical lateralization may constitute a neurophenotype for clinically meaningful stratification in autism.


Introduction
Autism spectrum disorder (autism) is a neurodevelopmental condition characterized by social-communicative deficits, restricted, repetitive behaviors and sensory abnormalities (1). One of the key characteristics of autism is its great phenotypic and biological heterogeneity (2,3). Particularly in the neuroimaging literature, results are mixed and inconsistent and have been attributed to this heterogeneity in the autism population. To provide a more coherent picture of the complex neuropathology of autism, it is critical to tackle its heterogeneity and identify the neural markers that are consistent across studies. Such markers can provide clinically relevant stratification of individuals with autism.
When it comes to identifying a consistently implicated neural feature in autism, a large body of literature converges towards a disruption of hemispheric specialization -one of the most fundamental biological properties of our brains (4,5). This basic organizational principle describes that the two hemispheres differ in their functional specialization and exhibit pronounced structural asymmetries (6,7). Functional specialization involves leftward lateralization of language and motor skills and rightward lateralization of spatial perceptual abilities (8). It is also evident in grey matter (GM) asymmetries with frontal opercular and temporal perisylvian regions, and hippocampus exhibiting 5 leftward asymmetries, whereas thalamus and posterior parietal cortex showing rightward asymmetries (9,10).
Individuals with autism exhibit impairments in left hemisphere skills such as social-communication, language and motor-related symptoms, whilst appearing relatively intact in right hemisphere functions such as visuospatial skills (11). This lateralized pattern of deficits and strengths in autism has given rise to theories trying to reconcile its complex clinical profile with atypical structural hemispheric specialization (11). In the largest autism cohort to date comprising over 3000 subjects individuals with autism presented with widespread leftward cortical reductions (12). Smaller studies are in line with this showing either a reduction or even reversal of typical leftward asymmetries in language-and motor-related regions (13)(14)(15)(16)(17)(18)(19)(20). Thus, atypical lateralization in autism is among the most replicated findings with moderate to high effect sizes (21) in the otherwise heterogeneous neuroimaging literature.
However, inconsistencies in results remain regarding regional specificity and direction of observed patterns. Such inconsistencies are usually attributed to sample heterogeneity and delineating heterogeneity still remains one of the central tasks in autism research.
The relationship between heterogeneity and brain asymmetry in autism may be further co-dependent on age, sex, handedness, different symptom profiles and comorbidities such as attention-deficit hyper-activity disorder (ADHD).
Age: Lateralization in neurotypicals becomes more pronounced through agerelated maturational processes (22). This typical trajectory is disrupted in 6 individuals with autism showing increasing reversed rightward lateralization (23). Sex: NT males are usually more strongly lateralized, while NT females have a more symmetric distribution pattern (24). How sex affects atypical asymmetry in autism is unknown. Handedness: Left-handed individuals have a higher chance of a different organization in the brain than right-handed individuals. Individuals with autism exhibit elevated rates of non-righthandedness, which has been attributed to atypical specialization in the brain (25). Language delay (LD): Individuals with autism and early LD show more pronounced deviations from typical asymmetry than those without LD (15,26).
ADHD: Some conditions that are also a common comorbidity in autism, such as ADHD, are also associated with atypical lateralization (27).
Despite increased recognition of heterogeneity in autism (28), there is little effort to address such challenges methodologically. One first example towards quantifying biological variation at the individual level in autism was recently demonstrated (29) using a novel normative modeling method (30,31). Similar to the use of growth charts in paediatric medicine, normative modelling aims to place each individual with respect to centiles of variation in the population and thereby facilitates a move away from classical case-control analyses that ignore individual differences. Applying normative modelling to cortical thickness estimates, Zabihi et al., 2018 (29) showed that individuals with autism exhibit highly individualized atypicalities of cortical development.
This study was designed to address heterogeneity in autism with regard to age, sex and core and co-occurring symptoms in the context of brain 7 lateralization using novel individualized analyses: 1) we transcend classical case-control analysis and address inter-individual variation by applying normative modeling (30,32). 2) We aim to identify laterality-related subtypes by considering co-occurring clinical symptoms in autism such as language development. Through capturing variation at the individual level in combination with addressing different sources of heterogeneity and using a consistent imaging feature in autism, our work provides a step towards precision neuroscience in autism.

Methods and Materials
Participants Participants were part of the EU-AIMS and AIMS-2-TRIALS Longitudinal European Autism Project (LEAP) (33,34) cohort -the largest European multicenter initiative aimed at identifying biomarkers in autism. Participants underwent comprehensive clinical, cognitive and MRI assessment at one of six collaborating sites ( Figure S1). All participants with ASD had an existing clinical diagnosis of autism. For details on participants, study design and exclusion criteria, see Supplemental Information (SI) and (34). The final sample comprised 352 individuals with autism (259 males and 93 females), and 233 NT controls (154 males and 79 females) between 6 and 30 years.
For details on demographic information, see Table 1. 8 IQ was assessed using the Wechsler Abbreviated Scales of Intelligence.

Clinical and cognitive measures
The Autism Diagnostic Observation Schedule (ADOS-G (35)) measured clinical core symptoms of autism. The Autism Diagnostic Interview-Revised (ADI-R) (36) was used to measure parent-rated autistic symptoms and LD, which was defined as having onset of first words later than 24 months and/or first phrases later than 33 months. ADHD symptoms were assessed with the DSM-5 ADHD rating scale. Handedness was assessed with the short version of the Edinburgh Handedness Inventory (37). For further details on all cognitive measures, see SI and (33).

Image Preprocessing
For MRI data acquisition parameters, see SI and Table S1. Structural T1weighted images were preprocessed according to a validated laterality pipeline (15,38,39) using the CAT12 toolbox (http://www.neuro.unijena.de/cat) (see Figure S2). All original images were segmented and affine registered to a symmetric tissue probability map before being reflected across the cerebral midline (x = 0). All segmented reflected and original (non-reflected) grey matter (GM) maps were then used to generate a symmetrical study-specific template via a flexible, high-dimensional nonlinear diffeomorphic registration algorithm (DARTEL) (40). They were next registered to the symmetrical study-specific template as per standard DARTEL procedures. An intensity modulation step was included to retain voxel-wise information on local volume

Normative Modelling
The normative modeling method has been described in detail previously (29,30,(41)(42)(43). In summary, we estimated a normative brain aging model at each GM laterality voxel by using Gaussian process regression (GPR) (44), a Bayesian non-parametric interpolation method that yields coherent measures of predictive confidence in addition to point estimates (for details see SI). With this method, we could predict both the expected regional GM asymmetry changes and the associated predictive uncertainty for each individual allowing us to quantify the voxelwise deviation of GM asymmetry from the NT range across the entire brain.
First, we trained a GPR model at each voxel on the NT cohort using age, sex and site as covariates to predict GM asymmetry resulting in a developmental model of GM asymmetry in NT individuals. To avoid over-fitting, assess generalizability and determine whether NT individuals fall within the normative range, we used 10-fold cross-validation in NT individuals before retraining the model in the entire sample to make predictions in individuals with autism (see SI).

0
We generated normative probability maps (NPM), which quantify the deviation of each participant from the normative model for GM asymmetry at each voxel. These subject-specific Z-score images provide a statistical estimate of how much each individual's true laterality value differs from the predicted laterality value with reference to the NT pattern at each voxel given the participants age, sex and site. NPMs were thresholded at an absolute value of Z>|2.6| (30,42). Based on this fixed threshold, we defined extreme rightwardlateralized (positive) and extreme leftward-lateralized (negative) deviations for each participant.
All extreme deviations per subject were summarized into (log-transformed) scores representing the percentage of extreme rightward and extreme leftward deviations per subject in relation to the total number of intracerebral voxels. These percentage scores were compared calculating a GLM including diagnosis and sex as the regressors of interest, and age as covariate. To compare our normative modeling against a conventional group-mean difference analysis we ran Permutation Analysis of Linear Models (PALM) on the LI images examining the sex-by-diagnosis(-by-age) interaction with site as covariate.

Spatial characterization of deviations
We applied two strategies to spatially characterize the extreme rightward and

Relative contributions of different cognitive and behavioral measures
We ran relative importance analyses within individuals with autism including variables related to lateralization such as LD, ADHD, handedness, sex and symptom severity (CSS). This was done in a sample with reduced sample size (N=305) without missing values on these variables. We used averaging over orderings (47) to rank the relative contribution of highly correlated regressors to the linear regression model. The variable showing strongest contribution to explained variance was followed-up with additional analyses.

Symptom Associations
An individual-level atypicality score was estimated for each individual through extreme value statistics by computing the 10% trimmed mean of the 1% top deviations (29). We then computed one-tailed, FDR-corrected Pearson's correlations between these individual atypicality scores and the ADI-R and ADOS-2 symptom severity scores. 1 2

Robustness and Replicability
To assess robustness, we estimated a separate normative model without including site as a covariate, but removing site effects using ComBat (48). For the sake of comparability, we also applied FDR-correction (49) to the NPMs.
We also included full-scale intelligence (FIQ) and handedness as nuisance covariates. To test whether the effects were robust against the influence of intellectual disability (ID), we reran analyses excluding individuals with FIQ <70. Finally, as individuals with autism with and without LD were not matched on certain demographic variables, we additionally created a sub-sample matched for age and symptom severity and re-ran second-level statistical analyses to assess robustness of results. To assess replicability, we selected a sample from the publically available Autism Brain Imaging Data Exchange (ABIDE) I and II (50,51) to examine extreme right-and leftward deviations in an independent datasets. For details, see the SI and Table S2.

Results
The classic PALM analysis to assess mean group differences did not yield any significant results. Figure

Symptom Associations
Across the whole brain, there were significant associations between extreme rightward deviations and ADI-R communication scores (r=0.14, p=0.004). No correlations survived FDR-correction in individuals with autism with and without LD. In males with autism, there were significant positive correlations between rightward deviations and core autism symptoms (ADI-R social: 1 6 r=0.18, p=0.002; ADI-R communication: r=0. 19, p=0.001). In females with autism, as well as in males and females with autism with and without LD, no results survived FDR-correction.

Replicability
In line with the results of the EU-AIMS LEAP sample, males and females with autism of the combined ABIDE dataset showed both significantly more extreme rightward (F (1) =7.7, p=0.006, d=0.14) and leftward deviations (F (1) =17.1, p<0.001, d=0.24) compared to NT individuals (see Figures S10a-b).

Discussion
In this study, we mapped extreme deviations in structural asymmetry at the individual level in comparison to a normative model of laterality development.
A further aim was to explore laterality as a stratification marker in autism in a large and deeply phenotyped cohort and an equally large replication cohort.
We found highly individualized patterns of both right-and leftward asymmetry 1 7 deviations in males and females with autism. In contrast, when using classical case-control analyses, we did not detect any significant group differences, emphasizing the need to move beyond group averages to capture an accurate representation of the phenotype at the individual level. Similarly, a recent study addressing heterogeneity dimensionally points out that traditional case-control analyses yield smaller effects and miss atypicalities detected hemisphere and therefore are more susceptible to environmental influences during neurodevelopment. Whether prenatal androgens and related early immune activation (55,57) contribute to atypical hemispheric development in autism, remains to be established.
Extreme rightward deviations were most pronounced in the motor network and frontal operculum and extreme leftward deviations in the visuospatial network. Accordingly, rightward shifts in asymmetry in language related regions, particularly in Broca's area, are frequently reported in autism (16,17,58). However, atypical lateralization of motor and visuospatial performance is underexplored in autism. Accumulating evidence suggests the important role of motor-related asymmetries in the neurobiology of autism (5,13,59,60). Despite mostly intact visuospatial performance in autism, atypical activation patterns have also been reported in individuals with autism while performing visuospatial tasks (61). More specific cognitive measures are needed to establish the functional relevance of these alterations.
Males and females with autism showed a similar degree of extreme deviations ( Figure 2). Specifically, in the TOFC, females with autism showed stronger rightward deviations, while males with autism and NT females both show fewer rightward deviations than NT males. This 'occipital face area' is strongly right-lateralized in NT individuals (62). Individuals with autism exhibit atypical face processing strategies and these are more pronounced in females with autism (63). Being one of the most reported impairments in social cognition in 1 9 autism, face processing and related atypical, sex-differential lateralization awaits further exploration.
When considering the spatial extent of deviations, females with autism show on average greater overlap across differently implicated regions than males with autism ( Figure 2 and Figure S10). This suggests that females with autism constitute a less heterogeneous group than males with autism who show less pronounced overlap of focal atypicalities in laterality. Also, when considering males and females separately, atypical asymmetry was only associated with more social-communicative symptoms in males with autism, but not females with autism, suggesting that phenotypically similar manifestations of atypical lateralization appear to have differential cognitive implications for males and females with autism.
We found that both extreme left-and rightward deviations were associated with LD, but only rightward deviations differentiated individuals with autism with and without LD from each other. This is in line with prior reports (15,26).
Two structural studies further showed differential morphological alterations in individuals with autism with different developmental language profiles (64)

Strengths and Limitations
We present analyses in a large-scale, deeply phenotyped and prospectively normative pattern. However, an alternative hypothesis that individuals with autism might lack specialization of either hemisphere (59) which would be expressed in reduced laterality might be less detected with the current approach.

Conclusion
We estimated a normative model of GM voxelwise asymmetry based on a large NT cohort and applied this to a large, deeply phenotyped and heterogeneous autism sample. Our results confirm that atypical asymmetry is a core feature of the autism neurophenotype, shows highly individualized patterns across individuals and is differentially related to different symptom profiles such as language delay. Further exploration of such associations has the potential to yield clinically relevant stratification markers needed for precision medicine.        Note that the standard VBM pipeline is adjusted to meet the needs for laterality analyses, i.e., using symmetric registration. Blue colors indicate a shift towards leftward asymmetry, whereas red colors indicate a shift towards rightward asymmetry. The regression line depicts the 3 6 predicted laterality values extracted from the language network based on neurosynth between 6 and 30 years of age along with centiles of confidence.
These are based on the normative model maps thresholded with positive Rho-maps. Blue dots are the true values for NT females in KCL.