We demonstrate that specific abnormalities in brain network structure are present in autism. Moreover, the topology of network-level abnormalities is consistent with prior morphological and functional work as well as clinical hallmarks of the disease. Specifically, the SN, anchored by the right FI, appears underdeveloped in autism. In addition, distinct nodal differences are apparent. Frontal regions are underrepresented in autism, whereas SMA covariance exceeds that of controls. In contrast, the DMN, anchored by PCC, may have both ‘overconnected’ and ‘underconnected’ components. Covariance beyond normal DMN topology was evidenced in autism, and this overrepresentation of the DMN was restricted to posterior brain regions. Many of the regions outside of DMN boundaries have been previously implicated in autism, including caudate, inferotemporal/fusiform cortex, auditory and temporal association cortex, and lateral postero-occipital regions. In contrast, covariance with frontal canonical DMN nodes was nearly absent in autism. These data suggest that distinct and network-specific alterations in structural architecture may underlie autism, providing a plausible neural substrate for clinical hallmarks of the disease. Furthermore, these data suggest an anatomic substrate for abnormal functional connectivity reported in autism, and demonstrate that this abnormal architecture is observable using standard anatomic MRI.
Our findings provide a resolution for previously reported inconsistencies in both structural and functional brain architecture in autism, as network-level effects could drive these apparent differences. Regional variability in GM and WM volume, WM integrity, cortical thickness, functional connectivity, and microscopic structure are predicted by network-level abnormalities. A framework of altered network-specific architecture unifies earlier divergent reports suggesting under- and overconnectivity within the same gross brain regions in autism and remains consistent with earlier hypotheses of abnormal connectivity 
. Similarly, previous reports of gross brain overgrowth can be interpreted in the context of selective network overdevelopment, as the gross anatomic overgrowth that has long characterized the disease may result from network-specific overgrowth or overconnectivity. The plausibility of coexisting over- and underconnectivity in autism is supported by recent fcMRI and structural studies. Noonan et al. 
reported overconnectivity in posterior regions including visual association and posterior inferotemporal cortex, but underconnectivity in prefrontal regions. Mengotti et al. 
described regional increased gray matter volumes in posterior regions of autistic subjects, as well as decreased gray matter volumes in frontal regions. Our findings are consistent with reports of both broad-based 
as well as selective 
connectivity abnormalities, as a network-level framework predicts widespread but specific alterations in connectivity.
Our data suggest restricted topology of the SN in autism. These results support the notion that autistic subjects lack the cognitive mechanism to disambiguate relevant social and emotional cues within complex environmental stimulus streams 
. Further, autistic individuals may not possess a robust neuroanatomic substrate for appropriate socio-emotional cognitive processing. Numerous reports suggest that DMN and ECN are functionally anti-correlated in healthy individuals, and their dynamic balance is modulated by the salience network 
. In this context, our data may reflect a structural basis for an as yet unreported functional imbalance between the DMN and ECN in autism.
Posterior overgrowth and ‘anterior-posterior disconnection’ of the autistic DMN may reflect a perturbed ‘division of labor’ within this network. Our data support an emerging theme in the literature of anterior-posterior underconnectivity 
, but also suggest coexisting specific regional overconnectivity within and outside of the DMN. However, as frontal structural covariance in canonical DMN nodes was modest in our control group, the finding of decreased frontal covariance in autism needs to be confirmed in older subjects, when a mature DMN is expected. However, medial prefrontal cortex has been shown previously to have decreased functional connectivity with other regions in the DMN 
, particularly with posterior nodes 
. This anterior-posterior disconnection is underscored by overconnectivity within posterior nodes themselves 
. Using a PCC seed, Monk et al. 
showed decreased anterior-posterior connectivity and
increased connectivity between PCC and other posterior regions, precisely consistent with our results using the same seed. Weaker connectivity between select DMN nodes such as prefrontal cortex and angular gyrus has also been shown 
. Moreover, posterior DMN connectivity was present in autistic subjects, but anterior nodes of the DMN were not apparent. Consistent with our results, they also detected connectivity with nodes outside of the canonical DMN. Kennedy et al. 
demonstrated that the DMN remains abnormally engaged during resting conditions in autism, and reported subnormal activity in medial orbitofrontal cortex with emotional stimuli. Weng et al. 
reported that poorer verbal and non-verbal communication scores in autistics correlated with stronger connectivity versus controls between PCC and bilateral temporal cortex including regions that overlap with our autistic DMN scMRI map. Social impairment in autism is associated with decreased connectivity between PCC and superior frontal as well as medial prefrontal regions 
, consistent with our results and plausibly relating to the dual function of FI in autism.
The neural systems that support social and emotional processing appear to be underdeveloped in autism, whereas those that mediate internally- versus externally-directed processing appear to be overrepresented in posterior nodes but isolated from anterior network nodes. The neuropathological processes underlying these selective abnormalities in large-scale brain networks in autism remain unknown. Plausible factors include early neuronal excess and later neuronal loss, abnormal microstructure, excess synapse formation, excessive dentritic outgrowth or hyperconnectivity, aberrant axonal pathfinding, overgrowth, or connections, and altered myelination 
. Each of these processes may be driven by aberrant gene expression, environmental factors, or both, and may progress by age-, network-, or domain-specific mechanisms.
How these processes relate to gray matter density is unclear. Our work suggests, however, that downstream effects of presumably disrupted molecular and cellular mechanics produce distinct and measurable alterations of normal brain network architecture within networks that underlie the core manifestations of the autistic disease state. Posterior DMN subnets may be overconnected, whereas rostro-caudal connectivity may be limited. In the SN, interconnections between critical nodes may be malformed, mature architecture not achieved, and the network left rudimentary and dysfunctional. Phenotypic features of autism could result as network ‘dysconnection’ leads to a deficit of ‘salience filtering’ and resultant inefficiencies in recruiting appropriate attentional, socio-emotional, behavioral, and higher-order cognitive resources 
. This could impact cognitive and behavioral functioning by at least three mechanisms. First, integration of external sensory information with visceral, autonomic, and hedonic status is aberrant, yielding breakdown of appropriate behavioral guidance from relevant internal and external stimuli. Second, social and emotional cues are not properly input to downstream processing pathways, resulting in misguided response-selection. Third, the modulator of externally- versus internally-directed stimulus processing pathways is ‘dysconnected’, resulting in inappropriate signal filtering and generating abnormal engagement of downstream cognitive processing streams and dysfunctional disengagement of the DMN.
We describe anatomic substrates consistent with altered functional connectivity reported in autism, and demonstrate that structural network-level abnormalities are quantifiable using standard anatomic MRI. It is plausible that multiple large-scale network architectures, including ICNs and SCNs, may be affected in autism. Our work predicts abnormal fcMRI covariance within large-scale networks commensurate with specific structural abnormalities that together disrupt emergence and maintenance of complex psychological and physiological functions in autism. However, although connectivity is often assumed by techniques measuring MRI signal covariance, neither fcMRI nor scMRI techniques directly measure anatomic connectedness. Moreover, direct associations between functional synchrony and underlying anatomic structure have yet to be established in autism.
Our whole-brain seed-based approach differs from most neuroimaging studies examining pairwise correlations between a priori
ROIs. Such studies limit resolution of network-level effects and allow limited conclusions regarding whole-brain or whole-network abnormalities. Whole-brain or whole-network fcMRI approaches indeed report both under- and overconnectivity, particularly within the DMN 
. In addition, our method is not biased by Euclidean distance between presumably connected neural substrates, as local and long-range structural covariance is captured equally by our analyses. Our results suggest decreased local and long-range connectivity in the salience network, as well as altered (both increased and decreased) local and long-range connectivity within the DMN. The fundamental causes of brain dysfunction and altered structure in autism may not reflect a global insult or gross morphological change, but may instead emerge via selective processes resulting in abnormal architecture of specific neural networks underlying the clinical manifestations of the disease.
Our data suggest that discrepancies in fcMRI over- and under-connectivity, WM and GM volume, cortical thickness, and WM integrity measures may be reconciled by a model of autism as a network-based disease. Recent fcMRI studies are consistent with the concept of autism as a disease with distinctive network distribution patterns 
. Abnormal network architecture may be evident from early stages in the disease, and may reflect genetic, molecular, or cellular neuropathology in specific networks or nodes 
. Whether these network alterations are permissive or reactive remains unknown. The age-dependent topology of large-scale structural networks in childhood has recently been identified 
, but the interrelationships of various structural and functional network-level measures have yet to be elucidated. Age-related patterns within these networks remain understudied in autism, but future work may reveal network- or node-specific abnormalities in genetic, molecular, or microstructural development in critical periods throughout development as well as later stages of the disease. More work in younger children is needed to clarify whether early disruptions may result in emergence of stable abnormal network architecture, further impacting network operation and cross-network interrelationships. Moreover, early overgrowth and later decline 
may not be homogenous throughout the brain; rather, network-specific growth trajectories may contribute to regional and whole-brain over- and underdevelopment across ages. Larger cross-sectional and longitudinal studies of network development in autism are necessary to further clarify distinct network trajectory patterns.
These results could guide future studies of network abnormalities in autism. For example, our findings predict decreased fcMRI and WM connectivity between SN nodes, as well as increased connectivity between specific posterior elements within as well as outside of the canonical DMN. Moreover, it remains unknown whether other large-scale networks show similar patterns of abnormal structural architecture. Distinctive frontal, temporal, and cingulate gray matter overgrowth in young autistics, with relative sparing of occipital cortices has been reported 
, suggesting involvement of other networks beyond SN and DMN. In addition, whether distinct network topologies could characterize autism subtypes remains to be studied. Future scMRI investigations of other large-scale function-critical brain networks, such as those involved in executive function, speech and language, semantics, and primary sensorimotor functions, may reveal abnormal structural architecture to be a fundamental characteristic of network neurobiology in autism.
Our study supports a model of autistic pathophysiology affecting domain-specific large-scale brain networks. Using standard anatomic MRI, we identified network-level structural abnormalities in the autistic brain, providing the first account of whole-brain, network-specific perturbations within autistic brain architecture. Our results suggest that structural brain abnormalities in autism may affect distinct large-scale networks. The SN appears underdeveloped in volume and extent, whereas the DMN demonstrates elements of both under- as well as over-development. These network-level perturbations are consistent with the clinical manifestations of the disease, and may provide targets for further study and intervention. Moreover, FI may represent an epicenter of perturbed structure and function in the autistic brain. Our work provides a unifying model of previously discordant findings based on structural and functional assessment, reconciles recent work with classic gross morphological findings in the disease, and reveals divergent network-dependent over- and underdevelopment in the same subjects. The diffuse specificity of our findings is consistent with emerging literature identifying regional abnormalities using varying techniques on microstructural as well as macrostructural levels.