Differential links in 16p11.2 deletion carriers reveal aberrant connections between large-scale networks

Abstract

Qualitatively different topographical patterns of connections are thought to underlie individual differences in thought and behavior, particularly at heteromodal association areas. As such, we hypothesized that connections unique to 16p11.2 deletion carriers compared to controls, rather than hyper- or hypo-connectivity, would serve as a better model to explain the cognitive and behavioral changes observed in individuals carrying this autism-risk copy number variation. Using a spatially-unbiased, data-driven approach we found that differential links clustered non-uniformly across the cortex—particularly at the superior temporal gyrus and sulcus, posterior insula, cingulate sulcus, and inferior parietal lobule bilaterally. At these hotspots, altered local connectivity that spanned across the borders of cortical large-scale networks coincided with aberrant distant interconnectivity between large-scale networks. This was most evident between the auditory and the dorsomedial default (DNb) networks—such that greater between-network interconnectivity was associated with greater communication and social impairment. Entangled connectivity between large-scale networks may preclude each network from having the necessary fidelity to operate properly, particularly when the 2 networks have opposing organization principles—namely, local specialization (segregation) versus global coherency (integration).

Introduction

Focus has shifted from explaining autism as a localized brain disorder of particular structures to one involving higher-order association areas—as a disorder of large-scale neural systems (Geschwind and Levitt 2007; Minshew and Williams 2007). This paradigm shift is concordant with neuropsychological testing. Basic information processing is intact (or even enhanced), whereas tasks that place high demand on information integration are impaired (Minshew and Williams 2007). Frith (1989) first synthesized the core problem under a unified cognitive model of “weak central coherence.” The model was reframed in early functional connectivity studies as “underconnectivity” by Just and colleagues, which asserts that reduced synchronization between brain areas is a reflection of underfunctioning information integration i.e., reduced correlation strength (Pearson’s r) between distant regions (Just et al. 2004). In contrast to weak central coherence which implicates a top-down problem (“a conductor problem at a symphony”), underconnectivity asserts an emergent property problem (“a jam session in which the musicians cannot hear each other”). However, in a meta-analysis by Müller et al., only 2 of 22 autism studies that showed support for global underconnectivity actually used a whole-brain field of view—the rest presumed that the effects were global by extrapolating from specific ROI analyzed (2011). In later functional connectivity studies, the focus shifted to a dysconnectivity model, in which a quantitative change in the number of local or distant links is the fundamental problem (Uddin et al. 2013). This posits that autism may spare brain regions that involve local specialization (segregation) while targeting highly distributed large-scale networks involved in global coherency (integration). Nonetheless, there have been conflicting results across studies, which have variously shown hyper- and hypo-connectivity (Vissers et al. 2012; Supekar et al. 2013).

Autism brain research has come under scrutiny for divergent results and lack of reproducibility across studies. This sentiment was captured by a prominent review article at the time, advocating to reconceptualize the disorder as “the autisms” in order to account for this plurality (Geschwind and Levitt 2007). Hahamy et al. (2015) reported that high inter-subject variability in connectivity is the idiosyncratic defining feature of autism spectrum disorder (ASD). In response, the autism brain imaging data exchange, an impressive grassroots consortium, aimed to address the variability that can arise from small sample sizes by pooling multiple international data sets (Di Martino et al. 2014). Methodological choices have also been identified as a major source of heterogeneous results (Müller et al. 2011). Others have reconciled conflicting data by accounting for changes that occur through development (Courchesne and Pierce 2005; Uddin et al. 2013). Another source of variance in ASD is the extensive genetic and etiologic heterogeneity. At least 71 ASD risk loci and 65 risk genes have been reported (Sanders et al. 2015), but none of the known causes of ASD account for >1 to 2% of cases (Abrahams and Geschwind 2008).

This study takes a genetics-first approach to ASD by focusing on individuals with microdeletion of one copy of the recurrent ~ 600 kb BP4-BP5 on 16p11.2. De novo copy number variations (CNVs) may account for up to 15% of sporadic cases of autism (The Simons VIP Consortium, 2012). 16p11.2 is among the most frequently occurring CNV associated with ASD, at 0.76% (95% CI, 0.51–1.12; Sebat et al. 2007). The locus contains MAPK3, a key regulator downstream from other ASD targets, but like other large de novo CNVs 16p11.2 likely contains multiple modest-effect ASD-risk genes (Sanders et al. 2015; Coe et al. 2019). Carrying a 16p11.2 deletion increases the odds of developing autism or developmental delay by 38.7-fold (95% CI, 13.4–111.8; McCarthy et al. 2009). High pleiotropy and variable penetrance are not rare in CNVs; in actuality they are characteristic of ASD-risk loci, including 16p11.2 (Buxbaum et al. 2012). The advantage of a genetics-first approach is that it is better able to isolate the contributions of a particular ASD-risk locus to a resulting phenotype—namely the connectome “endotype” and the resulting cognitive and behavioral changes. The present study investigates the resting-state functional connectivity of deletion carriers of this copy number variation.

Qualitatively different topographical patterns of connections are thought to underlie individual differences in thought and behavior, and are associated with variability in intrinsic functional connectivity (Seeley et al. 2007; van den Heuvel et al. 2009), particularly at heteromodal association areas (Mueller et al. 2013). As a cohort, autistic individuals report very different internal thoughts compared to control participants (Hurlburt et al. 1994; Frith and Happé 1999). As such, rather than a presumption of phenotype being a result of hyper- or hypo-connectivity, or quantitative differences in connectivity in general, we hypothesized that insight into cognitive and behavioral changes in 16p11.2 deletion carriers may be best found by searching for unique or differential connections between carriers and control participants—which was conducted using a spatially-unbiased, data-driven approach.

Methods

Participants

The Simons Foundation Autism Research Initiative developed the Simons Searchlight program to collect high-quality, standardized data on >175 rare genetic neurodevelopmental disorders, which includes 30 copy number variants (The Simons VIP Consortium 2012). Individuals were recruited in various ways including referral by clinical genetic centers or testing laboratories, web-based networks, or by self-referral of families who learned about Simons Searchlight (https://www.simonssearchlight.org/). Initial screening and review of medical records occurred at Geisinger and Emory Universities. Deletion carriers of the 16p11.2 CNV (DEL) were enrolled only of the recurrent ~ 600 kb BP4-BP5 16p11.2 CNV, and no other pathogenic CNVs or other known genetic diagnoses or syndromes were found. Individuals were excluded if they were not fluent in English, or had a history of environmental insults that could affect neurocognitive status, such as fetal alcohol syndrome, severe birth asphyxia, or severe prematurity. All participants were enrolled in an initial level of study that took place at 3 phenotyping core sites: University of Washington Medical Center, Baylor University Medical Center, and Boston Children’s Hospital. Assessments included the Autism Diagnostic Observation Schedule (ADOS), Autism Diagnostic Interview (ADI), broad screening measures of social impairment, such as the Social Responsiveness Scale (SRS), Vineland-II, and Social Communication Questionnaire (SCQ), as well as general measures of cognitive abilities with the Differential Ability Scale (DAS-II) and Wechsler Individual Achievement Test (WIAT-III) among others. A structural magnetic resonance imaging (MRI) scan was also obtained at the initial assessment.

A subset of these participants, 8 years and older, and able to tolerate MRI, were recruited to the second level of study focused on neuroimaging. Data acquisition occurred at 2 imaging core sites: University of California sites (UC) and Children Hospital of Philadelphia (CHOP). All aspects of the study at UC were performed at University of California—San Francisco, except for scanning, which was performed at University of California—Berkeley to maintain conformity with the CHOP scanner. Subjects had structural MRI, resting-state functional MRI, task-based functional MRI, diffusion-tensor imaging and magnetoencephalography collected on matched scanner platforms. The functional MRI scans obtained from the UC and CHOP neuroimaging sessions were the focus of this study.

CON participants (CON) were recruited at this point, directly by the 2 imaging core sites (CHOP and UC). Individuals with any major DSM-IV diagnosis were excluded (based on clinical psychologist review) or if an immediate family member was diagnosed with ASD, other developmental disorders, dysmorphic features, or genetic abnormalities. They had no familial relationships to the 16p11.2 deletion carriers. CON were also excluded for a Symptom Checklist 90 (SCL-90) score >62, an axis 1 psychiatric diagnosis on the Diagnostic Interview for Children (DISC), and on the Broader Autism Phenotype Questionnaire (BAPQ) an aloof subscale score > 4/3.5 (male/female), a pragmatics subscale score > 3.25/3.5, or a rigidity subscale score > 4/3.7. CON subjects also had a chromosome microarray, neurologic exam, a photograph evaluation for dysmorphology, and a clinical psychologist review. The study aimed to recruit CON who matched CNV carriers for age, sex, handedness, and nonverbal IQ (NVIQ). Data can be acquired at https://base.sfari.org/.

Ninety-seven individuals were initially screened, which consisted of 56 CON and 41 DEL. The MRI images were screened for quality control manually, as well as with automated measures (detailed below). As a result, 14 CON were excluded for poor quality of the functional image, and 1 additional subject was removed for poor quality of the T1 structural image. As for DEL, 14 subjects were also excluded for poor quality on the functional image, and 1 additional subject was removed for the quality of the structural image. Thus, the final groups consisted of 41 CON and 26 DEL for analysis.

MRI data acquisition

Functional MRI scans were acquired at UC and CHOP with matched 3T TIM Trio MRI scanners (Siemens) and 32-channel phased-array radio-frequency head coils. The functional images were acquired during a 6-min eyes-open resting state session with a gradient-echo echo planar imaging sequence sensitive to blood oxygenation level-dependent (BOLD) contrast (parameters: TR = 3,000 ms, TE = 30 ms, 3.0 × 3.0 × 3.0-mm voxels, FA = 85°, 47 axial slices with interleaved acquisition). A structural scan was also obtained with a T1-weighted multi-echo magnetization-prepared gradient echo image using the following parameters (TR = 2530 ms, TE = 1.64 ms, FA = 7°, 1 × 1 × 1 mm, and FOV = 256 mm). More scans were obtained at UC compared to CHOP, but no significant imbalance between groups was found (CON 28:13 or DEL 16:10, t = −0.56, P = 0.58).

Image processing

The functional images underwent the following preprocessing steps common to functional connectivity analyses: (i) first 4 volumes discarded to allow for signal stabilization; (ii) AFNI’s 3dDespike which dampens large fluctuations in the resting-state BOLD time series; (iii) slice-timing correction (SPM); (iv) rigid body correction for head motion with MCFLIRT in the FSL suite; (v) registration to MNI space by transforming the functional images to an EPI template (SPM); (vi) smoothed with 6-mm full-width half-maximum Gaussian kernel; (vii) band-pass temporal filtering (0.01 to 0.08 Hz) and regression of 6 motion parameters and signal from the white matter, ventricles, and whole brain.

MRI quality control and motion

All anatomical and functional images were manually inspected for cerebral cortex clipping, signal inhomogeneity, ghosting, and radio-frequency noise or spiking using a standard methodology (Harvard Center for Brain Science, http://cbs.fas.harvard.edu/science/core-facilities/neuroimaging/information-investigators/qc).

Special attention was placed on reducing bias from motion. Head motion is associated with decreased functional coupling in the default and frontoparietal CON networks, and increased local functional coupling—including left and right motor regions (Van Dijk et al. 2012; Power et al. 2014). In conjunction with manual inspection for macro-movements in the form of MRI artifacts, automated measures were calculated to assess for micro-movements. Slice-based temporal signal-to-noise ratio (tSNR), and micro-movements were calculated by determining the number of movements from 1 volume to the next that exceeded 0.5 mm of root mean squared displacement in 3D space (micro-movements > 0.5 mm). Participants were excluded based on automated quality control measures of tSNR < 100 and/or >15 micro-movements > 0.5 mm. As discussed above, a total of 15 CON and 15 DEL were excluded for poor quality. This resulted in the final groups of 41 CON and 26 DEL for analysis.

In the data analyzed no significant difference was found in tSNR between CON and 16p11.2 DEL (226.7 ± 73 vs. 201.0 ± 49 respectively; t = 1.72, P = 0.09), or voxel-based signal-to-noise (69.8 ± 11 vs. 65.6 ± 9; t = 1.62, P = 0.11). Additionally, no significant difference in micro-movements > 0.5 mm were seen (1.2 ± 2 vs. 1.0 ± 2, t = 0.26, P = 0.79). Furthermore, an additional preprocessing measure was undertaken, AFNI’s 3dDespike, which dampens large fluctuations in the resting-state BOLD time series that are likely due to head motion (Allen et al. 2014).

On a more fine-scale measure, the 2 groups did show a significant difference in the number of micro-movements > 0.1 mm with less motion in CON (22.0 ± 16) than DEL (38.3 ± 27; t = −2.78, P < 0.01, Cohen’s d = 0.73). As such, in the analysis motion was controlled for as a nuisance variable using linear regression.

Differential link analysis

Individual whole brain correlation matrices can discern group differences when analyzed link-by-link (or connection-by-connection). As previously described (Sepulcre et al. 2012), an MNI atlas mask of 4,968 voxels covering the whole brain was used to extract the time course of low-frequency BOLD fluctuations. This resulted in a 4968 x 4968 matrix of Pearson’s correlation coefficients for each individual. The connectivity maps were then standardized by converting to Z scores, z(r), to compare values cross individuals. Thus, each node (i) has a z-transformed-r value associated with every other node across the cortex (i-j, i-k, i-…n). An Analysis of covariance is run on each link, comparing i-j in all participants of the first group to the second group to assess whether the null hypothesis for that the given link can be rejected. For instance, if only links i-a, i-f, i-p are significantly different among the possible 4967 possible connections to node i, then a value of 3 is displayed on the cortical map at node i, which represents 3 differential links (DLs) at that topographical coordinate. Nodes within a 16-mm radius sphere (Sepulcre et al. 2010), were counted towards local connectivity (~32 possible links), and outside the radius sphere were counted towards distant connectivity (~2,454 links for each hemisphere). Of note, no values were removed from any participant on statistical testing; specifically, z(r) < 0.25 were included as to avoid applying nonparametric constraints to a parametric dataset. Inclusion of this these correlations is supported by evolving conceptions of functional connectivity. Dynamic functional connectivity, from seconds to minutes, has shown time-varying profiles for canonical resting-state network; a time-compressed average correlation value (used in our analysis), may be negative or positive depending on the time point sampled (Chang and Glover 2010; Hutchison et al. 2013b, 2013a). A false discovery rate (FDR) correction at 0.05 was performed to correct for multiple comparisons.

The advantage of this approach is that it is a hypothesis-free, data-driven method for isolating specific brain regions with distinct topographic patterns of connectivity between groups—simultaneously across the cortex. To visually demonstrate these distinct patterns of connectivity between groups, a seed was placed at the epicenter of brain regions with the greatest number of DLs (specifically locations with > 50 links at P < 0.001 and FDR correction). Mean seed-based correlation maps were displayed for both groups. Statistical testing was intentionally not performed here to avoid circularity.

As a final step all resulting maps were projected onto cerebral hemisphere of PALS surface (PALS-B12) provided by Caret Software (Van Essen 2005) using the interpolated voxel algorithm and average fiducial mapping. The resolution or granularity of voxel/parcel size sufficient for DL analysis was also explored (Supplemental Fig. 14).

Graph analysis of cortical parcels and spring-loaded graph

Changes in between-network architecture were investigated further using a cortical parcellation scheme independent of the nodes used in DL analysis. Schaefer et al. developed a cortical parcellation using a gradient-weighted Markov Random Field model with superior functional and connectional homogeneity (Schaefer et al. 2018), and these 400 cortical parcels were found to respect the boundaries of 17 large-scale networks (Yeo et al. 2011). For each subject, we resampled the mean time course of low-frequency BOLD fluctuations according to the predefined boundaries of each parcel (from the publicly available template: https://github.com/ThomasYeoLab/CBIG/tree/master/stable_projects/brain_parcellation/Schaefer2018_LocalGlobal). This resulted in a 200 × 200 matrix of z-transformed Pearson’s correlation coefficients, z(r), for each hemisphere. Only parcels from relevant large-scale networks were kept—as identified from the DL analysis (such as all 46 parcels of the left hemisphere default network [DN]). This resulted in a final 74 × 74 matrix from the left hemisphere, and 64 × 64 matrix from the right hemisphere. The mean group connectivity for each group average were displayed in a spring-loaded graph using Pajek software (Nooy et al. 2011). Only positive connections (>0.2) were displayed for clarity.

The composite network strength for a particular network was computed by averaging the connections, z(r), for that submatrix for each participant. Linear regression on nuisance variables was performed using SPSS software. The strength of connectivity, z(r), was covariance-corrected for age, sex, handedness, site, NVIQ and micro-movements 0.1 mm. The beta values for each variable were used to compute adjusted measures in the original units using the formula, as previously established (Buckner et al. 2004):

Z(r)_adj = Z(r)_nat—[β₁(Age_nat—mean Age_nat) + β₂(Sex_nat—mean Sex_nat) + β₂(Handedness_nat—mean Handedness_nat)+ β₄(Site_nat—mean Site_nat) + β₅(NVIQ_nat—mean NVIQ_nat)+ β₆(Motion_nat—mean Motion_nat)]

where z(r)_nat is the uncorrected, raw connectivity strength, Age_nat is the uncorrected age, mean Age_nat is the group mean, and so on. Final covariance corrected values are reported as adjusted z(r) for any given connectivity strength.

Results

Cognitive and behavioral characterization

Twenty-six child deletion carriers of the 16p11.2 CNV (DEL) and 41 age-matched control participants (CON) served as the comparison group (Table 1). Age, sex, and handedness did not significantly differ between groups (t = 1.32, P = 0.19; t = −0.98, P = 0.33; t = −0.60, P = 0.55, respectively). The study recruited control participants with an aim of matching NVIQ to that of 16p11.2 carriers. Despite best efforts, a 16-point difference was observed between groups (t = 5.21, P < 0.001; Cohen’s d = 1.30). As such, full scale and verbal IQ were also significantly different.

Table 1.

Demographics and behavioral characterization.

	CON	DEL
n	41	26
Mean age ± SD	12.4 ± 3	11.6 ± 2
Age range	8–16	8–16
Male: female	27:14	14:12
Left: right	7:34	6:20
FSIQ	109.8 ± 12	88.1 ± 15*
VIQ	111.3 ± 14	84.3 ± 18*
NVIQ	108.0 ± 11	92.0 ± 14*
D-KEFS	9.6 ± 2	5.8 ± 3*
CELF, ≤ 85	108.8 ± 12, 1/35	76 ± 20*, 16/23
SRS, ≥60	15.1 ± 9, 0/38	78.8 ± 40*, 19/26
Vineland-II, ≤ 85	107.1 ± 10, 0/38	82.2 ± 13*, 18/26
ADOS	-	8/26
ADI	-	11/19

Abbreviations: full scale intelligent quotient (FSIQ), verbal IQ (VIQ), Delis–Kaplan executive function system (D-KEFS), Clinical Evaluation of Language Fundamentals (CELF), Social Responsiveness Scale (SRS), Autism Diagnostic Observation Schedule (ADOS), Autism Diagnostic Interview (ADI).

^*Represent significant differences.

On phenotypic assessment (detailed previously in Hanson et al. 2010), 16p11.2 deletion carriers showed significant impairment on social and communication scales in comparison to CON: SRS (t = −7.97, P < 0.0001, Cohen’s d = 2.19), and the more granular Vineland-II (t = 8.49, P < 0.0001; Cohen’s d = 2.10). Additionally, 14/26 (54%) DEL met criteria for ASD on either the ADOS or the ADI.

All participants were also evaluated for mental disorders using the Diagnostic and Statistical Manual of Mental Disorders, fourth edition (DSM-IV; Table 2). A clinical diagnosis of ASD was given to 6 (23%) DEL. Neuropsychological testing was more sensitive in detecting cognitive deficits than the clinical evaluation across multiple domains. For instance, 16 DEL met the threshold for a language disorder on the Clinical Evaluation of Language Fundamentals, fourth edition (CELF) of < 85, whereas only 10 met a clinical diagnosis of language disorder with DSM-IV criteria. There was no overlap in individuals given a diagnosis of intellectual disability with those diagnosed with ASD.

Table 2.

Clinical diagnoses met on the diagnostic and statistical manual of mental disorders, fourth edition (DSM-IV).

	DEL n = 26 (%)
Phonological processing disorder (315.39)	15 (58)
Language disorder (315.31, 315.32, 307.9)	10 (38)
Coordination disorder, developmental (315.4)	9 (35)
ASD (299.00, 188,070)	6 (23)
ADHD diagnosis (315.00, 314.01, 314.9)	6 (23)
Intellectual disability (317, 318, 319)	5 (19)
Behavior disorder (312.9, 313.82)	3 (12)
Anxiety disorder (300.0, 300.02)	3 (12)
Learning disorders (315.0, 315.1, 315.2, 315.9)	2 (8)
Tic disorder (307.2, 307.22, 307.3)	2 (8)
Mood disorder (300.4, 300.9)	0 (0)

Results in this study were covariance-corrected for nuisance factors of age, sex, handedness, NVIQ, scanner site, and motion.

Hotspots of distinct functional connectivity

The differential link (DL) approach provides a hypothesis-free, data-driven method for identifying locations with distinct patterns of functional connectivity between groups. Only DLs reaching significance are displayed (P < 0.001), while controlling for the aforementioned nuisance variables. We found DLs to be non-uniformly distributed across the cortex (Fig. 1). Particular brain regions harbored the preponderance of DLs between groups—akin to “hotspots” of distinct functional connectivity. Brain regions circumscribed in black denote DLs that survived multiple comparisons from FDR correction (q = 0.05).

Fig. 1.

Differential links (DL) between groups distribute non-uniformly across the cortex, and reveal “hotspots” of distinct connectivity in 16p11.2 deletion carriers. Only links significantly different at P < 0.001 while controlling for nuisance variables are counted. The resulting sum of DL at each node is projected on an inflated surface brain map for each hemisphere. A) The local map displays nodes with 2 DL (yellow), 3 DL (orange), 4 DL (red), and ≥ 5 DL (dark red). B) The distant map displays nodes with ≥20 DL (yellow), ≥30 DL (orange), ≥40 DL (red), and ≥ 50 DL (dark red). Areas circumscribed in black denote DLs that survived FDR correction. Local and distant DLs overlap primarily at the superior temporal gyrus (STG) and the posterior insula, but can also be seen to some extent at the inferior precentral gyrus, inferior parietal lobule, the caudal middle frontal gyrus, and the medial cingulate sulcus.

Local connectivity (within a 16-mm radius sphere) demonstrated the greatest number of DLs at the superior temporal gyrus (STG), insula, and middle cingulate gyrus/sulcus bilaterally (Fig. 1A). For distant connectivity (beyond its 16-mm radius sphere), the STG and insula were again present. In addition, the left lower precentral gyrus, left middle frontal gyrus (MFG), left dorsal medial prefrontal cortex (dmPFC), right middle temporal cortex, bilateral inferior parietal lobule (IPL), posterior middle cingulate cortex, and the posterior cingulate sulcus were regions a high degree of DLs (Fig. 1B).

Functional connections across large-scale networks

The DL analysis only identifies where connectivity is significantly different between groups, not what is different at that location. To visualize these distinct patterns of connectivity a seed-based correlation was performed. Hotspots were visualized if that cluster exceeded > 50 DL. A seed was placed at the epicenter of each cluster—at the topographic coordinate with the greatest number of DLs. The left hemisphere cortical map of DLs was overlaid with boundaries of Yeo’and Buckner's cortical parcellation (Yeo et al. 2011) to appreciate connectivity patterns in relation to large-scale networks (Fig. 2, upper box). Brain regions functionally connected to that topographic coordinate are displayed for each group. Statistical testing was intentionally not performed here to avoid circularity.

Fig. 2.

Seed-based correlation was used to visualize distinct patterns of connectivity at DL hotspots. (upper box) The left hemisphere DL map has been overlaid with the boundaries of large-scale networks (left). Yeo’s 7 networks are displayed with seed locations (right). Relevant networks include the somatomotor network (blue), default network (DN) (salmon), cingulo-opercular network (CingOp) (magenta), and dorsal attention network (dATN) (green). The mean functional connectivity from that topographical coordinate are displayed for each group as z-transformed-r, or z(r), with positive correlations displayed from yellow to dark red (≥0.3 to ≥0.6) and negative correlation from teal to dark blue (≤ −0.2 to ≤ −0.5). A) In CON the auditory cortex (−62 16 0) has connectivity restricted locally to adjacent superior temporal gyrus (STG) and superior temporal sulcus (STS), and strong connectivity (≥0.5) respects the boundary of somatomotor network. In contrast, DEL have distant connections to the inferior parietal lobule (IPL), dorsomedial prefrontal cortex (dmPFC), precuneus and posterior cingulate cortex (PCC) of the DN. Strong connectivity also extends to the lateral surface of the temporal cortex, and negative correlations are observed at brain regions associated with the dorsal attention network and the parietal operculum (OP) of the CingOp. B) The dmPFC (−2 54 24) in CON shows connectivity to the canonical DN. In DEL the dmPFC also includes the same region of auditory cortex. Additionally, functional connectivity in DEL is less robust at the posterior IPL and the caudal middle frontal gyrus (MFG). C) The IPL (−52 56 40) in CON has strong connectivity to the caudal MFG, middle temporal cortex, precuneus/PCC, and dmPFC. In DEL the IPL again shows connectivity to the STG and STS in addition to regions associated with the DN. (Lower box) The DN fractionates into a lateralized dorsomedial subsystem (salmon), a medial temporal subsystem (navy blue), and a core DN (yellow). This is distinct from the language network (or temporoparietal network, blue). More recent densely-sampled, intra-individual data describes no core DN subsystem at all, but rather 2 parallel, interdigitated systems DNa (overlapping with the medial temporal subsystem), and DNb (overlapping with dorsomedial subsystem)—that only give the appearance of a core DN when averaged over a group. To be precise, the auditory node is functionally connected to cortical regions of the dorsomedial (DNb) network. D) The posterior insula (−34 20 8) in CON has connectivity restricted locally with some extension anteriorly to the middle insula, and strong connectivity (≥0.5) respects the boundary of somatomotor network. In DEL the posterior insula has strong connectivity with the OP, middle insula, and pre-supplementary motor area (pre-SMA) of the CingOp. DEL also show negative correlations to the IPL and precuneus/PCC.

The greatest number of DLs at the left hemisphere were found at the middle portion of the STG at approximately A2, or the secondary auditory cortex (Fig. 2A). CON showed an expected pattern functional connectivity restricted to nearby cortex—local connections across the STG, and inferiorly to the superior temporal sulcus (STS). Additionally, the strongest connections (z(r) ≥0.5) respect the boundary of the primary somatomotor network. In contrast, the auditory node in DEL showed distant connections to the dmPFC and precuneus and posterior cingulate cortex (PCC) on the medial surface, and to the IPL on the lateral surface—cortical hubs of the DN. This is not an artifact of anatomical displacement from projection of the seed location from volume-space to surface-space (Supplemental Fig. 1). The auditory node of DEL also showed negative correlations to the brain regions that constitute the dorsal attention (dATN) and cingulo-opercular (CingOp) networks, which are the expected anti-correlations of the DN (Fox et al. 2005). Conversely, when hotspots were located at nodes of the DN, the connectivity from the dmPFC to the temporal lobe extended superiorly to include the same auditory node in DEL—in addition to regions cannonically associated with the DN (Fig. 2B). The IPL also retained its functional connections to canonical DN nodes, while being differentially connected to the STG and STS (Fig. 2C). Collectively, probing these hotspots reveal that the auditory node is functionally connected to the distributed, parallel-organized (default) network as a whole in 16p11.2 deletion carriers.

Of note, the aforementioned seed-based functional connectivity maps are not positioned to show the full intrinsic connectivity of canonical networks. Rather they highlight only the particular coordinates in which connectivity is different between groups. In fact, global architecture was preserved in DEL in terms of: 1) not demonstrating disruption of any single large-scale network (Supplemental Fig. 2), and 2) retaining regional preferences of the cortex to harbor local (hierarchical) versus distant (parallel) processing (Supplemental Fig. 3).

To be more precise, the auditory node in DEL is functionally connected to temporoparietal junction (TPJ), the lateral temporal cortex, temporal pole, inferior frontal gyrus (IFG), caudal MFG, and dmPFC—which constitute the dorsal medial subsystem of the DN (Andrews-Hanna et al. 2010; lower box). The same seed correlations are shown with 17-network boundaries in Supplemental Fig. 4. Of note, densely-sampled intra-individual data has revealed more accurate topographic boundaries for this subnetwork, which has been reinterpreted as Default Network-B (Braga and Buckner 2017). Regardless, in both interpretations, this network is the principal network of metalizing, theory of mind, and broadly, social cognition (Andrews-Hanna et al. 2014; DiNicola et al. 2020). This is an important distinction—the auditory node is functionally connected to the dorsomedial DN (DNb) and not any other subnetwork, namely the medial temporal subsystem (DNa).

A high degree of DLs were also found at the posterior insula (Fig. 2D). CON demonstrated functional connectivity restricted locally to the somatomotor network, with some extension anteriorly to the middle insula. In contrast the posterior insula DEL demonstrated strong connectivity with the parietal operculum (OP), middle insula, and pre-supplementary motor area (pre-SMA)—nodes of the CingOp (Sadaghiani and D’Esposito 2015). DEL also show the expected negative correlations of the CingOp from the posterior insula seed –i.e. anti-correlations to the IPL and PCC/precuneus nodes of DN. This is a second instance a primary sensorimotor area demonstrating functional connectivity with a heteromodal association area in DEL.

The remaining hotspots at the left hemisphere (Supplemental Fig. 5), and the hotspots of the right hemisphere (Supplemental Fig. 6) also demonstrated functional connections across large-scale networks. This included connections between heteromodal networks - such as default and frontoparietal control (FPN) as well as dorsal attention and FPN networks.

We replicated our results with a dataset previously published data by Bertero et al. (2018). Both their data and ours are a part of a larger Simons Foundation consortium, with 44% overlap of CON and 61% of DEL cases. Our sample was chosen based on our manual and automated quality control measures, and they used equally valid quality control measures. Differences in degree connectivity are likely a result of methodological choices (Supplemental Fig. 7). It is reassuring that both datasets converge on concordant results (Supplemental Fig. 8). A direct comparison of degree connectivity versus DLs was also performed on the Bertero dataset (Supplemental Fig. 9).

As confirmation, the changes observed in the left hemisphere functional connectivity of DEL were not secondary to reduced inter-hemispheric lateralization (of homologous large-scale networks)—as no group differences were found (Supplemental Fig. 10).

A post hoc, subgroup analysis of DLs between the sexes was also explored (Supplemental Figs 11 and 12).

Connectivity between cortical parcels verify between-network connections

To verify that distinct between-network connections were present in 16p11.2 deletion carriers, the data was analyzed using an independent methodology. In the DL analysis, the boundaries of Yeo’s networks were used simply as a visual aid. Here, the BOLD data was resampled according to Shaefer’s 200 predefined left hemisphere cortical parcels. Only parcels from relevant large-scale networks (6 of Yeo’s 17 networks) were assessed. The mean connectivity between cortical parcels is displayed as spring-loaded graphs for each group (Fig. 3A). Broadly, CON displayed tight-knit, well-delineated networks - reflecting high within-network intrinsic connectivity at the lower somatomotor nodes, language nodes, dorsomedial DN nodes (DNb), core DN nodes, CingOp nodes and salience nodes. In contrast, the same networks in DEL are observed to be more nebulous and less constrained in appearance, most notably for dorsomedial (DNb) network. This observation was quantified by the mean intrinsic connectivity strength for DNb for each group (Fig. 3B). A significant difference in the DNb network was found between the 2 groups with a composite mean z(r) 0.43 in CON and 0.33 in DEL (t = 2.11, P < 0.05), while controlling for age, sex, handedness, NVIQ, site, and motion.

Fig. 3.

The mean connectivity between cortical parcels support between-network connections. A) CON have tight-knit, well-delineated networks at lower somatomotor nodes (teal), language nodes (blue), dorsomedial DN nodes (DNb, salmon), core DN nodes (yellow), CingOp nodes (magenta) and salience nodes (pink) at the left hemisphere. In contrast, the same networks in DEL are nebulous and less constrained, most notably for the dorsomedial network (DNb). The clouds behind the lower somatomotor, language and DNb networks highlight the intermingled nodes in DEL. In particular, the STG2 node of the auditory cortex (black outline, and right box) is more closely connected to IFG2 of DNb in DEL (red outline, and dashed line in the right box) with a mean z(r) of 0.22 vs. 0.01 in CON. Similarly, the STS2 node of the language network (white outline, and right box) is more closely connected to IPL2 of DNb in DEL (red outline, and dashed line in the right box) with a mean z(r) of 0.52 vs. 0.12 in CON. B The composite mean intrinsic connectivity strength for DNb is significantly different between groups for raw values (shadow) and when adjusted for covariates (color-filled bars; ^*P < 0.05. Errors bars represent SEM).

On a more granular level, the auditory nodes (STG) were tightly connected to the posterior insula (pIns) in CON. In contrast, DEL showed fractionation of the lower somatomotor network such that pIns nodes were more closely connected to the middle insula (mIns) and parietal operculum (OP1) of the CingOp network, whereas STG nodes were found to be more closely connected to DNb nodes. In particular, the STG2 node of the auditory network is situated far from DNb nodes in CON (mean z(r) = 0.01), in contrast STG2 is more closely connected to IFG2 of DNb in DEL (mean z(r) = 0.22). Immediately inferior at the STS2 node of the language network (also called the tempoparietal network) has strong connectivity to IPL2 in DEL, a cortical hub of DNb, whereas the connection is minimal in CON (mean z(r) of 0.52 vs. 0.12).

Connectivity and behavior

The composite connectivity between nodes of the auditory (STG) and dorsomedial (DNb) networks was significantly different between CON and DEL. The mean connectivity strength, z(r), between auditory-dorsomedial (Aud-DNb) networks was −0.05 ± 0.12 in CON—meaning that the networks were not connected (~0) or slightly opposing and independent (<1). In contrast, Aud-DNb z(r) was 0.11 ± 0.12 in DEL (t = −5.28, P < 0.00001, Cohen’s d = 1.33), showing some interconnectivity between the 2 networks. When controlled for nuisance variables (age, sex, handedness, NVIQ, site, and micro-movements > 0.1 mm) the adjusted mean z(r) was −0.01 ± 0.12 for CON and 0.06 ± 0.10 for DEL, which remained significantly different (t = −2.72, P < 0.01, Cohen’s d = 0.70; Fig. 4A).

Fig. 4.

Connectivity strength between Aud-DNb correlates with communication and social impairment. A) The mean composite connectivity significantly different between groups for raw values (shadow) and after regression of nuisance variables (color-filled bars; ^**P < 0.01; Errors bars represent SEM). Next, Aud-DNb connectivity is plotted against the B) Vineland-II communication subscale and C) Vineland-II social subscale. The majority of DEL are above z(r) = 0 (dashed line) while CON are below. The vertical red line signifies an abnormal score on the Vineland-II (<85). Bolded triangles on the communication plot are those DEL who were independently diagnosed with phonological processing disorder (on DSM-IV), and bolded triangles on the social plot are those who meet criteria for ASD (on ADOS or ADI). D) When the communication subscale is plotted against social subscale it can be observed that positive Aud-DNb connectivity strength (red) is seen more often when both scores are low, and negative Aud-DNb (blue) is seen more often when both scores are high. E) Both DSM diagnoses, phonological processing disorder and ASD, are observed together more frequently when Aud-DNb interconnectivity is more positive and the (overall) Vineland-II score is lower. A single (either) diagnosis is observed as Aud-DNb becomes more negative or Vineland-II scores become higher. Neither diagnoses was seen in 3 DEL.

The strength of connectivity between Aud-DNb correlated significantly with behavior. More severe social impairment was associated with greater network interconnectivity on the SRS—a broad measure of social impairment (r = 0.45, P < 0.001; and r = 0.27, P < 0.05 after regression of nuisance variables); as well as a more specific measure that identifies cognitive/social impairment severe enough to warrant special needs—the Vineland-II (r = −0.50, P < 0.0001; and r = −0.36, P < 0.01 after regression of nuisance variables). The Vineland-II consists of a subscale for communication (measuring receptive, expressive, and written language), and a subscale for socialization (measuring interpersonal, play, and coping skills). Greater network interconnectivity between Aud-DNb was associated with significantly more communication impairment (r = −0.39, P < 0.01; Fig. 4B) and social impairment (r = −0.31, P < 0.05; Fig. 4C) while controlling for the same nuisance variables. DEL have shifted as a group towards positive Aud-DNb connectivity, rather than being driven by a few outliers. If Aud-DNb connectivity strength is also corrected for group (CON vs DEL), then the results become null (r = −0.18, P = 0.17 for communication; and r = −0.05, P = 0.70 for social impairment)—as the main effect on connectivity is removed.

As these networks become more interconnected one would expect less specialization in both networks, such that communication and social issues would occur jointly more often. Indeed, when the communication subscale is plotted against the social subscale of Vineland-II (Fig. 4D) positive Aud-DNb interconnectivity was observed more often when communication and social functioning were both impaired in individuals; and negative Aud-DNb interconnectivity was seen more often in individuals with high scores in both communication and social scales. Furthermore, phonological processing disorder (auditory network) and severe social impairment at a threshold of ASD (dorsomedial DN) were observed together (as 2 diagnoses) more often when both: (i) Aud-DNb correlation strength was high, and (ii) the overall Vineland-II score was low (Fig. 4E).

Discussion

Between-network interconnectivity and behavior

A search for unique connections, or significantly different links, between 16p11.2 deletion carriers and control participants was conducted using a spatially-unbiased, data-driven method. DLs were not uniformly distributed across the cortex, but instead clustered at particular brain regions. A seed placed at the epicenter of these hotspots did indeed visually demonstrate distinct topographic patterns of functional connectivity between the two groups. The greatest differences were found at the STG, insula, cingulate sulcus, inferior parietal lobule, dorsomedial prefrontal cortex, precentral and middle frontal gyri. Many of these hotspots for distinct functional connections coincide with regions of cortical gray matter previously reported to be most affected by the 16p11.2 CNV: the bilateral insula, transverse temporal gyrus, calcarine cortex, precentral, superior and middle temporal gyri (Martin-Brevet et al. 2018).

The auditory cortex was functionally connected to the DN in the left hemisphere of 16p11.2 deletion carriers. This is extremely peculiar as the transformation of sensation-to-perception necessitates local and hierarchal processing at primary sensorimotor areas (e.g. A1 to A2 to unimodal association areas); in contrast heteromodal association areas instead integrate information in parallel, across a widely-distributed cortical network (Goldman-Rakic 1988; Mesulam 1998). The DN is at the apex of transmodal networks, does not directly encounter low-level sensations, and functions as a cortical hub of integration (Buckner and Krienen 2013; Buckner and DiNicola 2019). Specialization requires segregation (Van Essen and Maunsell 1983; Mesulam 1998), and in DEL a single process, entangled large-scale networks, may have precluded each network from having the necessary fidelity to operate properly.

The hotspot at the STG (secondary auditory cortex) abuts the transverse temporal gyrus (primary auditory cortex). Dysfunction at this location would be expected to produce abnormalities in the translation of sound frequencies into phonemes. This is what was seen phenotypically in this cohort. The most frequent DSM diagnosis given to this cohort was phonological processing disorder (15/26, 58%). Previous reports also describe a delay of auditory evoked cortical response (M100) in 16p11.2 carriers (Berman et al. 2016; Jenkins III et al. 2016). Social deficits are also characteristic of this CNV (Hanson et al. 2015). In the present cohort, social deficits were demonstrated on the SRS and Vineland-II, and were severe enough to meet ASD criteria on ADOS or ADI in 14/25 (54%). The DN engages in a variety of self-related thoughts, but the dorsomedial subsystem of the DN, or DN-B, has been more specifically linked to mentalizing and theory of mind in social cognition (Andrews-Hanna et al. 2014; DiNicola et al. 2020).

Positive between-network correlation between the auditory and dorsomedial DN (Aud-DNb) was observed in the DEL, whereas negative between-network correlation was seen in the CON. This is a measure of the extent to which all of the nodes of these 2 networks are connected to each other—and therefore represents the degree to which the 2 networks are functionally one. We observed a group-level effect on connectivity, such that carrying 1 copy of 16p11.2 (rather than 2 copies in controls) shifts Aud-DNb interconnectivity in a positive direction.

Greater Aud-DNb interconnectivity was observed more often in those who jointly had low scores on both the communication and social subscales of the Vineland-II. Additionally, 2 diagnoses—phonological processing disorder (dysfunction of the auditory network) and ASD on ADOS/ADI (dysfunction of DNb)—occurred more frequently when Aud-DNb interconnectivity was high, and the Vineland-II score was low. The magnitude of anti-correlation of large-scale networks has been correlated to cognitive performance (Kelly et al. 2008). Here greater anti-correlation of Aud-DNb was indeed proportional to better performance on the Vineland-II—as the 2 large-scale networks were better able to maintain fidelity in their operations.

A second instance of local (hierarchical) and distant (parallel) processing networks sharing functional connectivity was observed between the primary somatomotor network (at the posterior insula and lower precentral gyrus) and the CingOp network (at the middle insula, parietal operculum and pre-SMA). Interestingly, the left posterior insula was among the most robust cluster identified across methodologies in the largest aggregated dataset of ASD patients (Di Martino et al. 2014). A forward gradient of adjacent, but functionally disparate, insular cortex starts with the interoceptive sensation x (posterior insula), to the subjective feeling of x (middle insula), and then the meta-cognitive self-awareness of x as a site of multimodal integration (anterior insula; Carr et al. 2003; Wicker et al. 2003; Craig 2009). Individuals with ASD report distorted emotional experiences, and an impaired ability to recognize the emotional states of others correctly (Hill et al. 2004; Blakemore et al. 2006; Ramachandran and Oberman 2006). Unfortunately, our study did not collect relevant behavioral data to correlate interception with observed changes in connectivity.

Models of autism

Aberrant links between large-scale networks is a novel model to explain brain connectivity and behavior. Namely, direct functional links between low-level sensorimotor networks and higher-order transmodal networks leads to poor performance of both networks involved. Prior models explain parts of the data without capturing the salient overarching pathophysiology of connectivity found in 16p11.2 deletion carriers.

Global disruption of any single large-scale network was not found in DEL (including the DN; Supplemental Fig. 2). However, the case could be made that the dorsomedial DN (DNb) was underconnected (Just et al. 2004). We found that the DNb had significantly lower mean intrinsic functional connectivity in DEL. Interestingly, our analysis, revealed that the underlying reason for less intrinsic connectivity was more extrinsic connectivity—between DNb and nodes of other networks, particularly auditory and language network nodes in DEL (as apparent on the spring-loaded graphs).

Similarly, a trend for an increased number of local connections at the DN (IPL and MFG) and an increased number of distant connections at the auditory network (STG/STS) in DEL would align with the quantitative dysconnectivity model (Uddin et al. 2013; Supplemental Fig. 3). A closer look at cortical parcels revealed that STG2 and STS2 are functionally connected to DNb, which would be counted as a topographically “distant” link; and at the same time IPL1 and MFG1 have weaker intrinsic connectivity to DNb, which would be counted as a topographically “local” link. So, while a quantitative change in the number of links (degree centrality) may indeed describe a cursory feature of this cohort, a more meaningful understanding of the data came from a qualitative understanding of which two nodes are aberrantly connected.

Aberrant links between large-scale networks were not isolated to these 2 networks, but rather, it was a theme found bilaterally across the cortex—at brain regions with a high degree of DLs. Primary sensorimotor networks were connected to heteromodal networks (Aud-DNb and pIns/M1-CingOp), and various heteromodal networks showed functionally connections between them (DN-FPN and dATN-FPN). Rather than to presume a model of underconnectivty or dysconnnectivty, a first-principles approach using differential link analysis revealed aberrant connections between large-scale networks.

Variability in interpretation can also occur when many teams analyze a single neuroimaging dataset. On this topic, Botvinik-Nezer et al. found that no 2 teams (of 70) choose an identical workflow, and were found to have substantial variability with high disagreement in the results reported. In contrast, greater consistency was found in the underlying statistical parametric maps across teams (Botvinik-Nezer et al. 2020). On surface, our results appear discrepant with previous investigations on the functional connectivity of 16p11.2 carriers. Moreau et al., found hyperconnectivity at ventral attention (aka. CingOp), motor, and frontoparietal networks (2020). Our data similarly showed greater connectivity between the CingOp and lower somatomotor networks (see Fig. 2D and Supplemental Fig. 5e); as well as between-network interconnectivity of the frontroparietal control network with the default and dorsal attention networks (see Supplemental Fig. 5f and g). Conversely, Bertero et al., reported hypoconnectivity from the dmPFC. When we analyzed the sample used by Bertero et al., using degree connectivity, we were able to replicate the decrease in distant connectivity at the dmPFC. Interestingly, the complementary statistical parametric maps of our sample and theirs taken together converge on the principal results of the DL analysis (specifically, DEL showed increased distant connectivity at the STG/STS, PCC and IPL in our dataset [see Supplemental Fig. 3]; increased local connectivity at the posterior insula, parietal and frontal operculums, and increased distant connectivity at the lower motor cortex in the Bertero set [see Supplemental Fig. 7]). The advantage of a DL analysis as a model for explaining discrepancies is elucidated when we directly compared both the degree and DL analysis in this second dataset (see Supplemental Fig. 9). Our aim was to search for unique connections in 16p11.2 carriers, using an unbiased method across the cortex, and then to investigate the consequences in connectivity and behavior that resulted from it. We found that the DNb network (which contains the aforementioned region at dmPFC) had “underconnected” intrinsic connectivity, but this was secondary to greater connectivity extrinsic to its network. Our data is consistent with prior work on 16p11.2 deletion carriers, and expands upon what it means to be hyper- or hypo-connected.

Local link to potential underlying neurobiological mechanism

We have previously shown that the increase in gray matter volume is primarily due to an effect of the CNV on surface area in 16p11.2 deletion carriers rather than cortical thickness (Qureshi et al. 2014; Martin-Brevet et al. 2018). This is not completely surprising as the 16p chromosome segment is among the most extreme cases of positive selection among humanoid species. Positive selection of 16p was seen after humans and great apes separated from orangutans, and positive selection continued after the divergence of human and chimpanzee lineages (Johnson et al. 2001). An increase in the number of neuroprogenitor cells was corroborated in a zebrafish model of the 16p11.2 microdeletion (Golzio et al. 2012; Golzio and Katsanis 2013), which translates to an increase in the number of cortical minicolumns—the fundamental neural unit of information processing (Buxhoeveden and Casanova 2002a, 2002b). In autism, minicolumns have been reported to be increased in number, narrower in width, with reduced neuropil space and smaller neuron cell bodies and nucleoli (Minshew and Williams 2007; Casanova 2015). Reduced inhibitory control over excitatory activity (Rubenstein and Merzenich 2003) within and between minicolumns could undermine the development of functionally discrete minicolumns in autism, such that boundaries become blurred and under-selective (Courchesne and Pierce 2005). This would allow for local excitation to last too long, and spread too far across local patches of cortex. This leads to retaining an abnormal number and pattern of local connections, and reduces the temporal and spatial resolution of input from distant sources (Courchesne and Pierce 2005). In other words, minicolumn pathology may simultaneously disrupt local and distant circuit organization (Just et al. 2004). The DL analysis revealed that local hotspots often spanned across borders of large-scale networks; and importantly, the same 2 networks involved were the 2 networks showing between-network interconnectivity from distant hotspots (Supplemental Fig. 13). This includes links that cross the auditory and DN boundaries (at the STG/STS), lower somatomotor and CingOp network boundaries (at the posterior insula and lower precentral gyrus), default and FPN network boundaries (at the IPL, MFG, and right medial cingulate sulcus), and dorsal attention and FPN boundaries (at the SPL). Critical changes in local connectivity at these border zones coincide with similar aberrant functional connections between large-scale networks observed on distant connectivity. Given the abrupt, fine cortical boundaries of juxtaposed large-scale networks with vastly disparate functions—the topographic positioning of altered local connectivity itself could provide a simple mechanism that connects the microscale (minicolumns) with the macroscale (large-scale networks). We suspect that connectivity between large-scale cortical networks have a greater impact on behavior when the networks involved are those that process in categorically different ways—those involved in local specialization (segregation) are linked to networks involved in global coherency (integration).

Limitations

We aimed to remove the confound of genetic heterogeneity in the study of autism. The limitation of this approach is that high pleiotropy and variable penetrance is characteristic of CNVs. Nevertheless, the results were not driven by outliers, but rather group-level effects on connectivity. As with most fMRI studies of ASD low-functioning individuals may have been excluded from high motion. Additionally, no duplication carriers available meet ASD criteria on ADOS—perhaps from ascertainment bias. As such, no inferences can be made in regard as to the reciprocal nature of this CNV. Also, no inference can be made about the longitudinal course of DEL carriers through adulthood. Group differences in resting-state functional connectivity can arise artificially from differences in motion between groups, misregistration of anatomical differences, and heterogeneity of functional connectivity within individuals of a group. Concerted effort was made to address these issues but can always remain as alternate explanations to group differences in functional connectivity. The cortical parcels used were delineated by the boundaries of the dorsomedial DN. For all intents and purposes DNb is a modern reinterpretation of the dorsomedial DN, but without precision functional mapping of each individual one cannot truly say whether “DN-B” was involved. Moreover, dense-sampling of individualized resting-state networks could potentially reveal specific individual boundary differences as an alternative explanation for between-network connectivity. However, the aim of this study was to characterize group-level effects on connectivity from the 16p11.2 microdeletion—abstracted from the many other environmental, genetic and epigenetic factors that affect a particular individual in the DEL group. Finally, extrapolation of our results to an ASD cohort may be limited from the possibility of multiple genome-to-connectome pathways in a highly heterogeneous condition.

Conclusions

In conclusion, a search for connections unique to 16p11.2 deletion carriers revealed aberrant links between large-scale cortical networks. Greater between-network interconnectivity between the auditory network (with local and hierarchical processing) and the dorsomedial DN (with distributed and parallel processing) was associated with greater communication and social impairment. At these hotspots, altered local connectivity that spanned across the borders of large-scale networks (such as between the auditory and DN, or lower somatomotor with the CingOp network) coincided with aberrant distant interconnectivity between large-scale networks. Given the abrupt, fine cortical boundaries of juxtaposed large-scale networks with vastly disparate functions—the topographic positioning of altered local connectivity itself could provide a potential mechanism that connects the microscale (minicolumns) with the macroscale (large-scale networks).

Author contributions

Abid Y. Qureshi (Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Project administration, Resources, Software, Validation, Visualization, Writing—original draft, Writing—review & editing), Jared A. Nielsen (Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Project administration, Visualization, Writing—original draft, Writing—review & editing), and Jorge Sepulcre (Conceptualization, Data curation, Investigation, Methodology, Project administration, Supervision, Validation, Visualization, Writing—review & editing)

Funding

A.Q. was supported by National Institutes of Health (NIH)/NINDS 5R25NS065743 and CTSA grant number UL1TR002366; and J.S. by the NIH grant numbers R01AG061445 and R01AG061811.

Conflict of interest statement: None declared.

Data availability

We appreciate obtaining access to phenotypic data on SFARI Base. Data were also provided in part by the Brain Genomics Superstruct Project of Harvard University and the Massachusetts General Hospital (principal investigators: Randy Buckner, Joshua Roffman, and Jordan Smoller). Approved researchers can obtain the Simons Searchlight population dataset described in this study by contacting the Simons Foundation Autism Research Initiative https://base.sfari.org/. They can also obtain the Brain Genomics Superstruct data by going to https://dataverse.harvard.edu/dataverse/GSP.

Ethics approval

Consent was obtained, and study procedures were approved by the relevant institutional review boards at Children Hospital of Philadelphia and University of California Berkeley.