An unbiased Bayesian approach to functional connectomics implicates social-communication networks in autism

1. Introduction

Autism spectrum disorder (ASD) is characterized by impaired social–emotional reciprocity, communication deficits, and stereotyped patterns of behavior (Americal Psychiatric Association, 2013). Thus far, the etiological complexity and phenotypic heterogeneity of ASD has greatly limited our understanding of its pathogenesis. Neurologically, ASD cannot be attributed to a single unified brain dysfunction (Waterhouse and Gilberg, 2014; Aoki et al., 2013); rather, it emerges from complex interactions across the brain (Just et al., 2004; Geschwind and Levitt, 2007; Sullivan et al., 2014). Behaviorally, the manifestation and severity of symptoms vary considerably across individuals and over the lifespan of each patient. These challenges have impeded the discovery of robust neuroimaging biomarkers for the disorder (Sullivan et al., 2014; Hernandez et al., 2014). Despite immense clinical diversity, social dysfunction is considered a hallmark and unifying feature of ASD (Pelphrey et al., 2014). Social abnormalities are apparent in both verbal and nonverbal domains and manifest across simple (e.g., shared gaze) and complex (e.g., back-and-forth conversation) behaviors. Since these deficits emerge within the first years of life, the prevailing theory posits that ASD alters the developmental trajectory of inter-regional connections (Just et al., 2004; Geschwind and Levitt 2007; Just et al., 2012; Melillo and Leisman, 2011; Courchesne and Pierce, 2005) via experience-dependent processes. Specifically, individuals with ASD exhibit reduced long-range connectivity (Wolff et al., 2012). These global patterns likely correspond to integration processes (Wolff et al., 2012; Oberman and Ramachandran, 2008) essential for higher-order social, emotional and communication functions (Rippon et al., 2007). At the same time, ASD patients show normal or increased connectivity within local networks, which may contribute to enhanced sensory-motor processing in clinical sub-populations (Johnson et al., 2002).

Task-based fMRI studies reveal an inconsistent pattern of hyper- and hypo-connectivity between the cortex and limbic structures (Just et al., 2012; Koshino et al., 2005). Such analyses have often been limited to pre-defined neural networks and do not reflect whole-brain patterns. Similarly, temporal correlations within resting-state fMRI (rsfMRI) are believed to reflect the intrinsic functional connectivity of the brain (Biswal et al., 1995; Fox and Raichle, 2007; Buckner and Vincent, 2007). Robust rsfMRI findings have focused on the default mode (Kennedy and Courchesne, 2008; Cherkassky et al., 2006) and salience (von dem Hagen et al., 2013) networks, whereas global rsfMRI analyses have identified distributed abnormalities across multiple neural processes (Supekar et al., 2013) and have yet to be reproduced. Recently, there has been a push to model the brain as a system of interconnected nodes. However, at present, the network analysis of ASD is limited to aggregate measures (e.g., centrality and degree distribution) (Paakki et al., 2010; Cherkassky et al., 2006; Noonan et al., 2009), which do not pinpoint a concrete etiological mechanism for the disorder.

In contrast to prior work, we propose to aggregate connectivity differences between ASD patients and Typically Developing Controls (TDCs) in order to localize the central nodes of the altered functional networks, i.e., the disease foci. Our novel approach (Venkataraman et al., 2013; Venkataraman, 2012) effectively translates pairwise rsfMRI correlations into estimates of the affected brain regions. Within a Bayesian framework, we define a latent graph that characterizes K independent networks of abnormal functional connectivity, emanating from the disease foci. These latent variables specify an underlying architecture that we cannot directly access, but which generates population-level differences in the observed rsfMRI data. Our solution algorithm jointly infers the regions affected by the disease and the induced network topology. Hence, our model provides new insights into the disrupted neural pathways of ASD and promises to enhance our understanding of its clinical phenomenology.

We demonstrate our approach on data from the publicly available and multi-site Autism Brain Imaging Data Exchange (ABIDE) (DiMartino et al., 2014). Previous work on ABIDE has revealed widespread hypo-connectivity across nearly every functional system, coupled with hyper-connectivity between sub-cortical areas (DiMartino et al., 2014). Region abnormalities have been localized to the thalamus, insula and occipital gyrus (DiMartino et al., 2014), which are not traditionally associated with the social brain. In contrast, our Bayesian model identifies key areas in the default network and temporal lobe that are thought to be involved in processing social stimuli. Moreover, the extracted networks tend to exhibit both reduced distal connectivity and increased local connectivity. We selected four ABIDE sites with well-matched populations of children and adolescents (see Table 2). We used non-parametric permutation tests and a bootstrapping setup to evaluate the reproducibility of our region foci on random subsets of the data. Encouragingly, the identified regions are mainly in the temporal and frontal lobes, and our results show remarkable similarities when inferring one (K = 1) and two (K = 2) abnormal networks. As such, our model can reliably integrate multi-acquisition rsfMRI data to extract robust and clinically informative biomarkers.

2. Materials and methods

3. Results

4. Discussion

The localization of region hubs, which describe pathological brain networks, provides key insights into the neural basis of ASD. Our Bayesian model (Venkataraman et al., 2013; Venkataraman, 2012) facilitates a biological interpretation of the group differences, and the validation procedures confirm the reliability of our findings. We emphasize that the model is completely data driven, i.e., we do not guide or constrain the solution according to a priori theoretical speculations about the networks that should be involved in the disorder. Moreover, the analysis is performed on whole-brain rsfMRI data and treats all regions/connections equally. Our approach marks a new direction of study for ASD and has the potential to reveal unbiased and complex influences within the brain.

Figs. 3 and 4 depict the salient patterns, assuming that population differences in rsfMRI correlations can be explained by one or two abnormal networks, respectively. The primary network consists of four ASD-related foci: the right temporal pole, the left posterior cingulate cortex, the left supramarginal gyrus, and the left middle temporal gyrus. Automated decoding of this network by the Neurosynth meta-analytic framework implicates the concepts of “language” and “comprehension” as likely the functional correlates. Our model also identifies a secondary network that consists of the left banks of the superior temporal sulcus, the right poster STS, extending into the inferior parietal lobule, and the right middle temporal gyrus. The primary associated functions include “social” and “person”. We observe that several of the altered connections in Fig. 4 (bottom) extend from the temporal lobe foci to the prefrontal cortex. Accordingly, it has been suggested that genes associated with ASD congregate in the deep layer projection neurons of the prefrontal and primary motor cortex, and that they are expressed during midfetal development (Tebbenkamp et al., 2014). Hence, exploring the relationship between our abnormal brain networks and the underlying genetic influences of ASD is an important direction for future work.

Our approach has revealed abnormal resting state functional connectivity in a network of brain regions known to engage in language processing (left middle temporal gyrus, the left posterior cingulate, the left supramarginal gyrus, and the right temporal pole). Language and communication difficulties are defining features of ASD. Specifically, language plays a central role in enabling academic achievement, vocational options, independent living, and mainstream integration for individuals with ASD. Task-based fMRI has informed us about the system-level organization of language processing. For example, the upper bank of the STS responds preferentially to the human voice, in comparison to other, well-matched acoustic signals (Belin et al., 2000; Hickok, 2009). Price's review (Price, 2009) on the anatomy of language in typical individuals suggests that phonemic perception is relatively localized to the bilateral superior temporal gyri (STG). Similarly, perception of meaningful speech localizes to the middle and inferior temporal cortex, word retrieval to the left angular gyrus and pars orbitalis, and sentence comprehension to the bilateral STS. Hickok (Hickok, 2009) has suggested that fronto-basal ganglia circuits and the anterior temporal lobe/pole are also involved in sentence comprehension.

Task-based fMRI studies of language have also revealed lateralization differences in ASD. For example, an investigation of speech processing in sleeping 2–3 year olds found reduced activity in brain regions associated with language comprehension, along with greater activation in medial and frontal areas of the right hemisphere in ASD (Redcay and Courchesne, 2008). This trend towards greater recruitment of right hemisphere regions increases with age into the preschool period (Eyler et al., 2012). Functional neuroimaging studies in older speakers with ASD have also found evidence of hemispheric abnormalities. Reduced activation in the left inferior frontal gyrus (Just et al., 2004; Garey et al., 2007; Kana et al., 2006; Lai et al., 2012a,b; Mülller et al., 1998, 1999a) and atypical activation in the right frontal lobe (Mülller et al., 1999a,b) have been reported during language paradigms. A meta-analysis by Phillips (Philip et al., 2012) recently summarized these findings to suggest greater activation in the right precentral gyrus in children with ASD, as well as reduced activation in the bilateral STG, and generally decreased left lateralization of a variety of language functions (see (Boddaert et al., 2003; Dawson et al., 1989; Kleinhans et al., 2008; Knaus et al., 2008) for more details).

Diffusion tensor imaging (DTI) has demonstrated reduced anatomical connectivity for language processing (Just et al., 2004; Kana et al., 2006), as well as white matter disruptions in language pathways (Alexander et al., 2007; Barnea-Goraly et al., 2004; Keller et al., 2007; Lee et al., 2007) in children with ASD. Minshew and Keller (Minshew and Keller, 2010) suggest that speakers with ASD show greater than normal activation and connectivity in posterior areas and reduced connectivity to frontal systems typically involved in language processing. The authors interpret these trends as a sign of over-reliance on visuospatial abilities for verbal reasoning tasks.

Our Bayesian model also pinpoints dysfunction in a well known social perception network centered in the right posterior STS, extending into the inferior parietal lobule and the right middle temporal gyrus (Yang et al., 2015). This network receives and interprets direct input from primary visual and auditory regions in human and nonhuman primates (Belin et al., 2000; Brothers, 1996; Hoffman and Haxby, 2000; Jastorff et al., 2012; Kreifelts et al., 2009). The posterior STS is sensitive to and selective for social stimuli that signal intent in humans (Jastorff et al., 2012). It responds to static social (i.e., faces) vs. non-social stimuli (i.e., objects) as well as complex dynamic social information (Gobbini and Haxby, 2007; Watson et al., 2014). The posterior STS responds more strongly to socially meaningful human actions than to non-goal-directed movements (Bahnemann et al., 2010). In the auditory domain, the posterior STS responds to affective speech (Wildgruber et al., 2006). Ethofer et al. (2006) discovered that the posterior STS is the input region of the affective voice processing system, where such information is extracted and represented. In terms of somatosensory input, stroking skin areas with C-tactile (CT) nerves with a soft brush, thereby mimicking the type of touch during a close social interaction, has been shown to activate the posterior STS (Björnsdotterr et al., 2014; Gordon et al., 2013). Finally, the posterior STS is functionally interconnected to all key regions in the “social brain” (Yang et al., 2015).

There is strong evidence for atypical cortical and subcortical processing of visual, auditory, and somatosensory signals related to social perception in ASD. When young children with ASD were shown point-light displays of coherent versus scrambled biological motion, they exhibited reduced activity in the FFG, amygdala and posterior STS compared to their unaffected siblings and typically developing peers (Kaiser et al., 2010). In the auditory domain, individuals with ASD showed a similar activation pattern to controls for non-vocal sounds but failed to activate the voice-selective regions of the posterior STS in response to vocal sounds (Gervais et al., 2004). Finally, with respect to somatosensory signals, individuals that reported more autistic traits displayed reduced activity in the OFC and posterior STS in response to slow and gentile touch to the forearm (Voos et al., 2013).

It is worth mentioning that our method does not identify certain regions that have been previously reported in the ASD literature. Examples include the prefrontal cortex, as related to working memory and executive function (Just et al., 2004; Courchesne and Pierce, 2005; Baron-Cohen et al., 1999; Gilbert et al., 2008), and the visuomotor cortex (Koshino et al., 2008; Allen and Courchesne, 2003; Muller et al., 2003; Villalobos et al., 2005). The absence of these findings suggests weak evidence in the data for prefrontal and visual areas to act as foci. Contributing factors include subtleties of the rsfMRI paradigm, high levels of noise due inter-site variability, heterogeneity of ASD symptoms across subjects, unexplored genetic influences, or (most likely) a combination of the above. A future refinement to the Bayesian framework might include additional random variables to account for external influences. However, such modifications are beyond the scope of this paper.

A major limitation of prior connectivity analyses is the difficulty interpreting the results post hoc. For example, univariate tests select independent connections influenced by disorder. However, the bulk of our knowledge about the brain is organized around regions (e.g., functional localization, cortical thickness) and not the connections between them. Moreover, it is nearly impossible to design noninvasive experiments that probe an arbitrary connection in the brain. While region statistics attempt to localize the impairments, they also average the pairwise effects measured by rsfMRI. Hence, we can rarely tie statistical differences in region statistics to an underlying etiological mechanism. Our method overcomes these limitations by explicitly modeling the propagation of information in the brain from regions to connections. Given the rsfMRI correlation matrices, we can easily infer both the disease foci and the neural breakdown associated with ASD. These variables have a straightforward biological meaning and can be used to design follow-up studies for ASD.

We provide two forms of validation: statistical significance, as quantified by non-parametric permutation tests, and test–retest reliability on random subsets of the data. On average, we are able to find the original disease foci in Figs. 7 and 8. However, these experiments highlight an important point; namely, the most robust foci have among the lowest statistical significance in Figs. 3 and 4. This ties into an ongoing debate about the utility and reliability of p-values (Goodman, 2008). Furthermore, it is unclear whether to apply significance and multiple comparison correction to Bayesian models, since these formulations are not designed for conventional hypothesis testing. Encouragingly, both evaluation schemes produce similar results, which supports that the region foci correspond to robust phenomena within the highly heterogeneous ABIDE data.

It is important to note that our results may be influenced by the selection of regions. Due to the copious inter-site and inter-subject variability of rsfMRI within ABIDE, if the regions are too small, then group differences are likely driven by noise. In contrast, larger regions potentially blur relevant functional distinctions. This work relies on the Desikan–Killiany atlas in Freesurfer (Fischl et al., 2004), which provide anatomically meaningful correspondences across subjects that relate to functional divisions within the brain. We include the results for K = 1, 2, and 3 using the Freesurfer Destrieux atlas in Supplementary Results. As seen, there is a great deal of overlap between the two parcellations, in terms of identifying social brain regions. However, there are lso notable differences, which suggest that selecting the appropriate resolution for analysis is a nontrivial problem. We emphasize that the Bayesian framework readily applies to any set of consistent ROIs across subjects.

Finally, we address the model complexity parameter K, which specifies the number of abnormal networks to explain the data. Our results for K = 1 and 2 suggest an evolution of disease foci, such that additional clusters explain progressively weaker phenomena. Intuitively, the reproducibility goes down with increasing K, as the model possesses more degrees of freedom (see Fig. 7 vs. 8). This posits a tradeoff between robustness and our ability to pinpoint smaller effects. Due to the absence of ground truth connectivity information, we believe that meta-analysis and clinical relevance are the best measures to evaluate the impact of K. The present work uses a data-driven clustering scheme to estimate multiple abnormal networks. Alternatively, we might use clinical scores to tailor the networks to particular subtypes of the ASD population. Therefore, our Bayesian model promises to be a powerful analysis tool for future exploration.

5. Conclusion

We have demonstrated that a Bayesian analysis of whole-brain functional connectivity, as measured via resting-state fMRI, supports the theory of impaired social cognition in ASD. Our model captures multivariate dependencies between pairwise connections to infer both the region foci that are most affected by ASD and the corresponding networks of functional abnormalities. We leverage the multi-site Autism Brain Imaging Data Exchange and localize the regions implicated by ASD to the temporal lobe and default mode network. Reliability of these results is confirmed by non-parametric permutation tests and test–retest experiments. Our discoveries provide valuable insights into the underlying neural mechanisms of ASD and bring us one step closer to deciphering its complex pathogenesis.

Appendix A. Bayesian model formulation

This appendix highlights the mathematical details of our Bayesian model from Section 2.3. The complete generative process is depicted in Fig. 2. Additional details about the probabilistic formulation and inference algorithm can be found in Venkataraman et al. (2013) and Venkataraman (2012).

The multinomial variable R_i indicates whether region i is healthy (R_i = 0), or whether it belongs to an abnormal community k. We assume an i.i.d. multinomial prior for R_i:

where the unknown prior parameters π_k^r are shared by all nodes in the network.

The latent functional connectivity F_ij describes how region i and region j co-activate in the TDC population: positive synchrony (F_ij = 1), negative synchrony (F_ij = − 1), and no co-activation (F_ij = 0). This tri-state random variable is drawn from a multinomial distribution with parameters π_s^f. For convenience, we represent F_ij as a length three indicator vector with exactly one of its elements [F_ij,− 1 F_ij0 F_ij1]^T equal to one, i.e.,

This discrete representation of functional connectivity is a notable departure from conventional analysis. Essentially, we assume that rsfMRI correlations fall into one of three general categories and differences in bin assignment are the relevant markers of a disease. Our choice of three states is motivated by the rsfMRI literature. For example, most works specify a threshold to determine functionally connected areas, which corresponds to F_ij = 1 in our framework. Moreover, strong negative correlations are often found in rsfMRI data. Since there is no consensus about their origin and significance (van Dijk et al., 2010), we isolate negative connectivity (i.e., F_ij = − 1) as a separate category.

The latent functional connectivity ${\overline{F}}_{ij}$ of the ASD population is also tri-state and is based on F_ij and the region indicator variables R_i:

such that ε₁ = ηε + (1 − η)(1 − ε). The first condition in Eq. (A.3) states that if both regions are healthy (R_i = R_j = 0), then the edge 〈i, j〉 is healthy. Consequently, the functional connectivity of the ASD population is equal to that of the control population with probability 1 − ε, and it differs with probability ε. This relationship also holds if the two regions are foci in different abnormal networks (second line of Eq. (A.3)). The third term is similarly obtained by replacing ε with 1 − ε. Finally, the probability ε₁ in the fourth condition reflects the coupling between η and ε if one region is healthy and the other is diseased.

The rsfMRI correlation B_ij^l between regions i and j in the lth TDC subject is a noisy observation of the functional connectivity F_ij:

We fix μ₀ = 0 to center the parameter estimates. The likelihood for ${\overline{B}}_{ij}^m$ in the ASD population has the same functional form and parameter values as Eq. (A.4) but relies on the clinical template ${\overline{F}}_{ij}$. We have previously demonstrated (Venkataraman et al., 2012) that the Gaussian distribution in Eq. (A.4) is a reasonable approximation to the global histogram of rsfMRI data.

Appendix B. Supplementary data