Basal Ganglia Impairments in Autism Spectrum Disorder Are Related to Abnormal Signal Gating to Prefrontal Cortex

Abstract

Research on the biological basis of autism spectrum disorder has yielded a list of brain abnormalities that are arguably as diverse as the set of behavioral symptoms that characterize the disorder. Among these are patterns of abnormal cortical connectivity and abnormal basal ganglia development. In attempts to integrate the existing literature, the current paper tests the hypothesis that impairments in the basal ganglia's function to flexibly select and route task-relevant neural signals to the prefrontal cortex underpins patterns of abnormal synchronization between the prefrontal cortex and other cortical processing centers observed in individuals with autism spectrum disorder (ASD). We tested this hypothesis using a Dynamic Causal Modeling analysis of neuroimaging data collected from 16 individuals with ASD (mean age = 25.3 years; 6 female) and 17 age- and IQ-matched neurotypical controls (mean age = 25.6, 6 female), who performed a Go/No-Go test of executive functioning. Consistent with the hypothesis tested, a random-effects Bayesian model selection procedure determined that a model of network connectivity in which basal ganglia activation modulated connectivity between the prefrontal cortex and other key cortical processing centers best fit the data of both neurotypicals and individuals with ASD. Follow-up analyses suggested that the largest group differences were observed for modulation of connectivity between prefrontal cortex and the sensory input region in the occipital lobe [t(31) = 2.03, p = 0.025]. Specifically, basal ganglia activation was associated with a small decrease in synchronization between the occipital region and prefrontal cortical regions in controls; however, in individuals with ASD, basal ganglia activation resulted in increased synchronization between the occipital region and the prefrontal cortex. We propose that this increased synchronization may reflect a failure in basal ganglia signal gating mechanisms, resulting in a non-selective copying of signals to prefrontal cortex. Such a failure to prioritize and filter signals to the prefrontal cortex could result in the pervasive impairments in cognitive flexibility and executive functioning that characterize autism spectrum disorder, and may offer a mechanistic explanation of some of the observed abnormalities in patterns of cortical synchronization in ASD.

1. Introduction

When considering the results from the past 25 years of research on the biological basis of autism spectrum disorder (ASD), the choice of the “puzzle” as a symbol for ASD seems particularly appropriate. Integrating the individual research findings into a cohesive picture has proven to be a daunting task, because the neural abnormalities associated with the disorder seem to be as diverse as the list of behavioral symptoms that characterize it. Whether one or many mechanisms underlie the disorder (e.g., Happé, Ronald, & Plomin, 2006), integrating the existing findings within a theoretical framework that can account for diverse descriptions at both behavioral and brain levels will greatly enhance our understanding of the biological nature of ASD. The current article describes a test of one such framework, which posits that abnormal basal ganglia development in ASD results in abnormal neural signal routing to the prefrontal cortex.

1.1 Abnormal Cortical Synchronization in ASD

Theories that describe deficits in network-level functioning in ASD have become increasingly popular, as they can account for diverse psychological phenomena (Just, Cherkassky, Keller, Kana, & Minshew, 2007; Just, Keller, Malave, Kana, & Varma, 2012; Kana, Libero, & Moore, 2011; Kleinhans et al., 2008; Müller, 2007; Schipul, Keller, & Just, 2011). These network-level characterizations of function offer advantages over more focal descriptions in that they can explain general behavioral deficits through failures to integrate the outputs of key cortical processing centers (e.g., Just et al., 2012; Müller, 2007). However, evidence is divided about whether network-level connectivity problems in ASD results from under-connectivity (e.g., Just et al., 2012; Schipul, Williams, Keller, Minshew, & Just, 2012), over-connectivity (Mizuno, Villalobos, Davies, Dahl, & Müller, 2006; Noonan, Haist, & Müller, 2009; Supekar et al., 2013), or some combination of the two (e.g., Courschesne & Pierce, 2005; Kana et al., 2011; Uddin, Supekar, & Menon, 2013).

Whatever the direction of abnormal neural synchronization may be, another missing piece of the puzzle concerns the mechanism(s) underpinning the deviant patterns of connectivity. The most common explanation for under-connectivity in ASD has been that it reflects abnormalities in the white matter microstructure connecting frontal and posterior regions of the brain (e.g., Just et al., 2007; Kana et al., 2011). It is worth noting, however, that differences in structural connectivity do not necessarily relate to differences in functional synchronization, as measured by correlations of timecourses of BOLD signal fluctuations between regions. For example, Tyszka, Kennedy, Adolphs, & Paul, (2011) found normal bilateral functional connectivity in individuals without a corpus callosum, the major band of white matter tracts structurally connecting the two hemispheres. In addition, numerous experiments have shown that patterns of functional connectivity change as a function of varying task demands (e.g., Prat, Keller, & Just, 2007; Prat & Just, 2011), and varying methods used for assessment (e.g., Müller et al., 2011), whereas the underlying structural connectivity measured is presumably not influenced by these variables. The current investigation used Dynamic Causal Modeling, which has previously been used to characterize group differences in complex task performance (e.g., Becker, Prat, & Stocco, 2016) to explore a potential functional mechanism that has not been previously considered, namely that impaired signal gating in the basal ganglia can explain patterns of abnormal functional connectivity to the prefrontal cortex observed in individuals with ASD, providing a critical “missing piece” for solving the puzzle of the biological underpinnings of autism spectrum disorder.

1.2 The Role of the Basal Ganglia in Cortical Synchronization

The theory tested in this article is based on cross-disciplinary research demonstrating that the basal ganglia nuclei function as a dynamic “gate,” controlling the transmission of neural signals to the prefrontal cortex (Frank, Loughry, & O'Reilly, 2001; O'Reilly & Frank, 2006; Stocco, Lebiere, & Anderson, 2010). The basal ganglia are a set of interconnected nuclei that form a complex inhibitory circuit, controlling inputs to the frontal lobes through the thalamus. Of these nuclei, the dopamine-rich dorsal striatum (itself composed of two components, the caudate nucleus and the putamen) plays a particularly important role, as it functions as the input station of the circuit, receiving projections from the entire cortex. Thus, the striatum is in an ideal position to function like a dynamic “director” of information, monitoring signals from throughout the cortex, and prioritizing their transfer to the prefrontal cortex.

According to the Conditional Information Routing Model put forth by Stocco et al. (2010; Stocco & Lebiere, 2014), the basal ganglia operate as a system that can flexibly impose order over the highly overlapping exchange of signals between networks of cortical regions. Because the prefrontal cortex lies at the apex of a hierarchy of converging pathways, “conflict” situations, in which multiple regions compete to influence a single region, are likely to occur. In the absence of basal ganglia interventions, the flow of signals across the network, and resulting measures of cortical connectivity, are determined by the relative strength of cortico-cortical projections. Such relative strength is in turn shaped by previous practice and reward contingencies, and under normal conditions, is sufficient to produce the desired behaviors. In certain conditions however, this is not the case. For instance, during learning (where relevant preexisting cortical networks have not been established), or when previously successful responses become inappropriate (such as in the dimensional card sorting task or other tests of task switching), the basal ganglia can shape behavior by prioritizing the transmission of weaker signals to the prefrontal cortex. This can have the effect of overriding (or biasing) existing cortico-cortical connections, and thus modulating patterns of functional synchronization between these regions (Stocco et al., 2010; Stocco et al., 2012; Becker, Prat & Stocco, 2016; Verleger, Schroll, & Hamker, 2013). As a consequence, measures of cortico-cortical connectivity are in fact biased by the underlying patterns of basal ganglia activity. This has been empirically verified in patients with basal ganglia specific disorders, such as Parkinson's disease, who exhibit abrnormal cortico-cortical connectivity (Hammond, Bergman, & Brown, 2007; Lebedev et al., 2014; Verleger, Schroll, & Hamker, 2013). Because the basal ganglia selectively outputs to frontal regions, their effects on measures of functional connectivity are likely larger when measured on pairs of regions which include at least one frontal target—which, in turn, is the case for most types of cognitive tasks. Therefore, the Conditional Routing model provides a potential subcortical mechanism (i.e., abnormal signal gating to the prefrontal cortex) that might underlie the observed task-specific patterns of abnormal cortical synchronization observed in neuroimaging investigations of individuals with ASD.

1.3 Basal Ganglia Abnormalities in ASD

Both structural and functional abnormalities of the basal ganglia have been widely documented in individuals with ASD (e.g., Kriete, 2008). For instance, several studies have reported that the dorsal striatum is enlarged in ASD compared to neurotypical individuals (Estes et al., 2011; Haznedar et al., 2006; Herbert et al., 2003; Hollander et al., 2005; Langen et al., 2009; Rojas et al., 2006; Sears et al., 1999). This difference can be observed in young children (Estes et al., 2011), but seems to become exacerbated with age. Specifically, Langen et al. (2009) investigated 188 individuals ranging from 6 to 25 years of age and found abnormal development trajectories of the dorsal striatum in ASD, with neurotypicals showing decreased striatal volume (especially in the caudate nucleus) with increased age, whereas individuals with ASD showed increased striatal volume with increased age. Despite having a generally larger basal ganglia, a few neuroimaging experiments have shown less activation in the basal ganglia of individuals with ASD than in controls during learning and higher-level cognitive tasks (Haznedar et al., 2006; Schipul et al., 2011; Silk et al., 2006). In addition, patterns of connectivity between the basal ganglia and cortical regions also differ significantly between individuals with ASD and controls (Langen et al., 2012).

Although there is extensive literature showing that the basal ganglia are atypical in ASD, investigations of the functional implications of these differences have primarily focused on abnormal (repetitive or stereotyped) motor behaviors (Estes et al., 2011; Hollander et al., 2005; Langen et al., 2009; Rojas et al., 2006). Thus, another way in which the current research provides a missing piece to the puzzle of ASD, is by building a bridge between abnormal basal ganglia functioning, and a more broad range of cognitive and behavioral phenomena. For instance, the basal ganglia circuit has recently been implicated in language processes (Frederici, 2006; Prat & Just, 2011; Stocco, Yamasaki, Natalenko & Prat, 2014; Stocco & Prat, 2014), planning (Monchi, Petrides, Strafella, Worsley, & Doyon, 2006), and social cognition (Rojas et al., 2006), all of which are known to be impaired in ASD.

While it may be overly ambitious to propose a unitary biological cause for ASD, we propose that an understanding of the disorder can be advanced by the integration of seemingly inconsistent and/or unrelated findings in both the neuroscientific and behavioral domains. The proposed impairment in a mechanism for routing signals to the frontal lobes in ASD tested herein resolves some of these inconsistencies by bridging previously unrelated behavioral and neural phenomena.

1.4 Using Dynamic Causal Modeling to Investigate Impaired Basal Ganglia Signal Gating

The conditional routing model (Stocco et al., 2010) predicts that one of the main basal ganglia functions is to modulate the connectivity between pairs of cortical regions involving the prefrontal cortex through dynamic signal filtering, and the goal of the current investigation is to test whether this modulatory function is impaired in individuals with ASD. To test our theory, we must be able to separate the cause and directionality of abnormal cortical synchronization, while estimating the influence that basal ganglia activation has on these patterns of synchronization. To achieve this goal, we used Dynamic Causal Modeling (DCM: Friston, Harrison, & Penny, 2003), which can separate the modulatory effects of basal ganglia activation on cortical synchronization from the task-related influences that might drive changes in effective connectivity (Müller et al., 2011).

DCM can be seen as an application of dynamic systems theory to the analysis of neuroimaging data. As in conventional approaches, different task conditions are modeled as experimental factors, and are allowed to explain fluctuations in the activity of each region. In DCM, however, the experimenter also provides a model of how various brain regions are functionally connected using a network of directional connections. Thus, the analysis of each region's activity includes, not only the given task conditions, but also the effects that are propagated from other regions throughout the network. Because the network-level effects of signals traveling in one direction can be significantly different than those of signals traveling in the opposite direction, DCM can estimate the directionality of these effects (e.g., from region A to region B vs. from region B to region A), as well as their strengths. In addition to providing indices of effective and modulatory connectivity, DCM also yields indices of overall “fit” of the data. Thus, it is an excellent technique for comparing various models of the flow of information across cortical networks. As a proof of its robustness, DCM has been applied to known pathways in the brain and has replicated their directionality, such as the existence of modulatory backward connections from the superior parietal cortex to visual area V5 (Penny, Stephan, Mechelli, & Friston, 2004) or the subcortical effects of epileptogenic neurons in rat models (David et al. 2008). As a result of its success, DCM is being increasingly applied to the study of special populations (Posner et al., 2011; Seghier, Zeidman, Neufeld, Leff, & Price, 2010; Schlösser et al., 2008; Schlösser et al., 2010; Sladky et al., 2015).

In the current study, we applied DCM to an existing fMRI dataset to compare basal ganglia modulation of cortico-cortical connections while individuals with ASD and controls completed a task designed to measure executive functioning, the Go/NoGo task. Our theory that ASD might be related to a lack of basal ganglia-mediated modulation of cortico-cortical connectivity generates the following testable predictions: (1) that a model in which the basal ganglia modulate patterns of cortical connectivity to the prefrontal cortex will fit the data better than a comparable model where the basal ganglia do not modulate such connectivity, especially for control participants, and (2) that the modulatory effects of the basal ganglia on pairs of cortical centers involving the prefrontal cortex will differ between individuals with ASD and controls. More specifically, according to the Conditional Information Routing model, the modulatory effects of the basal ganglia on pairs of cortical connectivity should be negative. This prediction arises because the prefrontal cortex is receiving synchronous and overlapping signals from multiple posterior regions at any given time, and the ongoing activity drives the synchronization between frontal and posterior regions. The activity of the basal ganglia modifies this situation through its gating function, by selecting only a subset of the possible signals that converge over the prefrontal cortex, and only at appropriate moments during the task, while filtering out the others. Thus, unimpaired basal ganglia activity effectively reduces the average synchronization between most of the posterior cortical regions and the prefrontal cortex, resulting in a negative modulation value. On the other hand, impaired functioning of the basal ganglia in ASD may be associated with either the lack of a modulatory effect, or, with a pathological increase of synchronization between cortical regions. Although increased cortical synchronization is often viewed as advantageous in normally functioning populations (e.g., Prat et al., 2007), an increase in synchronization relating to basal ganglia functioning may indicate the lack of an ability to filter out irrelevant signals. Such increases in synchronization between cortical regions with basal ganglia damage have been observed in patients affected by basal ganglia pathologies such as Parkinson's disease (Moazami-Goudarzi, Sarnthein, Michels, Moukhtieva, & Jeanmonod, 2008; Stoffers et al., 2008). In summary, the current experiment uses DCM to test a prediction of the impaired signal gating hypothesis, that individuals with ASD and neurotypicals will differ in the extent to which basal ganglia activation modulates patterns of cortical connectivity to the prefrontal cortex.

2. Methods

2.1 Participants

Data were collected from 16 individuals with ASD (mean age = 25.3 years; 6 female) and 17 healthy controls (mean age = 25.6, 6 female). All participants provided informed consent using procedures approved by the Institutional Review Board for human subjects research at the University of Washington. Data from one individual with ASD had to be discarded because the estimation procedure for the corresponding Modulatory model did not converge to a single solution. Healthy controls and individuals with ASD were recruited through Internet postings and flyers. Both groups were tested with the Wechsler Abbreviated Scale of Intelligence (WASI: Wechsler, 1999) and had comparable measures of full-scale IQ [t(29) = 0.68, p > 0.48]. In addition, individuals with ASD were administered the Autism Diagnostic Interview-Revised (ADI-R: Lord, Rutter, & Couteur, 1994) and the Autism Diagnostic Observation Schedule (ADOS: Lord et al., 2000). Clinical diagnoses were confirmed using clinical judgment based on all available information and DSM-IV criteria (American Psychiatric Association, 1994). The ASD sample included Asperger's Syndrome, (N=9), Autistic Disorder (N=5), and Pervasive Developmental Disorder - Not Otherwise Specified (N=2). Psychometric properties of the ASD and control groups are summarized in Table 1.

Table 1.

Characteristics of the ASD and control groups used in the study.

Measure		ASD	Controls	Difference
Group size		16 (6 female)	17 (6 female)	NA
Age	M +/− SD	25.3 +/− 5.0	25.6 +/− 7.2	t(31) = 0.10, n.s.
	Range	18.5 – 35.7	19.1 – 44.5
WASI Full IQ	M +/− SD	107.4 +/− 16.7	111.0 +/−12.7	t(31) = 0.70, n.s.
	Range	67− – 129	83 – 132
WASI Verbal IQ	M +/− SD	105.9 +/− 20.8	109.8 +/ 12.6	t(31) = 0.65, n.s.
	Range	59 – 140	88 – 127
WASI Performance IQ	M +/− SD	107.3 +/− 13.0	109.5 +/− 12.8	t(31) = 0.51, n.s.
	Range	78 – 131	79 – 133
Com Q	M +/− SD	20.3 +/− 6.3	31.5 +/− 4.3	t(31) = 5.70, p < 0.001
	Range	13 – 37	23 – 37
Social Avoidance Distress	M +/− SD	14.6 +/− 6.5	4.8 +/− 4.2	t(31) = 4.95, p < 0.001
	Range	3 - 24	0 – 14

2.2 Materials and Procedure

Imaging data were collected while participants performed a Go/NoGo task (see Figure 1 for a schematic), a task designed to measure the ability to inhibit prepotent responses, and known to rely on prefrontal contributions (e.g., Durston, Thomas, Worden, Yang, & Casey, 2002; Hester, Fassbender & Garavan, 2004). The Go/NoGo task was chosen because it is a canonical measure of motor inhibitory control that has been previously shown to relate to indices of structural connectivity of the dorsal striatum (Langen et al., 2012). The task stimuli consisted of either letters or faces drawn from the NimStim (http://www.macbrain.org/resources.htm) database (e.g., Tottenham et al., 2009) which were presented sequentially for 1s each, with 500ms inter-trial intervals. Stimuli were presented in homogeneous blocks of 12 items each (e.g., a sequence of stimuli consisting of only letters, or only faces). Participants had to respond to each stimulus by pressing a button, unless it was either an “X” (for letter blocks) or a sad face (for face blocks). Blocks could be either “Go” or “Go/NoGo” blocks. In “Go” blocks, none of the letters were “X”, and all of the faces were happy, thus requiring a button press on every trial. In “Go/NoGo” blocks, on the other hand, 50% of the trials “X”s were embedded within the sequence of letters, or sad faces were embedded within a sequence of happy faces. Thus, “Go/NoGo” blocks demand additional mental control to successfully withhold a response for the “NoGo” stimuli, in the face of an otherwise prepotent tendency to respond to the frequent “Go” stimuli. In summary, the experiment employed a 2-by-2 factorial design, where the factors were Stimuli Type (Letters vs. Faces) and Block Type (Go vs. Go/NoGo). To give participants an opportunity to adjust their mental strategy, written instructions on the screen before each block informed participants whether the forthcoming sequence of stimuli was a “Go” or “Go/NoGo” block.

Fig 1.

Design of the experiment. The experiment consisted of blocks of stimuli that could be either single alphabetical letters or black-and-white face pictures. Participants were told to press a button with their right hand for every letter and face, unless the letter was “X” or the face was sad, instead of happy. Half of the blocks were “Go” blocks, where participants had to respond to all the stimuli. The remaining were “Go/NoGo” blocks, where a few “NoGo” trials (i.e., “X”s or sad faces) were randomly intermixed within the “Go” stimuli. Participants were told in advance whether the next block was a “Go” or a “Go/NoGo” block.

2.3 Imaging Data Acquisition and Preprocessing

The data were collected at the Diagnostic Imaging Sciences Center at the University of Washington with a Philips Achieva 3T scanner with an 8-channel coil, using an echo-planar imaging (EPI) sequence with a TR = 2s, TE = 21ms, and FoV = 220 × 220mm. Functional scans consisted of 38 axial slices with 3.5mm thickness, 0mm gap, and an in-plane matrix of 64 × 64 voxels with an in-plane resolution of 3.44 × 3.44mm. In addition to the functional images, a high-resolution MP RAGE structural scan was also acquired from each participant to improve the normalization process. The structural images consisted of 180 sagittal slices with 1mm thickness, 0mm gap, and an in-plane matrix of 256 × 256 voxels with an in-plane resolution of 0.86 × 0.86mm.

All of the analyses were conducted with the Statistical Parametric Mapping (SPM8) processing software (Friston et al., 2007). Before being analyzed, functional images were corrected for differences in slice acquisition time, spatially realigned to the first image in the series, normalized to the Montreal Neurological Institute (MNI) linear International Consortium for Brain Mapping (ICBM) 152 template, resampled to 2 × 2 × 2 mm voxels, and finally smoothed with an 8 × 8 × 8-mm Full-Width Half-Maximum Gaussian kernel to decrease spatial noise and accommodate for individual differences in anatomy. To minimize the effects of signal artifacts due to motion, the first-level models included special “outlier” regressors that modeled volumes during which participants moved more than 1 mm in any direction. An analysis of both absolute and inter-scan motion parameters found no significant differences between motion parameters in the ASD and the control subjects (p > 0.1).

2.4 Basic General Linear Modeling (GLM) Analysis

A General Linear Model analysis of functional images was carried with a mass-univariate statistical parametric mapping approach, as implemented in the SPM8 software package (Friston et al., 2009). This approach consists of two steps; first-level fixed-effects models of individual participants’ effects, and second-level mixed-effect models of group-level effects. First-level models included four non-orthogonal regressors which were created by convolving four non-orthogonal boxcar function with a realistic hemodynamic response function. The four boxcar functions corresponded to the time courses of the four experimental conditions (obtained by crossing our two manipulated factors: Go vs. No-Go and Faces vs. Letters). The time boxcar timecourses corresponded to the onsets and durations of each corresponding block of trials for each participant. No other factors nor nuisance variables were included in the first-level model.

The mixed-effects group level models were applied on the estimated parameters of the individual, first-level models, and included each subject as the random factor. To correct for multiple-comparisons, a height threshold of p < 0.001 was combined with a cluster-level threshold of k = 131 contiguous voxels. The value of k was estimated by means of a monte-carlo simulation procedure, as implemented in AfNI's 3dClusterSim software, to ensure that the combined thresholds results in a family-wise error corrected threshold of p < 0.05 (Forman et al., 1995).

2.5 Dynamic Causal Modeling Procedure

Our hypothesis about the nature of basal ganglia dysfunction in ASD was tested by comparing the fit of two different models of how the basal ganglia interact with the putative cortical network that underpins the Go/No-Go task. These two models were designed to be identical, with the exception of one key feature: how the basal ganglia interact with the other regions to perform the experimental task. Specifically, both models were implemented with the DCM10 software package using a deterministic estimation procedure. However, in the bi-linear “direct” model, the basal ganglia connections to the rest of the network are designed to propagate signals, no differently than all of the other connections in the model. On the contrary, in the non-linear “modulatory” model, the basal ganglia function to modulate existing connections to prefrontal cortex, rather than merely propagating signals. The latter, but not the former, embodies the theory of the signal routing function of the basal ganglia described by Stocco and colleagues (2010).

2.5.1 Basic Model Architecture

To provide an adequate test of our hypothesis, it is important that the two models compared differ only and exclusively in terms of the nature of connections that originate from the basal ganglia, and otherwise share a common network architecture. The regions included in the common network (depicted in Figure 2A) were defined a-priori, based on our hypothesis about the role of the basal ganglia, as well as the existing neuroimaging literature about performance in the Go/No-Go task. Because our hypothesis concerns cortical coordination, it is important that such a network is as exhaustive as possible while remaining simple. As a result, our network comprises seven regions, divided into three classes of regions:

Fig 2.

The two models used for comparison and analysis. In the Direct model, projections from the basal ganglia directly target the two prefrontal regions, while in the Modulatory model, the basal ganglia modulate the effect of the incoming projections to the prefrontal VOIs.

1. Input and output nodes of the network

Three of the regions in the network are involved in task demand encoding and motor output. Their inclusion is necessary for providing a clear entry point through which the task stimuli affect the network activity, as well as a clear output point where behavioral responses are produced. The input region for the visual stimuli (letters and faces) was in the primary visual cortex in the occipital lobe (OCC). A second region, located in the medial frontal cortex (MFC) extending into the anterior cingulate cortex, was chosen as the input region for the higher-level, “cognitive” signal warning participants whether the following block was either a consistent “Go” or a mixed “Go/No-Go” block. There is ample evidence that the medial frontal cortex plays a crucial role in adapting neural resources to changing task demands, either on-line (Botvinick, Nystrom, Fissell, Carter, & Cohen, 1999; McDonald, Cohen, Stenger, & Carter, 2000; Barber & Carter, 2005) or off-line and anticipatorily (Brass & Von Cramon, 2002; Rushworth, Hadland, Paus, & Sipila, 2002; Sohn, Albert, Jung, Carter, & Anderson, 2007) and play an essential role in Go/No-Go tasks (Simmonds, Pekar, & Mostovsky, 2008). Although different functions have been suggested for the ACC (e.g., Carter & Van Veen, 2007) and the medial frontal regions (e.g., Supplementary and pre-Supplementary motor areas: Nachev, Kennard, & Husain, 2008), they were approximated as single VOI in our dataset to simplify the model, as their roles were not functionally isolated in the GLM contrasts explored. Finally, the output region was represented by the left motor cortex, which controls the right hand with which participants responded.

2. The executive nodes of the network

Two regions in the left and right dorsolateral prefrontal cortex (PFC) were included as they are consistently reported in studies employing Go/No-Go paradigms (Garavan, Ross, Murphy, Roche, & Stein, 2002; Kelly et al., 2004). PFC activation is widely involved in cognitive tasks with proposed functions include working memory and executive functioning (e.g., Barber & Carter, 2005; Brass, Derrfuss, Forstmann, & Cramon, 2005; Carpenter, Just, & Reichle, 2000), primarily associated with left PFC, and inhibitory control (e.g., Aron, Fletcher, Bullmore, Sahakian, & Robbins, 2003; Aron, Robbins, & Poldrack, 2004), primarily associated with right PFC. Both the left and the right PFC are targets of projections originating in the basal ganglia (Middleton & Strick, 2002). For these reasons, the bilateral PFC regions provide a critical nexus that connects our hypothesis about the role of basal ganglia in signal gating to the cortical network required to perform the Go/No-Go task.

3. The signal gating nodes of the network

To test the signal gating hypothesis, our common network included two regions in the left and right basal ganglia (more specifically, in the dorsal striatum), which functions as the input station for signals entering the basal ganglia circuit. These two regions were specifically included as part of the hypothesis tested in this study; however, these regions have also been previously reported in neuroimaging experiments using Go/No-Go paradigms (e.g., Kelly et al., 2004).

As noted above, the two models tested share not only the same regions of interest, but also share the same connectivity between the five cortical regions. This connectivity is also illustrated in Figure 2A. The two input regions (MFC and OCC) project to the two prefrontal regions, as well as the “output” motor region. In addition, the prefrontal regions project back to the occipital region (reflecting spatial and feature-based modulation of visual processing (Gandhi, Heeger, & Boynton, 1999; Saenz, Buracas, & Boynton, 2002) and to each other (reflecting inter-hemispheric coordination). All of the three prefrontal regions (right and left PFC and MFC) project to the two basal ganglia regions.

Finally, both models share the same drives, that is, the time courses of task-relevant conditions that provide the inputs to the models. Specifically, two drives modeling the onsets and durations of Face and Letter blocks provide the input to the occipital region, while a single drive modeling the onset and duration of the mixed Go/No-Go blocks provides the input to the medial frontal region. Task drives had no modulatory effect on cortico-cortical connectivity.

2.5.2 Alternative Models

Two alternative models were built on top of the basic architecture to test alternative theories of the general role of the BG within the cortical network that is recruited by this specific task. In the Direct model (Figure 2B), the two basal ganglia regions project back to the three prefrontal regions (left and right PFC and MFC). In the Modulatory model (Figure 2C), the basal ganglia regions modulate the strength of the incoming connections to the three prefrontal regions. Thus, the Direct model simply reflects the existent anatomical connectivity, while the Modulatory model reflects the hypothesis that the two BG regions perform signal routing functions (Stocco et al., 2010). Note that the Modulatory model does not need to explicitly model the direct projections from the BG to the frontal VOIs, as their functional role in routing signals is effectively captured by the modulatory projections.

2.5.3 Localization of individual VOIs

As it is common practice in DCM studies (Friston, Harrison, & Penny, 2003; Mechelli et al., 2003; Schlösser et al., 2008), these regions were approximated as spherical Volumes of Interest (VOIs), whose approximate position and radius is described in Table 2. The precise locations of the VOIs were determined on an individual-by-individual basis, using initial “seed” coordinates as a starting point. The seed coordinates were determined for each VOI by identifying corresponding peaks within the group-level functional activation maps corresponding to the condition of Go/No-Go blocks with Faces. These blocks were chosen because they elicit the most robust imaging effects, both in terms of the number of voxels above threshold and the individual voxel intensities in the statistical map, and thus guarantees the most robust inter-subject reliability when identifying the seed regions. Note that the seed coordinates only provided a group-level starting point for the identification of each individual VOI. The precise position of each VOI was then determined for each individual through an automated procedure. The procedure used the seed coordinates as a starting point, and identified the closest activation peak on the corresponding individual activation map for the mixed Go/NoGo blocks across both types of stimuli (Letters and Faces). This condition was chosen because the Go blocks were much easier and did not reliably recruit the prefrontal monitoring regions of interest. In addition, by including both Letters and Faces, the identification of each individual VOI did not show a bias towards one particular stimulus type.

Table 2.

Anatomical locations of the individual Volumes of Interest (VOIs) used in the models. Anatomical labels were obtained from the Automatic Anatomical Labeling atlas (Tzourio-Mazoyer et al., 2002).

VOI	Seed Coordinates (MNI)	Mean Individual Coordinates (MNI)	Radius (mm)	Seed Anatomical Label (Brodmann area)
Left BG	−20, 11, 9	−21.0, 10.2, 9.6	6	Left dorsal striatum (NA)
Left Motor	−48, −32, 53	−47.6, −31.1, 53.1	6	Postcentral gyrus (BA1, BA2)
Left Prefrontal	−45, 4, 35	−44.2, 4.6, 34.1	6	Left inferior frontal gyrus (BA6, BA44)
Medial Frontal	−6, 5, 55	−4.2, 4.6, 55.1	10	Anterior cingulate cortex (BA24)
Occipital	2, −78, 2	5.2, −80.2, 2.1	10	Primary visual cortex (BA17/18)
Right BG	20, 9, 10	19.6, 8.9, 10.1	6	Right dorsal striatum (NA)
Right Prefrontal	45, 9, 30	44.4, 10.0, 29.7	6	Right inferior frontal gyrus (BA44)

To make sure that the peaks were within the anatomical limits of the regions, peaks whose distance exceeded the VOI radius (6mm for the lateralized regions, and 10mm for the bilateral occipital and medial frontal region) were manually inspected, and, if they were outside of the desired VOI, the second largest peak location was determined. The distribution of each VOI centers across participants is depicted in Figure 3. A series of t-tests found no significant difference in coordinates between ASD participants and controls for any of the VOIs (p > .10).

Fig 3.

Position of the seven VOIs centroids across all the 33 subjects in MNI space. VOI centroids are depicted at infinite search depths, using the method described in Stocco (2014).

3. Results

3.1 Behavioral Results

As a first step, we performed an analysis of the accuracy and response times collected from the task. This analysis was performed to ascertain whether or not the two groups differed significantly in their behavior. Significant behavioral differences could constrain the interpretation of any group differences observed in DCM analysis, for instance if it appeared that one group was unable to do the task, or that the two groups were performing the task in qualitatively different ways. While the use of different strategies cannot be ruled out even in the presence of identical patterns of accuracy and response times (Sohn et al., 2004), it is less likely in the presence of comparable behaviors.

Task performance was assessed by measuring accuracy scores, sensitivity scores (that is, d’ scores: McMillan & Creelman, 2004), and reaction times. Accuracy was calculated by counting responses to “Go” letters and faces (“Hits”) as well as the lack of responses to “X”s and sad faces (“Correct rejections”) as correct responses. Correspondingly, a lack of responses to letters and happy faces (“Misses”) and responses to “X”s and sad faces (“False alarms”) were counted as incorrect responses. Responses that occurred within 100ms of the stimulus onset were counted as incorrect as well, as insufficient time had passed to process the information related to the current stimulus. d’ scores were calculated as the differences between the standardized hit rate (ratio of hits over hits and correct rejections), and the standardized false alarm rate (ratio of false alarms over false alarms and misses). Reaction times were calculated from correct responses (i.e., “Hits”) only. The mean accuracy levels and reaction times for each condition are summarized in Table 3.

Table 3.

Accuracies, d’ prime scores, and reaction times of ASD and control subjects across the different task conditions. Individual cells contain mean values +/− SD.

	ASD		Controls		Difference
Accuracies	Faces	Letters	Faces	Letters	Faces	Letters
Go	0.87 +/− 0.22	0.94 +/− 0.17	0.84 +/− 0.28	0.87 +/− 0.28	t(31) = 0.69, p = 0.49	t(31) = 1.31, p = 0.20
Go/NoGo	0.88 +/− 0.15	0.88 +/− 0.16	0.90 +/− 0.15	0.91 +/− 0.14	t(31) = −0.85, p = 0.40	t(31) = −0.73, p = 0.47
d’	Faces	Letters	Faces	Letters	Faces	Letters
Go	1.66 +/− 0.51	2.04 +/− 0.22	1.60 +/− 0.80	1.84 +/− 0.63	t(31) = −.24, p = .81	t(31) = −1.16, p = .26
Go/NoGo	2.84 +/− 0.99	2.95 +/− 0.96	3.18 +/− 0.67	3.26 +/− 0.76	t(31) = 1.17, p = .25	t(31) = 1.02, p = .32
Response Times (ms)	Faces	Letters	Faces	Letters	Faces	Letters
Go	577 +/− 92	478 +/− 190	574 +/− 73	459 +/− 118	t(31) = −.10, p = .92	t(31) = −.168, p = .87
Go/NoGo	595 +/− 91	461 +/− 70	607 +/− 56	497 +/− 84	t(31) = −0.27, p = .98	t(31) = .11, p = .91

Accuracies, d-prime scores, and reaction times were analyzed with separate mixed-effects ANOVAs, all of which were carried out with the statistical software R (R Core Team, 2008) and included the type of stimuli (Letters vs. Faces) and the type of block (Go vs. Go/NoGo) as within-subject factors, the group (ASD vs. controls) as a between-subject factor, and subjects as the random factor. In the case of accuracy, the analysis uncovered a significant main effect of the stimulus type [F(1,31) = 8.46, p < 0.01], and a significant interaction between the stimulus type and the block type [F(1,31) = 7.21, p = 0.01]. These effects were due to the accuracy being greater for letters than for faces in all participants, with the difference being smaller in the Go/NoGo blocks (see Table 3). In the case of d’ scores, the ANOVA uncovered a significant main effect of stimulus type [F(1, 31) = 9.43, p = 0.004) and of block type [F(1, 31) = 127.73, p < 0.0001]. These effects were due to higher d’ scores for letters than for faces, and for mixed Go/NoGo blocks than for Go blocks (see Table 3). In the case of response times, the analysis uncovered a significant main effect of the stimulus type [F(1,31) = 231.91, p < 0.001] and a significant main effect of the block type [F(1,31) = 23.87, p < 0.001]. Neither analysis uncovered any effect of Group [F(1,31) < 0.22, p > 0.64], nor any interaction between Group and any other factor. Thus, these results suggest that, compared to controls, that ASD subjects were not significantly impaired in executing the Go/NoGo task.

3.2 Group Comparison using General Linear Model Analysis

In addition to the analysis of behavioral data, we performed a canonical analysis of the neuroimaging data using the General Linear Model (GLM). The goal of the GLM analysis was to test whether the two groups exhibit any difference at the level of neural functioning which could be indicative of the use of different cognitive strategies (Sohn et al., 2004), as the main thrust of this paper is dynamic causal modeling analysis. Thus, no within group contrasts (e.g., No-Go – Go) are reported herein. No reliable differences in activation between ASD and control subjects were found at an uncorrected threshold of p < 0.001, for any of the relevant conditions (Go Faces, Go Letters, No-Go Faces, No-Go Letters) or their contrasts with one another. Given the lack of group-related difference, data from the two groups were combined to investigate the network of regions involved in the execution of Go/No-Go blocks compared to neutral fixation periods (Figure 4 and Table 3). The results of this analysis are comparable with those of previous studies (e.g., Menon et al., 2001; Simmons, Pekar, & Mostovsky, 2008; Watanabe et al., 2002), with the exception of the frontal regions being somewhat more posterior. It is also worth nothing that this network, identified from exploratory data analysis, is highly overlapping with the network of regions defined a-priori and used in the DCM models (Table 3; Figures 2 and 3), thus supporting the internal validity of our modeling choices.

Fig 4.

General GoNo network as measured by canonical GLM analysis of the imaging data. No significant differences were found between groups at p < 0.001. Significant effects, however, were found in response to the different task conditions. When compared to baseline Fixation blocks, mixed “Go/No-Go” blocks elicited significant brain activity in left, right, and middle prefrontal regions, as well as the parietal and occipital regions (see Table 4). Numbers in the figure refer to the corresponding indices of Table 4. All the effects were corrected at p < 0.01 with a Family-Wise Error correction procedure.

In summary, a canonical analysis of neuroimaging data revealed no significant difference between individuals with ASD and controls. As described in the next sections, this lack of group-related differences between ASD and control participants revealed by traditional GLM analyses hides a deeper functional difference between the two groups, in terms of network-level functioning.

3.3 Model Comparison in ASD and Controls

To test our first prediction, namely that a model in which the basal ganglia modulate patterns of cortical connectivity to the prefrontal cortex will fit the data better than a comparable model where the basal ganglia do not modulate such connectivity, we first compared the fits of the Direct (Figure 2B) and Modulatory (Figure 2C) models on the entire group, collapsing across neurotypicals and individuals with ASD, using a random-effects Bayesian model selection procedure. Both the models and the comparison procedures were performed as implemented in the DCM10 software. Consistent with the signal routing hypothesis, the Modulatory model was a significantly better fit for the data than was the Direct model (exceedance probability of 66.0% vs. 34.0% respectively). Interestingly, when the fit of the Direct and Modulatory models was computed separately in individuals with ASD and controls, the Modulatory model was clearly a better fit for controls (exceedance probability of 68.7% vs. 31.3% respectively), but only marginally better than the Direct model for the ASD group (exceedance probability of 54.5% vs. 45.5% respectively). Figure 5 depicts the results of the Bayesian model selection procedures for all participants collapsed, and then for each group considered separately.

Fig 5.

Exceedance probabilities for the group (left), and for individuals with ASD (middle) and controls (right) independently for Direct and Modulatory models as calculated with Baysean model selection procedures.

The fact that the Modulatory model fits the data better than the Direct model is consistent with the general hypothesis that basal ganglia coordinate cognitive activity by modifying the effects of incoming signals to prefrontal cortex, as proposed by Stocco et al. (2010). The fact that the comparative fits of the Direct model and Modulatory model for participants with ASD were much closer than in controls supports the hypothesis tested herein, that impaired basal ganglia gating in ASD contributes to the abnormal patterns of cortical connectivity previously reported.

3.4 Modulatory Effects of the Basal Ganglia in ASD and Controls

To test our second prediction, namely that the modulatory effects of the basal ganglia on pairs of cortical centers involving the prefrontal cortex will differ between individuals with ASD and controls, we performed an analysis of the estimated parameters of the modulatory connections across the two groups. Based on the theory outlined by Stocco et al., (2010), effective basal ganglia modulation in a control task should typically result in a reduction of the functional connectivity between pairs of regions. This reduction occurs because, by “gating” signals to the prefrontal cortex, the basal ganglia are biasing (or overriding) cortico-cortical connections, selectively altering the signals received by the prefrontal cortex. As explained above, this prediction implies that the mean value of the modulatory parameters should be negative for control subjects (indicating successful modulatory effects), but less negative for ASD subjects (indicating lack of efficient filtering).

To test this prediction, the individual parameters of the six modulatory connections from the basal ganglia (see Figure 2C) were extracted from the Modulatory model of each participant. A mixed-effects 2 × 2 × 3 ANOVA was then performed on the parameters, using group (ASD vs. controls), laterality (Left vs. Right Basal ganglia), and origin of the incoming signals to PFC (Occipital, Medial frontal, and Contro-lateral Prefrontal region) as fixed-effects factors, and subjects as the random factor. The ANOVA uncovered a trend towards a main effect of group [F (1,31) = 3.75, p = 0.06), and significant interaction between group and origin [F(2, 62) = 4.17, p = 0.02]. Follow up analyses showed that the interaction was due to the fact that the modulatory effect of the basal ganglia for the signals coming from the Occipital region [t(31) = 2.03, p = 0.025, one-tailed] to the prefrontal cortex differed significantly between the groups (see Figure 6). In particular, and consistent with our hypothesis, the modulatory effects of the basal ganglia were negative for controls (M = −0.08 Hz), but positive for individuals with ASD (M = 0.16 Hz). No other effect or interaction in the ANOVA model was significant (F < 0.58, p > 0.56), indicating that there were no significant differences, in either group, between the modulatory effects of the left and the right basal ganglia.

Fig 6.

Estimates of the modulatory connectivity from the basal ganglia to prefrontal cortex, in ASD and controls. Bars represent mean values +/− SE.

3.5 Behavioral Relevance of Modulatory Parameter Estimates

To examine the functional relevance of patterns of direct and modulatory connectivity in individuals with ASD and controls, two follow-up analyses were conducted. First, the critical model parameters (direct and modulatory connections including PFC) were correlated (using Pearson's r) with task performance in the ASD and control groups separately. In controls, a highly significant negative correlation between the effect of MFC on right dorsal striatum and reaction times across conditions was observed, such that stronger influences of MFC on the right dorsal striatum were associated with faster responses [r(17) = −0.614, p = .009]. This effect was not observed in the group with ASD [p = .51]. In the ASD group, d-prime scores were significantly explained by effective connectivity between right PFC and OCC [r(16) = 0.549, p = .028] and a nonsignificant trend toward the influence of right PFC on right dorsal striatum and d-prime [r(16) = .43, p < .10].

The second follow up analysis, conducted on individuals with ASD only, correlated the modulatory effects of basal ganglia with symptoms of ASD as measured by the ADOS and ADI-R. Interestingly, this analysis revealed a significant correlation between the modulatory effect of right dorsal striatum on the connection between OCC and right PFC and repetitive behaviors, as measured by the ADI-R. Specifically, the larger the modulatory effect of right basal ganglia, the less severe the symptoms observed were [r(16) = −0.528, p = .036]. Additionally the modulatory effects of left dorsal striatum on the connection between MFG and left PFC were positively correlated with social symptoms of ASD as measured by the ADOS [r(16) = .533, p = .034].

3.6 Differences in Effective Basal Ganglia-Cortical Connectivity between ASD and Controls

Because the difference in fit between the Direct and Modulatory model was much smaller in the ASD than in the control group, a follow-up analysis was performed on the basal ganglia cortical connections that were specific to the Direct model (Figure 2B). This analysis is complementary to the previous analysis of modulatory influences of the basal ganglia, and was performed to ensure that the differential fit of the Direct model across group was not due to underlying differences in the direct basal ganglia connectivity, in addition to the expected differences in modulatory effects. To test this hypothesis, the parameters of the four direct connections from the basal ganglia (see Figure 2B) were extracted from the Direct model of each participant. A mixed-effects 2 × 3 ANOVA was then performed on the parameters, using group (ASD vs. controls), and target region (Left PFC, Right PFC, and MFC) as fixed factors, and subjects as the random factor. The ANOVA uncovered a significant effect of the target region [F(2, 62) = 4.07, p = 0.02], implying a differential strength of basal ganglia projections to the three frontal regions. However, the analysis uncovered no main effect of group [F(1, 31) = 0, p = 1.00] and no interaction [F(2, 62) = 0.26, p = 0.77], implying that the strength of the striato-cortical connections was neither different nor differentially modulated in the two groups. These results are summarized in Figure 7.

Fig 7.

Estimates of the direct connectivity from the basal ganglia to prefrontal cortex, in ASD and controls. Bars represent mean values +/− SE. LPFC = Left PFC, RPFC = Right PFC.

3.7 Differences in Effective Cortico-Cortical Connectivity between ASD and Controls

An important implication of our framework is that the disrupted modulatory effects of the basal ganglia might contribute to the abnormal patterns of cortico-cortical connectivity that have been frequently observed in ASD (Just et al., 2012; Kana et al., 2011; Kleinhans et al., 2008; Schipul et al., 2011). In the context of our DCM analysis, this hypothesis can be empirically tested by comparing the estimated parameters of the cortico-cortical connections (i.e., the white arrows in Figure 2) across the Modulatory and Direct models. If the estimates of connectivity between cortical regions do not vary as function of the modulatory effects of the basal ganglia, we should find no significant differences between the estimates derived from the two models. If, as we suspect, differences in connectivity between cortical regions are at least partially explained by the underlying modulatory effect of the basal ganglia, then the values of the cortical connectivity parameters estimated should be significantly higher in the Modulatory model than in the Direct model. This is because the modulatory effects of the basal ganglia are primarily negative (i.e., they filter cortico-cortical signal transmission), and thus the “baseline” value of connectivity between two different regions needs to be comparatively higher than in the Direct model to compensate for the desynchronizing effect of the basal ganglia.

To test this hypothesis, we performed an analysis of all existing connections between cortical regions that are common to the two models. The analysis was performed using a 2 × 2 mixed-effects ANOVA, with model (Direct vs. Modulatory) as a within-subjects factor, group (ASD vs. controls) as between-subjects factor, and subjects as the random factor. The analysis uncovered a significant main effect of the model [F(1, 31) = 8.47, p = 0.007], a trend towards a significant main effect of group [F(1, 31) = 3.75, p = 0.06], and no interaction [F(1, 31) = 1.17, p = 0.29]. Effective estimates of cortico-cortical connectivity were marginally higher for controls (M = 0.13) than for ASD subjects (M = 0.08), consistent with existing theories highlighting underconnectivity in ASD (e.g., Just et al., 2007; 2012; Kleinhans et al., 2008). Additionally, across participants, estimates of cortico-cortical connectivity were higher for the Modulatory model (M = 0.13) than for the Direct model (M = 0.09), consistent with the theory tested herein. The results of the analysis are visually summarized in Figure 8. The detailed estimates of cortico-cortical connectivity parameters for each model and group are reported in the Supplementary Materials.

Fig 8.

Estimates of cortico-cortical effective connectivity in ASD and controls. Bars represent mean values +/− SE.

In summary, an analysis of cortico-cortical connectivity has identified both a significant difference between ASD and control subjects and a significant difference in the estimates of cortico-cortical connectivity when using different models of network connectivity. While the first finding is consistent with many results previously established in the literature (e.g. Just et al., 2012; Schipul et al., 2011), the second finding illustrates some potential limitations of traditional functional connectivity analysis.

4. Discussion

This paper reports the application of Dynamic Causal Modeling (DCM) to test the hypothesis that deficient basal ganglia functioning in individuals with ASD results in abnormal gating of signals to the prefrontal cortex. The results described herein are consistent with this hypothesis, demonstrating that the influence of the dorsal striatum on cortical networks differs reliably in individuals with ASD and in controls. We believe that these results provide a missing piece of the “puzzle” of ASD that is critical to understanding some of the seemingly disparate behavioral and neural symptoms previously described.

4.1 Basal Ganglia Dysfunction and Information Overload in ASD

DCM is a method that permits the testing of hypotheses, such as the one described herein, about effective and modulatory connectivity at the network level. By applying DCM analysis to neuroimaging data collected from a Go/NoGo task, we were able to identify significant differences between individuals with ASD and controls that support our hypothesis. Specifically, when the parameters that regulate the modulatory effects of the basal ganglia were compared between groups, significant group differences in modulatory effects were observed (Figure 6). Interestingly, our follow-up analyses showed that basal ganglia activation was associated with a reduction in effective connectivity between occipital and frontal regions in controls, consistent with our proposed signal gating function of the basal ganglia. In contrast, basal ganglia activation was associated with a large increase in effective connectivity between occipital and frontal regions in individuals with ASD. This data is consistent with the possibility that rather than filtering signals to the PFC, the basal ganglia in individuals with ASD is non-selectively copying signals, thus overloading the PFC with information (including some that is irrelevant to the task at hand). This never-before-proposed hypothesis is consistent with reports of overstimulation and a preference for routine and sameness in ASD. It is also consistent with the recently proposed “Intense World” hypothesis (Markram, Rinaldi, & Markram, 2007; Markram & Markram, 2010), which poses that individuals affected by ASD are less capable of “filtering out” irrelevant information.

Although it might appear counterintuitive, the fact that the modulatory effect of the dorsal striatum is inhibitory in neurotypicals but excitatory in ASD is compatible with the functional anatomy of the circuit. The basal ganglia contains two distinct pathways that exert opposite effects on the thalamic outputs to PFC (Albin, Young, & Penney, 1989; DeLong, 1990). The two pathways normally balance each other out, and their relative effect is controlled by the release of dopamine and the expression of dopamine receptors. In fact, genetic abnormalities in the expression of D1 receptors have been reported in ASD (Hettinger et al., 2008).

4.2 Adaptation versus Abnormality: Considering the Functional Implications of Basal Ganglia Differences in ASD and Controls

The idea that ASD is associated with atypical basal ganglia function is not new. It was proposed as early as 1978 by Damasio and Maurer, who noted the similarity between certain motor symptoms of ASD, and those caused by basal ganglia degeneration such as those observed in Parkinson's Disease. Our theory is rooted, however, in what the past 25 years of research have contributed to our understanding of the functioning of the basal ganglia, which extends dramatically past motor planning into the realm of language (Friederici, 2006; Prat & Just, 2011) and higher-level cognitive functioning (Monchi et al., 2006). Specifically, our findings provide a potential neurocognitive mechanism for a core deficit in information processing in ASD. Consistent with the Conditional Information Routing Model of basal ganglia functioning proposed by Stocco et al., (2010), we propose that in neurotypical controls, the basal ganglia are responsible for strategically blocking irrelevant inputs to the prefrontal cortex (as indicated by the negative values of modulatory parameters), permitting more streamlined processing of relevant stimulus features, while ignoring otherwise irrelevant sensory information.

It is worth noting that the primary difference between individuals with ASD and neurotypical participants was found in how the basal ganglia modulated the signals that reach PFC from the occipital region (Figure 6). The explanation of this effect proposed herein is that the occipital region is the primary input for this task using visual stimuli, and that the basal ganglia in individuals with ASD are failing to select and route only the relevant signals to the prefrontal cortex (which represent the control structure for the task). These findings are also consistent with the results suggesting abnormal attention orientation to sensory information in individuals with ASD (see Marco, Kinley, Hill, & Nagarajan, 2011 for review).

In contrast, the modulatory effects of the BG on signals coming from either the middle frontal region or the controlateral prefrontal regions were negligible (Figure 6). This lack of modulation might reflect two possible mechanisms. First, although both lateral and middle frontal regions function as both input and output stations to the basal ganglia, the basal ganglia might not perform any form of signal routing (Stocco et al., 2010) or gating (O'Reilly & Frank 2006) within frontal regions, but only between posterior regions of the brain and frontal regions. Alternatively, this type of frontal-to-frontal modulation might be supported by the basal ganglia, but not be required by the specific paradigm used herein. Future neuroimaging studies employing a larger sample of paradigms will perhaps clarify this issue.

One limitation to the interpretation of these results is that they are based on characterizations of the functional coordination of functional activity during a task in which no group differences were observed. Thus, we have little evidence to directly infer that the differences in basal ganglia signal modulation are disadvantageous for the Go/NoGo task. Instead, the only reliable predictor of performance in ASD was the effective connectivity between the canonical inhibitory control region in the right PFC and the occipital lobe. It is interesting to note, however, that the modulatory effects of the dorsal striatum on the executive control network were reliably correlated with the severity of restricted and repetitive behavior and social symptoms in ASD. Extensions of this analysis on new populations and replications on new data sets is necessary before strong claims can be made about the functional implications of the observed differences in basal ganglia signal routing observed between individuals with ASD and controls.

4.3. Possible Mechanisms of Basal Ganglia Dysfunction in ASD

This study was motivated by one specific theory of basal ganglia function, and the models supporting it (O'Reilly & Frank, 2006; Stocco et al., 2010). Because of this, it is important to situate these findings within the broader literature on basal ganglia functioning. Previous research on basal ganglia impairments in ASD has primarily focused on two aspects of the basal ganglia circuit: (1) its role in dopamine-modulated reward-based learning, and (2) the nature of its connections to cortical regions.

Abnormal reward signals in the basal ganglia have been proposed as the underlying etiology of attention deficit disorders, Tourette's syndrome, schizophrenia, and addiction (e.g., Maia & Frank, 2011, Redish, 2004). These models share a common architecture based on reinforcement learning, that is, a family of algorithms capable of learning how to select appropriate actions to maximize future rewards based on previous experiences (Sutton & Barto, 1998). The correspondence between these algorithms and basal ganglia function is well established (Schultz, Dayan & Montague, 1997; Niv, Duff, & Dayan, 2008). The theory tested herein is not antagonistic to this framework. In fact, the models by Stocco et al. (2010) and O'Reilly & Frank (2006), which best embody the idea of the basal ganglia as a “routing” or “gating” systems, are built on top of reinforcement learning algorithms, and advance them by proposing that basal ganglia “actions” consist of routing signals that are selected, amplified, and delivered to prefrontal cortex.

Individuals with ASD, however, do not seem to exhibit behaviors that can be reproduced with pure reinforcement learning models, even in the decision-making domain that is typically well explained by such models (Solomon et al., 2011). If the specific signal gating impairment observed in this study cannot be related to the dopaminergic reward system, it must then be localized within the cortico-striato-cortical connectivity. Indeed, recent research has paid increasing attention to the role of these connections in a variety of disorders, including ASD (Shepherd, 2013; Südhof, 2008). Perhaps most relevant to this study, children with ASD exhibit increased striato-cortical resting-state functional connectivity (Di Martino et al., 2011), a pattern that is consistent with the task-related increased effective connectivity reported herein. However, while some computational models have made attempts to address the functional significance of these distinctions (Ashby et al., 2005; Stocco et al., 2010), at present they offer only a limited and highly simplified rendition of the full complexity of these pathways (Shepherd, 2013). Thus, while the results presented provide further evidence that abnormal cortico-striatal connections are related to ASD, more sophisticated models will be needed to adequately describe the mechanisms underlying this relationship.

4.4 Implications of Abnormal Signal Gating for Functional Connectivity Analysis

In addition to examining the modulatory effects of the basal ganglia, we also performed a group comparison of cortico-cortical connectivity (Figure 8). Our results indicate the DCM-based estimates of effective connectivity between cortical regions are indeed significantly lower in ASD subjects than controls, thus confirming some previous findings (Just et al., 2012; Kleinhans et al., 2008; Schipul et al., 2011). However, our analysis also indicates that these differences crucially depend on the nature of the underlying network model that is used when measuring connectivity (Figure 8), with the Modulatory model yielding higher estimates than the Direct model. This finding has important implications for interpreting results from functional connectivity studies in ASD. In particular, our results imply that the differences in cortical connectivity between individuals with ASD and controls change according to tasks and experimental conditions. As shown in Figure 6, the basal ganglia can have opposite modulatory effects on the patterns of cortical connectivity in individuals with ASD and in controls. Because of this, differences in baseline patterns of connectivity might be obscured, or even reversed, when experimental conditions involve basal ganglia contributions.

4.5 Limitations of DCM

One limitation of the Dynamic Causal Modeling results reported herein is that the parameter estimates obtained from DCM are contingent upon the model tested. Such models are necessarily over simplified, and thus are not full replications of the networks involved in the task at hand. As discussed in the previous section (4.3), changes to the assumptions of any model can have widespread effects on the remaining parameters of the model. Thus, although we tested the relative fits of two models in the current design, it is possible that an untested model could fit the data better, and may yield different results with respect to the group differences between individuals with ASD and controls. Thus, future work is needed including converging evidence from other methods before we can say with confidence that the basal ganglia are functioning abnormally in ASD.

4.6 Summary

Despite these limitations, the current analyses provide compelling evidence for a critical link between abnormal basal ganglia functioning in ASD and impairments in cortical synchronization. The results suggest that ASD is characterized by abnormal effective connectivity at two different levels. Specifically, consistent with existing underconnectivity accounts, we found that individuals with ASD exhibited weaker patterns of cortico-cortical connectivity than did controls overall. However, the basal ganglia's modulation of cortical connectivity was larger and in the opposite direction in individuals with ASD than in controls, consistent with the predictions of our model. Interestingly, this abnormal modulation was correlated with some of the classic symptomology of ASD. Future research is needed to explore the extent to which these two patterns are independent or inter-related, with either one causing the other, or due to a shared neurobiological cause. In either case, by separating the effects of cortical and subcortical structures on cortical connectivity patterns, DCM provides important additional evidence for understanding the network-level processing deficits that characterize ASD.

Table 4.

Clusters of voxels significantly more active during the Go/No-Go blocks than during the Fixation blocks.

Index	Peak Coordinates (MNI)	Brodmann Areas (% Cluster Volume)	Anatomical Locations (% Cluster Volume)	Size (voxels)	Peak Intensity (t value)
1	2, −80, 4	BA18 (14.73%); BA19 (4.92%); BA37 (3.90%); BA17 (3.26%); BA30 (1.20%)	Left and Right Occipital lobe (76.48%); Left and Right Cerebellum (6.93%); Left and Right Posterior Temporal Lobe (11.60%)	6,443	7.98
2	46, 8, 32	BA6 (14.50%); BA9 (13.43%); BA8 (0.21%)	Right Precental Gyrus (42.22%); Right Middle Frontal Gyrus (29.00%); Right Inferior Frontal Gyrus, (27.55%); Right Superior Frontal Gyrus	469	4.63
3	−44, 2, 36	BA9 (25.86%); BA6 (18.10%)	Left Precental gyrus (97.41%); Left Inferior Frontal Gyrus (2.59%)	116	4.65
4	30, −52, 54	BA7 (31.25%); BA40 (1.39%)	Right Inferior Parietal Gyrus, (55.56%); Right Superior Parietal Gyrus (22.22%); Right Angular Gyrus (13.19%)	144	4.75
5	−4, 4, 58	BA6 (33.56%); BA8 (5.38%); BA32 (4.31%)	Left Supplementary motor area (54.99%); Right Supplementary motor area (41.88%)	1,022	6.0

Highlights.

• Abnormal signal routing from basal ganglia to PFC in ASD was found

• DCM model suggests that basal ganglia filter signals in controls and copy them in ASD

• Offers another mechanism for abnormal cortical synchronization in ASD

• Results help to explain the link between abnormal motor and cognitive behaviors