Adaptive Prediction for Social Contexts: The Cerebellar Contribution to Typical and Atypical Social Behaviors

Abstract

Social interactions involve processes ranging from face recognition to understanding others’ intentions. To guide appropriate behavior in a given context, social interactions rely on accurately predicting the outcomes of one’s actions and the thoughts of others. Because social interactions are inherently dynamic, these predictions must be continuously adapted. The neural correlates of social processing have largely focused on emotion, mentalizing, and reward networks, without integration of systems involved in prediction. The cerebellum forms predictive models to calibrate movements and adapt them to changing situations, and cerebellar predictive modeling is thought to extend to nonmotor behaviors. Primary cerebellar dysfunction can produce social deficits, and atypical cerebellar structure and function are reported in autism, which is characterized by social communication challenges and atypical predictive processing. We examine the evidence that cerebellar-mediated predictions and adaptation play important roles in social processes and argue that disruptions in these processes contribute to autism.

INTRODUCTION

Social interactions require participants to comprehend the goals and beliefs that produce actions, predict what will emerge from those actions, and ultimately respond appropriately to those actions in real time (Knoblich & Flach 2001, Sebanz & Knoblich 2009). In addition, as a specific action may carry different meaning depending on context, person, or the internal state or history of the person performing the action, participants must be able to rapidly integrate vast quantities of information for optimal social interaction (Lupfer et al. 1990, Todd et al. 2011). Indeed, social behavior consists of many complex interactions that fundamentally require predictions and are naturally reliant on continued adaptability and precise timing. Consider an interaction with an infant versus an adult peer. Interaction with an infant might require getting on the floor, using expressive facial expressions, and simplifying verbal or nonverbal content. The same predictions, behavioral strategy, and interpretation of actions applied to an interaction with one’s supervisor are unlikely to achieve the desired level of social success. Moreover, specific predictions are required for each situation (a conversation with a prospective business client versus a well-known family member). Even interactions with the same individual require both consideration of enduring traits and anticipation of the impact of momentary states (e.g., being tired, getting a grant, losing a close family member). Thus, accurate and specific predictions integrating prior experience and current context are a fundamental requirement for effective social interaction.

Beyond theoretical justification, experimental evidence supports critical roles for predictions in social interactions (Brown & Brune 2012). Visual observations of another’s gaze, movements, or spoken words also result in the prediction of intentions, motivation, and experience (Teufel et al. 2010). Upon observing such actions, one is able to draw parallels from one’s own experiences to predict the goals, feelings, and mental states of the other person (Keysers & Perrett 2004, Van Overwalle et al. 2014). One clear example of higher-level social prediction is mentalizing or theory of mind (TOM), which characterizes an individual’s ability to see the world from another’s point of view. Understanding the mental states of others requires the development of a model of that person’s internal thoughts in order to predict their future behavior (Baron-Cohen et al. 1985). To do this, one must also differentiate between self-generated actions and feelings and externally generated actions (Gentsch & Schutz-Bosbach 2011, Yomogida et al. 2010), which itself requires a system of prediction. Finally, these predictions must be accurately timed (Allman et al. 2011, Breska & Ivry 2016), as exemplified by how the fluid give-and-take of a conversation breaks down when an unpredictable delay is introduced via poor phone or internet signal.

Therefore, social predictions require the integration of large quantities of information (prior knowledge about an individual, situation, current context) into a model that enables a fluid, well-timed, appropriate social response. Critically, this model must be adaptable to changing contexts and new information. This ability to identify and adapt to error is a fundamental property of all forms of learning, including social learning (Brown & Brune 2012, Rushworth et al. 2009), and data indicate that there are shared mechanisms between nonsocial associative and social learning (Behrens et al. 2008, Jones et al. 2011). Both previously learned action-outcome associations and reward prediction errors that encode discrepancies between expected and true reward code for violations of social predictions and act as instructive signals to inform decision-making and update social predictive models (Behrens et al. 2008, Jones et al. 2011, Schultz & Dickinson 2000). Failure of predictive models can lead to breakdowns in social communication, from fundamental understanding of temporal order in social interactions to more complex processes such as mirroring (simulation of action leading to action comprehension) and mentalizing (ascribing thoughts and intentions to others).

While prediction and adaptation of predictions involve multiple brain regions, specific neural substrates meet the criteria for the representation of adaptive internal models that enable automatic, rapid, well-calibrated behaviors. One such neural system is the cerebellum. Traditionally considered a motor structure, the human cerebellum is now thought to support a range of nonmotor behaviors (Schmahmann et al. 2019, Sokolov et al. 2017). Here we consider the evidence that the cerebellum is a critical neural substrate of predictive processing, that the cerebellum supports a range of social behaviors, and that disrupted cerebellar predictive processing could explain core deficits in autism.

CEREBELLAR INTERNAL MODELS ENABLE PREDICTION AND ADAPTATION

What does the cerebellum do and how is this relevant to social behaviors? The cerebellar cortical circuit is stereotypically organized (Figure 1; see the sidebar titled Cerebellar Circuitry), and this basic circuit is conserved throughout the cerebellar cortex, although it is likely that differences in morphology, gene expression, and functional behavior in cerebellar neurons introduce precision and specificity into cerebellar circuits (Apps et al. 2018, Cerminara et al. 2015). It is hypothesized that a similar computation is performed by any given cerebellar region, with regional variations in the timing of the circuit dynamics and anatomical connectivity producing functional subregions (Ito 2008, Raymond & Medina 2018, Schmahmann et al. 2019, Sokolov 2018). To better understand the cerebellar computation contributing to social behavior, we will borrow from the more established sensorimotor literature, where cerebellar disruption leads to poorly timed, uncoordinated movement that is overly reliant upon sensory feedback (Therrien & Bastian 2015).

Figure 1.

Cerebellar predictive modeling of social information. Supratentorial networks involved in mirroring, mentalizing, and emotion (based on Kennedy & Adolphs 2012) send context and goal-related information to the cerebellum via mossy fiber inputs. This information routes through the cerebellar circuit to produce internal models of the behavior, which, once trained, provide predictions to the cerebral cortex. These models can be adapted based on reinforcement (blue check mark) or error (blue x) information via plasticity driven by climbing fiber inputs from the inferior olive. These inputs modify the firing of Purkinje cells, altering the signal projecting back to supratentorial regions via the CN. These predictions and adaptations are critical for proper social behaviors, and breakdown of these adaptive prediction models contributes to the core symptoms of autism. Abbreviations: ACC: anterior cingulate cortex; CN, cerebellar nuclei; IFG, inferior frontal gyrus; IPL, inferior parietal lobule; mPFC, medial prefrontal cortex; PCC, posterior cingulate cortex; PreC, precuneus; STS, superior temporal sulcus; ToM, theory of mind; TP, temporal pole; TPJ, temporoparietal junction; vmPFC, ventromedial prefrontal cortex.

To perform any complex behavior, a system must know the current context or state of the system. For movement, sensory feedback provides critical information about the body’s state, but sole reliance on sensory feedback is problematic. Due to temporal delays related to signal conduction and multiple points of synaptic transmission (Schmolesky et al. 1998), the brain is always receiving out-of-date sensory information. Like the disrupted flow of conversation over a delayed phone signal, control of movement based on delayed feedback results in unstable, poorly timed, oscillatory movements (Wolpert & Miall 1996). This system would also operate too slowly to provide the information needed for fast, well-timed behaviors, as would be needed for complex behavioral paradigms such as social interactions. In addition, sensory systems are not perfect sensors and are subject to errors due to noise (Wolpert & Flanagan 2001). Actions and behaviors are also conducted in changing sensory environments, and updating actions to account for such dynamic contexts requires learning. Therefore, our brains must have a mechanism both to compensate for feedback delays and for adaptation and learning.

Another feature of finely tuned behaviors/actions is that they are performed implicitly. Whether behaviors/actions are initially learned explicitly or implicitly, performance becomes more implicit as skills are mastered. When finely honed, behaviors become automatic, freeing up cognitive resources and making the behavior more immune to disturbances. Experience-driven learning and updating in motor learning is largely (although not completely) implicitly driven (Haith & Krakauer 2018, Mazzoni & Krakauer 2006), and cerebellar-dependent implicit adaptation and learning has been demonstrated in many sensorimotor and nonmotor paradigms (Clausi et al. 2017, Kim & Thompson 1997, Tseng et al. 2007, Wolpert et al. 1998). This implicit learning can generalize to different contexts unconsciously, thereby extending its utility (Hoffmann & Koch 1998).

CEREBELLAR CIRCUITRY.

Excitatory input to the circuit comes from both mossy and climbing fibers (CFs) (Figure 1). Mossy fiber input derives from many sources external to the cerebellum—from multiple precerebellar nuclei, directly from the dorsal spinocerebellar tract, and from the cerebellar output nuclei. These inputs synapse onto the most numerous cells in the brain, the granule cells. Granule cells then send axons that bifurcate into parallel fibers and synapse onto molecular-layer interneurons and Purkinje cells (PCs). This diversity allows both direct excitatory information and feedforward inhibition to be relayed to PCs. PCs, which are the sole output neurons of the cerebellar cortex, receive inputs from tens to hundreds of thousands of granule cells. In addition, PCs receive excitatory input from a single CF derived from the inferior olive. Each CF can make hundreds of synapses across a PC’s dendritic arbor (Palay & Chan-Palay 2012). PCs also receive inputs from a variety of local interneurons that play key roles in cerebellar function (Prestori et al. 2019, Whitney et al. 2009). PCs are GABAergic and send inhibitory axonal output to downstream nuclei, including the cerebellar output nuclei, vestibular nuclei, and parabrachial nuclei.

One way to overcome the obstacles of delayed feedback and need for adaptation is to predict the consequences of the behavioral command: in essence, to generate an estimation of the future state of the system ahead of the behavior. To produce accurate predictions, the body requires a system that can not only simulate the body’s behavior but also simulate the impact of that behavior on the environment while allowing for the ability to update any established models. The cerebellum has long been considered to generate predictive internal models to support implicit and coordinated behavior (Ito 1970, Popa & Ebner 2018, Wolpert et al. 1998). Internal models can be inverse, with the model providing an action to fit a desired outcome, thereby avoiding the need for another brain area to regenerate all the components of that action, and data exist for the cerebellum providing an internal inverse model (Yamamoto et al. 2007). The other feedforward model is an internal forward model. This model examines the causal relationship between a command and the outcomes of that command, predicting the outcome of the performed action given the current context. Extensive evidence exists for the cerebellum to generate and maintain internal forward models (Imamizu et al. 2000, Kawato 1999, Popa & Ebner 2018), with the mossy fiber input thought to bring information about the motor command and the present context into the cerebellar circuit. Deficits in motor function and motor learning observed in individuals with cerebellar dysfunction or with transient disruption of the cerebellum via transcranial magnetic stimulation (TMS) are consistent with this hypothesis (Miall et al. 2007, Therrien & Bastian 2015). Studies in nondisordered cerebellum show changes in cerebellar activity after motor learning, with specific changes in activity with motor errors and when state estimation was required for the task (Imamizu et al. 2000, Schlerf et al. 2012). The cerebellar internal model predictive system thus helps address the temporal complexities inherent in action performance.

Another requirement for the generation of accurate predictions is that internal models can be updated in response to prediction errors generated by noise, unpredictability, and changes in experience. The traditional view of the adaptation of cerebellar internal models revolves around a supervised learning paradigm, guided by error-based corrections (see the sidebar titled Cerebellar Circuits Build and Adapt Internal Models). In this model, predictions from feedforward models can be compared to sensory feedback to evaluate the difference or error between the predicted outcome and the action that has been accomplished (Ito 1970, Mazzoni & Krakauer 2006, Shadmehr et al. 2010). If no difference is detected between the actual and expected behavior, no error is registered, and the prediction does not require updating. A prediction error would elicit a learning signal to adjust the prediction such that the expectation mirrors reality.

CEREBELLAR CIRCUITS BUILD AND ADAPT INTERNAL MODELS.

Traditionally, the climbing fiber (CF) from the inferior olive is thought to carry the teaching signal for internal models. Upon the presence of error, the CF produces a large dendritic calcium spike—complex spike (CSp)—in Purkinje cells (PCs) (Marr 1969), which results in heterosynaptic plasticity in the form of long-term depression (LTD) at the parallel fiber–PC synapse. LTD results in reduced PC simple spiking to a given stimulus and, as PC output is inhibitory, produces timed disinhibition of downstream cerebellar nuclear neurons (Heiney et al. 2014). For example, in eye-blink conditioning, the participant learns to associate a neutral stimulus [conditioned stimulus (CS) such as a tone] with a subsequent corneal air puff that results in eyelid closure [conditioned response (CR)] (Medina et al. 2000). In this paradigm, with learning, upon hearing the CS, the CF can instruct a predictive eyelid closure, accurately timed to precede the onset of the air puff. Eye blink–conditioning studies point to a critical role for pausing in PC activity in the timing of the CR (Halverson et al. 2015), and the mechanisms driving the timing of this CR reside either within the PC itself (Johansson et al. 2014) or within the cerebellar nuclei (Brooks et al. 2015, Medina et al. 2000).

Although a supervised learning paradigm such as this can work well when there is a linked stimulus and response, it requires the climbing fibers (CFs) (see the sidebars titled Cerebellar Circuitry and Cerebellar Circuits Build and Adapt Internal Models) to be able to represent the entirety of what might be deemed an erroneous action in order to improve the behavioral prediction. Thus, in complex nonmotor behaviors where a stimulus and response are not clearly related or when the meaning of sensory information differs based on context, this supervised error-based paradigm may be insufficient to produce the needed adaptation (Hull 2020). Recent studies shed light on potential mechanisms for more complex adaptation within the cerebellar circuit. These studies identify CF-generated complex spikes that do not appear to represent error. One study has demonstrated that CFs appear to provide a predictive signal, more consistent with a temporal-difference reinforcement learning paradigm (Ohmae & Medina 2015). In this study, the CF teaching signal varied depending on the expectations within the behavioral paradigm. Moreover, contrary to an error-based model, complex spikes were generated in response to the conditioned stimulus upon learning, which mirrors studies in humans showing increased cerebellar activity with the conditioned stimulus (Ernst et al. 2019). Such a reinforcement signal could allow for learning to occur without prior knowledge of a stimulus–action relationship. Consistent with this, several recent studies have demonstrated a cerebellar role in reward-based reinforcement learning, with complex spikes generated with accurately predicted rewards and mirroring reward predictions that varied based on expectation (Heffley & Hull 2019, Heffley et al. 2018). These principles provide a potential mechanism to update learning rules for an already-learned paradigm, which is important to facilitate learning of complex motor and nonmotor behaviors (Hull 2020).

One additional key property of the internal model is the capacity to produce behaviors that become increasingly automatic and thus implicit. Interaction with internal models during explicit learning promotes the generation of prediction error signals as new predictions and/or actions are acquired. Once acquired, internal models are refined and corrected through the cerebellar-generated learning framework, and feedback loops promote the shift from explicit to implicit learning with increased automaticity (Clausi et al. 2017, Wolpert et al. 1998). These mechanisms then translate through cerebellar-neocortical connections, with downstream impacts of cerebellar output on the vast majority of the neocortex, including regions critical to both sensorimotor and nonmotor functions (Fujita et al. 2020, Kelly & Strick 2003). Although traditionally a top-down view dominated, recent studies also support bottom-up contributions, with a noted cerebellar influence on neocortical areas (Gao et al. 2018, Proville et al. 2014). For example, cerebellar output plays a critical role in the regulation of ramping activity in the anterior lateral motor cortex, which impacts motor predictions but not motor output specifically (Chabrol et al. 2019, Gao et al. 2018).

The cerebellum thus provides the circuit properties to generate, update, and automatize predictions using both supervised and reinforcement learning. These predictive and adaptive internal models enable precise performance, timing, and learning of complex behaviors, including those relevant to social interactions. Cerebellar circuits, acting at the implicit level and using social information conveyed through extensive cerebrocerebellar connections, could build and hone internal models of complex behaviors, allowing for automatic information processing and appropriate behaviors in a range of socially relevant processes ranging from perception of biological motion to understanding the mental states of others. For example, an internal model of the dynamics of a social situation (e.g., a birthday party) would enable one to understand and accurately predict both the order of events (e.g., singing happy birthday prior to cake) and their emotional consequences (e.g., child excited to eat cake). Consistent with these roles, the cerebellum is engaged during social tasks, it is functionally connected to regions critical for social processing, and cerebellar disruption is associated with altered social behaviors in clinical populations and animal models (Sokolov 2018, Van Overwalle et al. 2020).

THE CEREBELLUM AND SOCIAL BEHAVIOR

The cerebellar–neocortical circuits provide a substrate to generate internal forward models, through which the cerebellum generates predictions that are supplied to cortical regions for integration into complex behaviors. Inputs from diverse neocortical regions project to the pontine nuclei and are relayed to the cerebellum via mossy fibers (Eccles et al. 1967). Cerebellar output nuclei project to multiple regions of the thalamus and midbrain, which in turn innervate neocortical and subcortical regions to form segregated cortico-cerebellar-cortical loops (Fujita et al. 2020, Leiner et al. 1991). Based on regional differences in connectivity (Buckner et al. 2011), functional subregions of the cerebellum support overt sensorimotor control and a range of nonmotor processes (Stoodley & Schmahmann 2010).

This extensive cortico-cerebellar-cortical connectivity allows for cerebellar predictive modeling to be applied to many types of social information (Figure 1). Investigators have identified specific cerebellar circuits that are involved in more fundamental processing of socially relevant information (e.g., action imitation or implicit processing of affect) versus others that have roles in the cognitively demanding TOM or mentalizing (Siciliano & Clausi 2020, Van Overwalle et al. 2020). For example, connections between cerebellar lobules VI and VIII and sensorimotor networks could support action imitation, whereas the connections between lobules Crus I/II, the medial prefrontal cortex (mPFC), the posterior cingulate/precuneus, and the temporoparietal junction support mentalizing processes (Molenberghs et al. 2012, Van Overwalle et al. 2020). Implicit socioaffective processing has been linked to posterior vermis networks with the anterior cingulate, hypothalamus, amygdala, and hippocampus. These regional differences are evident in meta-analyses of functional activation patterns during social task paradigms (Van Overwalle et al. 2020) and multitask imaging in large cohorts (Guell et al. 2018). Further organization of social information is evident in the lateralization of activation patterns, with signals related to biological motion left-lateralized in cerebellar Crus I/II, consistent with the structural and functional connections between left Crus I and the right superior temporal sulcus (Sokolov 2018, Van Overwalle et al. 2020). TOM tasks engage right Crus I/II, although activation often extends bilaterally (Van Overwalle et al. 2014). These patterns suggest that cerebellar subregions—in particular, lobule VII and the posterior vermis—receive socially relevant information from neocortical circuits to incorporate into predictive models, similar to how information about motor commands and sensory context are routed to cerebellar subregions that form internal models of movement.

Consistent with these findings, disruptions of these cerebellar circuits lead to specific deficits in social behaviors in clinical populations. Neuropsychiatric symptoms following cerebellar damage in patients include autism-like behaviors and social skill deficits (Hoche et al. 2016, Schmahmann et al. 2007). Patients with early-life cerebellar damage show social withdrawal and increased internalizing behaviors as well as high rates of autism diagnoses, with cerebellar injury being the greatest nongenetic risk factor for autism (Stoodley & Limperopoulos 2016, Wang et al. 2014). Adult patients with isolated cerebellar damage performed worse than did age-matched controls on a measure of emotional attribution and reported social skill deficits on a neuropsychiatric scale (Hoche et al. 2016). Similarly, patients with cerebellar atrophy are impaired in both emotion attribution and mentalizing (Clausi et al. 2018). Neither executive function nor ataxia scores correlated with social task performance in these patients (Clausi et al. 2018), ruling out confounding effects of intellectual or motor impairment. While voxel-level lesion data linking regional cerebellar damage to specific deficits are scarce, lesions of the left posterolateral cerebellum impaired performance on biological motion tasks, consistent with left Crus I activation during biological motion processing (Sokolov et al. 2010). Temporary alterations in cerebellar function with neuromodulation support the clinical literature, with TMS targeting the left cerebellum impacting the categorization of emotional faces and negative body language (Ferrari et al. 2018, 2019) and transcranial direct current stimulation over the right lateral cerebellum affecting social learning (L.C. Rice, A.M. D’Mello, C.I.C. Thomas, S.E. Martin & C.J. Stoodley, unpublished data).

CEREBELLAR CONTRIBUTION TO AUTISM

It is possible that the cerebellar contribution to social processing is particularly important as social skills are being acquired (see Wang et al. 2014, Stoodley 2016). Recent evidence suggests that developmental disruption of cerebellar predictive modeling could underpin the core behavioral features in autism. Autism spectrum disorder (ASD) is a neurodevelopmental disorder diagnosed based on atypical social relationships and the presence of repetitive and inflexible behaviors (Am. Psychiatr. Assoc. 2013). Consistent with a critical cerebellar role in social interactions, the cerebellum is highly implicated in the pathogenesis of ASD. Postmortem studies of brains from individuals with ASD report reductions in Purkinje cell (PC) numbers of 35–95% and reduced PC size (Fatemi et al. 2002, Whitney et al. 2008). Cerebellar output nuclear neurons also show changes in size and number in autism (Bauman 1991). Cytoarchitecture in the cerebellum remains mostly intact, however, suggesting that PCs develop and then die instead of failing to develop altogether (Whitney et al. 2009). Imaging studies in individuals with autism and rodent models of ASD-linked genes also report cerebellar structural and functional differences when compared to neurotypical populations (D’Mello et al. 2015, Kelly et al. 2020, Stoodley et al. 2017).

Many autism-linked susceptibility genes are expressed in the cerebellum, some exclusively so (Menashe et al. 2013). Syndromic forms of ASD, including Phelan-McDermid syndrome, fragile X syndrome (FXS), tuberous sclerosis (TSC), and 15q11 duplication syndrome, are also frequently associated with cerebellar alterations or dysfunction (Mosconi et al. 2015). In FXS and TSC, changes in volume or lesions of the posterior vermis distinguish between individuals with and without autism in these disorders (Eluvathingal et al. 2006, Kaufmann et al. 2003). Moreover, in TSC, changes in cerebellar activity mirror studies in idiopathic ASD and correlate with the presence or absence of ASD in TSC patients (Asano et al. 2001, Ryu et al. 1999). Preclinical studies in models of ASD susceptibility genes and autism-associated neurodevelopmental disorders have further demonstrated critical roles for the cerebellum in the regulation of autism-relevant behaviors. Models of cerebellar loss of ASD susceptibility genes, including Tsc1/2, PTEN, and Shank2, show social impairments as well as repetitive and inflexible behaviors (Cupolillo et al. 2016, Peter et al. 2016, Reith et al. 2013, Tsai et al. 2012). In mutant models of ASD susceptibility genes, evidence for impaired PC output mirrors clinical studies of decreased functional activity in autism (Peter et al. 2016, Ryu et al. 1999, Tsai et al. 2012, Whitney et al. 2008). Further, many of these ASD-relevant models display impaired long-term depression (LTD), in addition to disrupted LTD-dependent behaviors (Baudouin et al. 2012, Kloth et al. 2015, Piochon et al. 2014). These findings are consistent with the deficits in adaptation observed in autism (Forgeot d’Arc et al. 2020, Sears et al. 1994).

Structural magnetic resonance imaging studies in young children find overall larger cerebellar volumes in ASD, while cerebellar volumes are found to be reduced in adults with autism (Courchesne et al. 2011, D’Mello & Stoodley 2015, Stoodley 2014). Diffusion tensor imaging (DTI) studies have identified altered white matter pathways connecting the cerebellum to the neocortex (Sivaswamy et al. 2010). Volumetric differences are evident in specific cerebellar lobules, many of which are linked with cognitive functions (Duerden et al. 2012, Stoodley 2014) and are part of the neural circuitry of the social cerebellum, including the posterior vermis and bilateral lobule VII (Crus I/II). Hypoplasia of the posterior vermis is consistently identified in ASD, with the degree of vermis volumetric loss correlating with the severity of repetitive behaviors and social impairment (Courchesne et al. 1988, D’Mello et al. 2015). Similarly, volumetric changes in lobule VII correlate with ASD symptom severity and social deficits (D’Mello et al. 2015) and are specific to autism as compared with other neurodevelopmental disorders (Stoodley 2014).

Functional imaging studies also show altered cerebellar activation in autism, again highlighting the posterior vermis and lobule VII (Van Overwalle et al. 2020). For example, atypical vermis activation when processing facial expressions has been reported in ASD individuals (Wang et al. 2007). During the Frith–Happé triangle animations task, decreased activation in left Crus I was reported in the autism group (Kana et al. 2015), along with reduced functional connectivity between Crus I bilaterally and medial regions of the mentalizing network (Van Overwalle et al. 2014). During a social judgment task, the neurotypical group engaged bilateral lobule VII (Crus II) while the autism group showed significantly less cerebellar activation (bilateral VI and VII) (Stanfield et al. 2017). Preclinical studies using chemogenetic and optogenetic approaches have confirmed the importance of the posterior vermis and right lobule VII (Crus I) in the regulation of ASD-related behaviors (Badura et al. 2018, Kelly et al. 2020, Stoodley et al. 2017), and these studies point to differential regulation of social versus repetitive behaviors by right Crus I and the posterior vermis, respectively (Kelly et al. 2020, Stoodley et al. 2017).

Despite the growing evidence for a primary role of the cerebellum in the pathophysiology of ASD, how the cerebellum regulates ASD-relevant behaviors is not well understood. Studies show that cerebellar disruption has significant effects on neocortical regions to which the cerebellum projects (Stoodley & Limperopoulos 2016, Volpe 2009). Consistent with altered cerebrocerebellar signaling, individuals with ASD have atypical structural and functional cerebellar-neocortical connectivity (D’Mello & Stoodley 2015, Kelly et al. 2020, Stoodley et al. 2017). Altered patterns of connectivity and paired activation have been identified during many behaviors, including motor skills, language, and emotion processing (Critchley et al. 2000, Mosconi et al. 2015, Verly et al. 2014). For example, in adults with autism, left Crus II showed reduced functional connectivity with the right dorsal temporoparietal junction, a region important in TOM (Igelstrom et al. 2017); in language-impaired individuals with autism, right Crus I did not show the typical functional connectivity with cortical language networks (Verly et al. 2014). These studies indicate a disconnection between the cerebellum and the neocortex in autism. In preclinical studies, circuits connecting right Crus I with the mPFC are disrupted across many mouse models of ASD susceptibility genes and have been shown to be critical for the regulation of social and repetitive/inflexible behaviors (Kelly et al. 2020).

Disruption of these cerebellar-neocortical loops further ties into one of the most prominent theories in autism research—that of a disrupted neocortical excitatory/inhibitory (E/I) balance in ASD, with a shifted E/I balance toward increased excitation that has been observed in both human studies and preclinical animal models (Chez et al. 2006, Nelson & Valakh 2015). Output from the cerebellar cortex is exclusively derived from PCs and is exclusively inhibitory; thus, cerebellar output is a critical source of inhibitory tone to the cortex via cerebellar-neocortical connections (Chabrol et al. 2019). Although still an area of active study (Person & Raman 2012), disrupted inhibitory output from PCs results in increased output from the cerebellar nuclei, as is observed in individuals with ASD (Asano et al. 2001). In preclinical autism models, the observed reductions in PC output (Peter et al. 2016, Tsai et al. 2012) result in increased neocortical activity (Kelly et al. 2020, Stoodley et al. 2017). Mechanistically, impaired PC function may provide a critical contribution to the cortical E/I imbalance observed in ASD.

PREDICTION AND ADAPTATION ARE DISRUPTED IN AUTISM

Consistent with disrupted cerebellar predictive modeling, the ability to make accurate predictions in both social and nonsocial domains is impacted in ASD (Balsters et al. 2016, Sinha et al. 2014). One clear example of this disruption is the difficulty with mentalizing and TOM in ASD (Baron-Cohen et al. 1985). Inherently, TOM tasks depend on generating predictions about another’s mental state based on current observations, and individuals with ASD have significant difficulties with these paradigms (Baron-Cohen et al. 1985, Sinha et al. 2014) as well as difficulties in making accurate social predictions and judgements (Chambon et al. 2017). Individuals with autism also show disrupted interoception and an impaired ability to make predictions about their own internal state (DuBois et al. 2016), while also displaying inaccurate temporal predictions (Allman et al. 2011). These impairments in predictive abilities correlate with the severity of ASD (Greene et al. 2019), and prediction errors are evident in social behaviors in autism (Balsters et al. 2016, Greene et al. 2019, Kinard et al. 2020).

Impaired predictions and adaptation are not just limited to social domains in autism. While atypical motor performance is not a diagnostic feature of autism, motor skill impairment is very common (Lloyd et al. 2011), and individuals with autism make atypical sensorimotor predictions (Kinard et al. 2020, Van de Cruys et al. 2014). The high rate of sensory hypersensitivities in ASD is also consistent with impaired predictions (Leekam et al. 2007, Pellicano & Burr 2012, Sinha et al. 2014), as habituation to noise or nonsalient sensory stimuli is directly correlated with the predictability of sensory stimuli (Herry et al. 2007). Impaired habituation results in the sensation of continuous sensory salience, which has been shown to be anxiogenic (Plutchik 1959). Sensory processing and integration are known challenges for individuals with ASD (Leekam et al. 2007), and the ritualistic, repetitive behaviors characteristic of ASD could be a mechanism to reduce the resultant anxiety (Eilam et al. 2011). Consistent with a shared principle underlying social deficits and such repetitive behaviors, atypical sensory responses correlate with the severity of social deficits in individuals with ASD (Hilton et al. 2010).

Predictive behavioral models are a precursor to the flexible adaptation of those models, and atypical adaptation is also evident in autism. In sensorimotor paradigms that rely on cerebellar function, individuals with ASD display impaired accuracy and timing, have difficulty adapting predictions during sensorimotor tasks, and depend on slower, often explicit, feedback mechanisms (Allman et al. 2011, Mosconi et al. 2015, Van de Cruys et al. 2014). Individuals with ASD also demonstrate difficulties with adaptation in social domains (d’Arc et al. 2020, Pellicano et al. 2007). This impaired adaptation contributes to behavioral rigidity and cognitive inflexibility, the other core features of autism. Individuals with autism hold tightly to specific behavior patterns and have difficulty changing plans, deviating from routines, and adapting to new environments. These behaviors are often described as an insistence on sameness and can be categorized as expressions of cognitive inflexibility. Cognitive flexibility, the rapid switching between functions or behaviors, by nature requires adaptation to new contexts. While this type of perseverative behavior is more frequently associated with the prefrontal cortex (Sakai 2008), perseveration is also observed in patients with cerebellar damage (Schmahmann et al. 2007).

In individuals with ASD, however, impairments in prediction are nuanced and related to task demands. With fixed contingencies and nonopen-ended paradigms, individuals with ASD can produce predictions and adapt those predictions accordingly. Challenges emerge in situations where the environment or conditions are open-ended and volatile (Van Eylen et al. 2011), especially when multiple sensory cues converge and when predicted outcome of these cues cannot be classified. Navigation of these situations requires both explicit and implicitly derived prediction and adaptation, and individuals with ASD struggle with implicitly learned paradigms and emotional perception (Klinger et al. 2007). Individuals with ASD tend to rely on explicit strategies to navigate what would otherwise be implicit tasks (Callenmark et al. 2014). As with motor control, reliance on such strategies in social and cognitive domains significantly impairs function and contributes to the social deficits and restricted behaviors seen in individuals with ASD (Callenmark et al. 2014, Klinger et al. 2007).

That said, explicit learning strategies can assist social skill learning (Mazzoni & Krakauer 2006, Taylor et al. 2014). Researchers have found that individuals with ASD and higher verbal age benefit more from explicit tutelage and are better able to successfully navigate what would be intuitive tasks for typically developing children (Happé 1995). This explicit learning approach also underlies one of the most common and effective ASD treatment options, applied behavioral analysis (Ivy & Schreck 2016). Instructors have been able to help individuals with ASD successfully complete TOM tasks through teaching of explicit rules that rely on effortful verbal strategies (Happé 1995). These strategies suggest that the hard work of explicit learning can overcome deficits in implicit modeling of information.

Thus, in autism, there is evidence for impaired prediction, adaptation, and implicit learning/processing, resulting in the need to learn and perform tasks explicitly and a reliance on slower, feedback-based adaptation. These findings have led to the hypothesis that disruptions in these critical neural functions could be the underlying pathological mechanism in autism (Pellicano & Burr 2012, Sinha et al. 2014, Van de Cruys et al. 2014). Given the evidence that these processes rely in part on cerebellar function, it is possible that developmental disruption of cerebellar predictive mechanisms in autism leads to a failure to acquire or automatize the information necessary for effective sensory habituation, efficient social interaction, and flexible behavior (Stoodley 2016).

FUTURE CONSIDERATIONS

Social behaviors are extraordinarily complex. They require the integration of diverse processes, including the ability to understand the consequences of one’s own actions while anticipating the state, needs, and meaning of others’ actions, all in real time. A core principle underlying this understanding and anticipation is the ability to generate rapid, well-timed, and accurate predictions and seamlessly adapt those predictions to the dynamic behavioral landscape that makes up a social interaction. Here, we have highlighted the key role that the cerebellum plays in generating predictions, adapting predictions to rapidly changing environments, and performing these functions implicitly. These principles are required not only for social behaviors but also for the cognitive flexibility and adaptability required for organisms to negotiate rapidly changing and complex environments, consistent with the cerebellar contribution to both motor and nonmotor behaviors. We also highlight the evidence that disrupted predictions in autism contribute to sensorimotor and social deficits, while an inability to adapt those predictions is consistent with the insistence on sameness and cognitive inflexibility that challenge individuals with autism. Thus, we propose that the growing evidence for cerebellar dysfunction in autism is not coincidental but rather provides a parsimonious explanation for both the circuit dysfunction and behavioral challenges observed in ASD. The fundamental cerebellar-mediated functions—implicit prediction and adaptation—provide a core mechanism that ties together the seemingly disparate behavioral categories impacted in ASD, from sensorimotor challenges to social impairment and inflexible behaviors.

However, many gaps in our understanding remain. Although an improved understanding of cerebellar topography as it relates to social behaviors is emerging, incorporating increased specificity within these frameworks will further improve our understanding of the cerebellar contribution to social interactions. It has long been appreciated that a more refined modular organization exists beyond the anatomic boundaries of cerebellar lobules. From functional microzones to molecular characteristics both within the cerebellum and in cerebellar output nuclei, a greater understanding of these additional layers of organization and how they map onto contributions to social behaviors should reveal greater insights into the cerebellar contribution to social behavior and autism (Apps et al. 2018, Cerminara et al. 2015, Fujita et al. 2020).

In addition, our understanding of the circuit networks that are involved in the cerebellar regulation of social behaviors is fairly primitive. Recent studies using functional imaging in humans and neural tracing methods in preclinical models have combined with molecular tools to improve our understanding of cerebellar-neocortical networks (Buckner et al. 2011, Fujita et al. 2020). Advances in the measurement of neural activity in both humans and preclinical models afford a unique opportunity to better understand the specific contributions of the cerebellum to social brain networks. Such studies have begun to investigate the neocortical circuit network contribution to social behaviors (Cho et al. 2020) and the cerebellar contribution to these networks (McAfee et al. 2019). Delving into the function of these circuits with increased specificity will provide important insights into how the cerebellum regulates these networks and the computations performed by these networks to support prediction, adaptation, and ultimately social behavior.

Future research should also examine how and what role the cerebellum is playing in generating and adapting predictions, especially in more open-ended and nonmotor behaviors such as social interactions. In particular, studies of the contribution of the cerebellum to social behaviors should use tasks and manipulations designed to test hypotheses regarding the specific mechanism by which the cerebellum contributes to such behaviors (Van Overwalle et al. 2020). Important insights have emerged that raise the possibility for reinforcement learning paradigms to be encoded in the cerebellar circuit (Heffley & Hull 2019, Heffley et al. 2018, Ohmae & Medina 2015). Combined with recent discoveries of connections between the cerebellum and canonical reward circuits (Carta et al. 2019), studies showing cerebellar responses to reward (Wagner et al. 2017), roles for cerebellar molecular-layer interneurons in associative learning (Ma et al. 2020), and intrinsic roles for monoamines and monoamine receptors, including dopamine, within the cerebellum itself (Locke et al. 2018, 2020) further point to critical roles for the cerebellum in the regulation of reward processing and potentially novel frameworks through which the cerebellum regulates learned behavior beyond error-based supervised learning.

In addition, developmental cerebellar abnormalities have been shown to have long-range impacts on connected neocortical regions, a cerebello-neocortical diaschisis that can impact social cortical networks (Stoodley & Limperopoulos 2016, Volpe 2009). The fact that early cerebellar damage has a significant long-term impact on the developing cerebral cortex raises intriguing questions about the role of cerebellar processing in the developing brain (Stoodley & Limperopoulos 2016). If one considers the conceptualization of the cerebellum as a neuronal learning machine (Raymond & Medina 2018) and the rapid learning that takes place in early life, intact cerebellar circuitry could be critical to the acquisition of the internal models that enable implicit prediction and rapid adaptation of behavior in new contexts. If this processing module is affected in early development, either through genetic risk or overt damage, there could be significant long-term impacts on the optimization of activity-dependent cortical circuits supporting a range of behaviors. Although studies are beginning to examine critical and sensitive periods for this cerebellar contribution (Badura et al. 2018, Tsai et al. 2018), the developmental contribution to these behaviors remains an important avenue of further study.

Lastly, understanding the greater impact of cerebellar dysfunction on autism might offer new therapeutic opportunities for the treatment of social disorders such as ASD or even schizophrenia. In schizophrenia models, the modulation of cerebellar function impacts timing and neocortical physiologic signatures of schizophrenia (Parker et al. 2017), while noninvasive modulation using TMS has been shown to benefit treatment-refractive patients through increasing cerebellar-cortical connectivity (Brady et al. 2019). Preclinical studies in a model of an ASD-linked gene have demonstrated that cerebellar modulation is sufficient to improve both social and repetitive behaviors (Kelly et al. 2020, Stoodley et al. 2017). These manipulations have been performed in adult models and appear to bypass sensitive periods as defined by the underlying molecular mechanism (Kelly et al. 2020, Stoodley et al. 2017, Tsai et al. 2018), raising the possibility that circuit modulation could provide benefit even later in life. Lastly, modulation in mice alters parallel circuit pathways to a noninvasive modulation approach in humans (Stoodley et al. 2017), raising the distinct possibility that cerebellar neuromodulation may offer a therapeutic target for ASD and other neuropsychiatric processes.

The cerebellum has long been known to contribute to learning and adaptive behavior in the sensorimotor domain. Here, we detail the critical cerebellar contribution to adaptive social predictions and the evidence that disrupted cerebellar processing could lead to the characteristic challenges in autism. Future research must focus on linking circuit, computational, systems, and behavioral levels of inquiry to develop a comprehensive model of the specific cerebellar contribution to complex social behaviors, with the ultimate goal of developing novel therapeutic options that are appropriately targeted and timed for optimal outcome.

SUMMARY POINTS.

1. Social behaviors rely on accurate predictions and the ability to update those predictions to ensure appropriate behavior in dynamic contexts.

2. Cerebellar dysfunction is implicated in autism, while developmental cerebellar dysfunction is associated with high rates of autism-relevant behaviors.

3. The cerebellum forms internal models to generate and adapt predictions. These models enable accurate, well-timed, and automatic sensorimotor and social behaviors.

4. Individuals with autism demonstrate difficulties with both generating and adapting predictions.

5. Cerebellar dysfunction offers a parsimonious explanation not only for social challenges but also for inflexible behaviors and sensory impairment in individuals with autism.