Large-scale brain networks in affective and social neuroscience: Towards an integrative functional architecture of the brain

Affective Neuroscience: The Nature of Emotion

In the field of affective neuroscience, no topic has received more attention than the brain basis of emotion. Until recently, scientists were largely convinced that anger, fear, sadness, happiness, and disgust, as emotional faculties, arise from separate, innate, culturally universal neural modules in the brain (for a review see [5,9]). In the typical brain imaging study of emotion, participants are asked to cultivate an emotional experience from viewing images or movies, by remembering prior experiences or perceiving an emotion in posed facial expressions (such as smiles, scowls, pouts, etc.), in non-linguistic vocalizations (such as sighs, shouts, etc.) or in body postures during brain imaging. Recently, two large-scale statistical summaries (i.e., meta-analyses) of human neuroimaging studies (covering studies published between 1993 and 2011) have demonstrated that anger, sadness, fear, disgust, and happiness cannot be localized to activity in specific topographical regions of the human brain using such tasks [9,10]¹. Brain regions such as the amgydala, anterior insula, pre- and sub-genual anterior cingulate cortex, and orbitofrontal cortex (once considered to be the brain locations of fear, disgust, sadness, and anger, respectively) demonstrate remarkably consistent increases in activity during a variety of emotional states indicating that these regions lack the specificity that is the hallmark of an emotion faculty perspective (see Figure 6 in [9]).²

Nonetheless, the belief that emotions can be localized somewhere in the brain is very strong (see [11, 12] for discussions), and efforts at topographical localization have given way to the hypothesis that emotions can be localized to specific brain networks (e.g., [13,14]). According to a emotion faculty approach, emotions are homologous in non-human mammals and universally inherited in humans, so the corresponding hypothesis would be that emotion networks should be anatomically constrained (and can therefore be inherited) and are homologous with networks that exist within the brains of other animals. The science of “intrinsic” network connectivity represents a prime candidate for identifying emotion networks according to these criteria (see Box 1). Recent research from our lab casts doubt on the specific emotion network hypothesis, however. Using a “seed and discovery” process for revealing intrinsic brain networks in task-independent (i.e., resting-state) fMRI BOLD data from 239 individuals, we demonstrated that distributed, anatomically-inspired networks for each emotion do not exist within the intrinsic architecture of the human brain [15]. Instead, the most reliable sources of activation for anger, sadness, fear, and disgust (as obtained from meta-analytic results) each produced a discovery map whose spatial overlap yielded the intrinsic network referred to as the “salience” network [16] or the ventral attention network [17,18] (see Box 2). A recent fMRI study from our lab confirms that nodes within the “salience” network (see Table 1) independently track feelings of pleasure and distress (medial orbitofrontal cortex) and feelings of arousal (amygdala) during both pleasant and unpleasant instances of happiness, sadness, and fear [19].

The nodes within the “salience” network not only lack specificity for discrete emotions, but they are also not specific to the broader domain of emotion per se (see Figure 2), consistent with criticisms of the limbic system concept. For example, regions within the amygdala are not only engaged during emotional states, but also show reliable increases in activity when people are asked to process novel material, even when such material is neutral and not affectively evocative (e.g., [20-22]). The amygdala also shows degeneration that is equivalent to hippocampal degeneration in the early stages of Alzheimer’s Disease [23]. Together, these findings suggest a role for the amygdala in the processes supporting encoding and retrieval, even for material that is “non-emotional” or is not particularly potent affectively. Other research shows that the amygdala is important for processes supporting perception and attention [24]. Furthermore, the amygdala has also long been implicated in social cognition, both within humans (e.g., [25]) and in comparative studies of non-human primates [26]. Recent research has demonstrated that humans with larger amygdalae characterized by stronger intrinsic connectivity to other brain regions such as the ventromedial prefrontal cortex, superior temporal sulcus, and fusiform gyrus are able to form and maintain larger and more complex social networks, both in face-to-face ([27,28]) and online groupings [29]. There is even some indication that amygdala volume increases due to the pressures associated with living in larger social groupings [30].

More broadly, nodes within the “salience” network are engaged during a variety of psychological domains, including during empathy (e.g., [31-33]), language and executive function tasks [34], and during attention allocation [17]. These nodes are thought to play some role in the brain’s “switching” or “reorienting” between “internal” and “external” events [17,35]. The spatial topography of the “salience” network is highly similar to an “aversion” network whose major hub is in the amygdala (see Figure 3 in [27]). It is possible that the “salience” network supports a variety of different psychological mechanisms (cf. [8]). Such multiple functions makes it all but impossible to infer the presence of a single process from a neural response in nodes of the network (called reverse inference; [36]). In addition, the lack of domain specificity in “salience” network engagement suggests the possibility that brain imaging is too spatially and temporally coarse to permit meaningful functional interpretation.

Yet an intriguing possibility is that the anatomic connections within the nodes of the “salience” network, along with their temporal dynamics, together perform a domain-general function supporting a range of tasks crossing the boundaries of emotion, social cognition, and non-social cognition. For example, anatomical connections indicate that nodes within the “salience” network are important for autonomic and hormonal regulation and representation [37-42], making it reasonable to hypothesize that in a given moment, the “salience” network functions to orient the brain’s processing capacity towards the most homeostatically- and metabolically-relevant information – it performs a body-based source of attention within the human brain. If a stimulus or event has homeostatic or metabolic implications (or the brain predicts these implications based on prior experience; [43]), then the “salience” network will be engaged, helping to allocate processing resources towards that stimulus.

The idea that the “salience” network might be domain-general does not preclude the possibility that it contains spatial gradients or subnetworks that are relatively more engaged for certain types of tasks. For example, the dorsal subnetwork of the “salience” network (most clearly involving connections between the dorsal anterior insula and dorsal ACC) seems to be especially important for selecting motor actions (i.e., response selection) necessary for executing goal directed behavior (e.g., [44]); the ventral subnetwork (involving connections between ventral anterior insula and pregenual ACC extending to the subgenual ACC) appears to be important for directing selection in the “visceromotor” (or autonomic nervous system) responses that support motor actions and that maintain homeostasis. This hypothesis is consistent with our recent finding that the strength of intrinsic connectivity within the ventral “salience” subnetwork is uniquely correlated with the intensity of affective experience, whereas connectivity within the more dorsal subsystem is uniquely correlated to motor control and changing mental sets [45].

Evidence from our meta-analyses on the brain basis of emotion more broadly supports a domain-general approach to distributed structure-function mappings. Some of the voxel clusters showing consistent increases in activation during anger, sadness, fear, disgust, and happiness appear within nodes of several of the brain’s intrinsic networks that are typically not considered to fall within the emotion domain (Table 1; also see [9]). For example, during emotional states, activity consistently increases within the ventromedial and dorsomedial prefrontal cortex and the posterior cingulate cortex/precuneus regions -- key nodes within the brain’s “mentalizing” network [46], also referred to as the “default mode” network [47]; this network routinely and robustly engaged when remembering personal events (autobiographical memory), when imagining the future (prospection), during moral cognition and reasoning, when accessing memory for word meanings (semantic memory), during scene construction and context-based object perception [47-50] and during instances of social affiliation (discussed in [27]). In addition, our emotion meta-analysis revealed consistent increases of activity in the lateral prefrontal cortex, which is a key node within “central executive” and “language” networks [16,17,51]. These nodes within the “mentalizing”, “executive”, and “language” networks show a consistent increase in activation during a range of different emotions and more general affective states [9,52]. A recent fMRI study from our lab found that several of these nodes produced consistent increases in activation during anger and fear experiences that occur when imagining a variety of social threat and physical danger scenarios [5], and show increased network cohesion when watching movies designed to evoke sadness [54]. Moreover, a recent neuroimaging study that directly compared patterns of network activity during emotions, thoughts, and bodily feelings confirmed that nodes within these networks are engaged across these different psychological domains [55].

Taken together, the search for the distinct brain correlates of different emotions, inspired by faculty psychology, has produced evidence of an entirely different functional architecture of the brain – one with broadly distributed functional networks that interact to produce a range of emotional states. In the next section, we discuss how these networks are not limited to constructing instances of anger, sadness, fear, disgust, and happiness.

Social Neuroscience: Person Perception and the Self

In the domain of social neuroscience, a key question concerns how one person’s mind creates the perception of another person as having intentions, beliefs, morals, traits, etc. Much of this research has been guided by prominent social cognition theories developed in the 1980s and 1990s, referring to these abilities as “person perception”, “mind perception”, “ordinary personology”, or “mental state attribution” [56-59]. Although early research in social neuroscience attempted to locate social cognitive processes in individual brain regions, such as the dorsomedial prefrontal cortex or the temporoparietal junction, the move towards a systems neuroscience approach quickly developed as dozens of studies showed that the neural correlates of person perception spanned regions that appeared to organize themselves into networks. Several meta-analyses [60-61] indicate that studies of person perception consistently report activation in regions comprising the “mentalizing” network mentioned above, and in a “mirroring” network, (which shows an increase in activation when a person is detecting the goal-directed, voluntary movement of body parts, and includes the premotor cortex, superior temporal sulcus, and inferior parietal lobule; [62-63]) (see note in Table 1). .

Currently, an overarching goal in social neuroscience has been to understand the distinct functional contributions of the “mentalizing” and “mirroring” networks during instances of person perception. For example, studies suggest that the “mirroring” network appears to be more associated with representing instrumental motor movements (e.g. facial movements and bodily gestures) while making inferences about their mental meaning, whereas the “mentalizing” network is more strongly associated with representing context information that allows mental state inference to occur [64]. The “mirroring” network also appears to be relatively more engaged when physical movements are perceived as discrete behaviors (called “action identification”; e.g., someone shredding paper), whereas the “mentalizing” network is relatively more engaged when making inferences about the goals of those actions (e.g., “destroying evidence”) [65, 66]. This research has inspired hypotheses for how the two networks communicate and interact with each other. One hypothesis is that the networks act in parallel, at times cooperating and other times competing during person perception [64]. Another hypothesis is that the two are hierarchically related, with the ability to form abstract mental state attributions being dependent on the ability to observe and represent actions in more concrete terms first [65, 66]. Interestingly, observed actions appear to engage the “mirroring” system only after their meaning has been inferred (cf. [67]), perhaps within the neural context of an active “mentalizing” network.

Network approaches to understanding person perception have, in a sense, revived domain-general insights in social perception from over fifty years ago. For example, a recent study of emotion perception [68] finds that perceiving physical actions as emotional expressions involves an interaction between nodes of both networks, placing emotion perception within the more general domain of person perception (as suggested by [69]). This social perception framework has been recently extended to understand how people construct emotional experiences by making mental state inferences about bodily changes [11, 70]. The hypothesis is that people are engaging in a form of mental state inference when they simultaneously categorize their (interoceptive) bodily sensations using conceptual knowledge about emotion.³ Taking this perspective, a recent study has shown that making mental state attributions about one’s own affective states (e.g. “I feel good”) indeed involves nodes within the “mentalizing” network, the “executive control” network, and the “salience” network [71].

A sense of personal identity (referred to as “the self”) is another topic of interest in social neuroscience [e.g. 72, 73]. Early theorizing in social cognition conceived of the self as resulting from a powerful, domain-general memory system [74], and foreshadowed recent meta-analytic discoveries that report consistent activation in the “mentalizing” network when people create mental state inferences and trait judgments about the self (i.e. judging the self as an object, as when self-reflecting [60, 61, 75, 76], also see [77, 78]), including when processing one’s own name or seeing one’s own face (termed “self-specific activity”; [79]). Moreover, autobiographical memory (which routinely engages the “mentalizing” network; [50, 79]) has been discussed as a precondition for mental states that create a sense of self [80].

Just as with the “salience” network, a variety of functions have been attributed to the “mentalizing” network (see [47, 50]), but it is possible that the network is performing one basic function across social, affective, and cognitive domains. We hypothesize that the nodes of the mentalizing network interact to create an inferential, conceptual system [2] that creates situated conceptualizations: these are the multimodal simulations that are strongly situated in a particular background context and that make meaning of sensory input and support specific courses of situated action [11,70,81]. Our hypothesis is that the “mentalizing” network associatively recombines stored information (both abstract and embodied) to create not only cognitions but also emotions and perceptions of people and events in the world. Such a proposal is consistent with the view that every waking moment of life is an instance of the “remembered present” [82]. It is also consistent with the hypothesis that the “mentalizing” network constructs mental models or simulations that facilitate future behavior [50]. Remembering, thinking about the future, taking another person’s perspective, understanding the cause and consequences of your own bodily feelings, all depend on the ability to draw on stored experiences to create a meaningful mental moment in the present. Sometimes, the present moment is created with minimal input from the senses (as in mindwandering or prospection); studies of internally-driven or endogenous mentation (i.e. mentation that is not driven by a concrete task set forth by the experimenter) have robustly demonstrated the engagement of the “mentalizing” network. At other times, the brain’s task is to make meaning of the current sensory array (to understand what sensations stand for in the world); brain imaging evidence indicates that this network is also engaged when the brain’s task is to create predictions about what sensory input refers to in the moment (e.g., during object perception, person perception, or emotion; e.g., [9,48,83]).

By explicitly comparing “mentalizing” network activations during judgments of self vs. others [61, also see 73], it is possible to hypothesize that a dorsal “mentalizing” subnetwork is relatively more engaged in creating situated conceptualizations when abstract, “third-person” information is required to make a judgment (such as knowledge about mental state categories, scene construction, etc.), whereas a more ventral “mentalizing” subnetwork appears to be relatively more engaged when more embodied, “first-person” information is required to make a judgment (e.g., bodily sensations or feelings related to homeostasis), referred to as the “dorsomedial” and “medial temporal lobe” subnetworks, respectively [47]. In support of this view, there is a spatial gradient within the medial prefrontal cortex (a key node of the “mentalizing” network), with more ventral activations in evidence during self-related judgments, during autonomic and endocrine regulation, and during the processing of pain and rewarding stimuli, whereas more dorsal activations are relatively more in evidence during other-related judgments, as well as during memory and mind-wandering which often involves viewing and evaluating the self as an object [61, 84]. Similarly, nodes within the dorsal subnetwork of the “mentalizing” network are more engaged when perceivers create abstract representations of other people (sometimes termed “meta-cognition”), whereas the more ventral nodes are more engaged when people use their own internal state as a way of inferring what someone else is thinking or feeling (termed “self-projection”; [85]). Interestingly, we predict that during mental states that require first-person information, we might see more inter-network connectivity between the ventral “mentalizing” subnetwork and the “salience” network. Consistent with this hypothesis, two meta-analyses report increase in activation for several nodes within the “salience” network, including the insula and the anterior cingulate cortex, particularly for self-related judgments [61, 86].

A Constructionist Functional Architecture of the Brain

Figure 2 summarizes the transition in human neuroscience research away from the search for domain-specific neural modules towards the discovery of large-scale, domain general networks that are distributed in both their structure and function. There are three take-away hypotheses from this figure. First, a psychological faculty, such as fear, or the ability to perceive the traits or mental states of another person (or other faculties we have not discussed here, such as working memory [86]), is not a process that can be probed in an unadulterated way by a single task; each faculty represents a category of phenomena – a collection of instances - that are each constructed via the interaction of more basic, domain general processes, which themselves map to networks that emerge from neural integration across time and space within the brain. Second, the experimental tasks that are currently in use within affective, social, and cognitive neuroscience do not themselves reflect individual processes; the responses to those tasks are constructed via the interaction of more basic, domain-general processes and their corresponding distributed networks. Third, scientific progress will be speeded not by trying to localize psychological faculties to topographical brain regions or even to individual networks, but instead by modeling responses to experimental tasks as high dimensional brain states (reflecting the engagement of domain general networks, their internal operations, and their interactions).⁴

From this perspective, then, the distinction between social, affect, and cognitive neuroscience is artificial. There is no “affective” brain, “social” brain, or “cognitive” brain. Each human has one brain whose functional properties can be understood differently for different time scales and levels of organization. At the level of human brain imaging, we are suggesting the need for a single systems neuroscience framework that spans psychological domains. In such a framework, the brain contains a set of intrinsic networks that can be understood as performing domain-general operations; these operations serve as the functional architecture for how mental events and behaviors are constructed. Each process can be observed as a set of “functional motifs” arising from the structural motif that undergirds each network (for a discussion of motifs, see [88]). A similar approach has recently been proposed for improving the characterization and treatment of psychopathology and neurodegenerative diseases [89].

By shifting the empirical emphasis from the search for mental faculties as unified neurobiological categories towards developing a more componential, constructionist functional architecture of the human brain, the overlap in empirical findings across psychological domains is not a problem for reverse inference, but becomes the engine that drives a more valid approach to reverse inference. Debates about whether emotions are a special case of social cognition, or whether person-perceptions are a special case of memory recede to be replaced an approach that attempts to build a domain-general understanding of how the brain creates the mind.