Establishing a Role of the Semantic Control Network in Social Cognitive Processing: A Meta-analysis of Functional Neuroimaging Studies

Most leading models of socio-cognitive processing devote little discussion to the nature and neuroanatomical correlates of cognitive control mechanisms. Recently, it has been proposed that the regulation of social behaviours could rely on brain regions specialised in the controlled retrieval of semantic information, namely the anterior inferior frontal gyrus (IFG) and posterior middle temporal gyrus. Accordingly, we set out to investigate whether the neural activation commonly found in social functional neuroimaging studies extends to these ‘semantic control’ regions. We conducted five coordinate-based meta-analyses to combine results of over 500 fMRI/PET experiments and identified the brain regions consistently involved in semantic control, as well as four social abilities: theory of mind, trait inference, empathy and moral reasoning. This allowed an unprecedented parallel review of the neural networks associated with each of these cognitive domains. The results confirmed that the anterior left IFG region involved in semantic control is reliably engaged in all four social domains. This suggests that social cognition could be partly regulated by the neurocognitive system underpinning semantic control.


Introduction 1
The ability to comprehend and respond appropriately to the behaviour of others is 2 essential for humans to survive and thrive. A major challenge for the cognitive sciences, 3 therefore, is to characterise how we understand others and coordinate our behaviour to 4 > low semantic control, or comparisons between a task requiring semantic control and an 4 0 0 equally demanding executive decision in a non-semantic domain. We excluded studies with a 4 0 1 focus upon priming without an explicit semantic judgment (e.g., primed lexical decision), 4 0 2 bilingualism, episodic memory, or sleep consolidation. March 2020, and reference-tracing, and then applied our inclusion criteria to both newly 4 0 7 identified studies and those analysed by Noonan et al. (2013). This produced a yield of 96

Data Analysis 4 1 2
We performed coordinate-based meta-analyses using the revised activation likelihood 4 1 3 estimation (ALE) algorithm (Eickhoff et al., , 2009Turkeltaub et al., 2012) 4 1 4 implemented in the GingerALE 3.0.2 software (http://brainmap.org/ale). We used the 4 1 5 GingerALE software to conduct two types of analysis. The first were independent dataset 4 1 6 analyses, which were used to identify areas of consistent activation across particular sets of 4 1 7 experiments. These analyses were performed only on the experiment samples with a 4 1 8 recommended minimum of 17 experiments in order to have sufficient power to detect 4 1 9 consistent effects and circumvent the possibility of results being driven by single experiments 4 2 0 . The ALE meta-analytic method treats reported activation coordinates 4 2 1 as the centre points of three-dimensional Gaussian probability distributions which take into 4 2 2 account the sample size (Eickhoff et al., 2009). First, the spatial probability distributions of 4 2 3 all coordinates reported were aggregated, creating a voxel-wise modelled activation (MA) 4 2 4 map for each experiment. Then, the voxel-wise union across the MA maps of all included 4 2 5 experiments was computed, resulting in an ALE map that quantifies the convergence of 4 2 6 results across experiments (Turkeltaub et al., 2012).The version of GingerALE used in the 4 2 7 present study tests for above-chance convergence between experiments (Eickhoff et al.,4 2 8 2012) thus permitting random-effects inferences. 4 2 9 Following the recommendations of Eickhoff et al. (2016), for the main statistical 4 3 0 inferences, the individual ALE maps were thresholded using cluster-level family-wise error 4 3 1 (FWE) correction of p < 0.05 with a prior cluster-forming threshold of p < 0.001. Cluster-4 3 2 level FWE correction has been shown to offer the best compromise between sensitivity to 4 3 3 detect true convergence and spatial specificity . This was 4 3 4 complemented by a highly conservative voxel-level FWE correction of p < 0.05 which, 4 3 5 despite the decreased sensitivity to true effects, allows the attribution of significance to each 4 3 6 20 voxel above the threshold, offering increased spatial specificity . The 4 3 7 FWE-corrected cluster-level and voxel-height thresholds were estimated using a permutation 4 3 8 approach with 5000 repetitions . None of the meta-analyses that we 4 3 9 updated had used the recommended cluster-level FWE or the FWE height-based correction 4 4 0 methods.

1
The second set of analyses, conjunction and contrast analyses, were also performed in 4 4 2 GingerALE and were aimed at identifying similarities and differences in neural activation 4 4 3 between the different sets of studies. The conjunction images were generated using the 4 4 4 voxel wise minimum value (Nichols et al., 2005)  thresholding but did not survive extent-based thresholding. These results are largely 4 7 4 consistent with those of Molenberghs et al. (2016), with the difference being that they did not 4 7 5 find activation in SMA, left fusiform gyrus or cerebellum. In order to address concerns 4 7 6 regarding the validity of some other popular ToM tasks (Heyes, 2014;Quesque and Rossetti, 4 7 7 2020), we conducted a separate supplementary meta-analysis that was limited to the subset of 4 7 8 ToM experiments that employed false belief tasks (see Section 3.1 of SI1, The ALE meta-analysis of all 164 empathy experiments revealed 16 clusters of 4 9 1 convergent activation ( Figure S6a; Table S1.3.1), including in the bilateral IFG (extending survived both the height-based and extent-based thresholding, except for the anterior dmPFC, 4 9 6 right pMTG and brainstem clusters, which survived extent-based thresholding only. Two 4 9 7 clusters, one in the right cerebellum and one in the right hippocampus survived height-based 4 9 8 thresholding but did not survive cluster extent-based thresholding. These areas were also 4 9 9 implicated by Timmers et al. (2018). In contrast, however, we did not find convergent 5 0 0 activation in the left posterior fusiform gyrus, left SMG (although we found a cluster slightly 5 0 1 more posterior and inferior), left anterior ITG, right TP, precuneus, middle cingulate gyrus, 5 0 2 right superior parietal lobule, and right amygdala. 5 0 3 The separate ALE maps for empathy for pain and empathy for affective states are 5 0 4 displayed in Figure 1c and d. A conjunction analysis found activation common to empathy 5 0 5 for pain (Table S1.3.2) and empathy for affective states (Table S1.3.3) in the bilateral insula 5 0 6 (extending to the IFG), SMA, right precentral gyrus, right ITG, bilateral occipital cortex and 5 0 7 the brainstem ( Figure S6b; Table S1.3.4). Formal contrasts revealed that empathy for pain 5 0 8 and empathy for emotions also engage highly distinct brain areas ( Figure S6b; Table S1.3.4). 5 0 9 Clusters with increased convergence for empathy for pain were found in left IFG (pars 5 1 0 triangularis), right MFG, bilateral insula, middle cingulate gyrus, bilateral SMG, right IPL 5 1 1 23 and bilateral pITG. In contrast, increased convergence in empathy for affective states was 5 1 2 revealed in left IFG (pars orbitalis), PCC, left pMTG, right temporal pole and left anterior 5 1 3 MTG. Given these significant differences in their underlying neural networks, empathy for 5 1 4 pain and empathy for emotions were considered separately for all subsequent analyses. those obtained by Eres (2018), with the difference that we did not find convergent activation 5 2 3 in the left amygdala and right AG, and found additional clusters of convergent activation in 5 2 4 left MFG, bilateral anterior MTG, and right pMTG. activation resulting from independent meta-analyses of ToM studies (N=136), trait 5 2 7 inference (N= 40), empathy for pain (N=80) and emotions (N=75) and moral 5 2 8 reasoning (N=69). The ALE maps were thresholded using an FWE corrected cluster-5 2 9 extent at p < .05 with a cluster-forming threshold of p < .001 (red) and an FWE 5 3 0 corrected voxel-height threshold of p < .05 (yellow). The lateral views, which show 5 3 1 projections on the cortical surface, are accompanied by brain slices at the sagittal 5 3 2 25 midline and also coplanar with the peak of the left IFG cluster observed across all 5 3 3 social domains (X = -39; Table S1.5). 5 3 4 5 3 5 3.1.5. A common network for multiple sub-domains of social cognition 5 3 6 To identify brain areas consistently activated across multiple sub-domains of social 5 3 7 cognition, we performed an overlay conjunction analysis of the cluster-extent FWE-corrected 5 3 8 ALE maps associated with ToM, trait inference, empathy (for pain and/or emotions) and 5 3 9 moral reasoning (see Figure 2a, Because the conservative thresholding used in this analysis could have excluded smaller 5 4 6 clusters that nonetheless overlap across the sub-domains, we repeated the conjunction using 5 4 7 ALE maps treated with a more liberal statistical threshold of p<.001 uncorrected. This 5 4 8 revealed additional overlapping activation for all four social domains in the right IFG (pars 5 4 9 orbitalis) and bilateral ATL ( Figure S7). These brain areas have been implicated in a variety 5 5 0 of social-cognitive abilities by multiple previous meta-analyses (Alcalá-López et al., 2018).

1
The extent to which brain regions engaged in social cognition overlap with those engaged 5 5 2 in general semantic cognition (including both representation and control processes) is 5 5 3 illustrated in Figure 2b. Figure 2c shows that the brain regions engaged in social cognition 5 5 4 are largely non-overlapping with the network engaged by domain-general executive 5 5 5 processes (i.e., the MDN). The contrast maps were thresholded with a cluster-forming threshold of p < .001 and a 6 4 5 minimum cluster size of 200 mm 3 . The lateral views, which show projections on the 6 4 6 cortical surface, are accompanied by brain slices at the sagittal midline and also coplanar 6 4 7 Contributions to the left IFG cluster depended on the difficulty category (p < .001) 7 3 0 and pairwise comparisons indicated that the C>E experiments contributed 7 3 1 significantly less peaks compared to E>C (p = .001) and E>C (p = .046) experiments. 7 3 2 Contributions to the left anterior MTG cluster also depended on the difficulty 7 3 3 category (p = .043) and pairwise comparisons indicated that the C>E experiments 7 3 4 contributed fewer peaks compared to E>C, but this effect did not survive correction 7 3 5 for multiple comparisons (p = .06). These results suggest that the left IFG is 7 3 6 particularly sensitive to cognitively-challenging ToM processing. 7 3 7 The ALE maps were thresholded using an FWE corrected cluster-extent threshold at p 7 4 5 37 < .05 with a cluster-forming threshold of p < .001. The lateral views, which show 7 4 6 projections on the cortical surface, are accompanied by brain slices at the sagittal 7 4 7 midline and also coplanar with the peak of the left IFG cluster (X = -39) that 7 4 8 overlapped across all four social domains (Table S1.5) and the right pSTG cluster 7 4 9 between a) top-down goal-directed retrieval and b) post-retrieval selection of goal-relevant 7 9 6 semantic knowledge (Badre et al., 2005;Jefferies, 2013;Thompson-Schill et al., 1997), and it 7 9 7 has been suggested that both of these two semantic control mechanisms contribute 7 9 8 significantly to interpersonal interactions (Binney and Ramsey, 2020; Satpute and 7 9 9 Lieberman, 2006). Studies of semantic cognition suggest that 'selection' is engaged when cues (e.g., sarcasm). This causes semantic competition that can only be resolved by taking 8 0 4 into account the wider situational and linguistic context and/or prior knowledge about the 8 0 5 speaker (Aviezer et al., 2008;Pexman, 2008). Controlled retrieval processes, on the other 8 0 6 hand, are engaged when automatic semantic retrieval fails to activate the semantic 8 0 7 information necessary for the task at hand. This may occur frequently in social interactions, 8 0 8 and particularly with less familiar persons, because of a preponderance of surface features 8 0 9 (e.g., physical characteristics) over less salient features (e.g., personality traits, preferences, 8 1 0 and mental states). To avoid exchanges that are deemed superficial at best, controlled 8 1 1 retrieval must be used to bring to the fore person-specific but also context-relevant semantic 8 1 2 information. 8 1 3 There is now over a decade's worth of multi-method evidence that semantic control is Research is now aimed at understanding the neural mechanisms by which these regions 8 1 6 modulate semantic processing. One recent proposal is that it involves coordination of 8 1 7 spreading activation across the semantic representational system (Chiou et al., 2018). that represent modality-specific information (e.g. visual features, auditory features, verbal 8 2 2 labels, etc). Chiou et al., (2018) showed that the left IFG could be imposing cognitive control 8 2 3 by flexibly changing its effective connectivity with the hub and spoke regions according to 8 2 4 task characteristics; the IFG displayed enhanced functional connectivity with the 'spoke' 8 2 5 region that processes the most task-relevant information modality. A similar proposal has 8 2 6 been made for the contribution of domain-general cognitive control systems to social 8 2 7 information processing. Zaki et al. (2010) found that, in the presence of conflicting social 8 2 8 cues, IFG activity becomes functionally coupled with the brain areas associated with 8 2 9 processing the particular cue type the participant chose to rely on to make inferences about 8 3 0 emotional states. On this basis, they proposed that cognitive control areas upregulate 8 3 1 activation within systems that represent social cues that are currently most relevant to the 8 3 2 task. Consistent with this, a further study found evidence to suggest that the left IFG 8 3 3 downregulates neural activation associated with task-irrelevant self-referential information 8 3 4 when the task required reference to others (and vice versa) (Soch et al., 2017). 8 3 5 An important feature of semantic processing is the ability to accommodate new 8 3 6 information that emerges over extended periods of time and update our internal 8 3 7 representation of the current "state of affairs" in the external world according to contextual 8 3 8 changes. This is particularly important for navigating social dynamics which are liable to 8 3 9 abrupt and sometimes extreme changes in tone. For instance, imagine being in a bar and 8 4 0 having your attention drawn to someone standing suddenly and picking up a glass. One might 8 4 1 reasonably infer that this person is thirsty. That is until they proceed to walk towards a group 8 4 2 of noisy sports fans rather than the bartender. In this case, you will likely adapt your 8 4 3 interpretation and engage in a pre-emptive defensive stance. Recent research suggests that 8 4 4 this ability to update depends, at least in part, on the IFG, as well as the mPFC and ventral 8 4 5 IPL (also see Section 4.2.2) (Branzi et al., 2020). Likewise, Lavoie et al., (2016) showed that, 8 4 6 during a ToM task, activation of the left IFG and pMTG is associated with contextual 8 4 7 adjustments of mental state inferences (and also more general physical inferences) although 8 4 8 not the representation of mental states specifically. Left IFG activation has also been 8 4 9 observed when newly-presented information requires one to update the initial impression (Lambon Ralph et al., 2017;Binney & Ramsey, 2020). The latter includes top-down 8 5 7 attentional control and working memory systems that support goal-driven behaviour 8 5 8 irrespective of the task domain (i.e., perceptual, motor or semantic). These processes recruit 8 5 9 nodes of the MDN (Duncan, 2010), which include the precentral gyrus, MFG, IPS, insular 8 6 0 cortex, pre-SMA and adjacent cingulate cortex (Assem et al., 2020;Fedorenko et al., 2013). 8 6 1 In terms of organisation, the SCN appears to be nested among domain-general executive 8 6 2 systems  and could play a role in mediating interactions between the 8 6 3 MDN and the semantic representational system (Davey et al., 2016;Lambon Ralph et al., 8 6 4 2017). In line with this general perspective, we expected MDN regions to be reliably engaged 8 6 5 by all four social sub-domains explored in the present meta-analyses. While there was 8 6 6 evidence of engagement of the MFG, the pre-SMA, ACC, insula and IPS in some of the 8 6 7 social sub-domains, MDN regions were not part of the core social processing network 8 6 8 identified by the overlay conjunction analysis. This could reflect the fact that the majority of 8 6 9 contrasts included in our meta-analyses employed high-level control conditions that were 8 7 0 42 well-matched to the experimental conditions in terms of general task requirements, and thus, 8 7 1 most activation associated with general cognitive demands had been subtracted away. 8 7 2 Consistent with this notion is the fact that studies contrasting social tasks with lower-level 8 7 3 control conditions (e.g., passive fixation) find more extensive MDN activation in ToM 8 7 4 (Mason et al., 2008;Mier et al., 2010), trait inference (Chen et al., 2010;Hall et al., 2012), constraints to facilitate goal-driven aspects of processing that is not limited to the semantic 8 8 6 domain (Duncan, 2013;Fedorenko et al., 2013;Gao et al., 2020;Whitney et al., 2012). In 8 8 7 contrast, the engagement of the anterior ventrolateral IFG (pars orbitalis) and the left pMTG a single unifying framework Rice et al., 2018). Some clues already exist 1 0 1 8 in the aforementioned work of , who observed a broader symptom 1 0 1 9 complex comprising multimodal semantic impairments, including visual and auditory 1 0 2 0 associative agnosias, that might explain rather than just co-present with social-affective 1 0 2 1 disturbances. More recent work that leverages the higher spatial resolution of functional 1 0 2 2 neuroimaging in humans has revealed a ventrolateral ATL region that responds equally to all 1 0 2 3 types of concepts, including social, object and abstract concepts, be they referenced by verbal 1 0 2 4 and/or non-verbal stimuli Rice et al., 2018). Activation of the dorsal-1 0 2 5 polar ATL, on the other hand, appears to be more sensitive to socially-relevant semantic 1 0 2 6 stimuli Rice et al., 2018;Zahn et al., 2007b). These observations support 1 0 2 7 a proposal in which the broadly-defined ATL region can be characterised as a domain-1 0 2 8 general supramodal semantic hub with graded differences in relative specialisation towards 1 0 2 9 certain types of conceptual information (Binney et al., 2012;; Lambon 1 0 3 0 Ralph et al., 2017;Plaut, 2002;Rice et al., 2015). Our results reveal that the temporal poles 1 0 3 1 are reliably activated across four social domains -moral reasoning, empathy for emotions, 1 0 3 2 ToM and trait inference. They do not, however, provide support for the involvement of the 1 0 3 3 ventrolateral ATL. We argue this is likely due to technical and methodological limitations of 1 0 3 4 the fMRI studies included in the meta-analyses (see Visser et al., 2010). Most notably this 1 0 3 5 includes vulnerability to susceptibility artefacts that cause BOLD signal drop-out and 1 0 3 6 geometric distortions around certain brain areas, including the ventral ATLs (Jezzard and 1 0 3 7 , 1999;Ojemann et al., 1997). Studies that have used PET, which is not vulnerable to Because semantic control demands were not explicitly manipulated in the social 1 0 4 5 contrasts we included, our results cannot directly confirm our hypothesis regarding the 1 0 4 6 specific contribution made by the SCN in social cognition. Our conclusions rely on an 1 0 4 7 assumption that overlap reflects a generalised neurocomputation upon which both semantic 1 0 4 8 control and social processing rely. The alternative explanation is that overlapping activation 1 0 4 9

Clare
reflects tightly yet separately packed cognitive functions which may only dissociate when 1 0 5 0 investigated at an increased spatial resolution (Henson, 2006;Humphreys et al., 2020). studies employed non-verbal stimuli. Some of the differences between the associated 1 0 6 0 networks (e.g, in lateralization) might therefore be attributable to verbal processing demands. 1 0 6 1 As is the case with all meta-analyses, therefore, some aspects of our results should be treated 1 0 6 2 with caution. 1 0 6 3 Another limitation of this study is that most of the experiments included used control 1 0 6 4 conditions that were highly matched to their experimental conditions in terms of the demand 1 0 6 5 for domain-general processes such as cognitive control and semantic processing, and 1 0 6 6 therefore they may have subtracted away much of the activation we were aiming to explore. 1 0 6 7 Despite this, we did find consistent activation of the SCN, particularly the left IFG, across all 1 0 6 8 four social domains. This may be because, although a considerable subset of included 1 0 6 9 50 experiments had high-matching control conditions, not all may have properly controlled for 1 0 7 0 semantic control demands specifically. An alternative explanation is that processing socially-1 0 7 1 relevant conceptual knowledge may impose greater demands on the SCN. Consistent with 1 0 7 2 this, it has been shown that processing social concepts relative to non-social concepts led to 1 0 7 3 increased activation of the SCN even when controlling for potentially confounding 1 0 7 4 psycholinguistic factors . Elucidating the neural bases of social control and representation may help us 1 0 9 0 understand the precise nature of social impairments resulting from damage to different neural 1 0 9 1 systems. For example, our framework (Binney & Ramsey, 2020) predicts that damage to 1 0 9 2 representational areas such as the ATL will impair social information processing irrespective 1 0 9 3 of task difficulty or the need to integrate context. In contrast, we expect that damage to 1 0 9 4 control areas would lead to impaired social processing specifically when it requires selecting 1 0 9 5 from amongst alternative interpretations of social cues, and/or retrieving social information 1 0 9 6 that is only weakly associated with a person or a situation. Damage to perisylvian frontal 1 0 9 7 and/or temporo-parietal areas (comprising the SCN) leads to semantic aphasia, a disorder 1 0 9 8 characterized by impaired access and use of conceptual knowledge (Corbett et al., 2009; 1 0 9 9 Jefferies et al., 2008Jefferies et al., , 2007Jefferies and Lambon Ralph, 2006;Noonan et al., 2010). This  al., 2007;Rogers et al., 2004). As far as we are aware, the extent to which brain damage that 1 1 0 4 leads to semantic aphasia also affects social abilities has not yet been formally investigated. 1 1 0 5 Some insight can be found in neurodegenerative patients with prominent frontal lobe damage, 1 1 0 6 where social impairments can be linked to deficits in executive function (Healey and 1 1 0 7 Grossman, 2018; Kamminga et al., 2015). More generally, it will be interesting to discover 1 1 0 8 whether a distinction between knowledge representation and cognitive control can inform our 1 1 0 9 understanding of the precise nature of atypical or disordered social cognition in, for example, 1 1 1 0 the context of dementia, acquired brain injury, autism spectrum conditions and schizophrenia.