Domain‐general and domain‐specific functional networks of Broca's area underlying language processing

Abstract Introduction Despite abundant research on the role of Broca's area in language processing, there is still no consensus on language specificity of this region and its connectivity network. Methods The present study employed the meta‐analytic connectivity modeling procedure to identify and compare domain‐specific (language‐specific) and domain‐general (shared between language and other domains) functional connectivity patterns of three subdivisions within the broadly defined Broca's area: pars opercularis (IFGop), pars triangularis (IFGtri), and pars orbitalis (IFGorb) of the left inferior frontal gyrus. Results The findings revealed a left‐lateralized frontotemporal network for all regions of interest underlying domain‐specific linguistic functions. The domain‐general network, however, spanned frontoparietal regions that overlap with the multiple‐demand network and subcortical regions spanning the thalamus and the basal ganglia. Conclusions The findings suggest that language specificity of Broca's area emerges within a left‐lateralized frontotemporal network, and that domain‐general resources are garnered from frontoparietal and subcortical networks when required by task demands.

Studies of the functional, effective, and structural connectivity of Broca's area, generally employing resting-state fMRI, task-based fMRI coupled usually with dynamic causal modeling, and diffusion tensor imaging, respectively, have provided insights into the connectivity of Broca's area. Specifically, resting-state fMRI studies identified a largely left-lateralized functional connectivity pattern for Broca's area involving frontal, temporal, and parietal cortices, as well as several subcortical areas (e.g., the basal ganglia) (Tomasi & Volkow, 2012;Xiang et al., 2010). Structural connectivity research identified various white matter pathways including the superior longitudinal fasciculus (arcuate fasciculus), middle longitudinal fasciculus, inferior fronto-occipital fasciculus, extreme capsule, external capsule, and uncinate fasciculus that connect Broca's area with the superior and middle temporal gyri as well as with the inferior parietal lobe (supramarginal and angular gyri) (Axer et al., 2013;Glasser & Rilling, 2008;Kellmeyer et al., 2013;Parker et al., 2005;Powell et al., 2006;Saur et al., 2010). Studies of the effective connectivity of Broca's area also delineated the functional connectivity profile of this region both during resting state (Gao et al., 2020), and during various tasks including inhibitory control (Guha et al., 2020), speech production (Eickhoff et al., 2009a), and language processing (den Ouden et al., 2012;Schmithorst et al., 2007;Sonty et al., 2007), highlighting the causal associations between Broca's area and various cortical and subcortical regions. Importantly, this body of research underscored the connection between Broca's area and the posterior superior temporal cortex (Wernicke's area), providing convergent evidence, together with structural connectivity findings, for the primary role of this loop for language processing (den Ouden et al., 2012;Schmithorst et al., 2007;Sonty et al., 2007). Despite progress in understanding the functional network of Broca's area facilitated by this line of research, these techniques are not without limitations. First, resting-state fMRI and structural connectivity studies reveal taskindependent connectivity patterns of a given brain area, preventing the identification of domain-specific connectivity networks. Second, although effective connectivity can identify information flow within a functional network in a task-dependent manner, it is typically utilized in studies with a limited number of participants engaged in a specific task, limiting generalizability of the findings.
A recently developed technique in neuroimaging research that can circumvent these limitations is meta-analytic connectivity modeling (MACM) . MACM combined with activation likelihood estimation (ALE) can be used to identify functional connectivity of a given brain region by calculating its co-activation patterns using a database of neuroimaging experiments (BrainMap). Importantly, thanks to a detailed taxonomy of experiments enabling searching through the metadata of experiments including behavioral domains, categories, and subcategories Lancaster et al., 2012), BrainMap allows estimation of task-independent, or domain-general, (Erickson et al., 2017;Robinson et al., 2010), as well as task-dependent, or domain-specific, functional connectivity of a brain region (Ardila et al., 2016;Bernal et al., 2015;Viñas-Guasch & Wu, 2017). Given that large databases of experiments with various tasks and designs are utilized, MACM can produce highly generalizable findings (Samartsidis et al., 2020). Recent MACM investigations of IFG revealed a language network spanning largely left-lateralized frontal, temporal, and parietal regions, as well as several subcortical structures (Bernal et al., 2015;Bulut, 2022a). Furthermore, striking differences in the languagerelated functional connectivity patterns were observed among IFG subdivisions, with the left IFGop co-activating with a broad network of cortical, subcortical, and cerebellar structures (Bulut, 2022a).
Although MACM has been employed to identify functional connectivity of the left IFGop for language tasks (Bernal et al., 2015) and to parcellate distinct clusters within the left IFGop and identify their connectivity for different functional domains including language (Clos et al., 2013), no previous meta-analytic connectivity study directly compared functional connectivity of subdivisions of the broadly defined Broca's area, including the left IFGop, IFGtri, and IFGorb, for language and nonlanguage tasks. Although a recent study investigated language-related functional connectivity of bilateral IFGop, IFGtri, and IFGorb using MACM (Bulut, 2022a), these connectivity patterns were not compared with the connectivity patterns of the relevant regions for other, nonlanguage domains. Given that a functional network identified during language tasks may still involve domain-general processes, such as working memory and cognitive control, directly exploring divergence (through contrast analyses) and convergence (through conjunction analyses) between the functional network identified for language tasks and that identified for nonlanguage tasks may help disentangle the domain-specific (specific to language) and domaingeneral (shared between language and nonlanguage domains) neural circuitry of Broca's area. Against this background, the present study builds on and extends a previous MACM study on the functional connectivity of IFG for language tasks (Bulut, 2022a). Thus, the aim here is to explore language-related domain-specific and domain-general coactivation patterns of the left IFGop, IFGtri, and IFGorb by utilizing the MACM method and the BrainMap functional neuroimaging database.
To my knowledge, this is the first meta-analytic connectivity study directly investigating language-related domain-specific and domaingeneral functional connectivity networks of the opercular, triangular, and orbital parts of Broca's area in a broad sense. F I G U R E 1 Anatomical 3-D renderings of the ROIs used in the meta-analyses. The color bars indicate probability of capturing the relevant anatomical structure within the ROI.

MATERIALS AND METHODS
The MACM procedure employed in the current study involved defining regions of interest (ROIs) within the left IFG, using the ROIs in addition to several criteria to search the BrainMap database for neuroimaging experiments with language tasks and with tasks other than language that report activation in the relevant ROI, and carrying out ALE analyses to reveal co-activation network of each ROI for language and other tasks and to compute their contrast and convergence. 1

Regions of interest
Three ROIs were defined based on the probabilistic, cytoarchitectonic Julich-Brain atlas . As shown in Figure 1, the ROIs corresponded to IFGop, IFGtri, and IFGorb. The ROI maps were obtained from the European Human Brain Project (EHBP) website (https://ebrains.eu), which stores cytoarchitectonic maps for various brain regions  The thresholded maps were used to create the ROIs. To ensure that the intended brain regions were captured, the ROIs were visually inspected using the brain atlases in Mango and using different brain templates in MRIcron (https://www.nitrc.org/projects/mricron) (Rorden & Brett, 2000).

Database search
Database searches were conducted within the BrainMap functional database on 27 September, 2021 using Sleuth Version 3.0.4 (Fox & Lancaster, 2002;Fox et al., 2005;Laird et al., 2005 images of the ROIs," "experimental context: normal mapping," "behavioral domain: is not cognition.language, is not action.execution.speech," "subjects: normal," "experimental activation: activations only," "imaging modality: fMRI or PET," whereas the following search terms were used in the language query: "locations: MNI images of the ROIs," "experimental context: normal mapping," "behavioral domain: cognition.language," "experimental activation: activations only," "subjects: normal," "handedness: right," "imaging modality: fMRI or PET." Restriction of the searches to "normal mapping" and "normals" ensured that only the experiments conducted with healthy subjects were included. The ROIs defined as explained above were separately included as a search criterion in the database searches. The language search was intended to yield only language-relevant activations, hence, only the  Of note, since nonlanguage cognitive domains or categories were not added in the language search as exclusion criteria to ensure as broad coverage of language-related experiments as possible, a subset of the experiments identified in the language search also related to some other domains (e.g., perception.audition). Take, for example, a picture naming study included in the current meta-analyses (Wilson et al., 2009). In an overt picture naming paradigm, participants were asked to name pictures while they did nothing in response to scrambled pictures, which were created by randomly shuffling parts of many  (Beauregard et al., 1997). Two contrasts from this study contributed to the language analyses for IFGtri and IFGorb, but they were also categorized within the emotion.negative (unspecified) domain. Importantly, although these experiments contributed to the language meta-analyses in the current study, they were not included in the nonlanguage analyses as they were also categorized in the language domain (note that being categorized in the cognition.language or action.execution.speech domains is an exclusion criteria for the nonlanguage domain here). This meant that an experiment or contrast could not be in both comparison sets (language and nonlanguage).
The foci identified in each search were grouped by experiment using the most conservative approach (Turkeltaub et al., 2012); that is, foci reported in multiple experiments in a single study were combined and entered into the meta-analyses as a single experiment to prevent a single experiment from overinfluencing the results. The icbm2tal transform was implemented to automatically convert coordinates reported in Talairach space into MNI space Lancaster et al., 2007).

ALE analyses
Convergence of co-activations for each ROI was computed through ALE analyses using GingerALE 3.0.2 (Eickhoff et al., 2009b;Eickhoff et al., 2012). To that end, ALE analyses were performed using the activation coordinates identified for each ROI as a result of the language and nonlanguage searches. Standard procedures were implemented to carry out the ALE analyses as reported in previous research (Cieslik et al., 2015;Müller et al., 2017;Wojtasik et al., 2020). In particular, 3D Gaussian probability distributions centered at each foci group were generated using a full-width half-maximum, which was calculated based on the sample size in each experiment (Eickhoff et al., 2009b).
Then, the union of modeled activation maps was acquired to compute voxel-wise ALE scores. Afterwards, the union of these activation probabilities was compared against the null hypothesis of random spatial association between the experiments. Finally, the p-value distributions derived from these probabilities were thresholded at a voxel-level To compare language and nonlanguage co-activation patterns for each ROI, bidirectional contrast or subtraction analyses (language > nonlanguage, nonlanguage > language) as well as conjunction analyses (language ∩ nonlanguage) were performed using the Contrast Datasets utility in GingerALE. The network identified by the contrast of language > nonlanguage is interpreted as the domainspecific (language-specific) network, while the conjunction of language ∩ nonlanguage is taken to reflect the domain-general (shared between language and nonlanguage domains) network of the ROIs. The nonlanguage > language network, on the other hand, corresponds to co-activations of the ROIs specific to nonlanguage domains. In addition to these main analyses which combine foci from all domains other than language and speech into a general nonlanguage set, exploratory analyses were conducted to identify domain-specific and domaingeneral networks of the ROIs within individual behavioral domains.
The exploratory analyses were conducted to see, across different behavioral domains, whether the domain-specific co-activation patterns would be consistent and whether and how the domain-general networks would differ. Given that inclusion of at least 17-20 experiments in ALE meta-analyses has been recommended to obtain enough power for identification of small effect sizes and to ensure that the results are not overly influenced by individual studies Müller et al., 2018), the upper bound of this recommendation was adopted. Thus, exploratory analyses were performed for the individual behavioral domains in Table 2 which contributed 20 or more experiments. These are action.execution, cognition.attention, cognition.memory.working, and cognition.music for IFGop, and cognition.attention for IFGtri. It should be noted that the ALE subtraction analysis applies permutation significance testing, which controls for differences in the number of papers on each side of the comparison (Eickhoff et al., 2011;Erickson et al., 2017).
Since GingerALE conducts contrast analyses based on already thresholded single-dataset images, and since cluster-level inference is not currently available in GingerALE for contrast analyses (Hoffman & Morcom, 2018), an uncorrected threshold of p < .05 with an extent threshold (minimum cluster size) of 100 mm 3 was used for the contrast and conjunction analyses, as applied in previous research (Bulut, 2022a;D'Astolfo & Rief, 2017;Garrison et al., 2013;Hobeika et al., 2016;Kollndorfer et al., 2013;Papitto et al., 2020). The Talairach Daemon embedded in GingerALE was used to generate anatomical labels as the nearest gray matter within 5 mm for the activation peaks (Lancaster et al., 1997(Lancaster et al., , 2000. The Mango software

Results of main analyses
The contrast and conjunction results of the language and nonlanguage ALE analyses for the left IFGop, IFGtri, and IFGorb are summarized in Tables 3 and 4, and illustrated in Figure 2. The most widespread coactivation pattern for both contrast analyses and the conjunction analysis was observed for IFGop, followed by IFGtri and, lastly, by IFGorb.

Results of exploratory analyses
The results of contrast and conjunction analyses between language and four nonlanguage domains (action.execution, cognition.attention, cognition.memory.working, and cognition.music) for IFGop are illustrated in Figure 3 (see Supporting Information Tables SI1-4 for a detailed summary of the results), and those between language and cognition.attention domains for IFGtri are illustrated in Supporting Information Figure SI1 and summarized in Supporting Information and temporal (fusiform gyrus) structures ( Figure 3). Likewise, the domain-specific co-activation network of IFGtri (language > cognition.attention) closely matched the domain-specific network identified in the main analyses, involving primarily left frontotemporal regions (IFG, MFG, MTG, fusiform gyrus) (Supporting Information Figure SI1).

DISCUSSION
Using the MACM method, the present study investigated the differ- Broca's area that underlie language processing and that are, at least partially, distinct from domain-general networks (Campbell & Tyler, 2018;Fedorenko & Blank, 2020;Fedorenko et al., 2011). Although it is not possible to pinpoint relative contributions of different linguistic components (syntax, phonology, semantics, etc.) to the domain-specific functional network of the left IFG based on the present findings, it is probably not justifiable to attribute it solely to syntax as has been done in some previous research (Campbell & Tyler, 2018;Grodzinsky & Friederici, 2006;Grodzinsky & Santi, 2008), given that not only the left IFGop, which has more commonly been involved in syntactic processing , but also IFGtri and IFGorb, which have often been associated with semantic processing (Hagoort & Indefrey, 2014), revealed domain-specific networks that survived after subtraction of their respective nonlanguage networks. The present findings are incompatible also with the claims that attribute the role of Broca's area in language processing to its involvement in domaingeneral processes including cognitive control (January et al., 2009;Novick et al., 2005Novick et al., , 2010, or representation of complex structural and hierarchical relationships across domains including not only language, but also action and music (Fadiga et al., 2009;Fitch & Martins, 2014).
Exploratory analyses also revealed a closely similar left frontotemporal domain-specific network when the language domain was contrasted with action.execution, cognition.attention, cognition.memory.working, and cognition.music for IFGop, and cognition.attention for IFGtri.
This finding suggests that the domain-specific left frontotemporal network identified in the main analyses was not due to overinfluence of a given behavioral domain, but rather was consistent across different domains, at least for the domains that could be analytically investigated for IFGop (four domains) and IFGtri (one domain).
Given that the large number of experiments included in the nonlanguage analysis were related to various nonlanguage and nonspeech behavioral domains (action, cognition, emotion, interoception, and perception), it was possible to take into account extralinguistic factors that may accompany certain language tasks and that were previously associated with Broca's area such as emotional processing (Belyk et al., 2017), mathematical processing (Maruyama et al., 2012), action processing (Clos et al., 2013;Papitto et al., 2020), working memory (Clos et al., 2013;Makuuchi et al., 2009), cognitive control (Clos et al., 2013;Novick et al., 2005Novick et al., , 2010, and music (Heard & Lee, 2020;Koelsch, 2006). It should be acknowledged upfront that the current study did not undertake a characterization of the nonlanguage network, which, as mentioned above, spans multiple behavioral domains, each of which may have contributed to the obtained results differently. Therefore, the nonlanguage network was treated as a comprehensive, but heterogenous, baseline against which the language-related co-activation profile was compared to delineate the domain-specific network. Nevertheless, the exploratory analyses provided safeguard to a certain extent by showing consistent domain-specific co-activation patterns when compared to individual nonlanguage domains, as mentioned in the preceding paragraph, and enabled better characterization of the domain-general network for specific domains, as discussed below.
Interestingly, the frontal and parietal regions identified here as part of the nonlanguage and domain-general networks overlap with the frontoparietal network that has been highlighted as a domaingeneral system underlying a range of cognitive functions. Specifically, the frontoparietal network has been conceptualized as a control system incorporating various regions involved in cognitive control and decision making (Vincent et al., 2008) and as a multiple-demand system underlying various cognitive functions that drive intelligent, goaldirected behavior (Duncan, 2010;Duncan & Owen, 2000). Indeed, previous research associated parts of Broca's area with the domaingeneral frontoparietal multiple-demand network (Fedorenko & Blank, 2020;Fedorenko et al., 2011). Specifically, superior-posterior parts of the left IFG were found to be shared between domain-specific (language) and domain-general (verbal working memory and cognitive control) functions (Fedorenko et al., 2011). Likewise, the present study found overlap between language and nonlanguage networks in Relatedly, the domain-general network may also include coactivations arising from executive control-related functions that have been associated with bilingual language processing. In particular, bilingual language control and language switching have been associated with an inhibitory control network spanning frontoparietal and subcortical structures (Calabria et al., 2018;Luk et al., 2012;Rodriguez-Fornells et al., 2006;Sulpizio et al., 2020). Enhanced recruitment of parts of this network in bilingualism has been ascribed to inhibitory control over and competition among lexemes and lexicons, among others (Abutalebi & Green, 2007;Rodriguez-Fornells et al., 2006).
Moreover, neuroimaging studies directly comparing monolingual and bilingual language processing suggest that when faced with increasing cognitive demands, both groups recruit a common domain-general system, which may be more activated in bilinguals due to competition within and between languages (Abutalebi et al., 2013;Parker Jones et al., 2012). Therefore, domain-general frontoparietal and subcortical (left thalamus and basal ganglia) systems may play a role in language processing when additional cognitive resources are needed such as when resolving conflicts due to competition among lexical items or languages as in bilingual language processing.
The domain-general and nonlanguage networks identified in the current study also included several structures that constitute the resting-state default mode network. The default mode network spans the bilateral parietal (precuneus, IPL), posterior cingulate, medial prefrontal, and medial and lateral temporal cortices and has been associated with the brain's intrinsic activity (Raichle, 2015;van den Heuvel & Hulshoff Pol, 2010). Previous research showed that the precuneus interacts with both the frontoparietal and the default mode networks, potentially playing a crucial role in organizing task-and restrelated brain activity across these two systems (Utevsky et al., 2014).
Consistently, the present study identified the precuneus particularly within the nonlanguage network of IFGop and IFGtri, but also within the domain-specific and domain-general networks of IFGop. Furthermore, the exploratory analyses showed that the domain-general network of IFGop spanned the precuneus along with SPL and IPL only for the conjunction of language with cognition.attention and cognition.memory.working out of the four behavioral domains investigated.
These findings imply that the precuneus may serve as an interface not only between the two domain-general (frontoparietal multipledemand and default mode) networks, but also between the language and nonlanguage networks of the left IFG.
Another interface region revealed in the present study is the left orbitofrontal cortex, which demonstrated a lateral-to-medial gradient for IFGorb with lateral, lateral-medial, and medial co-activation patterns for the domain-specific, domain-general, and nonlanguage networks, respectively. Indeed, although the nonlanguage co-activation network of IFGorb included several limbic structures (right cingulate and parahippocampal gyri), the only overlap between language and nonlanguage connectivity of IFGorb was observed in the left orbitofrontal cortex, with the aforementioned lateral-to-medial gradient. This finding is consistent with a previous meta-analysis that associated lateral IFGorb with both emotion and semantics, and medial or opercular IFGorb with emotion alone (Belyk et al., 2017).
Taken together, these findings highlight the similarities and differences between the domain-specific and domain-general contributors to language processing, which may better be conceptualized as a gradient than a dichotomy.
Although the functional connectivity networks of the ROIs were predominantly cortical, several subcortical structures were also iden- whereas only IFGop exhibited distinct co-activation with the right cerebellum within its domain-specific network. The exploratory analyses also identified this domain-specific co-activation of the right cerebellum with IFGop when contrasted with action.execution, cognition.attention, and cognition.music. Previous research associated the cerebello-basal ganglia-thalamo-cortical system with a broad range of cognitive and sensorimotor functions including language (Bostan & Strick, 2010, 2018Bostan et al., 2013;Caligiore et al., 2017;Ford et al., 2013;Tomasi & Volkow, 2012). Accordingly, the present findings also underline the domain-specific and domain-general aspects of this loop for language processing.
Several potential limitations should be addressed. First, the experiments included in the analyses of the language network of IFG subdivisions pertain to multiple language components (syntax, semantics, phonology, orthography, speech), tasks (e.g., comprehension, production), and modalities (e.g., visual, auditory). This was intended as a generalization from specific tasks to language processing in general.
Nevertheless, divergence, but also convergence, in the neural representation of different language functions were shown in previous research, for example, between syntax and semantics (Hagoort & Indefrey, 2014;Rodd et al., 2015). Therefore, the domain-specific network identified in this study likely consists of subnetworks that may at least partially be dissociable from each other. Relatedly, it could be argued that certain language subdomains, such as syntax, are more domain-specific than others (Campbell & Tyler, 2018;Grodzinsky & Friederici, 2006;Grodzinsky & Santi, 2008 (Clos et al., 2013;Papitto et al., 2020), and within IFGorb along the lateral-medial axis (Belyk et al., 2017).
Indeed, to ensure that ROIs represented the intended brain regions, they were thresholded with minimum probabilities greater than .48.
Finally, although three subdivisions of IFG are scrutinized with regard to their domain-specific functional co-activation network, no claim has been made concerning language-specificity of any given IFG subregion. This is because the present study does not attempt a functional fractionation of Broca's area, and merely examines the languagespecific functional network of Broca's area, a region which has been conceptualized anatomically to include different combinations of IFG subdivisions. Indeed, in parallel with the literature, all three IFG subdivisions examined here exhibited domain-specific, language-related functional networks, albeit with certain differences. Therefore, this exploration does not address whether there are any language-specific subregions within Broca's area in general or within its subdivisions.
Future investigations of language-specific subregions within Broca's area and in other brain regions may benefit from more detailed parcellation techniques in ROI definition including connectivity-based parcellation (Fan et al., 2016;Wang et al., 2015) and multiple receptor mapping .

CONCLUSIONS
The present findings show that the language-related domain-specific functional network of Broca's area spans mainly left-lateralized frontotemporal regions including the middle and inferior frontal cortices as well as anterior, posterior, and inferior temporal cortices. Broca's area was also associated with domain-general frontoparietal and subcortical (thalamus and basal ganglia) networks shared between language and nonlanguage domains. The findings suggest that Broca's area, or Broca's complex in a broad sense, spanning the opercular, triangular and orbital parts of the left IFG, exhibits language-specificity as part of a functional connectivity network involving a left frontotemporal system, which recruits domain-general resources from frontoparietal multiple-demand and subcortical (thalamus and basal ganglia) networks based on task demands. Application of MACM to compare divergence and convergence between domain-specific and domaingeneral functional networks of brain regions, as in the present study, offers significant potential for explorations of specific and shared networks for other language-related regions and other behavioral domains.

ACKNOWLEDGMENTS
Open access funding enabled and organized by Projekt DEAL.

DATA AVAILABILITY STATEMENT
All data and analysis files associated with the study are available at https://doi.org/10.17632/tfg4pryhf9.1.

NOTES
1 The left-hemispheric ROIs and the neuroimaging experiments included in the language analyses in the present research were the same as a previous study on language-related connectivity of bilateral IFG (Bulut, 2022a). However, differently from that study, the present study included a nonlanguage contrast and compared functional networks of the left IFG subdivisions for language and nonlanguage domains in contrast and conjunction analyses in an attempt to identify language-related domain-specific and domain-general circuitry of these ROIs. 2 Minimum and maximum probabilities are given in the color bars in Figure 1. The mean probabilities (SD) and sizes of the ROIs were as follows: IFGop: 0.77 (0.09), 2409 mm 3 ; IFGtri: 0.64 (0.12), 2363 mm 3 ; IFGorb: 0.68 (0.12), 2353 mm 3 . 3 In the nonlanguage search with IFGop as the ROI, one duplicate experiment (Rypma et al., 2001), which reported the same coordinates of activation as another experiment (Rypma et al., 2001), was identified and eliminated from both Table 1 and the meta-analyses.