A SENtence Supramodal Areas AtlaS (SENSAAS) based on multiple task-induced activation mapping and graph analysis of intrinsic connectivity in 144 healthy right-handers

In this study, we selected right-handers from the BIL&GIN database because we previously demonstrated that these participants have a left hemisphere dominance for language at both the group level (Tzourio-Mazoyer et al. 2016) and the individual level (Zago et al. 2017), with only 5 (3%) of the 144 participants having a co-dominant right hemisphere. This sample group is optimal for selecting areas specific for sentence processing based on a conjunction of activations and leftward asymmetries. In addition, the inclusion of a fairly considerable number of participants (N = 144) provided us a high sensitivity for detecting supramodal sentence areas while minimizing the risk of overlooking some.

However, we must underline that the present atlas is not all-inclusive. First, we selected map regions involved in only high-order language processing and lexico-syntactic processing. Using the list of familiar words as the reference condition, we removed the dorsal route of language, including the phonological loop, responsible for articulation and sound-to-articulation mapping (Saur et al. 2008; Rauschecker and Scott 2009) In addition, the regions selected herein focused on the left hemisphere and did not account for right hemisphere-specialized aspects of sentence processing, such as emotional prosody (Beaucousin et al. 2007; Hurschler et al. 2012) and context processing (Grindrod and Baum 2003; Ferstl et al. 2005).

Second, the presence of susceptibility artefacts combined with averaging the large number of participants led to incomplete mapping of the inferior part of the temporal lobe, prohibiting us from documenting some areas, such as the basal language area in the anterior part of the fusiform gyrus. This essential language area, first identified using deep electrical recordings (Nobre et al. 1994), has been shown with positron emission tomography (PET) to be activated during both the production and auditory comprehension of language (Papathanassiou et al. 2000).

Third, small size regions may also be lacking in this atlas since we provided data at the hROI scale rather than at the voxel scale.

Concerning the clustering methods of correlation values at rest, a perfect match was observed between the Ward and average clustering methods (see supplementary material Table 1), and a good score was obtained with the Gaussian mixture model for global clustering at the 32 hROI levels. The weakest score was that of the k-means method, and such a difference in clustering observed with k-means compared to that in the other 3 methods is consistent with the fact that, as reported by Thirion et al., k-means forms clusters spatially close and connected but with poor reproducibility using the sample studied. By contrast, hierarchical clustering using Ward’s method, which we selected to segregate the networks, was reported to create connected clusters that are highly reproducible using the studied samples (Thirion et al. 2014).

We carefully designed each of the language tasks such that joint analyses were possible; the design was identical in the three tasks, and we chose to make them close enough to allow comparisons and conjunctions in terms of the number of words or the complexity of sentences. As mentioned above, the use of a high-level verbal reference task for controlling the involvement of primary areas (auditory, visual and motor) and removing phonological and automatic word processing kept the lexico-syntactic aspects common to all three tasks.

The first set of 32 hROIs provides left hemispheric regions that are dedicated to the monitoring and completion of tasks based on sentence processing. Although not all regions can be considered essential language areas, all were determined to be modulated by the verbal material with which they are associated (left activation and leftward asymmetry) and are thus part of an extended language network functioning during language tasks.

Unsupervised hierarchical clustering based on the resting-state connectivity between hROIs successfully segregated three different networks, including networks hosting core language, visual areas, and posterior areas of the DMN and posterior hippocampus. Within the systems to which they belonged, these networks hosted areas dedicated to the interaction/interface with language systems. For example, the current analysis extracted from among the visual areas involved in picture processing those areas specifically dealing with picture–sentence meaning integration.

Clustering the resting-state correlation between these 32 hROIs allowed the discrimination of SENT_CORE, a network of 18 strongly and positively correlated hROIs, including frontal and temporo-parietal hROIs located on the lateral surface of the left hemisphere and anterior insula areas. In particular, SENT_CORE included areas of the antero-posterior language networks, named in reference to the Broca–Wernicke model in aphasia literature and reported with consistency in meta-analyses of healthy individuals mapped during language tasks see Fig. 6 (Vigneau et al. 2006; Price 2010, 2012). Note that SENT_CORE was the largest network in terms of volume (in mm³), as it included more than half of the hROIs (18/32), all of which were strongly activated and leftward asymmetrical.

In the following, we discuss the potential roles of the identified areas in relation to the literature. However, it is now acknowledged that, apart from very specific regions where a lesion can be closely associated with a specific defect, the role of a given area documented with functional imaging must be understood as the combination of its functional properties with those of the regions with which it constitutes a network to complete a given cognitive task. For example, prec4 is not part of the regions commonly labelled as “language areas”. The present work shows that prec4, located at the junction between the precentral and middle frontal gyrus, is both strongly activated and leftward asymmetrical in the three sentence tasks. Language meta-analyses have reported prec4 as part of the language areas involved in lexico-syntactic processing (named F2p in Vigneau et al. 2006, Fig. 6), and in word selection and hierarchical sequencing (named dPrec in Price 2010, Fig. 6). Applying Neurosynth to prec4 coordinates (x = − 42.2, y = 0.7, z = 50, Table 3) with an association test reveals the greatest number of studies with the terms “sentence”, “comprehension”, “language”, and “sentences”, followed by “eye” and “premotor”. Jouen et al. (2018) propose that prec4 is “involved in the understanding of actions during verbal and non-verbal tasks”. In the present protocol, the sentences involved human actions, closely consistent with that role proposed by Jouen et al. (2018). Using a network approach, Saur et al. (2008) found that prec4 belongs to the sentence comprehension functional network. In Glasser’s atlas, prec4 corresponds to “language area 55b” (Glasser et al. 2016) and overlaps the posterior part of the “rostro-ventral module” described by Genon et al. (2018a, b). In this last work based on peaks meta-analysis, this module is connected with the inferior frontal gyrus, orbital frontal and inferior parietal as prec4 in the present work. The present study further demonstrates that prec4 is involved in a supramodal manner during sentence processing, is strongly leftward asymmetrical, and is strongly and positively connected at rest with the network of areas we named SENT_CORE that hosts essential language areas. These findings are consistent with the fact that prec4 is part of the areas that are conjointly atrophied in patients suffering from nonfluent primary progressive aphasia (PPA) (Mesulam et al. 2014) (Fig. 6).

Indeed, in the frontal lobe, SENT_CORE includes the whole inferior frontal gyrus (F3t and F3O) corresponding to Broca’s area according to most authors, whose lesion is responsible for conversational deficits, as is also the case for the anterior insula (INSa2, INSa3) (Borovsky et al. 2007). The posterior part of the inferior frontal sulcus (f2_2), also part of SENT_CORE, has been underlined as an area involved in lexico-syntactic processing (named F3opd in Vigneau et al. 2006, Fig. 6). Price also targeted f2_2 as being involved in word selection and hierarchical sequencing (named mFG in Price 2010, Fig. 6). Another indication of the important role of f2_2 in language is that together with F3t and prec4, it is the location of atrophy in nonfluent PPA [(Mesulam et al. 2014) see Fig. 6].

In the medial part of the frontal lobe, SENT_CORE includes both preSMA and the superior frontal gyrus (here, SMA2, SMA3 and F1_2), which have been reported in tasks involving sentences dealing with characters (Hervé et al. 2012), as in the present paradigm. The activation of these areas has been attributed to processing the social aspects of verbal material (Ferstl and von Cramon 2002). As discussed below, these medial frontal areas are strongly connected with the posterior part of the middle temporal gyrus (Turken and Dronkers 2011). Notably, these medial frontal areas are also the sites of atrophy in PPA (Wilson et al. 2010a, b; Tetzloff et al. 2018).

In the temporal lobe, SENT_CORE included the STS together with the posterior part of the middle and superior temporal gyri, extending to the angular and supramarginal gyri. This characterization is consistent with the proposal by Price that these areas are involved in amodal semantic combinations, a process common to the three sentence tasks (Price 2010). A recent work on the time course of sentence processing areas has also shown that the posterior temporal lobe, here corresponding to STS3 and T2_3, is involved in lexico-syntactic processing, with the processing of individual words in relation to the syntactic structure (Matchin et al. 2018). These posterior temporal areas have been documented to be essential language areas since the lesion of each results in specific deficits in sentence comprehension (Dronkers et al. 2004). In particular, lesion of the posterior parts of the middle temporal gyrus results in deep aphasia due to a loss of word meaning (Dronkers et al. 2004), consistent with the specific atrophy observed in logopenic PPA (Wilson et al. 2010a, b). STS4, located dorsally to STS3 and T2_3, is involved in sentence-level high-order processing, and Matchin, in his work on the time course of sentence processing investigated with MEG, proposed that “increased activation at the end of the sentence suggests a response associated with the interpretation of the sentence” (Matchin et al. 2018). The integration gradient of sentence meaning from lexical to conceptual progresses posteriorly with SMG7, which corresponds to the area labelled STSp in Vigneau et al.’s meta-analysis (Fig. 6). Vigneau et al. reported the following: “STSp role seems to process the semantic integration of complex linguistic material. This statement comes from the observation that it is recruited when subjects listen to coherent rather than syntactically or pragmatically incoherent sentences (Kuperberg et al. 2000; Luke et al. 2002), and it is involved in context processing and syntactic generation—more activated when subjects have to choose between two words to end a sentence or have to generate the final word of a sentence (Kircher et al. 2001). STSp activity is very likely related to the linkage of linguistic structure to meaning: it is more activated when sentences are linked as dialogue (Homae et al. 2002) or syllogisms (Goel et al. 1998) than when they are unlinked. It is also more activated during text comprehension, either presented auditory [compared to reverse speech (Kansaku et al. 2000; Crinion et al. 2003) or words (Jobard et al. 2007)] or visually [compared to words (Jobard et al. 2007) or pseudo-word reading (Vingerhoets et al. 2003)]”. Price made an alternative proposal in her meta-analysis in 2000 stating that the equivalent of SMG7 (vSMG and pPT) is activated by sentence processing, particularly when the sentences increase in difficulty, and she therefore suggested that SMG7 is involved in subvocal articulation. The present results demonstrating that SMG7 is involved and leftward asymmetrical not only during production and reading but also during sentence listening, which was cited as a easy task by the participants, suggests its involvement in meaning elaboration. The angular gyrus is the final location where this integration towards conceptual knowledge operates dorsally. Known to be involved in lexico-semantic processing, as identified by language task-induced activation studies [(Vigneau et al. 2006; Price 2010), see Fig. 6], the angular gyrus is specifically involved in conceptual knowledge as shown by neuropsychological studies. Lesion of this area leads to the inability to associate a sound or image related to the same concept (Saygin et al. 2004), and inactivation of the angular gyrus in the left hemisphere induces a deficit in semantic integration (Price et al. 2016). The angular hROI identified herein corresponds to Wang’s C5 and C6 parcels, involved in both language and theory-of-mind tasks (Wang et al. 2017). This last observation confirms that the role of the angular gyrus includes the interfacing between sentence processing and the understanding of human actions, a process that is part of the present sentence tasks.

The similitude between the areas composing the SENT_CORE network and the regions showing atrophy in all types of PPA [(Mesulam et al. 2014), see Fig. 6] is striking and a key element evidencing that SENT_CORE contains essential language areas, although the anterior part of the inferior temporal gyrus hosting the basal temporal language area (Nobre et al. 1994; Papathanassiou et al. 2000) is lacking for the methodological reasons discussed above. Finally, the language functional network, which shares a common molecular basis of interaction highlighted by Zilles et al. using neurotransmitter receptor fingerprints (Zilles et al. 2015), overlaps the hROIs of SENT_CORE, apart from AG2. Based on Zilles et al. report, we can hypothesize that the two other sentence-related networks we have identified are likely to have different fingerprints from those of SENT_CORE. However, because the hROIs of these networks were selected as both activated and leftward asymmetrical in the three sentence tasks, it would be of interest to investigate what subset of fingerprints the hROIs of SENT_VISU and SENT_MEM may share in common with SENT_CORE to deepen our understanding of network interactions.

The computation of betweenness and degree centralities allowed us to identify three hubs within SENT_CORE: one in the inferior frontal gyrus (F3t) and two others along the posterior part of the temporal cortex (STS3 and STS4). These high centrality figures demonstrate that these hROIs play a central role in communications with other parts of SENT_CORE, and the participation indices of F3t and STS3 further indicate their key role in communication with the other networks identified. Such properties are consistent with the definition of epicentres proposed by Mesulam et al. (2014), and these regions can be considered essential for sentence/test comprehension independent of the modality. According to Mesulam, the network epicentre specializes in a specific behavioural component, which is language in this study, and the destruction of transmodal epicentres causes global impairments. In fact, as noted above, these hROIs overlap the middle temporal gyrus targeted in the aphasics investigation by Dronkers et al. (2004) as well as the regions showing the highest atrophy in all types of PPA (Mesulam et al. 2014), confirming that they do correspond to epicentres. In addition, these hubs are distributed in the anterior and posterior cortices, constituting an antero-posterior loop across areas belonging to the same hierarchical level in terms of cortical organization, consistent with Fuster’s model for cognition (Fuster and Bressler 2012).

Other studies involving functional connectivity at rest have shown the strong connection of the left inferior frontal gyrus at rest with all SENT_CORE areas [(Margulies and Petrides 2013), Fig. 6]. The similarity between the regions found by these authors when seeding the posterior part of the middle temporal gyrus (here, STS3 and STS4) and the areas constituting SENT_CORE is also clear [(Turken and Dronkers 2011), Fig. 6]. These two observations are consistent with the present observation that these regions are hubs for sentence processing and, more generally, the language comprehension network. Such a framework is also thought to correspond to the Broca–Wernicke language model, with F3t serving as Broca’s area and STS3 and STS4 serving as regions involved in the supramodal integration of meaning, consistent with the location of posterior areas leading to comprehension deficits (Pillay et al. 2017). These posterior STS areas are also considered by Binder as areas supporting meaning integration during sentence comprehension (Binder 2017). Considering that a left deficit in this area leads to deficits in language comprehension, we propose to label it Wernicke’s area, although as reviewed by Binder, such a definition is different from the location currently proposed for Wernicke’s area in the superior temporal gyrus. In the present work, closely adhering to an anatomo-functional definition in reference to deficits in comprehension associated with Wernicke’s aphasia (Dronkers et al. 2004; Binder 2015, 2017), we propose that F3t, STS3 and STS4, regions strongly activated and asymmetrical during sentence processing in different modalities and hubs of the SENT_CORE network, are the epicentres of sentence comprehension.

Because there was a low inter-individual consistency in the segregation of the networks labelled SENT_VISU and SENT_MEM, we will discuss the potential role of the 14 hROIs that were not part of SENT_CORE together.

Since each event of the sentence tasks included task monitoring, such as shifting between the word list and sentence tasks when a picture was presented or providing a motor response at the end of sentence processing, the involvement of executive areas was expected. Conjunction analyses indeed revealed that, in addition to the anterior insula at play in the three tasks (see above), putamen hROIs were involved and are a key neural support of executive functions and task monitoring (Monchi et al. 2006; Sefcsik et al. 2009). Moreover, a recent meta-analysis involving connectivity modelling showed that the left and right putamen areas are different in terms of their respective co-activations, which are specifically co-involved in language areas (Viñas-Guasch and Wu 2017).

The posterior cingulate, precuneus, and paracentral lobule together with the posterior hippocampus are part of the DMN, which has been shown to be involved in both episodic thinking and processing of “self”. At rest, these areas constituted a network that had a very strong and negative correlation with SENT_CORE, confirming that, although they belong to different networks, they are related at rest. Task-induced activation studies have demonstrated that posterior DMN regions are part of mind-reading areas together with the angular gyrus (Spreng et al. 2009). Activated and leftward lateralized during language tasks, this set of areas likely interacts with SENT_CORE during language tasks to process the sentence content dealing with social interactions (Hervé et al. 2012). The angular gyrus is a likely candidate for such an interaction since it segregated with SENT_MEM areas in two clustering methods and is involved in both language and theory-of-mind tasks (Wang et al. 2017).

The fact that these posterior areas were segregated with the posterior hippocampus, involved in both the actual perception and the encoding of scenes (Zeidman and Maguire 2016), suggests that the SENT_MEM network is also involved in the image processing of the drawings, a component common to the three sentence tasks. In fact, all participants performed well in the recall of these sentences (more than nine out of ten images recalled).

Finally, three hROIs were located along the occipito–temporal junction. The posterior part of the inferior temporal lobe (T3_4), the inferior occipital gyrus (O3_1) and the mid-fusiform gyrus (FUS4), with activity and leftward asymmetry, are likely related to the processing of the sentence content in relation to the one-second drawing presentation (minus the scramble version of the word-list reference tasks) common to all tasks. In her meta-analysis, Price considered these areas to be involved in direct visual combinations (Price 2010), consistent with the design of the present paradigm wherein the participants dealt with images related to sentence content regardless of the sentence task modality and the fact that the mean value and asymmetry of this network did not vary with the language tasks. Furthermore, in her 2012 review, Price considered these areas to be involved in sentence processing depending on the task demand (Price 2012). FUS4 is a region of the ventral route that corresponds to the visual word form area (Mellet et al. 2018). It is notable that this region was activated and leftward asymmetrical during the three tasks and segregated with the other visual regions at rest. Such a behaviour is in accordance with Price and Devlin’s proposal that the mid-fusiform area is involved not only in word and visual processing but is a multimodal area whose function is determined by the set of interacting areas (Price and Devlin 2003). FUS4 involvement in the present study is likely to be in the association between the meaning conveyed by the pictures and the sentences and appears to be at play whatever the sentences’ modality. The fact that FUS4 did segregate with visual areas rather than sentence areas at rest supports this hypothesis (Fig. 2). In fact, visual hROIs were not correlated with SENT_CORE networks, meaning that they do not exchange information at rest. However, these areas involved in visual processing, as well as the other hROIs that did not constitute a network at rest with sentence core areas, showed a strong leftward activation during the language tasks, suggesting that their involvement during sentence is, at least in part, related to top-down influences from language networks.

To our knowledge, this study is the first to report a positive correlation between DC and task-induced activation values. Considering that the correlations between DC values and activation were positive for all hROIs, we suggest that the mechanism underlying this result may be that regions more highly connected within the SENT_CORE network are more highly recruited during language tasks. Such a hypothesis is supported by a very recent work conducted with electrocortical stimulation that demonstrated that critical language sites have significantly higher connectivity (Rolston and Chang 2018). It is also important to note that for 8 of the 13 hROIs, no significant effect of the nature of the task on the correlation strength was observed: the positive correlation did not differ regardless of whether the participants produced, read or listened to sentences compared to listing words. Such a relationship was also observed when plotting all SENT_CORE regions of all subjects, as reported by others in a different context (Buckner et al. 2009), meaning that the relationship between resting-state intrinsic connectivity and activation strength remains regardless of the participant and cortical area being considered. In Buckner’s study, the authors hypothesized that beta amyloid accumulates in high DC regions because their high metabolism makes them more vulnerable to the disease, consistent with the report that the DC calculated at the voxel level in the entire hemisphere is correlated with cerebral blood flow values (Liang et al. 2013). Similarly, DC values computed at the hROI level in the SENT_CORE network may also indicate the metabolism of language areas in a given individual, which may be of interest in the evaluation of pathological states. Mesulam et al. reported the selective atrophy of right hemisphere areas in a PPA patient with right hemisphere dominance for language, an observation suggesting that language networks are specifically targeted by this illness (Mesulam et al. 2005). DC measurements in SENT_CORE may thus be a valuable index with which to evaluate inter-individual variations in language area activities in relation to anatomical and clinical patterns in such pathologies.

Previous investigations dealing with the relationships between tasks and the resting state have compared the functional connectivity during cognitive tasks with that measured during the resting state (Cole et al. 2014; Gerchen and Kirsch 2017), as reviewed by Wig (2017). Reports more closely related to the approach used herein have compared resting-state connectivity and hemispheric activation asymmetries obtained during language production in healthy individuals (Joliot et al. 2016) and epileptic patients (Doucet et al. 2014), or during story listening (Raemaekers et al. 2018). Other investigations have found correlations between the task-induced and intrinsic connectivity asymmetries measured in selected sets of hROIs (Liu et al. 2009), such as the set of regions involved in a semantic decision task (Wang et al. 2014). Seeding specific language areas, a previous study involving epileptic patients demonstrated that the functional connectivity measured at rest was correlated with lateralization indices measured during language tasks. In this last work, the intrinsic connectivity most strongly correlated with asymmetry of task-induced fMRI was between the inferior frontal and temporo-parietal language areas, all regions constituting SENT_CORE (Teghipco et al. 2016). The set of consistent findings across the latter studies show that asymmetries of intrinsic connectivity partially explain the variability in activation asymmetries measured during language tasks. However, to our knowledge, there are no previous reports of a relationship between intrinsic DC and activation strength, probably because such a relationship is lacking when comparing whole-brain intrinsic connectivity to variations in activity triggered by cognitive tasks that are underpinned by specific networks. In fact, in the present study, this relationship was observed when the DC computation considered only the 18 hROIs of the SENT_CORE network and disappeared when the DC computation included all of the left hemisphere hROIs. This is an important observation because it underlines the necessity of exploring the properties of intrinsic connectivity within specific networks rather than at the whole-brain level.

Considering that the rs-fMRI acquisition was completed on average 11 months before the language fMRI session herein, it was surprising that the rs-fMRI-derived DC values explained up to 11% of the variance in the language fMRI-derived activation amplitudes in 12 of the 18 SENT_CORE network regions. Thus, DC values at rest in regions constituting SENT_CORE can be considered as proxies of their potential involvement during sentence processing. However, to generalize this observation, we need to both investigate how the SENT_CORE DC is modified in individuals atypical for language and to confirm that such a relationship between the DC value during the resting state and activation strength also exists for networks supporting other cognitive domains, specifically the attentional system, in the same participant.