The cognitive relevance of resting-state networks (RSNs) is poorly understood. To resolve this issue, we propose the Groupe d’Imagerie Neurofonctionnelle Network Atlas (GINNA), a comprehensive RSN atlas derived from the resting-state data of 1,812 participants, providing an exhaustive cognitive characterization of the human brain into 33 distinct networks reliably detected at the individual level.
We systematically analyzed the topographical similarity between GINNA networks and meta-analytic maps extracted from the Neurosynth database (Yarkoni et al., 2011). Although the method we propose relies on a simple measure of spatial similarity using Pearson correlation, here, we demonstrate its usefulness for investigating the cognitive processes potentially linked to RSNs. By relying on an approach of quantitative meta-analytic decoding of cognitive terms related to GINNA RSNs, we provide, to the best of our knowledge, the first empirical cognitive characterization of RSNs. Comparing task-derived meta-analytic maps from the literature and RSNs is particularly relevant if we consider that RSNs represent the prospective exploration of an available repertoire of cognitive functions (Deco et al., 2013). In addition, in the context of multivariate pattern analysis, decoding from Neurosynth-derived maps has been shown to perform similarly to more complex, multivariate decoders (Jabakhanji et al., 2022), and to allow to decode from short blocks of task fMRI the cognitive domains that a single participant was engaged in (Wegrzyn et al., 2018). This suggests that despite its simplicity and low computational cost, decoding based on topographical similarity with activation maps is theoretically justified.
To date, existing brain network parcellations that provide cognitive labeling (e.g., Yeo et al., 2011) have performed an association of function from visual similarity with networks obtained using task paradigms, a method that has proven to result in poor identifiability of RSNs (Uddin et al., 2023). Alternatively, it has been suggested that networks should be labeled according to their anatomical profile (Uddin et al., 2019). GINNA RSNs are provided with an anatomically grounded taxonomy accompanied by suggested cognitive process(es) to reconcile both views. Previous attempts of empirical cognitive characterization of RSNs have relied on topographically similar task-based networks as a proxy (Laird et al., 2011), or have done so with respect to broad cognitive domains extracted from BrainMap (Anderson et al., 2013). By contrast, our approach allows the direct assessment of networks obtained at rest, and benefits from the single cognitive term precision enabled by Neurosynth.
Positioning RSNs cognitive characterization with respect to specific cognitive processes is a much-needed endeavor for the field of cognitive neuroscience. For one, this referencing to well-defined psychological constructs allows to empirically test the predictions we make for each RSN, contrasting with broad cognitive domains (“visual”, “control”) that are not always informative. Second, as the rationale behind the inference of functions rests upon comparison with the neuroimaging literature, the present cognitive characterization is effectively an accurate summary of the current knowledge in both the conceptualization of cognitive concepts (as reflected by the terms present in studies and extracted by Neurosynth), as well as their brain underpinnings (as reflected by the topography of the meta-analytic maps). As such, if the goal is to understand the brain organization of cognition, it may prove more useful to describe RSN putative processes in terms related to those used in cognitive theories (e.g., theory of mind, visual perception) rather than using broad terms that do not necessarily relate to any psychological reality (e.g., limbic, visual). Moreover, because almost all attributed processes are referenced in the Cognitive Atlas Ontology (with the exception of “self-referential processing”, for which we found no equivalent), users can refer to the definitions provided in order to disambiguate the meaning of the concept and provide a common ground to all researchers (Poldrack et al., 2011). Cognitive atlas definitions are available at https://www.cognitiveatlas.org/concepts/categories/all.
The proposed atlas contrasts with existing atlases in its granularity; though rarely considered, this finer granularity might prove beneficial. This is supported by evidence showing that when grouping together a large span of networks constructs taken from psychology (e.g., “fear network”, “working memory network”) into higher-order, large-scale networks, markedly dissimilar cognitive processes become regrouped together (Thompson & Fransson, 2017). Additionally, while functional lateralization of brain circuits is a crucial organizational principle of the human brain, most existing atlases propose networks that are organized bilaterally. Bilateral RSNs are still observed, but the finer granularity in GINNA also resulted in the fragmention of some bilateral networks into homotopical counterparts. For instance, the hand somatomotor system is fragmented into two homotopical systems that correspond to the somatomotor homunculus of left-hand and right-hand motricity. The same can be observed for other networks, such as the FrontoTemporoParietal networks, that fragment into left and right counterparts, associated with distinct processes (e.g., L-FTPN01: sentence comprehension and R-FTPN01: self-monitoring-theory of mind). The identification of lateralized RSNs in GINNA supports the idea that they represent relevant functional units.
The cognitive processes attributed to GINNA RSNs range from low-order sensorimotor (visual, auditory, sensorimotor), up to more integrated, higher-order domains (decision-making, control, memory, social cognition, language, executive). This high diversity suggests that the GINNA atlas covers an extensive share of the known human cognitive repertoire. Of note, only 3 RSNs could not be significantly associated with any Neurosynth term. More importantly, though different in many aspects from existing atlases, some RSNs of the presently proposed atlas align with some of the main large scale networks described in the literature (Uddin et al., 2019). The closest resemblance is observed for RSNs related to the visual (ON-01 to ON-04, OTN), somatomotor (PcN-01 to PcN-03, L-PcN, R-PcN), and default mode systems (med-TN, med-FPN, pCing-medPN). D-FPN-03 - RSN08 associated with selective spatial attention corresponds to the dorsal frontoparietal network linked to attention. The mCingInsN – RSN22, which is associated with performance monitoring, resembles the MidCingulo-Insular network (Uddin et al., 2019), commonly referred to as the salience network. Performance monitoring implies detecting errors and conflicts during tasks and signaling the need for cognitive control adjustments, a role that seems in accordance with the functional definition of the salience network (Seeley, 2019; Seeley et al., 2007).
For a set of regions to be significantly associated with a given cognitive process, it is important that this association exhibits some specificity for this process: any set of regions that would systematically engage in many other tasks could lose its specificity and fail to be significantly more associated to a term than the others. Interestingly, this was the case for three networks (pCing-medPN- RSN01, R-FTPN-03 - RSN02, and ON-3- RSN26). The fact that two of these networks were the most consistently detected RSNs across all individuals may indicate their prime importance in diverse cognitive activities, although their exact contribution remains to be determined. pCingmedPN is a subpart of the classically defined default-mode network (Menon, 2023), and R-FTPN-03 shows a substantial overlap with the 'Multiple Demand (MD) system' (Duncan, 2013), although the latter is usually reported as bilateral rather than right-lateralized as in our case.
Detailed investigation of the terms decoded for well-studied networks, namely, default mode and language networks, highlights the precision of the method and its accordance with the literature. The cognitive labels that we associated with the DMN in its canonical definition (here, med-FPN - RSN07), almost exactly match the cognitive functions reported to be associated with increased activity within its nodes, namely autobiographical memory, self-referential cognition, and theory of mind, as recently reviewed (Menon, 2023).
The three networks that we uncover as related to language processes (L-FTN - RSN15, L-FTPN-01 - RSN19, TN-02 - RSN33) summarize well the current state of knowledge of the brain supports of language processes, and reveal the superiority of the meta-analytic decoding over visual attribution of function. Indeed, though they share a consequent amount of overlap, each RSN’s unique topographical pattern allows the segregation of their associated processes. Despite the overlap in the left superior temporal gyrus (STG), only the TN-02 – RSN33 is bilateral and, therefore, is associated with more perceptual aspects of speech, in line with the highest phonological specificity for the bilateral STG (Turker et al., 2023). L-FTPN-01 – RSN19, that we associate to sentence comprehension, encompasses regions that situated along the inferior bank of the superior temporal sulcus, the temporal pole, the angular gyrus, the left frontal pole, and the left superior frontal gyrus, all reported to be associated with semantic processing (Turker et al., 2023). This network is very similar to the core network of the SENtence Supramodal Areas AtlaS (SENSAAS) describing the essential areas for sentence reading, listening and production (Labache et al., 2019).
The present work is not without limitations. Our approach inherits all shortcomings from performing decoding from a meta-analytic database of task studies. Namely, our study is anchored in the risks associated with reverse inference: it cannot be concluded that because a cognitive process P engages a given brain region R, the activity in R implies the presence of the cognitive process P (R. Poldrack, 2006). In other terms, inferring cognitive processes to RSNs by analyzing their spatial similarity with task activations does not provide evidence of an explanatory relationship, but rather, of a coarse associative one (Mill et al., 2017).
Another limitation is related to the nature of the maps in the Neurosynth database: all task-fMRI studies proceed by contrasting some condition of interest to a control condition. By relying on this assumption of pure insertion (R. A. Poldrack & Yarkoni, 2016; Sternberg, 1969), any process that would be shared by the task of interest and the control condition would be masked out. Moreover, most task-based studies report results at the level of a restricted set of active brain regions or regions of interest. As our observed correlations rarely indicate a near-perfect match between RSNs and meta-analytic maps, our analysis does not allow to firmly determine that the decoded processes are implemented at the whole-network level, as opposed to a (subset of) region-level. The neural context hypothesis proposed that the functional relevance of a brain region relies on its co-activation with other brain regions (McIntosh, 2000). This means that a given region, reported to be implicated in, e.g., working memory, may perform markedly different computations when inscribed in a larger network comprising regions related to, e.g., language. Similarly, there is no guarantee that discrete cognitive processes map onto discrete brain representations. Therefore, it remains to be determined whether RSNs correspond to the prospective exploration of specific cognitive functions or, alternatively, to lower-level “cognitive building blocks” that cannot be isolated from the contrast logic,
We decided to rely on the qualitative attribution of networks’cognitive labels based on an expert consensus procedure. Establishing the relationship between the terms would necessitate a cognitive ontology (Francken et al., 2022; R. A. Poldrack & Yarkoni, 2016) that has yet to emerge despite significant efforts pushed in that direction (the most developed one being the Cognitive Atlas; Poldrack et al., 2011). As a consequence, and because there is to date no clear understanding of how distinct cognitive processes relate to one another (R. A. Poldrack & Yarkoni, 2016), a data-driven clustering of cognitive terms into broader cognitive domains (see, for instance, Wegrzyn et al., 2018), though it may provide an easily interpretable solution, is likely to be imperfect. The reader is invited to confront his/her own interpretation of the inferred processes attributed on the basis of the available results to the one proposed here. The fact that RSNs are labeled with respect to the Cognitive Atlas processes makes the present propositions amenable to further validation through empirical testing.
From the statistical standpoint, although we employed a method of spatial autocorrelation-preserving null hypothesis modeling that effectively reduces false positive rates as compared with spatial naive models (e.g., random shuffling of voxels), slightly inflated false positive rates may remain (Markello & Misic, 2021), leaving room for further methodological developments.
Finally, it is worth noting that the employed methodology does not account for the relevance of the dynamics in the expression of resting-state networks. Several studies have demonstrated that RSNs that appear over the course of relatively long resting-state acquisitions are, in fact, superordinate approximations of underlying dynamic states (Ciric et al., 2017; Sporns et al., 2021; Tagliazucchi et al., 2012). Therefore, the exact cognitive relevance of RSNs, seen as a prospective exploration of cognitive states (Deco et al., 2013), might be better understood in light of their instantaneous interactions with the rest of the brain, as observed at any given time.
Overall, we provide the Groupe d’Imagerie Fonctionnelle Network Atlas (GINNA), a 33 resting-state networks atlas of the human brain grounded in a meta-analytic decoding-based characterization of its cognitive relevance. Each resting-state network’s cognitive relevance is provided in terms of well-defined cognitive processes taken from the Cognitive Atlas ontology, and, as such, should represent better guides for future investigations. The atlas covers a broad spectrum of the human cognitive repertoire, with processes that align with brain laterality and display high associative precision. Potential use cases for GINNA include the selection of a priori regions of interest belonging to a specific network for neuroimaging analyses in the absence of task-derived functional data acquisition, as well as using the maps to analyze a posteriori whether significant regions/edges are distributed within specific cognitive systems.