A large-scale, coordinate-based meta-analysis of task-related deactivations was performed on studies archived in the BrainMap database to identify consistent nodes of the default mode network. Using activation likelihood estimation (ALE) (Eickhoff et al., 2009b
), DMN regions were identified in the precuneus, posterior and ventral anterior cingulate cortices, medial prefrontal cortex, bilateral inferior parietal lobules, bilateral middle temporal gyri, and left middle frontal gyrus. For each DMN node, behavioral profiles were constructed using BrainMap tools to quantitatively assess their functional attributes when active (not during rest), and meta-analytic connectivity modeling maps (MACM) were created to identify dissociable patterns of functional connectivity when unconstrained by any specific task. MACM maps were compared to determine which DMN nodes had the greatest degree of connectivity with other nodes, yielding a meta-analytic model of connectivity between default mode regions. Behavioral profiles of node sets were tested to determine significant functional properties of decomposed subnetworks in this model. Using this technique, affective and perceptual cliques of the DMN were identified, as well as the cliques associated with a reduced preference for motor processing.
The behavioral domain (BD) profiles and MACM maps definitively identified nine functional regions within the DMN. The PCC was observed as the central hub in six of the eight cliques identified in the DMN connectivity model (Table 3), in agreement with other studies that this cortical region is a critical node in the DMN (Raichle et al., 2001
; Grecius et al., 2003
; Fransson and Marrelec, 2008
). BD profiles revealed that even during tasks, functional specialization of DMN regions is limited: significant preference for two or fewer mental operations was observed for six regions. The precuneus and the right middle temporal gyrus (BA 39) displayed no domain preferences, while the PCC displayed the second least complex profile, with a simple decreased preference for action. In contrast, the RIPL and LIPL showed considerable functional specialization, with nine to fifteen significant domains, which suggests a different role within the DMN for these nodes.
In the comparison of individual MACM maps, the PCC and RMTG displayed the highest levels of functional coactivation with other regions, giving further confirmation of their crucial roles in the DMN. There was some, but not overly extensive, overlap across maps for the other seven regions. For all MACM maps, our analysis procedure was structured to first identify regions that are consistently deactivated
during tasks, and then analyze their functional connectivity across experiments in which they were observed to be active
. It has been postulated that characterizing the functional roles of DMN regions may be best undertaken by studying how these regions interact with other regions (Uddin et al., 2009
). The idea that the function of the DMN can be simplistically determined, or that each node contributes equally to this function is unlikely. This reasoning led us to investigate the functional specialization of the DMN regions with both modular and sub-network approaches. We believe that the MACM maps represent generalized functional connectivity for each region, which includes some component of interaction with other default mode regions as well as other components of other networks. Although it has been observed that connectivity of the DMN persists in passive and active tasks states (Greicius and Menon, 2004
; Fransson, 2006
; Buckner et al., 2009
), these regions are also known to be involved in various other behaviors in addition to default mode functioning. The inclusion of both DMN and non-DMN regions in the MACM maps is therefore not surprising.
Collective review of the MACM and BD profile results revealed that the degree of connectivity between DMN regions (i.e., number of coactivations across DMN regions) was inversely correlated with the complexity of their behavioral domain profiles (i.e., number of domain peaks) (P
<0.024); correlation was computed using square roots of the values to reduce the skew associated with the large number of peaks in the RIPL and LIPL. This observation indicates that the more critical default mode regions exhibit functional non-specialization, while more highly specialized nodes exhibit a reduced degree of default mode connectivity. Particularly, the RIPL and LIPL displayed low connectivity with other DMN regions. In contrast, these two regions exhibit strong connectivity with regions identified in a large-scale meta-analysis of executive function tasks (Minzenberg et al., 2009
), which includes regions that have previously been identified as being anti-correlated with the DMN (Fox et al., 2005
; Fransson, 2005
; Uddin et al., 2009
). This leads us to speculate that bilateral IPL are dynamic, bimodal regions that are self-referential during rest (consistent with significant behavioral profiles in interoception and somesthesis), and, upon receiving input from external stimuli, transition to a more extrospective functional role during the execution of goal-directed behaviors (consistent with significant behavioral profiles in action and attention). This type of transfer or facilitative node has previously been observed in association with resting state networks (Seeley et al., 2007
; Sridharan et al., 2008
; Uddin et al., 2009
), and may be a universal component of neural network architecture.
Application of the ALE method was a critical step in the analysis of concordance across studies; other coordinate-based meta-analysis methods are conceptually similar (Wager et al., 2007), and are likely to produce similar results (Salimi-Khorshidi et al., 2009
). This preliminary effort towards functional labeling requires further evaluation and may be limited by a potentially significant lack of data in BrainMap, in terms of the database sample size, study distribution, and the specificity of the behavioral taxonomy. Negative results for BD profiles in several DMN regions may not indicate a lack of functional specialization, but rather a lack of relevant dimensions of the taxonomy. However, analysis of regions outside this network conducted during development of the method consistently revealed more complex profiles, suggesting that many DMN regions actually do differ in their uniform domain distributions. We acknowledge that this analysis would yield more precise results if the granularity of the behavioral domains was increased. Testing for taxonomy-based differences in BD profiles may be a useful strategy for developing and validating a data-driven ontology of behavioral domains.
The behavioral profile analyses potentially provide a systematic method for evaluating the many-to-many mappings of brain regions to mental functions (Price and Friston, 2005
). Here, we pursued a modular formalism of networks by analyzing the behavioral profiles of individual regions, but also investigated a network-focused approach by analyzing cliques of multiple brain regions. This illustrates our first efforts at performing quantitative functional labeling of regions using the BrainMap database. Our goal is to establish a method for creating a probabilistic brain atlas (Laird et al., 2009
), similar to anatomical labeling using the Talairach Daemon (Lancaster et al., 2000
) and probabilistic cytoarchitectonic atlases (Eickhoff et al., 2005
). Such a tool could be useful in interpreting the observed results of any given functional neuroimaging study, and may potentially reduce the generally non-data-driven and impressionistic naming that has been heretofore applied when identifying networks of brain regions.
We investigated the DMN since this network appears to represent an archetypal mode of brain function. However, the general analysis strategy illustrated here can similarly be applied to unpack other modes of function, such as the resting state networks identified by ICA (DeLuca et al., 2006
; Damoiseaux et al., 2006
). The joint strategy of applying MACM and BD profile analyses yields valuable connectivity information in conjunction with a schema for functional labeling, and is unique to other techniques in its ability to assign functional attributes to distinct network circuits. Recent results indicate that resting state networks derived via ICA of resting state fMRI data explicitly match activation networks derived via ICA of BrainMap coordinate data (Smith et al., 2009
). These provocative results highlight the need to fully decompose and define RSNs since they potentially represent a foundational building block of the brain’s functional framework.
Other future work will involve testing the meta-analytic model generated here as a proposed model of effective connectivity in resting state data. If the model presented in is found to match the covariance structure of the DMN in resting data, which is reasonable given the recent results of Smith et al. (2009)
, then the analysis detailed in this study may be useful in developing more informed a priori
models of brain circuitry. Determining a priori
models is one of the most complex stages of structural equation modeling (SEM) or dynamic causal modeling (DCM), and the development of a method that refines this procedure in an unbiased fashion would represent significant progress for the community. If valid, this method could then be applied to determine how the DMN is disrupted in pathologies that are known to effect connectivity, such as Alzheimer’s disease and schizophrenia (Broyd et al., 2009
). Thus far, neuroimaging studies have analyzed temporal or spatial covariances separately, with little effort being made to integrate the results. Generating effective connectivity models using MACM to be tested in fMRI time series data offers the opportunity to link temporal and spatial covariance analyses, and may yield significant new insights into both resting and task-based brain activity.