Functional parcellation of the default mode network: a large-scale meta-analysis

The default mode network (DMN) consists of several regions that selectively interact to support distinct domains of cognition. Of the various sites that partake in DMN function, the posterior cingulate cortex (PCC), temporal parietal junction (TPJ), and medial prefrontal cortex (MPFC) are frequently identified as key contributors. Yet, it remains unclear whether these subcomponents of the DMN make unique contributions to specific cognitive processes and health conditions. To address this issue, we applied a meta-analytic parcellation approach used in prior work. This approach used the Neurosynth database and classification methods to quantify the association between PCC, TPJ, and MPFC activation and specific topics related to cognition and health (e.g., decision making and smoking). Our analyses replicated prior observations that the PCC, TPJ, and MPFC collectively support multiple cognitive functions such as decision making, memory, and awareness. To gain insight into the functional organization of each region, we parceled each region based on its coactivation pattern with the rest of the brain. This analysis indicated that each region could be further subdivided into functionally distinct subcomponents. Taken together, we further delineate DMN function by demonstrating the relative strengths of association among subcomponents across a range of cognitive processes and health conditions. A continued attentiveness to the specialization within the DMN allows future work to consider the nuances in sub-regional contributions necessary for healthy cognition, as well as create the potential for more targeted treatment protocols in various health conditions.

www.nature.com/scientificreports/ and subgenual cingulate increasingly seen at rest 20 . Alzheimer's patients not only show altered DMN responses at rest, but different task-induced deactivation patterns during a working memory task 21 and regional activation differences within the PCC 22 . In schizophrenia, both resting state and task-based DMN response changes have been associated with positive disease symptoms 3,15,23 . While the aforementioned disease states share the commonality of generally altered DMN function, the findings from these studies also suggest that altered activity in different DMN nodes may be specific to different health conditions; for instance, with the PCC specifically implicated in Alzheimer's and ADHD, and the MPFC in schizophrenia and anxiety 15 . However, despite the growing body of knowledge regarding the roles of these four DMN subregions in both the cognitive and psychiatric domains, it remains unclear whether further subdividing these regions would reveal distinct patterns of activity that show relative differences among one another in their contributions to various cognitive process and mental disorders alike.
Evidence for subregional specialization for a range of cognitive functions has been heavily documented in the MPFC in particular, exemplifying how a region may play different roles in cognitive function dependent upon subregional activation [24][25][26] . For instance, the MPFC as a whole appears to play an important role in self-referential processing, but shows a divergence in activity among its subregions: the ventral MPFC deactivates more during self-referential judgments while the dorsal MPFC activity increases 8 . This ventral-dorsal subspecialization within the MPFC appears to extend to other cognitive domains, with ventral MPFC responsible for emotional processing and dorsal for cognitive function 8 . A recent meta-analysis further delineated regional differences in MPFC's function, identifying anterior, middle, and posterior zones, responsible for episodic memory and social processing, cognitive control, and motor function, respectively 27 . This previous work suggests there may be overlap in function and sub-specialization within MPFC regions 27 , opening the door for further research into the more fine-grained aspects of MPFC's function.
Although these previous lines of investigation into DMN functional specialization have largely focused on the MPFC, the PCC and the bilateral TPJ may also serve a wide range of functions. For example, the PCC is thought to play a key role in focused attention during task-based activities 28,29 and continuous monitoring of internal and external stimuli at rest 2 . It has also been implicated in episodic memory retrieval 30,31 , emotional processing 32 , and self-referential processing 33 . Similarly, the TPJ 34 has been shown to play a role in self-referential processing 4 , and is important for social cognition along with other DMN regions (PCC and MPFC 11,35 ). Less is known about potential functional subspecialization within the PCC and TPJ, and it is possible that these DMN regions may also show intra-regional differences in cognitive processing similar to MPFC 36,37 .
To build a better understanding of whether cognitive processes and health conditions are linked to activation within distinct subunits of the DMN, we used the Neurosynth database and identified core DMN regions based on prior work and anatomical landmarks. We also used tools and methods developed in prior work 38,39 that have successfully characterized the functional substructure of the medial prefrontal cortex 27 and the lateral prefrontal cortex 40 . Specifically, we employed a naïve Bayesian classifier to quantify the association between activation in four particular DMN nodes-the PCC, TPJ, and MPFC activation-and topics pertaining to cognition and health. Our application of these analytical methods to four key regions of the DMN is important because it allowed us to quantify how those DMN regions are functionally associated with cognitive processes and disease states-thus extending and complementing prior efforts to fractionate the DMN 12,36,37,39,[41][42][43][44][45][46] . Our primary analyses focused on two key questions. First, are different psychological functions and disease states preferentially associated with distinct nodes of the DMN? Second, are there functionally distinct subregions within individual DMN nodes?

Results
To determine how the PCC, MPFC, and TPJ coactivated with the rest of the brain, we begun by searching for the term "DMN" in Neurosynth to create a functional mask, which we then constrained to the 4 anatomical regions of interest to the present study (PCC, MPFC, left and right TPJ) using the Harvard-Oxford probabilistic atlas in FSL (Fig. 1A). Note that our TPJ region corresponded to what was labeled as "posterior-TPJ" in a previous study 47 . After generating the DMN mask through functional mapping, we applied k-means clustering to determine whether the four subregions of the DMN showed functional segregation. When evaluating coactivation and functional preference profiles for the whole DMN, our contrasts reflect comparisons across DMN nodes; when investigating functional segregation of each DMN node, contrasts within a single DMN region represent comparisons among the clusters within an individual region. We note that our instantiation of our naïve Bayesian classifier, which was used to determine whether topics predicted activity in a region or cluster, used a fourfold cross-validation scheme, and classification accuracy was assessed by calculating the mean score across all four folds, scored using the area under the curve of the receiver operating characteristic (AUC-ROC). Permutationbased analyses accounted for the variability in the standard deviation values across folds.
Coactivation and functional preference profiles for the DMN regions. Next, we sought to characterize functional similarities and differences across DMN regions. To determine how these four regions within the DMN coactivate with voxels across the brain, we adopted an approach used by previous work 27,40 , and identified voxels with a greater probability of coactivating with each ROI relative to the other regions within the DMN. First, we found that the four ROIs within the DMN coactivate with each other: the PCC and MPFC strongly coactivated, and also showed coactivation with both the left and right TPJ (Fig. 1B). This pattern corresponds with long-standing evidence that coactivation among the various components within the DMN represents a typical characteristic of the network. In addition to the "collective" coactivation pattern among the four ROIs, we also found that the four regions showed coactivation overlap in ways that were not shared among all four ROIs. For instance, relative to the MPFC, the PCC and bilateral TPJ showed stronger coactivation with the hippocam- www.nature.com/scientificreports/ pus. Additionally, we found that relative to the PCC and bilateral TPJ, the MPFC showed stronger coactivation with both the amygdala and the ventral striatum, regions known for emotion processing and decision-making (Fig. 1B). Putting our coactivation results in perspective, these patterns demonstrate that there are both functional similarities and differences that characterize how these four regions within the DMN may operate relative to one another. To further explore functional properties among DMN regions, we performed a meta-analysis to select studies that activated a given ROI, and then trained a naive Bayesian classifier to predict which studies activated the region 27 . From our classifier, we extracted the log odds ratio to determine whether a topic predicts activation in a particular cluster, with positive values indicating that a topic is indeed predictive of cluster activation. We first described functional preference profiles based on cognitive processes, followed by those related to health conditions. Cognitive predictors were limited to 20 psychological topics-words that represent co-occurrence among fMRI study abstracts within the Neurosynth database, e.g., cognitive processes and health conditions-previously shown to be relevant to DMN function. We found that all four regions of interest within the DMN were primarily associated with the topics "social", "decision making", "awareness", and "memory" (Fig. 1C), consistent with previous evidence suggesting DMN involvement across these domains. We also found that activity among the four DMN regions were differentially associated with certain topics: fear, emotion and reward topics were associated with activity in the MPFC; emotion topics were associated with PCC activity; and math, semantics, and reading topics were associated with activity in the left TPJ. Next, we assessed the 8 disorder-related topics to examine whether regions within the DMN were differentially associated with various health conditions as well. We found that some disorders were associated with the MPFC or PCC: topics relating to smoking, eating disorder and depression were associated with activity in the MPFC, and topics pertaining to smoking and Alzheimer's/Parkinson's disease were associated with activity in the PCC (Fig. 1D). These results are consistent with observed coactivation patterns among regions within the DMN, again supporting the notion that there are functional similarities as well as differences among these four regions of the DMN, including function relating to psychiatric and health conditions. Next, we evaluated whether the strengths of associations between a topic and a particular region were significantly distinct compared to other regions within the DMN. Our post hoc analysis suggested that despite their associations, none of the topics showed a statistically greater preference for one region over another within the cat eg ori zat ion sw itch ing lan gu ag e inh ibit ion ari thm etic de cis ion ma kin g rea din g lea rni ng WM pri min g me mo ry em otio n

Functional distinctions within DMN subregions. We clustered individual voxels inside the MPFC,
bilateral-TPJ and PCC based on their meta-analytic coactivation with voxels in the rest of the brain to elucidate more fine-grained functional differences among subregions 27,48,49 . We used silhouette scoring to select optimal solutions for each region, and generated coactivation and functional preferences profiles for each subregion. We describe the results for four regions separately.
Within the MPFC, we identified two clusters based on the silhouette score ( Fig. 3A, left panel): a dorsal cluster ( Fig. 3A right panel, orange) and a ventral cluster ( Fig. 3A right panel, green). Our analysis did not reveal that either cluster coactivated more strongly with the rest of the brain. When exploring which topics predicted activation in these clusters, we found that both the dorsal and ventral MPFC clusters were associated with social, emotion, reward, decision-making, awareness, and memory; whereas the ventral cluster was associated with fear ( Fig. 3B). Additionally, both clusters in MPFC were associated with depression and smoking, and the ventral cluster was associated with eating disorders ( Fig. 3C; again, with no significant difference between the strength of association among cluster and topic associations).
Our silhouette score analysis revealed that a three-cluster solution was optimal for the right-TPJ (Fig. 4A, left panel): a dorsal cluster (Fig. 4A, green), a ventral one (Fig. 4A, yellow) and a posterior cluster (Fig. 4A, purple). All three clusters were strongly coactivated with the left-TPJ (Fig. 4B). The ventral cluster showed more coactivation with the PCC. Although topic-cluster associations did not differ amongst each other, activity in all three clusters were associated with social, memory, and awareness topics, with decision making associated with the dorsal cluster (Fig. 4C). No subregions showed any association with disorder-related topics (Fig. 4D). . We directly contrasted the coactivation patterns of each subregion. We found all three regions strongly coactivate with the right-TPJ. In addition, the posterior and ventral clusters showed stronger coactivation with the PCC (Fig. 5B). While both clusters also strongly coactivated with the MPFC, the posterior cluster showed strong coactivation with the ventral MPFC, whereas the ventral cluster coactivated with the dorsal MPFC more (Fig. 5B). Consistent with functional patterns typically associated with the DMN, all clusters in the left-TPJ were primarily predicted by social, decision-making, memory, and awareness (Fig. 5C). We found that the ventral cluster demonstrated an association with reading, semantics and emotion; the anterior cluster showed an association with reading and working memory; and the posterior cluster was associated with priming ( Fig. 5C). No subregions within the left-TPJ showed any association with disorder-related topics (Fig. 5D). Note that the parcellations within the left-and right-TPJ partially mirror each other (Table 1), which increases confidence in the clustering of coactivation patterns 27,40 .
We identified three clusters within the PCC with silhouette scoring (Fig. 6A, left panel): a dorsal cluster ( Fig. 6A right panel, blue), a medial cluster (Fig. 6A right panel, yellow) and a ventral cluster (Fig. 6A right panel,  red). The medial cluster showed stronger coactivation with the bilateral-TPJ and MPFC, whereas the ventral cluster more strongly coactivated with the hippocampus (Fig. 6B). Similar to this coactivation pattern, all clusters in the PCC were predicted by memory, awareness, and decision-making, with the medial region showing an association with social cognition and emotion topics (Fig. 6C). When looking at the disorder topics, we found that all PCC clusters were associated with Alzheimer's/Parkinson's disease; the dorsal cluster was associated with smoking; and the ventral cluster was associated with PTSD ( Fig. 6D), although these cluster-topic associations between the dorsal and ventral clusters were not significantly different in strength compared to each other.

Discussion
Previous research has linked the default mode network to a wide range of cognitive functions and disease states. Numerous studies have elucidated the functional architecture of the network through coactivation-or connectivity-based analysis 36  www.nature.com/scientificreports/ functions and different nodes within the default mode network remains in its early stages. To address this question, the present study builds upon previous work that used a high-powered meta-analysis of functional neuroimaging data to parcellate the MPFC and LFC by topic modeling of psychological functions 27,40 . Our findings help further characterize the functional specialization and subspecialization in two other key DMN regions, the PCC and TPJ. Here, we report coactivation among DMN regions and between the DMN and hippocampus, amygdala, and striatum. Overlap in function among DMN regions was observed with all four ROIs for social cognition, decision-making, awareness, and memory. DMN regions were also differentiated by function: the MPFC was associated with emotion, reward, and fear; the PCC was associated with emotion; and the left-TPJ was associated with math, semantics, working memory, and reading. Moreover, we found that two of the DMN regions were associated with psychopathology and neurological diseases: we found that the PCC was associated with Alzheimer's, Parkinson's disease, and smoking and the MPFC was associated with smoking, eating disorders and depression. Further examination of these DMN ROIs revealed that they could be divided into subregions based on cognitive and disease-related functional subspecialization. Across all four DMN nodes, we observed consistent patterns of involvement with social cognition, memory, decision making and awareness. Our analyses also showed strong coactivation patterns between PCC and TPJ, suggesting their interactive roles in support of functions, such as social cognition and memory processes 36,37,47 . These findings recapitulate the notion that nodes within DMN jointly contribute to multiple psychological processes 7,14,45 . In addition, our results also revealed that the MPFC, as compared to other nodes, has more heterogeneous functional characteristics. We show that the MPFC coactivates with subcortical regions, such as the amygdala and ventral striatum. Consistent with this pattern, our analysis further shows that the MPFC is strongly associated with fear, emotion and reward, consolidating the vast literature on MPFC's interactions with regions related to affect and reward learning 11,52 .
Within the MPFC, our finding of a strong association between the ventral cluster and fear is consistent with prior work implicating vmPFC in extinction of conditioned fear 53,54 . Our analyses further delineate psychological disorders associated with the MPFC. In particular, depression, which has been previously associated with altered www.nature.com/scientificreports/ DMN patterns 55 , loaded most strongly onto the MPFC. In patients with major depressive disorder, increased functional connectivity of the MPFC has been observed 56 , as well as increased MPFC activation during selfreferential processing 57 . The fact that both dorsal and ventral MPFC clusters were associated with depression lends further support to previous work suggesting involvement of both MPFC subregions in major depressive disorder, with dorsal MPFC activation during depressive self-comparisons and ventral MPFC activation during the attentional component of depressive self-focus 58 . Consistent with previous literature on the PCC, the present study shows associations between the PCC and social cognition. It has been shown that the PCC is a central node of the DMN for social cognition -e.g., ascribing mental states to others 11,12,59 and social feedback 60 -and has functional interactions at rest with inferior parietal and superior temporal regions 11 . While previously the role of the PCC in social cognition has not been well-delineated, our results suggest that a medial cluster of the PCC may be most strongly associated with  www.nature.com/scientificreports/ social cognition 61 . The same medial PCC cluster was shown to be specific to emotion; while there is relatively less evidence for the role of PCC in emotion, previous work has shown that the PCC activates in response to emotional words and may mediate interactions between emotional and memory processes 32 , consistent with our parallel finding of an association between the PCC and memory 62 . In addition to a primary association with memory, all PCC clusters loaded on Alzheimer's/Parkinson's disease, consistent with previous reports of PCC-DMN functional connectivity in patients with Alzheimer's disease and mild cognitive impairment 63 . This association may also be linked to our finding of PCC-hippocampal coactivation, as previous research has suggested disrupted resting state functional connectivity between PCC and hippocampus as a mechanism underling Alzheimer's disease 63 . Our analyses also highlight the association of both the dorsal and ventral PCC with smoking. Although most prior neuroimaging studies of smoking behavior focus on anterior cingulate and subcortical circuitry, nicotine has also been shown to consistently enhance functional connectivity between the PCC and medial frontal/ anterior cingulate cortex, as well as local connectivity between the dorsal and ventral PCC 64 . Additionally, PCC activity has been associated with craving 65 , viewing smoking cessation messages 66,67 , and suppressing cueinduced craving 68 . Taken together, these functional distinctions indicate that subregions within the PCC were differentially recruited by different cognitive processes, a pattern consistent with previous literature suggesting the multifaceted role of the PCC in cognition 43,69,70 .
We also report involvement with social cognition within the TPJ, with the left TPJ more strongly associated with math, semantics, reading, emotion, and working memory. Within the domain of social cognition, previous research has established the TPJ as most associated with mentalizing/heory of mind functions 71 . In addition, previous work has suggested functional heterogeneity within the TPJ on the basis of its functional and structural connectivity 11,41,72 . One study in particular attempted to map social cognition in the human brain, including parcellating the TPJ using diffusion-weighted imaging with comparison to non-human primates 47 . This work suggested that the posterior TPJ was most strongly associated with social cognition. Our analysis similarly Although our study provides a comprehensive characterization of the functional roles of the DMN, we note that our findings should be viewed with three caveats. First, the classifier used in our analysis did not distinguish activations from deactivations. However, it is well established that the DMN might be activated for some processes (e.g., social cognition 11,50,73 ; and deactivated for others (e.g., executive control [74][75][76]. It is thus conceivable that a dataset capable of detecting deactivations would potentially extend our current findings and provide a full account of the functional architecture of the DMN. Second, the coactivation maps may not be directly related to connectivity between brain regions because they are based on correlations. Indeed, correlation between brain regions can be driven by a number of factors that are not related to connectivity or coupling, including changes in signal-to-noise ratio in either region or a change in another brain region 77,78 . A thorough examination of connectivity would necessitate integrating behavioral tasks with effective connectivity measures, such as psychophysiological interactions analysis [78][79][80][81][82] (PPI), an avenue for future investigations. This alternative approach would provide insight into how specific tasks and processes drive connectivity with the DMN. Finally, the nature of the topic modeling oversimplifies the mapping between psychological ontology to complex, dynamic brain activity 40,83 . For example, each topic used in our analysis represents a combination of many cognitive processes operating at different levels. As a result, mappings between specific psychological concepts and brain activity may require further parcellation of complex cognitive processes.
In addition to these limitations, we also note that our approach for defining and fractionating the DMN merits additional consideration. For example, we defined the DMN based on a combination of anatomy and function and then parcellated individual nodes of DMN. Although an analogous approach has been used in prior studies 12 , we note that other studies have parcellated networks using responses from all nodes 61 . Both approaches assume a given network is composed of distinct nodes, which depends critically on the definition of those nodes 84,85 . To address this issue, some papers have defined networks using continuous maps (e.g., those derived from independent component analysis) and have examined connectivity with those maps using dual-regression analysis [86][87][88] . We believe that integrating this approach with our current analytical framework 27,40 and unthresholded whole-brain maps 89 will help future studies refine functional parcellations of the DMN. We also note that in our process of parcellating four regions within the DMN, our clusters had become relatively small, which may have limited our ability to find differences between cluster-topic associations among each DMN node. Future work may be able to address this limitation by using image-based meta analyses 90,91 , which may be possible as more researchers heed calls to share unthresholded statistical maps 82,89,92 .
To conclude, the present study applied a meta-analytic approach to characterize functional mappings among cognitive processes, health conditions, and the DMN. Although the DMN as a whole contributes to multiple cognitive processes, we identified distinct functional properties for each region, and extended these results by identifying further parcellation of functions for specific subregions. These results help clarify the functional roles of the DMN across a large corpus of neuroimaging studies and add to the body of literature that has focused on refining the theoretical and computational framework associated with the DMN 93 .

Methods
Our analysis was based on version 0.6 of the Neurosynth dataset (i.e., articles up to July 2015), which contains activation data from more than 12,000 fMRI studies 94 . DMN mask. First, to find regions that are functionally related but restricted to the four anatomical regions within the DMN that we are interested in, we created a DMN mask that constrained wider functional activation within an anatomical atlas 95 (Fig. 1A). Specifically, we performed a meta-analytical association test by searching the term "DMN" in Neurosynth to create a functional mask that specifically mapped the term "DMN" to brain regions. We chose an association test (formerly named a "reverse inference") because it can help estimate the relative specificity between brain activity and the DMN 94,96 . The resulting mask included 9,650 activations from 366 studies, and thus identified voxels in studies where the term "DMN" was mentioned in their abstract given brain activation. All of the images are thresholded with a false discovery rate of 0.01. Next, we constrained the mask to the four DMN regions of interest by using the Harvard-Oxford probabilistic atlas 97 at P > 0.25, which included the medial prefrontal cortex (MPFC), the posterior cingulate cortex (PCC), and the left and right temporal parietal junction (left-and right-TPJ). This constraint was applied in the interest of specifying the output of our functional mapping and subsequent parcellation to a precise set of standardized coordinates within the four regions of interest.
Coactivation-based clustering. To determine more fine-grained functional differences within the four primary DMN regions of interest, we applied a clustering method used by previous work 27 to cluster individual voxels inside each of the four regions based on their meta-analytic coactivation with voxels in the rest of the brain 38 (Fig. 1B). To assess coactivation patterns for each of the four regions, we correlated the activation pattern of each voxel with the rest of the brain across all 366 studies. The resulting coactivation matrix is large and would require intensive computational power; to mitigate this, the coactivation matrix was passed through a principal component analysis (PCA), where the dimensionality of the matrix was reduced to 100 components 27,40 . Once the coactivation was reduced, we calculated the Pearson correlation distance between every voxel with each whole-brain PCA component in each of the four DMN regions. We then grouped the voxels within each separate region into 2-9 clusters by applying a k-means clustering algorithm to the correlation distance coefficients 98  www.nature.com/scientificreports/ to measure cohesion within each cluster 99 . Previous reports caution that issues may arise when identifying the appropriate number of clusters due to variations in goals and levels of analysis used across investigations 83,100 ; however, recent work has confirmed that the silhouette score can successfully assess cluster solutions 27,40,101,102 . Coactivation profiles. Next, we analyzed how the whole-brain coactivation patterns differed between regions to reveal their functional networks 27 (Fig. 1C). To accomplish this, we began by setting up contrasts among each of the four regions' coactivation patterns (e.g., the MPFC against the PCC, left TPJ and right TPJ) to detect the coactivation differences between each region. Specifically, we performed a meta-analytic contrast between studies that activated the region of interest (ROI; "active studies") and studies that activated the other three regions of interest ("inactive studies"), to identify the voxels in the rest of the brain that had a greater probability of coactivating with a given ROI than other regions within the DMN. For instance, the purple voxels in Fig. 2B indicates that these voxels in particular (in this case, within the amygdala and striatum) tend to coactivate alongside the MPFC rather than the PCC, the left TPJ or the right TPJ. To generate coactivation images that highlight coactivation profiles that emerged as a result of contrasting the four DMN ROIs, we conducted a two-way chi-square test between two sets of studies-the active and inactive studies-and calculated p values to threshold the coactivation images using False Discovery Rate (q < 0.01). The resulting images were binarized and visualized using the NiLearn Library in Python 103 .

Meta-analytic functional preference profiles.
To map between functional states (e.g., cognitive processes and health conditions) and the MPFC, PCC, and TPJ regions of the DMN, we re-purposed a set of 60 topics-generated as a part of a previous study 27 -which represent words that co-occur among fMRI study abstracts in the Neurosynth database. The set of 60 topics, which were produced in the exact same manner as seen in Poldrack et al. 105 ), were derived from Dirichlet allocation topic modeling 104 , which is beneficial in reducing the level of redundancy and ambiguity in term-based meta-analysis maps in Neurosynth 27 . Per Poldrack et al. 105 and De La Vega et al. 27 , topics that were irrelevant to either cognitive processes or health conditions (for instance, topics pertaining to experimental methods and design; N = 23 topics) were excluded from the original set of 60 generated topics. After the exclusion step, we were left with 37 topics: 29 cognitive topics (e.g., mentalizing, attention, imagery) and 8 disorder-related topics (e.g., PTSD, ADHD, Autism). " We generated functional preference profiles by identifying which cognitive processes or health conditions best predicted activity within a given cluster across studies (Fig. 1D). Considering the redundancy and ambiguity in term-based meta-analytic maps (e.g., "memory" could refer to working memory or episodic memory), we followed the lead of Yarkoni et al. (2011), and trained a naive Bayesian classifier 27 to evaluate two sets of studies. One set of studies included those that activated a cluster and the other set did not; a study was defined as activating a given cluster if at least 5% of voxels in the cluster were activated in that study. The classifier was then trained to predict whether a study activated a particular cluster, given the semantic representation of the latent conceptual structure underlying descriptive terms featured in the study 27,40 . Our instantiation of the classifier employed a fourfold cross-validation scheme, the default variance smoothing in scikit-learn (i.e., 1e−09), and did not utilize any priors. Input variables included the 60 psychological and health topics to predict whether activation is likely given the latent concepts within a study. As a final measure of our model's performance, we calculated the mean score across all four folds, and scored them using a classification summary metric that considers the sensitivity and specificity: the area under the curve of the receiver operating characteristic (AUC-ROC).
Next, we extracted the log-odds ratio (LOR) of a topic, defined as the log of the ratio between the probability of a given topic in studies that activated a given cluster ROI and the probability of the topic in studies that did not activate the same cluster ROI, for each cluster separately. A positive LOR value indicates that a topic is predictive of activation in a given cluster. Based on LOR values, we identified 20 cognitive topics that loaded most strongly to the entire DMN mask for further analysis and then applied a procedure used in a previous study to determine the statistical significance of these associations 27 . To do so, we performed a permutation test for each region-topic log odds ratio 1,000 times. This resulted in a null distribution of LOR for each topic and each region. We calculated p values for each pairwise relationship between topics and regions and then adjusted the p-values using a False Discovery Rate at 0.01 to account for multiple comparisons within 20 selected cognitive topics and 8 disease-related topics (separately). We reported associations significant at the corrected p < 0.05 threshold. Finally, to determine whether certain topics showed greater preference for one region or cluster versus another, we conducted exploratory, post hoc comparisons by determining whether the 99.9% confidence intervals (CI) of the LOR of a specific topic for one region or cluster overlapped with the 99.9% CI of the same topic for another region. We use a more stringent threshold than that of previous work 27 due to the increased number of comparisons we make across DMN regions. We used bootstrapping to generate CIs, sampled with replacement, and recalculated log-odds ratios for each cluster 1,000 times.

Data availability
All data are publicly available on Neurosynth database (https ://neuro synth .org).