Beyond the Tripartite Cognition–Emotion–InteroceptionModel of the Human Insular Cortex

Lucina Q. Uddin1,2*, Joshua Kinnison3*, Luiz Pessoa3**,and Michael L. Anderson3,4**

Abstract

■ Functional MRI studies report insular activations across a widerange of tasks involving affective, sensory, and motor processing,but also during tasks of high-level perception, attention, andcontrol. Although insular cortical activations are often reportedin the literature, the diverse functional roles of this region arestill not well understood. We used a meta-analytic approach toanalyze the coactivation profiles of insular subdivisions—dorsalanterior, ventral anterior, and posterior insula—across fMRI studiesin terms of multiple task domains including emotion, memory,attention, and reasoning. We found extensive coactivation of eachinsular subdivision, with substantial overlap between coactivation

partners for each subdivision. Functional fingerprint analysesrevealed that all subdivisions cooperated with a functionally di-verse set of regions. Graph-theoretical analyses revealed that thedorsal anterior insula was a highly “central” structure in the co-activation network. Furthermore, analysis of the studies that acti-vate the insular cortex itself showed that the right dorsal anteriorinsula was a particularly “diverse” structure in that it was likely to beactive across multiple task domains. These results highlight thenuanced functional profiles of insular subdivisions and are con-sistent with recent work suggesting that the dorsal anterior insulacan be considered a critical functional hub in the human brain. ■

INTRODUCTION

The insular cortex is a functionally heterogeneous brainregion that has been characterized by its involvement insomatic and visceral sensory processes, autonomic regula-tion, as well as motor processing (Augustine, 1985, 1996).Earlier views of the insula as primarily a low-level, “limbic”structure have given way to more complex and multi-faceted views of insular function in recent years. In con-cert, it has been suggested that the anterior portion ofthe human insula may not have an equivalent in the rator monkey brain (Bud Craig, 2009; but see Nieuwenhuys,2012, for alternate views). Furthermore, citing evidencefor substantial variability with respect to extent, shape,gyral and sulcal patterns, and laminar organization of in-sular cortex, a recent study across several species suggests“it is not possible to identify a general model of organiza-tion for the mammalian insular cortex” (Butti & Hof, 2010),although earlier comparative anatomical work suggestssubstantial overlap between the human insula and that ofother hominoids (see Nieuwenhuys, 2012, for a review).These controversies regarding anatomical organizationhave made functional characterization of the insula in thehuman brain all the more challenging.

Recent critical thinking regarding insula function in hu-mans has focused on its role in the experience of emo-tion derived from information about bodily states(Critchley, Mathias, & Dolan, 2001), in line with earliertheories suggesting that signals from the autonomic ner-vous system shape emotional experience (Damasio,1996). Functional neuroimaging studies have shown thatthe right anterior insula (AI) in particular is active duringa wide variety of tasks involving interoception and thesubjective awareness of both positive and negative feel-ings, including anger, disgust, judgments of trustworthi-ness, and sexual arousal (see Craig, 2002, for a review).The AI is also involved in empathy or the “capacity to un-derstand emotions of others by sharing their affectivestates” (Singer, 2006). Whereas the posterior insula hasbeen shown to be activated when participants receivepainful stimulation, AI activates both during pain percep-tion and during witnessing a loved one experiencingpain, suggesting that AI activation contributes to the ex-perience that is linked to the understanding of the feel-ings of others (empathy) and ourselves (Singer et al.,2004). A study using alexithymia and empathy scales toassess emotional awareness of the self and others, re-spectively, reported that difficulties in emotional aware-ness were related to hypoactivity in the AI of bothautistic individuals and controls (Silani et al., 2008).Taken together, these studies reveal a role for the AI insocial and affective processes involving perception andintegration of information from multiple sources.

1Stanford University School of Medicine, Stanford, CA, 2Universityof Miami, Coral Gables, FL, 3University of Maryland, College Park,MD, 4Franklin & Marshall College, Lancaster, PA*,**These two authors contributed equally to this work.

© 2013 Massachusetts Institute of Technology Journal of Cognitive Neuroscience 26:1, pp. 16–27doi:10.1162/jocn_a_00462

In addition to the well-documented involvement of theAI in affective processing, reports of AI involvement incognition are emerging. In a recent special issue of thejournal Brain Structure and Function, the functionalroles ascribed to the human AI ranged from salience de-tection (Menon & Uddin, 2010) to attentional control(Nelson et al., 2010), risk prediction (Bossaerts, 2010),and subjective awareness or consciousness (Craig,2010). Partly because of the regionʼs participation in mul-tiple cognitive tasks, it has recently been argued that theAI is a key node in a large-scale brain network includingthe ACC (Seeley et al., 2007). This “salience network”is thought to detect salient events and initiate switchesbetween networks involved in self-related, internallyoriented processing and those involved in goal-directed,externally oriented processing (Menon & Uddin, 2010;Uddin & Menon, 2009; Sridharan, Levitin, & Menon,2008). It has been proposed that the AI can be consid-ered a “causal outflow hub” because of its role in coordi-nating other large-scale brain networks (Uddin, Supekar,Ryali, & Menon, 2011).One strategy to understanding the functional role of

a brain region is to focus on its functional connectivity.Recent reports have taken advantage of the existence oflarge databases of functional neuroimaging studies (www.brainmap.org/ and neurosynth.org/) to perform coactiva-tion meta-analyses of the human insula. These studiesoperationalize functional connectivity in terms of the ten-dency for different brain regions to be simultaneouslyactive across experiments (Anderson, Brumbaugh, &Suben, 2010; Toro, Fox, & Paus, 2008). A recent coactiva-tion meta-analysis of insular subdivisions suggests that thedorsal anterior subdivision is more involved in high-levelcognitive processes (e.g., switching, inhibition, and errorprocessing), the ventral anterior with affective processes(e.g., emotion and anxiety), and the posterior with sen-sorimotor processes (e.g., pain; Chang, Yarkoni, Khaw, &Sanfey, 2012). Indeed, both resting-state functional con-nectivity (Deen, Pitskel, & Pelphrey, 2010) and task-basedmeta-analytic approaches are beginning to convergeon the finding that the AI can be subdivided into dorsaland ventral (Chang et al., 2012; Kelly et al., 2012; Kurth,Zilles, Fox, Laird, & Eickhoff, 2010) as well as anteriorand posterior (Cauda et al., 2012) subregions.The parcellation suggested by these investigations has

shed light on the multifaceted nature of insular func-tion. However, several open questions remain regardingthe resulting tripartite framework of the AI: a cognitivedorsal AI, an affective ventral AI, and a sensory posteriorinsula. In particular, the putative distinction between adorsal “cognitive” subdivision and a ventral “affective”subdivision is potentially problematic given that a generalcognitive/affective dichotomy has become increasinglychallenged (Lindquist, Wager, Kober, Bliss-Moreau, &Barrett, 2012; Pessoa, 2008). Furthermore, the cognitive/affective distinction proposed for dorsal and ventral AI isnot always clearly observed. One of the first meta-analyses

investigating insular function provided evidence that thedorsal AI is involved in all task domains examined, withthe exception of somatosensation and motion, raisingthe possibility that this region may play a basic func-tional integrative role common to multiple tasks (Kurthet al., 2010). A recent study examining intrinsic func-tional connectivity of the dorsal and ventral AI foundthat, although the regions exhibit distinct patterns ofconnectivity consistent with purported cognitive andaffective functions, the two regions share connectivity withportions of OFC, frontal operculum, ACC, and ventralputamen (Touroutoglou, Hollenbeck, Dickerson, &Feldman Barrett, 2012). Thus, several open questions withregards to the functions of putative insular subdivisionsremain unresolved.

To quantitatively capture and assess the functionalcomplexity of the insula, we adopted and extended thenotion of a functional fingerprint: the observed activityor response properties of a region under various taskconditions (Passingham, Stephan, & Kotter, 2002). Weleverage large amounts of publicly available neuroimagingdata to perform a coactivation meta-analysis of the threefunctional subregions of the human insula. This allowedus to characterize the functional fingerprint of insularsubdivisions, as well as the fingerprint of their coactivationpartners (i.e., “target” regions). The determination of func-tional fingerprints allowed us to also assess functionaldiversity. A brain region with high diversity is engaged bytasks in multiple task domains, whereas a low-diversityregion is more specialized, being engaged by tasks in fewerdomains (Anderson, Kinnison, & Pessoa, 2013; Anderson& Penner-Wilger, 2013). Finally, we computed graph-theoretical metrics of centrality that index the relativeinfluence of each insular subdivision in the brain co-activation network. Combined, our approach allowed usto unravel both commonalities and differences of putativeinsular subregions. In particular, previous studies havenot characterized the degree of functional differentiationand functional overlap of insular coactivation patterns. Un-derstanding the broad, diverse set of brain regions withwhich the insula interacts during task performance is criti-cal to building models of insular function in the healthybrain and can potentially provide insights into disordersresulting from insular dysfunction.

METHODS

Insular Subdivisions

The general methodology for performing coactivationanalyses is discussed in detail in previous publications(see Anderson et al., 2010; Toro et al., 2008). The threemain steps are as follows. (1) choose a scheme for spatialsubdivision of the brain; (2) assign experimental activationsto the subdivisions; and (3) determine which regionsare statistically likely to be active at the same time (activeduring the same brain imaging experiments).

Uddin et al. 17

The insular cortex was defined using maps providedby Deen et al. (2010), who subdivide the insula intothree functional regions based on a clustering analysis ofresting-state functional connectivity of the whole insula(Figure 1). Mean coordinates of these insular subdivisionsare shown in Table 1. Note that although Deen et al.label the posterior-most subdivision as “posterior insula,”this subdivision also includes portions of midinsula re-gions that have been described as functionally distinct inprevious meta-analyses (Kurth et al., 2010).

fMRI Databases

Task-based coactivation was determined using a collectionof neuroimaging studies obtained from the Neurosynthdatabase (Yarkoni, Poldrack, Nichols, Van Essen, & Wager,2011). At the time of access (January 1, 2012), Neurosynthcontained 144,680 activation peaks from 4393 studies.Neurosynth does not categorize studies by domain ortask, and articles are included in the database regardlessof the type of task used or function investigated. We in-vestigated coactivation using both a voxel-wise “search-light” approach as well as an ROI-based approach. Forvoxel-wise analysis, a given voxel was considered activeduring an experiment if an activation peak was reportedwithin 10 mm of its location. When ROIs were considered,

a given ROI was considered active during an experimentonly if an activation peak fell directly in that ROI.To gauge the reproducibility of our results, we employed

an index commonly utilized in the literature (Raemaekerset al., 2007):

Roverlap ¼ 2VoverlapV1 þ V2

where V1 and V2 are the number of voxels significantlyactive in each half and Voverlap is the number of voxels sig-nificantly active in both halves. The ratio can thus varybetween 1 ( perfect reproducibility) and 0 (no reproduc-ibility). Here, we split our data set into two random halves.For the analysis shown in Figure 4, Roverlap values for allsix insula subdivisions (three per hemisphere) rangedfrom .52 to .61. These values are comparable, for instance,with those observed for test–retest Roverlap values of theposterior cingulate cortex—a key node of the defaultmode network—identified using independent componentanalysis of resting state fMRI data, which one studyreported between .49 and .59 (Meindl et al., 2009). Com-parable results were also observed for motor-relatedactivation (Kristo et al., 2012).For the investigation of functional fingerprints and

diversity, we analyzed studies from the BrainMap database(Laird, Lancaster, & Fox, 2005). As there are no widely ac-cepted ontologies of mental processes (Yarkoni, Poldrack,Van Essen, & Wager, 2010; Price & Friston, 2005), weemployed the BrainMap taxonomy, which has under-gone considerable refinement in the past decade (Laird,Lancaster, & Fox, 2009; Fox et al., 2005; Fox & Lancaster,2002). Thirty-two task domains were considered, span-ning perception, action, cognition, and emotion, an ap-proach similar to that employed in recent studies (Laird,Eickhoff, et al., 2009; Smith et al., 2009). All studies con-sidered involved healthy adults and used a within-subject,whole-brain, univariate design. That is, for all the studiesin the data set, brain activity during an experimental taskwas observed over the whole brain and compared voxel-wise to activity observed in the same participant duringa control task. Here, we use the term “observation” to

Figure 1. Insula subdivisions.The insula was dividedfollowing the parcellationscheme reported by Deenet al. (2010) into threesubdivisions: dorsal anterior(blue), ventral anterior(green), and posterior (red).

Table 1. Insula Subdivision Coordinates (from Deen et al.,2010)

Insular Subdivision

Mean Talairach Coordinates

x y z

Left dorsal AI −38 6 2

Left posterior insula −38 −6 5

Left ventral AI −33 13 −7

Right dorsal AI 35 7 3

Right posterior insula 35 −11 6

Right ventral AI 32 10 −6

18 Journal of Cognitive Neuroscience Volume 26, Number 1

refer to the pairing of a reported activation and a taskdomain. For example, for an experiment filed in the data-base under both the “working memory” and “vision”domains (because of the task manipulation), each re-ported activation would count as two observations (oneper domain) at its activation site.

Functional Fingerprint and Diversity Analyses

A functional fingerprint was defined as a 32-dimensionalvector, each dimension corresponding to a task domainfrom the Brainmap database. Each of the 32 values repre-sented the proportion of observations in the correspond-ing task domain (local number of observations dividedby global number of observations), normalized (i.e., all32 values summed to 1; Figure 2). We computed func-tional fingerprints of the coactivation partners of eachinsular subdivision as well as a “common” fingerprintconsisting of the average of all insular subdivisions. Spe-cifically, we constructed fingerprints by considering taskdomains observed in the “target” voxels coactivating witheach insula subdivision. Task coactivation was determinedbetween each insula subdivision and all other voxels inthe brain. To determine a measure of coactivation thatdescribed the relationships of each insula region separatefrom the rest, partial Pearson correlations were computed,which included the influence of the other five insularegions. Correction for multiple comparisons used falsediscovery rate correction (Benjamini, Drai, Elmer, Kafkafi,& Golani, 2001), such that correlations that did not survive(q = 0.05) were set to zero. In addition, negative correla-tions were set to zero.For each insular subdivision, a fingerprint was formed

by pooling all observations falling within the set of sig-nificantly correlated voxels (as if they comprised a single“target area”). The pooling procedure was weighted suchthat the contribution of a voxelʼs observations was basedon the partial correlation value. The weighting procedureinvolved a simple multiplication before pooling the ob-servations. Thus, for instance, a voxel with a partial cor-relation of 1.0 would contribute as normal whereas a

voxel with a partial correlation of .5 would have its obser-vations halved before summing (Figure 3A).

There are many measures of diversity in the literature,particularly in biology and economics (e.g., Magurran,

Figure 2. Determination of functional fingerprints. To illustrate theprocess, only three task domains are shown. The actual fingerprintsused in the paper were 32-dimensional. The label “regional” refersto voxels (via the searchlight) or ROIs. The final normalization stepensures that the fingerprint values all sum to 1.

Figure 3. Schematic diagram of data analysis. (A) Coactivation mapsand functional fingerprints. Using data from the Neurosynth database,task-based coactivations were determined for each insular subdivisionby moving a searchlight in a voxel-wise manner. The blue arrows denotethe strength of partial correlation between an insular subdivision andspecific voxels on the surface. All significant voxels contributed to thefunctional fingerprint of the “partner” regions that coactivated withthe specific insular subregion. Their contribution was proportional tothe value of the partial correlation (as indicated by the green arrowwidths). To determine the fingerprints BrainMap data that include taskdomain classification were used. (B) Network analysis. Using NeuroSynthdata, pairwise correlations between a set of ROIs were computed andused in the graph-theoretical analysis.

Uddin et al. 19

2004). Here, the Shannon diversity, H (Shannon, 1948),of a region of insula was defined as:

H ¼ −XS

t¼1

pi ln pi

where S = 32 was the number of task domains investi-gated and pi corresponded to the ith domain proportion.As Shannon diversity is negatively biased (i.e., it tends tounderestimate diversity), as proposed by Chao and Shen(2003), we used the correction term (S − 1)/(2n) addedto H, where n is the number of observations used in thedetermination of the fingerprint. This correction is sug-gested if n > S; however, the number of observationsfor each insula subdivision varied considerably, andsome were under 32. Thus, we report relative diversityvalues (percentiles) for insula as a whole and for the sub-divisions with sufficient observations to allow a reliableestimate to be made. Note that the lower bound on thenumber of observations is based on statistical consid-erations and does not imply that each subdivision mustbe active in each of the 32 domains. The 32+ observedactivations could be in a single or in several differentdomains.

ROI Generation

ROIs were automatically generated using a spatiallyconstrained clustering algorithm (Craddock, James,Holtzheimer, Hu, & Mayberg, 2012). This method cal-culated voxel-wise correlations among neighboring voxels,while ignoring long-distance correlations, to produce aset of spatially contiguous ROIs. In the present case, weapplied the algorithm to voxel-wise task-based coactiva-tion values. The process requires that one specify a targetnumber k of ROIs to produce. The main results presentedcome from an analysis using k = 160. To reduce “cross-contamination” between ROIs or activations on the bound-aries that could be assigned to either ROI, we eroded theboundaries of each ROI by e mm; thus, the ROIs we inves-tigated were physically isolated from one another.

To ensure that our results were not specific to the pa-rameter choices, we explored other settings in additionto those reported here. For k, results were stable betweenk = 120 and 200. For e, results were stable for erosionsbetween 2 and 5 mm. Overall, the results were relativelyinsensitive to the settings as long as the resulting ROIs were“reasonably sized”: briefly, on the one hand, large enoughto receive enough “hits” from studies and on the otherhand, small enough to ensure proper “granularity” (e.g.,very large ROIs would be functionally diverse simplybecause of their large size).

Subsequently, the six insular ROIs were added to theset generated via clustering, replacing any of these withwhich it overlapped. The hemisphere of each ROI wasdetermined according to its center of mass. Finally, we

only retained ROIs that contained a sufficient numberof observations (32; see above) to allow proper estima-tion of coactivation and functional diversity. Hence, ROIsin the posterior part of the brain were not included,resulting in a total of 97 ROIs in the network analysis(Figure S1).

Network Analyses

Although network analysis of neuroimaging data can beperformed in a voxel-wise manner, it is more commonlyperformed (and more stable) at the ROI level. Here, agraph was formed that comprised the ROIs describedin the previous section. For each ROI, “activity” was ob-served in the region if a reported location of task peakactivation fell within the region. Studies from the Neuro-Synth database were considered. We then performed apairwise ROI-to-ROI quantification of “activity” correla-tion, which captured the tendency of region pairs to beactive during the same experimental task. Specifically,coactivation was defined as the Pearsonʼs correlation ofthe “activation” histories. As before, we applied false dis-covery rate correction and removed negative correla-tions. Note that partial correlations were not employedin the network analysis because the goal here was tounderstand the properties of insular subregions in thecontext of all other ROIs (Figure 3B).We analyzed the resulting coactivation graph both as a

binary and separately as a weighted structure. A binaryundirected graph was constructed such that edges onlyexisted between ROIs with statistically significant correla-tions. A weighted undirected graph was constructed fromthe binary graph by assigning correlation coefficients be-tween ROIs as edge weights (Sporns, 2011). Nodes of abinary graph are primarily characterized by their degree(the number of other nodes they are connected to). Thedegree of a node is one way to measure its centrality in agraph, but there are other interesting measures, too. Forinstance, betweenness centrality evaluates the pathsthrough the graph that a given node lies along and pro-vides an indication of node importance to the overallinformation flow through a network. Similarly, nodestrength is an estimate of the importance of a node ina weighted graph based on the sum of weights of allits connections (Barrat, Barthelemy, Pastor-Satorras, &Vespignani, 2004). To compute degree, betweeness cen-trality, and strength, we used the Connectivity Toolbox(Rubinov & Sporns, 2010; Sporns, 2003).

RESULTS

Coactivation Maps and Functional Fingerprints

Task-based coactivation analysis was conducted to cap-ture the tendency of insular subdivisions to be activetogether with other brain areas during the same experi-mental task. Each of the six insular subdivisions served

20 Journal of Cognitive Neuroscience Volume 26, Number 1

as a seed region (Table 1), and coactivation was deter-mined by moving a searchlight across the brain in a voxel-wise manner (to account for the influence of the otherfive insular subdivisions, partial correlations were com-puted). For each insular subregion, extensive coactiva-tion was observed (Figure 4). Furthermore, coactivationoverlap was extensive (Figures 5 and 6).Next, we determined the functional fingerprint of the

coactivating partners of each insular subdivision. For

example, for the left dorsal AI, we considered “activa-tions” in all voxels shown in Figure 4 (top left), as if theywere a single “region.” Figure 7 displays the “common”fingerprint for the insular subdivisions. To obtain it, wefirst computed six fingerprints (one for each subregion;not shown) and took the minimum value for each taskdomain (to show that each subregion was engaged“at least that much”). The common fingerprint was func-tionally diverse and encompassed all 32 task domains

Figure 5. Overlap betweenconnection partners ofeach insular subdivision.To facilitate displaying overlap,the corresponding right andleft insular subregions werepooled together resulting inthree insular subregions(dorsal anterior, ventralanterior, posterior insula).Voxels shown in green-to-redcolors were coactive withtwo of the three subregions(the color bar indicates thestrength of overlap, specifically,the smallest value of the twostrongest partial correlations).Voxels in blue were coactiveacross all three subdivisions.

Figure 4. Coactivation of insula subdivisions. Using data from the Neurosynth database, task-based coactivation profiles were determined for eachinsular subdivision by moving a searchlight in a voxel-wise manner. The color bar indicates the partial correlation value with the specific insularsubregion “seed” when all other subdivisions were also considered.

Uddin et al. 21

probed. Furthermore, to understand what each insularsubdivision expressed to a greater extent relative to theirmean, we also determined profiles for each insular sub-region by first subtracting a “mean fingerprint” (notshown). Figure 8 shows that functional differentiation isalso evident in insular territories.

Network Analysis

To further characterize the coactivation structure ofinsular subregions, we performed a graph-theoreticalanalysis. In particular, given recent suggestions that theAI is an important hub governing information flow, wereasoned that patterns of coactivation involving the in-sula would also reflect those properties. To perform net-work analysis, the brain was subdivided into ROIs basedon local clusters of coactivating voxels. In addition, theinsular subregions were added to the set of ROIs con-sidered. The final set of ROIs comprised those regionswith sufficient observations and was used as the set ofnodes for graph-theoretical analysis (Figure S1). Table 2lists the six insular subregions examined in terms of per-centiles. Degree is the number of connections of a nodein a graph (in this case, significant coactivation linksbetween ROIs). Notably, the values for the dorsal AI werevery high, especially on the left hemisphere. Degree doesnot consider edge weight, so it is notable that differencesbetween left and right hemispheres were observed fornode strength, too. Strength percentile was determinedrelative to an ROIʼs hemisphere because the actual valuesfor the right hemisphere were relatively low when allROIs were considered; yet, percentiles within the righthemisphere were high for dorsal AI, especially. Between-

ness centrality attempts to capture a nodeʼs influence tothe “flow” of signals in a graph. Betweenness centralitywas highest for the left dorsal AI.Finally, we determined the functional diversity across

ROIs so as to compare them to that of insular subregions.Diversity reflects the range of tasks that a region is en-gaged by and was computed for the entire insular cortexas well as for its subdivisions with enough observationsto compute the measure. In previous work, we havedescribed functional diversity values across ROIs span-ning the entire brain (Anderson et al., 2013). Here,considering the entire left and right insula separately (ineach case, combining the subregions), we found that theywere more diverse than 86% and 100% of noninsulaROIs, respectively. The left and right dorsal AI subregionswere more diverse than 77% and 89% of noninsula ROIs,respectively.

DISCUSSION

Investigating the insula is a challenge for various reasons,including its functional complexity and uncertainty re-garding its internal structural organization. We addressedthese challenges by adopting the concept of a functionalfingerprint, which allows one to qualitatively describe aregionʼs functional repertoire as well as quantify the asso-ciated functional complexity. We took advantage of recentwork reporting a tripartite parcellation of the insula (Deenet al., 2010), which allowed us to characterize the potential

Figure 6. Overlap between connection partners of each insularsubdivision. As in Figure 5, voxels in blue were coactive across allthree subdivisions. The slices shown highlight several notable brainregions, including thalamus, caudate, putamen, ACC, left TPJ, andright lateral pFC.

Figure 7. “Common” functional fingerprint of insular subdivisions.The common fingerprint was determined by combining all six insularsubregion (see text). All task domains were engaged by each subregionat least some of the time. TOM = theory of mind; MemWork = workingmemory; MemOther = long-term memory.

22 Journal of Cognitive Neuroscience Volume 26, Number 1

functional differentiation of insular subdivisions andmore thoroughly investigate their roles relative to otherROIs in the brain.The seemingly ubiquitous nature of insular cortical

activation can in part be explained by taking into accountfunctional heterogeneity within this region, as has beendemonstrated by recent resting-state functional connectiv-ity (Deen et al., 2010; Taylor, Seminowicz, & Davis, 2009)

and meta-analytic analyses (Cauda et al., 2012; Changet al., 2012; Kelly et al., 2012; Kurth et al., 2010). However,several questions with regards to the functional repertoireof insular subdivisions, as well as their network properties,remain unanswered. In particular, it has not been inves-tigated whether purported subdivisions of the insular cor-tex demonstrate truly unique coactivation patterns. In thisstudy, we conducted meta-analytic coactivation, functionalfingerprint, diversity, and graph-theoretical analyses toaddress open questions about insular organization andprovide further insight into this increasingly popular brainregion (Behrens, Fox, Laird, & Smith, 2012).

Analysis of functional task coactivation revealed thatinsular subregions are engaged by a broad set of neuralpartners. On the lateral surface, the posterior insula co-activated with regions surrounding the lateral fissure,extending superiorly to somatomotor regions in thevicinity of the central sulcus. Notably, little or no coactiv-ity was observed with dorsolateral pFC or related areasin superior parietal cortex. On the medial surface, co-activation was quite extensive, especially for the lefthemisphere. For the ventral AI, coactivation with inferiorfrontal regions extending into the OFC as well as theanterior temporal lobe was evident. Finally, turning to

Table 2. Graph Metrics for Insula Subdivisions inEach Hemispherea

Insular Subdivision Degree StrengthBetweenessCentrality

Left dorsal AI 100 100 86

Left posterior insula 89 92 74

Left ventral AI 1 4 13

Right dorsal AI 84 93 67

Right posterior insula 84 81 52

Right ventral AI 46 43 34

aPercentile calculated within own hemisphere.

Figure 8. “Specific” functional fingerprints of insular subdivisions. When the “average profile” (not shown) is subtracted out, particular features areobserved in the fingerprints. The black circle represents zero, with points outside representing features found more in the given subdivision than theoverall insula and points within the black circle representing features found less in the given subdivision than the overall insula. Other abbreviationsas in Figure 7.

Uddin et al. 23

the dorsal AI, coactivation with lateral pFC regions, supe-rior parietal cortex, and midcingulate gyrus were con-spicuous. In broad strokes, the coactivation results are inagreement with the findings from resting-state connectivity(Deen et al., 2010) and indicate a more “cognitive” dorsalAI, a more “emotional” ventral AI, and a “somatomotor”posterior insula (see also Chang et al., 2012; Kelly et al.,2012; Deen et al., 2010; Kurth et al., 2010; Taylor et al.,2009). Yet, this is an oversimplification, as the intersectionresults of Figure 5 illustrate. Many of the regions high-lighted above were robustly coactive with at least two ofthe three insular subdivisions. Overlap between all threeinsular subregions was also identified across multiplebrain sites, including both cortical (e.g., ACC, left TPJ)and subcortical (e.g., thalamus) areas. Notably, these re-gions of overlap resemble the “salience network,” whichis involved generally in identifying relevant internal andexternal stimuli (Seeley et al., 2007). From this vantagepoint, the tripartite scheme previously proposed for thefunctional organization of the insula breaks down. In all,a more complete characterization of the insula requiresacknowledging the observed overlap, at least when taskcoactivation is considered.

The functional fingerprint analysis painted a similarpicture. When differences from the mean profile arehighlighted, it is possible to describe the dorsal AI as more“cognitive” and the ventral AI as more “emotional,” forinstance. But that characterization misses the strong func-tional overlap that is unveiled by the “common” finger-print. Thus, the neural partners of insular subregionswere active at least some of the time in all of the taskdomains investigated. Intriguingly, even the profiles withthe average removed revealed important similarities be-tween the ventral AI and the posterior insula—both ofwhich were quite dissimilar from the dorsal AI. Thus,whereas previous reverse inference mappings have fo-cused on functional dissociations within the insula (Changet al., 2012), our fine-grained analyses, which included amore detailed task domain structure, suggest a more com-plex and nuanced picture of the functionality of insularterritories.

The left and right dorsal AI displayed high functionaldiversity relative to noninsula brain regions. This findingis consistent with a recent meta-analysis reporting morewidespread participation of the anterior than posteriorinsula across task domains (Cauda et al., 2012). Further-more, here, betweeness centrality and node strengthmeasures were highest for AI subregions (within theirhemisphere) compared with the other insular subdivi-sions. High betweeness centrality is suggestive of nodeimportance to overall information flow through a graph.The current findings suggest that the dorsal AI may func-tion as an important node for integrating informationacross multiple brain networks. This is consistent withrecent work pointing to a central role for the dorsal AI ininfluencing switches between the central-executive net-work and the default-mode network (Uddin et al., 2011;

Sridharan et al., 2008), as well as previous work identifyingthe AI as a key cognitive control region (Levy & Wagner,2011; Cole & Schneider, 2007; Dosenbach et al., 2007).The AI is thought to be part of a larger “salience network”comprising this region along with dorsal ACC (Seeley et al.,2007). A proposed function of this network is to detectsalient events in the environment and initiate appropriatecontrol signals (Menon & Uddin, 2010). The current find-ings add further support to recently proposed theoriespositing a unique role for the AI and the salience networkin orchestrating access to attention and control systems(Menon & Uddin, 2010). They also align with emergingconceptualizations of large-scale brain networks that sup-port domain-general functions across a range of emotional,social, and cognitive tasks (Barrett & Satpute, 2013).More broadly, the findings highlight the need to care-

fully rethink our approach to function–structure mappingin the brain. There has been increasing awareness of theinadequacy of single task-based functional attributions toindividual brain regions (Anderson et al., 2013; Barrett &Satpute, 2013; Anderson, 2010; Poldrack, 2006, 2010).The current work further cements the need to take ac-count of three fundamental features of the functionalorganization of the brain: (1) the functional diversity ofindividual regions of the brain, (2) the functional differ-entiation of individual regions of the brain, and (3) thefrequent functional overlap between different brain net-works. Nodes in brain networks are active across a rangeof tasks in different domains, participate in multiple net-works, and yet remain functionally distinguishable. Howshould function–structure mapping be understood in away that is sensitive to all of these facts?One common theme in work addressing this question

has been a call for a reform of the cognitive ontology, theconcepts we use to organize and differentiate betweenpsychological tasks (Barrett & Satpute, 2013; Lindquistet al., 2012; Anderson, 2010; Poldrack, 2010; Poldrack,Halchenko, & Hanson, 2009; Pessoa, 2008). Some havesuggested that a cognitive ontology better suited to theneurosciences will reveal a set of primitive operations (orprocesses) that selectively engage individual nodes andnetworks (Poldrack, 2010). Others (Anderson et al.,2013; Barrett & Satpute, 2013; Pessoa, 2013; Anderson,Richardson, & Chemero, 2012; Lindquist et al., 2012) haveexpressed doubts that this will prove possible. One in-triguing possibility that would appear to respect the threeorganizational features of the brain emphasized aboveis that analysis of the multidimensional functional finger-prints of brain regions and networks will reveal a set ofprimitive psychological factors or “ingredients” that cap-ture the underlying functional contributions of regionsand networks to overall behavior (Barrett & Satpute,2013; Lindquist et al., 2012; Gold, Havasi, Anderson, &Arnold, 2011; Poldrack, 2010; Poldrack et al., 2009).According to this perspective, psychological states likeanger and fear as well as processes like attention and cog-nitive control would involve different mixtures of many of

24 Journal of Cognitive Neuroscience Volume 26, Number 1

the same domain-general ingredients. It is expected thatmore than one brain region will likely be involved in im-plementing or contributing each of those psychologicalingredients to a given state or process, and each brainregion will have a role in generating or contributing morethan one ingredient. That is, brain regions and networkswill differ in their loadings on this set of primitive psycho-logical factors (but see Pessoa, 2012).In the current case, such an approach could lead to the

identification of the set of functional factors that char-acterize insula function—both the common contributionsoffered by the insula subdivisions, as well as the contribu-tions that differentiate them. Whether or not each sub-division of the insula makes a single, domain-generalfunctional contribution—that is, loads on a single psy-chological factor or several—remains to be investigated,although the diversity of their functional partnershipsand fingerprints might seem to favor the latter possibility.For instance, the overlap between some nodes of the AInetwork and nodes in the so-called salience networkmight be taken to indicate their common contributionto one of the fundamental ingredients of attentional con-trol. At the same time, the contribution of the insula to avariety of emotional states, on the one hand, as well as toabstract domains like mathematics, on the other, mighthint at a repertoire of capacities broader than a singlefunction or ingredient.

Conclusions

We conducted an fMRI task coactivation meta-analysis toexamine functional and connectional properties of thehuman insular cortex. In broad strokes, our findings agreewith recent studies suggesting a functional differentiationwithin this cortical region. Yet, at a finer level of analysis,our findings question the legitimacy of a simple tripartitescheme. For example, both dorsal and ventral AI sharecoactivation partners with dorsolateral pFC, a region thatis often described as “cognitive.” This was the case thoughour procedure determined coactivation via partial correla-tion to account for the presence of all insular subdivisions.Our network analysis indicates the potential of the dorsalAI to be an influential structure because of its “centrality”properties, as well as its very high functional diversity.We hope that the present characterization of insula func-tion will provide a framework for future interpretationsof the role of this multifaceted brain structure in typicaland atypical behaviors.

Acknowledgments

This work was supported by awards K01MH092288 (to L. Q. U.)and MH071589 (to L. P.) from the National Institute of MentalHealth and award 0803739 from the National Science Founda-tion to M. L. A. This article was drafted while M. L. A. was a fellowat the Center for Advanced Study in the Behavioral Sciences atStanford University, on sabbatical leave from Franklin & MarshallCollege. He gratefully acknowledges the support. We are grateful

to Ben Deen for providing the insula subdivision maps from hisearlier work.

Reprint requests should be sent to Lucina Q. Uddin, Suite 220,1070 Arastradero Road, Palo Alto, CA 94304, or via e-mail: lucina@stanford.edu or Michael L. Anderson, Department of Psychology,Franklin & Marshall College, P.O. Box 3003, Lancaster, PA 17604,or via e-mail: michael.anderson@fandm.edu.

REFERENCES

Anderson, M. L., Brumbaugh, J., & Suben, A. (2010). Investigatingfunctional cooperation in the human brain using simplegraph-theoretic methods. In A. Chaovalitwongse,P. M. Pardalos, & P. Xanthopoulos (Eds.), Computationalneuroscience (pp. 31–42). Springer.

Anderson, M. L. (2010). Neural reuse: A fundamentalorganizational principle of the brain. Behavioral andBrain Sciences, 33, 245–266; discussion 266-313.

Anderson, M. L., Kinnison, J., & Pessoa, L. (2013). Describingfunctional diversity of brain regions and brain networks.Neuroimage, 73, 50–58.

Anderson, M. L., & Penner-Wilger, M. (2013). Neural reuse inthe evolution and development of the brain: Evidence fordevelopmental homology? Developmental Psychobiology,55, 42–51.

Anderson, M. L., Richardson, M. J., & Chemero, A. (2012).Eroding the boundaries of cognition: Implications ofembodiment(1). Topics in Cognitive Science, 4, 717–730.

Augustine, J. R. (1985). The insular lobe in primatesincluding humans. Neurological Research, 7, 2–10.

Augustine, J. R. (1996). Circuitry and functional aspects ofthe insular lobe in primates including humans. BrainResearch. Brain Research Reviews, 22, 229–244.

Barrat, A., Barthelemy, M., Pastor-Satorras, R., & Vespignani, A.(2004). The architecture of complex weighted networks.Proceedings of the National Academy of Sciences, U.S.A.,101, 3747–3752.

Barrett, L. F., & Satpute, A. B. (2013). Large-scale brain networksin affective and social neuroscience: Towards an integrativefunctional architecture of the brain. Current Opinion inNeurobiology, 23, 361–372.

Behrens, T. E., Fox, P., Laird, A., & Smith, S. M. (2012).What is the most interesting part of the brain? Trendsin Cognitive Sciences, 17, 2–4.

Benjamini, Y., Drai, D., Elmer, G., Kafkafi, N., & Golani, I. (2001).Controlling the false discovery rate in behavior geneticsresearch. Behavioural Brain Research, 125, 279–284.

Bossaerts, P. (2010). Risk and risk prediction error signals inanterior insula. Brain Structure & Function, 214, 645–653.

Bud Craig, A. D. (2009). How do you feel—Now? Theanterior insula and human awareness. Nature ReviewsNeuroscience, 10, 59–70.

Butti, C., & Hof, P. R. (2010). The insular cortex: A comparativeperspective. Brain Structure & Function, 214, 477–493.

Cauda, F., Costa, T., Torta, D. M., Sacco, K., DʼAgata, F.,Duca, S., et al. (2012). Meta-analytic clustering of theinsular cortex: Characterizing the meta-analytic connectivityof the insula when involved in active tasks. Neuroimage,62, 343–355.

Chang, L. J., Yarkoni, T., Khaw, M. W., & Sanfey, A. G.(2012). Decoding the role of the insula in humancognition: Functional parcellation and large-scalereverse inference. Cerebral Cortex, 23, 739–749.

Chao, A., & Shen, T. J. (2003). Nonparametric estimationof Shannonʼs index of diversity when there are unseenspecies in sample. Environmental and EcologicalStatistics, 10, 429–443.

Uddin et al. 25

Cole, M. W., & Schneider, W. (2007). The cognitive controlnetwork: Integrated cortical regions with dissociablefunctions. Neuroimage, 37, 343–360.

Craddock, R. C., James, G. A., Holtzheimer, P. E., III, Hu,X. P., & Mayberg, H. S. (2012). A whole brain fMRI atlasgenerated via spatially constrained spectral clustering.Human Brain Mapping, 33, 1914–1928.

Craig, A. D. (2002). How do you feel? Interoception:The sense of the physiological condition of the body.Nature Reviews Neuroscience, 3, 655–666.

Craig, A. D. (2010). The sentient self. Brain Structure& Function, 214, 563–577.

Critchley, H. D., Mathias, C. J., & Dolan, R. J. (2001).Neuroanatomical basis for first- and second-orderrepresentations of bodily states. Nature Neuroscience,4, 207–212.

Damasio, A. R. (1996). The somatic marker hypothesis and thepossible functions of the prefrontal cortex. PhilosophicalTransactions of the Royal Society of London, Series B,Biological Sciences, 351, 1413–1420.

Deen, B., Pitskel, N. B., & Pelphrey, K. A. (2010). Three systemsof insular functional connectivity identified with clusteranalysis. Cerebral Cortex, 21, 1498–1506.

Dosenbach, N. U., Fair, D. A., Miezin, F. M., Cohen, A. L.,Wenger, K. K., Dosenbach, R. A., et al. (2007). Distinctbrain networks for adaptive and stable task control inhumans. Proceedings of the National Academy ofSciences, U.S.A., 104, 11073–11078.

Fox, P. T., Laird, A. R., Fox, S. P., Fox, P. M., Uecker, A. M.,Crank, M., et al. (2005). BrainMap taxonomy of experimentaldesign: Description and evaluation. Human Brain Mapping,25, 185–198.

Fox, P. T., & Lancaster, J. L. (2002). Opinion: Mappingcontext and content: The BrainMap model. NatureReviews Neuroscience, 3, 319–321.

Gold, K., Havasi, C., Anderson, M., & Arnold, K. (2011).Comparing matrix decomposition methods formeta-analysis and reconstruction of cognitiveneuroscience results. Paper presented at the Proceedingsof the 24th Annual Conference of the Florida ArtificialIntelligence Research Society.

Kelly, C., Toro, R., Di Martino, A., Cox, C. L., Bellec, P.,Castellanos, F. X., et al. (2012). A convergent functionalarchitecture of the insula emerges across imagingmodalities. Neuroimage, 61, 1129–1142.

Kristo, G., Rutten, G. J., Raemaekers, M., de Gelder, B.,Rombouts, S., & Ramsey, N. (2012). Task and task-free fMRIreproducibility comparison for motor network identification.Human Brain Mapping. doi:10.1002/hbm.22180.

Kurth, F., Zilles, K., Fox, P. T., Laird, A. R., & Eickhoff, S. B.(2010). A link between the systems: Functional differentiationand integration within the human insula revealed bymeta-analysis. Brain Structure & Function, 214, 519–534.

Laird, A. R., Eickhoff, S. B., Kurth, F., Fox, P. M., Uecker,A. M., Turner, J. A., et al. (2009). ALE meta-analysisworkflows via the brainmap database: Progress towardsa probabilistic functional brain atlas. Frontiers inNeuroinformatics, 3, 23.

Laird, A. R., Lancaster, J. L., & Fox, P. T. (2005). BrainMap:The social evolution of a human brain mapping database.Neuroinformatics, 3, 65–78.

Laird, A. R., Lancaster, J. L., & Fox, P. T. (2009). Lost inlocalization? The focus is meta-analysis. Neuroimage,48, 18–20.

Levy, B. J., & Wagner, A. D. (2011). Cognitive control andright ventrolateral prefrontal cortex: Reflexive reorienting,motor inhibition, and action updating. Annals of theNew York Academy of Sciences, 1224, 40–62.

Lindquist, K. A., Wager, T. D., Kober, H., Bliss-Moreau, E.,& Barrett, L. F. (2012). The brain basis of emotion:A meta-analytic review. Behavioral and Brain Sciences,35, 121–143.

Magurran, A. E. (2004). Measuring biological diversity.Oxford, UK: Blackwell.

Meindl, T., Teipel, S., Elmouden, R., Mueller, S., Koch, W.,Dietrich, O., et al. (2009). Test-retest reproducibilityof the default-mode network in healthy individuals.Human Brain Mapping, 31, 237–246.

Menon, V., & Uddin, L. Q. (2010). Saliency, switching,attention and control: A network model of insula function.Brain Structure & Function, 214, 655–667.

Nelson, S. M., Dosenbach, N. U., Cohen, A. L., Wheeler,M. E., Schlaggar, B. L., & Petersen, S. E. (2010). Role ofthe anterior insula in task-level control and focal attention.Brain Structure & Function, 214, 669–680.

Nieuwenhuys, R. (2012). The insular cortex: A review.Progress in Brain Research, 195, 123–163.

Passingham, R. E., Stephan, K. E., & Kotter, R. (2002). Theanatomical basis of functional localization in the cortex.Nature Reviews Neuroscience, 3, 606–616.

Pessoa, L. (2008). On the relationship between emotion andcognition. Nature Reviews Neuroscience, 9, 148–158.

Pessoa, L. (2012). Beyond brain regions: Network perspectiveof cognition-emotion interactions. Behavioral and BrainSciences, 35, 158–159.

Pessoa, L. (2013). The cognitive-emotional brain: Frominteractions to integration. Cambridge, MA: MIT Press.

Poldrack, R. A. (2006). Can cognitive processes be inferredfrom neuroimaging data? Trends in Cognitive Sciences,10, 59–63.

Poldrack, R. A. (2010). Mapping mental function to brainstructure: How can cognitive neuroimaging succeed?Perspectives on Psychological Science, 5, 753–761.

Poldrack, R. A., Halchenko, Y. O., & Hanson, S. J. (2009).Decoding the large-scale structure of brain function byclassifying mental states across individuals. PsychologicalScience, 20, 1364–1372.

Price, C. J., & Friston, K. J. (2005). Functional ontologiesfor cognition: The systematic definition of structure andfunction. Cognitive Neuropsychology, 22, 262–275.

Raemaekers, M., Vink, M., Zandbelt, B., van Wezel, R. J.,Kahn, R. S., & Ramsey, N. F. (2007). Test-retest reliabilityof fMRI activation during prosaccades and antisaccades.Neuroimage, 36, 532–542.

Rubinov, M., & Sporns, O. (2010). Complex network measuresof brain connectivity: Uses and interpretations. Neuroimage,52, 1059–1069.

Seeley, W. W., Menon, V., Schatzberg, A. F., Keller, J.,Glover, G. H., Kenna, H., et al. (2007). Dissociableintrinsic connectivity networks for salience processingand executive control. Journal of Neuroscience, 27,2349–2356.

Shannon, C. E. (1948). A mathematical theory ofcommunication. The Bell System Technical Journal,27, 379–423, 623–656.

Silani, G., Bird, G., Brindley, R., Singer, T., Frith, C., & Frith, U.(2008). Levels of emotional awareness and autism: An fMRIstudy. Social Neuroscience, 3, 97–112.

Singer, T. (2006). The neuronal basis and ontogeny of empathyand mind reading: Review of literature and implications forfuture research. Neuroscience and Biobehavioral Reviews,30, 855–863.

Singer, T., Seymour, B., OʼDoherty, J., Kaube, H., Dolan,R. J., & Frith, C. D. (2004). Empathy for pain involves theaffective but not sensory components of pain. Science,303, 1157–1162.

26 Journal of Cognitive Neuroscience Volume 26, Number 1

Smith, S. M., Fox, P. T., Miller, K. L., Glahn, D. C., Fox,P. M., Mackay, C. E., et al. (2009). Correspondence of thebrainʼs functional architecture during activation and rest.Proceedings of the National Academy of Sciences, U.S.A.,106, 13040–13045.

Sporns, O. (2003). Graph theory methods for the analysis ofneural connectivity patterns. In R. Kötter (Ed.), Neurosciencedatabases. A practical guide. The Netherlands: Klüwer.

Sporns, O. (2011). The human connectome: A complexnetwork. Annals of the New York Academy of Sciences,1224, 109–125.

Sridharan, D., Levitin, D. J., & Menon, V. (2008). A criticalrole for the right fronto-insular cortex in switchingbetween central-executive and default-mode networks.Proceedings of the National Academy of Sciences, U.S.A.,105, 12569–12574.

Taylor, K. S., Seminowicz, D. A., & Davis, K. D. (2009). Twosystems of resting state connectivity between the insulaand cingulate cortex. Human Brain Mapping, 30,2731–2745.

Toro, R., Fox, P. T., & Paus, T. (2008). Functional coactivationmap of the human brain. Cerebral Cortex, 18, 2553–2559.

Touroutoglou, A., Hollenbeck, M., Dickerson, B. C., &Feldman Barrett, L. (2012). Dissociable large-scale networksanchored in the right anterior insula subserve affectiveexperience and attention. Neuroimage, 60, 1947–1958.

Uddin, L. Q., & Menon, V. (2009). The anterior insula in autism:Under-connected and under-examined. Neuroscience andBiobehavioral Reviews, 33, 1198–1203.

Uddin, L. Q., Supekar, K. S., Ryali, S., & Menon, V. (2011).Dynamic reconfiguration of structural and functionalconnectivity across core neurocognitive brain networks withdevelopment. Journal of Neuroscience, 31, 18578–18589.

Yarkoni, T., Poldrack, R. A., Nichols, T. E., Van Essen, D. C., &Wager, T. D. (2011). Large-scale automated synthesis of humanfunctional neuroimaging data. Nat Methods, 8, 665–670.

Yarkoni, T., Poldrack, R. A., Van Essen, D. C., & Wager, T. D.(2010). Cognitive neuroscience 2.0: Building a cumulativescience of human brain function. Trends in CognitiveSciences, 14, 489–496.

Uddin et al. 27