Three Systems of Insular Functional Connectivity Identified with Cluster Analysis

Study at a glance

• Authors: Ben Deen, Naomi B. Pitskel, Kevin A. Pelphrey
• Year: 2011
• Journal: Cerebral Cortex
• DOI: 10.1093/cercor/bhq186
• N: 30
• Design: Resting-state fMRI (3T, 200 volumes / 6 min 40 s) connectivity-based parcellation of the left AND right insula (k-means clustering of whole-brain voxelwise connectivity maps, k=3), with each cluster then used as a seed for whole-brain and cingulate-ROI connectivity; functionally validated on an independent task dataset (n=20) of passively viewed disgusting vs neutral IAPS images

Key findings

• Data-driven k-means clustering of whole-brain resting-state connectivity profiles parcellates each insula — left and right, separately — into the SAME three subregions arranged along the ventral-anterior→dorsal-posterior axis: ventral anterior insula (vAI), dorsal anterior-to-middle insula (dAI), and posterior insula (PI). This is the founding demonstration of the tripartite insula later replicated by Chang et al. (2013) and refined from Cauda et al.’s (2011) bipartite-plus-transitional split.
• Each subregion anchors a distinct whole-brain network: PI → primary and secondary somatomotor cortices (the bilateral somatomotor network); dAI → dorsal ACC and the Dosenbach cognitive-control network (pre-SMA, supramarginal gyrus, anterior IFG, plus weaker attention-network coupling); vAI → pregenual ACC, adjacent OFC/IFG and superior temporal sulcus (a limbic/affective pattern).
• A cingulo-insular gradient confirmed by ROI analysis: pregenual ACC is most strongly connected to vAI, dorsal ACC to dAI, and middle cingulate to PI — anterior cingulate regions couple to ventral-anterior insula, more posterior cingulate to more dorsal-posterior insula. Replicates and extends Taylor et al.’s (2009) two cingulo-insular systems, and matches macaque anatomical connectivity.
• Functional validation on an independent disgust task: both anterior clusters (vAI and dAI) responded selectively to disgusting > neutral images bilaterally; PI did not respond to images at all. dAI was more generally responsive to images (responded to neutral too), leading the authors to speculate vAI carries an earlier sensory/affective disgust response while dAI links it to attentional/executive mechanisms.
• Cluster-number robustness: the k=2 solution collapses vAI+dAI into a single anterior division (the salience network) vs posterior; the k=4 solution splits dAI into rostral/caudal halves with near-identical connectivity (t-map correlation r=0.84–0.89), so k=3 is the stable optimum. dorsal and ventral anterior insula are ‘closely linked but dissociable’ — a dorsal cognitive-control network and a ventral emotional network that combine into the salience network.

Limitations

• Resting-state connectivity is correlational and non-directional — it cannot say whether anterior insula drives posterior or vice versa, nor what the salience/sensorimotor/limbic labels compute.
• Clustering is imposed, not discovered: the authors fixed k=3 a priori (justified post hoc by higher-k solutions collapsing onto the three maps), and k-means on a dense distance matrix can return divergent parcellations across initializations — mitigated here by choosing the minimum-within-cluster-variance solution over 100 runs (cf. Nanetti et al. 2009’s warning).
• The disgust validation used passively viewed IAPS images with self-reported affect only; it shows the clusters are functionally distinct but not what each computes, and the emotion-regulation manipulation is reported elsewhere.
• Two slightly different structural pulse sequences across the 30 resting subjects (used only for coregistration); global-signal was among the 9 nuisance regressors, so any anticorrelation interpretation inherits the GSR caveat, though this paper leans on the positive parcellation.

Connections

• Methods: Resting-state functional connectivity
• Concepts: Insular cortex, Salience network, Anterior cingulate cortex (ACC), Disgust, Core affect
• Debates: Is interoception domain-general or organ-specific?, Do discrete emotions map onto distinct brain regions or networks?, inferential?
• Related studies: Chang et al. (2013), Cauda et al. (2011), Vercelli et al. (2015), Craig (2002), Craig (2009), Haruki & Ogawa (2023), Nord et al. (2021), Brooks et al. (2005), Lindquist et al. (2012)

This is the paper the wiki had been citing secondhand and mislinked: Chang et al. (2013) describe their own tripartite parcellation as replicating “the three-cluster solution of Deen et al. (2011),” and the Insular cortex page’s tripartite section is built on that replication — but the primary source was never held here. Deen, Pitskel & Pelphrey are the origin of the functional tripartition of the insula: two years before Chang, using the same tool (k-means clustering of resting-state connectivity), they split each insula into a ventral anterior, a dorsal anterior-to-middle, and a posterior subregion, then validated the split against an independent task. Where Cauda et al. (2011) recovered a bipartite insula (anterior/posterior plus a transitional zone), Deen et al. resolve the anterior end into two — the finding the wiki’s whole insula-subdivision thread rests on.

The parcellation: three systems, both hemispheres

The method is fully data-driven within the insula: every voxel in an anatomically drawn left- and right-insula ROI gets a whole-brain connectivity map (its resting time course regressed against every other voxel, averaged across 30 subjects), those maps are treated as vectors, and k-means clustering (k=3) groups the voxels by similarity of connectivity pattern. The result is the same three subregions on the left and the right, arranged along the insula’s ventral-anterior → dorsal-posterior trajectory:

• Posterior insula (PI) → the bilateral somatomotor network (primary and secondary somatosensory and motor cortex, SMA, central sulcus, area MT, medial thalamus). This is the interoceptive/somatosensory pole — the terminus the wiki’s lamina I account and Brooks et al.’s dpIns somatotopy assign to primary interoceptive cortex.
• Dorsal anterior-to-middle insula (dAI) → the cognitive-control network (Dosenbach et al. 2007): dorsal ACC, pre-SMA, supramarginal gyrus, anterior IFG, with weaker coupling to attention-network regions (frontal eye fields, intraparietal sulcus).
• Ventral anterior insula (vAI) → a limbic/affective network: pregenual ACC and adjacent dorsal ACC, OFC/IFG, and the superior temporal sulcus. Cytoarchitectonically closest to limbic cortex, most connected to allocortex, and (per Mutschler et al. 2009, cited here) the insular site most likely to co-activate with the amygdala and to track peripheral physiological responses.

This is exactly the dorsal-cognitive / ventral-affective seam in the anterior insula that Chang later grounds by reverse inference and that the wiki’s finer results keep re-finding — Haruki & Ogawa’s right dorsal anterior insula preferring cardiac/arousal attention, the EPIC model’s agranular ventral anterior insula as the visceromotor pole (see Insular cortex). Deen et al. drew it first.

The cingulo-insular gradient

Where Deen et al. add something neither Cauda nor Chang emphasize is the cingulate end of the connectivity. An ROI analysis regressing three cingulate spheres (pregenual ACC, dorsal ACC, middle cingulate) on the three insular clusters yields a clean gradient: pACC ↔ vAI, dACC ↔ dAI, MCC ↔ PI — the more anterior the cingulate region, the more ventral-anterior the insular partner it prefers, each pairing significant and dissociable by paired t-tests. This replicates Taylor et al.’s (2009) “two systems of resting-state connectivity between insula and cingulate cortex” and extends it to three, and it maps onto the standard cingulate functional split (pregenual/subgenual ACC = emotional, dorsal ACC = cognitive; Bush et al. 2000). The upshot for the Anterior cingulate cortex (ACC) page: the insula+ACC “limbic sensory + limbic motor” pairing the wiki states as a single unit is really two parallel insulo-cingulate loops — a ventral emotional one and a dorsal cognitive one — plus a posterior sensorimotor one.

The disgust validation, and why it fits the “not a disgust module” reframe

The distinctive methodological move is that Deen et al. don’t stop at rest — they take the three clusters, defined purely from resting connectivity, into a completely independent task dataset (20 subjects viewing disgusting vs neutral IAPS images) and ask whether the connectivity-defined regions are also functionally distinct. They are: both anterior clusters (vAI and dAI) responded selectively to disgusting > neutral images bilaterally, while the posterior insula did not respond to images at all. This is early, clean evidence that connectivity-based parcels are functionally meaningful — the same logic Chang uses with reverse inference, across a smaller data gap.

Two readings matter for this wiki. First, the anterior-insula disgust response is split: dAI is generally image-responsive (it responded to neutral images too), while vAI is more disgust-selective — so the authors speculate vAI carries an earlier sensory/affective disgust response and dAI links it to attentional/executive mechanisms. That is the dorsal-cognitive/ventral-affective division doing work on a specific emotion. Second, it sits comfortably inside the Lindquist “selective ≠ specific” reframe on Do discrete emotions map onto distinct brain regions or networks?: the anterior insula is selectively engaged by disgust here, but the same regions belong to networks (cognitive control, limbic/affective) that are anything but disgust-specific — so a disgust response is not a disgust module. See Disgust.

Where it sits in the wiki’s insula evidence

• It is the primary source for the tripartition Chang replicates. Read the two together: Deen (2011) established the three-cluster solution in both hemispheres and validated it with a disgust task; Chang (2013) reproduced it in the right insula, added the k=2/k=4 robustness the same way Deen did, and attached a function to each subdivision by reverse inference. The tripartition is now first-hand at both ends.
• It refines Cauda in the same direction Chang does. Cauda’s anterior pole is undivided (10 coarse ROIs, a-priori k=2); Deen’s voxelwise k=3 splits it into dorsal and ventral. Cauda’s bipartite-plus-transitional split is the coarse version of Deen’s three systems.
• It puts the cingulate on the map. The pACC↔vAI / dACC↔dAI / MCC↔PI gradient is the connectivity backbone under the wiki’s insula+ACC “an emotion” pairing, and shows it is really three parallel loops.

What it cannot say

Everything is correlational and cross-sectional — the standing Resting-state functional connectivity caveat. The clustering is imposed (k fixed at 3, justified by higher-k solutions collapsing onto the same three maps, not discovered by a validity index as in Chang), and k-means on a dense insular distance matrix can diverge across initializations — Deen et al. guard against this by taking the minimum-variance solution over 100 runs, but Nanetti et al. (2009), whom they cite, showed the concern is real. The disgust task shows the clusters are functionally distinct, not what each computes. And the parcellation is anatomical/functional, not causal: it says the resting insula is organized into three systems, not that any of them is necessary for anything — the lesion literature is where necessity is tested, and it does not confirm the tidy functional map.