Cortical and subcortical mapping of the human allostatic–interoceptive system using 7 Tesla fMRI

Abstract

The brain continuously anticipates the body’s energetic needs and prepares to meet them before they arise—a process called allostasis. To support allostasis, the brain continually models the body’s sensory state, a process known as interoception. Here we replicate and extend a large-scale system that supports allostasis and interoception in the human brain using ultrahigh precision 7 Tesla functional magnetic resonance imaging (n = 90), improving precision in subgenual and pregenual anterior cingulate topography and expanding brainstem nuclei mapping. Our functional connectivity analyses provide corroborating evidence for more than 96% of the anatomical connections documented in nonhuman animal tract-tracing studies. This system also includes regions of dense intrinsic connectivity throughout the system, some of which were identified previously as part of the backbone of neural communication across the brain. These results reinforce the existing evidence for a whole-brain system that supports the modeling and regulation of the body’s internal milieu.

The brain is constantly monitoring the systems in the body. Here the authors use 7 Tesla functional magnetic resonance imaging to map a large-scale brain system for body regulation in humans, including brainstem nuclei, and confirm many monosynaptic connections traced in nonhuman animals.

Main

The brain efficiently regulates and coordinates the systems of the body as it continually interfaces with an ever-changing and only partly predictable world. Various lines of research, including tract-tracing studies of nonhuman animals^1,2, discussions of predictive processing^3–6 and research on the central control of autonomic nervous system function^7–11, all suggest the existence of a unified, distributed brain system that anticipates the metabolic needs of the body and prepares to meet those needs before they arise, a process called allostasis¹² (for recent reviews, see refs. ^13,14). Allostasis is not a condition or a state of the body, but rather the process by which the brain efficiently coordinates and regulates the various systems of the body¹². Just as somatosensory and other exteroceptive sensory signals are processed in the service of skeletomotor control, the brain is thought to model the internal sensory conditions of the body (that is, the internal milieu) in the service of allostasis, a process known as interoception^15–18.

Using resting state functional magnetic resonance imaging (fMRI) in three samples totaling almost 700 human participants scanned at 3 Tesla¹⁹, we previously identified a distributed allostatic–interoceptive system consisting of two well-known intrinsic networks, the default mode and salience networks, overlapping in many key cortical visceromotor allostatic regions that also serve as ‘rich club’ hubs that have been implicated as the ‘backbone’ for neural communication throughout the brain (Fig. 1a). Our investigation was guided by the anatomical tracts identified in published studies of macaques and other nonhuman mammals (Table 2 in ref. ¹⁹). This study was more cortically focused, examining the functional connectivity of primary interoceptive cortex spanning the dorsal mid insula (dmIns) and dorsal posterior insula (dpIns), as well as key allostatic regions in the cerebral cortex that are directly connected to the brainstem regions that are known to be responsible for controlling the motor changes in the viscera (that is, visceromotor cortical regions), such as the anterior midcingulate cortex (aMCC), pregenual anterior cingulate cortex (pACC), subgenual ACC (sgACC) and agranular insular cortex (also known as ventral anterior insula (vAIns), which is also posterior orbitofrontal cortex), as well as the dorsal amygdala (dAmy) containing the intercalated bodies and the central nucleus (Fig. 1a). Our 3 Tesla analysis yielded a replicable, integrated system consisting of two well-known intrinsic networks, in addition to primary interoceptive cortex. We did explore some aspects of the system’s subcortical extent, including the thalamus, hypothalamus, hippocampus, ventral striatum, periaqueductal gray (PAG), parabrachial nucleus (PBN) and nucleus tractus solitarius (NTS), all regions known to have a role in control of the autonomic nervous system, the immune system and the endocrine system (for example, refs. ^20–26), but our ability to more extensively map the midbrain and brainstem extents of the system was limited by the use of 3 Tesla imaging.

Fig. 1. Key cortical and subcortical regions involved in interoception and allostasis.

a, Using 3 Tesla fMRI resting state connectivity, we showed a unified system consisting of the default mode network (in red) and salience network (in blue), which overlapped in many key cortical visceromotor allostatic regions (in purple) that also serve as ‘rich club’ hubs (labeled in yellow), in addition to a portion of primary interoceptive cortex (dpIns; left)¹⁹. We reported the system’s connectivity to some subcortical regions known to have a role in control of the autonomic nervous system, the immune system and the endocrine system, such as the thalamus, hypothalamus, hippocampus, ventral striatum, PAG, PBN and NTS (for example, refs. ^20–26; right)¹⁹. * denotes brainstem regions. Panel a is reproduced with permission from ref. ¹⁹. b, Expanded set of seed regions used in the present analysis. Hippo, hippocampus; hypothal, hypothalamus; IFG, inferior frontal gyrus; ITG, inferior temporal gyrus; PHG, parahippocampal gyrus; postCG, postcentral gyrus; STS, superior temporal sulcus; Thal, thalamus.

In the present study, we replicated and extended evidence for the allostatic–interoceptive system (Fig. 1b) using ultrahigh field (7 Tesla) MRI, which allows data acquisition with higher spatial resolution (1.1-mm isotropic), better signal-to-noise ratio (SNR^27–29) and increased sensitivity in mapping functional connectivity of brainstem nuclei involved in arousal, motor and other vital processes (for example, autonomic, nociceptive and sensory^30,31). This is particularly important given the increasing importance of the allostatic–interoceptive system as a tool for investigating interoception and allostasis in basic brain function, both in neurotypical samples and in specific populations (for example, refs. ^32–35). In addition, the research indicates that regions in this system are also important for a wide range of psychological domains, including cognition, emotion, pain, decision-making and perception (Fig. 5 in ref. ¹⁹; see also refs. ^36–38), suggesting the hypothesis that allostatic and interoceptive signals may have a more fundamental role in shaping basic brain dynamics (for discussion, see refs. ^5,39–41).

Thus, we tested within-system functional connectivity in 90 human participants (age range = 18–40 years, mean = 26.9 years, s.d. = 6.2 years; 40 females) using a fast low-angle excitation echo-planar technique sequence shown to reduce artifacts and improve temporal SNR^24,42. This approach allowed a more precise mapping of connectivity for regions with known signal issues at 3 Tesla, such as the sgACC (low SNR), amygdala (noise from adjacent veins⁴³), columns within the PAG (noise from adjacent aqueduct) and other small structures that could be particularly influenced by partial volume effects. We took advantage of the recently developed, much improved and validated in vivo brainstem and diencephalic nuclei atlases^44–48 to guide our analysis. This was crucial because our hypotheses were specifically derived from published tract-tracing studies of macaques and other nonhuman mammals that establish structural pathways carrying ascending interoceptive signals from the periphery, for example, through the vagus nerve, to subcortical and cortical regions of the allostatic–interoceptive system (Supplementary Table 1). Extending¹⁹, we more extensively examined the intrinsic connectivity of subcortical nuclei such as mediodorsal thalamus (mdThal), hypothalamus, dAmy, hippocampus, ventral striatum, PAG, PBN and NTS (in the medullary viscero-sensory-motor (VSM) nuclei complex, which also includes the dorsal motor nucleus of the vagus, nucleus ambiguus and hypoglossal nucleus), in addition to considering the connectivity of dorsal raphe (DR), substantia nigra (SN), ventral tegmental area (VTA), locus coeruleus (LC), superior colliculus (SC) and lateral geniculate nucleus (LGN). The DR, SN, VTA and LC are midbrain and pontine monoamine-producing nuclei that contribute to relaying the body’s metabolic status to the cortex⁴⁹. The SC and LGN are not traditionally considered to be directly involved in interoception and allostasis, but they share anatomical connections with key visceromotor regulation regions in the system (Supplementary Table 1; refs. ^50–55). For example, neurons in the intermediate and deep layers of the SC are connected to aMCC⁵⁶, hypothalamus^57,58 and PAG⁵⁹, and have been directly implicated in skeletomotor^60,61 and visceromotor^62,63 actions that facilitate approach or avoidance behaviors. The SC is also thought to be a major point of sensory-motor integration and is associated with affective feelings^64,65. The LGN receives interoceptive inputs from the PAG⁵² and PBN^55,66,67, and shares monosynaptic connections with the hypothalamus⁶⁸ and pACC⁶⁹. We also examined connectivity patterns for subregions of the PAG, hippocampus, SC and hypothalamus rather than as a single region of interest (ROI) as discussed in ref. ¹⁹ given their functional heterogeneity^70,71 and differential involvement in allostasis (for example, refs. ^72–74).

Results

We thus used a bootstrapping strategy to identify weak yet reliable signals that are important when examining cortical–subcortical connections in brain-wide analyses. For each of the 1,000 iterations, we randomly resampled 80% of the participants (n = 72) and identified, for each seed region, blood oxygen level dependent (BOLD) signal correlations for all voxels in the brain that survived a voxel-wise threshold of P < 0.05. We calculated discovery maps for each seed region that included both cortical and subcortical connections. We also calculated the similarities in the spatial topography among all the maps and subjected each resulting similarity matrix to k-means clustering analysis to characterize the allostatic–interoceptive network. We expected stronger connectivity among cortical seeds compared to subcortical seeds due to the latter’s noisier time courses and potential partial volume effects, which would result in lower correlations for smaller regions.

Cortico-cortical intrinsic connectivity

We first examined the hypothesized functional connectivity according to the published anatomical connections. As expected, we successfully replicated all of the cortico-cortical connections we previously observed with 3 Tesla imaging (Fig. 2 and Supplementary Table 1)¹⁹. In addition, we observed reciprocal intrinsic connectivity (that is, connectivity map of one region includes a cluster in the other region and vice versa) between the lateral vAIns and pACC, between the sgACC and aMCC, and between the dmIns and portions of cingulate cortex (sgACC and pACC; Fig. 2a and see Extended Data Fig. 1 for bootstrapped maps identifying connectivity surviving P value of <0.05 in more than 95% of subsampled analyses), extending the allostatic–interoceptive system to include more of the anatomical connections documented in tract-tracing studies in nonhuman mammals^75–78. All of these observations were confirmed by seed-to-seed connectivity strength calculation (Fig. 2b). Cortico-cortical functional connectivity within the allostatic–interoceptive system, as evidenced by the cortical maps and seed-to-seed connectivity matrix based on our sample, confirmed 100% of the monosynaptic connections documented in published tract-tracing studies of nonhuman animals (see Supplementary Table 1 and references therein).

Fig. 2. Cortico-cortical functional connectivity within the allostatic–interoceptive system.

a, Left column shows cortical seed locations and right column shows group-level t-value maps (n = 90) masked by voxels that showed positive connectivity (two-tailed t test, P < 0.05) with the seed in more than 950 iterations (of 1,000) by resampling 80% of the sample in each iteration (n = 72). b, Seed-to-seed functional connectivity matrix shows connectivity strength between each pair of the cortical seeds (two-tailed t test, P < 0.05, uncorrected; white color indicates correlation = 1; n = 90). c, The allostatic–interoceptive system showed connecting regions in all the a priori interoceptive and visceromotor control regions. Connecting regions belonging to the ‘rich club’ are labeled in yellow. ‘Rich club’ hubs image in panel c is adapted with permission from ref. ¹⁰⁵. lvAIns, lateral vAIns; mIns, mid insula; mvAIns, medial vAIns.

Extended Data Fig. 1. Bootstrapped cortico-cortical functional connectivity maps.

Bootstrapped functional connectivity maps depict all voxels whose time course was correlated (two-tailed t-test, p < 0.05) with that of the seed in more than 950 iterations (out of 1000) by resampling 80% of the sample in each iteration (n = 72). aMCC: anterior midcingulate cortex; dmIns: dorsal mid insula; dpIns: dorsal posterior insula; lvAIns: lateral ventral anterior insula; mvAIns: medial ventral anterior insula; pACC: pregenual anterior cingulate cortex; sgACC: subgenual anterior cingulate cortex.

Next, we binarized the cortical connectivity maps for all cortical seeds (P < 0.05) and computed their conjunction to identify the connecting cortical regions (Fig. 2c). A k-means clustering analysis (optimal k = 2 based on the Calinski–Harabasz criterion⁷⁹) on the cortical maps replicated¹⁹, such that the system included two subsystems, one corresponding to the default mode network (that is, the dorsomedial prefrontal cortex, posterior cingulate cortex (PCC), and dorsolateral prefrontal cortex) and the other corresponding to the salience (that is, anterior to MCC, anterior insula, supramarginal gyrus, supplementary motor area) and somatomotor networks (that is, precentral gyrus, postcentral gyrus, superior temporal gyrus (STG); see details in Fig. 3); this ensemble of brain regions is sometimes referred to as the cingulo-opercular network or the action-mode network⁸⁰. This procedure also enabled us to identify any regions that could be reliably included in the intrinsic connectivity of the system. We replicated all the connecting ‘hub’ regions reported at 3 Tesla discussed in ref. ¹⁹ (that is, portions of aMCC/pMCC, inferior frontal gyrus (IFG), vAIns, dpIns, temporal pole, inferior temporal gyrus, superior temporal sulcus, parahippocampal gyrus (PHG) and cuneus) with the exception of medial postcentral gyrus. We also newly identified the entire ACC (including subgenual and pregenual extents), PCC, a greater extent of the insula (including mid insula), as well as some portions of medial superior frontal gyrus (SFG) and middle frontal gyrus (MFG) as allostatic–interoceptive system ‘hubs’. A majority of the allostatic–interoceptive system’s connecting hubs have been identified as members of the ‘rich club’ in the connectomics literature, defined as high-degree nodes showing denser interconnections among themselves than are lower-degree nodes⁸¹. The rich club hubs have a key role in global information integration across the brain and therefore may serve as the backbone for global communication within the brain⁸², suggesting that allostatic and interoceptive processes may be at the core of the brain’s computational architecture.

Fig. 3. The two large-scale intrinsic networks composing the cortical allostatic–interoceptive system correspond to the default mode network and salience/somatomotor networks.

a, The cortical allostatic–interoceptive system is composed of two large-scale intrinsic networks. The k-means clustering (k = 2, 1,000 iterations) yielded the most optimal solution, where Network 1 (resembling the default mode network) included a cluster of maps seeded in the sgACC, aMCC, pACC and mvAIns, and Network 2 (resembling the salience network) included a cluster of maps seeded in the lvAIns, dmIns and dpIns. All displayed maps result from the conjunction of binarized maps (two-tailed t test, P < 0.05) in the same cluster. Cortical ROIs are outlined in yellow (ROI names are labeled in the top panel). b, We computed Dice overlap between network maps and the Yeo 7-network cortical parcellation¹⁴⁹ using the Network Correspondence Toolbox (https://github.com/rubykong/cbig_network_correspondence)¹⁵⁰. In the grids, cells with significant Dice overlap at P < 0.05 (that is, showing substantial correspondence) are denoted with an asterisk. Network 1 showed significant Dice overlap solely with the default mode network, while Network 2 showed significant Dice overlap with the salience/ventral attention network and somatomotor network.

Subcortico-cortical intrinsic connectivity

In a new analysis that was enabled by newly delineated subcortical seeds^45,48,83 and that presented considerable challenges at 3 Tesla¹⁹ due to its coarser spatial resolution and lower SNR²⁷, we assessed subcortico-cortical connectivity by visually inspecting cortical discovery maps of the subcortical seeds to confirm topography (Fig. 4a and see Extended Data Fig. 2 for bootstrapped maps identifying connectivity surviving P < 0.05 in more than 95% of subsampled analyses) and calculating seed-to-seed connectivity to quantify strength of connection (Fig. 4b). Subcortico-cortical functional connectivity within the allostatic–interoceptive system, as evidenced by the cortical maps and seed-to-seed connectivity matrix based on our sample, confirmed 96% of the monosynaptic connections documented in published tract-tracing studies of nonhuman animals (see Supplementary Table 1 and references therein). There were two exceptions as follows: we did not observe significant, positive functional connectivity between PAG and dmIns/dpIns, or PBN and sgACC, despite known anatomical connections (Supplementary Table 1; refs. ^84,85). In some instances, averaged time courses between seeds did not correlate significantly (that is, gray squares in Fig. 4b, for example, DR–sgACC); however, connectivity clusters could still be observed in the maps (for example, sgACC cluster in DR-seeded map). Such discrepancies can result from noisy signals within an ROI or specific subportions of an ROI showing significant connectivity. We tested specificity of the allostatic–interoceptive network using a region of superior parietal lobule not known for visceromotor function¹⁹. This region only showed consistent functional connectivity to the SC⁸⁶, VSM, the hippocampus and the dAmy (Supplementary Table 2).

Fig. 4. Subcortico-cortical intrinsic connectivity within the allostatic–interoceptive system.

a, Left column shows subcortical seed locations and right column shows group-level t-value maps (n = 90) masked by voxels that showed positive connectivity (two-tailed t test, P < 0.05) with the seed in more than 950 iterations (of 1,000 iterations) by resampling 80% of the sample in each iteration (n = 72). b, Seed-to-seed functional connectivity matrix shows connectivity strength between pairs of subcortical and cortical seeds (two-tailed t test, P < 0.05, uncorrected; gray color indicates subthreshold correlations; n = 90). c, Conjunction map shows the number of binarized maps (two-tailed t test, P < 0.05) with shared connecting regions (ranging from 9 to 14). dACC, dorsal ACC.

Extended Data Fig. 2. Bootstrapped subcortico-cortical functional connectivity maps.

Bootstrapped functional connectivity maps depict all voxels whose time course was correlated (two-tailed t-test, p < 0.05) with that of the seed in more than 950 iterations (out of 1000) by resampling 80% of the sample in each iteration (n = 72). dAmy: dorsal amygdala; DR: dorsal raphe; hippo: hippocampus; hypothal: hypothalamus; LC: locus coeruleus; LGN: lateral geniculate nucleus; mdThal: mediodorsal thalamus; NAcc: nucleus accumbens; PAG: periaqueductal gray; PBN: parabrachial nucleus; SC: superior colliculus; SN: substantia nigra; VSM: medullary viscero-sensory-motor nuclei complex, including the nucleus tractus solitarius, dorsal motor nucleus of the vagus, nucleus ambiguus, and hypoglossal nucleus; VTA: ventral tegmental area.

As with the cortico-cortical analyses, we conjoined the binarized discovery subcortico-cortical maps (P < 0.05) to identify the overlapping cortical connectivity across subcortical seeds (Fig. 4c). Subcortically seeded maps showed connecting regions in hypothesized cingulate and insular regions, as well as in some parts of the MFG and cuneus. We examined a range of k values that showed similarly optimal Calinski–Harabasz criterion (k = 2 to k = 9; Supplementary Note). We retained k = 3 for its interpretability. All three clusters included cortical nodes from the default mode and salience networks. Cluster 1 included discovery maps from seeds in the lower brainstem (LC, PBN, VSM), and primarily showed connectivity to the PCC, supramarginal gyrus and some medial and lateral occipital regions (Fig. 5). Cluster 2 included discovery maps from seeds in the upper brainstem (PAG, DR) and the hypothalamus, and showed connectivity to the aMCC and PHG. Cluster 3 included discovery maps from larger seeds in the mdThal, LGN, hippocampus, dAmy, nucleus accumbens (NAcc), SC, SN and VTA, and showed widespread connectivity to the dorsomedial prefrontal cortex, cingulate cortices (sgACC, pgACC, aMCC, isthmus), supplementary motor area, cuneus, insula (anterior, mid and posterior), SFG, central sulcus and angular gyrus.

Fig. 5. Clustering solution (k = 3) for cortical maps of subcortical allostatic–interoceptive seeds.

Cluster 1 included maps that were seeded in small lower brainstem ROIs (LC, PBN, VSM). Cluster 2 included maps that were seeded in small upper brainstem ROIs (PAG and DR) and the hypothalamus. Cluster 3 included maps that were seeded in larger subcortical seeds (mdThal, LGN, hippocampus, dAmy, NAcc, SC, SN and VTA). All displayed maps result from the conjunction of binarized maps (P < 0.05) in the same cluster. Cortical ROIs are outlined in yellow (ROI names are labeled in the top panel).

Subcortico-subcortical intrinsic connectivity

With our newly delineated subcortical seeds^45,48,83, we also assessed subcortico-subcortical connectivity by visually inspecting subcortical maps of the subcortical seeds to confirm topography (Fig. 6a and see Extended Data Fig. 3 for bootstrapped maps identifying connectivity surviving P < 0.05 in more than 95% of subsampled analyses) and by calculating functional connectivity between all subcortical seeds to quantify strength of connection (Fig. 6b). Again, this analysis presented considerable challenges with 3 Tesla scanning as discussed in ref. ¹⁹. Subcortico-subcortical functional connectivity within the allostatic–interoceptive system confirmed 96% of the monosynaptic connections documented in published tract-tracing studies of nonhuman animals (see Supplementary Table 1 and references therein). There were three exceptions as follows: we did not observe significant, positive functional connectivity between hypothalamus and PBN, hypothalamus and LC, or hypothalamus and VSM, despite known anatomical connections (Supplementary Table 1). In one case, averaged time courses between the VSM and NAcc seeds did not correlate significantly (Fig. 6b, gray square in matrix), but bilateral NAcc clusters could nonetheless be observed in the VSM-seeded map. As in the subcortico-cortical maps, such discrepancies can result from noisy signals within an ROI or specific subportions of an ROI showing significant connectivity. Seed-to-seed connectivity strength between PAG subregions and other subcortical ROIs is displayed in Extended Data Fig. 4. Seed-to-seed connectivity strength between hippocampal subregions and other subcortical ROIs is displayed in Extended Data Fig. 5. Seed-to-seed connectivity strength between layers of the SC and other subcortical ROIs is displayed in Extended Data Fig. 6. Seed-to-seed connectivity strength between hypothalamus subregions and other subcortical ROIs is displayed in Extended Data Fig. 7. Conjoined binarized subcortical discovery maps (P < 0.05) indicated that all but four subcortical seeds showed overlapping connectivity: connecting regions were identified in the mdThal, LGN, hippocampus, dAmy, NAcc, PAG, DR, SC, SN and VTA but hypothalamus, PBN, LC and VSM showed less widespread and dense connectivity throughout subcortical seeds (Supplementary Table 3). The k-means clustering analysis (k = 3) on the subcortical discovery maps from subcortical seeds yielded an almost identical solution as their cortical connectivity maps.

Fig. 6. Subcortico-subcortical intrinsic connectivity within the allostatic–interoceptive system.

a, Left column shows subcortical seed locations and right column shows group-level t-value maps (n = 90) masked by voxels that showed positive connectivity (two-tailed t test, P < 0.05) with the seed in more than 950 iterations (of 1,000 iterations) by resampling 80% of the sample in each iteration (n = 72). b, Seed-to-seed functional connectivity matrix showed connectivity strength between each pair of the subcortical seeds (two-tailed t test, P < 0.05, uncorrected; white color indicates correlation = 1 and gray color indicates subthreshold correlations; n = 90). Several seeds had functional connectivity with a subset of voxels within target ROIs, as shown by binarized maps at P < 0.05 (two-tailed t test; target ROI outline is shown in blue).

Extended Data Fig. 3. Bootstrapped subcortico-subcortical functional connectivity maps.

Bootstrapped functional connectivity depict all voxels whose time course was correlated (two-tailed t-test, p < 0.05) with that of the seed in more than 950 iterations (out of 1000) by resampling 80% of the sample in each iteration (n = 72). dAmy: dorsal amygdala; DR: dorsal raphe; hippo: hippocampus; hypothal: hypothalamus; LC: locus coeruleus; LGN: lateral geniculate nucleus; mdThal: mediodorsal thalamus; NAcc: nucleus accumbens; PAG: periaqueductal gray; PBN: parabrachial nucleus; SC: superior colliculus; SN: substantia nigra; VSM: medullary viscero-sensory-motor nuclei complex, including the nucleus tractus solitarius, dorsal motor nucleus of the vagus, nucleus ambiguus, and hypoglossal nucleus; VTA: ventral tegmental area.

Extended Data Fig. 4. Intrinsic connectivity of PAG and its subregions within the allostatic-interoceptive system.

a, Bootstrapped connectivity maps obtained from resampling 80% of the sample (n = 72) 1000 times. b, Connectivity strength between PAG seeds and all other seeds (two-tailed t-test, p < 0.05, uncorrected; white color indicates correlation = 1 and gray color indicates subthreshold correlations; n = 90). c, Contrasts obtained by paired-sample t-tests between subregional maps (p < 0.05, uncorrected). lPAG and vlPAG showed more robust and more extensive connectivity than dmPAG and dlPAG, with stronger connectivity especially with aMCC, mvAIns and dAmy. aMCC: anterior midcingulate cortex; dAmy: dorsal amygdala; dmIns: dorsal mid insula; dmPAG: dorsomedial periaqueductal gray; dlPAG: dorsolateral periaqueductal gray; dpIns: dorsal posterior insula; DR: dorsal raphe; hippo: hippocampus; hypothal: hypothalamus; LC: locus coeruleus; LGN: lateral geniculte nucleus; lPAG: lateral periaqueductal gray; lvAIns: lateral ventral anterior insula; mdThal: mediodorsal thalamus; mvAIns: medial ventral anterior insula; NAcc: nucleus accumbens; pACC: pregenual anterior cingulate cortex; PAG: periaqueductal gray; PBN: parabrachial nucleus; SC: superior colliculus; sgACC: subgenual anterior cingulate cortex; SN: substantia nigra; vlPAG: ventrolateral periaqueductal gray; VSM: medullary viscero-sensory-motor nuclei complex, including the nucleus tractus solitarius, dorsal motor nucleus of the vagus, nucleus ambiguus, and hypoglossal nucleus; VTA: ventral tegmental area.

Extended Data Fig. 5. Intrinsic connectivity of the hippocampus and its subregions within the allostatic-interoceptive system.

a, Bootstrapped connectivity maps obtained from resampling 80% of the sample (n = 72) 1000 times. b, Seed-to-seed connectivity strength between hippocampal seeds and all other seeds (two-tailed t-test, p < 0.05, uncorrected; white color indicates correlation = 1 and gray color indicates subthreshold correlations; n = 90). c, Contrasts obtained by paired-sample t-tests between subregional maps (two-tailed t-test, p < 0.05, uncorrected). Hippocampal head and body showed stronger connectivity to default mode nodes such as the MPFC, PCC, AG and lateral temporal cortex. Hippocampal body and tail showed stronger connectivity to salience nodes such as ACC, PCC, SMA, MFG and SMG. ACC: anterior cingulate cortex; aMCC: anterior midcingulate cortex; dAmy: dorsal amygdala; dmIns: dorsal mid insula; dpIns: dorsal posterior insula; DR: dorsal raphe; hypothal: hypothalamus; LC: locus coeruleus; LGN: lateral geniculate nucleus; lvAIns: lateral ventral anterior insula; mdThal: mediodorsal thalamus; MFG: middle frontal gyrus; mvAIns: medial ventral anterior insula; NAcc: nucleus accumbens; pACC: pregenual anterior cingulate cortex; PAG: periaqueductal gray; PBN: parabrachial nucleus; PCC: posterior cingulate cortex; SC: superior colliculus; sgACC: subgenual anterior cingulate cortex; SMA: supplementary motor area; SMG: supramarginal gyrus; SN: substantia nigra; VSM: medullary viscero-sensory-motor nuclei complex, including the nucleus tractus solitarius, dorsal motor nucleus of the vagus, nucleus ambiguus, and hypoglossal nucleus; VTA: ventral tegmental area.

Extended Data Fig. 6. Intrinsic connectivity of the superficial and deep layers of the SC within the allostatic-interoceptive system.

a, Bootstrapped connectivity maps obtained from resampling 80% of the sample (n = 72) 1000 times. b, Seed-to-seed connectivity strength between SC subregions and all other seeds (two-tailed t-test, p < 0.05, uncorrected; n = 90). c, Contrasts obtained by paired-sample t-tests between subregional maps (two-tailed t-test, p < 0.05, uncorrected). Superficial SC showed stronger connectivity to primary sensory cortices in the posterior insular, occipital and pericentral regions. Deep SC showed stronger connectivity to allostatic-interoceptive hubs in the vAIns and the entire cingulate cortex. aMCC: anterior midcingulate cortex; dAmy: dorsal amygdala; dmIns: dorsal mid insula; dpIns: dorsal posterior insula; DR: dorsal raphe; hippo: hippocampus; hypothal: hypothalamus; LC: locus coeruleus; LGN: lateral geniculate nucleus; lvAIns: lateral ventral anterior insula; mdThal: mediodorsal thalamus; mvAIns: medial ventral anterior insula; NAcc: nucleus accumbens; pACC: pregenual anterior cingulate cortex; PAG: periaqueductal gray; PBN: parabrachial nucleus; PCC: posterior cingulate cortex; SC: superior colliculus; sgACC: subgenual anterior cingulate cortex; SN: substantia nigra; VSM: medullary viscero-sensory-motor nuclei complex, including the nucleus tractus solitarius, dorsal motor nucleus of the vagus, nucleus ambiguus, and hypoglossal nucleus; VTA: ventral tegmental area.

Extended Data Fig. 7. Intrinsic connectivity of the hypothalamus and its subregions within the allostatic-interoceptive system.

a, Bootstrapped connectivity maps obtained from resampling 80% of the sample (n = 72) 1000 times. b, Seed-to-seed connectivity strength between hypothalamic subregions and all other seeds (two-tailed t-test, p < 0.05, uncorrected; n = 90; white color indicates correlation = 1 and gray color indicates subthreshold correlations; n = 90). c, Contrasts obtained by paired-sample t-tests between subregional maps (two-tailed t-test, p < 0.05, uncorrected). The middle tuberal-posterior and superior communities showed more robust and extensive connectivity to midcingulate cortex and ventral mid insula than the anteroventral-tuberal and anterior communities, and the middle tuberal-posterior community further showed more extensive connectivity to posterior cingulate cortex and posterior insula. aMCC: anterior midcingulate cortex; AT HT: anteroventral-tuberal hypothalamus; dAmy: dorsal amygdala; dmIns: dorsal mid insula; dpIns: dorsal posterior insula; DR: dorsal raphe; hippo: hippocampus; HT: hypothalamus; LC: locus coeruleus; LGN: lateral geniculate nucleus; lvAIns: lateral ventral anterior insula; mdThal: mediodorsal thalamus; MTP HT: middle tuberal-posterior hypothalamus; mvAIns: medial ventral anterior insula; NAcc: nucleus accumbens; pACC: pregenual anterior cingulate cortex; PAG: periaqueductal gray; PBN: parabrachial nucleus; PCC: posterior cingulate cortex; SC: superior colliculus; sgACC: subgenual anterior cingulate cortex; SN: substantia nigra; VSM: medullary viscero-sensory-motor nuclei complex, including the nucleus tractus solitarius, dorsal motor nucleus of the vagus, nucleus ambiguus, and hypoglossal nucleus; VTA: ventral tegmental area.

The allostatic–interoceptive system

We observed dense interconnectivity between all the cortical and subcortical seeds included in our analysis (Fig. 7a). Conjoined binarized discovery maps (P < 0.05) across both cortical and subcortical extents converged in the hypothesized allostatic–interoceptive system (Fig. 7b).

Fig. 7. Summary of the allostatic–interoceptive system based on 7 Tesla fMRI functional connectivity.

a, Circuit diagram indicates dense within-system connectivity between the 21 cortical and subcortical seeds. All seeds are shown as spherical nodes located at their respective centers of gravity. Pairwise connectivity strengths between ROIs are shown as edges between nodes (two-tailed t test, ranging from P < 0.05 in red to P < 10⁻¹⁰ in yellow, uncorrected; n = 90). Nodes and edges in the glass brain were visualized using BrainNet Viewer¹⁵¹. b, Conjunction map shows the number of binarized maps (two-tailed t test, P < 0.05) that shared overlapping regions (ranging from 15 to 21, total number of cortical and subcortical seeds = 21).

Discussion

Ultrahigh-field 7 Tesla fMRI with 1.1-mm isotropic voxel resolution combined with newly delineated 7 Tesla brainstem and diencephalic parcellations^44–48 revealed both cortical and subcortical components of an integrated allostatic–interoceptive system in humans. Our original study applying 3 Tesla fMRI¹⁹ used 10-min resting state scans in two subsamples of 270–280 participants each, as well as a third sample of n = 41, whereas the present study involved a greater duration of resting state scan time (30 min in total) in a sample of 90 participants. Using functional connectivity among 7 cortical ROIs and 14 subcortical ROIs in humans, we verified more than 96% of the anatomical connections identified in published tract-tracing studies of macaques and other nonhuman mammals. Our current 7 Tesla findings revealed reciprocal connectivity between sgACC/pACC and dmIns/dpIns regions that was previously unreported in 3 Tesla functional connectivity studies of the ACC^87–91 and the insula^92–95. The improvement in sgACC connectivity, in particular, was expected at 7 Tesla, as this region is a part of the medial/orbital surface that is typically susceptible to low SNR, partial volume effects and physiological aliasing. In the current study, these effects were mitigated by higher resolution image acquisition at 7 Tesla, minimal smoothing and more precise nuisance regression using signals from individual ventricles. We also expanded observations of the subcortical extents of the system. Several subcortical nodes (that is, mdThal, LGN, hippocampus, dAmy, NAcc, SC, SN and VTA) showed robust connectivity with all cortical nodes, whereas the smaller brainstem nuclei (that is, PAG, DR, PBN, LC and VSM (including the NTS)) showed weaker but reliable connectivity to these nodes, consistent with other studies that examined a subset of the nodes as seeds at 3 Tesla (for example, ref. ⁴⁹) and 7 Tesla (for example, refs. ^30,31,96,97). We also observed reliable connectivity between regions that had not been previously documented as having monosynaptic connections in tract-tracing studies. For example, the LGN has virtually no monosynaptic connectivity with cortical nodes of the allostatic–interoceptive system according to the tract-tracing studies (except for modest projections to the pACC⁶⁹), yet we observed reliable functional connectivity between the LGN and the aMCC, medial vAIns and pACC. The LGN receives interoceptive input⁶⁷ and there is some evidence that interoceptive signals gate visual sensory sampling⁹⁸, suggesting that LGN functional connectivity with other nodes of the allostatic–interoceptive system reflects polysynaptic connections that are functionally meaningful. In our study, the observation of a broad allostatic–interoceptive system is consistent with the confirmed monosynaptic connections between the a priori ROIs and the understanding that functional connectivity may reflect both monosynaptic and polysynaptic connections⁹⁹.

The connecting ‘hub’ regions of the allostatic–interoceptive system observed at 7 Tesla covered all hypothesized cortical regions of interest, including the full extent of primary interoceptive cortex (dpIns, dmIns¹⁵) and the primary visceromotor regions (vAIns, sgACC, pACC and aMCC¹⁰⁰). Several other connecting ‘hub’ regions (MCC, PCC, IFG, PHG, STG) were also observed and we confirmed their anatomical connections to documented allostatic regions in nonhuman animals^{2,84,101–104}. The remaining connecting regions (that is, MFG, SFG, isthmus of the cingulate, cuneus) have not been documented as having monosynaptic anatomical connections to our subcortical and cortical seed regions; their functional connectivity may reflect polysynaptic connections or new connections in humans. Notably, most of the additional connecting regions observed at 7 Tesla (that is, pACC, PCC, isthmus cingulate, SFG, MFG and mid insula; except the sgACC) belong to the ‘rich club’ (the most densely interconnected regions in the cortex and thought to serve as the ‘backbone’ that synchronizes neural communication throughout the brain¹⁰⁵), consistent with the hypothesized central role of the allostatic–interoceptive system as a high-capacity ‘backbone’ for integrating information across the entire brain¹⁰⁶.

The results of this study have several important functional implications. First, several brain regions within the allostatic–interoceptive network likely have a crucial role in coordinating and regulating the body’s systems, although they are also involved in other psychological phenomena. For example, the SC is typically studied for visuomotor functioning in humans but has been shown to be important for approach and avoidance behavior, as well as the accompanying changes in visceromotor activity in nonhuman mammals (for example, refs. ^62,107) through anatomical connections to ACC⁵⁰ and hypothalamus⁵¹. Similarly, the hippocampus is usually considered central to memory function, but evidence from nonhuman animals indicates that the hippocampus also has a role in the regulation of feeding behaviors and in interoception-related reward signals^108–111. There is also circumstantial evidence that interoceptive signals, relayed from the vagus nerve to the hippocampus via the NTS and septal nuclei, may have a role in event segmentation^112,113. Furthermore, the LGN is usually considered part of the visual pathway that relays visual information from the retina and the cerebral cortex. However, the current functional connectivity findings are consistent with tract-tracing evidence showing LGN’s monosynaptic connections with cortical (for example, pACC⁶⁹) and subcortical visceromotor structures (for example, hypothalamus⁶⁸, PAG⁵² and PBN¹¹⁴), suggesting a role for facilitating communication among brain structures implicated in bodily regulation, in addition to its role in integrating interoceptive and visual signals³⁹. The broad functional connectivity profile of the LGN is also consistent with evidence of tracts between the LGN and other hypothesized regions of the allostatic–interoceptive network, such as the hippocampus, amygdala, DR, SC, SN and LC (Supplementary Table 1).

Second, both the default mode and salience networks have been functionally implicated in cardiovascular regulation as well as in other aspects of allostasis^9,115,116, and they have also been implicated in mental and physical illnesses and their comorbidities. Not surprisingly, psychiatric illnesses (for example, depression¹¹⁷, schizophrenia¹¹⁸), neurodevelopmental illnesses (for example, sensory processing disorder/autism spectrum disorder¹¹⁹), neurodegenerative illnesses (for example, dementia/Alzheimer’s disease¹²⁰, Parkinson’s disease¹²¹) and physical illnesses (for example, heart disease¹²², chronic pain¹²³) present with symptoms related to altered interoception or visceromotor control, and some of these symptoms are transdiagnostic^124,125. Moreover, interoceptive and visceromotor symptomatology is often accompanied by altered neurobiology (for example, volume, structural connectivity, functional connectivity, evoked potential, task activation) in the allostatic–interoceptive system (for example, depression¹²⁶, autism¹²⁷, dementia^33,120, chronic pain¹²⁸). In addition, there is evidence showing that psychological therapies targeting interoceptive processes¹²⁹ and neuromodulations targeting distributed regions within the allostatic–interoceptive system^130,131 may be effective transdiagnostic interventions. Taken together, these findings suggest that altered function of the allostatic–interoceptive system may be a transdiagnostic feature of mental and physical illness that holds promising clinical utility. More fundamentally, the system identified in this paper provides a scientific tool for integrating studies across psychological and illness domains in a manner that will speed discovery, the accumulation of knowledge and, potentially, strategies for more effective treatments and prevention.

Finally, the findings reported here are consistent with the growing body of evidence that a number of subcortical and cortical brain regions are important during both the regulation of bodily functions and during cognitive phenomena, calling into question their functional segregation^132–134. Our findings suggest that the default mode and salience networks may be concurrently coordinating, regulating and representing organs and tissues of the internal milieu at the same time that they are engaged in a wide range of tasks spanning cognitive, perceptual, emotion and action domains (Fig. 5 in ref. ¹⁹). Therefore, our results, when situated in the broader published literature, suggest that the default mode and salience networks create a highly connected functional ensemble for integrating information across the brain, with interoceptive and allostatic signaling at the core. Regulation of the body has been largely ignored in the neuroscientific study of the mind, in part, because much of interoceptive modeling occurs outside of human awareness^18,125.

Several limitations within the current study should be addressed in future studies. First, we did not validate the connectivity strength within the allostatic–interoceptive network against signal-based measures of interoception (for example, heart-beat evoked potentials), although there is growing evidence that, even at rest, limbic regions of the brain continually issue allostatic control signals and there should be synchronous relationships between resting state BOLD signals and electrical signals from visceromotor movements¹³⁵. Second, we did not fully monitor participants’ wakefulness (for example, through video recording) during the three 10-min resting state scans. The default mode and salience networks are present during sleep¹³⁶, although the strength of within-network functional connectivity has been shown to vary (with evidence of both stronger and weaker connectivity) as a function of wakefulness^137–140. Third, susceptibility-induced field inhomogeneities and distortions are more pronounced in 7 Tesla echo planar imaging scans. Although we applied bias field correction on the T1-weighted echo planar imaging to help mitigate these effects, further systematic investigation is needed to evaluate their impact on functional connectivity analysis, particularly in regions with high susceptibility, such as the orbitofrontal cortex. Finally, we did not map every relevant subcortical area that may be involved in allostasis or interoception. For example, opportunities for further research include septal nuclei (with direct projections to limbic regions such as the hippocampus and implicated in temporal control of neurons that make up the allostatic–interoceptive network^141,142), circumventricular organs (for example, area postrema with unique access to peripheral signaling molecules through its permeable blood-brain barrier^143,144) and motor brainstem nuclei (for example, dorsal motor nucleus of the vagus and nucleus ambiguus whose neurons give rise to the efferent vagus nerve^145,146). The cerebellum is also likely involved in allostasis and interoception^147,148.

Methods

Participants and MRI acquisition

We recruited 140 native English-speaking adults with normal or corrected-to-normal vision and no history of neurological or psychiatric conditions. All participants provided written informed consent and were compensated in accordance with the guidelines set by the institutional review board of Massachusetts General Hospital. A total of 50 participants were excluded from the current analysis (19 withdrew before the MRI session, 3 withdrew during MRI acquisition due to discomfort, 6 did not complete scans due to online scan reconstruction failure, 3 did not complete scans due to time constraint, 4 were excluded due to other technical issues during acquisition, 10 were excluded due to scanner sequence error, 4 were excluded due to corrupted MRI data that could not be processed and 1 was excluded due to excessive artifacts in the structural scan). This resulted in a final sample of 90 participants (26.9 ± 6.2 years old; 40 females, 50 males). MRI data were acquired using a 7 Tesla scanner (Magnetom, Siemens Healthineers) with a 32-channel phased-array head coil and personalized padding to achieve a tight fit. Participants completed a structural scan, three resting state scans of 10 min each, three diffusion-weighted scans, as well as other tasks unrelated to the current analysis. At the beginning of each resting state scan, participants were instructed to keep their eyes open and indicated their readiness to start the scan by pressing a button. MRI parameters are detailed in Supplementary Note.

Preprocessing of fMRI data

The preprocessing pipeline began with reorientation, slice timing correction, concatenation of all three resting state runs, coregistration to the structural image and motion correction (framewise displacement mean = 0.17, s.d. = 0.14, with 98.7% of frames showing subvoxel motion¹⁵²). We then conducted nuisance regression to remove physiological noise due to motion (six parameters measuring rotation and translation), as well as due to non-BOLD effects evaluated in the white matter, ventricular cerebrospinal fluid and the cerebral aqueduct. We then conducted temporal filtering and normalization. Finally, we performed conversion to Freesurfer orientation/dimensions, detrending, spatial smoothing (1.25 mm) and resampling to cortical surfaces. Preprocessing details are provided in Supplementary Note.

Functional connectivity analysis

Seven cortical seeds (4-mm-radius spheres) were defined based on the previous fMRI studies of interoception using the procedure outlined in ref. ¹⁹. The 14 subcortical seeds were defined based on the Brainstem Navigator toolkit (https://www.nitrc.org/projects/brainstemnavig/; for example, ref. ⁹⁶), CANLAB Combined Atlas 2018 (github.com/canlab) and Freesurfer segmentation (for example, ref. ¹⁵³). See Supplementary Note for details about seed definition. We randomly resampled 80% of the sample (n = 72) 1,000 times. In each iteration, for each seed, we estimated cortical connectivity using Freesurfer-based analysis procedure as outlined in ref. ¹⁹. This yielded final group maps that showed regions whose fluctuations significantly correlated with the seed’s fMRI time series, which were binarized to retain positive connectivity surviving the threshold of P < .05 and summed across 1,000 iterations to obtain ‘bootstrapped connectivity’ maps. We also quantified seed-to-seed functional connectivity by computing Pearson’s correlation coefficient between all pairs of ROIs and applying Fisher’s r-to-z transform. Significance at the group level was assessed with a two-tailed one-sample t test.

Connecting regions and k-means cluster analysis

To visualize the connecting ‘hub’ regions, we combined binarized functional connectivity maps (P < 0.05) for all seeds. To replicate the previously discovered two-subsystem distinction within the allostatic–interoceptive network¹⁹, we first computed a similarity matrix capturing pairwise η² (ref. ¹⁵⁴) between the unthresholded bootstrapped group maps of cortical seeds and then applied k-means clustering algorithm (k-means, MATLAB) with a range of k values between 2 and 10 (for each k, we tested ten initializations with new centroid positions, each with a maximum of 1,000 iterations to find the lowest local minimum for sum of distances). We evaluated the optimal k using the Calinski–Harabasz criterion⁷⁹. To visualize each subsystem, we binarized the group connectivity maps (P < 0.05) and calculated the conjunction between maps within the same cluster.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Online content

Any methods, additional references, Nature Portfolio reporting summaries, source data, extended data, supplementary information, acknowledgements, peer review information; details of author contributions and competing interests; and statements of data and code availability are available at 10.1038/s41593-025-02087-x.

Author contributions

T.D.W., L.L.W., A.B.S., L.F.B. and M.B. designed the research. J.Z., D.C., J.E.T., L.H., K.M.L, K.M., A.B.S., K.S.Q., L.F.B. and M.B. performed the research. J.Z., D.C., P.D., T.S., L.F.B. and M.B. analyzed the data and wrote the original draft of the paper. J.E.T., P.A.K., L.H., K.M.L., K.M., T.D.W., L.L.W., A.B.S., K.S.Q. and S.W.-G. read and approved the final draft of the paper.

Peer review

Peer review information

Nature Neuroscience thanks Hugo Andres Caffaratti, Mara Mather, Christopher Petkov and Prokopis Prokopiou for their contribution to the peer review of this work.

Data availability

Raw and preprocessed data can be found at https://openneuro.org/datasets/ds005747.

Code availability

Analysis outputs and codes can be found at https://github.com/jiahez/7-Tesla-Allostatic-Interoceptive-System.

Competing interests

The authors declare no competing interests.

Extended data

is available for this paper at 10.1038/s41593-025-02087-x.

Supplementary information

The online version contains supplementary material available at 10.1038/s41593-025-02087-x.