Neural Circuits of Interoception (Berntson & Khalsa 2021)
The anatomical quadrant of the Trends in Neurosciences 44(1) special issue — computational is Petzschner et al., clinical is Bonaz et al., functional is Quigley et al., interventional is Weng et al., and this is the wiring. It is cited as reference 3 or 5 by most of the others as the source of the afferent/efferent anatomy they lean on, so it arrives at the end of the queue having already been quoted at the wiki four times.
It is also the paper that makes the wiki’s anatomy start one stage earlier than it has been.
The continuum, and why the wiki was starting in the middle
Figure 1 lays interoception out as five stages, each feeding the next:
sensors → pathways → networks → circuits → awareness
- sensors — mechano-, thermo-, chemo-, osmo-, gluco-, humoral receptors, free nerve endings
- pathways — vagal, cranial, sacral, spinothalamic, somatosensory
- networks — central and peripheral autonomic, enteric, thalamocortical, hypothalamic, limbic, sensorimotor, salience, default
- circuits — appetitive, affective, arousal, thermal, nociceptive, cognitive, social, threat
- awareness — detection, attention, insight, magnitude, discrimination, accuracy, sensibility
Two things follow that this wiki should absorb.
First, its coverage has been pathways-and-upward. lamina-i-spinothalamocortical-pathway begins at lamina I; insular-cortex and central-autonomic-network are networks; nearly every method page measures awareness. The transducers themselves had no page. That gap is now interoceptive-sensors, and it matters because several of these receptors were characterized in the last decade — the afferent inventory is not a settled classical anatomy the field can take as given.
Second, and more pointedly: awareness is the last stage, and most of the process never reaches it. The authors are explicit that humans “rarely have high-resolution access to what is happening in the body at any moment,” that conscious interoception arises mainly on perturbation (quickened breath, surging heartbeat, full bladder) or on threat (angina, nausea, a kidney stone), and that some organs — liver, kidney — are never felt at all. The whole apparatus of taxonomy and heartbeat tasks the wiki is built on operates on the thin conscious tip of a system whose business is conducted elsewhere.
Sherrington’s boundary, and where it has moved
The historical opening is worth keeping because the wiki’s definitions all inherit from it. Sherrington (1906) coined interoceptive for senses reporting the body’s internal state, contrasted with exteroceptive sensing that yields projicience — sensory experience projected outward onto an object in the world. He included taste.
Since then the term has “considerably broadened”: visceroception would name the classic view, and interoception now covers the emotional and cognitive sequelae of, and contributions to, internal states. The authors’ own glossary definition is essentially the roadmap’s — sensing, interpreting, integrating, “across conscious and nonconscious levels” — which is unsurprising with Khalsa as second author.
The interesting residue is gustation, which Sherrington counted in and the field mostly counts out. Berntson & Khalsa decline to settle it and instead offer a criterion worth recording: oral taste receptors sense stimuli outside the body’s surface boundary, but ingestion is a decision to move a substance into the body, so swallowing may be “a functional demarcation of the transition to an interoceptive signal.” Gustatory afferents project to the NTS like other visceral afferents. Meanwhile functional bitter-taste receptors (TAS2Rs) and olfactory receptors sit on lung smooth muscle, vascular smooth muscle, and cortical neurons — chemoreception detached from the mouth entirely, with ligands possibly bacterial. The interoceptive/exteroceptive line, on this evidence, is not drawn by receptor type.
Compare Petzschner et al.’s proposal to draw it by inferred state rather than channel. Berntson & Khalsa draw it, tentatively, by an act — the swallow. Neither is the channel-based criterion the wiki has been assuming.
The gut section is the strongest thing in the paper
The one place a genuine mechanism replaced a placeholder in the last five years, and the paper knows it: until 2015 no direct innervation of an epithelial sensory cell had been described. Gut luminal stimuli were assumed to reach nerves only by slow paracrine hormone action (CCK, 5-HT, PYY).
Enteroendocrine cells turn out to be sensory transducers with synapses, now called neuropod-cells. Over two-thirds contact sensory nerves in the intestinal mucosa; in the colon they synapse via serotonin onto spinal nerves; in 2018 they were shown to synapse onto nodose neurons, connecting the intestinal lumen to the NTS across a single vagal synapse, transducing a luminal stimulus in as little as 60 ms.
And the behavioural payoff (Buchanan et al. 2020): silence neuropod cells and a mouse can no longer distinguish sucrose from sucralose; stimulate them and it consumes sucralose as though it were caloric sugar. Two transmitters carry the distinction — glutamate for sugars, ATP for sweeteners.
Why this matters beyond gastroenterology is on neuropod-cells. In brief: the wiki’s slow, diffuse, hormonally-mediated picture of gut interoception now has a fast, wired, transmitter-specific channel running alongside it, and an innate preference turns out to be set by a sensory cell in the gut rather than by the tongue or by reward circuitry.
The efferent limb acquires cortical detail — rabies tracing
The wiki’s central-autonomic-network page has held the descending hierarchy as a table of levels. This review fills the top row with a technique the wiki had not met: retrograde transneuronal rabies virus tracing, which walks backward from an organ through however many synapses one waits for.
- Adrenal medulla, monkey (Dum et al. 2016): three distinct cortical networks on lateral and medial prefrontal surfaces, each with a human equivalent — movement, cognition, affect — reaching the gland in three synapses.
- Adrenal medulla, rat (Dum et al. 2019): far more limited, originating mainly in primary/secondary motor cortex with lesser sensory and insular contribution. The cortical areas that are the monkey’s major source of cognitive control are absent from the rat’s adrenal network.
- Stomach, rat (Levinthal & Strick 2020): parasympathetic-linked neurons project from rostral insula and mPFC in three synapses; sympathetic-linked neurons from motor and primary somatosensory cortex in four. Two cortical systems, one organ, different branches.
The species result deserves emphasis because the wiki’s circuit-level evidence is overwhelmingly rodent. If the primate brain–body network is categorically broader at exactly the cognitive-control areas that carry the field’s stress, appraisal and belief stories, then rodent negative results about cortical visceral control are weak evidence about humans, and rodent positive results about lower circuits may transfer better than either. Petzschner et al. flagged one instance of this (parabrachial→insula/vmPFC in rats but not monkeys); here it is a general finding with a method behind it.
The cross-species gap, stated as the field’s structural problem
Figure 2 is a two-panel diagram of an evidence asymmetry rather than of a brain:
| human work | animal work | |
|---|---|---|
| cortical/limbic | well characterized | thin |
| body → brainstem (NTS, PB) | thin | well detailed |
Neither literature covers the join. The wiki has been meeting this piecemeal for a dozen ingests — rodent circuit work quoted in the same paragraph as human patient findings (Bonaz et al.), species-mixed functional claims (Quigley et al.), the “afferents are too thin to stimulate” methodological complaint. This is the underlying shape of all of it: the two halves of the interoceptive hierarchy are known in different species, and the middle is inferred.
Three of the paper’s four Outstanding Questions are versions of this.
Efferent control of endocrine and immune systems
Endocrine. The hypothalamus as the survival hub — energy metabolism, thermoregulation, feeding, fluid balance, sleep–wake, reproduction, stress — with the parabrachial nuclei as the relay carrying visceral and nociceptive afference up and control signals back down. The item with theoretical weight: hypothalamic control of homeostasis is “classically viewed as reactive,” but there is growing appreciation it can be both reactive and predictive (Burdakov 2019, on orexin/hypocretin neurons) — with the caveat that this is established mostly for non-neuroendocrine functions, and that identifying a genuinely predictive hypothalamic controller of endocrine function “would require rethinking of the impact of this brain region on allostatic processes.” That lands squarely on the open seam recorded at homeostasis.
Immune. The reflex arc: peripheral inflammation → vagal afferents → NTS → ascending to hypothalamus, amygdala, striatum, cingulate, insula; and laterally to DMN/nucleus ambiguus/RVLM as the efferent limb. Stimulating mouse RVLM or vagal efferents suppresses innate immune responses and downregulates splenic proinflammatory cytokines cholinergically. Humoral mediators (IFN-α, IL-6) and immune cells enter directly through circumventricular organs. And a striking microglial result: reducing sensory neuronal activity — by general anaesthesia or by whisker trimming — increases microglial process surveillance, under noradrenergic β-adrenergic control and only in awake mice (Liu et al. 2019). Sensory traffic modulates immune surveillance in the brain. See sickness-behaviors.
The authors’ own verdict on the precise role of the ascending targets in immune regulation: “unclear.”
What it contributes to the predictive debate
Two items, pulling in opposite directions, which is why the page treats it as evidence rather than as a position.
For prediction, from an unexpected species. Hunger/thirst-dependent activity patterns in mouse insular cortex shift in anticipation of expected satiety of food or water signals (Livneh et al. 2020) — “suggesting that mice can generate and modify interoceptive predictions in response to ongoing contextual changes in the external environment.” Every other line of evidence on feedforward-vs-predictive-interoception is human imaging, cytoarchitecture, or theory. This is a behavioural-neural demonstration of anticipatory interoceptive representation in an animal that can be manipulated.
Against settling it. The paper’s own framing of the question is deliberately unresolved: determining “the relative contributions of bottom-up versus top-down signals to interoception and body regulation is thus an important challenge facing the field,” and it lists the manipulations available — increasing interoceptive input, dampening exteroceptive input (floatation-rest), synchronizing exteroceptive stimuli with cardiac timing, direct circuit stimulation, false feedback, environmental context. It then adds a methodological warning the wiki should carry: “perceptual sensitivity for one signal may not generalize to others” (Ferentzi et al. 2018), and interoceptive processing can affect exteroceptive perception (Motyka et al. 2019). And it doubts reductionism will suffice, recommending a dynamic systems approach — the first appearance in this wiki of a serious alternative to both the read-out and the predictive-coding pictures.
A quiet constraint on the animal-feelings question
“There are currently no animal models of interoceptive awareness,” though insular homology between humans and non-human primates may make one worth attempting.
Set that beside the Livneh mouse result above and the position is sharp: an animal can hold and update an anticipatory representation of its internal state, and we have no way to ask whether anything is felt. That is can-we-know-animal-feelings restated as an experimental fact rather than a philosophical one, and it is the same boundary Craig drew anatomically (direct lamina I→VMpo→insula only in primates) from the opposite direction — Craig inferring absence of feeling from missing hardware, Berntson & Khalsa reporting absence of a measure.
Reading it against Craig
The wiki’s afferent anatomy is Craig’s, and this review is broader and flatter. Where Craig (2002) identifies one privileged pathway (lamina I → VMpo → dorsal posterior insula) whose existence reclassifies pain, temperature and itch as interoceptive, Berntson & Khalsa list the lamina I/spinothalamic route as one of several alongside vagal, cranial (V, VII, IX), sacral (S2–S4), humoral, and direct central chemoreception by neurons and astrocytes — plus a direct NTS→insula/SII vagal-activated pathway in macaque (Strigo & Craig 2016, so Craig’s own later work).
This is not a contradiction and is recorded as none: Craig’s claim was about a dedicated, modality-labelled, phylogenetically new pathway, not about exclusivity. But it changes the emphasis the wiki has been carrying. On the Craig pages the lamina I pathway is the substrate of interoception; here it is the best-characterized lane on a road with several, some of which (astrocytic pCO2/pH/Na⁺ sensing, humoral action on circumventricular organs) bypass afferent nerves altogether. See lamina-i-spinothalamocortical-pathway, interoceptive-sensors.
Cannon, corrected
A small historical note the paper turns into a thesis. Cannon called the system “autonomic” because it acts “automatically, without direction from the cerebral cortex.” The review’s whole efferent section is the accumulated refutation: three prefrontal networks reach the adrenal medulla; distinct cortical networks drive the stomach’s two autonomic limbs; insular cortex shifts feeding by aversive sensing and anticipates satiety. The wiki holds Cannon on walter-cannon chiefly as the critic of James; this is a second and independent place his framework has been overturned, and it is the premise the central-autonomic-network page exists to replace.
Placement
Read as the wiki’s anatomical floor plan rather than as an argument. Its contribution is inventory, technique, and a clear-eyed statement of what is not known: the multisynaptic organ-to-cortex maps, the NTS-to-cortex relay in humans, how multiple signals with different temporal dynamics are integrated, and what determines whether any of it reaches awareness. Those four Outstanding Questions are a fair summary of the gaps the rest of this wiki keeps running into from above.