685: Layer-Dependent Insular Circuitry Supports Reinforcement Learning�Based Modulation of Pain

Topic: Experimental (Non-clinical) Human Studies
Subtopic: Non -specific to any diagnosis, Placebo/Nocebo, Pain mechanisms (e.g. descending inhibitory control, placebo, inflammation), Imaging (e.g. fMRI, PET, MRS)

Background and Aims: The Reinforcement Learning (RL) framework of pain proposes that nociceptive perception is dynamically optimized to minimize anticipated future harm. Emerging anatomical models, including the Embodied Predictive Interoceptive Coding (EPIC) framework, posit that this optimization is implemented within a hierarchical, layer-dependent architecture of the insular cortex. Cytoarchitectural work suggests that superficial layers of the hypergranular posterior insula (pINS) preferentially project to deep layers of the agranular anterior insula (aINS), supporting feedforward prediction signalling. While this laminar organization has been demonstrated in non-human primates, its functional implementation in humans—and its relevance to pain learning—remains unclear.

Methods: Two datasets were utilized for this study. First, ~20 hours of 0.84 mm isotropic whole-brain layer-fMRI were acquired from a single participant during a naturalistic movie-watching paradigm and layer-specific functional connectivity between pINS and aINS.
Second, five participants completed four experimental visits involving a two-choice associative pain-learning task. Visual cues probabilistically determined thermal stimulation under varying levels of certainty (pain probability) and controllability (free vs forced choice). Across visits, 270–285 minutes of 0.8 mm isotropic 2D-EPI BOLD data were acquired per participant. Learning dynamics were modelled using the Rescorla–Wagner algorithm to derive trial-wise Q-values indexing associative strength.

Results: In the naturalistic dataset, we observed significant layer-dependent asymmetry: superficial pINS layers showed robust functional coupling with deep aINS layers (r = 0.12–0.32, p < 0.0001), whereas the reverse pattern (superficial aINS to deep pINS) was weak or absent.
In the pain-learning dataset, mixed-effects modelling revealed a significant association between Q-values and pain ratings stratified by temperature (low: β = −4.770, p < 0.001; high: β = −5.648, p < 0.001), indicating expectation-driven attenuation of perceived pain.
Pain-evoked activation revealed significant recruitment of middle and superficial pINS layers and deep aINS layers during thermal stimulation (FDR-corrected p < 0.01).

Conclusions: Across independent datasets, we identify convergent evidence for a layer-dependent posterior-to-anterior insular signalling pathway in humans. Superficial pINS to deep aINS connectivity appears to provide a circuit substrate through which learned expectations modulate nociceptive processing. By linking reinforcement-learning dynamics to laminar functional architecture, this work advances a circuit-level account of predictive pain modulation grounded in hierarchical interoceptive processing.

References: Seymour B. Pain: A Precision Signal for Reinforcement Learning and Control. Neuron. 2019 Mar 20;101(6):1029-1041. doi: 10.1016/j.neuron.2019.01.055. PMID: 30897355.

Barrett LF, Simmons WK. Interoceptive predictions in the brain. Nat Rev Neurosci. 2015 Jul;16(7):419-29. doi: 10.1038/nrn3950. Epub 2015 May 28. PMID: 26016744; PMCID: PMC4731102.

Presenting Author: Gianluca Guglietti
Submitting Author: Gianluca Guglietti

Senior Author: Etienne Vachon-Presseau

IASP 2026 Global Year theme of Neuropathic Pain: No

• Gianluca Guglietti - McGill University
• Amelia Lozano-Beckman - McGill University
• Sacha Goodson - McGill University
• Matt Fillingim - McGill University
• Laurentius Huber - Harvard University
• Christine Tardif - McGill University
• Mathieu Roy - McGill University
• Etienne Vachon-Presseau - McGill University