Thank you for visiting nature.com. You are using a browser version with limited support for CSS. To obtain the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in Internet Explorer). In the meantime, to ensure continued support, we are displaying the site without styles and JavaScript.

# Representation of aversive prediction errors in the human periaqueductal gray

## Abstract

Pain is a primary driver of learning and motivated action. It is also a target of learning, as nociceptive brain responses are shaped by learning processes. We combined an instrumental pain avoidance task with an axiomatic approach to assessing fMRI signals related to prediction errors (PEs), which drive reinforcement-based learning. We found that pain PEs were encoded in the periaqueductal gray (PAG), a structure important for pain control and learning in animal models. Axiomatic tests combined with dynamic causal modeling suggested that ventromedial prefrontal cortex, supported by putamen, provides an expected value–related input to the PAG, which then conveys PE signals to prefrontal regions important for behavioral regulation, including orbitofrontal, anterior mid-cingulate and dorsomedial prefrontal cortices. Thus, pain-related learning involves distinct neural circuitry, with implications for behavior and pain dynamics.

This is a preview of subscription content

## Access options

### Supplementary Figure 4 Study 2: comparison of different pain levels

A) Experimental conditions. In both the control and placebo runs, participants (n = 50) are presented two predictive cues. The low cue is followed 50% of the time by low pain (46°C) and 50% of the time by medium pain (47°C). The high cue is followed 50% of the time by high pain (48°C) and 50% of the time by medium pain (47°C). In placebo runs the thermode is installed on a skin spot pre-treated with a cream participants are told has analgesic properties. B) Periaqueductal gray (PAG) region of interest (ROI). C) Continuous pain ratings from 30 independent subjects for 11-s thermal stimulations at 46.5°C, 47.5°C and 48.5°C. The window of analysis for aversive prediction error signals was set between 4 and 10 seconds, i.e. between the time the temperatures can be differentiated and the peak of pain. D) Axiomatic predictions for aversive prediction error. Axiom #1 stipulates that aversive prediction error signals should increase with temperature intensity, regardless of expectations. Axiom #2 stipulates that lower pain expectations should be associated with higher prediction errors, regardless of temperature. Therefore, axiom #2 can only be tested on the medium temperature. Moreover, if the same prediction error signals are also influenced by instruction-based expectations, we should observe higher activity for the placebo vs. control condition. E) Activity in the PAG during the PE window. The left panel shows a clear effect of temperature (low < medium, medium < high, all p’s < 0.001). The right panel shows effects of cues and condition for stimulations at 47°C, which are in conformity with axioms #2 and 3. Activity in the PAG ROI for medium pain stimulations is higher for low vs. high pain cues (F(1,49) = 4.39, p < 0.05) and for the placebo analgesia condition vs. control (F(1,49) =16.03, p < 0.001). Error bars represent standard errors of the mean.

### Supplementary Figure 5 Activity during decision-making

Activity during decision-making (number or participants = 23) correlating positively (red) or negatively (blue) with the expected value of the chosen option (warm/cold colors indicate low/high subjective (model-based) probability of pain. Displayed activations are cluster-thresholded (p<0.05, FWE, two-tailed) with cluster-defining thresholds of p<0.001, p<0.01 and p<0.05.

### Supplementary Figure 6 DCM optimizing the connectivity of the aversive prediction error structure: step 1a

(A) In all models, the driving inputs are the pain > no stimulus and expected value parametric modulators on outcome onsets. The models tested systematically varied the structure(s) receiving the expected value driving inputs and conveying this information to the midbrain. The model with the highest exceedance probability is highlighted in red. (B) Expected (expected posterior probability) and exceedance (probability compared with other tested models) probabilities associated with each model. Val = expected value, str = striatum, hipp = hippocampus, mb = midbrain. number or participants = 23.

### Supplementary Figure 7 DCM optimizing the connectivity of the aversive prediction error structure: step 1b

(A) In all models, the driving inputs are the pain > no stimulus and expected value parametric modulators on outcome onsets. The model selected in the previous step is the first one (top left). From left to right, the tested models varied hippocampus targets (vmPFC, PAG, or nothing), or added a link between the striatum and midbrain. Models in the second row additionally include expected value as a driving input to the vmPFC, and a link from the vmPFC to the striatum. (B) Expected (expected posterior probability) and exceedance (probability compared with other tested models) probabilities associated with each model. Val = expected value, put = putamen, hipp = hippocampus, PAG = periaqueductal gray. number or participants = 23.

### Supplementary Figure 8 DCM optimizing the connectivity of the aversive prediction error structure: step 2a.

(A) In all models, the structures generating PE signals (put, vmPFC, hipp, PAG) are arranged according to the best model selected from the previous model selection steps. The links between these structures and the pain-specific PE structures are systematically varied. The model with the highest exceedance probability is highlighted in red. (B) Expected (expected posterior probability) and exceedance (probability compared with other tested models) probabilities associated with each model. Val = expected value, put = putamen, hipp = hippocampus, PAG = periaqueductal gray, OFC = orbitofrontal cortex, aMCC = anterior cingulate cortex, dmPFC = dorsomedial prefrontal cortex. number or participants = 23.

### Supplementary Figure 9 DCM optimizing the connectivity of the aversive prediction error structure: step 2b.

(A) Modulatory influences were systematically added to the connections from the striatum or midbrain to the OFC or aMCC. (B) Expected (expected posterior probability) and exceedance (probability compared with other tested models) probabilities associated with each model. Val = expected value, str = striatum, hipp = hippocampus, mb = midbrain, OFC = orbitofrontal cortex, aMCC = anterior cingulate cortex, dmPFC = dorsomedial prefrontal cortex.

### Supplementary Figure 10 Probabilities associated with each option

The four sets of random walks used in the current study. Probabilities associated with each option (blue and red lines) varied independently and slowly from trial-to-trial according to random walks. Probabilities were bounded 20% and 80%, and had to cross at least once over the course of the experiment.

## Supplementary information

### Supplementary Text and Figures

Supplementary Figures 1–11 (PDF 1123 kb)

(PDF 512 kb)

(XLSX 40 kb)

(XLSX 36 kb)

(XLSX 31 kb)

## Rights and permissions

Reprints and Permissions

Roy, M., Shohamy, D., Daw, N. et al. Representation of aversive prediction errors in the human periaqueductal gray. Nat Neurosci 17, 1607–1612 (2014). https://doi.org/10.1038/nn.3832

• Accepted:

• Published:

• Issue Date:

• DOI: https://doi.org/10.1038/nn.3832

• ### Anatomical dissociation of intracerebral signals for reward and punishment prediction errors in humans

• Maëlle C. M. Gueguen
• Alizée Lopez-Persem
• Julien Bastin

Nature Communications (2021)

• ### The influence of subcortical shortcuts on disordered sensory and cognitive processing

• Raymond J. Dolan
• Marta I. Garrido

Nature Reviews Neuroscience (2020)

• ### Bidirectional control of fear memories by cerebellar neurons projecting to the ventrolateral periaqueductal grey

• Jimena Laura Frontera
• Hind Baba Aissa
• Daniela Popa

Nature Communications (2020)

• ### A neuronal circuit for activating descending modulation of neuropathic pain

• Junting Huang
• Gerald W. Zamponi

Nature Neuroscience (2019)

• ### Altered resting state functional connectivity of the cognitive control network in fibromyalgia and the modulation effect of mind-body intervention

• Jian Kong
• Emily Wolcott
• Chenchen Wang

Brain Imaging and Behavior (2019)