Consistent with previous studies using the same methods (25, 26), this analysis involved the following steps. (i) Identifying a suitable network of interest for pain and disgust: For this purpose, the automated meta-analysis toolbox Neurosynth ( (39) was used to a create a mask of regions preferentially implicated in studies about pain (18,759 coordinates at a 2 mm × 2 mm × 2 mm resolution) or by chemosensory disgust (combining the terms “disgust,” “olfactory,” and “taste”; 4443 coordinates). (ii) Identifying an independent dataset in which neural responses to pain and disgust are estimated in the absence of any previous dilemma (training sample): For this purpose, we took advantage of our previous study, in which an independent group of participants was subjected to thermal (painful) or olfactory (disgusting) responses, with three district unpleasantness levels (low, medium, and high) matched between the two modalities (18). The trial structure (expectancy cue, sniffing event, stimuli duration, ramp-up of thermal responses) was almost identical to that of the reference trials from the present study. (iii) Extracting pain and disgust data (18) from the meta-analytically defined masks. The extracted values were fed into a principal components analysis to identify a limited number of components that retained ~99.9% of the variance of the original data (95 components for pain, 96 components for disgust). (iv) Feeding the components to a machine learning algorithm for the prediction of individual unpleasantness ratings. The algorithm’s proficiency was assessed through leave-one-out cross-validation to ensure prediction for an independent group of subjects than the ones used for the modeling. Among the different algorithms implemented (see the Supplementary Materials), the best prediction of either pain or disgust was provided by a support vector machine regression under a radial basis function kernel as implemented in the LIBSVM 3.18 software ( (v) Identifying the regions contributing the most to each model, by assessing the impact of the removal of each feature on the model’s predictive ability (40). This led to a contribution map, whose values were assessed statistically through bootstrap techniques (10,000 resamples).

The estimated pain and disgust signatures (available under Open Science Framework at: were applied to fMRI data from the present experiment. As a first step, we assessed whether each model could distinguish between corresponding conditions in the reference trials, with the pain signature successfully distinguishing between HP and LP and the disgust signature distinguishing between HD and LD. Subsequently, each model was used to assess the neural activity in postdilemma trials for each of the corresponding modality, and the resulting prediction values were fed into a linear mixed model probing for differences between unpleasantness and dilemma.