SUPPLEMENTARY INFORMATION: Evidence for category-dependent and category-independent goal- value codes in human ventromedial prefrontal cortex Daniel McNamee1,2, Antonio Rangel2,3, John P. O’Doherty1,2,3 1 Trinity College Institute of Neuroscience, Trinity College Dublin, Dublin, Ireland 2 Computation and Neural Systems, California Institute of Technology, Pasadena, USA 3 Division of Humanities and Social Sciences, California Institute of Technology, Pasadena, USA Contact: Daniel McNamee Caltech, MC 136-93, Pasadena, CA 91125, USA E-mail: [email protected] CONTENTS SUPPLEMENTARY FIGURES ..................................................................................................... 2 Figure S1. Masks Covering Distinct Subregions of vmPFC .................................................. 2 Figure S2. Independent Replication of Main Results ............................................................. 3 Figure S3. Leave-One-Participant-Out Anterior/Posterior mOFC Gradient Analysis............. 5 Figure S4. Item Ratings . ....................................................................................................... 7 Figure S5. Leave-One-Participant-Out Ventral/Dorsal vmPFC Gradient Analysis ................ 9 Figure S6. Value Decoding Based on “Mean-Subtraction” Searchlight ............................... 10 SUPPLEMENTARY TABLES ..................................................................................................... 12 Table S1. fMRI Results ........................................................................................................... 12 Table S2. Item Ratings, Subject-Level Analysis ..................................................................... 14 Table S3. Item Ratings, Distribution Per Category Across Subjects ....................................... 14 Table S4. Item Ratings, Distribution Per Category Across Items ............................................ 14 Table S5. Items Used Organized By Category . ..................................................................... 15 Table S6. Items Used In Chib et al., 2009 Organized By Category ........................................ 17 1 SUPPLEMENTARY FIGURES x = –3 y = 43 mOFC mPFC Figure S1. Masks Covering Distinct Subregions of vmPFC. Based on previous functional and anatomical results, our a priori hypothesis was that distributed and univariate value encoding signals would be found in vmPFC extending from the orbital surface to more dorsal regions up to and including parts of Brodmann areas 10 and 32. Due to the similarity of the experimental design, we used univariate peak coordinates from a related study (Chib et al., 2009) to construct a medial prefrontal cortex (mPFC) mask as a sphere of radius of 9mm surrounding these peak coordinates (corresponding to the size of the multivariate searchlight sphere). A similar functional mask did not exist for the medial orbitofrontal cortex (mOFC); most likely due to the distributed nature of the value codes found there and the relative scarcity of MVPA studies in value-based decision-making thus the mOFC mask was constructed according to anatomical descriptions used previously in the literature (Beckmann et al., 2009). This mask encompassed the medial orbital and olfactory sulci bilaterally with the anterior and posterior limits defined by the extents of these sulci. The vmPFC mask was defined as a union of these two masks. 2 a b x = 9 x = –14 Food category-dependent goal value Univariate goal value (conjunction) Trinkets category-dependent goal value Category-independent goal value (conjunction) Figure S2. Independent Replication of Main Results. For an independent replication of our results, we applied our analysis procedures to the data acquired for a previous study (Chib et al., 2009) with a similar task paradigm but with some important differences. This study also used a BDM auction process to elicit the participants’ willingness-to-pay (WTP) on an integer scale from $0 to $4 for a variety of items drawn from three categories (food, monetary sums, and “trinkets”). However, the WTP bids (that is, the goal values) for all the items were recorded before the participant entered the scanner. Subsequently, on each trial in the scanner, participants were required to make binary choices between an item and a fixed reference sum of money (the median bid over all items placed during the pre-scan behavioral experiment). The motor response performed was a left or right button press and was completely uncorrelated with both the choice and the value of the item since the item and the reference sum of money were randomly presented to the left and right of a fixation cross. Choosing the item meant that the participant paid the reference price in exchange for an 80% chance of receiving the item. If they chose the reference amount of money, they would neither pay anything nor have the opportunity to play the lottery. The analyses in the original study indicated that the value of the lottery item on each trial was commonly represented (as a smoothed univariate BOLD response) in a dorsal portion of vmPFC for all three item categories. This value representation was interpreted as a “decision value” signal (as opposed to a goal value in the paradigm used in the current study) since it is being computed in order to make a binary decision choice. In light our results, we hypothesized 3 that distributed value signals, both category-dependent and category-independent, would accompany this smoothed value signal in the ventral and dorsal portions of vmPFC respectively. More specifically, we expected to see an anterior/posterior dissociation in category-dependent value signals along the medial orbital surface, whereby food value would be located more posteriorly and trinkets more anteriorly. We performed the same value decoding analyses as described in the main text on this dataset (19 participants; 15 male; mean age, 23.7; age range, 18-47). a, A food-category-dependent value representation was located in posterior mOFC (peak [x = 3, y = 33, z = –24], t = 2.86; P < 0.05 SVFWE, small volume familywise error corrected, within a 20mm-radius sphere centered on the peak food value coordinates from the study reported in the main text). A category-independent value signal (conjunction across training/testing on food/trinkets and trinkets/food respectively) was located in mPFC (peak [x = 6, y = 57, z = 12], t = 1.98, P < 0.05 uncorrected). b, A trinket category-dependent value representation was located in anterior mOFC (peak [x = –15, y = 57, z = –9], t = 2.94; P < 0.05 SVFWE within a 20mm-radius sphere centered on the peak trinket value coordinates from the study reported in the main text). No clusters were present in any unanticipated ROI (e.g. a trinket category-dependent value signal where food category-dependent signals were found in the primary dataset). Results are shown at P < 0.005 and P < 0.05 uncorrected and overlaid on an averaged structural image. These results provide a robust independent replication of the anterior vs. posterior vmPFC value coding effects observed in the main study, with the ventral vs. dorsal effect also evident (although the multivariate category-independent result in dorsal vmPFC was uncorrected for multiple comparisons). 4 Figure S3. Leave-One-Participant-Out Anterior/Posterior mOFC Gradient Analysis. Here we replicate the anterior/posterior mOFC gradients identified in the main text in a completely non-circular manner using ANOVA interaction tests applied to per-subject classification scores derived using a leave-one-participant-out approach. For each subject and item category, we first performed second-level mass-univariate t- tests on the classification maps for 12 participants only (leaving one participant out). The peak t- score coordinate within the mOFC ROI was identified and the classification score for the left-out participant at the peak coordinate was recorded. In addition, the peak t-score from the alternative item category analysis within a searchlight sphere of voxels (restricted to the mOFC ROI) surrounding that peak coordinate was also taken. For example, for each subject we recorded two food value classification scores: (1) one based on the peak coordinate in mOFC and (2) the other based on the peak coordinate within a searchlight sphere of the peak coordinate from the trinket value decoding. Similarly, two trinket value classification scores were also acquired for each subject. In this way, for each item category and subject, we independently derived a classification score and then also recorded a classification score for the alternative item category within the same locality. This process was repeated for every subject in both analyses being contrasted. The end result was a dataset comprised of four classification scores for each subject derived in a completely independent manner. 5 The data was entered into a repeated measures 2 (cid:3400) 2 ANOVA design (spatial location x item category) and there was a significant interaction between the two factors (P = 0.039) whereby the trinket-category-dependent value encoding signal was stronger in the region identified more anteriorly but not posteriorly and vice versa for the analogous food-related signal. This replicates the corresponding result in the main text (Fig. 2b) in a completely independent manner. In this figure, the simple main effect of spatial location on classification score is plotted across item category i.e. the distribution of the relative differences in t-scores between the anterior and posterior ROIs (food items in blue, trinkets in red). Error bars reflect standard error of the mean. 6 Figure S4. Item Ratings. We acquired post-hoc behavioral ratings of the food and trinket items used from 9/13 of the original participants. One participant did not complete the questionnaire leaving 8/13 to be analyzed. The items were rated on five scales – “valence”, “intensity”, “liking”, “accessibility”, and “familiarity” from a score of 1 to 7. Items were presented in a random fashion across categories. Specifically, the questions were: LIKING – how much do you like this item? A score of 1 means “I do not like this item at all”, a score of 4 means “I neither like nor dislike this item”, while a score of 7 means “I really like this item a lot”. FAMILIARITY – how familiar are you with this item? A score of 1 means “This item is unknown to me”, a score of 4 means that “I am somewhat familiar with this item”, while a score of 7 means “I’m completely familiar with this item”. INTENSITY – how intense are the feelings evoked by this item? A score of 1 indicates “This item evokes no feelings or emotion for me”, a score of 4 “I have some feelings towards this item”, while a score of 7 means “I have very intense feelings towards this item”. Note that for this question, it is irrelevant whether the feelings/emotions you have are positive or negative. 7 ACCESSIBILITY – how easy do you feel it is for you to obtain this item? A score of 1 means “It is almost impossible for me to get this item”, a score of 4 means “I can get this item without much difficulty”, while a score of 7 means “I would have no problem getting this item”. VALENCE – how pleasant or unpleasant is this item? A score of 1 means “It is a very unpleasant item”, a score of 4 means “This item is neither pleasant nor unpleasant”, while a score of 7 means “This is a very pleasant item”. The point-biserial correlation r is the Pearson correlation between item ratings and the (cid:2926)(cid:2912) dichotomous variable indicating whether the item is a food item or a trinket. It describes to what extent higher or lower ratings are correlated with trinkets or food items. Positive correlations indicate that higher ratings correlate with trinkets; negative correlations indicate that higher ratings correlate with food items. A zero correlation implies that the ratings are evenly matched across items. Results of statistical analyses can be seen in Table S2, Table S3, and Table S4. At P > 0.05, there was no significant difference between food and trinket items with respect to any rating (across subjects or across items). In two ratings (“intensity” and “familiarity”), there was a trend towards higher ratings in the food category. The subject-level point-biserial correlation showed that this was a weak effect within individual subjects with only one subject reaching a P < 0.05 significance threshold for each rating. The bar chart in this figure reflects the point-biserial correlation coefficients (cid:1870) for each (cid:3043)(cid:3029) subject between item ratings and a dichotomous variable which indicated whether the item was drawn from the food or trinket category. Repeated-measure statistical tests of any ratings difference between the food and trinkets category were not significant (P > 0.05). As can be seen from this figure, there is a high degree of variability within and across subjects in these ratings indicating that they are unlikely to account for the gradient effects reported in the main analyses. 8 Figure S5. Leave-One-Participant-Out Ventral/Dorsal vmPFC Gradient Analysis. Here we replicate the ventral/dorsal vmPFC gradients identified in the main text in a completely non-circular manner using ANOVA interaction tests applied to per-subject value representation scores derived using a leave-one-participant-out approach. For each analysis, we first performed second-level mass-univariate t-tests on the multivariate classification maps and general linear modeling beta maps for 12 participants only, leaving the 13th subject out. The peak t-score coordinate within each vmPFC ROI was identified and a value representation score (classification score for the multivariate analyses or first-level GLM t-score for the univariate analyses) for the left-out participant at the peak coordinate was recorded. This process was repeated for every subject in for both the food and trinket item categories. The end result was a dataset comprised of four classification scores for each subject derived in a completely independent manner. Since we seek to compare results across encoding strategies, we standardized these results by computing the distribution of standardized value signal differences between the ventral and dorsal ROIs for each item category and encoding strategy. That is, we subtracted the mPFC scores from the mOFC scores and then divided by the standard deviation across both ROIs. This data is plotted in this figure. The data was then entered into a repeated measures 2 (cid:3400) 2 ANOVA design (spatial location (cid:3400) encoding strategy) and there was a significant interaction between the two factors for both categories (P = 0.0008 for food, P = 0.0005 for trinkets) whereby there was a greater drop in signal strength in mOFC compared to mPFC for univariate encoding as opposed to multivariate encoding. This replicates the corresponding result in the main text (Fig. 4a) in a completely independent manner. 9