The medial frontal cortex, including anterior midcingulate cortex, has been linked to multiple psychological domains, including cognitive control, pain, and emotion. However, it is unclear whether this region encodes representations of these domains that are generalizable across studies and subdomains. Additionally, if there are generalizable representations, do they reflect a single underlying process shared across domains or multiple domain-specific processes? We decomposed multivariate patterns of functional MRI activity from 270 participants across 18 studies into study-specific, subdomain-specific, and domain-specific components and identified latent multivariate representations that generalized across subdomains but were specific to each domain. Pain representations were localized to anterior midcingulate cortex, negative emotion representations to ventromedial prefrontal cortex, and cognitive control representations to portions of the dorsal midcingulate. These findings provide evidence for medial frontal cortex representations that generalize across studies and subdomains but are specific to distinct psychological domains rather than reducible to a single underlying process.
Publisher’s note: Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations
We thank S. Fukudo, T. Muratsubaki and J. Morishita for assistance with data collection; K. Ochsner for sharing data from studies of negative emotion; T. Braver and J. Gray for sharing working memory data; and R. Poldrack for sharing response selection data (available at https://openfmri.org/). This research was supported by grants R01 HL089850 to P.J.G.; P01 HL040962 to S.B.M.; grants OCI-1131801, R01 DA035484, and R01 MH076136 to T.D.W.; JSPS-FWO grant VS.014.13 N to L.V.O. and S. Fukudo; JSPS-KAKENHI grant 26460898 to M.K.; R01 MH076137 and R01 AG043463 to K.O.; by the Direction de la Recherche Clinique of the University Hospital of Grenoble Alpes; and by the pharmaceutical labs Ferring and Cephalon. L.V.O. is funded by the KU Leuven Special Research Fund. T.E.N. is supported by the Wellcome Trust.
Integrated supplementary information
a Force-directed graph conveying the observed similarity of brain activation across MFC (see Fig. 1b) using the Fruchterman-Reingold algorithm. Each small circle corresponds to a brain activation pattern from a single subject. The number next to each circle indicates the study it came from, and the color of each circle indicates domain membership of each study. Large circles depict the mean location of contrasts from each domain. b Results from modeling the similarity of activation patterns spanning the full MFC. The model explains clustering of brain activation patterns as a combination of study (top 18 rows), subdomain (middle nine rows), and domain-specific effects (bottom three rows). Smoothed bootstrap distributions of the generalization index (b = 5,000 bootstrap samples, drawn by random resampling of n = 270 participants) are plotted for each term and indicate the extent to which contrasts that share this feature are more (or less) similar. In general, there is greater clustering of domains and individual studies. For example, the domain of negative emotion is distinct from the domains of pain and cognitive control, yet the subdomain of social emotion does not form a clear cluster, because studies 15 and 16 are dissimilar from one another. *FDR q < .05; + P < .05 uncorrected.
Supplementary Figure 2 Variance Inflation Factors (VIFs) for representational similarity-based models.
VIFs indicate how much multicollinearity in a design matrix has inflated the variance of parameter estimates. All values here are below the standard cutoff of 10 (see Ray, W.D. Applied Linear Statistical-Models, 3rd Edition - Neter,J, Wasserman,W, Kutner,Mh. J Oper Res Soc 42, 815–815 (1991).).
Supplementary Figure 3 Regional analyses using different model parameterizations yield similar results.
a Model parameterized to capture increases in dissimilarity (1–Pearson′s r, n = 270 participants) for patterns of brain activity observed in subjects engaged in different studies, subdomains, or domains. The average within study dissimilarity serves as a reference and is modeled with a constant term. b Model parameterized to capture increases in similarity (Pearson′s r, n = 270 participants) for patterns of brain activity from the same study, subdomain, or domain. The average between domain similarity serves as a reference and is modeled with a constant term. c Model parameterized to capture differences in similarity (Pearson′s r, n = 270 participants) for patterns of brain activity from the same study, subdomain, or domain versus those that come from different studies, subdomains, or domains. The average overall similarity serves as a reference and is modeled with a constant term. d-f Bootstrap distributions (b = 5,000 bootstrap samples, drawn by random resampling of n = 270 participants) of the generalization index for terms modeling ‘pain’, ‘cognitive control’, and ‘negative emotion’. *FDR q < .05 corrected. +P < .05 uncorrected.
Colors indicate Bayesian information criterion weights, which reflect the relative evidence in favor of generalizable representation of pain (red), cognitive control (green), and negative emotion (blue). The colormap ranges from values of 0 to 1 for each model and the colors are additive. Purple regions indicate brain regions that exhibit equivocal evidence for representation of pain and negative emotion, and transparent (gray) regions do not exhibit evidence for the representation of a single domain. For full details see Table S6.
Supplementary Figure 5 Evaluation of model bias and variance, as well as false positive rates using resting-state fMRI (rsfMRI) data.
a Null-hypothesis representational similarity analysis (RSA) modeling procedure for a single Monte Carlo iteration (this procedure was repeated 1,000 times). Mirroring our experimental procedure, rsfMRI data was sampled from 18 sites from the 1,000 functional connectomes project70 (15 subjects per site, total n = 270). In each iteration, a random GLM was fit for each subject, producing 270 independent activation maps. These maps were used to estimate a ‘null’ representational dissimilarity matrix that served as the outcome for RSA-based models. The mean and standard deviation of parameter estimates from these models (βnull) were computed to index the bias and variance of our modeling procedure. To provide an estimate of the false positive rate (FPR), P-values were calculated on each iteration using bootstrap resampling. The FPR was calculated as the proportion of significant effects across the 1,000 iterations. b Bias and variance of the modeling procedure. Violin plots indicate the distribution of parameter estimates from null models (across 1,000 MC iterations), error bars indicate the mean and bootstrap standard error. Results show little bias, with all bootstrap distributions centered near the expected value of zero. c Estimated FPR for each region of interest. Significant effects were identified using a threshold α = .05 on each MC iteration. Dashed lines indicate the nominal rate of 0.05, which is the expected frequency for a single test using a threshold of α = .05. Error bars reflect standard error of the mean based on a binomial distribution. The conservative false positive rates are likely due to the dependence structure in the dissimilarity measurements that we do not explicitly account for, but none the less is evidence of a valid statistical procedure.
Supplementary Figure 6 Evaluation of model sensitivity and false positive rates by generating synthetic data with Wishart noise.
a Mirroring our experimental procedure, synthetic data was generated with a covariance structure with unique effects for each of 270 “subjects” and either 3 domains or 18 studies (not shown). On each of 500 Monte Carlo (MC) iterations, a Wishart random matrix was generated from a pre-specified covariance structure and converted to a representational dissimilarity matrix that served as the outcome for representational similarity analysis-based models. P-values were calculated on each iteration using bootstrap resampling to provide estimates of the true and the false positive rates. This procedure was repeated three times with different levels of signal (the covariance across studies and domains was set to values of 0.1, 0.2, and 0.3). b Estimates of true and false positive rates for simulations where domain effects are present. Error bars reflect standard error of the mean based on a binomial distribution. c Estimates of true and false positive rates for simulations where study effects are present. Significant effects were identified using a threshold α = .05 on each MC iteration. Error bars reflect standard error of the mean based on a binomial distribution. Sensitivity increases monotonically with the true effect size, and false positives are at or below the nominal value.
Supplementary Figure 7 Methods for comparing the similarity of patterns of brain activity within and across studies, using bootstrap resampling to derive P-values and make inferences.
Unique pairwise correlations between patterns of brain activity from 270 subjects are displayed in each panel. This illustration shows actual data from the anterior midcingulate cortex. a Full inter-subject correlation matrix where the shaded dark region indicates across-domain correlations. The average similarity across domains is estimated as the average inter-subject correlation in this area. b Regions of the inter-subject correlation matrix averaged to compute domain-general effects of pain (red), cognitive control (green), negative emotion (blue). These effects are computed by averaging correlations within domains but across different subdomains. c Regions used to compute the average correlation within domains but across different studies. d Regions used to compute the average correlation within domains. e For bootstrap tests, we resampled individual correlation coefficients from all 10 regions of the correlation matrix (here shown for the region with pairs of contrasts that are within the pain domain, but from different subdomains). The solid line indicates the sample mean correlation coefficient; the dashed lines indicate 95% confidence intervals of the mean, computed by bootstrap resampling (full bootstrap distribution shown). f Comparisons of the means between different regions reveal generalizable domain effects b > a, average effects of subdomain c > b, and average effects of study d > c. Differences in the means of two regions of the correlation matrix are compared against zero using the same bootstrap procedure (here showing a comparison of the average similarity within the pain domain and from different subdomains versus the similarity between pain and other domains).