Introduction

Cognitive processing is tightly modulated by emotions1,2: cues with emotional value capture faster attentional resources, and they enhance long-term memory2. Furthermore, emotion and feelings have profound impact on processes of decision making and reasoning2. Based on these interactions, cognition and emotion are viewed as emerging from the function of integrated brain networks, rather than separate brain subsystems3. The issue of how such network integration is articulated for specific brain functions is a current topic of research, motivated by advances in functional neuroimaging techniques and computational tools3,4,5,6,7. In this study we investigated the brain circuit basis of interactions between sadness and working memory (WM).

Negative affect and anxiety are known to impair WM performance8,9,10,11,12,13. This has been associated in neuroimaging studies with deactivation of cortical areas typically associated with WM in the prefrontal and parietal cortices and with the inverse activation of ventral regions typically associated with emotional processing8,9,14,15. However, there is a strong integration between the “emotional brain” and the “cognitive brain” in most daily activities3,16,17 so that this segregation is generally blurry except for in the most extreme conditions such as in depressed patients18. Inducing recapitulation of a sadness experience in healthy subjects achieves a pattern of brain activations consistent with the strong opposition between emotional and cognitive circuits observed in depressed patients19,20. Dorsal areas are typically deactivated (dorsolateral prefrontal cortex (dlPFC), dorsal anterior cingulate cortex (dACC) and parietal cortex) and ventral and subcortical areas are activated (insula, orbitofrontal cortex (OFC), subgenual anterior cingulate cortex (sACC), amygdala, hippocampus), while some other areas may play the role of interconnecting these two networks (basal ganglia, thalamus, rostral ACC)19,20. At present, it is unclear whether these integration-segregation dynamics occur diffusely between all these participating areas, or whether there are specific areas that channel primarily these interactions. Deep brain stimulation (DBS) effectiveness in treatment-resistant depression but only at very specific points of the cingulo-frontal bundle suggests the presence of key nodes in this distributed network21,22,23. Here, we study the topology of sadness-WM interactions within these brain circuits using network analysis of blood-oxygen level dependent (BOLD) time series with graph theory, while participants engage in a WM task in control and sadness conditions.

Graph theory provides systematic procedures to define the internal structure of brain networks based on BOLD data of different nodes or regions of interest4,5,24. Specifically, modularity measures identify subnetworks of nodes (communities) by maximizing the contrast between internal and external couplings, and degree and participation coefficient label each node’s role in coordinating activity within or across communities. Of particular interest are connector hub nodes, characterized by having high degree and participation coefficient, which are likely to regulate the flow and integration of information between the communities25.

Given the tight interactions between emotion and cognition1,2 and the reported inconsistencies of small regions related to sadness26, we hypothesized that a paradigm with a strong conflicting emotional (sadness) and cognitive (WM) demand would enhance the modularity of implicated brain networks in healthy participants and thus reveal the cortical areas that act as network connector hubs.

We collected functional magnetic resonance imaging (fMRI) data from 22 healthy subjects performing this task. Based on these activation patterns we extracted 20 regions of interest (ROI) on a subject-by-subject basis and we computed the correlations between fMRI time series in pairs of ROIs, obtaining the matrix of functional connectivity for each subject on which we applied network measures from graph theory27. We found that the dlPFC acted as connector hub of the dorsal subnetwork, and medial fronto-polar cortex (mFP) was the connector hub of the ventral subnetwork, but they both interacted via the sACC in the ventral subnetwork, and these connectivity patterns were associated with the intensity of the sadness experienced by the participants.

Methods

Participants

Twenty-two healthy participants (28.9 ± 3.9 years of age, mean ± s.d., 10 males) without any psychiatric, neurological or medical illness were recruited. All participants were screened with the Mini-International Neuropsychiatric Interview (M.I.N.I.) to specifically ensure the absence of any ICD-10 psychiatric disorder28 as well as those using psychoactive medications. All subjects were screened with Charlson comorbidity index29. All volunteers had normal or corrected-to-normal vision and were right-handed, native Spanish speakers. The study was carried out in accordance with ethical guidelines and deontological criteria established by the Hospital Clínic of Barcelona, as approved by its Ethics Committee of Clinical Research (Ref: 2009/4886) and written informed consent was obtained from all participants.

Experimental Design

The study was composed of two different behavioral paradigms run immediately in sequence in the scanner (Fig. 1).

Figure 1
figure 1

The two behavioral paradigms: Neutral-WM1 and Sadness-WM2. (A) The first paradigm is composed of two tasks: a neutral epoch followed by 20 trials of spatial working memory with a filtering component. (B) In the second paradigm, the participants underwent a sadness provocation task. When maximal sadness was achieved, participants closed their eyes for a 2-min scan, which was then followed by another 20 trials of spatial working memory with a filtering component. Note that in both paradigms the same stimuli were presented, the only difference being the sadness induced before in the second paradigm. Gray lines mark scanner acquisition periods (320 s).

Full size image

In the first paradigm (Neutral-WM1), subjects were first instructed to rest in a “neutral emotional state” while keeping their eyes closed in a 2-min neutral-epoch scan. Following this resting condition, participants engaged in a spatial working memory task (WM1) with a filtering component30. In each of 20 working memory trials, 5 bright stimuli were presented on a black screen for 1 s, while participants were instructed to maintain fixation on a central cross. Stimuli appeared at random locations of equal eccentricity (6 deg. of visual angle) on a grid of 16 possible positions30. Of these stimuli, three were red dots and two were yellow dots. Participants were instructed to remember only the position of the red dots. Stimuli were followed by a dark screen with a fixation central cross during a delay of 4 s and then a probe stimulus was displayed for 4 s at one grid location (Fig. 1). At that point, participants pressed one of two buttons with the index or middle finger of their right hand to indicate whether one of the red dots had been presented at the location indicated by the probe or not, respectively.

In 25% of the trials, the probe appeared at a location previously occupied by a yellow dot. In these trials, an error response of the participants (yes response) indicated a failure to apply the filtering component of the task. We define this type of error as a cognitive inhibition error30,31.

In the second paradigm (Sadness-WM2), participants performed an emotional task, the sadness provocation task (SP)32. All subjects had prepared in advance a short autobiographical narrative of personal events in which they felt particularly sad, e.g. sad experiences most commonly centered on the loss of relatives, friends, or significant relationships. Before scanning, the narrative texts were presented on the screen and subjects were asked to generate a state of sadness comparable to that originally experienced. After the maximum mood intensity was achieved, participants pressed a button, closed their eyes and were instructed to stop visualizing, thinking, or ruminating on the text and to focus on their feelings of sadness while we acquired fMRI data in a 2-min sadness-epoch scan. Following the sadness epoch, and without stopping the scanner, participants performed a second spatial working memory task (WM2) with filtering component, following the description in the first paradigm. After the scan session, participants reported their subjective rating of the sadness experienced on a 0–7 point scale (5.6 ± 0.17, mean ± standard error of the mean (SEM))33.

All participants trained for all tasks before the scanner session by running successive WM blocks (20 trials) until they stabilized the number of errors (average of 3 training blocks per participant). This training phase was entered to minimize learning effects between the two WM blocks in our successive paradigms (indeed WM performance and reaction times did not improve in the population, see Results). As one additional control for training effects, we derived for each participant a training index as the difference in mean reaction times between WM2 and WM1 sessions, divided by their sum, which was used as an explanatory factor in our analyses. A positive value of this training index indicated faster responses in WM2 relative to WM1, and thus quantified the effective training for a given subject. We did not reverse the order of the paradigms to avoid that sadness, being a particularly persistent emotion34 affected a subsequent Neutral-WM1 paradigm35.

Behavioral analysis

Participants were divided into two groups depending on their sadness intensity: “high-sadness group” were those whose subjective sadness rating ranked above the overall mean (n = 12, mean = 6.14, median = 6, range 6–7) and “low-sadness group” were those participants whose rating ranked below the mean (n = 10, mean = 4.85, median = 5, range 4–5.5). The distribution of sadness scores showed indications of a possible bimodal structure (Hartigan’s dip-test of unimodality: dip = 0.114, p = 0.015; Supplementary Fig. S1). Two participants were excluded from the behavioral analysis because they reported difficulty in distinguishing the color of the dots during the WM tasks (1 from the high-sadness group, and 1 from the low-sadness group). We measured WM performance with the fraction of errors, the reaction times and the fraction of cognitive inhibition errors for each subject in WM1 and WM2.

fMRI acquisition

Brain images were acquired on a 3 Tesla TimTrio scanner (Siemens, Erlangen, Germany) using the 8-channel phased-array head coil supplied by the vendor. A custom-built head holder was used to prevent head movement, and earplugs were used to attenuate scanner noise. High-resolution three-dimensional T1-weighted magnetization prepared rapid acquisition gradient echo (MPRAGE) images were acquired for anatomic reference (TR = 2200 ms, TE = 3 ms, FA = 7°, 1.0 mm isotropic voxels). A T2-weighted scan was used in order to detect possible pathological features (TR = 3780 ms, TE = 96 ms, FA = 120°, voxel size 0.8 × 0.6 × 3.0 mm, 3.0 mm thick, 0.3 mm gap between slices, 40 axial slices). Functional data were acquired using a gradient-echo echo-planar pulse sequence sensitive to blood oxygenation level-dependent (BOLD) contrast (TR = 2000 ms, TE = 30 ms, FA = 85°, 3.0 mm isotropic voxels, 3.0 mm thick, no gap between slices). Presentation® software and data acquisition were synchronized to stimulus pulse sent by the scanner. Participants were requested to avoid moving during the whole MRI scan. The total duration of uninterrupted scanning time during each of the two behavioral paradigms (Fig. 1) was 320 s, i.e. 160 volumes.

fMRI data analysis

Preprocessing and statistical analysis were carried out with SPM8 (Wellcome Trust Centre for Neuroimaging, http://www.fil.ion.ucl.ac.uk/spm). The images were manually aligned along the anterior commissure-posterior commissure line. Preprocessing included the realignment of the scans for motor correction and the normalization to the Montreal Neurological Institute (MNI) template (interpolating to 3 mm cubic voxels). For GLM analyses we further applied spatial smoothing with a Gaussian kernel of 10 mm (we also tried Gaussian kernels of 8 mm and 6 mm, which did not affect the GLM results). For the functional connectivity analysis the spatial smoothing was not applied because we averaged all voxels within each ROI prior to the connectivity analyses.

A random-effect, epoch-related statistical analysis was performed in a two-level procedure. At the first level, a general linear model (GLM) was estimated by using regressors for each instruction condition (before WM trials), neutral and sadness epochs, and fixation period, memory stimulus, delay period and probe stimulus (Fig. 1). Regressors were convolved with the canonical hemodynamic response function in SPM8. The data were high-pass filtered (128 s cutoff) to remove low-frequency drifts. Images from contrasts of interest for each participant were used in a second-level analysis, treating participants as a random effect.

A paired sample t-test was used to investigate the resulting statistical maps for the contrast delay-fixation in WM1 and WM2. The voxel significance was evaluated in a whole-brain analysis testing the global null hypothesis that delay-fixation showed no significant activation. This analysis was corrected for multiple comparisons (false discovery rate (FDR), P < 0.05) and it identified 10 different cortical areas implicated in cognitive processing in this task (Table 1).

Table 1 Regions identified in the task-based analyses.
Full size table

Subsequently, a paired sample t-test was used to statistically assess the difference between delay activity in WM1 and WM2. A mask was created including the activated areas in both WM1 and WM2 in order to compare the difference in the level of activation between WM1 and WM2. The voxel significance was evaluated in the mask testing the global null hypothesis that delayWM1- delayWM2 did not show significant activation. This analysis was corrected for multiple comparisons in the working memory mask (false discovery rate (FDR), P < 0.05).

We used a 2-factor ANOVA to statistically assess the interaction between delay activity in the two tasks (delayWM1 and delayWM2) and sadness intensity (high-sadness and low-sadness groups). The voxel significance was evaluated in a mask testing the global null hypothesis that delayWM1 - delayWM2 and high-sadness - low-sadness groups did not show significant interactions.

We further used a 2-factor ANOVA to statistically assess the interaction between brain activity in the passive conditions (Neutral and Sadness) and sadness intensity (high-sadness and low-sadness groups). Voxel significance was evaluated in a whole-brain analysis testing the global null hypothesis that the interaction between Sadness/Neutral and high-sadness/low-sadness groups was not significant. This analysis led us to identify 9 different cortical and subcortical areas implicated in emotional processing in this task (Punc < 0.05) (Table 1).

For the sACC, which shows significant interaction (small volume correction, 5 mm square at −6 21 −9, FWE-corrected, p = 0.038), a paired sample t-test was used to statistically assess the difference between Sadness and Neutral epochs in the high-sadness group.

Functional connectivity

Regions of interest (ROIs) were determined from global analysis and literature comparison as indicated in Results (see list of ROIs in Table 1).

Some areas known to be related to emotional processing did not survive the correction by multiple comparisons; medial orbitofrontal gyrus left, subgenual anterior cingulate cortex right, medial frontal pole bilateral, hippocampus bilateral and amygdala bilateral. Medial orbitofrontal gyrus right did not show significant activation and was added as the laterally symmetric counterpart of the medial orbitofrontal gyrus left. The locations of the ROIs taken around the peak activations of our contrasts were consistent with coordinates in the literature32,36,37,38,39,40. Each ROI was defined as a 5 × 5 × 5 voxel cube centered around the detected peak activations (for coordinates see Table 1).

The ROI signals were obtained by linear detrending preprocessed data without spatial smoothing for each voxel, and then by averaging across all voxels within the ROI. We removed covariations common to all ROIs by applying a signal regression41,42. This procedure removes global fluctuations related to physiological artifacts such as heart rate, respiration, and scanner noise that are seen throughout the brain artificially, but it could also introduce spurious anticorrelations (See Discussion)41,43,44.

To focus on temporal fluctuations of the BOLD signal not related to the imposed structure of the paradigms (two tasks (SP/resting and WM), change of task in a time scale of 2 min) ROI signals were band-pass filtered in the range 0.018–0.26 Hz (i.e. maintaining temporal fluctuations in time scales from 4 seconds to 1 minute). Functional connectivity between areas was computed with the Pearson correlation coefficient between the 320-sec signals for each pair of ROIs in our database (Table 1), separately for each of the two behavioral paradigms (Fig. 1).

We tested the functional significance within our task of correlations between ROIs identified as hubs of the network (see below) with ANOVA tests. For sACCl-dlPFCl and sACCl-mFPl correlations, we ran a 3-factor ANOVA tests with the factors: paradigm (Neutral-WM1 vs. Sadness-WM2), sadness intensity (high-sadness vs. low-sadness groups) and participant identity as a random factor. The interaction between sadness intensity and paradigm was significant for both connections (false discovery rate (FDR), P < 0.05; p = 0.0006 for sACCl-dlPFCl; p = 0.0434 for sACCl-mFPl), so we separated the data for each group and we performed a paired sample t-test comparing correlations for different sadness groups and paradigms.

Graph analysis

For each subject and behavioral paradigm, the correlation matrix between our ROIs is the adjacency matrix of the weighted graph that represents the corresponding brain network4. The symmetrical adjacency matrix resulting from our undirected graph was characterized for having positive and negative weights. We used algorithms adapted to this type of data using standard graph theory methods on Matlab (Brain Connectivity Toolbox developed by O. Sporns, Indiana University, Bloomington, IN; (https://sites.google.com/site/bctnet/)27,45.

We calculated the community structure from the mean correlation matrix across subjects. We followed this approach, commonly used in the literature45, because it proved to be more robust than community analyses on single-participants. Indeed, when we tried to identify consensus community partitions from individual participants (similarity between partitions taken as the z-score of the Rand coefficient, as described in Doron et al.46, code in http://commdetect.weebly.com) we did not identify well-known anatomical distinctions (cortical vs. subcortical) that the mean correlation robustly detected. To identify the best partition in subnetworks (communities), we quantified its modularity by a quality function Q, which we optimized47,48. Q is a scalar value that measures the density of links inside communities as compared to links between communities. Large values of Q reflect more segregation, or equivalently, decreased integration, between different communities. The best partition is defined as the set of communities with the largest modularity Q. Our modularity analysis for signed adjacency matrices has two free parameters (resolution parameters γ+ and γ−, see ref.47) that allow different clustering thresholds for the positive and negative correlations in the connectivity. By using a range of values for γ− (with γ + = 1) we scanned different layers in the community structure. We illustrate this parametric analysis of modularity in Supplementary Figs S2 and S6. To choose a specific set of resolution parameters for further analysis of the community structure, a common approach is to use Q optimization or cross-validation of the resulting partition with independent structural data49. We used resolution parameters (γ+ = 1, γ− = 0.45) in our analyses (Results) because these provided separate partitions for bilateral cortical and subcortical structures and because this choice of parameters achieved the largest contrast between Q computed in the Neutral-WM1 and Sadness-WM2 periods (Supplementary Fig. S2). We validated that all our results are consistent within a broad range of values for the γ− parameter.

Once we calculated the community structure, we measured the degree and the participation coefficient of each node (i.e. ROI in Table 1). The degree of a node is the number of connections to that node. There are algorithms that take into account the weight of the connectivity links to evaluate the degree, and we validated that for our data they yielded parallel results, so we stayed with the binary definition for its simplicity. The degree has a straightforward interpretation: nodes with a high degree are interacting, structurally or functionally, with many other nodes in the network. In our weighted graph, we validated a connection if its corresponding absolute weight exceeded a pre-defined threshold relative to the absolute strongest correlations. We chose this threshold as the one that provided the largest variance in degree across nodes. For both Resting-WM1 and Sadness-WM2, the threshold providing maximal degree distinction between nodes fell around 35% (Supplementary Fig. S3). In the results, we consider high-degree areas when their degree is greater than the network mean degree plus one standard deviation. A permutation test on our data validated that this threshold selected areas with degree beyond the 95 percentile of resamples assuming the null hypothesis of equal degree for all areas. High-degree areas are candidates to be defined as hubs of the networks, as previously argued in the literature25. The Participation coefficient is a measure of diversity of intermodular connections of individual nodes, it compares the degree of a given node to the number of connections within its own subnetwork. This measure requires a previously determined community structure (see above)25.

Statistical significance of group differences was assessed with permutation tests, where we permuted randomly the assignment of data to each group and we repeated the measure 1,000 times. We reported a significant difference if the difference between the actual groups was larger that 95% of the samples generated randomly (p < 0.05).

Data availability statement

The datasets generated and analysed in the current study are available from the corresponding author on reasonable request.

Results

We recorded fMRI brain activity in 22 participants while they engaged in two identical working memory tasks (WM1, WM2), separated by a period in which a sadness state was induced by remembering a previously identified biographical sketch (Fig. 1). After the scan session, participants reported their subjective rating of the sadness intensity. We sought to identify the functional changes induced by the sadness state in brain networks engaged in regulating the interaction between cognition and emotion.

Behavioral analysis

Across participants, the mean number of error trials, of cognitive inhibition error trials (see Experimental design), and the mean reaction times did not change significantly from the working memory session before sadness induction (WM1) to the working memory session after sadness induction (WM2) (Table 2, n = 20).

Table 2 Behavioral measures.
Full size table

As defined in our analysis, we split participants according to their subjective report on sadness intensity (see Behavioral analysis and Supplementary Fig. S1). We found that participants in the high-sadness group diminished their WM performance following sadness provocation (Fig. 2A, 3-way ANOVA with factors high-sadness/low-sadness, WM1/WM2 and subject identity, n = 20, p = 0.04 for the interaction between high-sadness/low-sadness groups and WM1/WM2, paired sample t-test p = 0.033 for WM1-WM2 errors in the high-sadness group, p = 0.37 for WM1-WM2 errors in the low-sadness group). Neither reaction times, accuracies nor inhibition errors differed between the high-sadness and low-sadness groups prior to sadness induction (two-sample t-test p = 0.91, p = 0.79, p = 0.65, respectively). We checked that these results were unlikely to result from a random assignment of participants to the high-sadness/low-sadness groups with a permutation test. The interaction term of the 3-way ANOVA was more significant in the true classification than in 96.84% of the 10,000 random permutations, and the one-tailed paired sample t-test for WM1-WM2 errors in the high-sadness group was also more significant than in 96.3% of permutations. This underscores the validity of this subjective report, so that we used it in the following to identify effects associated specifically with the experience of sadness.

Figure 2
figure 2

Sadness intensity disrupted WM performance and activated the sACCl. (A) Mean number of error trials during working memory tasks: The high-sadness group presented more errors in WM2 relative to WM1. Error bars are SEM (B) A significant interaction between Sadness/Neutral and high-sadness/low-sadness groups in sACCl (ANOVA T-contrast, 6-voxel cluster, peak activation at −6 21 −9). (C) Significant sACCl activation during sadness relative to neutral states (Fig. 1) in the high-sadness group (20-voxel cluster, peak activation at −3 21 −9). The intersection of lines marks the peak cluster activation of B (coordinates −6 21 −9).

Full size image

In addition, we defined a training index (Methods) for each participant to test for possible training effects. Across participants, the training index was not correlated with sadness score (Pearson correlation R = 0.11, p = 0.65, N = 20). Also, the training index did not correlate with measures collected in WM1 (reaction time: R = −0.36, p = 0.11; accuracy: R = 0.29, p = 0.22; inhibition errors: R = −0.17, p = 0.47; N = 20), indicating it did not relate to intrinsic behavioral differences in the participants.

fMRI BOLD response during working memory

We first identified the cortical areas that supported the memory component of the working memory task in the cognitive network during WM1 and WM2. We did this by assessing BOLD activation in the delay period of the trial, compared to the fixation period before cue presentation. These task periods differed only in the mnemonic load of the delay and therefore their activation contrast was maximal in those areas putatively implicated in working memory. We conducted a whole brain analysis to find regions activated in a delay-fixation contrast (see fMRI data analysis). We found significant activation (False Discovery Rate, FDR, p < 0.05, Supplementary Fig. S4) in dorsal areas typically associated with cognition: dorsolateral prefrontal cortex (dlPFC), intraparietal sulcus (IPS), medial superior frontal gyrus (mSFG), postcentral gyrus (PCG) and inferior frontal gyrus (iFG) (Supplementary Fig. S4A for WM1 and Supplementary Fig. S4B for WM2, for coordinates see Table 1). We tested changes in delay activity in these areas by applying a mask on a delayWM1 − delayWM2 contrast (Supplementary Fig. S4C), and we found a significant reduction in activity during delayWM2 in all these areas (FDR, p < 0.05). However, we could not attribute this decrease unambiguously to sadness experience, as we could not find an interaction between the factors high-sadness/low-sadness and WM1/WM2 in none of these areas (2-way ANOVA, FDR p < 0.05, mask with areas in Supplementary Fig. S4A,B).

fMRI BOLD activity related to sadness experience

We then looked for the anatomical regions activated during the Sadness epoch (Fig. 1). Previous studies have pointed to the sACC as an area involved in sadness processing. Across all participants, we did not find a significant activation in the sACC, or in any other area, in the Sadness epoch relative to the Neutral epoch (2-way ANOVA whole brain analysis with factors epoch and sadness intensity, T-contrast, FDR p < 0.05). Nevertheless, it is known that factors associated with individual differences at both neuroanatomical and behavioral levels may account for the difficulty in detecting sACC activation26. We thus resorted to a region of interest (ROI) analysis, where we defined the sACC ROI (125-voxel cube, center in Table 1) based on available evidence from previous neuroimaging studies32,37,38. Using this ROI as a mask in the above analysis, we found a significant interaction between Sadness/Neutral epoch and high-sadness/low-sadness groups in the left hemisphere (sACCl, Fig. 2B, cluster of 6 voxels, peak activation at −6 21 −9, 5-mm-square small volume SVC and FWE corrections, p = 0.038). Sadness provocation evoked an increase in the BOLD signal in the sACCl during Sadness compared with Neutral in the high-sadness group (Fig. 2C, cluster of 20 voxels, peak activation at −3 21 −9, SVC and FWE, p = 0.014). In other words, subjects who achieved an intense sadness state activated the left sACC.

Thus, the association of sACC activations with sadness reports was in contrast with its lack of statistical significance in FDR-corrected whole-brain Neutral-Sadness contrasts. Based on this result, we decided to define the network of areas putatively involved in sadness processing by lowering our statistical threshold (whole brain uncorrected tests at punc < 0.05) in a 2-factor ANOVA testing the interaction between the factors epoch (Neutral/Sadness) and sadness intensity (high-sadness/low-sadness). This analysis led us to identify 9 different bilateral cortical and subcortical areas implicated in emotional processing in this task (see Table 1).

Community structure distinguishes ventral and dorsal subnetworks

Based on these BOLD activations consistent with previous literature32,36,37,38,39,40, we defined a set of ROIs (Table 1) that would be presumably implicated in the regulation of cognitive and emotional task demands, conforming dorsal and ventral subnetworks, respectively, and we set to determine how sadness experience was associated with functional changes in network topology.

For each pair of ROIs we estimated their functional connectivity as the linear dependence of the temporal fluctuations in the corresponding signals, as measured by the Pearson correlation coefficient. This led us to define a symmetrical connectivity matrix containing the correlation coefficients between all possible pairs of ROIs. This matrix consists of positive and negative correlations (see Functional connectivity and Graph analysis). We obtained one such connectivity matrix independently for each subject, and we then averaged together these matrices to obtain a matrix of the averaged connectivities across participants. We applied graph-theoretic analyses by considering ROIs as nodes and the functional connectivity between each pair of ROIs as the corresponding edge.

We first asked if the pattern of connectivities defined subnetworks of areas that had distinct connectivity within and across subnetworks. This can be determined through a community detection algorithm that finds the assignment of nodes (ROIs) in communities (subnetworks) by maximizing the modularity Q of the partition (see Graph analysis). This community detection algorithm applied to our experimental grand-average connectivity matrix identified two main communities that coincide with the results of our BOLD contrast analyses above: the dorsal community (areas mSFG, PCG, IPS, dlPFC and iFG, Supplementary Fig. S4) that was related to working memory, and the ventral community (areas sACC, medial Frontal Pole (mFP), medial orbitofrontal gyrus (mOFG), Amygdala (Amy) and Hippocampus (Hip)) that was related to sadness (Fig. 3A,B, Supplementary Figs S2A and S5A,B). The pattern of correlations shows that these subnetworks interact internally mainly through positive correlations (Figs 3A,B and S5A,B, red lines) and between them mainly through negative correlations (Figs 3A,B and S5A,B, blue dashed lines), as seen in the distributions of correlations (Figs 3C and S5C). In addition, the algorithm identified two sub-communities within the ventral community (Fig. 3B and Supplementary Figs S2A and S5B, red line). These sub-communities corresponded to emotion-related areas in the cortex (sACC, mOFG and mFP) and the limbic system (Amy and Hip), respectively. Consistent with this substructure in the ventral community, the mean correlation within the ventral community was lower than within the dorsal community (Figs 3C and S5C; 2-way ANOVA, p = 0.044 and p = 0.0062, respectively). The modularity Q in Sadness-WM2 was higher than in Neutral-WM1 (Q = 0.278 vs. Q = 0.255, permutation test, p = 0.054; Supplementary Fig. S2B). This suggested that the ventral and dorsal communities get more segregated following an episode of intense sadness. However, when we applied the community detection algorithm to the average correlation matrices obtained separately for the high-sadness and the low-sadness groups in the paradigm Sadness-WM2, the modularity Q did not show significant differences. Neither the community assignment of the different ROIs (Supplementary Fig. S6), nor the modularity Q changed based on sadness intensity (Q = 0.27 for the high-sadness group vs. Q = 0.272 for the low-sadness group, p = 0.55). However, we noted that community assignment changed between the two groups of subjects, in that the low-sadness group unequivocally segregated subcortical areas from the ventral network while the high-sadness group integrated these groups of areas for some conditions of the modularity analysis parameters (Supplementary Fig. S6). Taken together, this suggests that sadness experience could induce the segregation of cognitive and emotional networks, and may promote the integration of ventral cortical areas and subcortical structures within the emotional network, but our study was unable to find large structural changes that would manifest at the global community network level. For this reason we sought in the following to investigate quantitative changes occurring at the network node level.

Hub identification and their modulation by strong emotional demands

The network-level properties studied above are quite stable in our protocol, but they hint at a modulation in the interaction between the networks according to the emotional or cognitive demands. We wondered if specific areas mediating these interactions would be more acutely affected and reveal these modulations more strongly. We investigated this by measuring, in each participant’s connectivity matrix, two network parameters for each node: the degree, and the participation coefficient (see Methods). The degree of a node in the network is the number of connections it has to other nodes, and the participation coefficient compares this number of connections to the number of connections within the node’s own subnetwork. Nodes with a high degree and a high participation coefficient are known as connector hubs, and they are candidates to mediate interaction between subnetworks25. For these measures we considered strong absolute correlations, above a threshold of the absolute maximal correlation (selected by maximizing degree variance across areas, see Methods), for each subject. We identified connector hub nodes as those ROIs with degree one standard deviation above the network’s mean degree, and with participation coefficient above the network’s mean participation coefficient, following the criteria of previous studies25.

Figure 3
figure 3

Dorsal and ventral communities for Sadness-Working memory. (A) 3D-graphical representation of the networks, the ROIs are located according to real-world coordinates. Mean significant correlations are plotted; positive correlations in red lines and negative correlations in blue dashed lines, shading brain for schematic purposes. (B) Matrix of the mean correlations across subjects. The analysis identified two main modules (Fig. S2A), the dorsal and ventral communities separated by the dashed black line. In the ventral community, two sub-communities were found (separated by the red line), corresponding to the emotional areas in the cortex and the limbic system (subcortical areas). (C) Correlation distributions for all subjects. The correlations between dorsal and ventral modules (inter-modules), plotted in blue were mainly negative (mean ± SEM = −0.15 ± 0.004). The correlations within the dorsal module (intra-dorsal, plotted in orange, mean ± SEM = 0.089 ± 0.005) and the correlations within the ventral module (intra- ventral, plotted in red, mean ± SEM = 0.048 ± 0.009) were both positively biased.

Full size image

During the Neutral-WM1 paradigm we identified 5 connector hub ROIs: IPSr, IPSl, dlPFCl, mFPl and mSFGl (Fig. 4A). On the other hand, during the Sadness-WM2 paradigm we identified 3 connector hubs: IPSr, IPSl and mFPr (Fig. 4B). These connector hubs were consistently identified independently of the threshold applied to the correlation matrix (Supplementary Fig. S7A,B). We noticed that the IPS and mFP were present in both task paradigms, while dlPFCl and mSFGl appeared only in the Neutral-WM1 paradigm. The drop of dlPFCl from the list of connector hubs was significant, in that its degree decreased significantly from Neutral-WM1 to Sadness-WM2 (3-way ANOVA with factors: task paradigm, sadness intensity, and subject, threshold 35%, main effect of paradigm p = 0.014, all other effects and interactions not significant, Fig. 4C), independently of the applied threshold (range 30–45%, Fig. 5A). The dlPFCl participation coefficient also showed a marginally-significant decrease (Fig. 4C, main effect of paradigm p = 0.08). However, the degree of the dlPFCl was not significantly modulated by the intensity of sadness (3-way ANOVA, threshold 35%, interaction between task paradigm and sadness intensity p = 0.55). Instead, we found that the degree of the mFPl was modulated by sadness intensity in a range of thresholds: it increased for the high-sadness group and it decreased for the low-sadness group relative to Neutral-WM1 (3-way ANOVA, threshold 35%, no main effect of task p = 0.399 or sadness intensity p = 0.147, significant interaction between task paradigm and sadness intensity, p < 0.05, N = 22, Fig. 5B; 4-way ANOVA, including training index groups by median split, did not show significant main effect or interactions involving the training group, p > 0.05, N = 20). These results identified dlPFCl as a connector hub in the dorsal subnetwork that reduced its coupling following sadness induction, and mFPl as a connector hub in the ventral subnetwork that increased its coupling specifically in those participants that experienced a stronger emotional state after sadness induction.

Figure 4
figure 4

Hub identification: The degree of dlPFCl is reduced after sadness. Degree (left panels) and participation coefficient (right panels) analysis applying a threshold of 35% to the correlation matrices of each subject. (A) During the Neutral-WM1 paradigm 5 regions were identified as hubs (green bars): IPSr, IPSl, dlPFCl, mFPl and mSFGl. Their participation coefficients were above the mean, so they were classified as connector hubs. (B) During the Sadness-WM2 paradigm 3 regions were identified as hubs (green bars): IPSr, IPSl, mFPr, which were classified as connector hubs. Note that the ordering of areas is the same as in (A). (C) The dlPFCl was the only hub that presented a significant decrease in the degree from Neutral-WM1 to Sadness-WM2 and a marginally-significant decrease in participation coefficient. Error bars mark SEM.

Full size image
Figure 5
figure 5

The dlPFCl was modulated across paradigms and mFPl was modulated by sadness intensity, robustly across thresholds of the correlation matrix. (A) The decrease in the degree for the dlPFCl during Sadness-WM2 was stable across thresholds. Black asterisks mark significant main effect of paradigm (3-way ANOVA, p < 0.05). (B) The mFPl was the only hub that was modulated by sadness intensity. It increased its degree in the high-sadness group and it decreased it in the low-sadness group. Black asterisks mark significant interaction between task paradigm and sadness intensity (3-way ANOVA, p < 0.05). We mark with a triangle when the corresponding region is classified as a hub among the rest of areas. Error bars mark SEM.

Full size image

Changes in functional connectivity underlie behavioral differences and hub modulations

We investigated the mechanism underlying the hub modulations described above by first testing if a change in the functional connectivity between the 2 connector hub nodes, dlPFCl and mFPl, could explain the modulations of their degree. We analyzed the change in the correlation between these 2 hub nodes for each participant, task paradigm and sadness-intensity groups. The functional connectivity between dlPFCl and mFPl did not present either a main effect of task epoch (Neutral-WM1 vs. Sadness-WM2, 3-way ANOVA, p = 0.48) or an interaction between high-sadness/low-sadness and Neutral-WM1/Sadness-WM2, (3-way ANOVA, p = 0.12). Direct interactions between the 2 hub nodes were thus not a mechanism supported in our data for the network modulation operated by the sadness state.

Then, we looked for other nodes that could mediate the modulation of the connector hubs. We analyzed the change in correlations (as a measure of functional connectivity) between the 2 connector hub nodes (mFPl and dlPFCl) and all other network areas for each participant, task paradigm and sadness-intensity groups. We thus tested a total of 19 pairwise correlations for each connector hub node, and we corrected our tests for the multiple comparison problem by controlling the false discovery rate (FDR) at a level α = 0.05.

The functional connectivity between dlPFCl and sACCl presented a significant interaction between high-sadness/low-sadness and Neutral-WM1/Sadness-WM2, (3-way ANOVA, p = 0.0006, p (FDR-corr) = 0.036; 4-way ANOVA, including training index groups by median split, did not show significant main effect or interactions involving the training group, p > 0.1, N = 20). The correlations between dlPFCl and sACCl became more negative after sadness provocation only in the high-sadness group (Fig. 6A, paired sample t-test p = 0.0001 for high-sadness and p = 0.49 for low-sadness). In other words, only the group reporting more intense sadness presented a stronger anticorrelation between sACCl and dlPFCl after sadness induction. This suggests that the interactions of dlPFCl with sACCl could be associated with the reduction in dlPFCl network degree following sadness induction (Figs 4C and 5A).

Figure 6
figure 6

Sadness intensity increased the sACCl-dlPFCl anticorrelation and the sACCl-mFPl correlation. (A) Subject-by-subject correlations for sACCl-dlPFCl and sACCl-mFP during Neutral-WM1 and Sadness-WM2. High-sadness subjects are plotted with red lines, low-sadness subjects with blue lines and the population averages are plotted in thick lines, respectively. Left, High-sadness subjects and not low-sadness subjects presented significant increased anticorrelations between sACCl and dlPFCl during Sadness-WM2 compared to Neutral-WM1. Right, sACCl and mFPl correlations presented a significant interaction between sadness intensity (High-sadness vs. low-sadness subjects) and paradigm (Neutral-WM1 vs. Sadness-WM2). (B) For the high-sadness group (right) but not for the low-sadness group (left), we found a significant anticorrelation between sACCl (Sadness-Neutral) and dlPFCl (WM2-WM1) BOLD activity: the stronger the BOLD activity in sACCl during sadness, the weaker the BOLD activity in dlPFCl during WM2.

Full size image

The anticorrelation between sACCl and dlPFCl was in addition related to the functional activation of sACCl (Fig. 2B,C) and deactivation of dorsal areas (Supplementary Fig. S4C) as we found a significant inter-individual anticorrelation between the contrast Sadness-Neutral in sACCl and the contrast delayWM2-delayWM1 in the left dlPFC (RPearson = −0.5197, p = 0.0132). Moreover, this anticorrelation was significantly stronger (permutation test, p = 0.042) in the high-sadness group (RPearson = −0.6639, p = 0.0186), than in the low-sadness group (RPearson = −0.1264, p = 0.7278) (Fig. 6B). None of these effects were observed when median-splitting by training index (high training group RPearson = −0.55, p = 0.1; low training group RPearson = −0.32, p = 0.37, N = 20).

We then wondered if the functional connectivity between mFPl and sACCl was also modulated by sadness. Indeed, the connector hub mFPl presented a significant interaction (high-sadness/low-sadness vs. Neutral-WM1/Sadness-WM2) in its correlation with sACCl (Fig. 6A, right, 3-way ANOVA, punc = 0.043; 4-way ANOVA, including training index groups by median split, did not show significant main effect or interactions involving the training group, p > 0.2, N = 20). Correlations of mFPl with all other areas did not reach significance (p > 0.05). The correlation between sACCl and mFPl showed a marginally significant increase after sadness provocation only in the high-sadness group (Fig. 6A, right, paired sample t-test p = 0.083 for high-sadness and p = 0.26 for low-sadness). The correlation between mFPl and sACCl points at the association of the mFPl-sACCl connection with the sadness-dependent modulation of mFPl network degree (see Fig. 5B).

In summary, we found sadness-related effects at three different levels: behavioral, in functional activity, and in network structure. Importantly, we evaluated the association with sadness by testing the interaction between behavioral paradigm and sadness intensity report, which emphasizes the role of sadness experience in all these modulations. At the behavioral level, the subjects that reported highest emotional scores diminished their performance in the working memory task after sadness provocation (Fig. 2A). At the level of functional brain activity, we found an overall decrease in activation in the dorsal areas (Supplementary Fig. S4C), an increase in sACCl activity (Fig. 2B,C) and an inter-individual anticorrelation between sACCl and dlPFCl activity (Fig. 6B). Finally, the graph analysis showed a stronger segregation between ventral and dorsal networks following a strong emotional experience, with the connectivity degree of the dorsal connector hub dlPFCl being down-regulated after sadness provocation (Fig. 5A) and that of the ventral connector hub mFPl being up-regulated by sadness intensity (Fig. 5B). Sadness intensity also modulated the functional connectivity of these connector hubs, via sACCl: it increased the correlation between sACCl and mFPl and the anticorrelation between sACCl and dlPFCl (Fig. 6A). We summarize these findings schematically in Fig. 7.

Figure 7
figure 7

Schematic summary: Sadness-related effects in functional activity and in network structure. The sACC was a key area: following sadness provocation and specifically for high-sadness participants it showed an increase in the negative functional connectivity with dlPFCl and an increase in the positive functional connectivity with mFPl. These functional connectivity modulations could underlie the modulations of dlPFC and mFP as hubs of their respective networks: The dlPFCl degree was decreased after sadness provocation and the mFPl degree was modulated by sadness intensity. Red (blue) denotes degree increases (degree decreases) and dashed (continuous) arrows denote negative (positive) functional connectivity.

Full size image

Discussion

We applied here graph-theoretic network analysis to identify functional changes in brain networks associated with states of sharp emotional and cognitive contrast. Based on functional activations in concatenated working memory and sadness induction tasks, we identified two distributed brain networks, one comprising dorsal regions putatively associated with cognitive operation (dlPFC, IPS, iFG, mSFG, PCG) and another one comprising ventral regions linked to sadness processing (mFP, sACC, mOFG, Amy, Hip), which could be further decomposed in cortical and subcortical partitions50. The modularity of these brain networks increased following sadness induction, consistent with the hypothesis that a sadness state increases the competition between emotional and cognitive subnetworks19,51,52. Two areas emerged from our analysis as task-related connector hubs in the dorsal and ventral brain subnetworks based on their modulation by the sadness induction protocol: the dlPFCl in the dorsal community and the mFPl in the ventral subnetwork. Our data did not support direct interactions between these two hubs but instead coordination via an interposed area, the sACC, as the mediator of sadness-induced modulations in network structure.

In the dorsal network, the connector hub dlPFCl presented a significant decrease in connection degree following sadness induction (Figs 4C and 5A), suggesting that the sadness state reduced the effective coupling of this area and thus its ability to influence brain processing. Recent evidence has shown that the fronto-parietal brain network, which underlies cognitive control53,54, has especially high global connectivity (i.e., average connectivity with the rest of brain regions55), and the global connectivity of the left dlPFC was specifically identified as the mechanism by which the fronto-parietal network might control other networks56. Moreover, previous work also attributed a top-down control role to dlPFC in spatial working memory based on neuroimaging data and computational models52,57. Integrating previous literature and our results, the decreases in the degree of the dlPFCl could be related with the overall decrease of the BOLD activity in the dorsal network after sadness experience (Supplementary Fig. S4C) and the decline in the participants’ WM performance (Fig. 2A) based on the dlPFC diminished capability of exerting cognitive control during the WM task.

Several studies have found attenuated spatial WM performance during negative task-irrelevant affect8,9,10,11,14,58, associated with marked decreases in dlPFC activity8,9,14,15. However, these studies were performed using intervening task-irrelevant aversive stimuli and are thus subject to possible confounds due to attention capture by the noxious stimulus. To address this issue we designed here a paradigm with separate episodes of strong conflicting emotional and cognitive demands (Fig. 1), respectively. In these tasks the outcomes do not depend on the integration of emotional and cognitive aspects. We used a sadness provocation task32 to induce a sadness state, followed by a spatial WM task30 (Sadness-WM2 paradigm), and we compared with a control paradigm that concatenates a neutral epoch and spatial WM task (Neutral-WM1 paradigm) (Fig. 1). In the Sadness-WM2 paradigm, the cognitive modulations mediated by emotional demands were provoked by an emotional state elicited before the cognitive task. Therefore, unlike previous studies our results do not depend on external distractors or emotional stimuli during the WM task. To our knowledge, only one study before has used a similar strategy on medicated depressed patients36. They found WM performance deficits following sadness induction in both controls and depressed patients. This is in line with our finding of impaired WM performance in subjects with higher sadness scores, and supports the role of emotional states in conditioning cognitive function, without any confounds of possible acute attentional shifts by intervening cues as in previous studies.

Intensity and duration are two central characteristics of an emotional response59. In our task, participants provided a subjective rating (on a scale 0–7) of the sadness intensity reached after the scanning session. We confirmed (Fig. 2) that this report was indeed measuring effective sadness intensity by validating its correlation with cognitive performance35 and with the activation of sACC32. Negative emotional traits in healthy people are known also to be associated with increased sACC activation following sadness induction60, so our sadness report may be also associated with personality traits of the participants. We then used this report in all our analyses to confirm the unambiguous association of sadness with differences between our two behavioral paradigms, and thus overcome two confounds in our paradigm. For one, in our paradigm we contrasted the induction of an emotional memory with a non-emotional resting epoch without any biographical memory component. We reasoned that this would emphasize the competition between cognitive and emotional processing, while also simulating the rumination associated with depression, but it did incorporate a memory confound that had to be addressed. Secondly, we did not reverse the order of the paradigms, because it has been shown that sadness is the emotion with the longest persistence34 and a previous study argued that the sadness block may generate some residual effect in subsequent control blocks35. This could pose interpretation problems associated with the sequence of tasks (practice, tiredness). We addressed both of these confounds in our analyses by testing the relation of our effects with the intensity of the sadness reported by participants. Specifically, we tested the statistical significance of an interaction between the factors paradigm and sadness intensity in our analyses of variance (ANOVA) tests. Most of the changes in network structure and functional connectivity reported in this study are supported by such a significant ANOVA interaction, thus supporting their unambiguous association with a change in emotional state. In addition, we validated that no such effects appeared when modeling our network measures with a training index that measured changes in reaction times as a proxy of training effects in individual participants. All these checks confer plausibility to our interpretation in terms of sadness-related effects, but it remains a limitation of our study of a fundamental origin given the persistence of sadness experience34.

The emotional connector hub was identified as area mFPl based on its modulations by emotional demands. Sadness experience increased the degree of the mFPl in the high-sadness group relative to the low-sadness group (Fig. 5B), suggesting that intense sadness increases the influence of mFPl on other brain areas. The mFP (part of medial prefrontal cortex) has been described as part of the default mode network, which drives the self-reference processes61,62,63,64,65. The modulation in the mFPl degree by sadness intensity could be related with more intense self-reference processes in participants of the high-sadness group.

Modulations of the mFPl and dlPFCl degree in our study are in line with the flexible hub theory recently presented66 and they suggest that these hubs are capable of functional connectivity adaptations in order to balance cognitive and emotional demands. We found that these adaptations occur coordinated through the sACCl, as it showed more negative functional connectivity with dlPFCl and more positive functional connectivity with mFPl following sadness provocation, and specifically for high-sadness participants (Figs 6 and 7). Remarkably, we also found an inter-individual negative correlation between sACCl and dlPFCl BOLD activity, which was higher in the high-sadness group (Fig. 6B). Such result is similar to the inter-individual negative correlation found between the amygdala and inferior frontal gyrus during a working memory task with a negative task-irrelevant stimulus presented during the delay67. We note here that by focusing on a predefined set of regions of interest, our study cannot exclude the possibility of other brain areas also playing a coordination role between these two large subnetworks.

There has been substantial debate surrounding the appropriate interpretation of negative correlations observed with resting state functional connectivity when including a preprocessing step termed global signal regression41,42,43,44,68,69. This data processing step can improve the specificity of resting state correlations and the correspondence with anatomy41 and electrophysiology42, but there are mathematical concerns that anticorrelations could emerge as an analysis artifact44. In order to test if global signal regression generated artifactual correlation patterns in our data, we repeated our analyses without global signal regression preprocessing and we found that this did not affect the relative relationships between functional connectivity in our conditions of interest (Supplementary Fig. S8). However, the results with data preprocessed with global signal regression fit together more consistently and provided an easier interpretation. In particular, note that without global signal regression the correlation between sACC and dlPFC became practically zero after sadness induction (Supplementary Fig. S8A), which would be interpreted as sACC and dlPFC becoming decoupled. This decoupling does not fit with other results, in particular with the strong inter-individual correlation that we found between BOLD activity in sACC and dlPFC (see also20), especially in the high-sadness group (Fig. 6B). Also, a decoupling effect of sadness is at odds with the observed correlation between focal PFC stimulation (TMS) treatment outcome and dlPFC-sACC anticorrelation strength in depressed patients70. Because the results are qualitatively unchanged by global signal regression, but provide a much more direct interpretation, we favor here this preprocessing step.

Thus, in our analysis the sACC was not identified as a connector hub area but it did emerge as a key region that coordinates dorsal and ventral connector hub areas, and is thus capable of influencing the global functional network structure. These results provide a new perspective on the previously reported implication of sACC in sadness and depression. Previous studies consistently associate sACC with acute sadness, major depression and antidepressant treatment effects, suggesting a critical role for this region in modulating negative mood states20,32,71. In addition, sACC connections to the brainstem, hypothalamus, and insula have been implicated in the disturbances of circadian regulation associated with depression and it has been described as a visceral-motor region72,73,74. Reciprocal pathways linking sACC to orbitofrontal, medial prefrontal and various parts of the anterior and posterior cingulate cortices constitute the neuroanatomical substrates by which primary autonomic and homeostatic processes influence various aspects of learning, memory, motivation and reward72,75,76,77. In depressed patients, the resting-state sACC functional connectivity with the default mode network (DMN) was found stronger than in control participants, and it further correlated with the length of the patients’ depressive episodes78. All these data reinforce the idea that sACC is implicated in sadness regulation and our results indicate that this could be by means of its regulatory role in relation to two hub network areas, rather than a direct driving mechanism. In our study, participants that reported strong sadness experience had brain activity patterns similar to those previously reported for depressed patients: an activation of the sACC19,38, a deactivation of dlPFC38,79,80 and an increase in the anticorrelation between sACC and dlPFC20,70. This underscores the idea that the network dynamics underlying negative emotional state in healthy subjects could be, when pathologically exacerbated, responsible for behavioral symptoms in depressed patients19,20. Indeed, the three areas that we have identified have been repeatedly associated with antidepressant treatments: response to selective serotonin reuptake inhibitors is related to sACC activity81,82,83, TMS of dlPFC is most effective in sites strongly anticorrelated with sACC70, response to TMS of dorso-medial PFC depends on the connectivity of the mFP84, response to cognitive behavioral therapy is related to changes in the three areas83, and recent results of DBS (subgenual white matter stimulation) link its therapeutic effect to fibers reaching the mFP22,23.

The fact that some of these interventions are very focal and yet address symptoms supported by the activity of system-wide networks20, could be explained by our findings of a modular structure in these networks, coordinated by one area (sACC) regulating two connector hubs (dlPFC and mFP). Conversely, also a very focal dysfunction in the network, in one of these coordinating areas, could have profound impact in functions subserved by distant circuits, as recently shown in a computational network model of sACC and dlPFC51,52. Depression as a circuit-level manifestation of a sharply localized disease remains an enticing hypothesis, supported by distributed sadness circuits that interact via very specific coordinating cortical areas.