Graph analysis of guilt processing network highlights links with subclinical anxiety and self-blame

Abstract

Maladaptive forms of guilt, such as excessive self-blame, are common characteristics of anxiety and depressive disorders. The underlying network consists of multiple associative areas, including the superior anterior temporal lobe (sATL), underlying the conceptual representations of social meaning, and fronto-subcortical areas involved in the affective dimension of guilt. Nevertheless, despite understanding the circuitry’s anatomy, network-level changes related to subclinical anxiety and self-blaming behaviour have not been depicted. To fill this gap, we used graph theory analyses on a resting-state functional and diffusion-weighted magnetic resonance imaging dataset of 78 healthy adults (20 females, 20–35 years old). Within the guilt network, we found increased functional contributions of the left sATL for individuals with higher self-blaming, while functional isolation of the left pars opercularis and insula was related to higher trait anxiety. Trait anxiety was also linked to the structural network’s mean clustering coefficient, with the circuitry’s architecture favouring increased local information processing in individuals with increased anxiety levels, however, only when a highly specific subset of connections was considered. Previous research suggests that aberrant interactions between conceptual (sATL) and affective (fronto-limbic) regions underlie maladaptive guilt, and the current results align and expand on this theory by detailing network changes associated with self-blame and trait anxiety.

Introduction

During our lives, we find ourselves navigating through the complex world of interpersonal interactions on a daily basis. Accurate comprehension of the actions and messages conveyed by others, as well as the expectations and the norms of acceptable behaviour on our side, constitute the key building blocks for everyday functioning. As a result of these interactions, our perceptions of them, and the associated mental events, we experience a variety of social feelings. Their complexity is reflected in the diversity of circumstances that can evoke similar affective experiences. For example, we often face feelings of guilt when we believe we have transgressed the rules of acceptable behaviour or have behaved immorally (interpersonal guilt; Monteith 1993, Zahn et al. 2020). Nevertheless, the same feelings are also experienced by individuals surviving a near-death or traumatic event when their loved ones perished (survivor’s guilt), or those who believe they are not meaningfully contributing to society (existential guilt). On the one hand, the lack of guilt is associated with psychopathic traits and results in antisocial and exploitative behaviours (Waller et al. 2020). Simultaneously, the prevalence of maladaptive forms of guilt, such as excessive self-blaming, leads to mental health conditions associated with difficulties in social functioning, such as depression and anxiety (Kim et al. 2011; Cândea and Szentagotai-Tătar 2018), respectively, through the increased feelings of worthlessness and worrying not to repeat the transgressive actions in the future (Schoenleber et al. 2014, Harrison et al. 2022). Adequate functioning of this system is thus essential for healthy social behaviour. With the rising number of individuals diagnosed with mood and anxiety disorders (Yuan et al. 2022), it has become increasingly important to comprehend the processes that contribute to their onset. Given that subclinical populations may exhibit symptomatology similar to that of diagnosed individuals (Besteher et al. 2017), investigating the neural correlates of guilt in these populations can offer invaluable insights into the general functioning of this circuitry. Additionally, it may enhance our understanding of how these symptoms emerge and manifest in clinical populations.

Owing to the recent meta-analytic and review efforts (Bastin et al. 2016, Gifuni et al. 2017, Eslinger et al. 2021), it has been established which brain regions participate in processing the feelings of guilt. The anatomy of the circuitry reflects the complex character of guilt feelings, as it includes frontal, parietal, and temporal areas known to be engaged in various guilt-related processes, such as mentalization, self-reference, semantic cognition, morality, disgust, and conflict monitoring (Gifuni et al. 2017). Nevertheless, the previous works have pinpointed the importance of the particular parts of the network. For example, frontopolar cortex activity appears to be the most reproducible for guilt contrasted with other similarly unpleasant and complex feelings, such as indignation towards others (Eslinger et al. 2021). Furthermore, a line of work underlines the role of the subgenual cingulate cortex and adjacent septal region (SCSR), with its functioning linked to individual differences in guilt proneness and empathic concerns (Moll et al. 2011, Eslinger et al. 2021). Last but not least, special attention is drawn to the right superior anterior temporal lobe (sATL), given that the connectivity between this area and SCSR during guilt processing is altered in major depressive disorder (MDD) patients and has the ability to predict the illness recurrence (Green et al. 2012, Lythe et al. 2015). With sATL suggested to be involved in accessing the conceptual representations of social meaning, and the activation of fronto-subcortical areas reflecting the affective dimension of guilt, it has been proposed that aberrant interactions between semantic and emotional networks might underlie the prevalence of maladaptive guilt across mood and anxiety disorders (Green et al. 2012, Lythe et al. 2015, González-García and Visser 2023).

Nevertheless, despite the general understanding of the anatomy of the guilt processing circuitry and the functions the individual regions may play in the phenomenon, currently, it is unknown whether and how their network-level interactions are in fact associated with self-blaming behaviour. The works reported earlier have concentrated on task-related measures of local activity and connectivity between two specific regions rather than looking at the guilt processing areas as a functional and structural brain network. Describing the physiology and anatomy of the brain utilizing approaches such as graph theory is especially important, given that no single area functions in isolation. The more complex characteristics, apparent only at the network level, might be otherwise easily overlooked in traditional mass-univariate analyses. To fill this crucial gap, we used a resting-state functional magnetic resonance imaging (rs-fMRI) and diffusion-weighted MRI (dMRI) dataset to construct the functional and structural guilt processing networks. Subsequently, we investigated the associations of self-blaming emotion regulation strategy with their global and local graph theory parameters. Blaming oneself emerges as a result of multiple co-occurring mental processes, including self-awareness, theory of mind, and conflict monitoring (Gifuni et al. 2017). Therefore, it is plausible that variability in the functional or anatomical connections of brain regions governing any of these functions could shift the delicate balance in this complex psychological machinery and contribute to increased self-blaming tendency. Alternatively, such an observation could also be related to the characteristics of the network as a whole. Indeed, individual differences in other emotion regulation strategies have been associated with both nodal and global graph theory parameters (Pan et al. 2018, Jacob et al. 2019). Thus, we hypothesized to observe a similar pattern for the self-blaming behaviour.

Furthermore, despite the importance of maladaptive guilt in anxiety disorders (Cândea and Szentagotai-Tătar 2018), no study so far has examined how the neural correlates of guilt processing vary in regard to individual anxiety levels. To explore this uncharted area of research, we set out to investigate how the global and local graph theory parameters of the functional and structural guilt processing network would be associated with trait anxiety in a subclinical population. Anxious people have been reported to have differential structural and functional connectivity both within and across the canonical brain networks (Xu et al. 2019, Yang et al. 2020). The areas implicated in guilt processing map onto several of these networks, making it difficult to predict the combined impact of such distinctions on both the local and global graph parameters. Previous whole-brain and network-specific studies reported associations between trait anxiety and both types of network parameters (Tao et al. 2015, Zhu et al. 2017, Makovac et al. 2018, Guo et al. 2021). Taking into account the prevalence of excessive guilt in anxious individuals, we hypothesized to observe a similar pattern in our study.

Therefore, the objectives of the current study are to define the guilt processing network using the available literature and use graph theory to examine the characteristics of this circuitry that are associated with self-blaming tendency and subclinical anxiety. Based on previous research, we expected the guilt processing network, and in specific the sATL and medial frontal regions, to interact differently in individuals with high self-blame and anxious personality traits.

Materials and methods

Dataset

The project was performed using the fully anonymized MRI and behavioural data obtained from the Max Planck Institute Mind-Brain-Body Dataset (Babayan et al. 2019). After applying the initial inclusion criteria, i.e. right-handedness, secondary or higher level of education, no current medication use, and no history of neurological, psychiatric, and substance use disorders, 95 out of 153 available young participants (aged 20–35 years) were found eligible for the analyses. Data from 17 subjects were discarded due to excessive head motion, making the final sample of the fMRI analysis equal to 78 (20 females and 58 males). To achieve the correspondence between the functional and structural neuroimaging indices, we analysed the dMRI data only from the previously mentioned 78 individuals. Two subjects (one female and one men, both aged 20–25 years) were found ineligible due to the lack of fieldmap data and a visible scanning artefact, making 76 the final sample of the dMRI investigation.

Trait anxiety was assessed using the State-Trait Anxiety Inventory (STAI; Spielberger 1989). The scale is made out of 20 Likert-type items scored from 1 (almost never) to 4 (nearly always), making the outcome range from 20 to 80. Additionally, a four-item subscale of the Cognitive Emotion Regulation Questionnaire (CERQ) (Garnefski et al. 2001) was used to measure generalized self-blaming tendencies after having experienced a negative event. Each of these items was scored on a 5-point Likert scale from 0 (almost never) to 4 (almost always), making the possible range of scores from 0 to 16. In the case of both tools, the higher the individual score, the more pronounced the psychological feature. STAI and CERQ scores were moderately correlated in our sample (ρ = 0.403; P < .001). The full characteristics of the cohort are provided in Table 1.

Table 1.

Demographic summary of the sample.

Variable	Mean/median (s.d.)	Range
Age (years)	20–25, N = 42 25–30, N = 31 30–35, N = 5	20–35
STAI	37.22 (8.04)	21–56
CERQ self-blaming subscale	4 (n/a)a	0–11

Abbreviations: s.d., standard deviation; n/a, nonapplicable.

^aIt indicates non-normal distribution of the data and the sole use of median and range.

MRI data acquisition

Three types of neuroimages acquired with a 3 Tesla scanner (MAGNETOM Verio, Siemens, Healthcare GmbH, Erlangen, Germany) equipped with a 32-channel head coil were used in the course of the investigation. High-resolution anatomical data were collected with T1 MP2RAGE sequence (176 sagittal slices; 1-mm isotropic voxel size; repetition time (TR) = 5000 ms; echo time (TE) = 2.92 ms; flip angle 1/flip angle 2 = 4°/5°; GRAPPA acceleration factor 3). The T2*-weighted resting-state fMRI data were acquired with a gradient echo-planar multiband imaging sequence (64 axial slices acquired in interleaved order; 2.3-mm isotropic voxel size; TR = 1400 ms; TE = 30 ms; flip angle = 69°; multiband acceleration factor = 4). During the resting-state scan, the participants were instructed to lie awake with their eyes concentrated on a low-contrast fixation cross. A total of 657 volumes were collected for each individual, resulting in a scan duration of 15 min and 30 s. Additionally, the high angular resolution diffusion-weighted images of the whole brain were collected axially (88 slices in interleaved order) with a 1.7-mm isotropic resolution (60 diffusion-encoding gradient directions, b-value of 1000 s/mm², 7 b₀ images, TR = 7000 ms, TE = 80 ms, FA = 90°, GRAPPA acceleration factor = 2, 32 reference lines, multiband acceleration factor 2).

MRI data preprocessing

The preprocessing of T1-weighted data was performed in the Computational Anatomy Toolbox (CAT12; Gaser, et al., 2024). Individual anatomical images were skullstripped and segmented into grey matter (GM), white matter (WM), and cerebrospinal fluid (CSF). The skullstripped T1-weighted data were used later for the delineation of regions of interest (ROIs) in each participant and coregistration with the other MRI modalities.

The preprocessing stream of the functional data consisted of steps available in the FMRIB Software Library (FSL) (susceptibility distortion correction; Jenkinson et al. 2012), R (motion scrubbing using the fMRIscrub package; Power et al. 2012), and Analysis of Functional NeuroImages (AFNI) (all other steps; Cox 1996). The first five volumes of the fMRI time series were discarded to allow for signal equilibration. Then the data underwent despiking, and slice-timing, motion, and susceptibility distortion corrections. At this point, the functional neuroimages were coregistered with the skullstripped anatomical data. Subsequently, the fMRI data were prewhitened, detrended, denoised, and band-pass filtered in the 0.009–0.08-Hz range. The regressors in the denoising matrix included six motion parameters and their derivatives, as well as WM and CSF signals. Additionally, volumes with extensive motion (i.e. framewise displacement >0.25 mm) were censored. Participants with a mean framewise displacement of >0.2 mm or >30% of the censored volumes were excluded from the further analysis. The minimal length of the noncensored time series was thus 10 min and 51 s, which is sufficient for a reliable estimation of the functional connectivity (Van Dijk et al. 2010). In order to ensure the spatial independence of the signal in all the nodes of the network, no smoothing was applied.

The dMRI data were preprocessed using MRtrix3 software (Tournier et al. 2019). First, the images underwent denoising and unringing. These steps were followed by distortion, eddy currents, and motion correction using the dwifslpreproc wrap-up based on the dedicated functionalities available in FSL (Jenkinson et al. 2012). Subsequently, the diffusion images were bias field-corrected. The response functions for WM, GM, and CSF were estimated with the use of the dhollander algorithm (Dhollander et al. 2019). They served as an input for the single shell 3 tissue-constrained spherical deconvolution performed using MRtrix3Tissue (https://3Tissue.github.io), a fork of MRtrix3 (Tournier et al. 2019). Following intensity normalization, anatomically constrained tractography (ACT; Smith et al. 2012) was performed based on the coregistered 5 tissue image generated by the ‘5ttgen fsl’ function. Compared to the standard fibre tracking algorithms, ACT is a much more biologically reliable method. The connections are seeded from the GM–WM interface and terminated upon either entering cortical GM or within subcortical GM, which is in agreement with the characteristics of neuronal projections derived from histological studies. In each participant, 10 million streamlines were generated using the iFOD2 (improved second-order integration over fibre orientation distributions) algorithm. The default settings were applied with the exception of the maximal track length reduced to 100 mm (compared to the default 170 mm, given the data resolution) and the cut-off value set to 0.06. As the last step, the generated streamlines were refined with the SIFT2 algorithm (Smith et al. 2015) to allow for the use of the number of streamlines as a valid index of the structural connection density.

Network construction

The regions (nodes) of the guilt processing circuitry were selected based on a recent meta-analysis of studies investigating neural correlates of guilt (Gifuni et al. 2017) and a review of the functional neuroanatomy of moral social feelings (Eslinger et al. 2021). We identified 19 ROIs mapping onto the associative areas of several canonical brain networks (Uddin et al. 2019). Such heterogeneity reflects the complex character of guilt experience, which requires the tight coordination of multiple cognitive processes often governed by distinct circuitry. This characteristic is further supported by the existence of widespread functional and structural connectivity within the established guilt processing network (Supplementary Figs S1 and S2).

The ROIs were primarily planned to be centred on the reported peak coordinates (Gifuni et al. 2017, Eslinger et al. 2021). Note that for the peak coordinates of the Eslinger et al. (2021) review, we reported the values mentioned in the original studies (Zahn et al. 2009, Moll et al. 2006; Green et al. 2010). Given the fact that the peaks of five areas mapped onto WM and in the case of the fMRI analysis the majority of the signal of interest is believed to originate in GM, their coordinates were slightly adjusted to ensure the proper GM coverage. ROIs were primarily delineated in the Montreal Neurological Institute (MNI) space as 3.45-mm radius spheres (1.5 times the fMRI data voxel size; see Table 2 for their mass centre coordinates). To refine the functional ROIs to include only GM voxels, we (I) averaged the CAT12-derived normalized GM segments across all the participants; (II) excluded voxels with values lower than the threshold used in the volumetric anatomical analyses (i.e. 0.2); and (III) binarized the resulting image and used it to as an inclusive mask for the previously defined ROIs.

Table 2.

The nodes of the guilt processing network.

Location	MNI coordinates	References
R medial orbitofrontal cortexa	11, 51, −3	Eslinger et al. (2021), Zahn et al. (2009)
R subgenual cingulate cortex R septal area	4, 15, −5	Eslinger et al. (2021), Moll et al. (2006)
L subgenual cingulate cortex L septal area	−4, 15, −5	Eslinger et al. (2021), Moll et al. (2006)
L anterior cingulate cortex	−4, 28, 24	Gifuni et al. (2017)
L superior frontal gyrus	−4, 56, 22	Gifuni et al. (2017)
L medial frontal gyrus (BA 6)	−4, −6, 48	Gifuni et al. (2017)
L medial frontal gyrus (BA 10)a	−12, 46, 8	Gifuni et al. (2017)
L medial frontal gyrus (BA 9)	−4, 42, 40	Gifuni et al. (2017)
L insulaa L IFGpoa	−46, 8, 5	Gifuni et al. (2017)
R IFGpo	46, 12, 6	Gifuni et al. (2017)
L superior temporal gyrus	−46, −60, 20	Gifuni et al. (2017)
L middle temporal gyrus	−62, −12, −10	Gifuni et al. (2017)
L sATL	−50, 4, −10	Gifuni et al. (2017)
R sATL	57, −3, −6	Eslinger et al. (2021), Green et al. (2010)
L posterior cingulate cortexa,b L parieto-occipital sulcusa,b	−15, −56, 7	Gifuni et al. (2017)
L precuneus	−4, −54, 38	Gifuni et al. (2017)
L supramarginal gyrusa	−50, −38, 27	Gifuni et al. (2017)
L cuneus	−4, −92, 10	Gifuni et al. (2017)
R lingual gyrus	4, −78, 2	Gifuni et al. (2017)

Abbreviations: L, left; R, right; BA, Brodmann area.

^aThe area originally mapped onto WM and thus had its location slightly adjusted.

^bThe area was originally reported as L parahippocampal gyrus; nevertheless, according to the Brainnetome atlas (Fan et al. 2016), it was more closely located to the currently disclosed regions.

Delineation of ROIs in the individual brains was achieved by warping them from the MNI space. The spatial correspondence of the network nodes with the fMRI and dMRI data was achieved using the matrices created during the linear coregistration of T1 images with the respective images of other modalities.

In the case of the functional network, the nodes’ time series were extracted and the functional connectivity matrix was calculated using the Pearson correlation. Negative correlations were removed from the resulting matrices to ensure easier interpretability of the topological measures (Qian et al. 2018). To normalize the distribution of the correlation values, Fisher’s z transformation was applied. Such values served as the edges of the functional network. In the case of distance-based graph measures, the obtained z-scores were inverted (1/z) to ensure their proper calculation (Rubinov and Sporns 2010).

For the structural connectivity data, the ROIs were dilated by 1 voxel to ensure that they covered the border of GM and WM, given that in ACT this interface serves as the seed for tractography (Smith et al. 2012). In each subject, the resulting nodes were inspected visually and, if required, corrected manually to ensure that they were restricted to the areas of interest. The structural connectivity matrix was created by summing up the SIFT2-generated (Smith et al. 2015) weights of all the streamlines connecting every pair of nodes. The edges of the structural network were defined in the same manner as for the functional data.

Weighted, undirected graphs were created using the matrices derived from the fMRI and dMRI data in R (version 4.3.0) using the igraph package (Csardi and Nepusz 2006). As the graph measures are sensitive to network density (Yeh et al. 2021), they were estimated for sparsities ranging from 1% to 30% with 1% steps. The decision to use the aforementioned sparsity range was guided by the fact that numerous graph metrics reached ceiling effects with higher network densities. To better capture the associations with the graph metrics across the entire sparsity range, we calculated the area under the curve and treated it as the primary outcome of the study. The sparsity-level findings were regarded as secondary.

Graph theory measures and statistical analysis

The same graph parameters were calculated for both the functional and structural networks. On the nodal level, the weighted clustering coefficient and betweenness centrality were extracted. As for the global network parameters, modularity (Leiden algorithm; Traag et al. 2019), mean clustering coefficient, and global efficiency were calculated. The definitions of these measures are provided in Supplementary Table S1 (Freeman 1979, Latora and Marchiori 2001, Barrat et al. 2004, Traag et al. 2019).

The associations between the graph indices and psychometric measures were tested with the permutation-based linear regressions (5000 permutations) using the permuco package (Frossard and Renaud 2021) in R (version 4.3.0). The models treated the graph metrics as the dependent variables and were run separately for STAI and CERQ scores, always including gender and age as covariates. The nodal results were Bonferroni-corrected for the number of regions [family-wise error rate (FWE)] at the level of P_FWE < .05. As only one value of each global parameter was calculated per network, the original significance values are reported.

Apart from testing for the associations of the graph measures with the psychometric scores, the betweenness centrality of the nodes was used to examine which areas functioned as hubs of the guilt processing network. These results were, however, treated as secondary and they were thus included in the Supplementary material (Supplementary Table S2, Supplementary Figs S3 and S4).

Results

Functional network

For the functional network, significant associations were observed for two nodes: the left sATL (Fig. 1) and the left inferior frontal gyrus pars opercularis (IFGpo) and insula (Fig. 2). The clustering coefficient of the left sATL was positively associated with the strength of the self-blaming coping strategy (t = 2.92; P_FWE = .049), while for the left IFGpo and insula this metric was negatively related to the trait anxiety (t = −3.03; P_FWE = .027). The clustering coefficient of the left sATL was also nominally associated with trait anxiety; however, this result did not survive multiple comparison correction (t = 2.26; P_FWE = .513). Neither area was found to function as a hub of the functional network (Supplementary Table S2, Supplementary Fig. S3). Associations between the global graph properties and psychometrics were not found.

Figure 1.

The association between the clustering coefficient of the left sATL (L sATL) in the functional network and self-blaming tendency as measured with the CERQ (Garnefski et al. 2001). Abbreviation: AUC, area under the curve.

Figure 2.

The association between trait anxiety and the clustering coefficient of the left IFGpo (L IFGpo) and adjacent insula in the functional network. Abbreviation: AUC, area under the curve.

Structural network

In the structural analyses, there were no area-under-the-curve (i.e. primary) results reaching the significance threshold. However, the secondary sparsity-specific analyses revealed positive links between the mean clustering coefficient of the network and trait anxiety for the 14% (t = 2.53; P_FWE = .013), 15% (t = 3.12; P_FWE = .002), 16% (t = 2.03; P_FWE = .045), and 17% (t = 1.94; P_FWE = .049) sparsities (Fig. 3). Interestingly, at the discussed densities, the circuitry appeared organized into a semantic submodule localized in the left temporal lobe, and an affective subcircuit consisting mainly of bilateral subcortical and cortical midline nodes (Supplementary Fig. S6). Aligning with the methods used on the functional network, the influence of head motion was controlled. Maximal displacement across all the diffusion-weighted volumes was calculated for each participant and correlated with the highlighted significant results. The lack of such associations (individual P_uncorrected > .05) confirms their validity.

Figure 3.

The associations between trait anxiety and the mean clustering coefficient of the structural network for the 14%–17% sparsity range.

Discussion

The functional network analyses revealed that across the entire spectrum of densities of the guilt processing circuitry, self-blaming behaviour is more pronounced in individuals with a more strongly interconnected left sATL. In turn, higher levels of trait anxiety are associated with increased functional isolation of the left IFGpo and insula. Moreover, at a limited range of sparsities, anxiety is positively linked to the structural network’s mean clustering coefficient, indicative of a neural architecture favouring increased local information processing, however, only when a highly specific subset of anatomical connections is considered. These results thus partially confirm the hypotheses, linking self-blaming emotion regulation strategy and trait anxiety to nodal parameters of the functional network, with the latter phenomenon additionally related to the global characteristics of the structural network.

As for the functional graph, the lack of associations with global parameters suggests that the behavioural differences might be related to alterations in the particular cognitive processes governed by specific brain areas rather than differential functioning of the entire network. This notion is supported by the fact that the regions identified here form part of the network involved in semantic and social cognition, with the latter process drawing on semantics-related computations (Zahn et al. 2007; Lambon Ralph et al. 2017, Jackson 2021, Diveica et al. 2021). As such, it is not surprising that the very same areas have been implicated in cognitive emotion regulation (Brandl et al. 2019), mapping onto a neural network responsible for reinterpretation/reappraisal and distancing/perspective-taking strategies (Morawetz et al. 2017). Taking into account the above evidence, our study suggests that alterations in the functional properties of the regions belonging to this particular circuitry may serve as the biological underpinning linking self-blaming behaviour and anxiety. The subsequent parts of the discussion will now go into details regarding the exact roles attributed to each area.

The left sATL, together with its contralateral homologue, is believed to be the brain’s locus for social semantic knowledge (Zahn et al. 2007, Olson et al. 2013, Rice et al. 2015, Binney et al. 2016). The region’s prominent role in self-blame has been pinpointed by studies showing the aberrant connectivity of this area with SCSR during guilt processing in MDD patients (Green et al. 2012, Lythe et al. 2015). By reporting sATL’s associations with self-blaming behaviour, our study underlines the key role it plays in the studied circuitry, implicating that the differential functional interactions of this region with the fronto-limbic emotional networks could underlie the prevalence of maladaptive forms of guilt observed across both mood and anxiety disorders (Kim et al. 2011, Green et al. 2012, Cândea and Szentagotai-Tătar 2018, González-García and Visser 2023). The stronger interconnected left sATL in individuals with increased self-blame might reflect the tendency to ruminate and worry over related negative social situations, which would require the sATL to access the related social-semantic information. The ATL has been described as a hub region that interconnects modality-specific regions to obtain semantic (including social) information. The more frequently semantic items are processed, the stronger the connections become, meaning that less activity is required in the network to obtain the information (Lambon Ralph et al. 2016; Rogers et al. 2004, Rogers and McClelland 2004). As such, our findings align with the idea that the sATL ‘communicates’ more effectively with other semantic-related regions in individuals with higher self-blame. The currently held view states that the sATL transmits the social information via the uncinate fasciculus to the medial prefrontal cortex, which in turn integrates it with emotional processes to regulate social behaviour (Sakata et al. 2019, Arioli et al. 2021). In the case of our study, the left sATL node was structurally connected to the ipsilateral SCSR in only 38.15% of the participants (Supplementary Fig. S2). Nevertheless, we believed this finding to be related to the relatively small size of the node. Indeed, when larger ROIs were used, resembling the more widespread task fMRI activations, the fibres belonging to the tract were successfully identified in 85.52% of the subjects (Supplementary Fig. S5), confirming that the sATL node indeed belongs to this functionally relevant circuitry.

The left IFGpo and adjacent parts of insula are key elements of the brain system responsible for semantic control, i.e. the ability to access and manipulate meaningful information, for example by inhibiting the dominant or amplifying the less dominant aspects of concepts and resolution of ambiguity or incongruent meanings (Jefferies 2013, Jackson 2021). Given the previously mentioned association between social and semantic cognition, it is not surprising that both regions are also implicated in processing moral information, with IFGpo involved in moral reasoning and anterior insular activity related to moral disgust and advantageous inequity (Gao et al. 2018, Ying et al. 2018, Diveica et al. 2021). Previous neuroimaging research has linked the structure of these areas with anxiety and guilt (Belden et al. 2015, Shang et al. 2014, Hu and Dolcos 2017). Furthermore, in individuals with social anxiety disorder (SAD), the cognitive reappraisal network, which includes the left IFG as one of its nodes, exhibits decreased influence on the emotional reactivity network during reappraisal of interpersonal criticism (Jacob et al. 2019). Additionally, patients with obsessive-compulsive disorder (OCD) show decreased activity of the insula when experiencing deontological guilt, i.e. when one feels they have violated their own moral code (Basile et al. 2014). Our finding of a negative correlation between trait anxiety and the efficiency of information integration around these two areas extends the previous literature by showing that the functional characteristics present in anxiety patients are also observed in subclinical groups, which is in line with the spectral nature of mood and anxiety disorders. Hence, it suggests that similar alterations in the architecture of brain networks may contribute to guilt symptomatology across both subclinical and clinical disorder levels. As such, we interpret the functional disconnection of the left IFGpo and adjacent parts of insula reported here as a correlate of the deficiency in one’s ability to modulate the retrieval and selection of conceptual-level knowledge, diminishing the likelihood of successful regulation of social experiences labelled as guilt.

As for the structural graph, associations between anxiety and solely global parameters were found, suggesting that, unlike in the functional network, anxiety is to a larger extent linked to the architecture of the circuitry as a whole rather than anatomical connections of its particular nodes. The presence of positive links between the network’s clustering coefficient and anxiety scores at 14%–17% densities indicates that in the more anxious participants, we observe a neural architecture favouring increased local information processing. As mentioned before, at the discussed sparsities, the guilt processing circuitry appears organized into a semantic submodule localized in the left temporal lobe, and an affective subcircuit consisting mainly of bilateral subcortical and cortical midline nodes (Supplementary Fig. S6, Green et al. 2012, Lythe et al. 2015, Eslinger et al. 2021). Given that hypersensitivity to emotional stimuli has been reported across both anxiety patients and individuals with high-trait anxiety (Hajcak et al. 2003, Weinberg et al. 2010, Voegler et al. 2018), the increased communication within both modules could constitute a neural correlate of this characteristic and contribute to the prevalence of maladaptive forms of guilt observed in anxiety disorders (Cândea and Szentagotai-Tătar 2018). Nevertheless, as our findings describe the circuitry only when a highly specific subset of anatomical connections is considered, they cannot be generalized to the network as a whole and should be treated with caution.

It is worth pointing out that the current study uses a different approach to a vast majority of neuroimaging works, which typically probe associations of particular cognitive processes with characteristics of specific canonical resting-state networks, like the default mode network. However, the more complex mental processes, such as guilt, require cooperation between associative areas belonging to more than one such network. To reflect this, the current approach offers measures of regional and global integration within a cross-network circuitry restricted to brain regions governing a particular cognitive process, conceptually yielding the findings highly relevant to actual behaviour. Therefore, it appears particularly useful for investigating complex cognitions, such as social emotions and decision-making, and in theory could also be utilized with task fMRI data.

One limitation of our study comes, however, from the fact that the conclusions are drawn on a limited sample of 78 young adults, affecting the generalizability of findings. As such, before any potential application, larger lifespan subclinical and patient cohorts should replicate the reported associations and investigate the effects of age (Orth et al. 2010). The relatability of the current results to those derived from previous MDD and OCD studies (Green et al. 2012, Basile et al. 2014) is also partially restricted due to psychopharmacological treatment in some of the patients. Furthermore, on a more methodological note, we stress out the fact that the functional networks were constructed only using the significant positive correlations. While negative (inhibitory) connections between brain regions are also known to be an important characteristic of a functional network’s architecture, they are better suited to be investigated with task-related data. As such, the approach used in the current work makes the topological graph measures more easily interpretable and thus increases the utility of the reported findings (Qian et al. 2018). Last but not least, as the reconstruction of longer streamlines, including interhemispheric connections, is often less accurate, the contribution of the nonfrontal right hemisphere regions to the structural connectome may be underestimated, while the contributions of the bilateral frontal regions may be biased in the opposite direction (Yeh et al. 2021). As such, the hubs of the structural network presented in the Supplementary material should be treated with caution.

Conclusions

Recently, we have been observing a shift in paradigm towards personalized approaches in psychiatry with the hopes of increasing the efficiency of deployed treatments. Given the spectral nature of mood and anxiety disorders, our work maps onto this new framework by trying to find in a subclinical sample a link between a specific symptom (guilt) and the alterations in the underlying neural circuitry. Our findings provide an important insight that the prevalence of self-blame symptomatology may be primarily related to the increased functional interconnectedness of the left sATL within the guilt processing network, which may manifest itself as a more pronounced communication of the guilt ‘label’ to its functionally coupled regions. This process may be further reinforced in highly anxious individuals by the functional isolation of the left IFGpo and insula, a potential marker of deficiency in one’s ability to reinterpret the guilt-associated social experiences, and the network’s architecture favouring increased local information processing within the circuitry’s semantic and affective submodules, however, only when a highly specific subset of anatomical connections is considered.

Author contributions

Michal Rafal Zareba (Conceptualization, Formal analysis, Visualization, Writing—original draft), Krzysztof Bielski (Formal analysis, Writing—review & editing), Victor Costumero (Conceptualization, Writing—review & editing), and Maya Visser (Conceptualization, Supervision, Writing—review & editing)

Supplementary material

Supplementary material is available at SCAN online.

Conflict of interest

None declared.

Funding

This publication forms part of the following research projects: grant PID2021-127516NB-I00 funded by MICIU/AEI/10.13039/501100011033 and by ‘ERDF/EU’; grant RYC2019-028370-I funded by MICIU/AEI/10.13039/501100011033 and by ‘ESF Investing in your future’; grant RYC2021-033809-I funded by MCIN/AEI/10.13039/501100011033 and by ‘NextGenerationEU/PRTR’; grant CIAICO/2021/088 funded by Conselleria de Educación, Universidades y Empleo; and grant UJI-B2022-55 funded by Universitat Jaume I.

Data availability

The described analyses were performed with the use of an anonymized openly available dataset (Babayan et al. 2019).