Main

Diffuse gliomas are the most common and most deadly primary malignant brain tumors1,2. They are characterized by infiltrative growth and can cause a wide range of neurological, cognitive, and neuropsychiatric symptoms3,4. Patients with diffuse glioma commonly undergo surgical resection as a first treatment step4.

After any cancer diagnosis, patients often experience sadness and distress during a few weeks to months, and depressive symptoms usually peak shortly after cancer diagnosis5,6. While it is understandable that patients recently diagnosed with life-threatening diseases may exhibit certain depressive symptoms, severe depressive symptoms are burdensome and negatively affect quality of life7. Up to 54% of preoperative patients with brain tumors report depressive symptoms, which is more frequent than in other cancer types and may be explained by patient characteristics, psychometric properties of the questionnaires used, or possibly the tumor location8,9.

As some depressive symptoms are congruent with recent bad news of the incurable nature of the disease and information on treatment-related risks, an absence of depressive symptoms seems discrepant. An absence of depressive symptoms has not been reported on before but could possibly reflect a maladaptive response such as emotional blunting. Emotional blunting, the numbing of both positive and negative feelings, has been described in brain tumor patients, as has apathy10,11,12. Emotional blunting is related to impaired functioning and reduced quality of life13. We hypothesize that the presence of intracranial lesions that disturb brain functions could contribute to severe depressive symptoms and an absence thereof10,11.

For many brain diseases, including glioma, symptoms have been mapped to lesion locations to find relations between symptoms and brain structures14,15,16. This mapping is traditionally analyzed by voxel-based lesion symptom mapping, which suffers from poor correction of multiple testing and ignores spatial correlations between voxels17. Therefore, spatial brain relations are preferably taken into account as these are more likely to explain symptoms18,19,20. Mapping symptoms to brain regions as a whole and brain networks of connected regions can be a rich source of information that can be applied to routine clinical imaging21,22,23.

In stroke, depressive symptoms are related to the location of the lesion, such as the prefrontal cortex and the thalamocortical and dorsal frontal white matter tracts24,25,26. However, differences in pathophysiological mechanisms between stroke and diffuse glioma may result in distinct lesion symptom associations27. Few studies investigated the association between brain tumor location and depressive symptoms; none investigated the absence thereof. Previous studies found no association between depressive symptoms and lobar involvement but lacked more detailed brain parcellations8,28,29. One study summarized 18 published case reports of patients with depression and found 89% of the depression-related tumors were functionally connected to the left striatum, the putamen, and the pallidum, according to voxel-wise analysis30.

In this study, we hypothesize that tumor location is related to both severe depressive symptoms and an absence of depressive symptoms in preoperative patients with diffuse glioma. To gain insight into the neuroanatomy and etiology of these phenomena, we mapped depression scores to brain regions and networks with tumor infiltration and their corresponding disconnected regions. We applied several methods: (1) a data-driven cluster analysis without pre-defined regions, (2) a lesion load analysis including white matter and cortical and subcortical gray matter parcels, (3) probability of involvement of white matter tracts, and (4) modeled functional network impact; see Extended Data Figs. 1 and 2 for an overview of methods.

Results

Patient population

We identified 203 patients with a newly diagnosed diffuse glioma who completed the Center for Epidemiologic Studies Depression scale (CES-D) within a year before surgery. Two patients were excluded from the analysis due to poor registration of their magnetic resonance imaging (MRI) to standard brain space, which resulted in 201 patients with characteristics as listed in Extended Data Tables 1–3.

Depressive symptom distribution and categories

An absence of depressive symptoms (CES-D ≤ 6) occurred in 29% of patients and severe depressive symptoms (CES-D ≥ 23) in 14% of patients (Fig. 1).

Fig. 1: Distribution of depressive symptom scores and corresponding categories.
figure 1

Number of patients versus CES-D score.

Full size image

Clinical characteristics

Of the patient characteristics, female sex (mean regression coefficient (RCμ) = 1.10, 94% posterior highest density interval (HDI94%) = 0.18–2.03) was associated with severe depressive symptoms, and high Verhage educational level (RCμ = 0.95, HDI94% = 0.15–1.81) and high 36-Item Short Form (SF-36) physical functioning score (RCμ = 0.09, HDI94% = 0.05–0.14) were associated with an absence of depressive symptoms. None of the other patient or tumor characteristics was significantly associated with either CES-D category.

Tumor and disconnectome distributions

The maps of the segmented tumor locations are shown in Fig. 2a, for both the complete group and split by depressive symptom category. We also created disconnectomes based on normative tractography, which reflect white matter tracts that are likely affected by a tumor lesion, as shown in Fig. 2b. When tested using sparse canonical correlation analysis (Extended Data Fig. 1; cluster-based analyses), tumor and disconnectome locations were not significantly associated with CES-D.

Fig. 2: Distribution maps for the total population by depressive symptom categories.
figure 2

a, For tumor locations. b, For disconnectome locations. The color legends to the right indicate the number of patients with tumors overlapping at a voxel location; the maximum numbers vary by category. The plotted numbers below refer to Montreal Neurological Institute 152 standard space z-axis slice. L, left; R, right.

Full size image

Tumor and disconnectome load

We then tested tumor load and disconnectome load on pre-defined brain regions using Bayesian categorical regression (see Extended Data Fig. 2: Lesion load). The load was calculated by dividing the lesion volume within a region by the total region volume.

Tumor load in the right thalamus was significantly associated with an absence of depressive symptoms (RCμ = 7.60, HDI94% = 1.19–14.43) (Fig. 3). See Fig. 5 and Supplementary Fig. 2 for all associations. Tumor load in the white matter tracts or cortical functional regions and disconnectome load on white matter tracts were not associated with CES-D category.

Fig. 3: Tumor load significant associations with depressive symptom categories.
figure 3

The right thalamus is associated with an absence of depressive symptoms. Here visualized in two dimensions and three dimensions. L, left; R, right.

Full size image

White matter tract involvement

Next, we tested whether the probability of involvement, also known as the probability of disconnection, related to depressive symptom categories (see Extended Data Fig. 2: Probability of involvement). Severe depressive symptoms were associated with white matter tract involvement of the right corticospinal tract (CST; RCμ = 4.63, HDI94% = 1.02–8.43), the right fornix (RCμ = 11.31, HDI94% = 6.62–15.50), the right inferior fronto-occipital fasciculus (IFOF; RCμ = 5.21, HDI94% = 0.24–10.30), and the left uncinate fasciculus (RCμ = 6.30, HDI94% = 2.08–10.12) (Fig. 4a).

Fig. 4: White matter tract involvement has significant associations with depressive symptom categories.
figure 4

a, White matter tracts where involvement was associated with severe depressive symptoms. From top to bottom: a 3D overview of all associated tracts, the right corticospinal tract (CST), the right fornix, the right fronto-occipital fasciculus (IFOF), and the left uncinate fasciculus (UF). b, White matter tracts where involvement was associated with an absence of depressive symptoms. From top to bottom: a 3D overview of all associated tracts, the right temporal cingulum (CT), the left first segment of the superior longitudinal fasciculus (SLF 1), and the left UF. The color intensities represent the RCμ. L, left; R, right.

Full size image

Interestingly, involvement of the left uncinate fasciculus was also significantly associated with an absence of depressive symptoms (RCμ = 3.65, HDI94% = 1.16–6.00) (Fig. 4b). In addition, involvement of the right temporal cingulum (RCμ = 5.88, HDI94% = 1.96–9.61) and the left superior longitudinal fasciculus (RCμ = 1.98, HDI94% = 0.33–3.81) were significantly associated with an absence of depressive symptoms. See Fig. 5 and Supplementary Fig. 2 for all correlations.

Fig. 5: Overview of significant associations for each analysis.
figure 5

a, For severe depressive symptoms. b, For an absence of depressive symptoms. Bubble size corresponds with effect size.

Full size image

Modeled functional network impact

Last, we modeled the potential impact of the tumor on functional networks using large normative functional network data. We calculated impact on graph measures local efficiency and eigenvector centrality to evaluate its relation to depressive symptom categories (Extended Data Fig. 2: Functional network impact). We found no significant associations.

Discussion

We explored how severe depressive symptoms and an absence thereof related to tumor and disconnectome locations and modeled functional network impact in patients with supratentorial diffuse glioma. The locations of tumors, in addition to established patient-related risk factors, were related to these extremes of depression scores.

Our findings confirmed that severe depressive symptoms occurred in 14% of patients with brain tumors and that female sex is a risk factor31. In addition, we found structures of the limbic system to be related to severe depressive symptoms (the right fornix and the left uncinate fasciculus)32 as well as regions involved in movement and in language, goal-oriented behavior, and visuospatial attention (the right CST and the right IFOF)33.

These results are in line with findings in major depressive disorder, where the limbic–thalamo–cortical circuit is pivotal34. Specific depressive symptoms have been associated with specific regions and circuits35, for example, the fornix in memory, disrupted cognitive control, and self-referential thought32,36,37. The uncinate fasciculus is involved in integrating visceral and emotional information, and dysfunction can result in cognitive and behavioral problems32. Furthermore, brain–depression relations have also been investigated in depression secondary to neurologic disorders. Post-stroke depression is most widely examined, and studies have found anterior structures, the reward circuit, and limbic structures to be involved26,38. Studies into depression after brain diseases other than stroke have identified similar regions, including the uncinate fasciculus, IFOF, CST, and frontolimbic circuits39,40,41.

In addition to these limbic structures, the right IFOF was related to severe depressive symptoms and has been implicated in non-verbal semantic processing and visuospatial awareness42,43. Possibly, lesions in this region result in more disability and consequently depression44. Indeed, the IFOF has been related to post-stroke depression and major depressive disorder, possibly as a result of cognitive changes45,46. A similar secondary mechanism could drive involvement of the CST with voluntary movement. Perhaps lesions in these areas result in more physical disability, resulting in more depressive symptoms although we did consider functional impairment as a covariate47.

We identified an absence of depressive symptoms in 29% of patients, which we considered remarkably high. Higher educational level related to an absence of depressive symptoms, corroborating earlier research31,48. In addition, higher scores on physical functioning related to an absence of depressive symptoms. Taking educational level, sex, and physical functioning into account, regions involved in an absence of depressive symptoms included the limbic regions right temporal cingulum, right thalamus and left uncinate fasciculus, and left SLF 1.

An absence of depressive symptoms is likely to be multifactorial in origin. In general, after a cancer diagnosis, a delayed realization of illness can occur, possibly contributing to an absence of depressive symptoms in the first phase49. In addition, resilience and even thriving can develop after cancer diagnosis and can relate to an absence of depressive symptoms. However, this generally occurs later in the disease and would therefore not explain the high prevalence we observed50. In addition, denial of cancer diagnosis could contribute to an absence of depressive symptoms. In newly diagnosed lung cancer patients, high levels of denial occurred in 3% of patients and indeed related to fewer depressive symptoms51,52. Indeed, brain tumor patients have been described to underestimate their problems compared with their caregivers, which has been hypothesized to be due to reduced insight or denial53. Another neuropsychiatric explanation may be anosognosia: the absence of awareness of a disease or dysfunction. Anosognosia occurs in around 10% of acute stroke patients and is related to worse functional outcomes54. However, the literature remains inconclusive on whether anosognosia relates to more or fewer depressive symptoms55,56,57,58,59. An absence of depressive symptoms could also reflect dysfunction of the mood circuitry. Indeed, the fornix, which in our study was related to severe depressive symptoms, has previously been associated with apathy60,61,62. Moreover, we found the left uncinate fasciculus to be associated with both severe depressive symptoms and an absence of depressive symptoms, suggesting the potential for a dual response to its dysfunction.

Finally, an absence of depressive symptoms could indicate emotional blunting, which is an absence of emotional response to an emotional stimulus such as a cancer diagnosis, or apathy, defined as diminished goal-directed behavior and decreased emotion or feelings or interest63,64. Apathy does have characteristics in common with depression, such as no desire to pursue reward, but it is a unique entity with decreased emotion rather than feelings of depression65,66. Both emotional blunting and apathy are related to worse functional outcomes and reduced quality of life13,67. Emotional blunting and apathy have been reported after brain lesions and are common in neurodegenerative diseases. Regions associated with an absence of depressive symptoms in our cohort have previously been described in apathy and emotional blunting: the thalamus is involved in decreased emotional responses and cognitive processes68, and lesions in this region are related to post-stroke apathy and worse self-reported cognitive functioning69,70. In neurodegenerative diseases, apathy has been associated with white matter alterations in limbic structures: the temporal cingulum and the uncinate fasciculus62,71,72,73, which are involved in motivation63. In schizophrenia, negative symptoms have also been associated with alterations in the temporal cingulum74,75. This could be due to memory formation and retrieval74 or due to absence of interest or emotional reactivity62. The superior longitudinal fasciculus has also been associated with apathy in various populations62,71,72,73,76,77, the SLF 1 specifically with attention78.

In summary, an absence of depressive symptoms could stem from various mechanisms. Considering the high prevalence of an absence of depressive symptoms in our glioma population, a neuropsychiatric basis such as a dysfunction of the mood circuitry or emotional blunting seems plausible. We encourage future research to explore this phenomenon and its etiologies.

Our findings confirm the importance of the limbic system both in depressive symptoms and in an absence of depressive symptoms. However, contrary to associations found in (post-stroke) depression and apathy, we found no evidence of neocortical regions interfering with the limbic system in glioma patients. Specifically, the dorsolateral prefrontal cortex, as part of the frontoparietal network, seems crucial in depression and apathy following stroke26,68,79. However, in our study, tumors in this region were not associated with depression scores. In general, depression could be the result of hypoactivity of the cortex or hyperactivity of the limbic system80. Brain tumors possibly increase the activity of the limbic system rather than decrease cortical regulation of the limbic system as brain tumors have been shown to increase neuronal activity81. This may explain why different brain lesion etiologies can cause different lesion symptom results:27,30 in patients with gliomas, the brain is functionally disturbed at distance from the tumor82,83,84,85. The interaction between tumors and the surrounding brain seems to be bidirectional: not only does increased neuronal activity promote tumor growth86,87, the tumor also induces both neuronal death and neuronal hyperexcitability and can increase and decrease peri-tumoral connectivity88,89,90,91,92. In support of this, glioblastomas have been demonstrated to modify neural circuits by overactivating brain areas around language regions found in healthy individuals90. We speculate that neural circuits involved in mood regulation, such as depressive symptoms, may also be remodeled by glioma by both hyperactivity and hypoactivity. Further research is required in this emerging field of cancer neuroscience.

In this Article, we adopted a rigorous design with several methods to investigate the association between depressive symptoms and tumor location. Not all methods identified a lesion–depression correlation. First, the cluster-based analysis yielded no regions associated with depression scores, possibly because the tumors are too large to identify small regions: LESYMAP has been developed in the context of stroke, in which the lesions are possibly smaller. Moreover, we used the continuous CES-D to relate to clusters, and a nonlinear relation between a brain region associated with both extremes of depression scores may not become apparent by assuming linearity. Furthermore, the statistical power to detect an association is not equally distributed over brain regions, varying by preferential locations of tumors, so that relevant brain regions may remain undetected. Second, the disconnectome of a tumor may incorrectly conceptualize an entire loss of function from disconnection. Similarly, the lesion load does not necessarily reflect a complete loss of function of the tumor-infiltrated brain region. In addition, especially in white matter tracts, a specific location within the tract may be more important than lesion load, for example, in a location with higher axonal density93. The dysfunction of a white matter tract is presumably better reflected by the probability of a tumor involving that tract, capturing switch-like functional intactness94. Third, modeled functional network impact also did not relate to depressive symptoms. Brain tumors presumably interfere more dynamically with the network than our modeled one-time differences, which is probably more fitting with stroke27. However, a better model of the effects of a growing tumor on the functional network does currently not exist to our knowledge. Nevertheless, an alternative approach using voxel-wise or peak cluster analysis to lesion network mapping, as described in other work21,24,79,95, may illuminate new associations between the functional network and depressive symptoms and could be considered in future work.

Limitations

Our study also highlights some challenges of tumor lesion symptom mapping. First, tumors do not have an acute onset with complete loss of function of normal brain tissue, but develop gradually over time and may result in partial loss of function given their infiltrative nature. This gradual growth often induces neuroplasticity, and therefore lesion symptom mapping in brain tumors and stroke could result in different findings27,30,96,97. Nevertheless, most of our findings corroborate the brain–depression literature. Second, tumor segmentation is subject to inter- and intra-rater variability, although this is usually not a major source of variation98. Third, tumor extent may be underestimated as there is diffuse involvement outside the focal lesion. Fourth, the unknown combination of expansive and infiltrative growth of the tumor makes registration to standard space inherently challenging99. To minimize these errors, all segmentations and registrations were visually verified in 3D Slicer. Fourth, the combined error from previous arguments may also lead to spatial bias and thus potential misinterpretation of tumor location in atlas space. In addition, the parcels of brain atlases do not necessarily represent units of depressive symptoms. Furthermore, we disregarded the anticorrelations as the interpretation of a brain region safeguarding against a dysfunction on tumor infiltration remains elusive21,95. In addition, we used the CES-D as a single measure of depressive symptoms within a month of a major surgical procedure. The CES-D focuses primarily on symptom frequency rather than intensity100, and this may be temporarily elevated before surgery, in particular in patients with neurologic deficits. For a thorough diagnosis of mood (dys)function in the complex circumstances following a grave tumor diagnosis and facing a major surgical procedure, diagnostic interviews by psychiatric professionals would be required. Moreover, no consensus on a cut-off for an absence of depressive symptoms exists. We undertook our best efforts, using several methods, to establish a convincing cut-off, and our findings align with clinical observations of apathy at presentation in patients with newly diagnosed glioma101. Still, verifying this phenomenon by comparing CES-D with other populations who just received a grave diagnosis would be advisable. Last, our population may be subject to selection bias. Patients with insufficient capacity to fill out the CES-D questionnaire were unavoidably missing from this analysis, possibly underestimating mood dysfunction.

Conclusions

Patients with supratentorial diffuse glioma experience both severe depressive symptoms (14%) and an absence of depressive symptoms (29%) before surgery. These possible mood dysfunctions are partly explained by tumor location. Severe depressive symptoms are associated with tumors in the right corticospinal tract, fornix, and inferior fronto-occipital fasciculus and the left uncinate fasciculus, whereas absence of depressive symptoms is associated with tumors in the left uncinate fasciculus and first segment of the superior longitudinal fasciculus and the right temporal cingulum and thalamus. Awareness of this phenomenon may be important for the identification of at-risk patients and patient counseling.

Methods

Patients

The study population consisted of 203 patients with a newly diagnosed supratentorial diffuse glioma included in several observational studies at the Amsterdam UMC location of Vrije Universiteit Amsterdam between 2009 and 2022. All participants gave their informed consent for inclusion before study participation. Patients received no financial compensation for study participation. The study was conducted in accordance with the Declaration of Helsinki, the protocol was approved by the Medical Ethical Committee of the Amsterdam UMC location of Vrije Universiteit Amsterdam (2008.52; 2009.189; 2010.126; 2014.297), and all patients signed informed consent. We included patients of a combined sample that has partly been previously reported on102. This specific subpopulation and the MRIs have not been analyzed before. We followed the STROBE (Strengthening the Reporting of Observational Studies in Epidemiology) guidelines.

Inclusion criteria were patients who (1) were ≥18 years old, (2) had an MRI before surgery, (3) completed the CES-D within a year before surgery, and (4) had a histopathologic diagnosis of supratentorial diffuse glioma WHO (World Health Organization) 2016 grade II–IV following surgery. The MRI included at least a 3D T1-weighted scan without gadolinium (T1w) and a contrast-enhanced T1-weighted scan after gadolinium administration (T1c). If available, T2-weighted (T2w) and fluid attenuated inversion recovery (FLAIR)-weighted sequences were also collected. Only patients without a previous resection were included. Patients typically undergo several MRIs before surgery: initially when a lesion indicative of a diffuse glioma is suspected and again just before surgery to assist in procedural planning. Additional MRIs may be performed to monitor the lesion’s progression. For the purpose of this study, we selected the preoperative MRI closest to the date of surgery for tumor segmentation.

We collected patient and tumor characteristics, including age, sex, handedness, educational level (low, middle, or high as classified according to the Verhage scale)103, Karnofsky Performance Scale104, SF-36 physical functioning subscale (higher scores represent better physical functioning)105, whether patient had a history of a neuropsychiatric disease such as depression, information on the use of antiepileptic drugs or antidepressants, type of surgery (biopsy or resection), tumor grade and tumor type according to the WHO 2016 classification1, and various time intervals: the time interval between first brain MRI with a lesion suspect for a diffuse glioma and CES-D measurement, the time interval between the CES-D measurement and the preoperative MRI used for tumor segmentation, and the time interval between CES-D measurement and surgery. We categorized antiepileptic use into three groups: none, levetiracetam, and other antiepileptics, given the potential negative impact on mood from levetiracetam106. All patients had surgery as first cancer treatment, so radiotherapy and chemotherapy were not included as confounders in the analysis.

Statistics and reproducibility

This study combined data from multiple observational studies done at our institute. We combined these raw datasets and extracted variables and MRIs of interest (see the preceding paragraph), which had not been used for publication before. The MRIs were processed as described in Tumor locations. The MRIs were anonymized to ensure blinding during segmentation. No statistical method was used to predetermine sample size.

Depressive symptoms

Depressive symptoms were assessed with the CES-D, which is a widely used patient-reported outcome measure for depressive symptoms100. It has good validity and reliability in cancer patients and consists of 20 items about feelings and behaviors during the past week107. Scores for each item range from zero points (rarely or none of the time (less than 1 day)) to three points (all the time (5–7 days)), adding to a total score of 0–60. Higher scores indicate more depressive symptoms. We considered a score of ≥23 as severe depressive symptoms. An alternative cut-off of ≥21 yielded similar results in a sensitivity analysis. An overview of the number of patients per category can be found in Extended Data Tables 1–3.

Absence of depressive symptoms

A cut-off for an absence of depressive symptoms is lacking, therefore we based an arbitrary cut-off on several methods. First, as our population faced challenging circumstances, we expected some level of symptoms measured through the CES-D, that is, feeling fearful, restless sleep, decreased appetite, concentration difficulties, and/or feeling sad, similar to other newly diagnosed cancer patients. If patients scored the maximum three points for these five items, they would have a CES-D of at least 15. Conservatively, we decided half of these points (7.5, rounded conservatively to seven) would be a remarkable absence of depressive symptoms, especially considering the 15 other items on which patients could score additional points. Second, the mean CES-D score is eight to nine in the general population108. Third, we visually compared the CES-D distribution of our population with another newly diagnosed cancer population without intracranial metastases (Supplementary Fig. 1), consisting of 100 women who received a breast cancer stage 0–II diagnosis in the previous three months and were recruited for a psychoeducational intervention aimed at improving post-diagnosis distress109. In these patients also receiving a tumor diagnosis, 19% demonstrated a CES-D below 7, which is well below 29% and therefore corroborated our arbitrary cut-off.

Tumor locations

To determine tumor location, tumors were segmented using an automated nnU-Net algorithm110 that can deal with missing pulse sequences followed by verification and manual editing (P.C.d.W.H., M.N.G.v.G., V.B.) under the supervision of an experienced neuroradiologist (F.B.) using 3D slicer v.5.0.2 (refs. 111,112). The segmentations included non-enhancing tumor parts from T2/FLAIR sequences combined with contrast-enhancing tumor parts and non-enhancing enclosed necrosis and cysts from T1w/T1c sequences. Segmentations were transformed from patient space (T1c) to Montreal Neurological Institute 152 standard space and resampled to 2 × 2 × 2 mm spatial resolution. Before transformation, skull stripping was performed using HD-BET brain extraction tool followed by a nonlinear registration with cost-function masking to standard space using ANTsPY v.0.3.2 (refs. 113,114). Tumor distribution maps in standard space were constructed by summing the tumor segmentations over all patients (Extended Data Fig. 1, upper right panel).

Disconnectome locations

The disconnectome represents white matter tracts likely affected by a (tumor) lesion. For each patient, a disconnectome was made using the BCBtoolkit v.4.2.0 (ref. 115). The BCBtoolkit creates disconnectomes from diffusion-weighted imaging data of 178 healthy subjects from the 7T Human Connectome Project22,116. In short, tumor segmentations in standard brain space were registered to each healthy participant’s space using affine and diffeomorphic deformations117,118. The registered tumor segmentations were then used as seed for tractography in Trackvis for each healthy participant119. From the tractographies, a binary visitation map was constructed from each healthy participant for each patient, showing for each voxel whether it was intersected by a tract. Then the map was registered back to standard space and an averaged disconnectome was constructed for each patient from the visitation maps of all healthy participants. The resulting patient-specific disconnectome thus represents interindividual variability of tract reconstructions from healthy participants, with each voxel representing a probability of involvement ranging from 0 to 100%. To obtain the patient-specific binary disconnectome, we thresholded the disconnectome with a probability of 50% or more115. We then summed the binarized disconnectomes over all patients to create a structural disconnectome distribution map (Extended Data Fig. 1, upper right panel).

Lesion load analysis

The lesion load of tumors on specific brain structures was determined for each patient by considering parcels from standard brain atlases. The load was calculated as the tumor volume within a given parcel divided by the parcel’s total volume. Three atlases were used: (1) the XTRACT probabilistic white matter atlases to define the subcortical white matter pathways using a tract probability of >0% to create binary tract masks, resulting in 41 tract structures120, (2) Yeo’s network atlas to define seven cortical parcels representing conjoined functional networks in each hemisphere, resulting in 14 cortical functional regions121, and (3) the Harvard–Oxford atlas, resulting in 14 deep gray nuclei122,123,124,125. Similarly, disconnectome load on XTRACT white matter atlases was calculated.

Probability of involvement of white matter tracts

In addition, the probability of involvement—also known as the probability of disconnection—of a white matter tract by tumor location was determined by mapping the tumor lesions of each patient to XTRACT white matter atlases120. The probability was measured by determining the maximum probability of the tract crossing a tumor lesion using Tractotron software available in BCBtoolkit115. We considered a tract involved when the probability was larger than 0.5 (ref. 126).

Statistics

First, we determined the association between the CES-D category and the patient and tumor characteristics by applying a Bayesian categorical multiple regression model. We included age, sex, handedness, education level, antiepileptic drugs, SF-36 physical functioning subscale, tumor grade, tumor volume in milliliters, and the time interval between CES-D measurement and the first brain MRI with a suspected diffuse glioma (Extended Data Fig. 1, upper left panel). For six patients, the SF-36 questionnaire was missing. We used median interpolation of the SF-36 to include these patients in the model. The model with patient and tumor characteristics that were significantly associated with the CES-D category was considered the core model and used as the base for the following analysis.

The Bayesian regression model had vaguely informative priors as we had no specific prior knowledge on effect size of variables. The Markov chain Monte Carlo settings were at 2,000 draws and four chains. The means of the posterior distributions were considered as estimates of the regression coefficients. Coefficients were considered statistically significant if the HDI94% excluded zero. This threshold was originally chosen as a reminder that Bayesian credibility interval cut-offs are an arbitrary value determined by consensus. Models were generated and run using the PyMC package through the Bayesian model-building interface (Bambi) v.0.9.3 in python v.3.8.0 (ref. 127).

To determine the association between continuous CES-D scores and the locations of tumors and disconnectomes without assumptions on delineation of separate brain regions, we performed voxel-wise multivariate lesion symptom mapping using sparse canonical correlation analysis (SCCAN) (Extended Data Fig. 1, lower panel). This analysis was conducted using LESYMAP package v.0.0.0.9221 in R v.4.1.3 (ref. 128). The SCCAN method optimizes voxel weights that maximize the multivariate correlation between voxel values and CES-D scores. Fourfold within-sample cross-validation was used to evaluate the significance of the map. Using this method avoids some pitfalls associated with voxel-based methods as the significance of the entire map is tested at once128. We excluded voxels with fewer than three lesions.

To examine whether tumor or disconnectome load was related to CES-D category, we performed a Bayesian categorical regression as described in the preceding. The CES-D category was used as the dependent variable and tumor or disconnectome load per region for each of the specified atlases as independent variables, in addition to the core model (Extended Data Fig. 2, upper row).

To investigate whether white matter tract involvement was related to the CES-D category, we performed a categorical Bayesian regression with the involvements per tract as independent variables in addition to the core model (Extended Data Fig. 2, middle row).

Modeled functional network impact

We modeled potential tumor impact on functional networks by virtually lesioning normative functional network data according to patients’ tumor locations. We utilized eigenvector centrality (EC) and local efficiency (LE) as graph theoretical measures to determine the importance of brain regions—regarded as nodes in the network129. The EC measures the importance of a node by the number of its own and its neighboring nodes’ connections;130 LE determines how connected the neighbors of a node are131.

As a normative reference, we used processed132 connectivity matrices of 1,000 publicly available Human Connectome Project participants. The matrices contain Pearson correlations that describe the pairwise correlations or co-activations between brain regions from resting-state functional MRIs. Brain regions consisted of 400 cortical parcels in Yeo’s seven networks from the Schaefer surface atlas and 16 subcortical gray matter parcels from the Tian surface atlas133,134. We averaged all matrices to create a normative connectivity matrix and calculated graph measures LE and EC for each parcel (node) using the correlations as weights.

Next, using Freesurfer v.7.3.2 and Connectome Workbench v.1.5.0, we brought the tumor segmentations from volume to fsLR 32k surface space and overlaid these segmentations with the atlases to identify tumor-infiltrated brain region nodes. To create a synthetically lesioned matrix for each patient, we removed completely tumor-infiltrated parcels from the normative connectivity matrix and lowered the weights of partially tumor-infiltrated parcels according to the percentage of tumor load. We then calculated LE and EC for each remaining parcel of each patient and subtracted these from the normative matrix. To acquire the LE and EC difference of the seven networks due to the tumor, we calculated the median over all parcels within a network for each patient.

Finally, to determine the association between the CES-D category and modeled functional network impact, we used the median absolute graph measure differences in a Bayesian categorical regression. We performed a separate analysis for each measure, with the modeled difference measures of the seven networks as independent variables in addition to the core model (Extended Data Fig. 2, lower row).

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.