Main

Multiple sclerosis (MS) is a chronic immune-mediated disease associated with demyelinating white matter lesions in specific regions of the central nervous system. Depression is one of the most common and debilitating complications of MS1,2 and is associated with greater impairment in cognitive function3, physical function4 and quality of life5. Determining the neuroanatomical substrate for MS depression could be useful for identification of treatment targets for transcranial magnetic stimulation (TMS) and deep brain stimulation (DBS) or other brain stimulation techniques6,7,8. This is particularly important given that people with MS are less responsive to antidepressant pharmacotherapy, and there are no clinical trials of brain stimulation for MS depression.

While depression and overall white matter lesion load both increase with disease progression4, attempts to relate MS depression to specific lesion locations have shown conflicting results9. Different studies have reported associations between MS depression and lesions in the frontal lobes10, temporal lobes11 and parietal lobes12. Due to this heterogeneity, it remains unclear whether or how lesion location is related to depression in MS4.

When lesions causing a specific symptom do not consistently localize to a specific region, they may still localize to a specific distributed brain circuit. These circuits can be identified using lesion network mapping, a technique that uses a normative connectome database (n = 1,000) to compare the functional connectivity of brain lesions rather than just their location13. This technique is traditionally used to study grey matter lesions, as in a recent study of depression associated with focal lesions caused by stroke and penetrating head trauma6,14. We found that grey matter lesions that cause depression are functionally connected to a common brain circuit. Depression improved after TMS to the positive parts of this circuit or DBS to the negative parts of the circuit, suggesting that lesion-derived circuits can be effective targets for therapeutic brain stimulation6. We identified a convergent neuroanatomical substrate for depression across 14 datasets of lesions, TMS sites and DBS sites6. This convergent circuit included positive connections to the dorsolateral prefrontal cortex and negative connections to the subgenual cingulate and ventromedial prefrontal cortex.

Of note, functional connectivity has been used primarily to study grey matter as it yields poor signal-to-noise ratio in the white matter15,16,17. As a result, it remains unclear whether this circuit applies to white matter lesions such as those seen in MS. It may be possible to overcome this limitation with large sample sizes, which can provide enough signal to overcome the noise21. This is consistent with increasing evidence that white matter signal can still reveal relevant network neuroanatomy15,16, as summarized in a recent exhaustive review of white matter functional connectivity findings17.

In this Article, we use functional connectivity between MS lesion locations and our previously described a priori depression circuit to test the hypothesis that MS lesions connected to our a priori circuit are more likely to cause depression.

Results

Participants with MS (n = 281) were identified from a longitudinal clinical and radiological database. Participants had complete magnetic resonance imaging (MRI) data, depression scores (Quality of Life in Neurological Disorders (Neuro-QoL) depression subscale) and overall disability scores at enrolment. Dataset characteristics are summarized in Fig. 1a. Fluid-attenuated inversion recovery (FLAIR) hyperintense lesions were defined using an automated white matter segmentation pipeline as previously described18. Lesions were distributed throughout the white matter, with high density in periventricular regions as expected (Fig. 1b). On one-way analysis of variance, choice of immunomodulatory therapy was not associated with depression severity (P = 0.35) but was associated with lesion load (P = 0.0001) and overall disability (P = 0.0007) (Supplementary Table 1).

Fig. 1: Dataset characteristics.
figure 1

a, Participant demographics. b, Distribution of MS lesions across the brain.

Full size image

Using these data, we conducted lesion network mapping, spatial correlation analyses with our a priori circuit, permutation testing and clinical outcome prediction following the same methods as in our previous work8,13. We estimated the whole-brain connectivity of each participant’s lesion location using a normative connectome database (Fig. 2a,b) and assessed the similarity of each lesion’s connectivity profile to our a priori depression circuit using spatial correlations (Fig. 2c). We compared the Fisher-transformed spatial correlations with depression scores (Fig. 2d), both of which were normally distributed (Supplementary Fig. 1), using Pearson correlation. As hypothesized, participants with MS whose lesions were more connected to our a priori depression circuit had higher depression scores, independent of the effect of age, sex, overall disability and total lesion volume (r = 0.15; P = 0.013). This relationship was specific to depression compared with 11 other Neuro-QoL metrics and overall disability (P = 0.0058; 25,000 permutations) (Fig. 2e). This relationship also survived after additionally controlling for fatigue and cognitive symptoms, which are common neuropsychiatric comorbidities of MS depression4,19 (r = 0.16; P = 0.0075).

Fig. 2: Comparing MS lesions with the a priori depression circuit.
figure 2

a, MS lesions were automatically segmented and warped to a standard brain atlas (n = 281 participants). Here we show three representative lesions. b, Lesion connectivity was estimated using a normative connectome database (n = 1,000). Normative connectivity maps of the three representative lesions are shown here. c, Each lesion’s connectivity profile was compared with our a priori depression circuit using spatial correlations. d, These spatial correlations were compared with depression severity and 12 comparator symptoms, including the Expanded Disability Status Scale and the remaining 11 Neuro-QoL symptoms, using Pearson correlations. e, Lesion connectivity to our a priori depression circuit was more associated with depression than with other symptoms (n = 281 participants and 12 symptoms; P = 0.0058 on permutation testing with 25,000 permutations). The asterisk denotes that association with depression was significantly stronger than other metrics.

Full size image

Next, we derived a brain network for MS depression in a data-driven fashion by comparing lesion connectivity profile with depression severity across all participants. The resulting circuit map represents the connectivity of MS lesion locations that are associated with greater depression severity. To investigate the full topography of this circuit, we intentionally did not apply a statistical threshold, following our previous work6,8,20. The topography of our MS circuit (Fig. 3a) showed high spatial correlation with the topography of our convergent a priori depression circuit (spatial r = 0.63) (Fig. 2c), and permutation testing confirmed that this relationship was stronger than chance (P = 0.015; 25,000 permutations).

Fig. 3: Convergence across lesion etiologies associated with depression.
figure 3

a,b, Our data-driven circuit derived from MS lesions (n = 281 participants) (a) was compared with a previously published depression circuit derived from other lesion etiologies, including stroke and trauma (n = 461 participants across five datasets) (b). Permutation testing confirmed that these maps were more similar than expected by chance (spatial r = 0.51; P = 0.04 on permutation testing with 25,000 permutations). Areas of similarity (white circles) included (1) the medial temporal lobe, (2) the medial prefrontal cortex, (3) the retrosplenial cortex and (4) the intraparietal sulcus. Areas of difference (black arrows) included (5) the sensorimotor cortex, (6) the visual cortex and (7) the posterior lateral prefrontal cortex.

Full size image

We also compared the topography of our MS depression circuit (Fig. 3a) with the topography of each individual component of our a priori depression circuit, which was derived from lesions, TMS sites and DBS sites. The MS depression circuit aligns well with the depression circuit that was derived from 461 stroke and penetrating head trauma lesions (Fig. 3b) (r = 0.51; P = 0.04; 25,000 permutations). Our MS depression circuit (Fig. 4a) also aligns with the brain circuit for therapeutic improvement in depression symptoms derived from 151 TMS sites in people with major depressive disorder (MDD) (Fig. 4b) (r = 0.62; P = 0.009; 25,000 permutations). Finally, our MS depression circuit aligns with the brain circuit derived from 101 DBS sites in people who were treated for MDD, Parkinson disease or epilepsy, some of which induced greater changes in depression than others (Fig. 4c) (r = 0.68; P = 0.002; 25,000 permutations).

Fig. 4: Convergence between MS depression and brain stimulation sites that modulate depression.
figure 4

a,b, Our data-driven circuit derived from MS lesions (n = 281 participants) (a) showed similar topography to our previously published depression circuit derived from TMS sites (n = 151 participants across four datasets) that reduce depression severity (b) (spatial r = 0.62; P = 0.009 on permutation testing with 25,000 permutations) and our previously published depression circuit derived from DBS sites (n = 101 participants across five datasets) that modify depression severity (c) (spatial r = 0.68; P = 0.002 on permutation testing with 25,000 permutations).

Full size image

Finally, we searched for the connection in the brain most strongly associated with MS depression. We found that negative connectivity between lesion locations and the ventral midbrain, including the ventral tegmental area (VTA), was the strongest association with MS depression, independent of the effect of age, sex, lesion volume and overall disability (Fig. 5a) (P value corrected for familywise error (PFWE) < 0.05; peak t = −3.80; cluster size = 1,016 mm3). This relationship was not driven by fatigue or cognitive symptoms as it was even stronger after controlling for the fatigue and cognitive subscales of the Neuro-QoL (Fig. 5b) (PFWE < 0.05; peak t = −4.61; cluster size = 3,536 mm3) (Fig. 5b). Both of these voxel clusters were substantially larger than the minimum detection threshold of 240 mm3 at P < 0.001. These results were driven primarily by greater depression in people whose lesions were negatively connected to the VTA, not by decreased depression in people whose lesions were positively connected to the VTA (Supplementary Fig. 2).

Fig. 5: The peak of our data-driven MS depression circuit is in the ventral midbrain.
figure 5

a, The peak association in the MS depression circuit was a negative correlation between depression severity and lesion connectivity (green) to the ventral tegmental area (violet) (peak t = −3.80; cluster size = 1,016 mm3). b, This association was even stronger after controlling for age, sex, lesion volume, overall disability, fatigue and cognitive symptoms (peak t = −4.61; cluster size = 3,536 mm3). The VTA was defined as in ref. 53 (freely available via Lead-DBS Toolbox54). P, posterior.

Full size image

Of note, these results were nearly identical when repeating the analysis after excluding the two participants who were not scanned on a Siemens Skyra for clinical reasons (Supplementary Table 2).

Discussion

Lesion-based localization of MS depression has yielded inconsistent results, leading to lack of consensus regarding whether depression is related to lesion location in MS4. To our knowledge, this is the largest study to compare lesion locations to depression severity in people with MS21. Using lesion network mapping, we found that MS lesions associated with depression are preferentially connected to the same circuit as stroke lesions, TMS sites and DBS sites that modify depression severity6. This demonstrates not only that MS depression is associated with lesion location, but also that MS depression may share some neuroanatomical features with other depression etiologies and therapeutic neuromodulation sites.

MS depression localized to the same a priori circuit as multiple different lesion types associated with depression, including ischaemic stroke, haemorrhagic stroke and penetrating head trauma. Although the a priori circuit was derived primarily from grey matter lesions, the present work shows that this circuit also applies to white matter lesions. This result may seem surprising as functional MRI (fMRI) fluctuations have relatively low signal-to-noise ratio in white matter15,16,17. We observed positive results despite this limitation, possibly because the large sample size of our connectome database (n = 1,000) improved our signal-to-noise ratio. This allows us to see functional connectivity with white matter voxels, similar to previous work using this same connectome21. Our results are also consistent with accumulating evidence suggesting that resting-state functional connectivity with white matter voxels can identify relevant network neuroanatomy15,16,17. Our data provide further evidence for a convergent brain circuit that is causally associated with depression across different diseases. However, independent of this convergence, each disease entity is probably also associated with some unique patterns of circuitry as several regions in our MS depression map were different from our a priori depression circuit. Furthermore, MS depression may uniquely be associated with immune-mediated or other systemic factors4,22.

MS depression also localized to a circuit that was similar to the circuit derived from other lesions and therapeutic brain stimulation sites that modify depression severity. This cross-modality convergence strengthens causal inference8 as it minimizes the possibility of reverse causality. Specifically, it is possible that our circuit is driven by depression-induced lesions rather than lesion-induced depression as depression may modify the pathophysiology of MS and the probability of lesions occurring in specific locations. However, this same depression circuit is apparent in data types in which this argument would be implausible. For example, a depression circuit with similar topography can be derived from penetrating brain lesions (the location of which is determined mostly by the direction from which the trauma came), different TMS sites (the location of which is determined mostly by scalp dimensions) and different DBS sites (the location of which is determined mostly by the surgeon’s approach)6. While each of these factors may follow a non-random distribution that is influenced by premorbid depression, it is highly unlikely that they all converge on a similar non-random distribution. This illustrates how convergence across different modalities, methods and disorders can increase the strength of causal inference as the confounders from one approach can be mitigated by a different approach8.

In addition to strengthening causal inference, this cross-modality convergence also suggests that MS depression may respond to the same TMS targets as primary MDD and post-stroke depression, both of which are supported by multiple randomized clinical trials23,24. This is consistent with recent findings that lesion connectivity can help identify therapeutic targets for various neuropsychiatric disorders, including MDD6, Parkinson disease6,25 and Tourette syndrome26. Future clinical trials should thus specifically investigate the antidepressant efficacy of TMS in people with MS, which has not yet been thoroughly studied27. Furthermore, future work can test whether this principle may be extended to identify circuit-based TMS targets for other symptoms that are commonly found in people with MS.

The peak of our data-driven MS depression circuit was in the ventral midbrain, including the VTA. This brainstem nucleus is the source of dopamine in the mesolimbic reward system, which is believed to be affected in some people with depression28. The implication of the VTA is consistent with multiple studies showing that reward processing and dopaminergic neurotransmission are affected in people with MS29,30,31. Several antidopaminergic and dopaminergic drugs are effective for unipolar or bipolar depression, such as aripiprazole32, lurasidone33,34, bupropion35 and selegiline36. However, despite the modest success rate of serotonergic antidepressants in MS4,9, to our knowledge there are no clinical trials of dopamine-modulating drugs for MS depression9. Although highly speculative, our results suggest that these medications might be worth exploring in MS depression.

Of note, our study relied on the Neuro-QoL, which does not categorically distinguish between presence or absence of major depression. Rather, it is a continuous measure that quantifies variance in mood across the full distribution of patients. We used the term ‘depression’ in a manner that is consistent with its use in the Neuro-QoL and in the MS literature37,38,39,40,41,42,43 but is not synonymous with a diagnosis of MDD. One advantage to using a continuous outcome is that our results are sensitive to lesions associated with either higher or lower mood. This raises the intriguing possibility that some MS lesion locations could protect against depression, a concept that has a strong precedent in the stroke literature13. The present study was designed only to localize the neuroanatomy, but future work may further disentangle how different effects on this circuit may have different effects on depression.

Strengths of this study include independent replication of an a priori circuit using causal sources of information8 in a unique patient population. There are also several limitations, many of which could lead us to underestimate our effect size due to increased sampling error. First, our analyses were based on a single depression score, which may not capture the full breadth of different clinical presentations. Future studies with more detailed psychiatric phenotyping may reveal distinct circuits for distinct symptom clusters20. Second, there remains active debate about how effectively resting-state fMRI can be used to measure white matter signal. This probably introduced noise into the analysis, although some of this noise can be overcome with large sample sizes as in the connectome database that we employed15,16,17. Third, we used a normative database to estimate lesion connectivity, consistent with the methods used to derive our a priori depression circuit6. Participant-specific connectivity measurements may explain additional variance but would add additional noise to the analysis44. Future work may also include multimodal approaches that incorporate other imaging techniques such as diffusion connectivity, but these combined approaches require further methodological validation. Fourth, our results represent only a cross-sectional view and do not consider how different lesion patterns may influence longitudinal progression or efficacy of different pharmacological interventions. Although our circuit was independent of immunomodulatory drugs on cross section, we did not have longitudinal data on therapeutic response. Furthermore, our dataset did not include information on antidepressant use. All of these limitations introduce noise that would bias our analysis in favour of a false negative. For example, if fMRI recordings from white matter were composed of pure noise (with no signal), then our permutation statistics would show no association between white matter lesion location and depression. Future studies that address these limitations may detect stronger effects than those reported here as effect sizes are inversely proportional to noise45. Because this method is not designed to precisely measure the effect size, it is of limited utility for prognostication.

Overall, our results demonstrate that MS lesion locations associated with depression are connected to a specific brain circuit, with a peak in the ventral midbrain. This circuit topography aligns with previous depression circuits derived from other lesion etiologies or brain stimulation sites. These results may be useful for identification of pharmacological and neuromodulatory treatment targets.

Methods

Dataset characteristics

The SysteMS study included neuroimaging and behavioural data collected for the purpose of identifying factors associated with disease severity and progression in MS. SysteMS participants were enroled between November 2015 and September 2017 if they met McDonald criteria (2010 revision46) for MS or clinically isolated syndrome. This study was approved by the Partners Human Research Committee, and the study complied with all relevant ethical regulations. All study investigators were authorized by the Human Research Committee before accessing identifiable information. All participants signed written informed consent for identifiable information to be accessible to investigators at Brigham and Women’s Hospital and for de-identified information to be shared with outside investigators for studies related to MS and other inflammatory diseases. No formal power analysis was conducted, but our sample size is larger than any other lesion-mapping study of depression in MS21.

Symptom severity was measured using tools and scales that have been previously validated in people with MS37,38,39,40,41,42,43,47. Most symptoms were assessed using Neuro-QoL, a self-report inventory of 12 neuropsychiatric symptom clusters, including ability, anxiety, emotional/behavioural dyscontrol, fatigue, lower-extremity function, upper-extremity function, stigma, positive affect/well-being, satisfaction with social roles/activities, cognitive function and sleep. All measures were collected using a computerized adaptive test except for sleep, which was collected using the short-form inventory. The primary outcome in this study was the Neuro-QoL depression subscale. Covariates included the Neuro-QoL fatigue subscale, the Neuro-QoL cognitive subscale and overall disability as measured by the Expanded Disability Status Scale (EDSS)47. As there is no clear consensus on how Neuro-QoL may be used to categorically distinguish between presence and absence of depression or other symptoms (Supplementary Table 3)37,38,39,40,41,42,43, all measures were treated as continuous variables.

Lesion localization

Nearly all participants (n = 279) completed structural imaging on a 3T Siemens Skyra MRI scanner. For clinical reasons, one participant was scanned on a 3T Siemens Verio, and one was scanned on a 1.5T Siemens Avanto. All Skyra scans were conducted using a 20-channel head coil to collect three-dimensional high-resolution T1 MPRAGE (TE/TR = 2.96/2,300 milliseconds; TI = 900 milliseconds; flip angle = 9°), T2 spin echo (TE/TR = 300/2,500 milliseconds; echo train length = 160) and T2 FLAIR (TE/TR = 389/5,000 milliseconds; TI = 1,800 milliseconds; echo train length = 248) with 1 mm isotropic voxels.

FLAIR hyperintense lesions were defined using an automated white matter segmentation pipeline as previously described18. This pipeline includes bias field correction; spatial co-registration between different sequences; skull stripping; anatomic parcellation/segmentation; intensity normalization; quality cross-check with T1 and T2; delineation of white matter, grey matter and CSF; expectation maximization of tissue class and white matter lesion segmentation; and post-processing to remove islands and minimize false positives. The resulting lesion segmentations were warped into standard Montreal Neurological Institute atlas space using the B-spline registration function in the Elastix Image Registration Toolbox48.

Statistical methods

Except as otherwise specified, all statistical analyses were conducted using MATLAB R2021b, using the same algorithms that were used to derive and cross validate the a priori depression circuit6. For correlations across participants, two-tailed P values were determined analytically using standard parametric methods as each participant is independent of every other participant. For spatial correlations across voxels, P values were determined using permutation testing as parametric P values would yield elevated false-positive rates due to inter-dependence of brain voxels. As in our previous work6, spatial correlations were recomputed after randomly permuting each participant’s clinical outcomes with a different participant’s neuroimaging. After 25,000 iterations, a P value was computed as the percentage of cases in which the real spatial correlation was stronger than the permuted spatial correlation. This method provides a rigorous way to control for Type I errors (false positives) due to noise introduced by physiological, clinical or methodological factors as these factors are equally applicable in the permuted sample6,20,49. However, this method increases the chance of Type II errors (false negatives), which can occur if the preceding sources of noise are large relative to the signal of interest50.

Associating lesion connectivity with depression scores specifically

We hypothesized that depression severity in MS would be correlated with each lesion’s connectivity to our a priori depression circuit. As in our previous work, connectivity was estimated on the basis of the Fisher-transformed spatial correlation of each lesion’s connectivity profile to our convergent depression circuit6. For our primary analysis, we compared these spatial correlation values with the Neuro-QoL depression scale using partial Pearson correlations, controlling for the effect of age, sex, lesion volume and overall physical disability (based on EDSS). To rule out the effect of clinical confounders and investigate specificity, we also repeated this analysis with additional covariates known to be independently associated with depression in MS4, including fatigue and cognitive symptoms as measured by Neuro-QoL.

To further assess specificity, we repeated the primary analysis for each of the other Neuro-QoL measures and for the EDSS. We hypothesized that the convergent depression circuit would be more associated with depression than with other symptoms. To quantify this, we computed the difference between the r value for depression and the mean r value across other symptoms; we refer to this difference as Δr. To test for significance, we used a permutation test in which Δr was recomputed after each participant’s clinical outcomes were randomly re-assigned to a different participant’s neuroimaging. A P value was computed as the percentage of cases in which the real Δr value exceeded the randomly permuted Δr value.

Derivation of the MS depression circuit map

Lesion network mapping was conducted following the same procedure that was previously used to derive a convergent depression circuit from stroke lesions, penetrating head trauma lesions, TMS sites and DBS sites6. First, normative connectivity of each participant’s lesion was estimated using a human connectome database that includes resting-state fMRI scans from 1,000 healthy volunteers. This yielded a map depicting the connectivity of each participant’s lesion location to every voxel in the brain. At each voxel, this connectivity value was compared with the participant’s depression severity using partial Pearson correlations, controlling for the same covariates as the preceding. This analysis yielded a whole-brain ‘MS depression circuit map’ representing the positive and negative connectivity of MS lesions associated with increasing depression severity.

As in the preceding, to rule out the effect of clinical confounders and investigate specificity, we also repeated this analysis after additionally controlling for fatigue and cognitive symptoms as measured by the Neuro-QoL.

Comparison with convergent depression circuit

We hypothesized that the MS depression circuit would be similar to the convergent depression circuit map derived from stroke lesions, penetrating head trauma lesions, TMS sites and DBS sites6. We quantified the similarity between these two maps using a spatial correlation. Because the voxels in each map are not statistically independent of one another, significance was assessed using non-parametric permutation testing as in our previous work6,20. In this permutation test, the spatial correlation was recomputed 25,000 times after randomly re-assigning each participant’s connectivity map to a different participant’s clinical variables. A P value was defined as the percentage of randomly permuted spatial correlations that were stronger than the real spatial correlation.

We also hypothesized that the MS depression circuit map would be independently similar to each of the three circuit maps derived from brain lesions, TMS sites and DBS sites, respectively. To test this hypothesis, we repeated the preceding analysis for each of these three maps independently.

Identification of peak anatomical associations

In an exploratory analysis, we localized the peak positive and negative regions in the MS depression circuit map after controlling for the relevant covariates. Relative peaks were identified using the FSL cluster tool51 with FWE-corrected threshold of P < 0.05, detection threshold of P < 0.001 and minimum cluster size of 240 mm3, following standard recommendations for the spatial resolution (2 mm isotropic) and smoothness (6 mm full-width at half-maximum) of our resting-state functional connectivity data52.

Reporting summary

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