This was a secondary analysis of MRI and clinical testing data, obtained from either the National Institutes of Mental Health Data Archive (NDA), or ConnectomeDB, the data access portal for the Human Connectome Project (HCP) (47). NDA data can be accessed through this study-specific identifier: 3090. Three studies (shown in Table 1 ) were identified from the NDA based on the availability of high-quality diffusion MRI data and substantial numbers of MDD and HC subjects: Boston Adolescent Neuroimaging of Depression and Anxiety (BANDA) (48){Hubbard, 2020 #2}; Establishing Moderators/Biosignatures of Antidepressant Response (EMBARC) (49) the contribution of aberrant anticipatory processing to spectrum depression and mania, and cognitive and emotional dysfunction in major depressive and bipolar disorders (Aberrant) (50). We created a fourth cohort from the HCP S1200 release (see Supplemental Data for subject identifiers). All imaging and clinical data were de-identified prior to being uploaded to the NDA or ConnectomeDB. All subjects consented to participating in the original studies and consented to anonymously sharing their research data. All research was conducted according to the principles of the Declaration of Helsinki.

Subjects from the NDA were clinically diagnosed with MDD prior to their participation in the original studies. Subjects from the HCP cohort were not previously diagnosed with MDD. Instead, we characterized HCP subjects using the DSM-Oriented Guide for the Achenbach System of Empirically Based Assessment (51). HCP subjects with depression T-scores above 69, the diagnostic threshold for clinical relevance, were classified as MDD subjects. We matched HC to MDD subjects within each originating study, prioritizing in order of importance: sex, age, and self-identified race. For subjects younger than 18 years, we matched age within 12 months. For subjects older than 18 years, we matched within 36 months. The average age of subjects was approximately 25 years for both the MDD group (25.4 years) and HC group (25.7 years). We matched MDD and HC subjects from the NDA studies 1:1. We leveraged the large number of healthy controls in the HCP to generate 1:3 matches (MDD: HC). In total, the combined cohorts included 266 subjects, comprising 116 MDD and 150 HC subjects.

We categorized MDD subjects based on their depression severity using established depression scales. The EMBARC and Aberrant studies used the Hamilton Depression Rating Scale (52). The BANDA study used the depression subscale from the Revised Children’s Anxiety and Depression Scale (RCADS) (53). Severity categories for the Hamilton scale were: 8-13, Mild Depression; 14-18, Moderate Depression; 19-22, Severe Depression; >22 Very Severe Depression. Severity categories for the RCADS were: 18-20, Mild Depression; 21-23, Moderate Depression; 24-27, Severe Depression; 27-30, Very Severe Depression. For the HCP cohort, we categorized subjects based on the Achenbach T-scores as: 70-77, Mild Depression; 78-85, Moderate Depression; 85-92, Severe Depression; 92-100, Very Severe Depression.

Participants from all studies were imaged using 3T whole-brain diffusion tensor imaging (DTI) and T1 (MPRAGE) protocols. Additional scan protocols are available in the studies associated with the originating data (47–50). While some subjects were scanned multiple times in a single study, we only used MRI data from the first time point. The BANDA study consisted of 86 subjects scanned at a DTI resolution of 1.5 mm isotropic. In the EMBARC study, there were 56 subjects scanned at 1.87 x 1.87 x 2.5 mm and 6 subjects scanned at 2.5 mm isotropic. The Aberrant study had 50 subjects scanned at 1.5 mm isotropic. The HCP study consisted of 68 subjects scanned at 1.25 mm isotropic. We matched the study of origin and image resolution when generating MDD: HC matches to account for differences in resolution and image acquisition site.

We processed DTI data using the FSL FMRIB Diffusion Toolbox with standard parameters (54). We performed skull stripping and brain extraction using bet2 and corrected for eddy current distortion and subject movement during imaging acquisition using eddy_cuda10.2 (55). We used dtifit to assign diffusion tensors to each voxel and bedpostx_gpu to model crossing fibers and calculate diffusion probability estimates at each voxel (56). We generated registration matrices from the MNI152_T1 space into FMRIB58_FA diffusion space, and then into native diffusion space using flirt and fnirt. These registration matrices were used to register all MNI-space regional segmentations into each subject’s native diffusion space. The Aberrant and HCP studies had MRI data scanned in both anterior-posterior (AP) and posterior-anterior (PA) directions, which allowed us to correct for susceptibility artifacts using topup. Our other source studies included AP but not PA acquisitions. We performed probabilistic tractography using probtrackx2 in each subject’s native diffusion space.

We performed automated whole-brain segmentation of each subject’s structural (T1-MPRAGE) image using the FreeSurfer toolkit (57), yielding each subject’s estimated total intracranial volume (eTIV). We measured the volumes of the caudate and putamen in native diffusion space using fslstats and normalized by the eTIV to control for inter-subject differences in head size. To assess potential tissue injury or maldevelopment, we generated radial diffusivity (RD) by averaging the second and third eigenvalues from the output of dtifit. We then extracted RD values using fslstats.

We identified striatal voxels with matrix-like or striosome-like structural connectivity using probabilistic diffusion tractography to perform connectivity-based parcellation (58). We created masks of extrastriate target regions that, based on tract-tracing studies in animals and/or diffusion MRI in humans, selectively project to either the matrix or striosome compartments (46). Specifically, we generated matrix-favoring and striosome-favoring composite “bait” masks by summing regions whose structural connectivity was biased toward either the matrix (inferior frontal gyrus pars opercularis, primary motor cortex, supplementary motor area, primary somatosensory cortex, and superior parietal cortex) or the striosome (posterior orbitofrontal cortex, anterior insula, basolateral amygdala, basal operculum, and posterior temporal fusiform cortex). We previously utilized these regions in a separate MRI dataset to map compartment-like connectivity (23). We used composite bait regions for each compartment because while injected tract tracers in animals identified compartment-favoring biases, these were not absolute: regions could still include voxels with connectivity to the other compartment (59). Similarly, a region may favor one compartment within a specific striatal somatotopic zone but lose that bias in other striatal areas (25, 46, 60). By using summed bait masks, we minimized these sources of somatotopic variability when assessing compartment-like bias. The striatal mask included the caudate and putamen but excluded the posterior half of the caudate tail, as this region contains very little striosome (61) and its narrow structure in the coronal plane reduced registration accuracy and led to partial volume effects with the surrounding white matter. We excluded the nucleus accumbens due to the absence of prior tract-tracing studies to guide the selection of matrix- and striosome-favoring bait regions. The diffusion voxels in our study were larger than the branches of the human striosome (maximum cross-sectional diameter: 1.25 mm (29, 30), so even highly biased striosome-like voxels included some matrix tissue. Highly biased matrix-like voxels, in contrast, could include only matrix tissue given that the matrix compartment is both substantially larger and likely to occur in segregated parts of the striatum. We used the terms “striosome-like” or “matrix-like” to remind readers that these are indirect and probabilistic parcellations that should not be equated with matrix and striosome identified through the gold standard for tissue-level parcellation, histochemical staining.

We performed striatal parcellation using FSL’s probtrackx2 in classification targets mode, setting each striatal voxel as a seed and the composite compartment-favoring bait region masks as targets. Probtrackx2 used the following parameters: curvature threshold=0.2; step length=0.5 mm; 2,000 steps per sample; 5,000 streamlines per seed voxel. To prevent mis-tracked interhemispheric streamlines, we parcellated each hemisphere separately, using a contra-hemispheric exclusion mask. All tractography was performed in the subject’s native DTI space.

We utilized the FSL tool proj_thresh to calculate the voxelwise probability of connecting to matrix-favoring or striosome-favoring target masks. This approach generates two superimposable striatal probability maps, one for the matrix-like and one for the striosome-like distribution, each with P=0-1. At each striatal voxel, the sum of these two probabilities was always one. Within each distribution we defined voxels with P≥0.55 as reaching bias threshold, and thus meeting criteria to be defined as matrix-like or striosome-like. Voxels between the lower limits of the compartment-like distributions (P=0.45-0.55) were defined as indeterminate. Across the full range of thresholds (0.55<P<1.0), the matrix-like and striosome-like voxels consistently replicate known compartment tissue characteristics from histological literature, such as relative abundance and spatial distribution. A threshold of P>0.55 quantifies all biased voxels for the purposes of performing group level comparisons, while a stricter threshold of P>0.87 can be used for targeting voxels with stronger compartment-like bias – and therefore, a purer reflection of the properties of one compartment. We quantified the volume of matrix-like and striosome-like voxels separately in caudate and putamen. We used fsl-cluster to measure the size of the largest matrix-like and striosome-like clusters.

In animal and human tissue, the ratio of striosome:matrix cross-sectional area is approximately 15:85 (62–64). Our MRI-based striatal parcellations, in other human datasets, produced similar ratios (23, 25, 40, 46). While analyses that include all compartment-like voxels are useful for some measures, the natural differences in size between the compartments can bias probabilistic measures, such as tractography. Since histology predicts that matrix-like volume will be approximately 5-fold larger than striosome-like volume, we created equal-volume masks from the most-biased voxels from each distribution. We set the target size for these masks at 13% of total striatal volume: below the expected lower limit for striosome-like volume, and equal to the volume 1.5 standard deviations above the mean in a normal distribution. To account for differences in diffusion MRI resolution between the studies, the ratio of compartment-like volume to striatum volume was kept constant between studies. For cases in which one compartment’s distribution was smaller than the target volume, we equalized inter-compartment volume by reducing the size of the larger compartment mask to match that of the smaller compartment mask; the volume of these high-bias matrix-like and striosome-like voxels was always equal. We extracted the cartesian location of each voxel in these high-bias masks relative to the centroid of its nucleus of origin (caudate or putamen). We measured RD within each compartment’s high-bias mask.

To assess the robustness and specificity of structural connectivity between compartment-like voxels and their corresponding bait regions, we performed inverse seed-to-target tractography between compartment-like voxels and their associated bait regions. We performed both A-to-B and B-to-A streamline tractography, first setting compartment-like voxels as seeds and bait regions as targets, then setting the composite bait region masks as seeds and compartment-like voxels as targets. We used the same probtrackx2 parameters described above for striatal parcellation (Methods 2.5), save for the use of streamline mode rather than classification targets mode. We used waypoint and avoid flags to include one compartment-specific target and exclude the other. We normalized the streamline counts to the volume of the seed mask to account for the marked size differences in bait region and compartment-like masks.

To quantify the influence of each bait region on compartment-like bias, we performed 10 subsequent rounds of probabilistic tractography, each using only nine bait regions, rotating the left-out bait region for each round. These 10 “N-1” parcellations allowed us to quantify connectivity with the omitted region and generate masks identifying the discrete areas most-influenced by each bait region. We subtracted each “N-1” parcellation from the original (all bait regions) parcellation; the differences between these parcellations quantified the influence of the left-out bait region at a voxelwise level. We thresholded these difference maps (using fslmaths) to ensure that each somatotopic zone of influence was distinct, with no overlap between other bait regions’ somatotopic zones, and that each somatotopic zone was similarly sized. We then quantified compartment-like volumes (P≥0.55) within each of these 10 somatotopic zones. Volumes were presented as a ratio of compartment-like volume differences which allowed us to represent changes in both compartments as a single data point. In zones influenced by matrix-favoring bait regions we assessed region-specific volume as V o l ( M a t r i x )   −   V o l ( S t r i o s o m e ) V o l ( M a t r i x )   +   V o l ( S t r i o s o m e ) and for zones influenced by striosome-favoring bait regions as V o l ( S t r i o s o m e )   −   V o l ( M a t r i x ) V o l ( M a t r i x )   +   V o l ( S t r i o s o m e ) .

We performed a series of ANOVAs to test for significant group differences between MDD and HC subjects in nuclei size, compartment-like volumes, and radial diffusivity. Scanner type, study of origin, and age were included as covariates of no interest to control for potential confounding effects. We corrected for multiple comparisons using the family-wise error (FWE) method of Benjamini and Hochberg (65). We defined the following families of tests for these FWE corrections: Nuclei Size (2 tests), Radial diffusivity of nuclei (2 tests), Compartment-like volumes between MDD and HC groups (4 tests), Radial diffusivity of high-bias voxels (2 tests), Compartment-like volume varying by severity in the caudate (2 tests), Somatotopic zone compartment-like volumes (10 tests).

We performed two-tailed, two-sample t-tests with unequal variance to assess several approaches to validating our striatal parcellations: the mean location of high-bias compartment masks, compartment cluster size, and streamline counts in the inverse seed-to-target analysis. We defined the following families of tests for these FWE corrections: High-bias compartment masks locations (3 tests), Compartment cluster size (2 tests), Inverse seed streamline counts (4 tests).

In addition to these family-specific t-tests, we used the FSL tool randomise to perform voxel-wise nonparametric permutation inference testing of the whole striatum to identify voxels with significant differences in compartment-like bias between MDD and HC. We performed a second iteration of randomise to test if compartment-like biases correlated with depression severity, solely within our MDD cohort. While we generated tractography in each hemisphere independently, we combined the hemispheres for each subject prior to randomise testing to reduce the number of comparisons. We performed randomise with 5,000 permutations, 2 mm variance smoothing, threshold-free cluster enhancement, and masked by the striatum. We calculated the center of gravity and size for each significant cluster from randomise using fslstats. For testing MDD vs. HC group level differences and MDD severity effects, we included study-of-origin, scanner, and age as covariates of no interest within randomise.

MDD subjects exhibited a lower total intracranial volume (eTIV) relative to healthy control subjects (MDD: 1381 cm³ vs. HC: 1438 cm³; F_1,531, p = 0.048). We then compared volumes of the caudate and putamen, normalized to eTIV, between MDD and HC subjects. The normalized putamen volume in MDD subjects was 3.7% smaller (F_1,531, p = 3.4x10^-4), a pattern that was also observed in raw putamen volume (MDD: 3857 cm³ vs. HC: 4207 cm³, 8.3% smaller; F_1,531, p< 1.0x10^-6). In contrast, normalized volume differences in the caudate were nonsignificant between groups, showing a 5.6% increase in MDD (F_1,531, p = 0.089). Raw caudate volumes also exhibited no significant difference between groups (MDD: 3235 mm³ vs. HC: 3246 mm³; F_1,531, p = 0.73). We compared radial diffusivity (RD) between MDD and HC subjects for the caudate and putamen. RD was increased by 4.7% in the MDD putamen but only trended towards significance (MDD: 5.1x10^–4 vs. HC: 4.9x10^-4; F_1,531, p = 0.027) after multiple comparisons correction. Caudate RD differences were nonsignificant (MDD: 6.2x10^–4 vs. HC: 5.8x10^-4; F_1,531, p = 0.39).

In histologic sections, the striosome appears as separated “islands” while the matrix is a contiguous structure (27, 30). If connectivity-based striatal parcellations truly approximate the organization of matrix and striosome, compartment-like voxels should follow this schema (contiguous vs. separated). We used fsl-cluster to assess the volumes of the largest matrix-like and striosome-like clusters at a compartment-like probability bias of P>0.55. Matrix-like voxels were far more likely to occur in large, contiguous clusters; the mean volume of the largest cluster of matrix-like voxels was 571 mm³, while the largest striosome-like cluster was 162 mm³, a 3.5-fold difference (p = 6.3x10^-101). This indicates that matrix-like and striosome-like voxels followed the same organizational schema (contiguity vs. separation) seen in matrix and striosome tissue.

To demonstrate the similarities between compartment-like voxels and compartment characteristics derived from histology, we extracted the location of each voxel in our high-bias compartment masks relative to the centroid of its corresponding nucleus (caudate or putamen). Striosome-like voxels were, on average, located within 1 mm of matrix-like voxels in the medial-lateral plane, but were positioned 5.5 mm more rostral (p<1.0x10^-100) and 5.6 mm more ventral (p<1.0x10^-100) than matrix-like voxels. The mean location of matrix-like and striosome-like voxels mirrored the compartment location biases found in human histology (27, 30).

Our striatal parcellations were executed in classification targets mode, which quantified connectivity between individual striatal voxels and the bait region targets. As a further validation of the connectivity biases underpinning these parcellations, we carried out streamline tractography with compartment-like voxels as seeds and the compartment-favoring bait masks as targets. We then inverted seed and target (composite bait region masks as seeds, compartment-like voxels as targets), allowing us to assess bias in two modes (A-to-B, and B-to-A). Streamlines-per-seed followed the biases predicted from animal tract tracing studies (46) and by our connectivity-based parcellations. Normalized streamline counts were significantly larger when the compartments were matched (striosome-to-striosome, matrix-to-matrix) than when they targeted the opposing compartment ( Figure 1 ). This was true with both striatal compartment-like voxels as seeds, and when we set compartment-favoring bait regions as seeds. Normalized streamline counts were significantly different when seeds and targets were matched vs. unmatched. Specifically, matrix-favoring bait regions routed more streamlines to matrix-like voxels (995) than to striosome-like voxels (171; p = 4.2x10^-35). Striosome-favoring bait regions similarly projected more to striosome-like voxels (283) than to matrix-like voxels (25.2; p = 4.1x10^-37). When using compartment-like voxels as seeds, matrix-like voxels sent more streamlines to matrix-favoring bait regions (246) than to striosome-favoring bait regions (3.92; p = 4.1x10^-37). Striosome-like voxels sent more streamlines to striosome-favoring bait regions (48.3) than to matrix-favoring bait regions (48.3; p = 8.0x10^-29).

We compared compartment-like volumes in the caudate and putamen, normalized by eTIV, between MDD and HC groups ( Figure 2 ). In the MDD putamen, normalized matrix-like volume was 8.2% smaller (F_1,531, p = 2.25x10^-8) while normalized striosome-like volume was 15% larger (F_1,531, p = 9.0x10^-3). Raw volume measures in the putamen also exhibited this shift from matrix-like to striosome-like volume: Matrix-like (MDD: 2785 mm³ vs. HC: 3197 mm³; F_1,531, p = 3.9x10^-9); Striosome-like (MDD: 857 mm³ vs. HC: 757 mm³; F_1,531, p = 0.021).

In the caudate, there were no significant differences in normalized compartment-like volume between MDD and HC subjects. In MDD, normalized matrix-like volume in the caudate was 8.7% larger (F_1,531, p = 0.31) while normalized striosome-like volume was 1.7% larger (F_1,531, p = 0.42). Raw volume differences between groups were also nonsignificant in both matrix-like and striosome-like voxels: Matrix-like (MDD: 2075 mm³ vs. HC: 2025 mm³; F_1,531, p = 0.55); Striosome-like (MDD: 2075 mm³ vs. HC: 2025 mm³; F_1,531, p = 0.55).

We analyzed RD within our high-bias compartment-like masks to investigate whether abnormalities in compartment-like volume might be attributable to compartment-specific injury or maldevelopment. When we removed the influence of age, sex, and scanner type, we found no significant differences in RD between MDD and HC groups in either matrix-like or striosome-like voxels: Matrix-like (MDD: 5.5x10^–4 vs. HC: 5.0x10^-4; F_1,531, p = 0.65); Striosome-like (MDD: 6.0x10^–4 vs. HC: 5.3x10^-4; F_1,531, p = 0.14).

We identified clusters of voxels with significant differences between MDD and HC subjects in compartment-like bias at a voxelwise level using FSL’s nonparametric inference testing utility, randomise ( Figure 3 ). Because each voxel’s compartment-like biases always sum to one, changes in the amplitude of matrix-like bias are always countered by a comparable shift in striosome-like bias. Therefore, we included the full probability distribution (P=0-1) of the matrix-like striatal probability maps, thereby assessing both compartments in one iteration of randomise.

In the bilateral rostral putamen, the connectivity bias in MDD subjects shifted from matrix-like toward striosome-like bias. The volume of putaminal voxels with significantly shifted bias in MDD was 236 mm³ in the left hemisphere and 683 mm³ in the right hemisphere. The left and right putaminal clusters had highly symmetrical centers of gravity (COG); the MNI coordinates (x-, y-, z-planes) for the left significant cluster were (11, 21, 5) while those for the right were (11, 22, 4).

In the caudate, the left and right hemispheres exhibited opposing abnormalities in compartment-like bias. The right hemisphere in MDD had a large cluster of voxels in which bias was shifted from striosome-like toward matrix-like (1834 mm³), while the left hemisphere had a small volume in which voxels shifted from matrix-like toward striosome-like bias (116 mm³). These significant clusters were located similarly in the x and y planes but differed substantially in the z (axial) plane: left COG (10, 17, 1) vs. right COG (11, 11, 13). Notably, these interhemispheric differences in bias may explain why our measures of caudate compartment-like volume (Section 3.5) did not differ between MDD and HC.

We used FSL randomise to test if compartment-like connectivity varied with depression severity ( Figure 4 ). We identified a region in the right caudate where matrix-like bias was significantly correlated with depression severity (higher matrix-like volume in subjects with higher standardized depression scores). Both matrix-like (F_1,531, p = 0.043) and striosome-like (F_1,531, p = 4.9x10^-4) volumes in the caudate were significantly correlated with MDD severity. Compartment-like volume in the putamen ( Figures 2 , 3 ) did not correlate with MDD severity. We performed a post-hoc analysis, extracting compartment-like volumes in the caudate for each of our severity subgroups to identify the drivers of this group-level effect ( Figure 5 ).

We compared eTIV-normalized compartment-like volumes in the whole-caudate between a combined mild/moderate severity group and a combined severe/very severe group. In the caudate, the severe/very severe group showed a 26% increase in matrix-like volume (p = 1.2x10^-3) and a 29% decrease in striosome-like volume (p = 1.1x10^-3) relative to the mild/moderate group. Raw compartment-like volume measures also exhibited this difference: Matrix (Mild/Moderate: 1936 mm³ vs. Severe/Very Severe: 2338 mm³; p = 1.4x10^-3); Striosome (Mild/Moderate: 894 mm³ vs. Severe/Very Severe: 611 mm³; p = 8.0x10^-5).

We mapped the locations where each of our 10 bait regions had the greatest influence on compartment-like bias. To identify these somatotopic zones, we parcellated the striatum with nine out of our ten bait regions, leaving one bait region out. We then subtracted this “N-1” parcellation from the original striatal parcellation (which was based on all ten bait regions) to identify the influence of the left-out bait region. Although we parcellated the left and right hemispheres separately, the somatotopic zones attributable to each bait region were highly similar in size and location between hemispheres ( Figure 6 ). Each somatotopic zone was spatially distinct with no overlap between zones. Notably, the somatotopic zones in these subjects are highly similar to the somatotopic zones we previously described in other cohorts (25, 40, 46).

We measured matrix-like and striosome-like volume (P≥0.55) within each somatotopic zone ( Figure 7 ). We expressed these volume measures as a normalized ratio of volume change to capture changes in both compartments as a single value: ( V o l ( D o m i n a n t )   −   V o l ( N o n d o m i n a n t ) V o l ( D o m i n a n t )   +   V o l ( N o n d o m i n a n t ) ) . The dominant compartment refers to matrix-like volume for matrix-favoring bait regions and striosome-like volume for striosome-favoring bait regions.

In MDD, there were significant increases in matrix-like connectivity in the somatotopic zones dominated by the posterior orbitofrontal cortex (F_1,531, p = 2.0x10^-4) and basal operculum (F_1,531, p = 2.6x10^-4). Notably, these particular zones occupied the same area of rostral putamen in which our voxelwise assessment identified a shift toward striosome-like bias in MDD (compare Figures 3 and 6 ). It is important to note that each bait region also contributed to compartment-like bias in areas outside these somatotopic zones. However, these zones represent the area where a bait region’s contributions were stronger than those of all other regions.