The organization of frontostriatal brain wiring in non-affective early psychosis compared with healthy subjects using a novel diffusion imaging fiber cluster analysis

Alterations in brain connectivity may underlie neuropsychiatric conditions such as schizophrenia. We here assessed the degree of convergence of frontostriatal fiber projections in 56 young adult healthy controls (HCs) and 108 matched Early Psychosis-Non-Affective patients (EP-NAs) using our novel fiber cluster analysis of whole brain diffusion magnetic resonance imaging tractography. Using whole brain tractography and our fiber clustering methodology on harmonized diffusion magnetic resonance imaging data from the Human Connectome Project for Early Psychosis we identified 17 white matter fiber clusters that connect frontal cortex (FCtx) and caudate (Cd) per hemisphere in each group. To quantify the degree of convergence and, hence, topographical relationship of these fiber clusters, we measured the inter-cluster mean distances between the endpoints of the fiber clusters at the level of the FCtx and of the Cd, respectively. We found (1) in both groups, bilaterally, a non-linear relationship, yielding convex curves, between FCtx and Cd distances for FCtx-Cd connecting fiber clusters, driven by a cluster projecting from inferior frontal gyrus; however, in the right hemisphere, the convex curve was more flattened in EP-NAs; (2) that cluster pairs in the right (p = 0.03), but not left (p = 0.13), hemisphere were significantly more convergent in HCs vs EP-NAs; (3) in both groups, bilaterally, similar clusters projected significantly convergently to the Cd; and, (4) a significant group by fiber cluster pair interaction for 2 right hemisphere fiber clusters (numbers 5, 11; p = .00023; p = .00023) originating in selective PFC subregions. In both groups, we found the FCtx-Cd wiring pattern deviated from a strictly topographic relationship and that similar clusters projected significantly more convergently to the Cd. Interestingly, we also found a significantly more convergent pattern of connectivity in HCs in the right hemisphere and that 2 clusters from PFC subregions in the right hemisphere significantly differed in their pattern of connectivity between groups.


INTRODUCTION
In this study, we assess frontostriatal structural connectivity (i.e., brain wiring) in early psychosis-non-affective (EP-NA) subjects.Frontostriatal connectivity is the initial component of frontrostriatal circuitry, which modulates functions that are abnormal in schizophrenia such as executive functions [1].Diffusion magnetic resonance imaging (dMRI) tractography [2][3][4][5] allows for an in vivo method to measure the local variation in brain connectivity patterns in human subjects.Here, we have applied dMRI tractography in a novel manner to investigate putative structural connectivity disturbances (brain miswiring) in frontostriatal connections in early psychosis patients.We have previously published normative data about brain wiring in frontostriatal connectivity [6] using this approach and here apply it in early psychosis subjects where the confounds of medication and chronicity are relatively small.In addition, as brain wiring is formed during early development [7], the presence of its altered pattern in adult subjects lends support to developmental hypotheses of schizophrenia [8][9][10][11].Further, measures of brain miswiring in schizophrenia could serve as developmental biological markers to identify those early psychosis patients who might benefit from early intervention.Schizophrenia (SZ) is a psychotic disorder which manifests both positive and negative symptoms such as hallucinations, delusions and avolition, as well as prominent cognitive abnormalities including impaired executive function.It is thought to be associated with disrupted white matter connectivity between brain regions [12,13].Based on the evidence of abnormal brain functional connectivity in SZ [14,15], it is important also to examine brain structural connectivity.The primary focus of this paper, the frontostriatal circuit, modulates higher cognitive functions and behaviors, including executive function and goal directed behavior, which are core clinical deficits in schizophrenia [1,16,17].
The striatum is the main input nucleus of the basal ganglia for cortical projections and channels feedback to the cortex, via the thalamus, through the output nuclei of the basal ganglia [18,19].A topographic, tripartite, functional organization of the striatum with limbic, associative and sensorimotor subloops has been described [20].Corticostriatal projections, which contain frontostriatal circuits, have a generally topographic, functionally segregated, anatomic organization [18].However, non-topographic, functionally integrative cortical projections, which anatomically overlap at the level of the striatum, also exist [19,21].Animal tract tracing and human imaging studies [19,[21][22][23] both support the existence of corticostriatal projection patterns yielding integrative and segregated corticostriatal target zones.The function of striatal target zones that receive overlapping cortical projections is to allow for integration of information from different cortical functional subregions [19,24].
Prior studies have used dMRI to assess frontostriatal white matter pathways in humans with developmental brain disorders.Such studies have shown reduced fractional anisotropy (FA) in attention deficit hyperactivity disorder (ADHD) in right frontostriatal projections [25,26], increased fiber density in fragile X syndrome in left ventral frontostriatal tracts [27] and increased FA with decreased RD in multiple frontostriatal tracts in 22q11.2deletion syndrome [28].In a prior study in first-episode schizophrenia patients using dMRI tractography, we showed a reduction in FA in several frontostriatal tracts, and in one of these tracts, a reduction in FA was correlated with impaired executive function [29].In addition, in our prior studies of chronic schizophrenia, using dMRI tractography to assess diffusion measures and connectivity wiring patterns, we showed reduced FA as well as reduced streamline counts in frontostriatal circuitry [17].In a subsequent dMRI tractography study, using a different approach to evaluate brain wiring, we assessed the degree of overlap in frontal projections to the striatum and showed a less integrative pattern of connectivity in chronic schizophrenia vs healthy controls (HCs) [30].
Using our current fiber clustering methodology in a prior study [6], we assessed HC subjects to establish a normal wiring pattern which can later be contrasted with other neuropsychiatric conditions.We showed that the normative PFC wiring pattern projecting from the PFC to the caudate, deviates from a strictly topographic, parallel, organization, due to a pattern of convergence in regionally specific anatomic fiber clusters connecting the PFC and striatum.We found that these clusters originated in subregions in the ventrolateral, dorsolateral and orbitofrontal PFC.We are interested in quantifying the degree of convergence of efferent projections to the striatum from the cortex as such anatomic convergence allows for the functional integration of information between different cortical subregions at the level of the striatum.Stated otherwise, such convergence allows for greater circuit integration, greater cross-talk between cortical subregions and greater synchrony of systems, serving as potential connection hubs at the level of the striatum.(e.g., see Averbeck et al. [21].) In the current study, we compared the frontostriatal wiring pattern in EP-NA subjects with that of HC subjects.We employed our novel way of using tractography using a fiber clustering method [31].This approach utilizes machine learning to generate clusters of streamlines, or fibers, based on the similarity of their trajectories.With this approach, we identified and analyzed all fiber clusters connecting regions of interest (ROIs) in the frontal cortex and the caudate for their degree of geometric convergence.More specifically, to assess the geometry of the input from multiple fiber clusters, we measured the mean distance between the endpoints of streamlines of fiber clusters at the level of frontal cortex (i.e., cortical distance) and the mean distance between the endpoints of the corresponding streamlines of fiber clusters at the level of the caudate (i.e., caudate distance; see Fig. 1).We consider our measures to reflect frontostriatal connectivity, but we note that the target zone for our projections from the frontal cortex is the dorsal striatum and more specifically the caudate nucleus.The primary focus of our approach is on the associative striatum as we and others have shown that the caudate nucleus is a brain region showing important group differences between schizophrenia and healthy controls.See [17,32,33].In addition, technically, we found it simpler to exclude the putamen, even though the anterior part of which is part of the associative striatum see [34], as a target zone for our efferent projecting streamlines from the cortex, for the reason described below in our Methods, fiber clustering, cluster analysis section.
Based on animal work [21] and our prior human work [6,30], we hypothesized (1) frontostriatal connectivity in both groups would show a pattern of deviation from a strictly anatomically topographic arrangement with selective PFC subregions showing greater convergence; (2) certain clusters in both groups would show greater patterns of convergence coming from subregions of the frontal cortex, which we surmise may subserve functions that benefit from greater circuit integration; and, (3) HCs would show a more convergent Frontal Cortex (FCtx)-caudate connectivity pattern compared to EP-NA subjects.

METHODS Subjects
This study was comprised of data from 108 young adult EP-NA subjects and 56 HCs from the Human Connectome Project for Early Psychosis (HCP-EP) (MPI: Shenton, Breier).Subjects were recruited from 4 HCP-EP sites (Indiana University, Massachusetts General Hospital, McLean Hospital and Beth Israel Deaconess Medical Center-Massachusetts Mental Health Center).MRI data were acquired from 3 HCP-EP sites, 1 in Indiana, 1 at Brigham and Women's Hospital (with subjects from McLean, Massachusetts General Hospital, and Beth Israel Deaconess Medical Center-Massachusetts Mental Health Center) and later MRI data were also acquired at McLean Hospital.The study was approved by the Brigham and Women's Hospital IRB and the VA Boston healthcare system R&D.Informed consent was obtained from all subjects in the study.The unprocessed brain imaging data and behavioral data are in a publicly available data set from the HCP-EP Project stored at The National Institute of Mental Health Data Archive (NDA).

Image acquisition and preprocessing
The HCP-EP dMRI data were acquired using a modified HCP-Lifespan protocol to cut down on the overall time in the scanner (https:// www.ncbi.nlm.nih.gov/pmc/articles/PMC6649668/)A 3T Siemens Prisma Fig. 1 A model of our method for obtaining cortical and caudate mean distances between the endpoints of streamlines of two frontostriatal fiber clusters.Adapted from Levitt et al. [6].
scanner was used and images were processed using a well-designed pipeline (https://github.com/pnlbwh)that included motion correction, eddy current correction and EPI distortion correction.The acquisition parameters were: TE = 89.2ms, TR = 3230 ms, partial Fourier of 6/8, multiband factor of 4 and voxel size = 1.5 × 1.5 × 1.5 mm 3 .Data were acquired along 206 gradients in both AP and PA phase encoding directions at b values of b = 500, 1500 and 3000 s/mm 2 with 6 gradient directions at b = 500 s/mm 2 and 100 gradients along each of b = 1500 and 3000.

Structural image post-processing
Diffusion imaging data harmonization.The dMRI data were harmonized using our harmonization methodology [35] (https://github.com/pnlbwh/dMRIharmonization).Briefly, data were acquired from four different sites, and dMRI data were acquired on a 3T Siemens Prisma scanner from two (Brigham and Women's Hospital, Indiana University) and then a third scanner was added (McLean Hospital).Control subjects from each site were used to learn a mapping between the sites, with the BWH used as the reference site.Data from all other sites were then harmonized to the BWH site.As shown in [35], post-harmonization, the statistical differences due to scanner effects were removed.We note that despite the use of HC subjects to learn the mapping between sites for harmonization, we have previously demonstrated differences between the groups are preserved.We have shown that harmonization does not change the effect sizes or differences between the groups.For example, we performed several different experiments that demonstrate that our method preserves group differences as described in the following references [35][36][37].

Diffusion imaging two-tensor whole brain tractography post-processing
In the harmonized dMRI data, whole brain tractography was computed using a two-tensor model Unscented Kalman Filter (UKF) method [38,39] that accounts for crossing fibers, as implemented in the ukftractography package (https://github.com/pnlbwh/ukftractography).A two-tensor model was chosen to account for the many crossing fibers [40,41] found in the architecturally complex white matter tracts.We used the UKF tractography method as it has been shown to handle crossing fibers much better than many tractography methods.In particular, the method was chosen as one of the winners of the Fiber Cup Tractography challenge (community challenge).For a review of all the competing algorithms see [42].We also note that the UKF method was used in several of our clustering algorithms and had very high reproducibility [31,43].The two-tensor model associates the first tensor with the main direction of the fiber tract that is being traced, while the second tensor represents fibers crossing through the tract.The UKF method is highly consistent in tracking fibers using dMRI data from independently acquired populations across ages, health conditions and image acquisitions [31], and it is more sensitive than standard single-tensor tractography [44][45][46].Visual and quantitative quality control of the tractography was performed using the quality control tool in the whitematteranalysis (WMA) software (https://github.com/SlicerDMRI/whitematteranalysis).We have previously shown that our multi-tensor model detects fiber crossings at a very small angle (e.g., less than 30 degrees) quite robustly across datasets [38].This method has been used by several groups across the world with open-source code available at https://github.com/pnlbwh/ukftractographyand it is also widely distributed along with the popular 3D Slicer software (http://www.slicer.org).The UKF Tractography method uses correlation in diffusion along the fiber bundles to robustly estimate parameters of the multi-tensor model and has shown excellent performance at b = 3000 despite higher noise.Several comparisons and reproducibility of the method is shown in the following publications: [31,38,39,42,44,47].We extracted the b = 3000 shell of 92 gradient directions and collected 18 b = 0 images, 9 in the AP and 9 in the PA direction interspersed during the scan, in scans for each subject, as applied in our previous studies that perform tractography-based analysis using HCP data [31,43,47,48].Angular resolution is better and more accurate at high b-values such as 3000 [49,50] and this single shell was chosen for reasonable computation time and memory use when performing tractography.We here use only a single b = 3000 shell as it provides better angular precision in separating the fiber bundles as shown in prior studies [51,52].Further, including the use of multiple shells (e.g., using two shells) involves using twice as many gradient directions which makes the UKF tractography method slower.Thus, to reduce computational load without sacrificing the angular precision to detect fibers, we have used only the b = 3000 shell when performing tractography.
Specifically, tractography was seeded in all voxels within the brain mask where FA was greater than 0.1.For deterministic tractography, the "bruteforce" or "holistic" approach that seeds tractography in the entire brain and then performs a selection of tracts of interest has been shown to be more robust and reproducible than seeding from specific ROIs [53][54][55][56].
Tracking stopped where the FA value fell below 0.08 or the normalized mean signal (the sum of the normalized signal across all gradient directions which was employed to robustly distinguish between white/ gray matter and cerebrospinal fluid) fell below 0.06, as recommended for HCP data in [31].Fibers that were longer than 40 mm were retained to avoid any bias towards implausible short fibers [57][58][59].For each of the subjects under study, there were about 700,000 fibers in the whole brain tractography [31].
A deterministic algorithm was used to generate our tractography data.In addition, we clarify that we did not perform additional data preprocessing such as spherical-deconvolution informed filtering of tractograms (SIFT) [60].The tractography data was parcellated according to our atlas where the curation of biologically meaningful white matter tracts in our atlas is performed by a neuroanatomist [31].As this process does not count number of streamlines, applying methods such as SIFT are not expected to have a major effect on the results.However, we acknowledge that applying SIFT might improve the anatomical tract identification result and would be of interest for a potential future investigation.As for the total number of streamlines in the tractography data, we note that 700,000 fibers are sufficient to generate a good whole brain white matter coverage based on several of our previous studies [61,62].

Fiber clustering methods
Cluster analysis.To enable the identification of fiber tract parcels (i.e., fiber clusters) from orbital, lateral and medial prefrontal cortical regions projecting to the caudate, we used a data-driven fiber clustering atlas [31].This atlas allows for a whole brain tractography parcellation into 2000 fiber clusters according to the white matter anatomy (i.e., fiber geometric trajectory).The clusters themselves were comprised of fibers, or streamlines, that reflect the trajectory of white matter axons (See Fig. 1).
In brief, the atlas was generated by creating dense tractography maps of 100 individual HCP subjects (an independent population from the HCP subjects used in the present study) and then applying a fiber clustering method to group the tracts across subjects according to their similarity in shape and location.For each cluster in the atlas, the tract anatomical profile (TAP) [31], i.e. the set of segmented brain FreeSurfer regions through which the cluster passed, is provided.The TAP was calculated based on the 100 HCP-atlas subjects, as described in [31].Briefly, for each cluster, the set of intersected FreeSurfer regions per atlas subject was computed.Then, the set of regions (here, a given ROI in the PFC and the caudate) intersected by at least 40% of fiber streamlines of this cluster across all atlas subjects was used to define the cluster's TAP.In this work, fiber clusters of interest from the cortex projecting to the caudate were identified according to their connected anatomical brain regions as defined in the TAP, in the same way as used in our previous study [6].In this way, we identified 17 white matter fiber clusters that connected the PFC and the caudate in both left and right hemispheres.The following FreeSurfer ROIs [63] were used to identify the clusters of interest to generate the streamlines of interest: the caudalanteriorcingulate, caudalmiddlefrontal, lateralorbitofrontal, medialorbitofrontal, parsopercularis, pars orbitalis, pars triangularis, rostralanteriorcingulate, rostralmiddlefrontal, superior frontal, frontal pole cortical ROIs and the caudate nucleus.We note that in using the dorsal striatum as our target zone for efferent streamlines projecting from the frontal cortex, we excluded the putamen.This was done because our inter-cluster endpoint distance measures are Euclidian distances and inter-cluster streamlines projecting to the putamen and the caudate might be measured as near to each other when they are far apart based on their non-Euclidian distances.Hence, to avoid ambiguity about how to measures such distances projecting to caudate and putamen, we instead used the simpler more banana shape of the caudate as the target for our streamlines from the frontal cortex.
Inter-cluster streamline endpoint distance analysis.To quantify the topographical relationship of these fiber clusters, we used our strategy of inter-cluster endpoint distance analysis [6].To achieve this, between each pair of fiber clusters we measured the mean Euclidean distance between the endpoints of streamlines at the level of the frontal cortex (i.e., the cortical inter-cluster, pairwise, endpoint distance) and the mean Euclidean distances between the endpoints of the streamlines in the corresponding fiber cluster pair at the level of the caudate (i.e., the caudate inter-cluster, pairwise, endpoint distance; see Fig. 1).This, in turn, allowed us to quantify the degree of convergence or divergence, i.e., deviation from a parallel, strictly topographic organization, among the 17 frontostriatal fiber cluster projections.Our method does not quantify the shape, or topology, of specific fiber clusters themselves.

Statistical analysis
Fiber clusters, and the distances between these clusters, were determined from HCP-EP Project diffusion scans on 108 non-affective subjects and 56 controls.No randomization or blinding was employed.Only one scan was available for each subject.Data for at the time all available subjects were used in the analyses.Consequently, although we had no control over the inclusion and exclusion criteria used by the HCP-EP project, we avoided any risk of introducing selection bias beyond what might already have been present.Since we used all available relevant data in the HCP archive, the number of subjects in our retrospective analyses were not based on a-priori consideration of statistical power.We performed t-tests, chi-square tests, and mixed model regression.All tests were at the 5% significance level, two-sided except for the chi-square tests, which were one-sided.Bonferroni adjustment for multiple comparisons was used for the 17 clusters within a hemisphere, where appropriate.The chi-square statistics were calculated from the 136 pairings of 17 clusters, and hence necessarily not independent.Therefore, we used permutation to determine the chisquare p values.For all statistical tests, we reported p values, test statistics, and degrees of freedom.Central tendency was measured using means, and error bars are deviations of plus or minus one standard error or of ½ standard deviation.
To determine the pattern of frontostriatal connectivity in both groups, first, we generated scatter plots for each hemisphere (see Fig. S1).based on the 17 fiber clusters (with 136 pairs of fiber clusters, yielding 136 data points), that showed the relationship between the cortical distances and the corresponding caudate distances of the fiber cluster pairs that connect the prefrontal cortex and the caudate.An exponential model describing the connectivity pattern was then fit to the data points as previously shown in Levitt et al. [6].Second, in both groups, we generated scatter plots (see Fig. S1).for each of the 17 clusters.For each cluster in each group, we performed a paired t-test of the distance from that cluster to the other clusters in the hemisphere, comparing these mean inter-cluster streamline endpoint distances at the level of the frontal cortex and at the level of the caudate, adjusting for the 17 tests performed, per hemisphere, using a Bonferroni correction.
Third, we categorized the mean differences in cluster endpoint distances (EP-NA minus HC) between groups in each hemisphere and plotted them with a convex hull (i.e., the smallest convex polygon containing all of the data points in the plot) yielding 4 quadrants with the point of origin signifying no group difference (see Fig. 2A).We used the signs of the differences of the endpoint distances in Cd and FCtx for each cluster pair for the 4 quadrants to calculate the number of cluster pairs in each quadrant (++; +−; −+; −− quadrants) for each hemisphere.We performed chi-square tests of the null hypothesis of equal counts per quadrant for each hemisphere, with p values of the correlated cluster pairs determined by permutation.Figure 2B illustrates how subtracting endpoint distances can result in a point in the upper left quadrant (i.e., −,+ quadrant) in Fig. 2A; points in other quadrants can be explained similarly.Upper left quadrant values occur when endpoint distances from a relatively more divergent pattern from an EP-NA subject (blue lines) are subtracted from a second endpoint pair with a relatively more convergent pattern (brown lines) from a HC.
Lastly, we assessed the between-group difference for each cluster pair in each cluster in the degree of convergence, reflected by a convergence quotient (CQ).Our CQ was calculated as: (Cortex Distance − Caudate Distance )/ (Cortex Distance + Caudate Distance ).For each cluster, we employed a mixed model regression analysis of the degree of cluster pair convergence, i.e., the cluster pair CQs, in both left and right hemispheres separately.For each of 17 clusters in each hemisphere, we assessed the between-group difference for the 16 pairings with all other clusters in the same hemisphere.For each cluster pair, the degree of convergence was quantified by a CQ.We fit 17 mixed model regressions for each hemisphere, each with convergence quotient as response, and subject as a random effect.Fixed effects in these models were all 16 pairings with a specific cluster in a hemisphere, group (EP-NA vs HC), and the interaction of pairing with group.

The overall pattern of frontostriatal connectivity
For both groups in both hemispheres, to evaluate the overall pattern of connectivity between the FCtx and striatum, we generated scatter plots in each hemisphere (see Fig. 3A).These plots were based on the 17 fiber clusters per hemisphere (with 136 pairs of fiber clusters, yielding 136 data points) that connect the FCtx and the caudate.These scatter plots show the relationship between the cortical cluster endpoint distances at the level of the cortex and corresponding cluster endpoint distances at the level of the caudate for each of the 136 fiber cluster pairs.We fit an exponential model to the data points which proved to be superior to a linear model for both groups in both hemispheres (see Fig. 3A).
We found a non-linear relationship between inter-cluster cortical distances and caudate distances in both HCs and EP-NA    groups in both left and right hemispheres which was driven by the results from the same 10 cluster pairs highlighted by the green circles located in on the right lower portion of the scatter plot [below] the exponential curve in Fig. 3A.In Fig. 3A, the 16 green circles represent clusters that include cluster number 6 as one its cluster pairs.Of note, a cluster originating in the inferior frontal gyrus, pars triangularis, fiber cluster number 6, was significantly over-represented in these 10 cluster pairs as can be seen by its being a member of each of these 10 cluster pairs for both groups in both hemispheres.Of further note, we found that the correlation curve in the right hemisphere showed a qualitatively different, and more flattened, non-linear relationship in the EP-NA patient group vs HCs, visually suggesting that the wiring pattern in the RH differed between groups (see Fig. 3A (d)).

Cluster variation in the pattern of frontostriatal connectivity
To determine the local variation of the pattern of frontostriatal connectivity in both left and right hemispheres, in both groups, we generated scatter plots for each of the 17 cluster pairs that show the relationship between cortical and caudate endpoint distances between that cluster and the other 16 (See Fig. S1).Paired t-tests comparing the mean inter-cluster endpoint distances in cortex and caudate between each cluster with the other 16 clusters in each group, revealed in both groups, bilaterally, similar clusters projected significantly convergently to the Cd.More specifically, we found that that for both groups in both hemispheres fiber clusters coming from the ventrolateral (clusters 6, 8 in LH and RH) and medial and lateral orbital frontal cortex, (cluster 10) showed significant convergence (i.e., caudate endpoint distance < cortex endpoint distance) after correcting for multiple tests (adjusted p value <0.05).Also, for both groups in the RH, a cluster coming from rostral middle frontal gyrus and superior frontal gyrus (cluster 1) showed significant convergence and in HCs a cluster from LH rostral middle frontal gyrus (cluster 9) showed significant convergence after correcting for multiple tests (adjusted p value <0.05).See Supplementary Table 1. for the t-test p values for fiber clusters showing significant patterns of frontostriatal convergence.See Fig. 3B for the scatter plot of cluster 6 and see Fig. S1 for the scatter plots of all fiber clusters.We note that Fig. 3B showing the scatter plots of cluster pairs that include cluster 6 as one of the cluster pairs, represent the same data points as the green circles in Fig. 3A (a)-(d).Also, as both age and FSIQ significantly differed between groups, we performed the following statistical analyses to test whether age or FSIQ influenced our inter-cluster distance measures, more specifically, the ratio of cortical to caudate distances for each pair of fiber cluster.Thus, for every pair in each hemisphere, we performed a linear regression of the ratio of cortical to caudal distances, first, with group, age, and their interaction as predictors and next, with group, FSIQ, and their interaction as predictors.We found no p values were significant for age or for FSIQ after a Bonferroni correction for the 136 cluster pairs per hemisphere.

Group comparison of the degree of convergence in frontostriatal connectivity
We used two between-group approaches to quantify the group difference in frontostriatal connectivity reflected in the difference in the fitted curves that we observed in the right hemisphere shown in Fig. 3A.First, we categorized the group mean difference in cluster endpoint distances in both FCtx and Cd areas by subtracting the HC distances from the EP-NA distances between groups in each hemisphere.We, in turn, used the signs of the differences of the endpoint distances in Cd and FCtx for each cluster pair, and demonstrated that there was a significantly greater number of cluster pairs with positive Cd and negative FCtx differences in the RH, but not in the LH.See Fig. 2A showing a plot enclosed by a convex hull perimeter with cluster pair 5, highlighted in red, against a background of all cluster pairs represented by the black plus signs.Also, see Fig. 2B visually demonstrating how we calculated the plot quadrant locations (++; +−; −+;quadrants) for each group difference in endpoint distances.Then, using a chi-square test with 1 degree of freedom  with p values determined by permutation, we showed for these pairwise cluster comparisons that there was an overall more convergent pattern in HCs compared with EP-NAs in the right (RH, adjusted for lack independence, p = 0.03) but not left (LH, adjusted for lack of independence, p = 0.13) hemisphere (see Fig. S2).
Second, using a mixed model regression analysis of the degree of cluster pair convergence, i.e., the cluster pair CQs, for both left and right hemisphere separately, we showed a significant group by fiber cluster pair interaction for 2 RH fiber clusters.These fiber clusters were numbers 5 and 11, which withstood Bonferroni significance correction for the 17 clusters in each hemisphere (F (1,15) =2.8; p = .00023;F (1,15) =2.8; p = .00023).See Fig. 4 (See Fig. S3 for similar plots of each cluster in both hemispheres).We note that fiber cluster number 5 originates in the frontal pole and the rostral middle frontal gyrus (rMFG), and fiber cluster number 11 originates in rMFG and inferior frontal gyrus (IFG), pars orbitalis (see Fig. 5).

DISCUSSION
In this study we used our novel diffusion imaging fiber cluster tractography method to assess the organization of frontostriatal brain wiring in EP-NA subjects and in HCs.We had four major findings.First, we found that the overall F-Cd wiring pattern, bilaterally, in both HCs and EP-NAs, deviated from a strictly topographic organization primarily driven by a cluster from IFG, pars triangularis.This was shown by non-linear, convex curves between inter-cluster cortical distances and caudate distances driven by the results from 10 cluster pairs in the left (LH) and right hemispheres (RH) in both HCs and in EP-NAs, all of which included a cluster coming from IFG, pars triangularis.Of note, however, we found a group difference such that in the RH in EP-NAs, the convex curve was more flattened.Second, for both groups in both hemispheres, we found that similar clusters showed significantly greater cluster pair convergence.More specifically, we found FCtxcaudate cluster projections with significantly more convergent patterns originating from PFC ventrolateral, dorsolateral, and orbitofrontal subregions.Third, we found a more convergent FCtxcaudate wiring pattern in the RH in HCs compared to EP-NAs.More specifically, we showed that there was a greater number of FCtx-caudate cluster pairs with a more convergent projection pattern in HCs vs EP-NA subjects which was significant in the right, but not left, hemisphere.Fourth, we showed a significant group by fiber cluster pair interaction for 2 right hemisphere fiber clusters projecting from the frontal pole and rostral middle frontal gyrus (cluster 5) and from rostral middle frontal gyrus and inferior frontal gyrus, pars orbitalis (cluster 11), respectively.We note that the contribution to the group difference in the overall wiring pattern of clusters 5 and 11 may be more than that of other clusters, although Clusters 5 and 11 do not specifically show a significantly greater group difference in convergence than do other clusters (See Fig. S2).However, based on the findings shown in Fig. 2, and the permutation chi-square statistic findings in Fig. S2, we believe that there is a more general pattern of disturbance in the degree of convergence in the right hemisphere in FC-Cd connectivity in EP-NA subjects compared with HCs.
Our finding showing a bilateral pattern of more localized deviation from a general topographic pattern of connectivity between the FCtx and caudate is both a replication and an Fig. 4 A mixed model regression analysis of convergence quotients.Using a mixed model regression analysis of convergence quotients (CQs) for both left and right hemispheres separately, we showed a significant group by fiber cluster pair interaction for RH Cluster 5 (p = 0.00023) and RH Cluster 11 (p = 0.00023).We note that fiber cluster 5 originates in the frontal pole and the rostral middle frontal gyrus and fiber cluster 11 originates in the rostral middle frontal gyrus and inferior frontal gyrus, pars orbital is (See Fig. 5).extension of our prior study in healthy subjects Levitt et al. [6].We conclude that such an anatomical pattern of connectivity is generally present in human brains and thus might be considered a transdiagnostic feature of brain organization.Further, we surmise that regions that deviate from a strictly topographic pattern of connectivity (i.e., either increased or decreased convergence), add greater circuit complexity and both promote greater cross-talk and circuit integration at the level of the striatum.Despite some similarities between groups in connectivity patterns between frontal cortex and caudate, our findings also importantly suggest group differences in the degree of convergence of such connectivity, what we call the CQ score, which occur both at an overall level of organization as well as at a more localized level.Specifically, we find in pairwise comparisons of all 136 clusters that more clusters in HCs are convergent in HCs than in EP-NAs which is significant in the right hemisphere.Also, we find that cluster 5, coming from the frontal pole and rostral middle frontal gyrus and the contiguous cluster 11, coming from the right hemisphere inferior frontal gyrus, pars orbitalis and rostral middle frontal gyrus, are the 2 localized subregions of the PFC whose patterns of projecting streamlines, with all other clusters, differ between EP-NA and HC subjects.Such specific clusters can be conceived of as specific subcircuits within the total PFC-caudate circuitry.Our method, thus, allows for a highly refined way of measuring localized brain wiring deviations from a healthy control pattern.These FCtx regions and the subcircuits projecting from them, as discussed below, subserve higher cognitive functions, such as executive functions, impaired in schizophrenia.

Right Rostral Middle
Our finding is novel as it isolates a specific brain circuit (the FCtx-caudate loop within the cortico-basal ganglia circuitry) and demonstrates using dMRI tractography in vivo group differences between EP-NA subjects and HCs.Such a finding is consistent with current thoughts that SZ is caused by dysconnectivity in brain circuits.For example, multiple studies have shown that functional and structural connectivity is disrupted in SZ [14,15,17,64].
As brain wiring occurs early in development, its disruption comports well with a neurodevelopmental hypothesis of schizophrenia [8,11,65].Further, as macroscale (i.e., long tract) brain wiring is established developmentally and then is enduring [66], dMRI tractography measures of brain wiring should serve as strong candidate trait biomarkers for disorders such as SZ with abnormal brain wiring [30].Moreover, genes affecting neuronal migration and axonal growth which can disrupt white matter long tract brain connectivity, have been found to be associated with schizophrenia [67,68].Further, prenatal stress, a risk factor for SZ, as shown in a non-mammalian vertebrate model via maternal immune activation, can influence circuit formation and normal axonal development [69].
With regard to prefrontostriatal brain wiring development in non-human primates, autoradiographic tract tracing studies in monkeys have shown a distinct pattern of corticostriatal connectivity that is established in the final third of pregnancy prior to birth [66].Importantly, this pattern can be shown in newborn monkeys as well as in adult monkeys [70].More specifically, it has been shown that PFC-Cd fibers terminate in a pattern in which terminals are concentrated in patches surrounded by areas without PFC input, described as fenestrated by Goldman [70], and that a transformation from a diffuse pattern of distribution during early gestation to a fenestrated pattern occurs by the final third of pregnancy [66, see Fig. 6].As adult monkeys continue to show the fenestrated pattern [70], the above strongly implies that in monkeys after the fenestrated pattern emerges during the final third of pregnancy, it persists into adulthood.Thus, the above monkey data suggest that our finding of group differences in FCtx-caudate wiring patterns between HCP-EP subjects and HCs reflects wiring patterns that emerged prenatally and persisted throughout postnatal development into adulthood.
The organization of corticostriatal anatomic connectivity has been thoroughly investigated through the use of animal tract tracing studies and in human brain imaging studies [22,30,71].Animal tract tracing studies have shown projection zone overlap in the striatum of cortical projections e.g.[21,72,73].For example, in a more recent monkey tract tracing study, Averbeck et al. [21] found the pattern of corticostriatal connectivity deviated from a strictly topographic one.They compared the distance between pairs of injection sites in the frontal cortex in monkeys with the degree of overlap in the projection zones of these cortical injection sites.They found an exponential decrease in overlap in striatal projection zones as a function of greater distance between pairs of injection sites.Such non-human primate studies suggest that projection zone overlap is a characteristic anatomic feature of corticostriatal connectivity.In healthy human subjects, Draganski et al. [22], using probabilistic diffusion imaging, reported projection zone overlap in the striatum coming from prefrontal, premotor and motor cortices.Although our data do not determine where fibers terminate inside the striatum, we interpret our convergence measures to reflect a pattern of projection convergence from the prefrontal cortex to the caudate as similar to a pattern of projection zone overlap described by Averbeck et al. [21] and Dranganski et al. [22].
The cortico-basal ganglia circuitry has been described to influence a number of important higher cognitive functions, in addition to its traditional role in influencing motor activity.As our data shows clusters 5 and 11 in the right hemisphere differentiate groups, it is of interest to review the function of the subregions from which these clusters project, i.e., the frontal pole, and rMFG (cluster 5), and the rMFG and IFG, pars orbitalis (cluster 11).
The rMFG is located in the dorsolateral PFC.It is a critical component of the frontoparietal control network and subserves processes that include goal directed behavior, cognitive flexibility, such as mental set shifting, and working memory [20,[74][75][76].The IFG has been shown to be engaged by word retrieval and the updating of working memory [77].In addition, the IFG, in particular in the right hemisphere, has been shown to subserve inhibitory control.For example, damage to the right hemisphere IFG in human subjects and rodents has been shown to interfere with inhibitory control [78][79][80].Further, the frontal pole has been described as involved in cognitive functions including multitasking, prospective memory, and mentalizing (i.e., theory of mind), functions which have been subsumed under the term metacognition [80,81].Deficits in such functions might cause significant difficulties in social cognition and employability, skills deficient in patients with schizophrenia [82,83].Of note, impairments in the above higher cognitive functions have been described in schizophrenia [1,29,[84][85][86][87].
A further interpretation of our finding implicating the RH might be that as we believe our brain wiring measures reflect development, inhibitory control, which is subserved by the RH in particular, might also be a developmental cognitive feature of psychosis.This could lead to the testable hypothesis that SZ subjects showing the greatest RH group differences in wiring pattern, i.e., the greatest degree of developmental disruption, will show greater deficits in inhibitory control processes, consistent with literature suggesting that cognitive deficits precede florid psychotic symptoms [88,89].Our finding might also suggest that inhibitory control deficits precede positive symptoms in psychosis.Further, consistent with the idea that RH findings may occur early in the course of illness and may reflect a neurodevelopment process, i.e., a failure of typical maturation, we note both a prior study of subjects at clinical high risk of psychosis that showed more pronounced abnormalities in WM microstructure in the right hemisphere [90] and a study showing microstructural abnormalities in the RH in early-course schizophrenia [91].
Limitations of the paper include that the design of the study is a cross-sectional one.The idea that our brain wiring measures reflect normal neurodevelopment and its deviations should be confirmed in longitudinal studies across the lifespan.Other limitations include that medication and illness chronicity confounds cannot be ruled out as many of the patients were receiving antipsychotics and illness durations were variable.Also, both age and FSIQ differed between groups, although as described above, our correction for them for the ratio of cortical to caudate distances for each pair of fiber cluster yielded no adjusted p values that were significant.As the difference in age was quite small, we suspect that such a difference is unlikely to alter our brain wiring measures.FSIQ and psychosis are related, consistent with the lower FSIQ in our EP-NA subjects, but fully disentangling the effect of FSIQ and illness effects on brain wiring is beyond the scope of our current study.A further potential limitation is that individual streamlines within fiber clusters need not terminate directly onto the caudate in order to be counted in the endpoint calculations as described above in the Methods section.Lastly, we acknowledge the risk of false negative and false positive streamlines using dMRI tractography [92,93].For future studies, it will be important to apply these measures in subjects over the lifespan, from early childhood to old age, to test their stability.Further, we plan to explore this circuitry in other neuropsychiatric disorders, such as early psychosis affective subjects, and to explore sex as a potential variable affecting brain wiring.Finally, FCtx-caudate brain wiring behavioral associations should also be explored.
In summary, employing a novel use of dMRI tractography, we found for both HCs and EP-NAs that the overall FCtcx-caudate wiring pattern similarly deviated from a strictly topographic relationship and had similar clusters that projected to the caudate in a significantly convergent pattern of connectivity.Conversely, we found an overall significantly more convergent pattern of connectivity in HCs in the right hemisphere and that 2 specific clusters from selective PFC subregions in the right hemisphere significantly differed in their pattern of connectivity between HC and EP-NAs.We surmise that regions showing group differences impact certain higher cognitive functions disrupted in schizophrenia including cognitive control, inhibition and metacognition.Lastly, we believe the importance of our brain wiring measures is that they reflect trait biomarkers which can help to identify subjects with schizophrenia, early in their development, who would benefit from early treatment intervention.

Fig. 2
Fig. 2 Pairwise Distances for Fiber Cluster 5 (Right Hemisphere).A A plot with a convex hull of the mean difference in cluster endpoint distances (EP-NAs minus HCs) between groups in the right hemisphere.Points (plus signs) and numbers in the upper left quadrant (i.e., −,+ quadrant) show a more convergent pattern in HCs vs EP-NAs.Red numbers, which substitute for a plus sign, label specific cluster pairs for cluster 5. B An illustrative example of 2 fiber cluster pairs, one with a more divergent pattern from an EP-NA subject (blue lines) and one with a more convergent pattern from an HC subject (brown lines).Subtracting cluster endpoint distances of an HC subject from an EP-NA subject would result in a point location in the upper left quadrant (i.e., −,+ quadrant) in the plot shown in (A).More specifically, in this manner, subtracting inter-cluster endpoint distances at the level of the Frontal Cortex (a, b) yields a negative value and subtracting inter-cluster endpoint distances at the level of the Caudate (c, d) yields a positive value.This example yields a data point located in the upper left quadrant (i.e., −,+ quadrant) of the plot shown in (A).Frontal Cortex represents the origin of the fiber cluster streamline endpoints at the level of the frontal cortex; Caudate represents the termination of the fiber cluster streamline endpoints at the level of the caudate.
subtype data was unavailable for 17 subjects.NA data not applicable, CPZ chlorpromazine, DSMV-TR Diagnostic and Statistical Manual of Mental Disorder, Fifth Edition, Text Revised.*p < 0.05, **p < 0.001.a The sample size (n) differs among variables owing to unavailability of data in some participants.b FSIQ = Composite Score Estimate (based on Wechsler Abbreviated Scale of Intelligence -second Edition (WAS-II)).c Vocabulary T Score (based on (WASI-II)

Fig. 3
Fig. 3 Graphical representation of the relationship between cortical and caudate distances.A Scatter plots showing the relationship between Frontal Cortex (cortical distance) and Caudate inter-cluster endpoint distances in LH HCs (a), in RH HCs (b), in LH EP-NAs (c) and in RH EP-NAs (d).Note that the EP-NAs convex curve in the RH (d) is more flattened in comparison to the other curves.In this figure, the 16 green circles in each subplot represent clusters that include Cluster 6 as one its cluster pairs.B Plots for Cluster 6 showing the relationship between Frontal Cortex (cortical distance) and Caudate inter-cluster endpoint distances in LH HCs (a) and RH HCs (b) and LH EP-NAs (c) and RH EP-NAs (d).We note that the data points in the scatter plots here correspond to the same data points as the green circles in Fig.4.A See Fig.S1.for the scatter plots of all fiber clusters.