ENIGMA‐DTI: Translating reproducible white matter deficits into personalized vulnerability metrics in cross‐diagnostic psychiatric research

Abstract The ENIGMA‐DTI (diffusion tensor imaging) workgroup supports analyses that examine the effects of psychiatric, neurological, and developmental disorders on the white matter pathways of the human brain, as well as the effects of normal variation and its genetic associations. The seven ENIGMA disorder‐oriented working groups used the ENIGMA‐DTI workflow to derive patterns of deficits using coherent and coordinated analyses that model the disease effects across cohorts worldwide. This yielded the largest studies detailing patterns of white matter deficits in schizophrenia spectrum disorder (SSD), bipolar disorder (BD), major depressive disorder (MDD), obsessive–compulsive disorder (OCD), posttraumatic stress disorder (PTSD), traumatic brain injury (TBI), and 22q11 deletion syndrome. These deficit patterns are informative of the underlying neurobiology and reproducible in independent cohorts. We reviewed these findings, demonstrated their reproducibility in independent cohorts, and compared the deficit patterns across illnesses. We discussed translating ENIGMA‐defined deficit patterns on the level of individual subjects using a metric called the regional vulnerability index (RVI), a correlation of an individual's brain metrics with the expected pattern for a disorder. We discussed the similarity in white matter deficit patterns among SSD, BD, MDD, and OCD and provided a rationale for using this index in cross‐diagnostic neuropsychiatric research. We also discussed the difference in deficit patterns between idiopathic schizophrenia and 22q11 deletion syndrome, which is used as a developmental and genetic model of schizophrenia. Together, these findings highlight the importance of collaborative large‐scale research to provide robust and reproducible effects that offer insights into individual vulnerability and cross‐diagnosis features.


| INTRODUCTION
The Enhancing Neuro Imaging Genetics through Meta-Analysis (ENIGMA) Consortium was conceived in 2009 with the goal of performing large-scale neuroimaging genetic studies and has since grown into a collaboration of more than 1,400 scientists worldwide . The ENIGMA diffusion imaging working group was organized in 2009 to develop analytic workflows that analyze the effects of genes, environment, and neuropsychiatric disorders on white matter microarchitecture. The initial focus was on the multisite analysis of fractional anisotropy (FA) images, as this is the most commonly studied scalar parameter extracted from diffusion tensor imaging (DTI) (Basser, Mattiello, & LeBihan, 1994;. The absolute FA values are sensitive to fiber coherence and organization, myelination levels, and axonal integrity and have been widely used as an index of white matter health (Thomason & Thompson, 2011). FA has emerged as a sensitive index of normal white matter maturation and aging (Penke, Munoz Maniega, Houlihan, et al., 2010;. Prior to the ENIGMA studies, microstructural abnormalities were reported in many neuropsychiatric illnesses and brain disorders including schizophrenia spectrum disorder (SSD) (Alba-Ferrara & de Erausquin, 2013;Friedman et al., 2008;Mandl et al., 2013;Nazeri et al., 2013), bipolar disorder (BD) (Barysheva, Jahanshad, Foland-Ross, Altshuler, & Thompson, 2013;Sprooten et al., 2011), major depressive disorder (MDD) (Carballedo et al., 2012) and others. To date, the ENIGMA-DTI protocols have been used in the largest studies ranking effect sizes for case-control differences in six common neuropsychiatric disorders and a genetic microdeletion syndrome (Table 1). We review the workflow used to derive these findings, and how their high reproducibility provides a basis for individual-level measurements of microstructural signatures, thereby enabling neuropsychiatric research across diagnostic boundaries .

| ENIGMA-DTI WORKFLOW
The ENIGMA-DTI workflow provided a generalizable analysis approach to extract phenotypes from DTI data collected by imaging groups around the world . This workflow is based on tract-based spatial statistics (TBSS) (Smith et al., 2006), that uses a skeleton of major white matter tracts as the basis for determining statistical differences in regional FA values. The ENIGMA-DTI protocol adapts the TBSS approach for performing ROI-based multisite research by providing a custom protocol that includes QA/QC steps, a custom ENIGMA-DTI minimal deformation warping target along with the skeleton of major white matter tracts, and steps to extract tract-average FA values . Diffusion measures extracted using the ENIGMA-DTI workflow showed excellent reproducibility in both test-retest (McGuire et al., 2017) and longitudinal data (Acheson et al., 2017).
The inaugural aim of the workflow was to perform multisite heritability analyses of these quantitative DTI-based phenotypes. We demonstrated that tractwise diffusion measures extracted using this workflow were consistently heritable (h 2 = 0.42-0.75)-regardless of the data collection protocol and study designs that included twins and siblings, extended families and pedigree-based cohorts Kochunov, Fu, et al., 2016). The regional heritability patterns in data collected using different DTI protocols were likewise strongly correlated with each other (r~0.6-0.9) (Kochunov, Fu, et al., 2016;Kochunov et al., 2015;Kochunov, Jahanshad, et al., 2014;Kochunov, Patel, et al., 2019). The high reproducibility and consistent heritability of ENIGMA-DTI measures across diverse study designs and data collection protocols provided a strong rationale for disorderoriented ENIGMA working groups to use this workflow to map deficit patterns in studies of several major neuropsychiatric illnesses (Table 1).
The number of subjects and cohorts that were used to derive disorder specific patterns for patient control differences   (Dennis et al., 2018) Abbreviations: BD, bipolar disorder; ENIGMA, Enhancing Neuro Imaging Genetics through Meta-Analysis; MDD, major depressive disorder; OCD, obsessive-compulsive disorder; SSD, schizophrenia spectrum disorder; PTSD, posttraumatic stress disorder; TBI, traumatic brain injury.

| ENIGMA-DTI FINDINGS IN NEUROPSYCHIATRIC DISORDERS
To date, the ENIGMA-DTI workflow was used to elucidate regional patient-control differences in brain microstructure in SSD (Kelly et al., 2018), MDD (van Velzen et al., 2019), BD (Favre et al., 2019), obsessive compulsive disorder (OCD) (Piras et al., 2019), traumatic brain injury (TBI) (Dennis et al., 2018), posttraumatic stress disorder (PTSD) (Dennis et al., 2019), and 22q11 deletion syndrome (Villalón-  (Table 2). Patients with SSD, BD, and MDD also showed a pattern of significant regional reductions in FA values. The comparison of regional effect sizes across the disorders provided a unique opportunity to summarize the impact of these illnesses across diagnostic categories (discussed in Section 4.2). Other illnesses did not show significant patient-control differences for the average FA values. Patients with OCD showed a modest number of regions, including the sagittal stratum (SS) and posterior thalamic radiation (PTR), where cases on average, had lower FA than controls (Table 2).
Subjects with PTSD showed no difference in either average (Cohen's d = −0.02, p = .7) or regional FA values (  (Table 2). This finding was interpreted as a possible marker of recovery by the original study (Dennis et al., 2018).
Together, these findings provide the first opportunity to evaluate the cross-disorder similarity, especially if these findings are reproducible in the independent samples.

| REPRODUCIBILITY OF ENIGMA FINDINGS IN NEUROPSYCHIATRIC DISORDERS
Research findings in neuropsychiatric illnesses have historically suffered from a substantial variability and heterogeneity both within and across disorders including genetics, environmental risk factors, mean age of onset, symptom presentations, treatment response, and longterm prognosis. The sources of heterogeneity have long remained elusive to clinicians and scientists and have contributed to a surprisingly poor reproducibility of neuroanatomical, functional, and genetic findings in neuropsychiatric illnesses. Meta-analysis has always offered a principled approach to screen studies for false positive findings by overcoming the "chasing of significance" observed in some discovery studies (Ioannidis, 2014). The big data analyses performed by ENIGMA differ from the traditional meta-analytic studies that derive the mean effect from group-level comparisons based on previously published effect sizes and often fall prey to the heterogeneity of the underlying methods used in the original studies. Instead, ENIGMA analyses are more akin to the "multisite-study analytic" approaches that directly coordinate the analysis of many data sets, by a group of collaborating scientists using the methods vetted for multisite research. However, ENIGMA does not enforce an a priori selection of image acquisition protocols and behavioral or diagnostic assessments.
Instead, ENIGMA pays considerable attention at each participating site to ensure the quality, integrity, and homogeneity of the underlying data, validity of the outcomes, and reproducibility of the deficit patterns.
A study by the ENIGMA-schizophrenia workgroup on subcortical deficits was the first validation of large-scale cooperative analyses of neuroimaging data in a severe mental illness. It used standardized methods to assess a sample of 2,028 patients and 2,540 controls from 15 centers worldwide (van Erp et al., 2015). This was the first study to show that the effect size for the smaller hippocampus in SSD patients was greater than that for the well-known enlargement of the lateral ventricles, refocusing attention on the neurological basis of this disorder. It also provided the first opportunity to test the premise that Big Data neuroimaging approaches could improve the reproducibility of findings in a disorder known for its heterogeneity. In a recent editorial, we observed that the effect sizes for patient-control group differences for volumes of subcortical structures reported by the ENIGMAschizophrenia group were in remarkable correlation (r 2 > 0.9) with two studies performed since then in largely independent cohorts (Alnaes et al., 2019;Kochunov, Thompson, & Hong, 2019;Okada et al., 2016).

T A B L E 2
Meta-analytical effect sizes (Cohen's d-values) (with group-wise significance in parentheses) of the patients versus control differences in disorders studied by ENIGMA disorderoriented workgroups. The sample information for each disorder is provided in Table 1 Region Anterior limb of internal capsule (ALIC) Body of corpus callosum (BCC) Genu of corpus callosum (GCC) Internal capsule (IC) Abbreviations: BD, bipolar disorder; ENIGMA, Enhancing Neuro Imaging Genetics through Meta-Analysis; FA, fractional anisotropy; MDD, major depressive disorder; OCD, obsessive-compulsive disorder; SSD, schizophrenia spectrum disorder; PTSD, posttraumatic stress disorder; TBI, traumatic brain injury.
The ENIGMA-schizophrenia group followed up with the study of white matter alterations based on a sample of 1,963 patients and 2,359 healthy controls from 29 independent international cohorts (Kelly et al., 2018, Holleran et al., 2020. This study was the first to report a regional localization of deficits in this illness as the regional pattern of effect sizes. The associative white matter tracts that connect the frontal, parietal, temporal, and limbic areas such as the anterior corona radiata (ACR) and the body and genu of the corpus callosum (GCC), showed significantly lower FA values in individuals with schizophrenia compared with controls. In contrast, cerebral pathways that carry sensorimotor fibers, such as the corticospinal tract and posterior limb of the internal capsule, showed no detectable patient-control group differences (Table 2). Importantly, patients diagnosed with schizophrenia also had significantly lower integrity of the fornix (FX)-the primary tract connecting the hippocampus with the frontal brain regions, consistent with the anatomical specificity observed for the subcortical volumetric deficits.
The regional pattern of white matter deficits reported by the correlations increased from 0.55 to 0.81 (Kochunov, Dickie Erin, et al., 2018). Kochunov and colleagues showed that the ENIGMAschizophrenia pattern was very highly correlated (r = 0.92) with the measured deficit pattern in another cohort and partly explained the two chief cognitive deficits in SSD: processing speed and working memory . A report on findings from the Beijing Connectome Project (BCP) found a high correlation (r = 0.86) between regional effect sizes observed in a sample of Han Chinese and the ENIGMA-schizophrenia pattern. A COCORO study used the ENIGMA-DTI workflow and an independent cohort collected across 12 sites in Japan to calculate regional effect sizes for SSD (N = 696 patients), BD (N = 211 patients), and MDD (N = 398 patients) using N = 1,506 healthy controls (Koshiyama et al., 2019). We report a very high correlation in regional effect sizes by ENIGMA and COCORO for SSD (r = 0.94), high correlation for effect sizes of BD (r = 0.79) and moderate correlation for MDD (r = 0.47) (Figure 1). The magnitudes of regional effect sizes reported by ENIGMA and COCORO showed no significant differences (paired t-test) for SSD (p = .9). ENIGMA regional effect sizes were significantly higher for both BD (average showed excellent-to-good consistency and reproducibility across geographically and ethnically diverse cohorts. The deficit patterns for MDD showed a moderate consistency, likely due to more modest effect sizes; however, this may improve once more independent studies are conducted.

| Translating ENIGMA findings to the individual level
The excellent agreement observed between the ENIGMA regional deficit patterns provides a novel perspective on big data neuroimaging findings. The inclusive worldwide nature of these studies has likely removed site-specific variances in diagnosis, medication, and environment, yielding deficit patterns that remain even after treatment with existing therapies and are shared by patients worldwide. The remarkable agreement across cohorts within each SSD meta-analysis, and F I G U R E 1 Scatter plot of regional effect sizes (Cohen's d coefficients) calculated for SSD (left), BD (center) and MDD (right) by COCORO consortium (y-axis) versus ENIGMA workgroup reports (x-axis). The effect sizes calculated in nonoverlapping cohorts showed very strong correlation for SSD (r = 0.94), strong correlation for BD (r = 0.79) and moderate correlation for MDD (r = 0.47). BD, bipolar disorder; ENIGMA, Enhancing Neuro Imaging Genetics through Meta-Analysis; MDD, major depressive disorder; SSD, schizophrenia spectrum disorder their subsequent independent replication studies, indicates that the profile of these regional effect sizes may be a signature, or a vector, that is related to the signature of the common physiopathological processes in schizophrenia or currently unmet treatment targets including cognitive deficits, treatment resistance, symptoms, and others. We first utilized the ENIGMA-schizophrenia DTI pattern as a predictor in the structural equation modeling of two major cognitive deficits, processing speed and working memory, that are affected in SSD patients. We found that the individual similarity with the ENIGMAschizophrenia deficit pattern mediated the association between white matter abnormalities in individual patients and the severity of cognitive deficits . The same pattern of structurefunction association was also observed in controls. This suggested that the regional pattern of the schizophrenia-related white matter deficits predicted the association between white matter and cognition even in the controls, indicating that the cognitive effects in schizophrenia are likely driven by reduced white matter integrity that are not secondary effects of antipsychotic medications .
The next step is to translate the findings from ENIGMA studies to enable predictions of vulnerability at the individual level. Can we use the characteristic patterns of regional deficits as predictors to link individual brain scans to vulnerability for a disorder and to its frequencies (Choi et al., 2018). PRS was shown to be a better predictor of risk than any single candidate risk allele (Colodro-Conde et al., 2018).
The regional vulnerability index (RVI) was developed as a simple correlational approach to quantify the agreement between an individual's brain and the expected pattern for the disorder. In contrast to PRS, the RVI approach is based on effect sizes derived from ethnically diverse samples and therefore RVI values are translatable across ethnicities (Kochunov, Huang, et al., 2019). RVI is a correlation coefficient between the normalized regional measures in an individual, such as tractwise FA or cortical gray matter thickness values, and the pattern of regional effect sizes reported by ENIGMA. A normalization process is used before computing the index, which includes a linear regression to remove effects of covariates, such as age and sex, from the individual's data, followed by z-transforming the residuals using the average and SD calculated from the healthy controls. For each subject, this produces a vector of regional measurements that captures the deviation from the normative values for each brain region and therefore mimics the contrast captured by the Cohen's d-values reported by ENIGMA. Higher RVI values (with a maximum of 1.0) indicate a better correlation with the expected disorder pattern. We hypothesized that higher similarity to the expected pattern is indicative of individual vulnerability to a disorder (Kochunov, Huang, et al., 2019).
We evaluated the RVI calculated for white matter DTI as a marker of treatment resistance in SSD (Kochunov, Huang, et al., 2019). The link between treatment resistance and cerebral white matter in SSD was suggested by previous white matter volume reduction findings (Molina et al., 2005) and reduced FA values (Holleran et al., 2014;Vanes, Mouchlianitis, Wood, & Shergill, 2018 ity, but individual regional measures do not. Higher RVI-SSD values likely reflect the contrast between the high vulnerability of associative and the lower vulnerability of motor and sensory brain regions to SSD (Kochunov, Ganjgahi, et al., 2016;Weinberger, 1996;Weinberger & Lipska, 1995). We hypothesize that by considering findings across the whole brain, RVI accentuates the regional effects specific to SSD.
Therefore, higher RVI values are identified in the individuals with more severe patterns of neurodevelopmental damage, who, in turn, are more vulnerable to developing cognitive deficits and negative symptoms.
In Kochunov et al. (2020), presented in this issue, we show that RVI can be calculated as a multimodal index by considering cortical thickness, subcortical gray matter volumes, and white matter microstructure measurements. Combining phenotypes across diverse neuroimaging modalities to derive a meaningful index of vulnerability is challenging, but the ENIGMA-schizophrenia findings provided a common denominator to combine these data. We first showed that RVI derived from cortical gray matter thickness, subcortical gray matter volume, and white matter integrity can inform patient-control differences and provide insight into the timeline for establishing these deficits in SSD. Elevated cortical RVI was readily detectable in the early diagnosis group (≤5 years since diagnosis) and remained stable with illness duration. This suggests that cortical deficits may develop before the onset of illness and do not change with illness duration. In contrast, white matter RVI was significantly elevated between early and chronic patients, suggesting ongoing illness progression. However, the multimodal RVI showed both the highest effect sizes among all measurements for all groups and was higher in chronic patients. While these findings are preliminary and are based on cross-sectional analyses, they demonstrate the potential for translating ENIGMA patterns to the individual level. We expect that novel analytic approaches, including machine learning, will take advantage of the ENIGMA datasets to derive more comprehensive measures that translate statistics from a large group to make predictions about an individual.

| ENIGMA-DTI: Facilitating cross-diagnostic analyses
The patterns of patient-control deficits derived using the ENIGMA-DTI workflow by neuropsychiatric disorder-oriented workgroups provide a "bottom-up" approach to evaluate the "integrative" versus "diagnostic silos" heuristics in neuropsychiatric research (Bzdok & Meyer-Lindenberg, 2017;McEwen, 2017). The integrative heuristic argues that risk factors, including genetics, stress, and others, are shared across major neuropsychiatric illnesses, while the diagnostic silos heuristic argues for separation of etiopathological factors while accepting potential co-morbidity of these illnesses. Big data psychiat-  Table 2). This suggests some anatomical specificity and partially replicates the findings of shared genetic risk factors among SSD, BD, and MDD (Brainstorm et al., 2018;Docherty et al., 2016). Strong correlations in regional effect sizes between SSD and BD (r = 0.75) and SSD and MDD (r = 0.82) were later replicated in COCORO data, however, the MDD and BD patients also showed a strong correlation in that cohort (r = 0.73) (Koshiyama et al., 2019).
The deficit pattern of PTSD showed moderate correlation with the deficit pattern of BD (r = 0.43), OCD (r = 0.43), and SSD (r = 0.39) and a very weak correlation with MDD (r = 0.22). This further supports anatomical specificity of the white matter deficits and partially F I G U R E 2 The correlation in regional deficit patterns among common neuropsychiatric disorders. **Indicates strong correlation coefficients. *Indicates moderate correlation coefficients The scatter plot of regional effect sizes for (a) BD versus SSD; (b) MDD versus SSD, and (c) MDD versus BD. BD, bipolar disorder; MDD, major depressive disorder; SSD, schizophrenia spectrum disorder replicates genetic correlation patterns among these illnesses (Brainstorm et al., 2018). We observed no significant correlation between disorders with a strong genetic component (SSD, BD, MDD, PTSD, OCD, and 22q11) and TBI-which is presumed to have mainly causes of acquired injury and environment, though individual genetics likely affects the recovery and preexisting psychiatric disorders are associated with a worse outcome after TBI (Gerring et al., 1998).
However, we observed a moderate negative correlation between TBI and OCD (r = 0.40) but this finding is difficult to interpret.
We observed no correlation (r = 0.04) between the 22q11 deletion pattern of regional effect sizes and that of SSD (Figure 4). 22q11 deletion is used as a developmental and genetic animal model for SSD (Mancini et al., 2019;Sumitomo et al., 2018) because people born with this deletion are 20-30 times more likely to develop psychosis. In striking similarity with SSD, the onset of psychosis is preceded by development of cognitive deficits, chiefly in the verbal learning and working memory domains (Vorstman et al., 2015). While subjects with the 22q11 deletion showed higher FA values in frontal areas, both SSD and 22q11 showed significantly lower integrity of the FX and FX/ST tracts ( Figure 4), which is supported by findings of lower hippocampal volumes in both conditions compared to controls. This difference in regional deficits is also mirrored by the pattern of cognitive deficits between the two conditions. The deficits in processing speed are pervasive in SSD and are linked to lower integrity of associative white matter tracts (Kochunov et al., 2010;Kochunov et al., 2017), but these deficits are minored in 22q11 deletion syndrome. Conversely, both disorders showed significant deficits in verbal learning and working memory domains (Chawner et al., 2017;Vorstman et al., 2015).
To summarize the regional deficit data, we performed a hierarchical clustering analysis and measured the Euclidean distance among clusters ( Figure 5, Table 3). Ward's minimum variance method was used to cluster the illness-specific patterns based on the half-square Euclidean distance among the deficit vectors. The disorder patterns were separated into three clusters based on their proximity. MDD, SSD, and BD were clustered together with the average distance between them of 0.56 ± 0.07. PTSD and OCD likewise were clustered together with TBI with an average distance between them equal to 0.73 ± 0.28. The pattern for 22q11 deletion syndrome was given its own cluster based on large distances from the MDD, SSD, and BD (distance = 2.9 ± 0.09) and PTSD, OCT, and TBI (distance = 2.36 ± 0.14).

| Limitations
This summary of ENIGMA cross-disorder analyses demonstrates significant limitations of the biological interpretations that can be derived from DTI data within and across disorders. We observed that patientcontrol differences can be both negative and positive indicating that neuropsychiatric conditions are associated with both lower and higher FA values in affected individuals. This signifies the general limitation of the DTI approach to quantify diffusion behavior of water . FA is a convenient statistical parameter produced by fitting a tensor that assumes a nonanisotropic Gaussian diffusion process and does not carry explicit biological information. While FA is often interpreted as an index sensitive to the degree of axonal myelination (Song et al., 2003;Song et al., 2005), it is neither a direct nor a specific measurement (Beaulieu, 2002 (Kochunov, Rowland, et al., 2016) as well as affect the fit of DTI model due to incomplete quantification of the diffusion process (Kochunov, Chiappelli, et al., 2014).

| CONCLUSION
The ENIGMA-DTI workflow was developed for imaging genetic analysis and validated by demonstrating uniform and reproducible heritability patterns across regional phenotypes. It was used across multiple brain disorders by ENIGMA workgroups and other studies for its ability to run the same analysis protocol worldwide, thus allowing multiple regional phenotypes to be aggregated and to deduce salient, consistent, and robust deficit patterns across illnesses. Abbreviations: BD, bipolar disorder; ENIGMA, Enhancing Neuro Imaging Genetics through Meta-Analysis; MDD, major depressive disorder; OCD, obsessive-compulsive disorder; SSD, schizophrenia spectrum disorder; PTSD, posttraumatic stress disorder; TBI, traumatic brain injury.