nature neuroscience Article https://doi.org/10.1038/s41593-023-01404-6 Regional, circuit and network heterogeneity of brain abnormalities in psychiatric disorders Received: 4 March 2022 Ashlea Segal 1,2 , Linden Parkes 3,4, Kevin Aquino1,2,5,6, Seyed Mostafa Kia7,8,9,10, Thomas Wolfers7,11,12, Barbara Franke 13,14, Accepted: 13 July 2023 Martine Hoogman13,14, Christian F. Beckmann8,15,16, Lars T. Westlye 11,17,18, Published online: 14 August 2023 Ole A. Andreassen 11,18, Andrew Zalesky 19,20, Ben J. Harrison19, Christopher G. Davey 21, Carles Soriano-Mas 22,23,24, Narcís Cardoner , 23,25,26 Check for updates Jeggan Tiego 1,2, Murat Yücel 1,27, Leah Braganza1, Chao Suo1,2,28, Michael Berk29,30,31,32, Sue Cotton30,31, Mark A. Bellgrove 1, Andre F. Marquand 7,8,15,33,34 & Alex Fornito 1,2,34 The substantial individual heterogeneity that characterizes people with mental illness is often ignored by classical case–control research, which relies on group mean comparisons. Here we present a comprehensive, multiscale characterization of the heterogeneity of gray matter volume (GMV) differences in 1,294 cases diagnosed with one of six conditions (attention-deficit/hyperactivity disorder, autism spectrum disorder, bipolar disorder, depression, obsessive–compulsive disorder and schizophrenia) and 1,465 matched controls. Normative models indicated that person- specific deviations from population expectations for regional GMV were highly heterogeneous, affecting the same area in <7% of people with the same diagnosis. However, these deviations were embedded within common functional circuits and networks in up to 56% of cases. The salience–ventral attention system was implicated transdiagnostically, with other systems selectively involved in depression, bipolar disorder, schizophrenia and attention-deficit/hyperactivity disorder. Phenotypic differences between cases assigned the same diagnosis may thus arise from the heterogeneous localization of specific regional deviations, whereas phenotypic similarities may be attributable to the dysfunction of common functional circuits and networks. The neurobiological mechanisms of mental illness are elusive. Thou- pathophysiological processes are poorly understood and clinically sands of neuroimaging studies have documented diverse brain changes useful biomarkers are lacking. associated with specific psychiatric diagnoses, and meta-analyses One reason for this limited progress may be a continued reliance have identified the brain regions that are most consistently affected on case–control designs, which compare group averages and ignore the by each condition, revealing both diagnosis-specific and transdiag- considerable clinical heterogeneity often shown by individuals with the nostic effects1–7. However, despite this substantial research effort, same diagnosis8,9. Indeed, recent magnetic resonance imaging (MRI) A full list of affiliations appears at the end of the paper. e-mail: ashlea.segal@gmail.com; alex.fornito@monash.edu Nature Neuroscience | Volume 26 | September 2023 | 1613–1629 1613 Article https://doi.org/10.1038/s41593-023-01404-6 Normal Deviation Circuit Network Fig. 1 | Characterizing neural heterogeneity at the level of brain regions, reveals areas that are functionally coupled to the deviant loci. In this work, functional circuits and extended networks. A schematic showing how neural we define a functional circuit as the set of regions that show significant FC with heterogeneity can be characterized at different scales. Nodes represent different a specific deviant region (orange). In this example, the two deviant areas are brain regions, edges represent functional coupling (FC) between nodes and coupled to a common region (black arrow) despite being located in different colored areas correspond to different functional networks of the brain. At the areas themselves. These circuits can be embedded within extended networks regional level (left), deviations from normative model predictions are localized (right) that include regions that may not be directly coupled to the deviant to specific brain regions in each individual. Red nodes show the locations of regions, but which nonetheless participate within the same functional such deviations mapped in two different people. A circuit-level analysis (middle) system (yellow). studies investigating person-specific patterns of brain deviations show We developed a new framework to integrate normative models of that group average differences are not representative of individual GMV with elements of lesion network mapping to map the functional cases10. These individual-specific inferences are often performed using circuits and extended networks within which regional GMV devia- normative modeling11,12, which involves modeling normative expecta- tions are embedded. We used this approach to derive a multiscale tions for a brain phenotype, such as gray matter volume (GMV), given characterization of neural heterogeneity across 1,294 individuals an individual’s age, sex or other relevant characteristics. The model diagnosed with one of six disorders: ADHD, ASD, BP, depression predictions can then be used to define a normative range of variation (MDD), obsessive–compulsive disorder (OCD) and SCZ. Inspired by against which new individuals are compared. Fitting the model to data studies of patients with brain lesions22–24, we tested the hypothesis for multiple brain regions yields a personalized deviation map quantify- that anatomically heterogeneous regional GMV deviations in each ing the extent to which each person deviates from population norms, disorder are functionally coupled with common sites, either within a thus enabling identification of areas associated with unusually small or functional circuit or an extended functional network (for a schematic large phenotypic values, termed extreme deviations, in an individual. explanation, see Fig. 1). Our transdiagnostic, multiscale approach Normative modeling studies of diverse MRI-derived phenotypes in allowed us to comprehensively understand the extent of neural het- attention-deficit/hyperactivity disorder13 (ADHD), bipolar disorder erogeneity within each disorder while also revealing commonalities (BP), schizophrenia10,14,15 (SCZ) and autism spectrum disorder16–18 (ASD) and differences between disorders. have found that, while cases often have more extreme deviations than controls (HCs; defined as the absence of any clinical diagnosis), the Results specific location of these deviations varies considerably across indi- Sample characteristics viduals with the same diagnosis. We examined data for 1,465 HCs (54.47% male) and 1,294 cases, taken This extreme regional heterogeneity of individual brain deviations from 14 different studies and 25 different scan sites. The clinical sam- aligns with the well-described clinical heterogeneity often associated ple comprised 202 individuals with ASD (100% male), 153 individu- with specific psychiatric diagnoses8,9, but raises an important ques- als with ADHD (41.18% male), 228 individuals with BP (47.37% male), tion: if cases show little overlap in the anatomical locations of their 161 individuals with MDD (34.16% male), 167 individuals with OCD GMV deviations, what then explains phenotypic similarities between (50.30% male) and 383 individuals with SCZ (62.14% male). The scan- people assigned the same diagnostic label? It seems reasonable to ner details, sample size and demographic characteristics of each scan assume that such similarities are driven by some common aspect of site, after various exclusions based on data quality and other criteria neural dysfunction across individuals, but the findings of normative (Methods), are presented in Supplementary Table 1 (for age distribu- modeling studies suggest otherwise. tions, see Supplementary Fig. 1). One possible explanation is that these regionally heterogeneous deviations aggregate within common circuits or neural systems. The Normative modeling brain is a connected network, and pathological processes often affect We used an established pipeline to obtain voxel-wise estimates of GMV distributed, interconnected systems19–21, meaning that it is possible in each participant25,26 (Methods), which we aggregated into regional for deviations in disparate loci to impact the function of common, estimates for 1,032 brain areas27,28 (1,000 cortical and 32 subcortical connected areas. This principle has been demonstrated by lesion regions; Fig. 2a). We then trained a normative model based on hierar- network mapping studies of neurological patients sharing a common chical Bayesian regression29 (HBR), fitted separately in each region, in motor, perceptual or cognitive syndrome22,23. Such patients typi- a training set comprising 1,196 HCs (55.02% male; Fig. 2a; HCtrain) to cally show little overlap in the anatomical location of their lesions, establish a normative GMV range given an individual’s age, sex and but the lesioned sites are often functionally coupled to common scan site (Fig. 2b; for model fit statistics and site effects, see Supple- areas (for a recent review, see ref. 24). The clinical expression of mentary Fig. 3 and Supplementary Table 2). HCtrain spanned the age these syndromes is, therefore, more closely related to dysfunction range of cases, enabling predictions for people aged between 18 and of loci coupled to the damaged area, rather than dysfunction of the 64 years (Supplementary Fig. 2). The remaining 269 controls (52.04% lesioned region itself. male; age range 18–62 years) were held out as a test set (HCtest) to estab- In this Article, we considered whether a similar process is at lish a normative benchmark for comparison with each clinical group play in psychiatric disorders by investigating whether anatomi- (Fig. 2c). For each individual in the clinical groups and HCtest, we quan- cally heterogeneous regional brain deviations within psychiatric tified the degree to which regional GMV estimates deviated from nor- disorders are functionally coupled to common areas and networks. mative model predictions as a z score (termed Deviation z maps; Nature Neuroscience | Volume 26 | September 2023 | 1613–1629 1614 Article https://doi.org/10.1038/s41593-023-01404-6 a GMV estimates c d Deviation z maps f Overlap map g ∆ Overlap map Increased GMV ADHD Decreased GMV Cases ASD HCtrain HCtest Threshold BP 5% More overlap b e Extreme deviation in HCtest maps HCtest MDD More overlap in 0% cases GMV OCD SCZ Age (years) h Group-based permutation test ADHD ASD BP MDD OCD SCZ P < 0.05, uncorrected P < 0.05, FDR corrected Fig. 2 | Regional heterogeneity of extreme negative GMV deviations in each z > |2.6| to identify extreme deviations (e). For the HCtest and each clinical group, disorder. a–f, Workflow for characterizing regional-level heterogeneity. GMV maps we quantified the proportion of individuals showing an extreme deviation in a given for each individual were parcellated into 1,000 cortical and 32 subcortical regions brain region, yielding an extreme deviation overlap map (f). We subtracted the HCtest (a). The training dataset, HCtrain, was used to train a normative model to make overlap map from each clinical group’s overlap map to obtain an overlap difference predictions about regional GMV values given an individual’s age, sex and scan site (b). map ( Δ overlap map) for each clinical group and then evaluated the magnitude of The predictions for held-out controls (HCtest) and cases were then compared with this difference (for details, see Extended Data Fig. 3a–d) (g). Cortical and subcortical empirical GMV estimates. Model predictions for one region, showing individuals in surface renderings showing regions with significantly greater overlap of extreme the training set (HCtrain; light blue) and the held-out control (HCtest; dark blue) and negative GMV deviations in cases compared with controls, as identified using clinical groups (red). Solid and dashed lines indicate the 99th and 95th centiles, group-based permutation tests (pink corresponds to Puncorrected < 0.05, red respectively (c). For each individual, deviations from model predictions were corresponds to PFDR < 0.05; two tailed, cases > controls) (h). Data used to generate quantified as a deviation z map (d). This deviation map was then thresholded at this figure can be found in Supplementary Data 1 (Regional_neg_thr26). Methods and Fig. 2d), with extreme deviations defined as z > |2.6| Supplementary Table 3 shows that, across all groups, >75% of (Fig. 2e), corresponding to P < 0.005, uncorrected, as per prior work10 participants show at least one extreme negative deviation. The propor- (for a discussion of thresholding, see Methods). tion of such participants was ~5–12% higher in MDD, SCZ, BP and OCD, Nature Neuroscience | Volume 26 | September 2023 | 1613–1629 1615 Article https://doi.org/10.1038/s41593-023-01404-6 compared with controls. Over 65% of participants showed at least one themselves (Extended Data Fig. 4). For instance, the maximum positive extreme deviation, and the proportion of such participants circuit-level overlap observed across regions was 33% in HCtest and was ~4–10% higher in ADHD, ASD, BP and OCD compared with controls. ranged between 39% (ADHD) and 53% (SCZ) in the clinical groups. People with BP, MDD, OCD and SCZ showed a higher burden of extreme To a certain extent, one should expect overlap to be higher at the negative deviations, defined as the total number of extreme deviations level of FC union maps (Fig. 3d,e) compared with regional deviation identified in each person, compared with HCtest ( P = 0.003; Extended maps (Fig. 2e,f) since any single deviant area can show FC with multiple Data Fig. 1 and Supplementary Table 3). Only people with ASD showed other regions, increasing the likelihood that common areas will be a higher positive deviation burden than controls ( P < 0.001; Extended implicated across individuals. For this reason, regional circuit-level Data Fig. 1 and Supplementary Table 3). Scan quality (for a definition, overlap maps in each clinical group must be contrasted with the HCtest see Methods) was not correlated with extreme deviation burden (HCtest: overlap map, which provides a normative benchmark for the expected ρ = 0.04, P = 0.51; cases: ρ = 0.03, P = 0.34). In the following, we first level of overlap in FC union values (Fig. 3f). A critical consideration in focus on characterizing negative GMV deviations (that is, GMV values these contrasts concerns the effect of total deviation burden, which dif- lower than normative expectations) before considering positive fers between groups (for example, Extended Data Fig. 1). For instance, GMV deviations. the total number of extreme deviations identified in individuals diag- nosed with SCZ was 4,951 compared with only 1,410 in HCtest. Heterogeneity at the level of brain regions This discrepancy means that more deviant-related FC maps will We quantified regional heterogeneity in GMV deviations (Fig. 1, left) be used in the FC union maps of SCZ individuals, thereby increasing as the proportion of individuals showing an extreme deviation in each the chance of observing higher overlap. On the one hand, this higher parcellated brain area, estimated separately for each diagnostic group circuit-level overlap will have real phenotypic consequences, since and the HCtest cohort (Fig. 2f). Despite most people showing at least one a higher deviation burden is an intrinsic and expected characteris- deviation in each group, the maximum percentage overlap within any tic of psychiatric disorders and these deviations are likely to impact of the 1,032 brain regions never exceeded 7% (ADHD: 3.27%; ASD: 4.95%; circuit-level function and behavior. On the other hand, it is informative BP: 5.26%; MDD: 6.21%; OCD: 4.19%; SCZ: 4.96%; HC: 2.97%). Hence, to determine whether the overlap is driven simply by group differ- individual extreme deviations were common, but they were rarely ences in deviation burden or reflects a preferential targeting of the found in consistent locations in individuals with the same diagnosis circuit by the disorder in question. We therefore evaluated the statisti- (Extended Data Fig. 2). cal significance of regional group differences in circuit-level overlap We next compared the spatial overlap of each clinical group with using two different permutation tests designed to disentangle these controls by subtracting their respective percentage overlap values in effects (for details, see Methods and Extended Data Fig. 3). The first, each region (Fig. 2g). The statistical significance of the observed Δ group-based permutation test, relied on shuffling the group labels of overlap values in each region was assessed with respect to an empirical the individual-specific FC union maps (Extended Data Fig. 3a–d) to null distribution generated by shuffling group labels (Methods and assess differences in overall circuit-level overlap regardless of group Extended Data Fig. 3a–d). While each disorder showed isolated areas differences in total deviation burden, thereby characterizing group of greater overlap at uncorrected levels, only ASD (32 regions) and BP differences in their ‘natural’ state. The second, spatial permutation (45 regions) showed differences that survived false discovery rate (FDR) test, evaluated group differences with respect to a null distribution correction ( PFDR < 0.05, two tailed; Fig. 2h). These differences were that preserves the number of deviation-related FC maps contributed by scattered throughout the cortex and rarely aggregated into spatially each individual, thus matching differences in the total deviation burden structured clusters. Few regions showed significantly greater overlap of each group (Methods and Extended Data Fig. 3e–k). This analysis in controls compared with cases (Supplementary Fig. 4). Repeating tests whether observed group differences in circuit-level overlap are the analyses using a method that avoids reliance on a single threshold greater than expected for FC maps generated from the same number for defining extreme deviations30 yielded similar findings (Methods of randomly chosen seeds, with the random seeds being selected from and Supplementary Figs. 5–7). Collectively, these results extend past a deviation map with the same underlying spatial autocorrelation as reports10,13,14 to indicate that minimal spatial overlap in the location of the empirical deviation maps (for further details and interpretation, person-specific extreme GMV deviations is a general characteristic of see Methods). In simple terms, the group-based permutation test psychiatric illness. identifies naturally occurring group differences in overlap regardless of deviation burden, whereas the spatial permutation test identifies Heterogeneity at the level of functional circuits differences in overlap beyond variations in overall deviation burden, We next asked whether the regionally heterogeneous extreme devia- implying a preferential targeting of specific functional circuits. tions identified in each clinical group show significant functional cou- Using the group-based permutation tests, we observed signifi- pling (FC) with common, remote areas, thus yielding greater cantly greater overlap across wide swathes of cortex for people with interindividual consistency at the level of distinct functional circuits SCZ and MDD ( PFDR < 0.05 , two tailed) compared with controls (Fig. 1, middle). To this end, we took each region showing an extreme (Fig. 3g). Regions with significantly greater overlap in SCZ were dis- deviation in each participant (Fig. 3a) and mapped its pattern of tributed diffusely and included ~75% of cortical areas. In MDD, ~31% of whole-brain FC in an independent sample of 150 unrelated HCs (HCP150) areas were implicated, predominately localized to regions of visual, to establish the normative pattern of expected FC for the deviant parietal, somatomotor, frontal and insula cortices. A similar spatial region (Fig. 3b). We thresholded ( PFWE < 0.025 ) and binarized each pattern was observed in ASD, OCD and BP when considering uncor- deviant-related FC map (Fig. 3c), and took the union of these thresh- rected results, although fewer regions survived FDR correction. Very olded maps across all extreme deviations for a given person (Fig. 3d; few areas showed greater overlap of deviant-related FC in controls for a discussion of thresholding issues, see Discussion), yielding a map compared with cases (Supplementary Fig. 8a). An alternative method that represents all areas showing significant FC with at least one for mapping FC results to parcellated regions yielded comparable extreme deviation in that individual. Next, we estimated, for each findings (Methods and Supplementary Figs. 9, 10a and 11a). region, the proportion of individuals within each group for whom that Extended Data Fig. 5 confirms that, for nearly all areas where we region showed significant FC with an extreme deviation (Fig. 3e). This identified greater circuit-level overlap in cases compared with con- analysis revealed that the overlap observed in brain regions function- trols (Fig. 3g), the magnitude of the difference exceeded the overlap ally coupled to deviant loci was much higher, in absolute terms, than differences observed at the regional level. This result aligns with our the overlap observed in the locations of the extreme deviations main findings to indicate that the level of overlap across participants Nature Neuroscience | Volume 26 | September 2023 | 1613–1629 1616 Article https://doi.org/10.1038/s41593-023-01404-6 a Extreme deviation b Map seed c Map to parcel d FC union map e Overlap map f ∆ Overlap map maps FC HCP150 P1 P1 P2 Threshold P2 Case P3 Case P3 0% 50% More More overlap overlap in cases in controls C1 Threshold C1 C2 C2 HCtest C3 C3 HCtest g Group-based permutation test h Spatial permutation test ADHD ASD ADHD ASD BP MDD BP MDD OCD SCZ OCD SCZ P < 0.05, uncorrected P < 0.05, FDR corrected Fig. 3 | Functional circuit heterogeneity of extreme negative GMV deviations areas showing direct FC with one or more deviant areas for a given individual (d). in each disorder. a–f, Workflow for characterizing circuit-level heterogeneity. For the HCtest and each clinical group, we quantified the proportion of individuals For each participant in the HCtest and each clinical group, we took each brain showing significant FC in a given region, yielding an extreme deviation FC overlap region showing an extreme deviation (a). For each individual in an independent map (e). We subtracted the HCtest FC overlap map from each clinical group’s FC sample of controls (HC150), we extracted a representative time course from each overlap map to obtain an FC Δ overlap map for each clinical group. Group deviant region and mapped the areas to which it is functionally coupled to using a differences in circuit-level overlap were evaluated with respect to two empirical seed-related FC analysis. Shown here are three participants in a clinical group null models (for details, see Extended Data Fig. 3) (f). g,h, Cortical surface (Case: P1, P2, P3) and three participants in the control group (HCtest: C1, C2, C3). renderings of regions with significantly greater overlap in cases compared with Two FC maps for two different deviant loci identified in P3, and one FC map for controls in areas functionally coupled to extreme deviations identified using one deviant loci identified in C3 are depicted (b). We thresholded and binarized group-based (g) or spatial permutation tests (h) (pink corresponds to each FC map associated with a given extreme deviation (c). Note, that no Puncorrected < 0.05, red corresponds to PFDR < 0.05, two tailed, cases > controls). subcortical regions survived this thresholding procedure. We took the union of Data used to generate this figure can be found in Supplementary Data 1 the thresholded maps across all deviant FC maps to obtain a single map of all (Circuit_neg_parc50). was greater at the circuit compared with regional level, supporting evidence of greater overlap in regions of left inferior and middle frontal the conclusion that GMV deviations in all psychiatric disorders except gyri at uncorrected thresholds (Fig. 3h). However, only three areas in ADHD are part of common functional circuits, despite being located BP and one in MDD survived FDR correction. Few results survived FDR in anatomically heterogeneous areas. correction when considering regions showing greater overlap in con- We next used the spatial permutation test to determine the degree trols (Supplementary Fig. 8b) and an alternative method for mapping to which the above differences in circuit-level overlap were attributable FC results to our regional parcellation yielded comparable findings to deviation burden. The results indicate that all disorders show some (Methods and Supplementary Figs. 10b and 11b). Thus, while these Nature Neuroscience | Volume 26 | September 2023 | 1613–1629 1617 Article https://doi.org/10.1038/s41593-023-01404-6 a Extreme deviation maps b c Extreme network d Overlap map e ∆ Overlap map deviation map Visual P1 Vis Vis Somatomotor P1 s SM HCtest s SM Dorsal attention Ba 50 Case Ba 20 P2 40 10 Case Salience/ventral attention P2 30 0 Limbic Tha Tha P3 20 DA DA –10 Frontoparietal 10 Assign P3 Default mode deviation SAL SAL Medial temporal e e MeT MeT to network / / Thalamus VA VA Basal ganglia C1 DM L DM L C1 F F C2 HCtest C2 C3 C3 f Group-based permutation test g Spatial permutation test ADHD ASD ADHD ASD Vis Vis Vis Vis s SM SM SM SM Ba 2.5 Ba s 2.5 Ba s 2.5 Ba s 2.5 2.0 2.0 2.0 2.0 1.5 1.5 1.5 1.5 **A 1.0 1.0 1.0 1.0 Tha D Tha ** * Tha Tha DA DA DA 0.5 0.5 0.5 0.5 SAL * SAL SAL SAL ** e MeT e e e MeT MeT MeT ** /VA /VA /VA /VA L DM L L L DM DM DM F F F F * BP MDD BP MDD Vis ** Vis Vis s SM * Vis ** SM SM Ba 2.5 s 2.5 SM Ba s 2.5 Ba s 2.5 Ba 2.0 2.0 2.0 2.0 1.5 1.5 1.5 1.5 ** 1.0 1.0 1.0 1.0 Tha DA Tha Tha Tha DA DA DA 0.5 0.5 0.5 0.5 SAL SAL SAL SAL e e MeT e e MeT * MeT MeT /VA /VA ** /VA /VA L L DM DM L L ** DM DM ** F F F F * ** OCD SCZ OCD SCZ Vis ** Vis Vis s SM * Vis ** s SM s SM Ba 2.5 Ba s 2.5 SM Ba 2.5 Ba 2.5 2.0 2.0 2.0 2.0 1.5 1.5 1.5 1.5 1.0 1.0 1.0 1.0 Tha DA Tha ** Tha Tha DA DA DA 0.5 0.5 0.5 0.5 SAL * SAL SAL SAL * e e e e MeT MeT MeT MeT * * /VA /VA /VA /VA ** ** DM L DM **L DM L DM L F ** F F F ** Fig. 4 | Functional network heterogeneity of extreme negative GMV null models (for details, see Extended Data Fig. 3). f,g, The network-level −log10 deviations in each disorder. a–d, Workflow for characterizing network-level P values associated with the difference in percent overlap for extreme negative GMV heterogeneity. For each individual in the HCtest and each clinical group (a), GMV deviations between each clinical group and the HCtest cohort (gray) under we assigned each brain region showing an extreme deviation to one of seven group-based (f) or spatial permutation (g) testing, respectively. ** corresponds to canonical cortical functional networks or three subcortical nuclei (b), such that PFDR < 0.05, two tailed, cases > controls, * corresponds to Puncorrected < 0.05, the entire network was considered deviant if it contained at least one region with two tailed, cases > controls. The solid black line indicates −log10 P = 1.6 (P = 0.05, an extreme deviation. The cortical surface renderings show the resulting two tailed, uncorrected). VIS, visual; SM, somatomotor; DA, dorsal attention; network-level extreme deviation maps (c). We quantified the proportion of SAL/VA, salience/ventral attention; L, limbic; F, frontoparietal; DM, default mode; individuals in each group showing a deviation within each network and MeTe, medial temporal; Tha, thalamus; Bas, basal ganglia). Data used to generate compared these proportions to the network overlap in HCtest (d). e, Group this figure can be found in Supplementary Data 1 (Network_neg_10network). differences in network-level overlap were evaluated with respect to two empirical Nature Neuroscience | Volume 26 | September 2023 | 1613–1629 1618 Article https://doi.org/10.1038/s41593-023-01404-6 findings offer preliminary evidence for the preferential involvement of the salience/ventral attention and dorsal attention/medial temporal of neural circuitry involving the lateral prefrontal cortex (PFC) in each systems, respectively. disorder, the dominant factor explaining greater circuit-level overlap in cases is total deviation burden. In other words, cases show greater over- Analysis of positive GMV deviations lap at the circuit level largely because they are more likely to express For completeness, we repeated the same analyses as above for extreme extreme GMV deviations, which in turn increases the probability with positive GMV deviations, representing areas where volume was higher which sites functionally coupled to deviant areas will be implicated. than normative expectations. At the regional level, extreme devia- tion overlap never exceeded 6% (ADHD: 5.23%; ASD: 3.96%; BP: 4.82%; Heterogeneity at the level of extended functional networks MDD: 5.59%; OCD: 4.19%; SCZ: 5.22%; HCtest: 2.60%) and very few regions Our analysis thus far indicates that the locations of regional GMV showed significant case–control differences in overlap (Extended Data deviations show marked individual heterogeneity, that cases show Figs. 6 and 7 and Supplementary Fig. 14). substantially greater overlap when considering the functional circuitry Circuit-level overlap was higher, reaching a maximum of 40% of these deviant loci and that this overlap is largely driven by total devia- across all regions and disorders (Extended Data Fig. 8). Group-based tion burden. However, our circuit-level analysis only focused on areas permutation testing revealed significantly greater circuit-level overlap showing significant FC with deviant regions, and these circuits often in individuals diagnosed with ASD compared with controls in ~20% form part of larger, extended functional networks that may not be fully of regions, predominately in visual, parietal and frontal cortices. No mapped by a circuit-level characterization (Fig. 1, right). We therefore other differences survived FDR correction (Extended Data Fig. 9a examined the location of person-specific GMV deviations in relation and Supplementary Fig. 15a). Similarly, spatial permutation testing to canonical functional networks, defined according to a widely used only identified isolated areas in pregenual cingulate and right lateral and validated classification of brain areas into one of seven cortical PFC as showing significantly greater overlap in MDD (Extended Data networks27,31 or three subcortical regions28, to derive a comprehensive, Fig. 9b and Supplementary Fig. 15b). multiscale characterization of GMV heterogeneity (Fig. 4a–c). In this At the network level, group overlaps were as high as 48% (Sup- way, if an individual showed an extreme deviation in at least one region plementary Table 5), with group-based permutation tests identifying affiliated to a given network, the entire network was considered devi- significantly greater overlap in the basal ganglia in SCZ and all cortical ant (Fig. 4c). Once again, we quantified the proportion of individuals networks except the default mode network in ASD ( P < 0.05 , two within each group showing a deviation in each network (Fig. 4d) and tailed), compared with controls (Extended Data Fig. 10a). Only the compared these proportions between the HCtest and each clinical group former difference was also observed with spatial permutation tests, (Fig. 4e). We then evaluated group differences in network-level overlap and was also accompanied by greater overlap in the salience/ventral using both group-based and spatial permutation tests, as was done in attention network (Extended Data Fig. 10b). The medial temporal lobe the circuit-level analysis (Extended Data Fig. 3). showed significantly greater overlap in controls compared with SCZ The results of these analyses are shown in Fig. 4d–f (for a summary (Supplementary Fig. 16). of the degree of spatial overlap (%) in each network for each group, see In summary, elevated circuit- and network-level overlap was par- Supplementary Table 4). Using group-based permutation testing, SCZ ticularly prominent in ASD under group-based permutation testing, and MDD cases showed significantly greater overlap in multiple net- and implicated areas of medial and lateral parietal, temporal and PFC works compared with controls ( PFDR < 0.05 , two tailed; Fig. 4f). For at the circuit level, and all cortical systems except the default mode SCZ individuals, the differences included all networks except for the network at the network level. These differences were not apparent thalamus and basal ganglia. In MDD, all cortical networks were impli- with spatial permutation testing, indicating that they were largely cated. The dorsal attention network and medial temporal lobe were driven by the elevated positive GMV burden of individuals diagnosed implicated in ADHD and the ventral attention network in OCD. No with ASD. Differences in circuit- and network-level overlap for positive networks survived multiple comparison correction for ASD and BP. At GMV deviations in other disorders were less pronounced. uncorrected levels, the salience/ventral attention network was impli- cated across all disorders except for ASD. There were no networks in Discussion which controls showed greater overlap (Supplementary Fig. 12a). We adapted lesion network mapping22,23 for use with normative mod- Using spatial permutation tests with FDR correction, only the eling11,12 to characterize individual heterogeneity of brain deviations salience/ventral attention network showed greater overlap in SCZ, at the regional, circuit and network level in different psychiatric disor- and the medial temporal lobe and dorsal attention networks showed ders. We showed that heterogeneity between individuals in regional greater overlap in ADHD (Fig. 4g). No deviant-related networks showed GMV deviations is a general feature of psychiatric illness, but that greater overlap in controls (Supplementary Fig. 12b). these regionally heterogeneous loci are often embedded within We repeated the same analyses using a 20-network parcellation common functional circuits and networks. Regional heterogeneity (17 cortical networks and 3 subcortical nuclei). The results of these thus offers a plausible explanation for the well-described clinical analyses are shown in Supplementary Fig. 13 (also see Supplementary heterogeneity observed in psychiatric disorders8,9, while circuit- and Table 5). Broadly, the findings align with those obtained using the network-level aggregation of deviations is a putative neural substrate ten-network parcellation to indicate that SCZ and MDD are associated for phenotypic similarities between individuals assigned the same with greater network overlap in negative GMV deviations across most diagnosis. Using different null models, we showed that much of networks, with the effects being more circumscribed in the other dis- the elevated overlap observed at the circuit and network levels is orders. Using a 20-network parcellation, we found robust evidence for attributable to total deviation burden. This result challenges promi- preferential involvement for temporoparietal and control networks in nent models in which specific disorders are assumed to arise from ADHD and MDD, respectively. dysfunction of characteristic neural systems selectively targeted by Together, these results align with the circuit-level analysis to the disease process32. indicate that SCZ and MDD are associated with greater network-level overlap of negative GMV deviations. Group differences in overlap are Patient-specific deviations are regionally heterogeneous more circumscribed in ADHD, ASD, BP and OCD, although the salience/ Few areas showed significantly greater regional overlap in cases ventral attention network is implicated across most disorders. These compared with controls, and no single area in any disorder showed an effects are largely driven by total deviation burden, with only SCZ and extreme deviation in more than 7% of cases across the 1,032 regions ADHD showing robust evidence supporting preferential involvement investigated. This substantial regional heterogeneity aligns with Nature Neuroscience | Volume 26 | September 2023 | 1613–1629 1619 Article https://doi.org/10.1038/s41593-023-01404-6 past normative modeling studies of GMV in SCZ, ADHD and BP10,13,14, Generic and preferential targeting of neural systems confirming that this heterogeneity is a general feature of psychi- Our spatial permutation test allowed us to evaluate the degree to which atric illness. The variable involvement of different regions across circuit- and network-level overlap was explained by group differences individuals may yield distinct clinical profiles and drive phenotypic in total deviation burden. At the circuit level, prefrontal regions were heterogeneity in people with the same diagnosis8,9. An important identified as showing greater overlap compared with controls in all clini- avenue of future work will involve precisely characterizing the rela- cal groups at uncorrected levels, consistent with extensive evidence tionship between GMV deviations and interindividual differences for prefrontal dysfunction in each condition47–49. However, only regions in symptom expression. of right and left lateral PFC in MDD and BP, respectively, survived FDR More generally, the high degree of regional heterogeneity correction. The right PFC showed 21% overlap in MDD cases, whereas observed indicates that group mean comparisons are not representa- the left PFC showed 14–15% overlap in BP. Extensive literature implicates tive of the specific profile of GMV deviations apparent in any individual lateral PFC dysfunction in affective disorders48,50,51, and the dorsolat- case. Mean comparisons may thus offer an incomplete account of eral PFC is a popular brain stimulation target in both MDD52 and BP53. causal pathophysiological mechanisms, unless the broader network The region has also been implicated by lesion network mapping as a context of any identified group differences is considered. Notably, most core site explaining the emergence of depression following stroke54. participants showed a low deviation burden (Supplementary Table 3). Given that our spatial permutation test did not find any evidence of This result is expected, since the failure so far of psychiatric neuroimag- greater network-level overlap in MDD and BP, our findings point to a ing to identify robust diagnostic biomarkers of illness indicates that any strong specificity for functional circuits coupled to these particular disease-related brain changes are likely to be subtle and complex. We prefrontal areas. Spatial permutation testing also revealed greater should, therefore, be circumspect about our ability to identify strong network-level overlap for SCZ in the salience/ventral attention network neurobiological signatures of psychiatric disorders. and for ADHD in the dorsal attention network and medial temporal regions, suggesting that GMV deviations preferentially target these Deviations aggregate in common circuits and networks systems in a disorder-specific way. Despite considerable heterogeneity at the regional level, deviations Lesion network mapping of neurological cases indicates that were often coupled to common functional circuits and networks. In symptom expression is often driven by dysfunction of remote sites some cases, >50% of people with the same diagnosis showed a devia- connected to a lesion, implying that these remote sites represent tion implicating at least one of these systems, with group-based per- viable treatment targets22,23,55,56. Following this logic, the findings from mutation testing indicating that all disorders showed some evidence our spatial permutation analysis imply that viable treatment targets of greater circuit-level overlap than controls. This higher overlap at may be located in the right PFC for MDD, in the left PFC for BP, in the circuit and network levels parallels lesion network mapping studies salience/ventral attention network for SCZ, and in the dorsal attention of neurological syndromes, which suggest that many clinical pheno- and medial temporal networks for ADHD. However, these targets may types are not caused by dysfunction of the lesioned area itself but by only be relevant for a subset of patients, with the observed overlaps its impact on remote, functionally coupled regions22–24. Our findings ranging between 10% and 50% across disorders. It thus follows that the indicate that a similar process may occur in psychiatric disorders, with most current approaches, which attempt to identify a single common anatomically distributed GMV deviations often being coupled within therapeutic target for each diagnosis, will only have limited success. similar circuits and networks. The consequent dysfunction of these A more comprehensive understanding of patient-specific brain common circuits and networks may drive clinical similarities between changes, and their network context, will be necessary to develop more people with the same diagnostic label, despite extreme heterogeneity effective, and personally tailored, interventions. in the locations of the deviations themselves. Multiple mechanisms In most other cases, we failed to reject the spatial null hypothesis. may explain these circuit- and network-level effects, ranging from This result indicates that many case–control differences in circuit/ transient diaschisis-like effects on distributed circuit/network func- network overlap observed under group-based permutation cannot tion through to more prolonged transneuronal dysfunction caused be attributed to the preferential accumulation of deviations within a by aberrant interregional signaling or disrupted axonal transport of particular circuit/network, since the differences are consistent with trophic factors20. comparisons of the same number of randomly selected seeds in each Areas of frontal, parietal, insula and temporal cortex showed group. When taken with our group-based permutation analysis, our greater circuit-level overlap across most conditions, and cross-disorder findings indicate that accumulating deviations are more likely to be differences were more of degree rather than kind. For instance, differ- coupled to areas in prefrontal, temporal, parietal and insula cortices, ences in circuit-level overlap were spatially circumscribed in ADHD simply because these areas are known connectivity hubs of the brain but encompassed nearly the entire brain in SCZ (Fig. 3g). These find- and have a higher probability of being functionally coupled to other ings challenge classical views that distinct psychiatric diagnoses are regions57–59. While this circuit- and network-level accumulation will associated with dysfunction in specific circuits (for example, ref. 32,) still have real phenotypic consequences, the spatial permutation and suggest that each disorder is associated with complex changes test allowed us to identify an underlying random process as a can- that affect diverse neural systems33,34, often transdiagnostically35–37. didate generative mechanism. This finding challenges the implicit Accordingly, our network-level analysis revealed greater overlap in assumption in many studies that any brain changes observed in a the salience/ventral attention network in five of the six disorders that disorder result from a targeted pathophysiological process. Further we considered (Fig. 4f). The salience/ventral attention network plays work investigating the genetic and environmental contributions to a central role in cognitive control38, interoceptive awareness39, and individual-specific deviations should elucidate the mechanisms driv- switching between internally and externally-focused attention40. Its ing their anatomical distribution. dysfunction has been implicated in a diverse range of psychiatric disor- ders41–44, is associated with increased levels of general psychopathology Heterogeneity of positive deviations in youth45 and shows cross-disorder abnormalities in meta-analyses of The locations of positive GMV deviations were more heterogeneous classical case–control voxel-based morphometry (VBM) studies and than negative deviations, and showed less overlap across people. The functional neuroimaging studies35–37. Our findings support this past exceptions were ASD and BP at the regional level, which showed signifi- work to suggest that salience/ventral attention network dysfunction cantly greater overlap in predominately frontal, parietal areas. At the may play a critical role in the expression of general psychopathological circuit level, ASD showed significantly greater overlap in visual, parietal processes common to diverse diagnoses46. and frontal cortices, and at the network level, all cortical networks Nature Neuroscience | Volume 26 | September 2023 | 1613–1629 1620 Article https://doi.org/10.1038/s41593-023-01404-6 except the default mode network were implicated in ASD. ASD has been Our cases were diagnosed according to Diagnostic Statistical associated with dysregulated and accelerated brain growth, particularly Manual of Mental Disorders (DSM) or the International Classification in the temporal, parietal and frontal association cortices during early of Diseases (ICD) criteria70,71. Given the widespread application of DSM childhood60. Whether these increases persist into adulthood and can and ICD in clinical and research contexts, we deemed it important to explain the present findings remains unclear. understand heterogeneity with respect to these constructs and to Differences in circuit-level and network-level overlap for positive gain insight into the neural correlates of phenotypic similarities and GMV deviations in other disorders were less consistent. In MDD, greater differences between cases assigned specific diagnoses. Nonetheless, circuit-level overlap in left pregenual cingulate and anterior right lat- focusing on specific syndromes rather than traditional diagnoses, as eral PFC is consistent with the known roles of these areas in regulating in the lesion network mapping literature23,24, may yield a more precise emotion61 and cognitive control62, respectively. The greater overlap mapping between regional deviations, their network context and observed in the basal ganglia of people with SCZ may be attributable behavior. Establishing sufficiently large databases for conducting such to the effects of antipsychotics, which can cause volumetric expansion syndrome-focused analyses will be a key challenge for future work. in this region63. Conclusions Limitations Our multiscale analysis of neural heterogeneity across six psychiatric While deviation burden was higher in cases than controls, and most disorders confirms that extreme regional heterogeneity of GMV devia- people showed at least one deviation, a substantial fraction of clinical tions is a general characteristic of mental illness. Further, we showed participants (~50%) showed a relatively low deviation burden (<3 devia- that these deviations are often coupled to common functional circuits tions). This result implies that there is considerable overlap between and networks, offering a putative neural substrate for phenotypic simi- cases and controls, as noted for many other phenotypes64, and that brain larities among individuals assigned the same diagnosis. The common changes associated with psychiatric illness may be subtle and complex. involvement of prefrontal and parietal circuits, and the salience/ventral This subtlety, combined with considerable interindividual heterogene- attention network, across disorders may be a marker of transdiagnostic ity, may partly explain the failure of the field to identify pathognomonic psychological distress, with the variable involvement of other systems biomarkers for psychiatry, despite decades of research. The result may explaining phenotypic differences among disorders. More broadly, also reflect the limited sensitivity of MRI-derived GMV estimates to map our findings underscore the need to consider the network context of pathophysiologically relevant brain changes. Normative modeling of disorder-related markers of pathophysiology20, and indicate that the other phenotypes, such as those obtained with functional or molecular clinical expression of disease is not driven solely by sites of primary imaging, may reveal stronger separation between groups. pathology, but also by the effect of this pathology on remote, con- Sample sizes in single-site psychiatric neuroimaging studies are nected systems. often small, so we pooled data from multiple sites to generate a suf- ficiently large cross-disorder dataset. As a result, the data were col- Mapping neural heterogeneity at the regional lected with different acquisition, recruitment and clinical assessment level protocols. To avoid introducing a dependence between scan site and Our analysis aimed to characterize neural heterogeneity at the level diagnostic group, which can confound case–control comparisons, of individual brain regions, neural circuits and extended brain net- we focused primarily on control data obtained on the same scanners works (Fig. 1). To map heterogeneity at the regional level, we obtained as the clinical data. While our model diagnostics indicated adequate person-specific deviation maps, representing the extent to which the fit to the data, future work could extend the techniques used here so GMV of a given person within each brain region deviates from norma- that they can be used with normative models trained on larger samples tive predictions. This analysis involved quantifying regional GMV for (for example, refs. 65,66) to obtain more reliable deviation estimates. each individual and evaluating the observed measures with respect to Although we used stringent quality control, and our hierarchical an underlying normative model, as outlined in the following. Bayesian model29 appropriately parsed site-related variance (Supple- mentary Table 2), investigating correlations with symptom profiles or Anatomical data other clinically relevant variables such as age of onset, illness duration, We estimated regional GMV using VBM of T1-weighted anatomical medication exposure or disease severity across disorders was beyond MRI scans25,26. We focused on GMV because it is one of the most fre- our scope. Indeed, while our methods are useful for characterizing quently studied neural phenotypes in psychiatry. VBM is also arguably circuit- and network-level overlap across individuals, they are not well the most widely used tool for measuring GMV in this context. Image suited for analyses of individual differences in behavior. A valuable acquisition parameters for each dataset are provided in Supplementary extension of our work would involve directly testing the hypothesis Table 1. The following outlines our quality assurance procedures and that circuit- and network-level overlap relate to clinical similarities data processing pipeline, which was applied to all raw T1-weighted between patients. Addressing this goal would require appropriate images obtained for each dataset. methods for aggregating deviation-related circuit- and network-level data within and between individuals, and extensive measurement of Quality control diverse symptom dimensions to allow adequate assessment of clinical All T1-weighted images were visually inspected and evaluated for the phenotypes67. Extensive symptom sampling that allows for personal- presence of artifacts72,73, resulting in the exclusion of 53 images with ized models would be particularly useful in this context68. Such data gross artifacts or abnormalities. Next, we used the Computational are rarely available in single-site clinical neuroimaging studies. Future Anatomy Toolbox25 (CAT12 r1113, ref. 74) to generate a weighted overall studies may work toward developing harmonized, transdiagnostic and image quality rating (IQR) for every scan. This metric combines ratings multisite clinical protocols that can be used across different disorders, of basic image properties, including the level of noise and geomet- such as those informed by the Hierarchical Taxonomy of Psychopa- ric distortions, into a single score that quantifies the overall image thology69. The specific age at which the measures are acquired should quality of a participant’s T1-weighted scan (for more information, see also be considered, given that the nature and magnitude of case– ref. 74). On this metric, lower scores denote higher image quality. As per control differences may vary across the lifespan4. Such an approach previous work10, we excluded 153 images with an IQR >2.8. An additional may facilitate data-driven strategies to identify biological subtypes 63 images were excluded due to a failure of our processing pipeline. based on patient-specific deviation maps that cut across traditional Collectively, this quality assurance process resulted in the exclusion diagnostic boundaries. of a total of 269 scans Nature Neuroscience | Volume 26 | September 2023 | 1613–1629 1621 Article https://doi.org/10.1038/s41593-023-01404-6 MRI preprocessing samples. To account for these site-related effects, we built our norma- Regional GMV was estimated using the CAT12 VBM pipeline25,26, which tive model using HBR, which successfully accommodates signal and is included as an extension of Statistical Parametric Mapping software noise variance in multisite data by estimating different but connected v777175 in MATLAB v9.8 (ref. 76). Briefly, the T1-weighted images were mean and variance components through shared prior distributions first corrected for intensity nonuniformities, and segmented into gray across sites29. Before modeling, regional GMV estimates were first matter (GM), white matter and cerebrospinal fluid tissue probability subjected to a Box–Cox transformation to ensure normality82. The maps. Then, using the high-dimensional Diffeomorphic Anatomical optimal lambda parameter for minimizing skewness was estimated for Registration Exponentiated Lie Algebra77, the segmented scans were each brain region independently using maximum likelihood. normalized into standard Montreal Neurological Institute IXI555 space. For the model to accurately parse variance attributable to age, sex Lastly, the images were bias-field corrected and modulated by the lin- and site, sufficient observations are required for any given combination ear and nonlinear components of the Jacobian determinant obtained of these variables. We therefore excluded data cells containing less than from the Diffeomorphic Anatomical Registration Exponentiated Lie ten HCs for a particular sex at a given site. If this exclusion procedure Algebra deformation fields to obtain voxel-wise estimates of GMV. resulted in less than ten cases in total from any given site, the entire site To constrain our analyses to GM voxels, we generated a mean image was excluded, resulting in the exclusion of 217 scans. from all the normalized GM maps and retained voxels with a tissue The training data for the normative model (HCtrain) were created by probability ≥0.2. randomly selecting 90% of HC individuals from each site, provided that the site included data for ≥30 HCs. For sites with smaller HC samples, all Brain parcellation HC data were included in the training set. The test data (HCtest), which We employed HBR to estimate normative models of regional GMV was completely independent of the normative model, comprised the (see the Normative Model section, below). To limit computational bur- 10% of HCs from sites with ≥30 HCs and all case scans. The HCtest data, den, we parcellated the brain into 1,032 cortical and subcortical regions comprising 269 individuals (140 male), offered a normative benchmark by combining well-validated parcellations of the cortex27 comprising for assessing case-specific model deviations, as outlined below. 1,000 regions, and of the subcortex28 comprising 32 regions. The corti- Following stratification of our sample into training and test sub- cal parcels have been mapped to well-described, canonical functional sets, for each of the 1,032 parcellated brain regions, we fitted sep- networks of the brain31, facilitating our analysis of network-level overlap arate HBRs29. The HBRs modelled site and sex effects with random in deviations. slopes, intercepts and noise, to model GMV as a function of age, sex Recent work indicates that group-based parcellations can be tai- and site in the training data, yielding estimates of normative regional lored to capture individual variability in regional borders (for example, GMV variance and predictive uncertainty. Importantly, based on a ref. 78). We used a group-based parcellation to ensure a consistent partial-pooling approach, shared prior distributions were imposed approach across cortex and subcortex (methods for individual sub- over site-specific and sex-specific model parameters. These shared cortical parcellations have not yet been developed), to allow direct priors assume that, while the model parameters for each site and sex comparison with extant literature (which have overwhelmingly relied are different, they are drawn from a common distribution. These shared on group parcellations), and to ensure that region-of-interest seeds priors thus regularize the model parameters and prevent the model remained approximately the same size among participants, as the from overfitting small batches for a given sex or site. effect of varying seed sizes between individuals has not yet been stud- To quantify deviations from the normative model predictions for ied extensively. Investigating the effect of individually tailored parcel- each participant in the test data, we generated deviation z maps for lations on our findings will be an important goal for future work. Voxels each participant. Specifically, for each participant i, at each brain region that overlapped between atlases were assigned to the corresponding j , we combined the predicted GMV yˆij , true GMV yij , the predictive subcortical region. Regional GMV estimates were obtained using freely uncertainty σij and the normative variance σnj to calculate a z score, available code79. zij , as The choice of parcellation will necessarily affect the degree of overlap observeable in any given brain area, such that coarser parcella- yij − yˆij zij = , tions will be associated with higher levels of overlap. Comparison with 2 2 √σij + σnj the control data thus provides a critical normative benchmark against which to evaluate the levels of overlap observed in each clinical group. We chose our 1,032-region parcellation to offer high spatial resolution which quantifies the extent to which an individual’s regional GMV while ensuring computational feasibility. The same parcellation was estimate deviates from the model prediction, given the uncertainty used in regional-, circuit- and network-level analyses, facilitating direct of the model. comparison across scales. For any given brain region, we are interested in participant values that show large deviations from the normative expectations set by the Normative modeling model, under the assumption that these deviations represent patho- We used normative modeling to obtain person-specific GMV devia- physiologically relevant features of the disease11. We used two comple- tion maps in relation to an underlying model of normative expecta- mentary approaches to characterize large positive and negative tions for regional GMV variations (see refs. 11,12,29, PCNtoolkit v0.16 deviations from the normative model at each region and for each (ref. 80)). Normative models estimate the mean and variance (referred individual. First, we thresholded the deviation maps at z > |2.6| (that to as a normative range) of a response variable (for example, GMV) from is, P < 0.005 ) to identify regions showing extreme deviations from a set of clinically relevant covariates (for example, age and sex) across model predictions. These deviation maps were derived for each indi- a large healthy sample, referred to as a training set. These estimates are vidual in the clinical and the held-out control (HCtest) groups and were then used to quantify the deviations of samples in the test subset, which used for subsequent analyses. We used this approach to be consistent typically consists of cases sampled from the normative demographic with prior work10,14 and because alternative methods, such as the FDR83, range of the training set. rely on adaptive thresholds that can vary between individuals. Applying When using multisite data, scanner- and site-related variability a fixed z threshold allows for an absolute definition of extreme devia- introduce artefactual variance that confounds the results of any subse- tions. Second, to ensure that our findings were not driven by this spe- quent analyses81. These confounding effects present as site-correlated cific choice of threshold, we adapted a threshold-weighted approach30, biases that cannot be explained by biological heterogeneity between as detailed below. Nature Neuroscience | Volume 26 | September 2023 | 1613–1629 1622 Article https://doi.org/10.1038/s41593-023-01404-6 Note that scaling deviation z scores by predictive uncertainty We complemented this analysis with an alternative approach, which means that we are less likely to observe deviations in data regimes does not yield similarly intuitive metrics but which integrates results where model uncertainty is high. This may be a consideration for devia- across a range of thresholds30, thus allowing us to determine the degree tion estimates in older participants, given that our training sample was to which our findings depend on a specific threshold choice. First, for skewed toward younger ages. However, we found no correlation each diagnostic group, we thresholded z maps across the threshold between age and deviation burden ( ρ = −0.004, P = 0.86). range 1.64 < z < 3.10 , in 100 equal log-space increments, separately for positive and negative extrema. Second, we obtained a region- and Evaluation of model performance threshold-specific percentage overlap map quantifying the proportion To assess model generalizability, we used five-fold cross-validation of individuals within each group showing an extreme deviation for any applied to the HCtrain cohort (n = 1,196 HC individuals, 55% male). Spe- given region and threshold. Third, we applied a weighted function to cifically, we partitioned the HCtrain cohort into five folds. Within each penalize less conservative thresholds. Fourth, we obtained a final fold, we trained HBR models on 80% of participants, using age, sex and threshold-weighted overlap map by taking the area under the curve of site as covariates, withholding 20% of the participants for estimating the cumulative histogram of the proportion of participants at each generalization performance. This procedure was repeated five times threshold for each region across participants. We used the following so that regional GMV values for all participants in the HCtrain group were linear weighted function proposed by Seghier and Price30, which ensures predicted once. As above, we estimated deviation z maps for every indi- that overlap map values range between 0 and 1, with 1 indicating an vidual and identified extreme deviations. This procedure is standard effect in a region is present in each participant at each threshold: in machine learning and provides approximately unbiased estimates of the true generalization ability. th − zmin Wth = 2 × , For each fold in the cross-validation, we assessed model fit for each zmax − zmin brain region by evaluating three performance metrics: (1) explained variance, (2) the mean standardized log-loss and (3) the standard- where zmax and zmin respectively correspond to the maximum ized mean-squared error29, as shown in Supplementary Fig. 3. We also and minimum z-value thresholds, and th is the given statistical thresh- evaluated the model’s efficacy in partitioning site-related variance in old. We then calculated the difference between the threshold-weighted the data using linear support vector machines (LSVMs). Specifically, percentage overlaps between cases and the held-out HC cohort, result- we used a series of one-versus-all LSVMs (with default slack parameter ing in an overlap difference map for each disorder. We performed infer- of 1) trained separately on the z maps from the HCtrain and the HCtest ence on these differences using the same group-based permutation subsets to classify scan sites. For each site, we ran a two-fold LSVM procedure described above with statistically significance defined using classifier to obtain the mean balanced accuracy score for the given a threshold of PFDR < 0.05, two tailed. site. Here, a balanced accuracy at chance level (50%) indicated that the resulting deviations were not contaminated by residual site effects, Mapping neural heterogeneity at the circuit level as confirmed in Supplementary Table 2. To assess whether individual Having investigated neural heterogeneity at the level of individual variations in scan quality affected the deviation z maps, we calculated brain regions, we next evaluated whether regions showing extreme the Pearson’s correlation between the total number of extrema and deviations in individual cases are functionally coupled to common CAT12’s IQR rating in each group. areas, which we term circuit-level heterogeneity (Fig. 1, middle). Where prior work has identified spatial correlations between GMV difference Characterizing regional heterogeneity of extreme deviations maps and canonical functional networks estimated at the level of group We characterized the heterogeneity of case-specific thresholded means86, here we leverage normative modeling to move beyond such deviation maps for each disorder using a nonparametric approach. means to investigate whether person-specific GMV deviations aggre- Specifically, for each clinical group and the HCtest cohort, we computed gate within common functional circuits. Evidence for such aggregation the proportion of individuals in each group showing an extreme devi- offers a plausible explanation for phenotypic similarities between ation within each region, resulting in group-specific overlap maps individuals assigned the same diagnosis, despite reports of extreme estimated separately for positive and negative extrema (Fig. 2f). We heterogeneity in deviation location at the regional level10,12–18. We define used the proportion of individuals rather than raw counts to account the functional circuitry of a brain region showing an extreme deviation for sample size differences between groups. Next, we subtracted the as the set of brain areas that show significant FC with that region in HCtest overlap map from each disorder’s overlap map, resulting in an an independent cohort of healthy individuals. To map this circuitry, overlap difference map for each disorder (Fig. 2g). We then permuted we adapted elements of the lesion network mapping methodology group labels (that is, HC and case) and repeated the procedure 10,000 used extensively in neurological disorders22–24 for use with normative times to derive an empirical distribution of overlap difference maps modeling, as outlined in the following. under the null hypothesis of random group assignment (Extended Data Fig. 3a–c). For each brain region, we obtained P values as the pro- fMRI data and processing portion of null values that exceeded the observed difference (Extended The functional circuitry of deviant regions was mapped using an Data Fig. 3d). The tails of the null distribution (that is, values associated independent cohort of 150 resting-state functional MRI (rs-fMRI) with P < 0.10) were approximated using a generalized Pareto distribu- scans (71 males, age 21–35 years) from the S900 release of the Human tion84, as implemented in the Permutation Analysis of Linear Models Connectome Project87 (HCP). These individuals corresponded to the software package85 (alpha116), to allow inference at arbitrarily high 150 people with the lowest total head motion, as estimated using frame- levels of precision. Statistically significant effects were identified using wise displacement88, from the broader set of 282 unrelated participants an FDR-corrected83 threshold of PFDR < 0.05, two tailed. with the same fMRI reconstruction in the S900 release. We focused on this subsample to minimize the effects of head motion in our FC maps Threshold-weighted deviation mapping and to minimize computational burden due to the large number of The procedure to characterize the regional heterogeneity of extreme analyses required. Mapping deviation-related functional circuity in deviations above used an arbitrary threshold ( z < |2.6|) on each indi- an independent cohort is essential because we seek to understand vidual z map. The practical benefit of this approach is that it allows the circuit context of deviant regions under normative conditions. calculation of intuitive metrics, such as the proportion of cases within Our approach aligns precisely with the logic used in lesion network a group showing a supra-threshold deviation within a given brain region. mapping of neurological patients22,23. Nature Neuroscience | Volume 26 | September 2023 | 1613–1629 1623 Article https://doi.org/10.1038/s41593-023-01404-6 The rs-fMRI data for each participant consists of four runs (rfMRI_ approximation of the tail of the null distribution84 and threshold-free REST1_LR, rfMRI_REST1_RL, rfMRI_REST2_LR and rfMRI_REST2_RL) cluster enhancement (TFCE)103, run separately for the cortical and acquired at two different sessions (REST1 and REST2) using two dif- subcortical areas, with a threshold of PFWE < 0.025, family-wise error ferent directions of phase coding (LR: left to right, and RL: right to corrected (that is, 0.05/2 to account for cortical and subcortical anal- left). Here, we mapped deviation-related FC using the rfMRI_REST1_LR yses). This procedure was repeated for each region in which at least data. All rs-fMRI data were obtained using a 32-channel Siemens 3T one extreme deviation across individuals was observed (Fig. 3b). connectome-Skyra scanner. The imaging parameters for rs-fMRI were To quantify circuit-level overlap, we developed an approach that as follows: repetition time 720 ms, echo time 33.1 ms, flip angle 52°, field parallels traditional lesion network mapping methodology22,23 but of view 208 × 180 mm2, matrix 104 × 90, slice number 72, slice thickness which allows statistical inference on observed case–control differences 2 mm, voxel size 2 × 2 × 2 mm3, multiband factor 8 and 1,200 volumes. in overlap scores at each region. Specifically, we thresholded each Participants were required to stay awake, relaxed and to keep their eyes deviant-related FC map by applying TFCE to the grayordinate maps, as open and fixed on a bright cross-hair projected on a dark background. described above. We then classified an area in our 1,032-region parcel- A detailed description of the HCP data is available in previous work87. lation as showing significant FC with the seed if more than 50% of its The rs-fMRI data underwent the HCP’s minimal preprocessing grayordinates survived the TFCE threshold. We repeated our analysis pipeline89, which includes gradient-nonlinearity-induced distortion, using a mapping threshold of 75% to ensure that our findings were motion correction to the single-band reference image using FMRIB’s not driven by this specific choice (Supplementary Figs. 9–11). Note, Linear Image Registration Tool (FLIRT), echo-planar imaging (EPI) the fslr32k projection of the Schaefer cortical parcellation is missing image distortion correction using TOPUP90; registration into standard two parcels (533 and 903, for more details, see ref. 104) so these were space using a customized boundary-based-registration algorithm excluded from the analysis. and single-step spline interpolation using all transforms; intensity We mapped FC at the grayordinate level and thresholded the data normalization and bias field removal to resample the original EPI into in this way to leverage the superior statistical sensitivity of TFCE and Montreal Neurological Institute space. Minimal high-pass filtering was to map the spatial architecture of the seed-related FC patterns more applied with a cutoff of 2,000 ms. Artifacts were then removed using accurately. Our procedure resulted in a binary map representing the independent component analysis X-noiseifier (ICA-FIX)91. This involves specific brain regions that comprise the functional circuitry of each employing an automatic classifier, specifically trained for HCP data, seed (Fig. 3c). Then, for each individual, we took the union of the binary to identify ICA components due to measurement noise, additional FC maps across that person’s set of extrema, resulting in a single map motion or physiological artifacts such as cardiac pulsation and respi- identifying areas showing significant FC to any extreme deviation ration. Next, the volume time series were mapped into the standard expressed by the person (Fig. 3d). Separate union maps were obtained CIFTI grayordinate space87 and smoothed to 2 mm full width at half for positive and negative deviations. The overlap of these union maps maximum (where the smoothing was on the surface for the cortex and was then taken for each clinical group and the HCtest sample, represent- in volume space for subcortex). This approach results in a standard set ing, for each region, the proportion of individuals in that group for of grayordinates in every participant that encompasses surface vertex whom significant deviant-related FC was identified. We refer to this data and subcortical volume voxel data. The mean GM signal was then image as a circuit-level overlap map (Fig. 3e). removed from each grayordinate’s time series to remove residual wide- Here we mapped the network context of GMV deviations using spread signal deflections that were not removed by ICA-FIX92–94. Global FC estimates of normative brain architecture mapped in an independ- signal regression (GSR) has been criticized because the global signal, ent sample, following the approach typically used in lesion network by construction, contains neural signal and its removal shifts the distri- mapping22,23. It is possible that patterns of FC are altered in diagnosed bution of correlations so that it is approximately centered on zero95,96. individuals, and that the underlying network architecture differs from Nonetheless, it is a highly effective denoising procedure, successfully normative expectations. The net effect of such alterations would be removing contributions to fMRI signals from head motion and respira- to affect coupling between deviant and other areas, and thus our tory variations97,98. Invasive recordings in animal models indicate that normative benchmark still offers an important reference point for GSR increases the correspondence between hemodynamic signals and understanding the distributed effects of regional deviations. Further neuronal activity99, and FC estimates obtained after GSR in humans investigation of how altered patterns of FC may moderate these rela- are more predictive of a wide range of behavioral measures than FC tionships remains an important topic for future investigation. Alterna- estimates obtained without GSR100. GSR also dramatically enhances tive approaches could involve mapping the network context of GMV the anatomical specificity of seed-based correlation maps101, which deviations with estimates of structural connectivity obtained with is essential for precise circuit mapping. We therefore used GSR in the diffusion MRI105,106 or estimates of interregional structural covariance107. present analysis to enable direct comparison with the vast majority of seed-based FC studies in the literature. Alternative strategies for Characterizing circuit-level heterogeneity denoising (for example, refs. 92,94) may address the limitations of We evaluated group differences in the circuit-level overlap maps of GSR and may provide more refined network maps. each clinical group relative to the HCtest group in two ways (for a sche- matic overview, see Extended Data Fig. 3). First, following the analysis Mapping the functional circuitry of extreme deviations of regional overlaps, we computed the difference in the case and con- We used seed-based FC analysis to map the functional circuitry of each trol overlap maps ( Δ overlap map, Fig. 3e) and evaluated the statistical region showing an extreme deviation in any individual within the significance of these differences with respect to an empirical null dis- clinical and HCtest groups. Specifically, each deviant region was used tribution generated by permuting the group labels of the as a seed, from which the average time course was extracted for each individual-specific FC union maps 10,000 times, with a generalized individual in the HCP150 sample (Fig. 4b). These seed time courses were Pareto tail approximation84. Statistically significant differences in then correlated with all other brain grayordinates and the resulting deviation FC overlap were identified using a threshold of PFDR < 0.05, correlation maps were subjected to Fisher’s r-to-z transformation102. two tailed (Extended Data Fig. 3a–d). The transformed correlation maps for each HCP150 participant were This group-based permutation test identifies regions where then aggregated using a one-way t-test at each grayordinate, as imple- the overlap in deviant-related functional circuitry differs between mented in Permutation Analysis of Linear Models. Thresholded maps cases and controls, but it does not preserve group differences in the representing the functional circuitry of the seed were obtained using total number of deviations. It therefore cannot distinguish whether permutation testing (500 permutations) with a generalized Pareto any observed differences arise from a preferential aggregation of Nature Neuroscience | Volume 26 | September 2023 | 1613–1629 1624 Article https://doi.org/10.1038/s41593-023-01404-6 deviations within specific circuits in one group relative to the other, or Evaluating heterogeneity across spatial scales whether the group differences are driven by variations in total devia- To examine heterogeneity across spatial scales, we subtracted the tion burden. Note that differences in deviation burden should not be regional Δ overlap map (Fig. 2g) from the circuit-level Δ overlap thought of as a confound in this context, since they are an intrinsic map (Fig. 3f) for each clinical group separately and evaluated the mag- feature of the disease that will have real phenotypic consequences. nitude of the resulting differences using the group-based permutation Nonetheless, it is important to determine which findings might be testing. The Δ overlap maps quantify the difference in overlap observed driven by differences in deviation burden compared with a preferential at each region between each diagnostic group and controls. Thus, by accumulation of pathology within a specific circuit for a given disorder. subtracting the two Δ overlap maps, we directly test whether case– To disentangle these possibilities, we considered a second, spatial control differences in overlap are greater at the circuit compared with null model that preserves group differences in deviation burden. Spe- the regional level. cifically, we generated an ensemble of null cortical deviation maps for each individual in the test data by spatially rotating their empirical, Mapping neural heterogeneity at the network unthresholded deviation maps (Extended Data Fig. 3e) using Hungarian level spherical spin tests108,109 (Extended Data Fig. 3f) and then thresholding The above procedure offers a means for understanding heterogeneity these rotated maps in the same way as the observed data (that is, at the level of neural circuits that show strong FC with a deviant region. z > |2.6|; Extended Data Fig. 3g). The Hungarian method was used as However, it is still possible for some pairs of brain regions to be affili- it preserves the original values and spatial autocorrelation of the origi- ated with the same extended functional network despite being weakly nal map, which ensures that the surrogate data yields the same number coupled themselves (Fig. 1, right). We therefore sought to characterize of extrema as originally observed. This procedure could only be applied neural heterogeneity at the level of the broader functional networks to the cortex, which is topologically comparable to a sphere. The choice within which a given deviant region may be embedded. of an appropriate null model for subcortical areas is more complicated. While model-based procedures (for example, ref. 110) can generate Assigning regions to networks surrogates with comparable spatial autocorrelation, they require We assigned each cortical region to one of seven canonical functional parameter tuning, can show variable fits across individuals and do not cortical networks using a well-validated network parcellation27,31 and preserve the exact values of the original data, making thresholding assigned each subcortical region to either the medial temporal lobe challenging. We therefore randomly shuffled deviation values across (amygdala and hippocampus), thalamus or basal ganglia (nucleus subcortical areas before thresholding them. Note that this null model accumbens, globus pallidus, putamen and caudate nucleus), as done is more lenient than a spatially constrained test since the model pre- previously28, resulting in a total of ten distinct functional networks serves fewer features of the data. As we found no significant differences (Fig. 4a–c). The subcortical regions were assigned to coarse ana- using this test, our results will not change under a more stringent, tomical areas that were not aligned with the finely mapped corti- spatially constrained null model. cal function networks because the correspondence between these We then obtained individual-specific surrogate FC union maps subcortical nuclei, as parcellated here, and canonical cortical func- using the same procedure described above (Extended Data Fig. 3h), tional networks has not been extensively investigated. An alterna- and calculated surrogate within-group overlap maps (Extended Data tive network assignment may influence network-level subcortical Fig. 3i) and between-group overlap difference maps (Extended Data overlap values. Fig. 3j). This procedure was repeated 10,000 times to generate a null distribution of overlap difference maps for each disorder, again using Characterizing network-level heterogeneity the generalized Pareto tail approximation (Extended Data Fig. 3k). To characterize network-level heterogeneity, we examined the degree Statistically significant differences were identified using a threshold to which GMV deviations aggregated within each network. For each net- of PFDR < 0.05, two tailed. work, we estimated the proportion of individuals within each diagnostic To summarize, group differences in circuit-level overlap were group that showed at least one extreme deviation in a region assigned assessed with respect to two null models, one based on permutation to that network, separately for positive and negative deviations of group labels and one based on spatial permutation of individual (Fig. 4d). We then computed case–control differences in overlap pro- deviation maps. The group-based permutation test identifies regions portions for each network (Fig. 4e). showing differences in circuit-level overlap regardless of deviation As in the circuit-level analysis, these observed group differences burden. The spatial permutation test can be used to assess the degree in network-level overlap were evaluated with respect to two comple- to which such differences are driven by variations in deviation burden. mentary null models. The first involved a group-based randomization More specifically, if a given brain region shows a significant difference test in which a null distribution was generated by permuting group for both tests, we have evidence to indicate that extreme deviations labels 10,000 times (Extended Data Fig. 3a–d). The second null model within the disorder preferentially aggregate within the functional cir- used the 10,000 spatially rotated maps generated in the circuit-level cuitry of that region. If, on the other hand, a region shows a difference analysis to obtain surrogate estimates of group differences in with the group-based permutation test but not the spatial permuta- network-level overlap while preserving variation in deviation burden tion test, then the observed group differences are consistent with (Extended Data Fig. 3e–g,l–o). In both cases, the tail of the null distri- expectations from randomly selecting of the same number of seeds bution was approximated using a generalized Pareto distribution and in each group, indicating that the effects observed under group-based we used a significance threshold of PFDR < 0.05, two tailed. As in the permutation can be attributed to the fact that one group has a higher circuit-level analysis, these two null models allowed us to distinguish deviation burden than the other. Note that this scenario still yields preferential network involvement from the effects of total deviation important information about the extent of circuit-level heterogene- burden. Note that comparisons in overlap values between networks ity in the disorder, since we generally expect cases to have a higher are complicated by differences in network size. Our analysis therefore deviation burden and any resulting circuit-level overlap will still have focuses primarily on comparisons between cases and controls within phenotypic consequences regardless of whether such overlaps are networks since the control data provide a critical normative bench- explained by deviation burden or not. The spatial permutation test mark against which cases can be compared. We repeated the same simply offers additional insights into the potential mechanisms that analyses using a 20-network parcellation (17 cortical networks and may drive group differences in circuit-level overlap. 3 subcortical regions). Nature Neuroscience | Volume 26 | September 2023 | 1613–1629 1625 Article https://doi.org/10.1038/s41593-023-01404-6 Reporting summary 17. Zabihi, M. et al. Dissecting the heterogeneous cortical anatomy Further information on research design is available in the Nature Port- of autism spectrum disorder using normative models. Biol. folio Reporting Summary linked to this article. Psychiatry Cogn. Neurosci. Neuroimaging 4, 567–578 (2019). 18. Bayer, J. M. M. et al. Accommodating site variation in neuroimaging Online content data using normative and hierarchical Bayesian models. Any methods, additional references, Nature Portfolio reporting sum- NeuroImage https://doi.org/10.1016/j.neuroimage.2022.119699 maries, source data, extended data, supplementary information, (2021). acknowledgements, peer review information; details of author contri- 19. Carrera, E. & Tononi, G. Diaschisis: past, present, future. Brain 137, butions and competing interests; and statements of data and code avail- 2408–2422 (2014). ability are available at https://doi.org/10.1038/s41593-023-01404-6. 20. Fornito, A., Zalesky, A. & Breakspear, M. The connectomics of brain disorders. Nat. Rev. Neurosci. 16, 159–172 (2015). References 21. von Monakow, C. Die Lokalisation im Grosshirn und der Abbau der 1. Opel, N. et al. Cross-disorder analysis of brain structural Funktion durch kortikale Herde (JF Bergmann, 1914). abnormalities in six major psychiatric disorders: a secondary 22. Boes, A. D. et al. Network localization of neurological symptoms analysis of mega- and meta-analytical findings from the ENIGMA from focal brain lesions. Brain 138, 3061–3075 (2015). Consortium. Biol. Psychiatry 88, 678–686 (2020). 23. Fox, M. D. Mapping symptoms to brain networks with the human 2. Hibar, D. P. et al. Cortical abnormalities in bipolar disorder: an MRI connectome. N. Engl. J. Med. 379, 2237–2245 (2018). analysis of 6503 individuals from the ENIGMA Bipolar Disorder 24. Joutsa, J., Corp, D. T. & Fox, M. D. Lesion network mapping Working Group. Mol. Psychiatry 23, 932–942 (2018). for symptom localization: recent developments and future 3. Boedhoe, P. S. W. et al. An empirical comparison of meta- and directions. Curr. Opin. Neurol. 35, 453–459 (2022). mega-analysis with data from the ENIGMA Obsessive- 25. Gaser, C. & Dahnke, R. CAT—a computational anatomy toolbox for Compulsive Disorder Working Group. Front. Neuroinform. 12, the analysis of structural MRI data. Hbm 2016, 336–348 (2016). 1–8 (2019). 26. Ashburner, J. & Friston, K. J. Voxel-based morphometry—the 4. Hoogman, M. et al. Brain imaging of the cortex in ADHD: methods. NeuroImage 11, 805–821 (2000). a coordinated analysis of large-scale clinical and population- 27. Schaefer, A. et al. Local–global parcellation of the human based samples. Am. J. Psychiatry 176, 531–542 (2019). cerebral cortex from intrinsic functional connectivity MRI. Cereb. 5. van Erp, T. G. M. et al. Cortical brain abnormalities in 4474 Cortex 28, 3095–3114 (2018). individuals with schizophrenia and 5098 control subjects via 28. Tian, Y., Margulies, D. S., Breakspear, M. & Zalesky, A. Topographic the Enhancing Neuro Imaging Genetics Through Meta Analysis organization of the human subcortex unveiled with functional (ENIGMA) Consortium. Biol. Psychiatry 84, 644–654 (2018). connectivity gradients. Nat. Neurosci. 23, 1421–1432 (2020). 6. Schmaal, L. et al. Cortical abnormalities in adults and adolescents 29. Kia, S. M. et al. Hierarchical Bayesian regression for multi-site with major depression based on brain scans from 20 cohorts normative modeling of neuroimaging data. Lect. Notes Comput. worldwide in the ENIGMA Major Depressive Disorder Working Sci. 12267 LNCS, 699–709 (2020). Group. Mol. Psychiatry 22, 900–909 (2017). 30. Seghier, M. L. & Price, C. J. Visualising inter-subject variability 7. Van Rooij, D. et al. Cortical and subcortical brain morphometry in fMRI using threshold-weighted overlap maps. Sci. Rep. 6, differences between patients with autism spectrum disorder and 1–13 (2016). healthy individuals across the lifespan: results from the ENIGMA 31. Yeo, B. T. et al. The organization of the human cerebral cortex ASD Working Group. Am. J. Psychiatry 175, 359–369 (2018). estimated by intrinsic functional connectivity. J. Neurophysiol. 8. Fried, E. I. & Nesse, R. M. Depression is not a consistent syndrome: 106, 1125–1165 (2011). an investigation of unique symptom patterns in the STAR*D study. 32. Insel, T. R. Faulty circuits. Sci. Am. 302, 44–52 (2010). J. Affect. Disord. 172, 96–102 (2015). 33. Cortese, S. et al. Toward systems neuroscience of ADHD: 9. Young, G., Lareau, C. & Pierre, B. One quintillion ways to have a meta-analysis of 55 fMRI studies. Am. J. Psychiatry 169, PTSD comorbidity: recommendations for the disordered DSM-5. 1038–1055 (2012). Psychol. Inj. Law 7, 61–74 (2014). 34. Fornito, A., Zalesky, A., Pantelis, C. & Bullmore, E. T. 10. Wolfers, T. et al. Mapping the heterogeneous phenotype of Schizophrenia, neuroimaging and connectomics. NeuroImage schizophrenia and bipolar disorder using normative models. 62, 2296–2314 (2012). JAMA Psychiatry 75, 1146–1155 (2018). 35. Goodkind, M. et al. Identification of a common neurobiological 11. Marquand, A. F., Rezek, I., Buitelaar, J. & Beckmann, C. F. substrate for mental Illness. JAMA Psychiatry 72, 305–315 (2015). Understanding heterogeneity in clinical cohorts using normative 36. McTeague, L. M. et al. Identification of common neural circuit models: beyond case–control studies. Biol. Psychiatry 80, disruptions in cognitive control across psychiatric disorders. 552–561 (2016). Am. J. Psychiatry 174, 676–685 (2017). 12. Rutherford, S. et al. The normative modeling framework for 37. McTeague, L. M. et al. Identification of common neural circuit computational psychiatry. Nat. Protoc. 17, 1711–1734 (2022). disruptions in emotional processing across psychiatric disorders. 13. Wolfers, T. et al. Individual differences v. the average patient: Am. J. Psychiatry 177, 411–421 (2020). mapping the heterogeneity in ADHD using normative models. 38. Ham, T., Leff, A., de Boissezon, X., Joffe, A. & Sharp, D. J. Psychol. Med. 50, 314–323 (2019). Cognitive control and the salience network: an investigation 14. Wolfers, T. et al. Replicating extensive brain structural of error processing and effective connectivity. J. Neurosci. 33, heterogeneity in individuals with schizophrenia and bipolar 7091–7098 (2013). disorder. Hum. Brain Mapp. 42, 2546–2555 (2021). 39. Chong, J. S. X., Ng, G. J. P., Lee, S. C. & Zhou, J. Salience 15. Lv, J. et al. Individual deviations from normative models of network connectivity in the insula is associated with individual brain structure in a large cross-sectional schizophrenia cohort. differences in interoceptive accuracy. Brain Struct. Funct. 222, Mol. Psychiatry 26, 3512–3523 (2021). 1635–1644 (2017). 16. Floris, D. L. et al. Atypical brain asymmetry in autism—a candidate 40. Menon, V. & Uddin, L. Q. Saliency, switching, attention and for clinically meaningful stratification. Biol. Psychiatry. Cogn. control: a network model of insula function. Brain Struct. Funct. Neurosci. Neuroimaging 6, 802–812 (2021). 214, 655–667 (2010). Nature Neuroscience | Volume 26 | September 2023 | 1613–1629 1626 Article https://doi.org/10.1038/s41593-023-01404-6 41. Uddin, L. Q. et al. Salience network-based classification 62. Miller, E. K. The prefrontal cortex: categories, concepts, and and prediction of symptom severity in children with autism. cognitive control. Nat. Rev. Neurosci. 1, 59–65 (2000). JAMA Psychiatry 70, 869–879 (2013). 63. Chopra, S. et al. Differentiating the effect of antipsychotic 42. Gürsel, D. A., Avram, M., Sorg, C., Brandl, F. & Koch, K. medication and illness on brain volume reductions in Frontoparietal areas link impairments of large-scale intrinsic first-episode psychosis: a longitudinal, randomised, triple-blind, brain networks with aberrant fronto-striatal interactions in OCD: placebo-controlled MRI study. Neuropsychopharmacology 46, a meta-analysis of resting-state functional connectivity. Neurosci. 1494–1501 (2021). Biobehav. Rev. 87, 151–160 (2018). 64. Winter, N. R. et al. Quantifying deviations of brain structure and 43. Koch, S. B. J. et al. Aberrant resting‐state brain activity in function in major depressive disorder across neuroimaging posttraumatic stress disorder: a meta‐analysis and systematic modalities. JAMA Psychiatry 79, 879 (2022). review. Depress. Anxiety 33, 592–605 (2016). 65. Rutherford, S. et al. Charting brain growth and aging at high 44. Wang, X., Zhang, W., Sun, Y., Hu, M. & Chen, A. Aberrant spatial precision. elife 11, e72904 (2022). intra-salience network dynamic functional connectivity 66. Bethlehem, R. A. I. et al. Brain charts for the human lifespan. impairs large-scale network interactions in schizophrenia. Nature 604, 525–533 (2022). Neuropsychologia 93, 262–270 (2016). 67. Tiego, J. et al. Precision behavioral phenotyping as a strategy 45. Shanmugan, S. et al. Common and dissociable mechanisms of for uncovering the biological correlates of psychopathology. executive system dysfunction across psychiatric disorders in Nat. Ment. Health 1, 304–315 (2023). youth. Am. J. Psychiatry 173, 517–526 (2016). 68. Kraus, B. et al. Insights from personalized models of brain and 46. Caspi, A. et al. The p factor: one general psychopathology factor behavior for identifying biomarkers in psychiatry. Neurosci. in the structure of psychiatric disorders? Clin. Psychol Sci. 2, Biobehav. Rev. 152, 105259 (2023). 119–137 (2014). 69. Kotov, R. et al. The hierarchical taxonomy of psychopathology 47. Ahmari, S. E. & Rauch, S. L. The prefrontal cortex and OCD. (HiTOP): a dimensional alternative to traditional nosologies. Neuropsychopharmacology https://doi.org/10.1038/s41386-021- J. Abnorm. Psychol. 126, 454–477 (2017). 01130-2 (2021). 70. American Psychiatric Association. Diagnostic and Statistical 48. Pizzagalli, D. A. & Roberts, A. C. Prefrontal cortex and depression. Manual of Mental Disorders, Fifth edition (American Psychiatric Neuropsychopharmacology https://doi.org/10.1038/s41386-021- Association, 2013). 01101-7 (2021). 71. The ICD-10 Classification of Mental and Behavioural Disorders 49. Zhou, Y., Fan, L., Qiu, C. & Jiang, T. Prefrontal cortex and the (World Health Organization, Geneva, 1992). dysconnectivity hypothesis of schizophrenia. Neurosci. Bull. 31, 72. Bellon, E. M. et al. MR artifacts: a review. Am. J. Roentgenol. 147, 207–219 (2015). 1271–1281 (1986). 50. Kenwood, M. M., Kalin, N. H. & Barbas, H. The prefrontal cortex, 73. Zaitsev, M., Maclaren, J. & Herbst, M. Motion artifacts in MRI: a pathological anxiety, and anxiety disorders. Neuropsycho­ complex problem with many partial solutions. J. Magn. Reson. pharmacology https://doi.org/10.1038/s41386-021-01109-z (2021). Imaging 42, 887–901 (2015). 51. Koenigs, M. & Grafman, J. The functional neuroanatomy of 74. Gaser, C., Dahnke, R., Thompson, P. M., Kurth, F., & Luders, depression: distinct roles for ventromedial and dorsolateral E. CAT-a computational anatomy toolbox for the analysis of prefrontal cortex. Behav. Brain Res. 201, 239–243 (2009). structural MRI data. BioRxiv https://doi.org/10.1101/2022.06. 52. O’Reardon, J. P. et al. Efficacy and safety of transcranial magnetic 11.495736 (2022). stimulation in the acute treatment of major depression: 75. SPM12. Univ. College London http://www.fil.ion.ucl.ac.uk/spm/ a multisite randomized controlled trial. Biol. Psychiatry 62, software/spm12 (2020). 1208–1216 (2007). 76. The MathWorks Inc. MATLAB version 9.4.0.813654 (R2018a). 53. Nguyen, T. D., Hieronymus, F., Lorentzen, R., McGirr, A. & (The MathWorks Inc., 2018). Østergaard, S. D. The efficacy of repetitive transcranial magnetic 77. Ashburner, J. A fast diffeomorphic image registration algorithm. stimulation (rTMS) for bipolar depression: a systematic review and NeuroImage 38, 95–113 (2007). meta-analysis. J. Affect. Disord. 279, 250–255 (2021). 78. Kong, R. et al. Individual-specific areal-level parcellations improve 54. Padmanabhan, J. L. et al. A human depression circuit derived from functional connectivity prediction of behavior. Cereb. Cortex 31, focal brain lesions. Biol. Psychiatry 86, 749–758 (2019). 4477–4500 (2021). 55. Siddiqi, S. H. et al. Brain stimulation and brain lesions converge 79. get_totals.m. Univ. College London http://www0.cs.ucl.ac.uk/ on common causal circuits in neuropsychiatric disease. Nat. Hum. staff/gridgway/vbm/get_totals.m Behav. 5, 1707–1716 (2021). 80. amarquand. GitHub https://github.com/amarquand/PCNtoolkit.git 56. Siddiqi, S. H., Kording, K. P., Parvizi, J. & Fox, M. D. Causal mapping 81. Focke, N. K. et al. Multi-site voxel-based morphometry—not quite of human brain function. Nat. Rev. Neurosci. 23, 361–375 (2022). there yet. NeuroImage 56, 1164–1170 (2011). 57. Buckner, R. L. et al. Cortical hubs revealed by intrinsic functional 82. Box, G. E. P. & Cox, D. R. An analysis of transformations. J. R. Stat. connectivity: mapping, assessment of stability, and relation to Soc. B 26, 211–243 (1964). Alzheimer’s disease. J. Neurosci. 29, 1860–1873 (2009). 83. Benjamini, Y. & Hochberg, Y. Controlling the false discovery rate: 58. van den Heuvel, M. P. & Sporns, O. Network hubs in the human a practical and powerful approach to multiple testing. J. R. Stat. brain. Trends Cogn. Sci. 17, 683–696 (2013). Soc. B 57, 289–300 (1995). 59. Arnatkeviciute, A. et al. Genetic influences on hub connectivity of 84. Winkler, A. M., Ridgway, G. R., Douaud, G., Nichols, T. E. & Smith, the human connectome. Nat. Commun. 12, 4237 (2021). S. M. Faster permutation inference in brain imaging. NeuroImage 60. Schumann, C. M. et al. Longitudinal magnetic resonance imaging 141, 502–516 (2016). study of cortical development through early childhood in autism. 85. Winkler, A. M., Ridgway, G. R., Webster, M. A., Smith, S. M. & J. Neurosci. 30, 4419–4427 (2010). Nichols, T. E. Permutation inference for the general linear model. 61. Palomero-Gallagher, N. et al. Human pregenual anterior cingulate NeuroImage 92, 381–397 (2014). cortex: structural, functional, and connectional heterogeneity. 86. Vanasse, T. J. et al. Brain pathology recapitulates physiology: Cereb. Cortex 29, 2552–2574 (2019). a network meta-analysis. Commun. Biol. 4, 301 (2021). Nature Neuroscience | Volume 26 | September 2023 | 1613–1629 1627 Article https://doi.org/10.1038/s41593-023-01404-6 87. Van Essen, D. C. et al. The WU-Minn human connectome project: 102. Silver, N. C. & Dunlap, W. P. Averaging correlation coefficients: an overview. Neuroimage 80, 62–79 (2013). should Fisher’s z transformation be used? J. Appl. Psychol. 72, 88. Jenkinson, M., Bannister, P., Brady, M. & Smith, S. Improved 146–148 (1987). optimization for the robust and accurate linear registration and 103. Smith, S. M. & Nichols, T. E. Threshold-free cluster enhancement: motion correction of brain images. NeuroImage 17, 825–841 (2002). addressing problems of smoothing, threshold dependence 89. Glasser, M. F. et al. The minimal preprocessing pipelines for the and localisation in cluster inference. NeuroImage 44, 83–98 Human Connectome Project. NeuroImage 80, 105–124 (2013). (2009). 90. Andersson, J. L. R., Skare, S. & Ashburner, J. How to correct 104. Missing parcel in Schaefer 1000 conte69. GitHub susceptibility distortions in spin-echo echo-planar images: https://github.com/ThomasYeoLab/CBIG/issues/10 (2020). application to diffusion tensor imaging. NeuroImage 20, 105. Kuceyeski, A., Maruta, J., Relkin, N. & Raj, A. The network 870–888 (2003). modification (NeMo) tool: elucidating the effect of white 91. Griffanti, L. et al. ICA-based artefact removal and accelerated matter integrity changes on cortical and subcortical structural fMRI acquisition for improved resting state network imaging. connectivity. Brain Connect. 3, 451–463 (2013). NeuroImage 95, 232–247 (2014). 106. Sperber, C., Griffis, J. & Kasties, V. Indirect structural 92. Aquino, K. M., Fulcher, B. D., Parkes, L., Sabaroedin, K. & Fornito, disconnection-symptom mapping. Brain Struct. Funct. 227, A. Identifying and removing widespread signal deflections from 3129–3144 (2022). fMRI data: rethinking the global signal regression problem. 107. Alexander-Bloch, A., Giedd, J. N. & Bullmore, E. Imaging structural NeuroImage 212, 116614–116614 (2020). co-variance between human brain regions. Nat. Rev. Neurosci. 14, 93. Burgess, G. C. et al. Evaluation of denoising strategies to address 322–336 (2013). motion-correlated artifacts in resting-state functional magnetic 108. Kuhn, H. W. The Hungarian method for the assignment problem. resonance imaging data from the Human Connectome Project. Nav. Res. Logist. Q. 2, 83–97 (1955). Brain Connect. 6, 669–680 (2016). 109. Markello, R. D. & Misic, B. Comparing spatial null models for brain 94. Glasser, M. F. et al. Using temporal ICA to selectively remove maps. NeuroImage 236, 118052–118052 (2021). global noise while preserving global signal in functional MRI data. 110. Burt, J. B., Helmer, M., Shinn, M., Anticevic, A. & Murray, J. D. NeuroImage 181, 692–717 (2018). Generative modeling of brain maps with spatial autocorrelation. 95. Chang, C. & Glover, G. H. Effects of model-based physiological NeuroImage 220, 117038–117038 (2020). noise correction on default mode network anti-correlations and correlations. NeuroImage 47, 1448–1459 (2009). Publisher’s note Springer Nature remains neutral with 96. Saad, Z. S. et al. Trouble at rest: how correlation patterns regard to jurisdictional claims in published maps and and group differences become distorted after global signal institutional affiliations. regression. Brain Connect. 2, 25–32 (2012). 97. Parkes, L., Fulcher, B., Yücel, M. & Fornito, A. An evaluation of the Open Access This article is licensed under a Creative Commons efficacy, reliability, and sensitivity of motion correction strategies Attribution 4.0 International License, which permits use, sharing, for resting-state functional MRI. NeuroImage 171, 415–436 (2018). adaptation, distribution and reproduction in any medium or format, 98. Power, J. D. et al. Methods to detect, characterize, and remove as long as you give appropriate credit to the original author(s) and the motion artifact in resting state fMRI. NeuroImage 84, 320–341 (2014). source, provide a link to the Creative Commons license, and indicate 99. Matsui, T., Murakami, T. & Ohki, K. Transient neuronal if changes were made. The images or other third party material in this coactivations embedded in globally propagating waves underlie article are included in the article’s Creative Commons license, unless resting-state functional connectivity. Proc. Natl Acad. Sci. USA indicated otherwise in a credit line to the material. If material is not 113, 6556–6561 (2016). included in the article’s Creative Commons license and your intended 100. Li, J. et al. Global signal regression strengthens association use is not permitted by statutory regulation or exceeds the permitted between resting-state functional connectivity and behavior. use, you will need to obtain permission directly from the copyright NeuroImage 196, 126–141 (2019). holder. To view a copy of this license, visit http://creativecommons. 101. Fox, M. D., Zhang, D., Snyder, A. Z. & Raichle, M. E. The global org/licenses/by/4.0/. signal and observed anticorrelated resting state brain networks. J. Neurophysiol. 101, 3270–3283 (2009). © The Author(s) 2023 1 Turner Institute for Brain and Mental Health, School of Psychological Sciences, Monash University, Melbourne, Victoria, Australia. 2Monash Biomedical Imaging, Monash University, Melbourne, Victoria, Australia. 3Department of Bioengineering, School of Engineering and Applied Science, University of Pennsylvania, Philadelphia, PA, USA. 4Department of Psychiatry, Rutgers University, Piscataway, NJ, USA. 5School of Physics, University of Sydney, Sydney, New South Wales, Australia. 6BrainKey Inc, Palo alto, CA, USA. 7Donders Centre for Cognitive Neuroimaging, Radboud University, Nijmegen, the Netherlands. 8Donders Institute for Brain, Cognition and Behaviour, Radboud University, Nijmegen, the Netherlands. 9Department of Psychiatry, University Medical Center Utrecht, Utrecht, the Netherlands. 10Department of Cognitive Science and Artificial Intelligence, Tilburg University, Tilburg, the Netherlands. 11Norwegian Centre for Mental Disorders Research, Division of Mental Health and Addiction, University of Oslo and Oslo University Hospital, Oslo, Norway. 12Department of Psychiatry and Psychotherapy, Tübingen Center for Mental Health (TÜCMH), University of Tübingen, Tübingen, Germany. 13 Department of Psychiatry, Donders Institute of Brain, Cognition and Behaviour, Radboud University Medical Center, Nijmegen, the Netherlands. 14 Department of Human Genetics, Donders Institute of Brain, Cognition and Behaviour, Radboud University Medical Center, Nijmegen, the Netherlands. 15 Department of Cognitive Neuroscience, Radboud University Medical Centre, Nijmegen, the Netherlands. 16Centre for Functional MRI of the Brain, Nuffield Department of Clinical Neurosciences, Wellcome Centre for Integrative Neuroimaging, University of Oxford, Oxford, UK. 17Department of Psychology, University of Oslo, Oslo, Norway. 18KG Jebsen Centre for Neurodevelopmental Disorders, University of Oslo and Oslo University Hospital, Oslo, Norway. 19Melbourne Neuropsychiatry Centre, Department of Psychiatry, The University of Melbourne and Melbourne Health, Melbourne, Victoria, Australia. 20Department of Biomedical Engineering, The University of Melbourne, Melbourne, Victoria, Australia. 21Department of Psychiatry, University of Melbourne, Melbourne, Victoria, Australia. 22Department of Psychiatry, Bellvitge University Hospital, Bellvitge Biomedical Research Institute, Barcelona, Spain. 23Centro de Investigación Biomédica en Red de Salud Mental, Carlos III Health Institute, Madrid, Spain. 24Department of Social Psychology and Nature Neuroscience | Volume 26 | September 2023 | 1613–1629 1628 Article https://doi.org/10.1038/s41593-023-01404-6 Quantitative Psychology, Universitat de Barcelona, Barcelona, Spain. 25Sant Pau Mental Health Research Group, Institut d’Investigació Biomèdica Sant Pau, Hospital de la Santa Creu i Sant Pau, Barcelona, Spain. 26Department of Psychiatry and Forensic Medicine, Universitat Autònoma de Barcelona, Barcelona, Spain. 27QIMR Berghofer Medical Research Institute, Brisbane, Queensland, Australia. 28Australian Characterisation Commons at Scale (ACCS) Project, Monash eResearch Centre, Melbourne, Victoria, Australia. 29Institute for Mental and Physical Health and Clinical Translation School of Medicine, Deakin University, Geelong, Victoria, Australia. 30Orygen, Melbourne, Victoria, Australia. 31Centre for Youth Mental Health, University of Melbourne, Melbourne, Victoria, Australia. 32Florey Institute for Neuroscience and Mental Health, Parkville, Victoria, Australia. 33Department of Neuroimaging, Centre of Neuroimaging Sciences, Institute of Psychiatry, King’s College London, London, UK. 34These authors contributed equally: Andre F. Marquand, Alex Fornito. e-mail: ashlea.segal@gmail.com; alex.fornito@monash.edu Nature Neuroscience | Volume 26 | September 2023 | 1613–1629 1629 Article https://doi.org/10.1038/s41593-023-01404-6 Methods generated programmatically130, with cortical surface renderings gen- Original study protocols were approved by the local ethics committee erated using plotSurfaceROIBoundary (https://github.com/StuartJO/ of each dataset, and written informed consent was obtained from each plotSurfaceROIBoundary.git). All code used to produce the figures for participant. This current study was approved by the Monash University this study can be found at https://github.com/ashlea-segal/crossdis- Research Ethics Committee (project ID: 23534). In this study, data col- order_multiscale_gmv_heterogeneity.git. lection and analysis were not blind to the conditions of the experiment because participants were not placed into experimental groups. References 111. Thompson, P. M. et al. The ENIGMA Consortium: large-scale Participants collaborative analyses of neuroimaging and genetic data. This study included 3,746 individuals (1,865 HCs; 1,833 cases across Brain Imaging Behav. 8, 153–182 (2014). six different diagnostic categories) from 14 separate, independently 112. Di Martino, A. et al. The autism brain imaging data exchange: acquired datasets and 25 scan sites. No statistical methods were used towards a large-scale evaluation of the intrinsic brain architecture to predetermine sample sizes but to our knowledge, our dataset is one in autism. Mol. Psychiatry 19, 659–667 (2014). of the largest aggregations of case–control data outside the ENIGMA 113. Di Martino, A. et al. Enhancing studies of the connectome in consortium111. Full details on the study design and clinical charac- autism using the autism brain imaging data exchange II. Sci. Data teristics have been described previously for each dataset (for basic 4, 1–15 (2017). characteristics and relevant references, see Supplementary Table 1). 114. Loughland, C. et al. Australian Schizophrenia Research Bank: Each study was approved by the relevant ethics committee and written a database of comprehensive clinical, endophenotypic and informed consent was obtained from each participant. Participants genetic data for aetiological studies of schizophrenia. Aust. N.Z. J. were reimbursed for their time in accordance with the local research Psychiatry 44, 1029–1035 (2010). protocols and ethics committee guidelines. 115. Dandash, O. et al. Differential effect of quetiapine and lithium The final sample was drawn from a larger pool of individuals on functional connectivity of the striatum in first episode mania. recruited across the 14 datasets. In addition to the specific quality Transl. Psychiatry 8, 59 (2018). control procedures used within each study (references in Table 1), we 116. Sabaroedin, K. et al. Functional connectivity of corticostriatal performed a series of quality control checks and exclusions for our circuitry and psychosis-like experiences in the general analysis. Specifically, we excluded participants if they were below community. Biol. Psychiatry 86, 16–24 (2019). 18 years or above 64 years of age (N = 346); did not have the necessary 117. Parkes, L. et al. Transdiagnostic variations in impulsivity clinical data (clinical diagnosis or, for HCs, absence of any clinical diag- and compulsivity in obsessive-compulsive disorder and nosis) or demographic information (age, sex and scanner site, N = 53); gambling disorder correlate with effective connectivity in if their T1-weighted structural MRI scan did not survive our stringent cortical-striatal-thalamic-cortical circuits. NeuroImage 202, manual and automated quality control procedure, as explained below 116070–116070 (2019). (N = 269); or if the data came from a site with less than ten individuals 118. Real, E. et al. Brain structural correlates of obsessive–compulsive in the same group and sex (described in the Normative model section, disorder with and without preceding stressful life events. below; N = 217). Our final sample available for analysis thus comprised World J. Biol. Psychiatry 17, 366–377 (2016). 1,465 HCs and 1,294 cases. Demographic and other details of this cohort 119. Davey, C. G., Cearns, M., Jamieson, A. & Harrison, B. J. Suppressed are presented in Supplementary Table 1. activity of the rostral anterior cingulate cortex as a biomarker for depression remission. Psychol. Med. https://doi.org/10.1017/ Data availability S0033291721004323 (2021). Summary of data availability for each dataset used in this study is 120. Hoogman, M. et al. Nitric oxide synthase genotype modulation described below. Autism Brain Imaging Data Exchange I (ABIDE I)112 of impulsivity and ventral striatal activity in adult ADHD and ABIDE II113 datasets are available through the ABIDE repository, patients and healthy comparison subjects. Am. J. Psychiatry 168, http://fcon_1000.projects.nitrc.org/indi/abide/. The ASRB114 dataset is 1099–1106 (2011). available through the ASRB repository, subject to approval of the ASRB 121. Doan, N. T. et al. Distinct multivariate brain morphological Access Committee (https://www.neura.edu.au/discovery-portal/asrb/). patterns and their added predictive value with cognitive and First Episode Mania Study115, Monash Cohort116, Obsessive–Compulsive polygenic risk scores in mental disorders. NeuroImage Clin. 15, and Problematic Gambling study117, SPAINOCD118 and YoDA119 datasets 719–731 (2017). are available from the principal investigators of the respective studies, 122. Lepping, R. J. et al. Neural processing of emotional musical subject to evaluation of the request and local ethics committee require- and nonmusical stimuli in depression. OpenNeuro. [Dataset] ments. The International Multicentre persistent ADHD CollaboraTion (2018). (IMpACT-NL)120 and TOP15121 datasets are not publicly available due to 123. Lepping, R. J. et al. Neural processing of emotional musical and privacy or ethical restrictions. Kansas Musical Depression Study (KAN- nonmusical stimuli in depression. PLoS ONE 11, 1–23 (2016). MDD; ds000171)122,123, Massachusetts Institute of Technology Autism 124. Koster-Hale, J., Saxe, R., Dungan, J. & Young, L. L. Decoding moral Study (MITASD; ds000212.v1.0.0)124,125, Russia fMRI Depression Study judgments from neural representations of intentions. Proc. Natl (RUSMDD; ds002748.v1.0.5)126,127 and University of Washington ASD Acad. Sci. USA 110, 5648–5653 (2013). Study (WASHASD; ds002522.v1.0.0)128,129 datasets are available through 125. Young, L. et al. Moral judgments of intentional and accidental the OpenNeuro repository, https://doi.org/10.18112/openneuro). The moral violations across Harm and Purity domains. OpenNeuro. HCP dataset is available in the Human Connectome Project repository [Dataset] https://doi.org/10.18112/openneuro.ds000212.v1.0.0 (https://www.humanconnectome.org/study/hcp-young-adult). (2019). 126. Bezmaternykh, D., Melnikov, M., Savelov, A. & Petrovskii, E. Resting Code availability state with closed eyes for patients with depression and healthy The PCN toolkit (version 0.16, https://github.com/amarquand/PCN- participants. OpenNeuro. [Dataset] https://doi.org/10.18112/ toolkit.git) was used to generate the person-specific z-score deviation openneuro.ds002748.v1.0.5 (2020). maps. The HCP Pipelines: Seed fMRI Analysis (https://github.com/ 127. Mel’nikov, M. E. et al. fMRI response of parietal brain areas Washington-University/HCPpipelines.git) was adapted to map the to sad facial stimuli in mild depression. Bull. Exp. Biol. Med. 165, regions functionally coupled to extreme deviations. Figures were 741–745 (2018). Nature Neuroscience Article https://doi.org/10.1038/s41593-023-01404-6 128. Kolodny, T., Schallmo, M.-P. & Murray, S. O. Contrast response Medical Research Council of Australia, the Pratt Foundation, Ramsay functions. OpenNeuro. [Dataset] https://doi.org/10.18112/ Health Care, the Viertel Charitable Foundation and the Schizophrenia openneuro.ds002522.v1.0.0 (2020). Research Institute. Additional data were provided (in part) by the 129. Kolodny, T., Schallmo, M. P., Gerdts, J., Bernier, R. A. & Murray, S. O. Human Connectome Project, WU-Minn Consortium (Principal Response dissociation in hierarchical cortical circuits: Investigators: David Van Essen and Kamil Ugurbil; 1U54MH091657) a unique feature of autism spectrum disorder. J. Neurosci. 40, funded by the 16 NIH Institutes and Centers that support the NIH 2269–2281 (2020). Blueprint for Neuroscience Research; and by the McDonnell Center 130. Chopra, S., Labache, L., Dhamala, E., Orchard, E. R. & Holmes, A. for Systems Neuroscience at Washington University. Finally, we would A practical guide for generating reproducible and programmatic like to thank Nathan Spreng, Michel Thiebaut de Schotten and the neuroimaging visualizations. Preprint at OSF Preprints https://doi. anonymous reviewer who provided valuable feedback. org/10.31219/osf.io/mtvy9 (2023). Author contributions Acknowledgements A.S., A.F. and A.F.M. conceived the study and wrote the manuscript, This work was supported by the MASSIVE HCP facility (http://www. with valuable revision from all authors. A.S. performed the formal massive.org.au). We thank B. Fulcher for helpful discussions about this analysis, with contributions from L.P., K.A. and S.M.K. Results were manuscript, S. Oldham and S. Chopra for sharing code to generate interpreted by A.S. and A.F., with contributions from A.F.M., L.P., the cortical and subcortical brain surface renderings for the figures, S.M.K., A.Z. and J.T. The following provided data: T.W., B.F., M.H., C.F.B., and M.L. Seghier for sharing code to generate threshold-weighted L.T.W., O.A.A., A.Z., B.J.H., C.G.D., C.S.M., N.C., J.T., M.Y., L.B., C.S., M.B., overlap maps. This research was supported by the following grants: S.C., M.A.B. and A.F. The project administrator was A.F. US Department of Health & Human Services | NIH | National Institute of Mental Health (NIMH) K99MH127296,. L.P.; 2020 NARSAD Young Competing interests Investigator Grant from the Brain & Behavior Research Foundation, K.A. is a scientific advisor to and shareholder in BrainKey Inc., a medical L.P.; German Research Foundation (DFG; project number(s): image analysis software company. B.F. has received educational 513851350 and 390727645), T.W.; Nederlandse Organisatie voor speaking fees from Medice GmbH. C.F.B. is director and shareholder Wetenschappelijk Onderzoek (Netherlands Organisation for Scientific of SBGNeuro Ltd. O.A.A. is a consultant to HealthLytix and received Research) 016130669, B.F.; grant U54 EB020403 to the ENIGMA speaker’s honorarium from Lundbeck and Sunovion. N.C. participed Consortium from the BD2K Initiative, a cross-NIH partnership, in advisory boards and received speaker’s honoraria from Angelini, European College of Neuropsychopharmacology (ECNP) Network Esteve, Janssen, Lundbeck, Novartis, Pfizer and Viatris. Furthermore, ‘ADHD Across the Lifespan’. B.F.; Nederlandse Organisatie voor they have been awarded research grants from the Ministry of Health, Wetenschappelijk Onderzoek (Netherlands Organisation for Scientific Ministry of Science and Innovation (CIBERSAM), and the Strategic Plan Research) 91619115. M.H.; Wellcome Trust (Wellcome) 215698/Z/19/Z, for Research and Innovation in Health (PERIS) for the period 2016– C.F.B.; Nederlandse Organisatie voor Wetenschappelijk Onderzoek 2020, as well as from Recercaixa and Marato TV3. M.Y. has received (Netherlands Organisation for Scientific Research) Vici grant no. 17854 philanthropic donations from the David Winston Turner Endowment and NWO-CAS grant no. 012-200-013. C.F.B.; Norges Forskningsråd Fund, Wilson Foundation, as well as payments in relation to court, (Research Council of Norway) 249795. L.T.W.; Norges Forskningsråd expert witness and/or expert review reports. Finally, he has received (Research Council of Norway) 298646, L.T.W.; Norges Forskningsråd funding to conduct sponsored Investigator-Initiated trials (including (Research Council of Norway) 300767, L.T.W.; South-Eastern Norway Incannex Healthcare Ltd). These funding sources had no role in the Regional Health Authority (2018076, 2019101), L.T.W.; European design, management, data analysis, presentation or interpretation Research Council (‘BRAINMINT’ 802998), L.T.W.; Norges Forskningsråd and write-up of the data. M.Y. also sits on the Advisory Boards of (Research Council of Norway) 223273. O.A.A.; Norges Forskningsråd Centre of The Urban Mental Health, University of Amsterdam; Enosis (Research Council of Norway) 276082, O.A.A.; Department of Health Therapeutics; and Monash Biomedical Imaging Centre. M.B. has | NHMRC 1118153. A.Z.; Department of Health | NHMRC 1064643. received grant/research support from the NIH, Cooperative Research B.J.H.; Department of Health | NHMRC 1124472, B.J.H.; Department Centre, Simons Autism Foundation, Cancer Council of Victoria, of Health | NHMRC 1024570. C.G.D.; Department of Health | NHMRC Stanley Medical Research Foundation, Medical Benefits Fund, National 1141738, C.G.D.; Generalitat de Catalunya (Government of Catalonia) Health and Medical Research Council, Medical Research Futures Fund, SLT006/17/00249, C.S.; Carlos III Heath Institute (PI16/00889 and Beyond Blue, Rotary Health, A2 milk company, Meat and Livestock PI19/01171), C.S.; Secretaria d’Universitats i Recerca del Deparament Board, Woolworths, Avant and the Harry Windsor Foundation; has been d’Economia i Coneixement (2021 SGR 01017), C.S.; Secretaria a speaker for Abbot, AstraZeneca, Janssen and Janssen, Lundbeck d’Universitats i Recerca del Deparament d’Economia i Coneixement and Merck; and served as a consultant to Allergan, AstraZeneca, (2021 SGR 00832), N.C.; Carlos III Heath Institute (PI18/00036 Bioadvantex, Bionomics, Collaborative Medicinal Development, Eisai, and PI21/01756), N.C.; Turner Impact Fellowship from the Turner Janssen and Janssen, Lundbeck Merck, Pfizer and Servier—all unrelated Institute for Brain and Mental Health. J.T.; Department of Health | to this work. M.B. has received grant/research support from National NHMRC 1117188. M.Y.; Department of Health | NHMRC 1156072. M.B.; Health and Medical Research Council, Wellcome Trust, Medical Department of Health | NHMRC 1136344. S.C.; Department of Health Research Future Fund, Victorian Medical Research Acceleration Fund, | NHMRC 1154378. M.A.B.; Department of Health | NHMRC 1146292, Centre for Research Excellence CRE, Victorian Government Department M.A.B.; Department of Health | NHMRC 1045354, M.A.B.; Department of Jobs, Precincts and Regions and Victorian COVID-19 Research of Health | NHMRC 1006573, M.A.B.; European Research Council Fund. He received honoraria from Springer, Oxford University Press, (ERC, grant ‘MENTALPRECISION’ 10100118), A.F.M.; Sylvia and Charles Cambridge University Press, Allen and Unwin, Lundbeck, Controversias Viertel Charitable Foundation (Viertel Charitable Foundation), A.F.; Barcelona, Servier, Medisquire, HealthEd, ANZJP, EPA, Janssen, Department of Health | NHMRC 1197431, A.F.; Department of Health | Medplan, Milken Institute, RANZCP, Abbott India, ASCP, Headspace and NHMRC 1146292, A.F.; Department of Health | NHMRC 1050504, A.F. Sandoz. The other authors report no conflicts of interest. We thank all the sites and investigators who have worked to share their data through data sharing repositories including ABIDE and Additional information OpenNeuro. This study was supported by the Australian Schizophrenia Extended data is available for this paper at Research Bank (ASRB), which is supported by the National Health and https://doi.org/10.1038/s41593-023-01404-6. Nature Neuroscience Article https://doi.org/10.1038/s41593-023-01404-6 Supplementary information The online version Peer review information Nature Neuroscience thanks R. Nathan contains supplementary material available at Spreng, Michel Thiebaut de Schotten, and the other, anonymous, https://doi.org/10.1038/s41593-023-01404-6. reviewer(s) for their contribution to the peer review of this work. Correspondence and requests for materials should be addressed to Reprints and permissions information is available at Ashlea Segal or Alex Fornito. www.nature.com/reprints. Nature Neuroscience Article https://doi.org/10.1038/s41593-023-01404-6 Extended Data Fig. 1 | Distribution of person-specific positive (Z > 2.6; blue) and negative (Z < -2.6; red) deviation burden scores (that is, total number of extreme deviations) in each diagnostic group. * Indicates clinical groups showing a statistically significant difference in extreme deviation burden compared to the HCtest group (Mann Whitney U-test, p < .05, two-tailed). Nature Neuroscience Article https://doi.org/10.1038/s41593-023-01404-6 Extended Data Fig. 2 | Spatial overlap of extreme negative GMV deviations (Z < - 2.6) in each group. a) Cortical and subcortical surface renderings showing spatial of overlap in 1032 brain regions, and b) the distribution of overlap percentages observed across all regions. Data used to generate this figure can be found in Supplementary Data 1 (Regional_neg_thr26). Nature Neuroscience Article https://doi.org/10.1038/s41593-023-01404-6 Extended Data Fig. 3 | Permutation tests for evaluating region-level, circuit- regions, the surrogate maps were generated using Hungarian spin tests19,20. level, and network-level overlap. We used two types of permutation test to For subcortical regions, we randomly shuffled deviation values across all evaluate different hypotheses. Group-based permutation tests were used to subcortical areas (see Methods). (g) We then thresholded the null deviation maps evaluate group differences in region-level, circuit-level, and network-level (Z> |2.6|) to generate surrogate extreme deviation maps. To evaluate circuit-level overlap, regardless of total deviation burden. These tests involved repeating each group differences in overlap, (h) we obtained individual-specific surrogate FC analysis 10,000 times after shuffling case and control labels. (a) At each iteration, union maps using the same procedure described in Fig. 3a-d. (i) For the HCtest we obtained a new grouping of person-specific deviation maps according to the group and each clinical group, we calculated surrogate within-group overlap shuffled group labels. At the regional level, we focused on extreme deviation maps. ( j) We subtracted the HCtest surrogate FC overlap map from each clinical maps (Fig. 2e); at the circuit-level, we focused on union FC maps (Fig. 3d); and group’s surrogate FC overlap map to obtain a surrogate overlap difference at the network level we focused on extreme network deviation maps (Fig. 4c). map (Overlap map). Steps (f) – ( j) were repeated 10,000 times to generate (k) a (b) For each brain region, we computed an overlap map for the HCtest and each null distribution of circuit-level overlap difference maps for each disorder. clinical group under shuffled group assignment (Overlap map). (c) We then To evaluate network-level group differences in overlap, (l) we obtained surrogate subtracted the surrogate HCtest overlap map from the surrogate clinical group’s network-level extreme deviation maps using the same procedure described in overlap map to obtain an overlap difference map (Overlap map). Steps (b) and Figure 5a-c. (m) For each clinical group and the control group, we quantified the (c) were repeated 10,000 times to derive an empirical distribution of overlap proportion of individuals showing a surrogate deviation within each network difference maps under the null hypothesis of random group assignment (d). (Overlap map). (n) We subtracted the HCtest surrogate network overlap map For each brain region, we obtained p-values as the proportion of null values that from each clinical group’s surrogate overlap (Overlap map). Steps (f) – (n) were exceeded the observed difference. The second type of permutation test we used repeated 10, 000 times to generate (o) a null distribution of overlap difference was a spatial permutation test. (e) We used the unthresholded deviation maps maps for each disorder. For all tests, statistically significance differences were of each person derived from the normative model to generate an ensemble of identified using a threshold of pFDR < 0.05, two-tailed. surrogate deviation maps for each individual in the test data (f). For cortical Nature Neuroscience Article https://doi.org/10.1038/s41593-023-01404-6 Extended Data Fig. 4 | Spatial overlap of regions functionally coupled surface renderings showing spatial overlap and b) the distribution of overlap (vertex-wise threshold pFWE < 0.025), to extreme negative deviations percentages observed across all regions. Data used to generate this figure can be (Z < - 2.6) across groups, using a parcel-mapping threshold of 50%). a) Cortical found in Supplementary Data 1 (Circuit_neg_parc50). Nature Neuroscience Article https://doi.org/10.1038/s41593-023-01404-6 Extended Data Fig. 5 | Comparison of regional and circuit-level to puncorrected < .05, red corresponds to pFDR < .05, two-tailed, regional < circuit heterogeneity. (a) Statistical maps showing regions where the case-control level overlap). (b) shows results for the reverse contrast (that is, regional > circuit differences in overlap observed at the circuit level (that is, Fig. 3g) is significantly level overlap). Data used to generate this figure can be found in Supplementary greater than the overlap difference observed at the regional level (for example, Data 1 (RegionalvCircuit). Fig. 2h), identified using the group-based permutation testing (pink corresponds Nature Neuroscience Article https://doi.org/10.1038/s41593-023-01404-6 Extended Data Fig. 6 | Spatial overlap of extreme positive GMV deviations (Z > 2.6) in each group. a) Cortical and subcortical surface renderings showing spatial of overlap in 1032 brain regions, and b) the distribution of overlap percentages observed across all regions. Data used to generate this figure can be found in Supplementary Data 1 (Regional_pos_thr26). Nature Neuroscience Article https://doi.org/10.1038/s41593-023-01404-6 Extended Data Fig. 7 | Regional heterogeneity of extreme positive GMV permutation tests (light blue corresponds to puncorrected < .05, dark blue deviations in each disorder. Cortical and subcortical surface renderings corresponds to pFDR < .05, two-tailed, cases>controls). Data used to generate showing regions with significantly greater overlap of extreme positive GMV this figure can be found in Supplementary Data 1 (Regional_pos_thr26). deviations in cases compared to controls identified using group-based Nature Neuroscience Article https://doi.org/10.1038/s41593-023-01404-6 Extended Data Fig. 8 | Spatial overlap in regions functionally coupled surface renderings showing spatial overlap, and b) the distribution of overlap (vertex-wise threshold pFWE < 0.025), to extreme positive deviations percentages observed across all regions. Data used to generate this figure can be (Z > 2.6) across groups, using a parcel-mapping threshold of 50%). a) Cortical found in Supplementary Data 1 (Circuit_pos_parc50). Nature Neuroscience Article https://doi.org/10.1038/s41593-023-01404-6 Extended Data Fig. 9 | Regions showing greater overlap in areas functionally compared to controls in areas functionally coupled to extreme deviations coupled to extreme positive GMV deviations in cases compared to controls. identified using group-based or spatial permutation tests, respectively (light Group differences in circuit-level overlap were evaluated with respect to two blue corresponds to puncorrected < .05, dark blue corresponds to pFDR < .05, empirical null models (see Extended Data Fig 3 for details). (a) and (b) show two-tailed, casescontrols. The solid black line indicates -log10 in extreme positive GMV deviations in cases compared to controls. The p = 1.6 (p=0.05, two-tailed, uncorrected). (a) and (b) identify networks showing network-level -log10 p-values associated with difference in percent overlap for significant differences under group-based or spatial permutation testing, extreme positive GMV deviations between each clinical group and the HCtest respectively. Data used to generate this figure can be found in Supplementary cohort. ** corresponds to pFDR < 0.05, two-tailed, cases>controls, * corresponds Data 1 (Network_pos_10network). Nature Neuroscience nature portfolio | reporting summary Corresponding author(s): Ashlea Segal, Alex Fornito Last updated by author(s): May 30, 2023 Reporting Summary Nature Portfolio wishes to improve the reproducibility of the work that we publish. This form provides structure for consistency and transparency in reporting. For further information on Nature Portfolio policies, see our Editorial Policies and the Editorial Policy Checklist. Statistics For all statistical analyses, confirm that the following items are present in the figure legend, table legend, main text, or Methods section. n/a Confirmed The exact sample size (n) for each experimental group/condition, given as a discrete number and unit of measurement A statement on whether measurements were taken from distinct samples or whether the same sample was measured repeatedly The statistical test(s) used AND whether they are one- or two-sided Only common tests should be described solely by name; describe more complex techniques in the Methods section. A description of all covariates tested A description of any assumptions or corrections, such as tests of normality and adjustment for multiple comparisons A full description of the statistical parameters including central tendency (e.g. means) or other basic estimates (e.g. regression coefficient) AND variation (e.g. standard deviation) or associated estimates of uncertainty (e.g. confidence intervals) For null hypothesis testing, the test statistic (e.g. F, t, r) with confidence intervals, effect sizes, degrees of freedom and P value noted Give P values as exact values whenever suitable. For Bayesian analysis, information on the choice of priors and Markov chain Monte Carlo settings For hierarchical and complex designs, identification of the appropriate level for tests and full reporting of outcomes Estimates of effect sizes (e.g. Cohen's d, Pearson's r), indicating how they were calculated Our web collection on statistics for biologists contains articles on many of the points above. Software and code Policy information about availability of computer code Data collection No software was used for data collection. Neuroimaging and behavioural data were from existing datasets (detailed below) whose acquisition's are presented in detail in previous work. Data analysis T1-weighted MRIs were preprocessed using the default settings of the Computational Anatomy Toolbox (CAT12, r113, http://dbm.neuro.uni- jena.de/cat/) and Statistical Parametric Mapping (SPM12, http://www.fil.ion.ucl.ac.uk/spm/software/spm12), using MATLAB v9.8. This involved using Diffeomorphic Anatomical Registration Exponentiated Lie Algebra (DARTEL) to normalise the segmented scans. Example code and preprocessing report can be found here: https://github.com/ashlea-segal/multiscale-heterogeneity-brain-abnormalities rs-fMRI HCP data was preprocessed using the HCP minimal preprocessing pipeline (Glasser et al., 2013, Neuroimage). This included FSL FLIRT, TOPUP, ICA-FIX, and HCP Connectome Workbench (v1.2.3). HCPpipelines TaskfMRIAnalysis was adapted for the rs-fMRI seed-based analysis. Code can be found here: https://github.com/Washington-University/HCPpipelines. Normative Modelling Software: PCNtookit (version= 0.16, https://github.com/amarquand/PCNtoolkit) to generate person-specific z-score deviation maps. March 2021 Permutation Analysis of Linear Models software package (PALM alpha116, https://github.com/andersonwinkler/PALM) for statistical inference Specific packages used within Anaconda virtual environment and code to generate figures can be found here: https://github.com/ashlea- segal/multiscale-heterogeneity-brain-abnormalities For manuscripts utilizing custom algorithms or software that are central to the research but not yet described in published literature, software must be made available to editors and reviewers. We strongly encourage code deposition in a community repository (e.g. GitHub). See the Nature Portfolio guidelines for submitting code & software for further information. 1 Data nature portfolio | reporting summary Policy information about availability of data All manuscripts must include a data availability statement. This statement should provide the following information, where applicable: - Accession codes, unique identifiers, or web links for publicly available datasets - A description of any restrictions on data availability - For clinical datasets or third party data, please ensure that the statement adheres to our policy This study used neuroimaging data collected across multiple independent datasets, with varying levels of data accessibility. Below summarizes the availability of each dataset. Autism Brain Imaging Data Exchange I (ABIDE I) and ABIDE II data used in this study came from the ABIDE 1000 Functional Connectomes Project repository, http:// fcon_1000.projects.nitrc.org/indi/abide/. Data usage is unrestricted for non-commercial research purposes. As per INDI protocol, users must register with the NITRC and 1000 Functional Connectomes Project to gain access. The Australian Schizophrenia Research Bank (ASRB) is a medical research database and storage facility that links clinical and neuropsychological information, blood samples and strcutral MRI brain scans from people with schizophrenia and healthy non-psychiatric controls. Data is available in the ASRB repository, subject to approval of the ASRB Access Committee https://www.neura.edu.au/discovery-portal/asrb/ First Episode Mania Study (FEMS) - Available from the principal investigator of the study, subject to local ethics committee requirements. Monash Cohort (MON) - Available from the principal investigator of the study, subject to local ethics committee requirements International Multi-centre persistent ADHD CollaboraTion (IMpACT-NL) - Not publicly available due to privacy or ethical restrictions. OpenNeuro - Kansas Musical Depression Study (KANMDD) - Available in the OpenNeuro repository, https://openneuro.org/datasets/ds000171 OpenNeuro - Massachusetts Institute of Technology Autism Study (MITASD) - Available in the OpenNeuro repository, https://doi.org/10.18112/ openneuro.ds000212.v1.0.0 Obsessive-compulsive and problematic gambling study (OCDPG) - Available from the principal investigator of the study, subject to local ethics committee requirements OpenNeuro - Russia fMRI Depression Study (RUSMDD) - Available in the OpenNeuro repository, https://doi.org/10.18112/openneuro.ds002748.v1.0.5 SPAINOCD - Available from the principal investigator of the dataset, subject to local ethics committee requirements TOP15 - The data are not publicly available due to privacy or ethical restrictions. OpenNeuro – University of Washington ASD Study (WASHASD) - Available in the OpenNeuro repository, https://doi.org/10.18112/openneuro.ds002522.v1.0.0 YoDA - Available from the principal investigator of the study, subject to local ethics committee requirements Human Connectome Project (HCP) - Available in the Human Connectome Project repository, https://www.humanconnectome.org/study/hcp-young-adult. Users must agree to data use terms for the HCP before being allowed access to the data and ConnectomeDB, details are provided at https:// www.humanconnectome.org/study/hcp-young-adult/data-use-terms. No new data were collected for this manuscript. Across all datasets, we did not use any specific software for downloading the data. For details on data collection for each data, please see relevant references in Supplementary Table 1. Field-specific reporting Please select the one below that is the best fit for your research. If you are not sure, read the appropriate sections before making your selection. Life sciences Behavioural & social sciences Ecological, evolutionary & environmental sciences For a reference copy of the document with all sections, see nature.com/documents/nr-reporting-summary-flat.pdf Life sciences study design March 2021 All studies must disclose on these points even when the disclosure is negative. Sample size No sample size calculations were performed. This projects uses open-access data and data made available through collaborators. Combined, the sample consists of 1465 healthy controls (HC; 54.47% male) and 1294 cases, taken from 14 different studies and 25 different scan sites. The clinical sample comprised 202 individuals with ASD (100% male), 153 individuals with ADHD (41.18% male), 228 individuals with BP (47.37% male), 161 individuals with MDD (34.16% male), 167 individuals with OCD (50.30% male), and 383 participants with SCZ (62.14% male). See manuscript for extended discussion. 2 HCP data was used because of the large number of subjects and the relatively equal male/female balance. For HCP data, only unrelated subjects were included. nature portfolio | reporting summary Data exclusions All T1-weighted images were visually inspected and evaluated for the presence of artifacts, resulting in the exclusion of 53 images with gross artifacts or abnormalities. Next, we used the Computational Anatomy Toolbox (CAT12 r113, http://dbm.neuro.unijena. de/cat/) to generate a weighted overall image quality rating (IQR) for every scan. This metric combines ratings of basic image properties, including the level of noise and geometric distortions, into a single score that quantifies the overall image quality of a participant’s T1-weighted scan (see here for more information www.neuro.uni-jena.de/cat/index.html#QA ). On this metric, lower scores denote higher image quality. As per previous work21, we excluded 153 images with an IQR > 2.8. An additional 63 images were excluded due to a failure of our processing pipeline. Collectively, this quality assurance process resulted in the exclusion of a total of 269 scans. Following this quality control protocol, we performed a series of additional quality control checks and exclusions for our analysis. Specifically, we excluded participants if they were below 18 years or above 64 years of age (N=346); did not have the necessary clinical data (clinical diagnosis or, for healthy controls, absence of any clinical diagnosis) or demographic information (age, sex and scanner site, N=53); if their T1-weighted structural magnetic resonance imaging (MRI) scan did not survive our stringent manual and automated quality control procedure, as explained below (N=269); or if the data came from a site with less than 10 individuals in the same group and sex (described in the Normative model section, below; N=217). Our final sample available for analysis thus comprised 1465 HCs and 1294 cases. Replication We have included numerous quality control checks to verify the stability of our findings, including: • Cross-validation of the normative models to assess model generalisability • Multiple linear support vector machines to evaluate the model’s efficacy in partitioning site-related variance • Pearson’s correlation between number of extreme deviations and image quality to assess whether individual variations in scan quality affected the deviation zmaps • Threshold-weighted analysis of deviation mapping to show that our results do not depend on the choice of a specific threshold for defining deviations • The inclusion of a held-out control group for all analyses to establish a normative benchmark for assessing % overlap in the clinical groups • Mapping circuit-level overlap using two mapping thresholds (50% and 75%) to ensure our findings were not driven by this specific choice • Statistical inference at the circuit- and network- level using two different null models to disentangle whether any differences arise from a preferential aggregation of deviations or whether group differences are driven by variations in total deviation burden Randomization Randomization was not performed because participants were not placed into experimental groups - case and control groups are defined based on diagnosis. Participants without a diagnosis of a psychiatric disorder were in the control group. Participants diagnosed with a psychiatric disorder were placed in the clinical group, which was further divided by diagnosis (ADHD, ASD, BP, OCD, MDD, SCZ). Blinding Blinding is not relevant to this study because participants were not placed into experimental groups. Reporting for specific materials, systems and methods We require information from authors about some types of materials, experimental systems and methods used in many studies. Here, indicate whether each material, system or method listed is relevant to your study. If you are not sure if a list item applies to your research, read the appropriate section before selecting a response. Materials & experimental systems Methods n/a Involved in the study n/a Involved in the study Antibodies ChIP-seq Eukaryotic cell lines Flow cytometry Palaeontology and archaeology MRI-based neuroimaging Animals and other organisms Human research participants Clinical data Dual use research of concern Human research participants Policy information about studies involving human research participants Population characteristics See above in Life sciences study design: Sample size. All participants were aged between 18-64 years old. Recruitment See the relevant reference for each dataset for specific details of each dataset. In general, the sample was collected through March 2021 a range of different recruitment resources including media advertisements, inpatient, outpatient and community mental health service providers, non-government organizations, rehabilitation services, and cold telephoning using the electoral rolls. Ethics oversight The study was approved by the local ethics committee of each dataset, and written informed consent was obtained from each participant. This study was approved by the Monash University Research Ethics Committee (Project ID: 23534). Note that full information on the approval of the study protocol must also be provided in the manuscript. 3 Magnetic resonance imaging nature portfolio | reporting summary Experimental design Design type resting-state fMRI, structural (VBM) MRI Design specifications HCP: The rs-fMRI data for each subject consists of four runs (rfMRI_REST1_LR, rfMRI_REST1_RL, rfMRI_REST2_LR, rfMRI_REST2_RL) acquired at two different sessions (REST1 and REST2) using two different directions of phase coding (LR: left to right, and RL: right to left). Behavioral performance measures Behavioral performance measures were not applicable in this study Acquisition Imaging type(s) resting-state fMRI, structural (VBM) MRI Field strength Described in Supplementary Table 1 for each dataset. Sequence & imaging parameters Described in Supplementary Table 1 for each dataset for structural data. HCP rs-fMRI data were obtained using a 32-channel Siemens 3T connectome-Skyra scanner. The imaging parameters for rs-fMRI were as follows: repetition time (TR) =720 ms, echo time (TE) = 33.1 ms, flip angle (FA) = 52°, field of view (FOV) = 208 × 180 mm2, matrix = 104 × 90, slice number = 72, slice thickness = 2 mm, voxel size = 2 × 2 × 2 mm3, multiband factor = 8, and 1200 volumes. Area of acquisition Whole brain Diffusion MRI Used Not used Preprocessing Preprocessing software Preprocessing of the structural and functional data was done using a suite of tools. Structural preprocessing GMV was estimated using the CAT12 VBM pipeline, which is included as an extension of Statistical Parametric Mapping software (SPM12, http://www.fil.ion.ucl.ac.uk/spm/software/spm12) in MATLAB v9.8 using the default settings. Briefly, the T1-weighted images were first corrected for intensity nonuniformities, and segmented into GM, white matter (WM), and cerebrospinal fluid (CSF) tissue probability maps. Using the high- dimensional Diffeomorphic Anatomical Registration Exponentiated Lie Algebra (DARTEL), the segmented scans were normalized into standard Montreal Neurological Institute (MNI) space. Lastly, the images were bias-field corrected and modulated by the linear and nonlinear components of the Jacobian determinant obtained from the DARTEL deformation fields to obtain voxel-wise estimates of GMV. Functional preprocessing The rs-fMRI data used underwent the HCP’s minimal pre-processing pipeline, which includes gradient-nonlinearity-induced distortion, motion correction to the single-band reference image using FLIRT, EPI image distortion correction using TOPUP, registration into standard space using a customized boundary-based-registration (BBR) algorithm, and single step spline interpolation using all transforms, intensity normalization and bias field removal to resample the original EPI into MNI space. Minimal high pass filtering was applied with a cutoff of 2000ms. Artifacts were then removed using ICA-FIX. This involves employing an automatic classifier, specifically trained for HCP data, to identify ICA components due to measurement noise, additional motion or physiological artifacts like cardiac pulsation and respiration. Next, the volume timeseries were mapped into the standard CIFTI grayordinate space and smoothed to 2mm FWHM (where the smoothing was on the surface for the cortex and in volume space for subcortex). This results in a standard set of grayordinates in every subject, with surface vertex data and subcortical volume voxel data. The mean GM signal was then removed each grayordinate’s time series to remove residual widespread signal deflections that are not removed by ICA-FIX Normalization See above Normalization template MNI Standard Space Noise and artifact removal See above March 2021 Volume censoring See above Statistical modeling & inference Model type and settings Mass univariate across multiple brain scales (regional, circuit, network; see manuscript) Effect(s) tested As the primary aim of this study was to investigate anatomically heterogeneous brain deviations across multiple brain scales (regional, circuit, network) across six psychiatric disorders, multiple effects were examined. The main effect - observed group 4 differences between each patient group and controls was evaluated with respect to two complementary nulls detailed in Supplementary Figure 6. nature portfolio | reporting summary Specify type of analysis: Whole brain ROI-based Both Parcel-level analyses utilised the Schaefer et al., 2018 Cerebral Cortex 1000 region parcellation for Anatomical location(s) cortical regions, and the Tian et al Nature Neuroscience 32 region parcellation for subcortical regions. Statistic type for inference Non-parametric inference was used across multiple neuroanatomical scales including regions of interest, and networks (see (See Eklund et al. 2016) manuscript) Correction Results were presented at p<0.05 uncorrected and following False Discovery Rate (FDR, q=0.05) correction. Models & analysis n/a Involved in the study Functional and/or effective connectivity Graph analysis Multivariate modeling or predictive analysis Functional and/or effective connectivity Pearson correlation March 2021 5