Abstract
BACKGROUND Left-right asymmetry is an important organizing feature of the healthy brain that may be altered in schizophrenia, but most studies have used relatively small samples and heterogeneous approaches, resulting in equivocal findings. We carried out the largest case-control study of structural brain asymmetries in schizophrenia (N = 11,095), using a single image analysis protocol.
METHODS We included T1-weighted data from 46 datasets (5,080 affected individuals and 6,015 controls) from the ENIGMA Consortium. Asymmetry indexes were calculated for global and regional cortical thickness, surface area, and subcortical volume measures. Differences of asymmetry were calculated between affected individuals and controls per dataset, and effect sizes were meta-analyzed across datasets. Analyses were also performed with respect to the use of antipsychotic medication and other clinical variables, as well as age and sex. Case-control differences in a multivariate context were assessed in a subset of the data (N = 2,029).
RESULTS Small average differences between cases and controls were observed for asymmetries in cortical thickness, specifically of the rostral anterior cingulate (d = -0.08, pFDR = 0.047) and the middle temporal gyrus (d = -0.07, pFDR = 0.048), both driven primarily by thinner cortices in the left hemisphere in schizophrenia. These asymmetries were not significantly associated with the use of antipsychotic medication or other clinical variables. Older individuals with schizophrenia showed a stronger average leftward asymmetry of pallidum volume than older controls (d = 0.08, pFDR = 9.0 × 10-3). The multivariate analysis revealed that 7% of the variance across all structural asymmetries was explained by case-control status (F = 1.87, p = 1.25 × 10-5).
CONCLUSIONS Altered trajectories of asymmetrical brain development and/or lifespan asymmetry may contribute to schizophrenia pathophysiology. Small case-control differences of brain macro-structural asymmetry may manifest due to more substantial differences at the molecular, cytoarchitectonic or circuit levels, with functional relevance for lateralized cognitive processes.
Introduction
Schizophrenia is a serious mental illness characterized by various combinations of symptoms that may include delusions, hallucinations, disorganized speech, affective flattening, avolition, and executive function deficits1. Left-right asymmetry is an important feature of human brain organization for diverse cognitive functions – for example, roughly 90% of people present with a left-hemisphere dominance for language and right-handedness2-5. A possible role of altered structural and functional brain asymmetry in schizophrenia has been studied for several decades6-10. Theoretical work has especially focused on disrupted laterality for language in relation to disorganized speech and auditory hallucinations11-13. Individuals with schizophrenia have been reported to show a decreased left-lateralized language dominance14,15, as well as an absence or even reversal of structural asymmetries of language-related regions around the Sylvian fissure (which divides the temporal lobe from the frontal and parietal lobes)13,16-18. Furthermore, the rate of non-right-handedness in schizophrenia is elevated compared to the general population13,19-23. Interestingly, some genomic loci that influence aspects of structural brain asymmetry or hand preference overlap with those associated with schizophrenia24-27. Thus, there might be an etiological link between altered brain asymmetry and schizophrenia.
However, alterations in structural brain asymmetry in schizophrenia have so far only been reported in studies with relatively small samples13,16-18,28-34; to our knowledge, the largest case-control sample consisted of 167 affected individuals and 159 controls31. Many of the existing findings are inconsistent and/or remain unreplicated, which is possibly due to the low statistical power which limits the sensitivity to detect true effects, and also increases the risk of overestimating effect sizes35,36. The reproducibility of findings may be further affected by the heterogeneity of clinical and demographic characteristics across studies. Moreover, varying approaches to process and analyze magnetic resonance imaging (MRI) data limit the possibility to reproduce results and/or to perform meta-analyses. For example, in studies targeting specific regions of interest, regions have been inconsistently defined, while studies that involved cortex-wide mapping used different image analysis protocols. Studies of subcortical volumetric asymmetries in schizophrenia have generally suffered from similar issues37-39, with the notable exception of a study in 884 affected individuals and 1,680 controls that used a single image analysis pipeline40. This study found an increased leftward asymmetry of the pallidum in schizophrenia (driven by a larger pallidum volume in the left hemisphere) compared to controls, which was also detectable in adolescents with subclinical psychotic experiences40,41.
The Enhancing Neuro Imaging Genetics through Meta-Analysis (ENIGMA, http://enigma.ini.usc.edu) consortium aims to perform large-scale analyses by combining imaging data from research groups across the world, processed with standardized protocols42,43. Previously, this consortium reported large-scale cortical thinning, smaller surface area, and altered subcortical volume in individuals with schizophrenia compared to controls44,45. However, asymmetry was not measured in these previous ENIGMA studies, and no tests were performed to assess whether case-control effects were different in the two hemispheres. The ENIGMA consortium has investigated structural brain asymmetries in other disorders46: major depressive disorder (MDD)47, autism spectrum disorder (ASD)48, obsessive compulsive disorder (OCD)49, and attention deficit/hyperactivity disorder (ADHD)50, but not schizophrenia to date. Case-control group-level effects were small for all of these disorders, with ASD showing the most widespread asymmetry differences – mostly involving regional cortical thickness measures – with a maximum Cohen’s d of 0.1348. Similar effect sizes may be anticipated for schizophrenia. Therefore, a large sample size is likely required to detect and accurately measure any effects. Although small group-average differences of brain macro-anatomy are unlikely to have clinical uses by themselves, they may help to identify brain regions and networks that have clinically relevant disruptions at other neurobiological levels – for example molecular or cytoarchitectonic – which can be investigated in future studies. Of note, the ENIGMA consortium has recently reported on asymmetry alterations with respect to subcortical shape (2,833 individuals with schizophrenia versus 3,929 controls), based on an automated approach quantifying local concave versus convex surface curvature51, but this study did not address subcortical volume asymmetries, and omitted the cerebral cortex.
Therefore, we designed the current study focusing on cortical and subcortical structural asymmetries in schizophrenia using the largest sample to date, containing 5,080 affected individuals and 6,015 controls from 46 separate datasets. The datasets were collected originally as distinct studies over approximately 25 years, using different recruitment schemes, MRI scanning equipment and parameters. Importantly, for the current study, all primary MRI data were processed through a single pipeline for cortical atlas-based segmentation/subcortical parcellation and quality control. Case-control differences were tested separately for each regional asymmetry index (AI) within each dataset, and effects were then combined across datasets using meta-analysis methodology. This approach was taken due to restrictions on sharing individual-level data for many of the primary datasets. We also explored possible associations of structural brain asymmetries with medication use and other disorder-specific measures: age at onset, duration of illness, as well as total, positive, and negative symptom scores. For 14 datasets, for which individual-level data were available, we tested for a multivariate association of case-control status simultaneously with regional AIs across the brain. Together, these analyses aimed to provide novel insights into the extent and mapping of structural brain asymmetry alterations in schizophrenia, and how they relate to key clinical variables.
Methods
Datasets
Structural MRI data were derived from 46 separate datasets (45 case-control and one case-only) via researcher participation in the ENIGMA schizophrenia working group, totaling 11,095 individuals. Of these, 5,080 were affected with schizophrenia and 6,015 were unaffected controls (Table 1, Supplementary Table 1A). The datasets came from various countries around the world (Fig. 1). Data collection was performed after written informed consent and was approved by local institutional review boards. Sample size-weighted mean age across datasets was 33.3 (range 16.2-44.0) years for individuals with schizophrenia and 33.0 (11.8-43.6) years for controls. Affected individuals and controls were 67% and 52% male, respectively. Diagnostic interviews were conducted by registered clinical research staff using different diagnostic criteria (either the Diagnostic and Statistical Manual of Mental Disorders [DSM]-III, DSM-IV, DSM-5 or International Classification of Diseases [ICD]-10), and hand preference was obtained through assessment scales (mainly the Edinburgh Handedness Inventory and Annett Handedness Scale) or self-report (Supplementary Table 2). No controls had present or past indications of schizophrenia.
Image acquisition, processing and quality control
T1-weighted structural brain MRI scans were acquired at each study site. Site-specific scanner information, field strengths (1 T, 1.5 T, and 3 T), and image acquisition parameters are provided in Supplementary Table 2. For data from all sites, image processing and segmentation were performed using FreeSurfer (see Supplementary Table 2 for software versions)52. For each individual, using the ‘recon-all’ pipeline, cerebral cortical thickness and surface area measures were derived for 34 bilaterally paired Desikan-Killiany (DK) atlas regions, as well as whole hemisphere-level average cortical thickness and surface area measures53. Volumes for 8 bilaterally paired regions from a neuroanatomical atlas of brain subcortical structures54 were derived using the ‘aseg’ segmentation command in FreeSurfer. A standardized ENIGMA quality control procedure was applied at each participating site (described in full here: http://enigma.ini.usc.edu/protocols/imaging-protocols/).
Briefly, this included outlier detection in the derived cortical and subcortical measures and visual inspection of segmentations projected onto the T1-weighted image of each individual. For cortical measures, predefined guidelines for visual inspection were followed. Measurements from regions with poor segmentation were excluded, as well as individuals whose data failed overall quality checks.Data sharing limitations did not allow the central analysis group to have access to individual-level data for the majority of participating study sites. For further processing and analyses of the data, a script running in R software (R Foundation for Statistical Computing, Vienna, Austria, www.R-project.org)55 was prepared and distributed among participating sites, to ensure coordinated collection of descriptive and summary statistics for subsequent meta-analysis by the central analysis team.
Asymmetry index calculation
For each bilaterally paired brain regional measure, we used the left (L) and right (R) hemispheric measurements to calculate an asymmetry index , where the denominator corrects for automatic scaling of the index with the magnitude of the bilateral measure. This formula for AI calculation has been widely used2,49,56-58. A negative value of the AI reflects a larger right hemispheric measurement (R > L), and a positive value a larger left hemispheric measurement (L > R).
Distributions of AIs were plotted using histograms to allow for visual inspection. Left or right measurements equal to 0 were set to missing, as these most likely reflected data entry errors. Furthermore, when a left or right measurement was missing, the corresponding measurement in the opposite hemisphere was also set to missing. The standardized pipeline from raw image data through FreeSurfer does not introduce left-right flipping errors, but to ensure that such errors were not introduced during processing of raw imaging data by non-standard processes (e.g. during the conversion of DICOM to NIFTI files with bespoke scripts), we compared mean regional asymmetries for all datasets against grand sample-size adjusted means. If we noticed a large proportion of reversed average asymmetries for a dataset, we contacted the relevant site to re-check and correct their process (Supplementary Table 3).
Asymmetry differences between individuals with schizophrenia and unaffected controls
Group differences were examined separately for each brain regional AI and each case-control dataset, using univariate linear regression implemented in R. Our primary analysis model included diagnosis (case-control status) as the main binary predictor, and sex and age as covariates (model 1 in Supplementary Note 1). For ten datasets where more than one scanner had been used (Supplementary Table 2), we added n-1 binary dummy covariates (where n is the number of scanners in a given dataset), to statistically control for scanner effects. Sex was not included as a covariate for the RSCZ dataset, as this dataset included only males. Collinearity between predictor variables was assessed using the R-package usdm (v1-1.18)59, and high collinearity (variance inflation factor > 5) was not found for any dataset. Linear regression analysis for any structural AI was not performed if the total sample size of a given dataset was lower than ten plus the number of scanner covariates, or if one of the diagnostic groups had a sample size lower than five. For each brain regional AI and each case-control dataset, we extracted the t-statistic for the ‘diagnosis’ term to calculate its corresponding Cohen’s d effect size, standard error and 95% confidence interval, using , and 95% CI = [d - 1.96 * sed, + 1.96 *sed]60. In these equations, d is the Cohen’s d effect size, t is the t-statistic, se is the standard error, n1 is the number of unaffected controls, n2 is the number of individuals with schizophrenia, and df the degrees of freedom in the linear model.
Random-effects meta-analysis
For each brain regional AI (Supplementary Fig. 1-3), effect sizes for ‘diagnosis’ from each case-control dataset were meta-analyzed in a random-effects model fitted with a restricted maximum likelihood (REML) estimator, using the function ‘rma’ in the R package metafor (v3.0-2)61. Meta-analyzed effect sizes were projected on 3D meshes of inflated cortical or subcortical models from Brainder (www.brainder.org/research/brain-for-blender/), using Matlab R2020a (version 9.8.0.1323502; MathWorks, Natick, MA, USA). We calculated false discovery rate (FDR) corrected p-values using the Benjamini-Hochberg method to account for multiple tests62 (i.e., separately for testing 35 cortical thickness AIs, 35 cortical surface area AIs, and eight subcortical volume AIs).
Effects with pFDR < 0.05 were considered statistically significant. For AIs that showed significant group differences between cases and controls, the group differences for the corresponding left and right measurements separately were also assessed post hoc (again using linear modelling with diagnosis, age and sex as predictors), to help describe the asymmetry differences.
Sensitivity analyses
For any AI that showed a significant case-control group difference in the primary meta-analysis, we identified datasets within which the 95% CI of the diagnosis effect did not overlap with the 95% CI of the meta-analyzed effect – using the ‘find.outliers’ function in the R package dmetar (v0.0.9)63 – and then repeated the meta-analysis after excluding such outlier datasets. In addition, for any AI that showed a significant case-control group difference in the primary meta-analysis, we also applied models that included the same covariates as the primary analysis, but also included either handedness (right-handed vs. non-right-handed), intracranial volume (ICV), both handedness and ICV, or age2 (models 2-5 in Supplementary Note 1).
Medication group differences
For AIs that showed significant case-control group differences in the primary analysis, we explored associations with antipsychotic medication use at the time of scanning, through between-group comparisons of AIs of unmedicated individuals with schizophrenia, affected individuals taking only first-generation (typical) antipsychotics, affected individuals taking only second-generation (atypical) antipsychotics, and those taking both first- and second-generation antipsychotics. Sex and age were included as covariates (model 6 in Supplementary Note 1) and derived Cohen’s d effect sizes were again meta-analyzed across datasets in a random-effects model. Applying a minimum group size threshold of 5 within any given dataset, sufficient data on the presence/absence of antipsychotic medication use for at least one comparison were available for 31 of the datasets (Supplementary Table 1B), and the sample sizes for each between-group comparison are in Supplementary Table 9. We calculated FDR corrected p-values to correct for all of the multiple subgroup comparisons and structural asymmetries tested.
Correlations of asymmetries with clinical variables
For AIs that showed significant case-control group differences in the primary analysis, we assessed relationships between these AIs and clinical variables within affected individuals only: age at onset, duration of illness, chlorpromazine equivalent medication dose (at the time of scanning), as well as positive, negative, and total symptom severity scores from the Positive and Negative Symptom Scale (PANSS)64, or the Scale for the Assessment of Positive Symptoms (SAPS)65 and Scale for the Assessment of Negative Symptoms (SANS)66 (separately depending on data availability, see Supplementary Table 1A). Partial correlations between brain AIs and these quantitative measures were estimated using the ‘pcor.test’ function in the R package ppcor (v1.1)67. Age and sex were included as covariates (model 7 in Supplementary Note 1). The same minimum sample size requirement for dataset inclusion was applied as in the linear regression analyses (above). Correlation coefficients were meta-analyzed across datasets in a mixed-effects model including dataset as a random effect. We calculated FDR corrected p-values to control for all of the clinical variables and structural asymmetries tested. Sample sizes for each model are shown in Supplementary Table 10.
Secondary analysis of age-or sex-specific effects
For all AIs in all case-control datasets we applied models which were the same as the primary analysis but additionally included either diagnosis-by-age or diagnosis-by-sex interaction terms. We then carried out meta-analyses of the interaction effect estimates across datasets to assess possible AI differences between affected individuals and controls that were relatively specific to either males or females, or differed with age (models 8-9 in Supplementary Note 1). In the same way as our primary analysis, we calculated FDR corrected p-values to account for multiple regional asymmetries tested.
Multivariate analysis of case-control asymmetry differences
To examine case-control group differences across all brain regional AIs simultaneously in one model, we conducted a multivariate analysis based on 14 datasets for which individual-level data were available to the central analysis team. For this analysis, we only retained individuals with complete data for all bilateral measures of cortical and subcortical structures, which were 935 individuals affected with schizophrenia and 1,095 unaffected controls (Supplementary Table 1C). We separately adjusted the left and right measurements using ComBat harmonization (an empirical Bayesian method) to remove dataset effects68, where each dataset (and each scanner within multi-scanner datasets) was treated as a distinct ‘batch’. Diagnosis, age and sex were used as covariates when finding the data harmonization parameters in ComBat. After ComBat adjustment, one additional control individual was removed due to being assigned a negative corrected right hemisphere lateral ventricle volume (Supplementary Fig. 4). AIs for cortical and subcortical measures were then calculated using the same formula as above, and collinearity between AIs was assessed by calculating a correlation matrix. AIs did not show higher pairwise correlations than 0.5 (Supplementary Fig. 5-6) A multivariate analysis of covariance (MANCOVA) using the ‘manova’ function in R was applied, testing all 76 regional structural brain AIs simultaneously against case-control status, with age and sex as covariates. We ran one million label-swapping permutations of case-control labels and calculated a permutation p-value by assessing the number of times the F-statistic of an analysis with permuted data was equal to or larger than the F-statistic of the analysis with real data, divided by the total number of permutations. When permuting case-control labels, we conserved case-control numbers within each dataset (and within scanner for multi-scanner datasets). To help interpret the MANCOVA results, we also derived univariate case-control association statistics for each separate structural AI from the multivariate association analysis output, using ANCOVA (‘summary.aov’ function in R).
Results
Asymmetry differences between individuals with schizophrenia and unaffected controls
In our primary analysis (model 1), total hemispheric average cortical thickness asymmetry (d = -0.053, z = -1.92, p = 0.055) and surface area asymmetry (d = 0.027, z = 1.23, p = 0.22) did not significantly differ between affected individuals and controls. At a regional level (Fig. 2, Supplementary Table 4, Supplementary Fig. 1-3), there was a small but significant case-control difference in cortical thickness asymmetry of the rostral anterior cingulate cortex (d = -0.083, z = -3.21, p = 1.3 × 10-3, pFDR = 0.047, reversal from leftward average asymmetry in controls to rightward average asymmetry in cases), and also in cortical thickness asymmetry of the middle temporal gyrus (d = -0.074, z = -2.99, p = 2.8 × 10-3, pFDR = 0.048, increased average rightward asymmetry in cases) (Fig. 3, Supplementary Fig. 7-8, Supplementary Table 5). Post hoc analysis of unilateral effects showed that both of these regional asymmetry differences were driven primarily by thinner left than right cortex in individuals with schizophrenia compared to controls (Table 2, Supplementary Table 6). The middle temporal cortex is a core language network region69, and left-hemisphere thinning is compatible with disrupted leftward laterality of brain organization for language in schizophrenia10,11.
Nominally significant regional case-control associations (i.e. which did not survive multiple testing correction), were found for the AIs of inferior parietal cortex thickness, cuneus surface area, parahippocampal gyrus surface area, and nucleus accumbens volume (Fig. 3, Supplementary Table 5).
Sensitivity analyses
For rostral anterior cingulate thickness asymmetry, there were three datasets in the primary meta-analysis which had outlier case-control effect sizes when compared to the meta-analyzed effect (see Methods). After excluding these datasets and repeating the meta-analysis for this AI, the case-control difference remained, with the same directionality (d = -0.073, z = -3.51, p = 4.5 × 10-4) (Supplementary Table 7). For middle temporal gyrus thickness asymmetry, the exclusion of two outlier datasets also yielded a similar result compared to the primary analysis (d = -0.079, z = -3.44, p = 5.9 × 10-4), again with the same directionality (Supplementary Table 7). In models that included either handedness, ICV, both handedness and ICV, or age2 as additional covariates (models 2-5), the results for both of these regional AIs remained nominally significant, although the case-control effect for middle temporal gyrus thickness asymmetry decreased in most of these models, whereas it increased for rostral anterior cingulate thickness asymmetry (Supplementary Table 8).
Medication group differences
Rostral anterior cingulate thickness asymmetry did not differ between affected individuals across medication groups (model 6) (Supplementary Table 9). For the middle temporal gyrus, there was a nominally significant increase in average rightward asymmetry in affected individuals taking first generation versus second generation antipsychotics at the time of scanning (d = -0.21, z = -2.56, p = 0.011, pFDR = 0.13), i.e., this was not significant after multiple testing correction (Supplementary Table 9).
Correlations of asymmetries with clinical variables
We found nominally significant correlations between rostral anterior cingulate thickness asymmetry and negative symptom severity measured with SANS (r = 0.049, z = 2.08, p = 0.038, pFDR = 0.32, decreased rightward asymmetry with higher negative symptom rate) (Supplementary Table 10A) and between middle temporal gyrus thickness asymmetry and duration of illness (r = -0.048, z = -1.97, p = 0.049, pFDR = 0.32, increased rightward asymmetry with longer duration of illness) (Supplementary Table 10B), but these correlations did not remain significant when correcting for multiple testing. No correlations with chlorpromazine-equivalent medication dose, age at onset, PANSS scores (total or positive and negative subscales), or SAPS or SANS scores, were found for either the rostral anterior cingulate thickness asymmetry or middle temporal gyrus thickness asymmetry (Supplementary Table 10).
Age-and sex-specific effects
In secondary analyses across all AIs using models with interaction terms, we found a significant diagnosis-by-age interaction (model 8) for pallidum volume asymmetry (d = 0.081, z = 3.26, p = 1.1 × 10-3, pFDR = 9.0 × 10-3, stronger leftward asymmetry with higher age in cases) (Supplementary Table 11-12A, Supplementary Fig. 9). This association was driven by a significantly decreased average leftward asymmetry with increasing age in controls (r = -0.077, p = 1.1 × 10-3) that was not present in affected individuals (Supplementary Table 12B; Supplementary Fig. 10). In terms of the corresponding unilateral effects, left and right pallidum volume decreased with increasing age in individuals with schizophrenia (L: r = -0.17, p = 4.7 × 10-9; R: r = -0.20, p = 4.7 × 10-21) and unaffected controls (L: r = -0.27, p = 2.1 × 10-22; R: r = -0.24, p = 6.2 × 10-17), but the two groups differed with respect to the side showing the stronger effect (Supplementary Table 12B). No significant sex-by-diagnosis interactions were found (model 9) (Supplementary Table 13).
Multivariate analysis of case-control asymmetry differences
Considering all 76 regional structural brain AIs simultaneously in a multivariate model, applied to the 14 datasets for which individual-level data were available to the central analysis team (935 affected individuals and 1,094 controls), there was a significant multivariate structural brain asymmetry difference between cases and controls that accounted for roughly 7% of the variance considered across all 76 AIs (Wilks’ Λ = 0.932, approximate F(76, 1950) = 1.87, p = 1.25 × 10-5). Only three of the F-statistics resulting from one million label-swapping permutations (see Methods) were larger than the F-statistic from the true analysis, resulting in a permutation p = 3.0 × 10-6. We also derived univariate (ANCOVA) association statistics from the multivariate model to understand which AIs contributed most to the significant multivariate association. The structural AIs that showed nominally significant, univariate case-control differences in the 14 datasets available for this analysis were those for pallidum volume, nucleus accumbens volume, and eight regional surface area or thickness measures distributed widely over the cerebral cortex (Table 3). These did not include the two cortical regional AIs that showed significant case-control differences in the meta-analysis over all 45 case-control datasets, but did include AIs of other language-related regions of the temporal lobe: superior temporal sulcus surface area asymmetry and transverse temporal gyrus thickness asymmetry (Table 3).
Discussion
In this study, we investigated group differences in structural brain asymmetries between individuals with schizophrenia and unaffected controls, in the largest sample to date. The large sample size offered unprecedented statistical power to identify group differences, and to measure their effect sizes35. Subtle differences of regional asymmetry were found for rostral anterior cingulate thickness, middle temporal gyrus thickness, and pallidum volume (the latter in older individuals). The Cohen’s d effect sizes were less than 0.1; i.e., very small70. In light of previous large-scale analyses of bilateral cortical and subcortical alterations in schizophrenia44,45, our results suggest that morphometric alterations in this disorder are largely the same for the left and right hemispheres, involving only subtle asymmetrical effects at the group average level. This suggests that effect sizes of brain asymmetry differences in schizophrenia reported in earlier, much smaller studies (see Introduction), are likely to have been overestimated. Nonetheless, in a multivariate context, 7% of the total variance across all regional asymmetries was explained by case-control status.
Subtle group differences of asymmetry in terms of macro-anatomic features, such as those studied here, may reflect substantial effects at other neurobiological levels with more functional relevance for disorder symptoms – for example molecular, cytoarchitectonic and/or circuit levels71-73. The middle temporal gyrus is prominently involved in the brain’s language network69, so that our univariate finding of lower left-sided cortical thickness in schizophrenia in this region is broadly consistent with a prominent theory in the literature: that left-hemisphere language dominance may be affected in this disorder10,11. Cortical thinning of the left-hemispheric middle temporal gyrus has been associated with auditory verbal hallucinations in schizophrenia74, and reported to be present also in individuals with first-episode schizophrenia and high familial risk for the disorder75,76. In terms of grey matter volume, an opposite pattern (reduced right, increased left) has been reported for the middle temporal gyrus in putatively at-risk children compared to typically developing children77. However, volume measures confound cortical thickness and surface area, and since these two aspects of cortical anatomy are known to vary substantially independently26,78,79, it is unclear how these earlier volume-based findings may relate to the present findings based on cortical thickness asymmetry. Again, earlier findings in smaller samples may have been false positives with over-estimated effect sizes.
The rostral anterior cingulate cortex is an important hub in emotional and cognitive control80, both of which are often affected in schizophrenia. In this region we observed a thinner left-sided cortex in affected individuals than controls on average, which was more pronounced than on the right side. This may be consistent with a previous study where adolescent/young adult relatives of individuals with schizophrenia showed a longitudinal decline of gray matter volume in the left rostral anterior cingulate cortex compared to controls81. It is therefore possible that asymmetrical differences in this region emerge before schizophrenia onset, although the previous study included only 23 relatives, so its reported effects remain equivocal, and it used volume rather than thickness measures. In the present study, we saw no evidence for an age*diagnosis interaction effect for this regional thickness asymmetry, which is consistent with a pre-onset alteration that subsequently remains stable through adulthood.
Multivariate analysis in 14 of the datasets, for which individual-level data were available, resulted in a highly significant case-control difference. Various regional asymmetries contributed to this multivariate association, with pallidum volume asymmetry showing the largest individual contribution. Pallidum volume asymmetry was especially associated with schizophrenia in older individuals, as observed in secondary testing of univariate interaction models across all 45 case-control datasets. Larger pallidum volume in schizophrenia compared to controls – with a stronger effect in the left hemisphere – has been reported before40,41,45,82, although some datasets in our analysis partly overlapped with three of these studies40,41,45. An age-dependent relationship between familial risk for schizophrenia and larger left pallidum volume has also been described in a small study of young adults83 – this suggests that alterations of pallidum asymmetry might already be present in a prodromal stage of the disease. However, in the present study, the group difference in pallidum volume was absent in younger individuals and became more apparent in older adults. This also explains why the association was not significant in the primary univariate meta-analysis of all datasets together, i.e. it was driven by a subset of datasets that included older individuals, and that were also available for multivariate analysis (Supplementary Fig. 10). The pallidum is prominently involved in reward and motivation84, and impaired reward anticipation and a loss of motivation are well-known negative symptoms of schizophrenia85. However, how pallidum structural asymmetry may relate to functional disorder-relevant changes remains unknown.
We found a tentative difference of middle temporal gyrus thickness asymmetry between individuals who were taking first-generation versus second-generation antipsychotics, with a relatively larger effect size (d = -0.21) than any case-control differences from the primary analysis. This result suggests that first-generation antipsychotics have a stronger effect on left middle temporal gyrus thickness than second-generation antipsychotics, which could relate to molecular or cellular hemispheric differences. Alternatively, this regional asymmetry may relate to clinical differences of disorder presentation which affect treatment choices. Indeed, we saw nominally significant evidence that this same asymmetry relates to illness duration. However, the medication subgroup analyses were limited by relatively small sample sizes compared to the primary case-control analysis, and this particular association did not survive multiple testing correction. Also, medication status was based on medication use at the time of scanning, and did not include information on previously used antipsychotics. This association therefore remains tentative until replicated.
Various brain regional asymmetries have shown significant heritability in a recent genome-wide analysis of general population data26, including rostral anterior cingulate thickness asymmetry and pallidum volume asymmetry (but not middle temporal gyrus thickness asymmetry). When polygenic risk for schizophrenia was assessed with respect to these heritable asymmetries in a multivariate analysis27, one of the strongest associations was with rostral anterior cingulate thickness asymmetry. The direction of that effect was consistent with the present study, i.e. a rightward shift of asymmetry with increased polygenic risk for schizophrenia. In contrast, pallidum volume asymmetry showed little relation to schizophrenia polygenic risk27, suggesting non-heritable contributions to this association. These genetic findings were established with adult general population data (UK Biobank)27, but together with the current case-control findings, they indicate that altered rostral anterior cingulate thickness asymmetry may be a link between genetic susceptibility and disorder presentation. Left-right asymmetry of the brain originates during development in utero71,86-91, and specific genomic loci that affect brain asymmetry have recently been identified26,92. Some of the implicated genes may be involved in patterning the left-right axis of the embryonic or fetal brain, and genes expressed at different levels on the left and right sides of the embryonic central nervous system were found to be particularly likely to affect schizophrenia susceptibility86. However, other genes may affect brain asymmetry as it changes throughout the lifespan2,93 and therefore may affect susceptibility to asymmetry-associated disorders later in life.
This was the largest study of structural brain asymmetries in schizophrenia to date, and made use of a single image processing and analysis pipeline to support analysis across multiple datasets. However, several limitations also apply. First, schizophrenia is a highly heterogeneous disorder covering a range of possible symptoms, which may correspond to differing underlying disease mechanisms. Our primary analysis only considered case-control group average differences, and in secondary analyses, we did not find significant correlations of asymmetries with major clinical variables within cases after adjusting for multiple testing -including age at onset, duration of illness, and symptom scores. More detailed clinical data would be useful to gather in future large-scale studies of structural asymmetries. Second, we used cross-sectional datasets, limiting the possible interpretation with respect to cause-effect relations or longitudinal changes. Third, we used macro-anatomical brain atlases for both the cortical and subcortical structures, which is the most feasible approach for large-scale analysis across multiple datasets, but limits spatial resolution. With higher resolution mapping, regions that showed negative results in our study may harbor more focal case-control asymmetry differences, which could be revealed for example through vertex-wise cortical mapping58,92,94, or subcortical partitioning into subfields or nuclei. Fourth, data for several variables were only available from a limited number of study sites (medication, handedness, clinical variables), reducing the sample size and thus statistical power in these secondary analyses. Fifth, this study focused on group average differences, but individual-level deviations in affected individuals may be highly heterogeneous and not well captured by group-average approaches95. Future studies may investigate individual or patient subgroup asymmetry deviations from a normative range or structural pattern, which may deliver clinical utility, for example through contributing to diagnosis or prognosis. This concept has shown promising results in recent studies even in smaller samples95,96. The small group-average effects that we identified in the present study are unlikely to have clinical utility when considered in isolation, although they may contribute to multivariate prediction models in future research, for example when considering brain features across multiple imaging modalities.
In summary, we performed the largest study of asymmetry differences between individuals with schizophrenia and unaffected controls to date. Effect sizes were small, but several regional case-control asymmetry differences in cortical thickness and subcortical volume were suggested, and multivariate analysis indicated that 7% of variation across all regional asymmetries could be explained by the case-control group difference. Our findings therefore support a long-standing theory that the brain’s asymmetry can be different in schizophrenia10,11, even if earlier studies in smaller samples were likely to have over-estimated the effect sizes in relation to structural asymmetry.
Altered asymmetry in schizophrenia may conceivably occur during development through disruption of a genetically regulated program of asymmetrical brain development, and/or through different trajectories of lifespan-related changes in brain asymmetries. The specific regions implicated here provide targets for future research on the molecular and cellular basis of altered lateralized cognitive functions in schizophrenia, which may ultimately help to identify pathophysiological mechanisms.
Data Availability
This study made use of 46 separate data sets collected around the world, under a variety of different consent procedures and regulatory bodies, during recent decades. Requests to access the data sets will be considered in relation to the relevant consents, rules and regulations, and can be made via the schizophrenia working group of the ENIGMA consortium: http://enigma.ini.usc.edu/ongoing/enigma-asd-working-group/
Competing interests
O.A.A. is a consultant to HealthLytix. J.B. has received royalties from UpToDate. D.H.M. is a consultant for Recognify Life Sciences Inc., and Syndesi Therapeutics. A.B. received consulting fees by Biogen and lecture fees by Otsuka, Janssen, and Lundbeck. C.M.D-C. has received honoraria from Exeltis and Sanofi. C.Ar. has been a consultant to or has received honoraria or grants from Acadia, Angelini, Boehringer, Gedeon Richter, Janssen Cilag, Lundbeck, Minerva, Otsuka, Pfizer, Roche, Sage, Servier, Shire, Schering Plough, Sumitomo Dainippon Pharma, Sunovion and Takeda. S.K. receives royalties for cognitive test and training software from Schuhfried, Austria. B.F. has received educational speaking fees from Medice GmbH.
Author contributions
D.S., M.C.P., S.E.F., B.F., D.C.G., R.C.G., R.H., N.J., S.E.M., P.M.T., J.A.T., T.G.M.vE., and C.F. conceived and designed the study; D.S., M.C.P., M.F., J.M., K.M., S.M.C.dZ., N.E.M.vH., W.C., H.E.HP., R.S.K., R.A-A, V.O-GdlF., D.T-G., J.V-B., B.C-F., D.A., A.D., L.T.W., I.A., O.A.A., E.G.J., P.K., J.M.B., S.V.C., P.T.M., B.J.M., Y.Q., P.E.R., U.S., R.J.S., V.J.C., M.J.G., F.A.H., C.M.L., C.P., C.S.W., T.W.W., L.dH., K.B., J-K.P., K.G.R., F.St., A.Jan., T.T.J.K., I.N., B.K., O.G., T.D.S., J.B., D.H.M., A.P., V.D.C., J.M.F., S.G.P., J.C., Y.T., Z.W., H.X., F.F., S.E., P.F-C., MA.G-L., A.G-P., R.S., S.S., E.P-C., V.C., F.P., D.V., N.B., G.S., S.M., T.vA., E.W.D., A.N.V., K.S., S.Ci., P.D., R.M.M., W-S.K., Y-C.C., C.An., A.Sc., S.B., A.M.McI., H.C.W., S.M.L., S.dP., H.K.L., F.Sc., R.E., D.G., R.L., U.D., J.T.E., K.R-M., J.M.S., A.R.M., L.A.A., L.F., G.P., A.B., C.M.D-C., J.J., N.G.L., C.Ar., A.S.T., I.L., S.Ce., C.M.S., F.G., M.Kir., S.K., T.H., A.Sk., F.Sp., M.Kim, YB.K., S.O., JS.K., A.Jam., G.B., C.K., M.S., V.O., A.U., F.M.H., D.J.S., H.S.T., A.M.D-Z., J.A.P-Z., C.L-J., S.H., E.J., W.S., P.H., D.C.G., R.C.G., R.H., J.A.T., and T.G.M.vE. acquired and/or analyzed data; D.S. and C.F. drafted the manuscript; D.S., M.C.P., M.F., J.M., K.M., S.M.C.dZ., N.E.M.vH., W.C., H.E.HP., R.S.K., R.A-A, V.O-GdlF., D.T-G., J.V-B., B.C-F., D.A., A.D., L.T.W., I.A., O.A.A., E.G.J., P.K., J.M.B., S.V.C., P.T.M., B.J.M., Y.Q., P.E.R., U.S., R.J.S., V.J.C., M.J.G., F.A.H., C.M.L., C.P., C.S.W., T.W.W., L.dH., K.B., J-K.P., K.G.R., F.St., A.Jan., T.T.J.K., I.N., B.K., O.G., T.D.S., J.B., D.H.M., A.P., V.D.C., J.M.F., S.G.P., J.C., Y.T., Z.W., H.X., F.F., S.E., P.F-C., MA.G-L., A.G-P., R.S., S.S., E.P-C., V.C., F.P., D.V., N.B., G.S., S.M., T.vA., E.W.D., A.N.V., K.S., S.Ci., P.D., R.M.M., W-S.K., Y-C.C., C.An., A.Sc., S.B., A.M.McI., H.C.W., S.M.L., S.dP., H.K.L., F.Sc., R.E., D.G., R.L., U.D., J.T.E., K.R-M., J.M.S., A.R.M., L.A.A., L.F., G.P., A.B., C.M.D-C., J.J., N.G.L., C.Ar., A.S.T., I.L., S.Ce., C.M.S., F.G., M.Kir., S.K., T.H., A.Sk., F.Sp., M.Kim, YB.K., S.O., JS.K., A.Jam., G.B., C.K., M.S., V.O., A.U., F.M.H., D.J.S., H.S.T., A.M.D-Z., J.A.P-Z., C.L-J., S.H., E.J., W.S., P.H., S.E.F., B.F., D.C.G., R.C.G., R.H., N.J., E.L., S.E.M., P.M.T., J.A.T., T.G.M.vE., and C.F. critically revised the manuscript.
Data availability
This study made use of 46 separate data sets collected around the world, under a variety of different consent procedures and regulatory bodies, during recent decades. Requests to access the data sets will be considered in relation to the relevant consents, rules and regulations, and can be made via the schizophrenia working group of the ENIGMA consortium: http://enigma.ini.usc.edu/ongoing/enigma-schizophrenia-working-group/.
Acknowledgments
The ENIGMA project is in part supported by the National Institute of Biomedical Imaging and Bioengineering of the National Institutes of Health (NIH) (Grant No. U54EB020403). The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health. D.S., M.C.P., S.E.F. and C.F. [ENIGMA-Laterality] were funded by the Max Planck Society (Germany). R.A-A. [PAFIP] is funded by a Miguel Servet contract from the Carlos III Health Institute (CP18/00003). J.V-B. [PAFIP] has received unrestricted research funding from Instituto de Investigación sanitaria Valdecilla (IDIVAL): INT/A21/10, INT/A20/04. D.A. [TOP] is funded by the South-Eastern Norway Regional Health Authority (grants 2019107 and 2020086). L.T.W. [TOP] is funded by The Research Council of Norway (223273, 300767), the South-Eastern Norway Regional Health Authority (2019101), and the European Research Council under the European Union’s Horizon 2020 research and Innovation program (ERC StG, Grant 802998). O.A.A. [TOP] is funded by the Research Council of Norway (#223273, #275054) and KG Jebsen Stiftelsen, South East Norway Health Authority (2017-112, 2019-108). P.K. [MPRC, Huilong] received support from NIH grants R01MH123163 and R01EB015611. M.J.G. [ASRB, IGP] received funding from NHMRC Project Grants 630471, 1051672, 1081603. C.P. [ASRB] was supported by a National Health and Medical Research Council (NHMRC) Senior Principal Research Fellowship (1105825), and NHMRC L3 Investigator Grant (1196508). V.D.C. [FBIRN, COBRE] was funded by NIH grants R01MH118695 and NSF 2112455. J.M.F. [FBIRN] was funded by a Senior Research Career Scientist Award, Department of Veterans Affairs. P.F-C. [FIDMAG] is funded by CIBERSAM and by Instituto de Salud Carlos III, co-funded by European Union (ERDF/ESF, ‘Investing in your future’): Sara Borrell Research contract (CD19/00149). G.S. [RomeSL] was funded by Italian Ministry of Health RC17-18-19-20-21/A grants. A.N.V. [CAMH] currently receives funding from the National Institute of Mental Health, Canadian Institutes of Health Research, Canada Foundation for Innovation, CAMH Foundation, and the University of Toronto. K.S. [IMH] received support from research grants from the National Healthcare Group, Singapore (SIG/05004; SIG/05028), and the Singapore Bioimaging Consortium (RP C009/2006). Y-C.C. [JBNU] was supported by a grant of the Korean Mental Health Technology R&D Project, Ministry of Health & Welfare, Republic of Korea (HL19C0015) and a grant of the Korea Health Technology R&D Project through the Korea Health Industry Development Institute (KHIDI), funded by the Ministry of Health & Welfare, Republic of Korea (HI18C2383). U.D. [FOR2107 Münster] was funded by the German Research Foundation (DFG, grant FOR2107 DA1151/5-1 and DA1151/5-2; SFB-TRR58, Projects C09 and Z02) and the Interdisciplinary Center for Clinical Research (IZKF) of the medical faculty of Münster (grant Dan3/012/17). J.M.S. [COBRE] was funded by NIH grant 1P20RR021938-01. A.R.M. [COBRE] received funding from NIH grants P30GM122734 and R01MH101512. C.M.D-C. [Madrid] has received funding from Instituto de Salud Carlos III, Spanish Ministry of Science and Innovation (PI17/00481, PI20/00721, JR19/00024). S.Ce. [KaSP] received funding from the Swedish Research Council (Grant No. 523-2014-3467). M.Kir. [Zurich] acknowledges funding from the Swiss National Science Foundation (P2SKP3_178175). T.H. [ESO] was supported by funding from the Canadian Institutes of Health Research (142255), Ministry of Health of the Czech Republic (16-32791A, NU20-04-00393), and Brain & Behavior Research Foundation Young and Independent Investigator Awards. A.Jam. [Oxford] was supported by MRC grant G0500092. P.H. [SWIFT] is supported by a NARSAD grant from the Brain & Behavior Research Foundation (28445) and by a Research Grant from the Novartis Foundation (20A058). R.C.G. [UPENN] received funding through NIH grants 1R01MH117014 and 1R01MH119219. N.J. [ENIGMA] is funded by NIH grant R01MH117601. S.E.M. [ENIGMA] is supported in part by Australian NHMRC APP1172917. J.A.T. [FBIRN, MCIC, COBRE] is supported by NIH grant R01MH121246. Further acknowledgments specific to datasets are listed in the Supplementary Information.
Footnotes
Additional author added; Updated incorrect reference to a supplementary figure; Minor textual (spelling) corrections; Updated acknowledgments.