The representation of mental illness along dimensions that cross diagnostic boundaries is progressively replacing the acceptance of distinct psychiatric diagnoses [1]. This paradigm shift is supported by the high comorbidity between disorders [2, 3], their clinical overlap [4], and their shared heritability [5]. This transition holds promise to identify more reliable biomarkers [6] and develop novel treatments [7] for psychiatric diseases. A crucial step in this endeavor is to unravel shared biological mechanisms underpinning different psychiatric illnesses. One hypothesis is that psychiatric illness reflects largely common disruptions in brain circuits encoding core dimensions of cognition and affect [8]. In support of this hypothesis, case-control neuroimaging studies have identified brain circuit disruptions shared across psychiatric disorders [9–11] and a recent study has shown associations between white matter properties and a transdiagnostic behavioral measurement of psychopathology [12]. However, the etiology and interpretation of these associations between neuroimaging findings and diagnostic categories or behaviors remain unclear. They may reflect shared genetic risks, environmental risks, consequences of the disease or behaviors, attempted adaptive responses to illness, or a combination thereof [13]. Understanding causal mechanisms of cross-disorder disease connectomics therefore calls for study designs that go beyond comparing patients with healthy controls [13]. For specific disorders, a few studies have shown that connectomic alterations are present in at-risk individuals (i.e., individuals with a high clinical risk for the disorder, or with an affected parent) providing initial evidence that some connectomic changes may precede the onset of illness [14, 15]. Whether these findings form part of a more general overarching vulnerability to psychiatric illness across diagnostic categories and whether this vulnerability can be imputed to genetic factors, remain unknown.

To begin to address this question, we conducted an integrated analysis of structural connectomes (i.e., comprehensive maps of the white matter connections in the brain) and single nucleotide polymorphisms (SNPs) in healthy children. Our analysis focused on SNPs with an established risk for psychiatric illness or for phenotypes related to brain connectivity (i.e., neurological diseases, behavior-cognition, and brain morphology/function). The aim of this analysis was to assess whether genetic risk factors for psychiatric illness are associated with specific patterns of brain connectivity. We used data from 678 children and adolescents in the Pediatric Imaging Neurocognition and Genetics (PING) cohort—one of the largest neuroimaging and genetic datasets in healthy children [16]. Importantly, the use of a healthy population factors out the impact of illness state, its sequelae, and treatments, while the use of a pediatric population limits the impact of different environmental exposures on the brain [17]. We used canonical correlation analysis (CCA), which has been applied previously to identify a rich relationship between functional connectomes, demographics, and behavior [18]. The primary input to CCA were data for 1877 SNPs and whole-brain structural connectomes obtained from diffusion MRI (dMRI). CCA seeks modes of population covariation which goes beyond testing whether a predetermined linear combination of SNPs for a specific disorder correlates with a brain imaging phenotype. To complement this approach, we also explored relationships between psychiatric polygenic risk scores (PRS) and the structural connectome.

Our analysis revealed five main findings.

First, there was a single, highly significant, mode of genome-connectome covariation (correlation between genetic and connectomic canonical scores: r = 0.74, permutation test: p<10⁻⁴ corrected for multiple comparisons across all modes estimated). It explains 6.43 times more of the variance in connectomes, and 3.33 times more of the variance in SNPs, than does the null distribution of CCA modes. All subsequent modes were non-significant (p>0.1).

Second, SNPs related to a range of psychiatric and behavioral-cognitive phenotypes were significantly correlated with the mode of covariation (Fig. 1A; FDR-corrected p-value < 0.05). SNPs correlated with the mode of covariation were distributed across the genome (Fig. 1B). The allele that significantly correlated positively with the genetic canonical score for the top SNP of each psychiatric phenotype encodes an increased risk for that phenotype. In other words, high canonical scores were associated with an increased risk for a wide range of psychiatric phenotypes. These SNPs map to a range of loci (Supplementary Table 3) which in turn map to a range of ontologies (Table 1). In contrast to these associations with psychiatric disorders, SNPs related to neurological disorders showed no such correlations (FDR-corrected p-value > 0.05) with the notable exception of migraine without aura (r=0.19, FDR-corrected p=0.0015) which parallels the observation that this is the only neurological condition to share heritability with psychiatric disorders [5]. In terms of brain morphology and function, SNPs for facial width and white matter hyperintensities (despite the absence of actual hyperintensities in the PING population) were negatively correlated with the CCA mode. The canonical brain network was found to be virtually identical when SNPs related to brain structure, function, and neurological phenotypes were excluded from the analysis (correlation with the original canonical network r=0.995, p<0.001) demonstrating that these phenotypes do not drive the results. To further explore whether the identified mode of genome-connectome covariation is specific to psychiatric phenotypes, we computed its correlation with SNPs associated with two non-psychiatric diseases with a well-established genetic component—asthma and type 2 diabetes—and found no significant association (r=0.08 and 0.12, FDR-corrected p=0.84 and 0.066 respectively).

At the polygenic level (Fig. 2 and Supplementary Fig. 3 and Supplementary Table 4), the connectomic canonical score was found to be positively and significantly associated with a PRS for schizophrenia (R²=1.1%, uncorrected p=0.0042, permutation-corrected p=0.030), major depressive disorder (R²=1.2%, uncorrected p=0.0026, permutation-corrected p=0.016), alcohol use disorder (R²=1.4%, uncorrected p=0.00088, permutation-corrected p=0.023), a cross-disorder PRS (R²=0.9%, uncorrected p=0.011, permutation-corrected p=0.030), and a combined bipolar-schizophrenia PRS (R²=2.1%, uncorrected p=4.8×10⁻⁵, permutation-corrected p=0.0010). It was also found to be positively associated with a PRS for bipolar disorder (R²=1.1%, uncorrected p=0.0045, permutation-corrected p=0.068) and for autism spectrum disorders (R²=0.7%, uncorrected p=0.020, permutation-corrected p=0.21) but these associations were not statistically significant after correction for multiple p-thresholds. It was not found to be significantly associated with attention deficit hyperactivity disorder (R²=0.2%, uncorrected p=0.27, permutation-corrected p=0.91) or eating disorders (R²=0.2%, uncorrected p=0.17, permutation-corrected p=0.79). These results strongly suggest that the mode of genome-connectome covariation is driven by SNPs associated with psychiatric phenotypes, and that high canonical scores relate to an increased genetic vulnerability to psychiatric illness.

Third, the brain network identified by CCA was made of both positive and negative canonical connection strengths (Fig. 3A–B). By definition, positive (respectively negative) canonical connections tend to be higher (respectively lower) in individuals with high canonical scores. Overall, high canonical scores were negatively correlated with average connection strength (r=−0.17, 95% C.I. [−0.24,−0.09], permutation test p<10⁻⁴) implying that individuals with high canonical scores have on average weaker connectivity. Hubs in the network overlap with three different systems (Fig. 3C; Supplementary Table 5 and 6): the default mode network (including the angular gyri, posterior cingulate cortices, precuneus, and supramarginal gyri), a network of cognitive control [9, 10] (including the anterior and mid-cingulate cortices, the inferior frontal gyri, and the supplementary motor area), and the occipital cortex. Further details on how the anatomy of the brain vulnerability network overlaps with known systems are provided in the Supplementary Information and Supplementary Fig. 4.

Fourth, the canonical brain network was broadly reproduced when the input SNPs were restricted to specific subgroups of psychiatric phenotypes related to psychosis and autism (correlation with the original canonical network r=0.80, p<0.001), mood disorders (r=0.77, p<0.001), and addiction (r=0.54, p<0.001). This finding supports the transdiagnostic nature of the canonical brain network by rejecting the explanation that it is a juxtaposition of subnetworks related to specific categories of psychiatric disorders.

Fifth, connectomic canonical scores were significantly correlated with measurements of social and emotional function and substance exposure, despite these variables being independent from the inputs to CCA (Fig. 4 and Supplementary Fig. 5; permutation test: p=0.029). Canonical scores were significantly associated with increased marijuana, alcohol, sedatives, and caffeine exposure and misuse, panic disorder score, perceived stress, and impulsive behavior while being negatively associated with academic satisfaction. Mediation analysis revealed that the canonical brain network mediates the influence of genes on these variables (permutation test: p=0.006). This finding implies that even though the PING cohort is a healthy population, individuals with higher canonical scores exhibit behaviors that are related to psychiatric vulnerability.

We also conducted various analyses to test the robustness of the results. The mode of population covariation was found to be robust when reproduced in two mutually exclusive subgroups of participants randomly sampled from the population (correlation between the original and reproduced canonical genetic strengths: 0.89 and 0.93; 95% C.I. [0.87,0.91] and [0.92,0.95] respectively; p<0.001 for both; and correlation between the original and reproduced canonical connection strengths: 0.65 and 0.68; 95% C.I. [0.62,0.68] and [0.65,0.71] respectively; p<0.001 for both). It was also found to be robust when both genetic and connectomic data (rather than connectomic data alone) were adjusted for admixture genetic variables (correlation between canonical genetic strengths: 0.86; 95% C.I. [0.83, 0.89]; and between canonical connection strengths: 0.95; 95% C.I. [0.95, 0.96]; p<0.001 for both) and when the dataset was restricted to a more homogenous population of European genetic ancestry (correlation between canonical genetic strengths; 0.78; 95% C.I. [0.74, 0.82]; and between canonical connection strengths: 0.74; 95% C.I. [0.71, 0.76]; p<0.001 for both). Furthermore, it was found to be robust when participants exhibiting behaviors related to psychiatric vulnerability were excluded from the analysis (correlation between the original and subject-reduced canonical genetic strengths: 0.92, p<0.001 and between the original and subject-reduced canonical connection strengths: 0.96, p<0.001) demonstrating that these participants do not drive the results but can rather be considered as extremes on the vulnerability spectrum. A breakdown of the mode of genome-connectome covariation by age, sex, scanner manufacturer, and genetic ancestry is also provided in Supplementary Fig. 6. Finally, it was also robust to the number of principal components (from 50 to 150) used in the preprocessing of the connectomic and genomic data (correlation between canonical genetic strengths: 0.86–0.98; and between canonical connection strengths: 0.85–0.96; p<0.001 for both; Supplementary Fig. 7).

The findings presented here support the existence of a structural brain network encoding genetic vulnerability to psychiatric illness across diagnostic categories. Meta-analyses of case-control studies linking brain functional connectivity to psychiatric disorders had suggested that neuroimaging findings are in part shared across psychiatric diagnoses [9]. But such studies are irrevocably limited by the confounders of illness state, its treatment, and its sequelae. The emergence of common psychopathology factors [4, 12], their heritability [12, 50] and neuroimaging correlates [12, 51, 52] further suggested that a shared neurobiology may explain the clinical overlap between psychiatric diagnoses. However, the etiology of these factors remains unclear as they are only partially heritable [4, 12], and their interpretation in terms of vulnerability to psychiatric illness is merely postulated since prospective cohort studies are lacking.

By contrast, the present findings are solely driven by SNPs that confer an established risk for psychiatric disorders. This enables interpretation of the resulting brain network as a ‘vulnerability network’ and the corresponding connectomic canonical scores as ‘brain vulnerability scores’. Unlike functional brain networks, this vulnerability network is made of white matter fascicles connecting cortical and subcortical areas which are thought to be state-independent [53]—a necessary condition to qualify as an endophenotype [54, 55]. Whether it represents a true endophenotype or is rather a biomarker [55] will ultimately require longitudinal studies following cohorts with high brain vulnerability scores. But the fact that brain vulnerability scores mediate the impact of genetic vulnerability onto behaviors related to psychiatric illness suggests that the vulnerability network forms part of its pathogenesis [54]. The mechanism by which this shared endophenotype turns into different clinical presentations remains to be discovered. Clinical presentations likely emerge from complex pathophysiological pathways involving interactions between environmental and genetic risks. In such intricate pathways, the vulnerability network might be a central element funneling genetic risks through a common neurobiological root.

The vulnerability network overlaps with a network of cognitive control identified by a meta-analysis of case-control studies across major psychiatric diagnoses [9] and with the occipital cortex. The latter is consistent with the growing evidence for the involvement of occipital cortices in psychopathology as part of a speculated disruption in the selection and processing of sensory information [51]. These overlaps thereby replicate previous findings whilst helping elucidate their etiology. Conversely, previous transdiagnostic findings which diverge from the vulnerability network may result from non-genetic causes such as environmental factors [55]. For instance, the finding of a relation between decreased fronto-occipital white matter connectivity and a common psychopathology factor [12] may reflect its known association with adverse childhood experience [56–58]. Finally, the vulnerability network presents two important specificities which had not yet been clearly identified as phenotypes shared between psychiatric diagnoses. First, the association between the mode of covariation and a decreased average connectivity suggests that hypoconnectivity is a shared phenotype which is, at least to a degree, genetically driven and which confers an increased vulnerability to psychiatric illness. This may explain why meta-analyses of case-control studies of white matter integrity more consistently find hypoconnectivity than hyperconnectivity in schizophrenia [59], depression [60], and bipolar disorder [61]. Second, the overlap with the default mode network may explain why the latter is frequently found to be altered in psychiatric disorders. This overlap supports the suggestion that the default mode network may also represent a transdiagnostic phenotype [62] while proposing that its disruption may be present before the onset of symptoms and be attributable in part to genetic factors.

The evolution of the brain vulnerability network through the lifespan and its changes in disease remain to be elucidated. Complex gene-environment interactions may substantially alter the association between the vulnerability network and the genetic architecture of psychiatric disorders in adults (indeed this was a major motivation for conducting the present study in a pediatric population). For instance, it might be that the cumulative effect of environmental risk factors for psychiatric illnesses also results in changes to the brain vulnerability network. In that case, high brain vulnerability scores in adults might occur in the absence of genetic predispositions. Conversely, a child with high brain vulnerability score developing in a favorable environment might reach adulthood without ever receiving a psychiatric diagnosis, perhaps reflecting an effect of the favorable environment upon their white matter microstructure. In an initial attempt to explore this question, we used data from 20,827 participants of the UK Biobank in whom suitable MRI data were available [63, 64] (see Supplementary Information and Supplementary Table 7 for details on the methods). Three imaging-derived phenotypes in UK Biobank overlap with the brain vulnerability network (Supplementary Fig. 8) and were predicted to be negatively associated with the genetic canonical score (if the mode of genome-connectome covariation remains unchanged in adulthood): the FA of the forceps major (most significantly) and the two acoustic radiations. We found that the FA of the forceps major was significantly and negatively associated with the genetic canonical scores (r=−0.02; p=0.0088, Bonferroni-corrected p=0.026). Moreover, it was significantly lower among participants who had a self-reported psychiatric diagnosis compared to other UK Biobank participants (mean standardized FA −0.056 in ‘patients’ vs 0.012 in ‘controls’; 95% C.I. of the difference [0.034, 0.11]; Bonferroni-corrected p=0.00042). These results thus replicate the association of lower occipito-occipital connectivity with genetic predisposition to psychiatric illness (and, even more strongly, with actual psychiatric diagnosis). Further analyses of imaging data from well-phenotyped and genotyped population samples will help clarify the temporal stability and/or dynamic nature of the vulnerability network across the lifespan.

In summary, we discovered a pattern of structural brain connectivity in children which robustly encodes a genetic vulnerability to psychiatric illness which crosses diagnostic boundaries. This brain network shows overlap with a network of cognitive control, the default mode network, and the occipital cortex, and mediates the impact of genetic risk on behaviors related to psychiatric vulnerability. These findings help explain the shared neurobiology underpinning various psychiatric diagnoses and may contribute to understanding the mechanisms underlying the emergence of psychiatric illness and the interventions aimed at preventing it.