Neurodevelopmental risk and adaptation as a model for comorbidity among internalizing and externalizing disorders: genomics and cell-specific expression enriched morphometric study

Mental disorder diagnoses were determined by using parent or guardian responses to the computerized Kiddie Schedule for Affective Disorders and Schizophrenia (KSADS) based on the Diagnostic and Statistical Manual of Mental Disorders, fifth edition (DSM-5) criteria [31]. Lifetime (past or present) diagnoses of the 18 disorders were used [21]. Based on the definition of broad diagnostic families adopted in recent studies [32], two broad diagnostic families including externalizing disorders (attention-deficit/hyperactivity disorder, oppositional defiant disorder, conduct disorder), internalizing disorders (dysthymia, major depressive disorder, disruptive mood dysregulation disorder, agoraphobia, panic disorder, specific phobia, separation anxiety disorder, social anxiety disorder, generalized anxiety disorder, post-traumatic stress disorder) were used in our analysis (Fig. 1A). We excluded children with thought disorders (hallucinations, delusions, associated psychotic symptoms, bipolar disorder, obsessive–compulsive disorder) in the manuscript for two reasons: (1) the sample size of thought disorders is too small (N = 347) compared to that of externalizing/internalizing disorders, and (2) involving thought disorders would make the single diagnostic families (externalizing or internalizing) contain children with comorbidity. We also consider the influence of thought disorders in Additional file 1 [33‐54].

We consider 3 broader diagnostic groups: pure internalizing and externalizing disorders and their comorbidity (see Fig. 1B). Healthy control preadolescents were those who did not have any mental disorders diagnoses (including unspecified or other specified disorders, eating disorders, alcohol use disorder, substance-related disorder, sleep problems, suicidal ideation or behavior, and homicidal ideation or behavior).

Preadolescents aged 9–10 years (N = 11,878) are recruited from 22 research sites across the USA from the Adolescent Brain Cognitive Development (ABCD) Study® (release 3.0, November 2020), which contains physical and mental health, cognition, genetic, and neuroimaging data. The ABCD study group obtained written and oral informed consent from parents and children, respectively. Lifetime psychiatric diagnoses were determined using K-SADS-5. Demographic data of the ABCD sample was listed in Additional file 2: Table S1 and Table S2. We consider four groups: externalizing, internalizing, comorbidities between internalizing and externalizing disorders, and healthy control groups.

T1-weighted structural MRI data were gathered on 3-T MRI systems (Siemens Prisma, General Electric MR 750, Philips). On the basis of standardized processing pipelines [23], structural MRI data processing was collected using FreeSurfer version 5.3.070. All scan sessions completed radiological review whereby scans with incidental results were identified. Participants were removed who could not pass the visual inspection of T1 images and FreeSurfer quality control [55] (imgincl_t1w_include =  = 1). According to the Desikan-Killiany Atlas, the current analysis used post-processed SA and CT data which were mapped to 34 cortical parcellations per hemisphere (68 brain regions in total) [56].

A general psychopathology factor (p-factor) and three sub-factors, externalized disorder (Ext), internalized disorder (Int), and thought disorder (Tho) were modeled using the parent-rated K-SADS-5 [57, 58]. Based on a prior observation from the ABCD study [35], employing a hierarchical model of externalizing (ADHD, ODD, CD), internalizing (MDD, GAD, PTSD, PD, SEP, SAD), thought (hallucinations, delusions, OCD, BP) disorder scores, we derived the p-factor using confirmatory factor analysis (R v4.0, cfa function of the lavaan package). This analysis was based on the whole sample (N = 11,878). We also performed association analyses between the total CBCL scores and the morphometric variables SA extracted from regions affected in the comorbid group compared to healthy subjects. For the CBCL symptom correlations, all children with symptom scores (N = 7570) were included irrespective of the diagnostic classifications, after regressing out the same covariates using LMM.

Before GWAS, we performed genetic ancestry inference, genotype imputation, and strict quality control on genotype data and filtered 4468 genetically unrelated preadolescents with European ancestry who passed structural image quality control (see Additional file 1 and Additional file 2: Tables S3-S6 for details). To control the confounding effects introduced by population stratification, we performed principal component analysis (PCA) on genotype and calculated the top 10 genetic principal components (Pcs) as covariates in GWAS. We identified altogether 15 regions with significantly altered CT in a single diagnosis family (internalizing: 10, externalizing: 5) and 29 regions with significantly altered SA in a comorbidity diagnosis family. To explore the genetic underpinnings of these abnormal regions of CT and SA, respectively, we therefore performed GWAS on the CT (15 regions) and SA (29 regions) using plink V2.0 [59]. Age, sex, mean CT (for regional CT) or total SA (for regional SA), top 10 genetic Pcs, and study sites were included as covariates.

Genomic risk loci were defined using the FUMA [60] online platform (version 1.3.6a). Independent significant single nucleotide polymorphisms (IndSigSNPs) were defined as variants with a p value < 5 × 10⁻⁸ and independent of other significant single nucleotide polymorphisms (SNPs) at r² < 0.6. Lead SNPs were also identified as those independent from each other (r² < 0.1). LD blocks for IndSigSNPs were then constructed by tagging all SNPs with MAF ≥ 0.0005 and in LD (r² ≥ 0.6) with at least one of the IndSigSNPs. The reference panel population was European of the 1000 Genomes Project Phase 3.

On the one hand, SNPs were mapped to genes by a combination of positional, expression quantitative trait loci (eQTL) and 3-dimensional (3D) chromatin interaction mappings. Specifically, positional mapping was mapping SNPs to locus based on their physical positions. In eQTL mapping, SNPs were mapped to candidate genes according to significance criteria (p < 0.05) eQTL associations from Genotype-Tissue Expression (GTEx) [61] v8, the UK Brain Expression Consortium [62] (http://​www.​braineac.​org/​), the Common Mind Consortium [63], and PsychENCODE [64] (http://​resource.​psychencode.​org). We included the major histocompatibility complex region in our FUMA analyses due to the links between the brain, psychiatric disorders, and immune system [65, 66]. Other parameters were consistent with Makowski et al. [67]. On the other hand, to combine the cumulative effects of SNPs assigned to a gene, gene-based analysis was carried out using MAGMA [68] implemented in FUMA. SNPs were mapped to genes within 50 kb upstream and downstream of the gene, a window size that has been used in previous cortical GWAS [69]. Then, the gene-based p values were calculated by the GWAS summary statistics of mapped SNPs, indicating the association between the gene and the GWAS phenotype. Genes significantly associated with each ROI with diagnostic effect on SA and CT were determined by Bonferroni correction (q = 0.05).

To further identify the biological processes underlying regional SA and CT, we performed gene set enrichment analyses on regional SA and CT based on KEGG, GO, and GWAS catalog gene sets. All genes were set as background genes. Bonferroni correction for all analyses was applied through FUMA. Other parameters in these analyses were set as default.

To test whether genetic risk variants for regional SA and CT converge on a specific cell type, we performed cell type specificity analysis [70] using 7 single-cell RNA sequencing datasets from human brain tissue (Additional file 2: Table S7) and pre-computed MAGMA results, which builds the relationships between cell type-specific gene expression and trait–gene associations. We used Bonferroni (q = 0.05) correction for multiple testing per dataset to identify significantly associated cell types.

We consider four groups: externalizing disorders (total N = 883, including attention-deficit/hyperactivity disorder (N = 624), oppositional defiant disorder (N = 369), and conduct disorder (N = 61)), internalizing disorders (total N = 1961, including dysthymia (N = 3), major depressive disorder (N = 59), disruptive mood dysregulation disorder (N = 4), agoraphobia (N = 10), panic disorder (N = 7), specific phobia (N = 1379), separation anxiety disorder (N = 310), social anxiety disorder (N = 172), generalized anxiety disorder (N = 84), post-traumatic stress disorder (N = 24)), comorbid internalizing and externalizing disorders (N = 1054), and healthy control (N = 3672) groups. Demographic data for the ABCD sample is listed in Additional file 2: Tables S1 and S2. Lifetime psychiatric diagnoses were determined using KSADS based on DSM-5. To keep our healthy control group free from overlapping disorders, subjects with any recorded unspecified or specified disorders (eating disorders, alcohol use disorders, substance-related disorders, sleep problems, suicidal ideation or behavior, and homicidal ideation or behavior) were excluded from this group. A total of 1392 subjects were excluded on this basis.

Children with comorbidity had pronounced SA reduction across the brain compared to the controls, while the single diagnostic groups had only a few regions with significant differences compared to the controls (Fig. 2). In particular, 29 out of 68 cortical regions demonstrated significantly lower SA in comorbid children, including the left precuneus (t =  − 4.23, p = 2.4 × 10⁻⁵), right middle temporal gyrus (t =  − 3.45, p = 5.6 × 10⁻⁴), left supramarginal (t =  − 3.16, p = 1.6 × 10⁻³) and prefrontal areas (left pars orbitalis (t =  − 2.96, p = 3.1 × 10⁻³), right pars orbitalis (t =  − 3.15, p = 1.6 × 10⁻³)), and sensory motor regions (right postcentral gyrus (t =  − 3.07, p = 2.1 × 10⁻³), left precentral gyrus (t =  − 2.82, p = 4.8 × 10⁻³)); all p values passed FDR correction (FDR q = 0.05) (see Additional file 2: Table S8). When children with externalizing disorders were compared with healthy children, only 2 temporal regions (right inferior temporal gyrus (t =  − 3.88, p = 1.1 × 10⁻⁴) and left superior temporal gyrus (t =  − 3.21, p = 1.4 × 10⁻³)) demonstrated a significant SA reduction, while no SA reduction was notable in the internalizing disorder group (Figs. 2 and 3). The two temporal regions with reduced SA in externalizing disorder group also showed SA reduction in the comorbid group.

The omnibus ANOVA analysis contrasting the 4 groups (internalizing, externalizing, comorbidity disorder, and healthy control groups) revealed significant changes in accordance with the above case–control results (Additional file 2: Table S9). A post hoc contrast revealed SA reduction affecting left precuneus and right pars triangularis in the comorbid group compared to the internalizing disorders and the control group, while this did not reach significance in comparison with the externalizing disorder group. Taken together, these observations indicate that the SA differences in comorbidity include those changes that are seen in externalizing disorders, at a somewhat greater magnitude; furthermore, SA differences in the comorbid group are more extensive than the minimal, insignificant deviations seen in internalizing disorders. This is also reflected in Fig. 3 in which the comorbid group is more “similar” to the externalizing disorder group than to the internalizing disorder group.

Children with either internalizing or externalizing disorders had significant alterations in CT when compared to healthy children (externalizing disorders: 5 regions, internalizing disorders: 10 regions, see Figs. 2 and 4). The comorbid group had no significant differences in CT in any of the 68 brain regions compared with the healthy children. For externalizing disorders family, auditory (left transverse temporal gyrus (t = 3.30, p = 9.9 × 10⁻⁴) and left superior temporal gyrus (t = 3.11, p = 1.9 × 10⁻³)), sensory-motor (left postcentral gyrus (t = 3.11, p = 1.9 × 10⁻³)), visual (right lingual (t = 2.89, p = 3.8 × 10⁻³)), and prefrontal cortex (left pars orbitalis (t = 2.87, p = 4.1 × 10⁻³)) showed significant CT differences. For internalizing disorders, sensory-motor (left precentral gyrus (t = 4.05, p = 5.1 × 10⁻⁵), right precentral gyrus (t = 3.29, p = 1.0 × 10⁻³) and left paracentral lobule (t = 2.93, p = 3.4 × 10⁻³)), temporal (right inferior temporal gyrus (t = 3.44, p = 5.9 × 10⁻⁴) and right banks of superior temporal sulcus (t = 2.84, p = 4.5 × 10⁻³)), and frontal-parietal cortices (left pars opercularis (t = 3.18, p = 1.4 × 10⁻³), left caudal middle frontal gyrus (t = 2.88, p = 4.0 × 10⁻³), right inferior parietal gyrus (t = 2.89, p = 4.0 × 10⁻³), and left superior frontal gyrus (t = 2.74, p = 6.1 × 10⁻³)) showed significant differences (Additional file 2: Table S10).

The ANOVA contrasting the 4 groups (internalizing, externalizing, comorbidity disorders, and control group) revealed significant differences in the bilateral precentral gyrus, left pars opercularis, temporal cortex (right inferior temporal gyrus and right bank of superior temporal sulcus), and right inferior parietal gyrus (Additional file 2: Table S11). A post hoc contrast revealed higher CT in bilateral paracentral lobules in children with externalizing disorders, compared to both the healthy and the comorbid groups. Children with internalizing disorders had higher CT in the left paracentral lobule and bilateral precentral gyrus compared to the other 2 groups. Taken together, these observations indicate that the CT differences in internalizing and externalizing disorders are extensive and variously distributed, but in the presence of comorbidity, these differences do not co-occur. Instead, they diminish in magnitude, leading to a pattern that is indistinguishable from healthy controls. Finally, ANOVA analysis reveals less significant changes when comparing the comorbid group to the externalizing disorder group than to the internalizing disorder group, suggesting that the comorbid group is more similar to the externalizing disorder group (Additional file 2: Table S11).

We found that alterations of surface area for externalizing and comorbid (internalizing and externalizing) group may reflect an additive effect. In other words, the SA differences in the comorbid state appear to be partially an aggregate of the individual differences that occur in internalizing and externalizing disorders. In contrast, CT in the comorbid group does not satisfy the expectations under an additive model, as the internalizing/externalizing group each had significant CT differences (compared to controls), but the comorbid group did not have more pronounced CT changes, as one would expect under the additive model; instead, comorbid subjects did not differ from healthy controls in their CT.

We also explored if the observed patterns of SA/CT alterations for comorbidity and single diagnostic family still hold true if we included the smaller group with “thought disorders,” a group that was not considered in the original analysis due to the small sample size. With this group included, we still observe a similar pattern of SA/CT alterations (i.e., a larger number of regional changes in the comorbidity group than single diagnostic families for SA and a reversed pattern for CT) (see Additional file 2: Table S12).

We finally explored if the observed patterns of SA/CT alterations for comorbidity and single diagnostic family simply reflect the total “burden” of psychiatric diagnoses (i.e., comorbidity within diagnostic families) or a specific combination of internalizing and externalizing disorders, by examining if there is any difference in CT/SA between children diagnosed with different number of within-family disorders. Specifically, we compared children diagnosed with 1, 2, 3, and > 3 internalizing disorders (or externalizing disorders) and did not find any difference (see Additional file 1). Based on these results, SA/CT alterations were not driven by comorbidity within a single diagnostic family, but rather by comorbidity across diagnostic families.

We found that the p-factor, which reflects an overarching susceptibility to any mental disorder [57, 71, 72], was significantly higher in children with a notable reduction in the SA (irrespective of diagnostic status) in the same cortical regions that were prominently affected in the comorbid group (bilateral precuneus, superior and inferior temporal gyrus, and the left superior frontal gyrus; all p < 0.05, FDR corrected (FDR q = 0.05), see Fig. 5A). However, the p-factor was not correlated with the cortical thickness (Fig. 5B).

Children with lower SA in the cortical regions with a pronounced comorbidity effect also had higher CBCL externalizing problems scores (including rule-breaking behavior scores, aggressive behavior scores, oppositional defiant problems scores, and conduct problems scores) and internalizing problems scores (including withdrawn/depressed scores and depressive problems scores) with p < 0.05, FDR corrected (FDR q = 0.05) (see Fig. 6). In particular, the SA of prefrontal and temporal regions related to both externalizing and internalizing problem scores.

Children with higher CT in the cortical regions affected by externalizing disorders rather than internalizing disorders had higher CBCL externalizing problems scores (including attention scores, aggressive behavior scores, oppositional defiant problems scores, and conduct problems scores) with p < 0.05, FDR corrected (FDR q = 0.05) (see Fig. 7). In particular, the CT of temporal cortex related to externalizing problem scores.

To understand the genetic underpinnings of the difference between SA and CT alterations across single diagnostic families (internalizing or externalizing disorders) and the comorbid group, we performed GWAS, gene set enrichment analysis, and cell type specificity analysis. We first performed GWAS for the 29 ROIs with significant SA differences and 15 ROIs with significant CT differences in the patient groups, using 4468 European-ancestry unrelated individuals. Under the classic genome-wide threshold of p < 5 × 10⁻⁸, we identified 76 genome-wide significant SNPs (after clumping) across 6 regions for CT and 139 genome-wide significant SNPs (after clumping) across 6 regions for SA (see Additional file 2: Tables S13-S14). After correcting the multiple comparisons for all 44 regions, 24 SNPs for CT and 80 SNPs survived the Bonferroni adjustment (p < 1.1 × 10⁻⁹ = 5 × 10⁻⁸/44, 44 is the number of all regions). Next, SNPs were mapped to genes with a combination of positional, eQTL, and 3D chromatin interaction mappings by FUMA. MAGMA gene-based association also identified several significantly associated genes in FUMA. Then, we performed enrichment analyses using all the above genes (Additional file 2: Table S15-16).

The major biological pathways that relate to SA differ from those related to CT. SA-related genes are enriched primarily in craniofacial microsomia (p = 1.72 × 10⁻⁷) and multiple system atrophy (p = 3.72 × 10⁻⁵), while CT-related genes are related to immunoglobulin light chain (AL) amyloidosis (p = 2.02 × 10⁻⁵) (Additional file 2: Tables S17-S18). Makowski et al. [67] have also found genes of the regional cortical area are significantly enriched in multiple system atrophy. Furthermore, in the gene set enrichment analysis, distributed CT alterations were associated with common genes that are related to immune-related biological processes (Additional file 2: Table S19). For example, the AHR gene was linked to CT of the left postcentral gyrus and left precentral gyrus; the TRPM8 gene was linked to CT of the right banks of the superior temporal sulcus and left precentral gyrus; the SKAP2 gene was linked to CT of the bilateral precentral gyrus. AHR is a physiological regulator of myelination and inflammatory processes in the developing central nervous system [73], implicated in psychiatric disorders like major depressive disorder [74]. TRPM8 channel augments T-cell activation and proliferation, which has been shown to be involved in mucosal sensory neurons in the regulation of innate inflammatory responses [75]. SKAP2 is a new regulator of migration and myelin sheath formation [76]. To validate our results in datasets with larger sample sizes, we also used the GWAS summary results from a previous GWAS [36] of 3144 functional and structural brain imaging phenotypes from the UK Biobank. This GWAS used 8428 subjects in the discovery dataset and 3456 subjects in the replication dataset, most of whom were of European ancestry. We also looked up the pathways associated with the same regions of CT and SA in UK Biobank using gene set enrichment analysis (see Additional file 1 and Additional file 2: Tables S20-S23).

Finally, cell-type specificity analysis for the genes associated with SA or CT alterations (Additional file 2: Tables S24-S25) uncovered further differences. SA-related genetic pathways specifically relate to the inhibitory neurons while CT-related cell types included astrocytes and oligodendrocytes. For SA, genes related to regions most affected in the comorbid group (left postcentral gyrus (p = 1.5 × 10⁻⁵) and right fusiform gyrus (p = 6.4 × 10⁻⁴)) had significant associations with In1c (inhibitory neurons). For CT, two internalizing-specific regions (left fusiform gyrus (p = 8.2 × 10⁻⁴) and right banks of superior temporal sulcus (p = 5.0 × 10⁻³)) showed significant associations with astrocytes. Moreover, an internalizing-specific region (left caudal middle frontal gyrus (p = 5.1 × 10⁻⁴)) showed significant associations with oligodendrocyte progenitor cells (OPC).

Our study has several strengths as well as limitations. We examined comorbidity in one of the largest developmental neuroimaging cohorts studied to date; we used multilevel analysis linking genetic variants, cell types, and distinct morphometric variables. Nevertheless, we lacked direct measures of myelination or microglial activity to infer the mechanistic processes in CT alterations in more detail. We also lacked sufficient data to resolve the temporal relationship between brain-based metrics and the behaviors of interest. More direct evidence and verification are needed in future analysis. As the ABCD database has some recently recognized issues with the diagnosis of certain disorders—particularly ADHD and MDD—we remove the children with ADHD and MDD and performed the same analysis, i.e., estimate the difference in cortical thickness (CT) and surface area (SA) between each of three transdiagnostic groups (externalizing, internalizing, and comorbid) and the healthy children group. We still observed a similar pattern of CT/SA changes, i.e., for SA, more brain regions were affected in the comorbidity group than the single diagnostic families, while for CT, more regions were affected in single diagnostic families than the comorbidity group (see Additional file 2: Table S28). Finally, the sample size (N = 4468) of GWAS is relatively small compared with current large-scale GWAS (N > 10 k). Larger datasets with neuroimaging data from adolescent samples are needed to validate our results.