Mapping the cortico-striatal transcriptome in attention deficit hyperactivity disorder

Abstract

Despite advances in identifying rare and common genetic variants conferring risk for ADHD, the lack of a transcriptomic understanding of cortico-striatal brain circuitry has stymied a molecular mechanistic understanding of this disorder. To address this gap, we mapped the transcriptome of the caudate nucleus and anterior cingulate cortex in post-mortem tissue from sixty individuals with and without ADHD. Significant differential expression of genes was found in the anterior cingulate cortex and, to a lesser extent, the caudate. Significant downregulation emerged of neurotransmitter gene pathways, particularly glutamatergic, in keeping with models that implicate these neurotransmitters in ADHD. Consistent with the genetic overlap between mental disorders, correlations were found between the cortico-striatal transcriptomic changes seen in ADHD and those seen in other neurodevelopmental and mood disorders. This transcriptomic evidence points to cortico-striatal neurotransmitter anomalies in the pathogenesis of ADHD, consistent with current models of the disorder.

Introduction

Attention deficit hyperactivity disorder (ADHD) is a prevalent, heritable neurodevelopmental disorder characterized by impairing symptoms of inattention, hyperactivity and impulsivity ¹. Considerable progress has been made in identifying the genetic variants that confer risk for the disorder ^{2, 3}. Further advances toward a molecular understanding of ADHD would benefit from a map of the brain’s transcriptome in the disorder. In the absence of post-mortem brain studies, prior studies have relied on imputation from genotype to predict gene expression in different brain regions, and provide indirect evidence for genes related to monoaminergic neurotransmitters, including dopamine and norepinephrine ^{4, 5}. Genes showing altered expression in peripheral blood in ADHD are enriched among networks related to nervous system development, along with lipid metabolism and galactose transcriptome ^6–8. While such work is informative, RNA sequencing of post-mortem brain tissue from those with histories of ADHD and matched, non-psychiatric controls, would provide a more definitive understanding of the how altered gene expression is tied to the disorder. Here, in the first study of the ADHD transcriptome, we uncover the expression of genes and gene networks tied to ADHD.

We focused on the caudate and anterior cingulate cortex (ACC) for four reasons: First, in vivo imaging has demonstrated ADHD-related changes in both regions, with reports of a smaller caudate and a reduction in anterior cingulate cortex area and thickness ^{9, 10} Second, the caudate and ACC are key hubs in networks supporting many cognitive processes implicated in ADHD. For example, functional MRI studies report those with ADHD show hypoactivation of the caudate during inhibitory processing and some attention-demanding tasks and hypoactivation of the ACC as part of the frontoparietal cognitive network ¹¹. Third, both regions are enriched for neurotransmitters that are perturbed in ADHD. Dopaminergic neurotransmission in the caudate has long been implicated in ADHD by pharmacological and in vivo receptor imaging ^{12, 13}. Additionally, genetic studies of both rare and common variation point to metabotropic glutamatergic neurotransmission in ADHD and this excitatory neurotransmitter is richly expressed in both the caudate and ACC ^{2, 14}. Finally, an influential developmental model of ADHD, grounded in cognitive and neuroimaging data, ties its onset to early life alterations in the caudate, whereas the course of the disorder is linked to prefrontal cortical plasticity (including the ACC) in adolescence and adulthood ^15–17. Following this model, it is predicted that ADHD related changes to the caudate transcriptome would be enriched in genes preferentially expressed during early life in the caudate, whereas changes in the ACC transcriptome would be found in genes expressed at later developmental stages.

We also test the hypothesis that differentially expressed genes in the post-mortem ACC or caudate will overlap with genes implicated in ADHD through genome wide association studies, consistent with genetic contributors to transcriptional change ¹⁸. Lastly, since ADHD shares clinical, epidemiological, cognitive, and genomic features with other mental disorders, we asked if transdiagnostic overlap also exists for the transcriptome ^19–21.

Methods

Clinical diagnoses.

Psychiatric diagnosis was conducted by teams at each study site, through DSM-based, clinician interviews with the next of kin and details are given in the Supplementary Methods. Cases were defined as those with a diagnosis of any presentation of ADHD and the principal exclusion criteria were the presence of any major neurological disorder or schizophrenia. Controls were those determined to have no mental illness during the diagnostic process. Twenty-six (12 cases and 14 controls) originated from the National Institute of Mental Health Human Brain Collection Core (HBCC) and 34 from the Neurobiobank (24 originated from the University of Maryland Brain and Tissue Bank [10 cases and 14 controls], and ten samples [5 cases and 5 controls] came from the Brain Tissue Donation Program at the University of Pittsburgh). Procedures at the NIMH Human Brain Collection Core procedures are approved by an oversight committee and the NIH Department of Bioethics and all study sites acquired postmortem tissue and conducted all related procedures under protocols approved by their local IRB. Details of the 60 subjects providing caudate and/or ACC samples used in the final analyses are given in Table 1 (for each region Supplementary Table 1, and for brain bank differences Supplementary Table 1C). Neither race/ethnicity (X²(2)=4.3, p=0.12), sex (X²(2)=.95, p=0.62), nor age at death (F(2, 57)=1.6, p=0.21) significantly differed among processing batches/brain banks. Following quality control procedures, 49 provided both caudate and ACC tissue, four provided ACC only and seven caudate only (Supplementary Figure 1 shows quality control steps).

Table 1:

Demographic and clinical details of the 60 donors. Following quality control procedures, 49 provided both caudate and ACC tissue; 7 ACC only and 4 caudate only.

		ADHD	UNAFFECTED	TEST OF GROUP DIFFERENCE
Total		27	33
Age	Mean (SD; range)	21.3 (8.4; 6.7 to 38)	22.9 (8.0; 13.2 to 38.8)	t(55)=−.078, p=0.44
Sex	Male	23 (85.2%)	24 (72.7%)	X2(1)=0.72, p=0.40
Race/ethnicity	White (White Non-Hispanic)	21 (77.8%)	13 (39.4%)	X2(1)=7.4, p=0.006
	Other	6 (22.2%)	20 (60.6%)
Evidence level	Confirmed	15 (55.6%)	27 (81.8%)	X2(1)=3.7, p=0.054
	Investigator impression	12 (44.4%)	6 (18.2%)
Comorbidities	Depression	8	NA
	GAD/panic disorder	2	NA
	Adjustment disorder	2	NA
	Bipolar affective disorder	1	NA
	Autistic spectrum disorder	1	NA
Substance use	No	14 (51.9%)	33 (100.0%)
	Yes	13 (48.1%)”	NA
Manner of death	Accident	10	8	Exact p=.04
	Natural	8	12
	Homicide	1	9
	Suicide	8	4
Post-mortem interval	Mean (SD; range) hours	27.3 (15.4; 7 to 69)	20.4 (10.5; 5 to 56.5)	t(44)=−1.99, p=0.05
RINe	Mean (SD; range)	ACC: 5.2 (.8; 4 to 6.7) Caudate: 6.0 (.5; 5.3 to 6.9)	ACC: 5.4 (.7; 4 to 6.5) Caudate: 5.9 (.7; 4.5 to 7.1)	ACC: t(46)=−1.2, p=0.22 Caudate: t(54)=.88, p=0.38

Tissue preparation

Brain tissues were dissected from frozen coronal slabs cut at autopsy. Anterior cingulate cortex tissue was targeted anteriorly, above the genu of the corpus callosum, and caudate tissue came from the head of the structure (see Supplementary Figure 2). Tissue was frozen at −80C and stored at −80C long-term. For dissections, they were held in −80C and retrieved in small batches and placed on dry ice, where all photography and dissections took place. Each sample was on dry ice for approximately 30 minutes.

There were 60 donors. Sufficient RNA was extracted for sequencing for the ACC in 56 subjects and for the caudate in 59 subjects. Total RNA was isolated from approximately 50–100 mg homogenized tissue in QIAzol using QIAGEN RNeasy Lipid Tissue Mini Kit according to manufacturer protocol. RNA Integrity Number (RINe) was analyzed using the Agilent 4200 TapeStation system and the Agilent RNA ScreenTape assay. RINe did not differ by diagnosis (Supplementary Table 1), nor by brain bank of origin (F(2, 106) =1.2, p = .30).

RNA sequencing

RNA-seq was performed at the National Intramural Sequencing Center using Illumina NovaSeq 6000, 2×150bp following Ribo-Zero GOLD treatment to remove mitochondrial RNA and cytoplasmic rRNA. Samples were sequenced in four batches with samples run in single lane within each batch. The RNA-seq read mapping used STAR, an alignment software package capable of mapping spliced RNA-seq reads to a genomic reference sequence (GRCh38). Gencode gene annotations (V31) were used as gene models during mapping of RNA-seq reads. The software package QoRTs ²² was used to obtain quality control metrics from the RNA-seq data. This included assessment of read quality (Phred quality scores), nucleotide composition of sequence reads, and mapping quality of RNA-seq reads. The software package RSEM was used to count the RNA-seq data. After sequencing, one caudate sample was removed due to high splice junction rate and outlying gene-body coverage metrics.

Differential gene expression

The gene expression data were subjected to principal components analysis. Visual inspection of the first 2 principal components (Supplementary Figure 3) revealed five outliers that were removed from further analysis. The final dataset was comprised of 53 ACC samples (24 ADHD and 29 unaffected) and 56 caudate samples (24 ADHD and 32 unaffected).

Gene read counts for ACC and Caudate were analyzed separately in R (version 4.0.3). Genes with low expression were removed following Chen et al. ²³, implemented in function filterByExpr in edgeR (version 3.32). DESeq2 (version 1.30.1) was used to estimate diagnostic differences in gene expression, through negative binomial generalized linear models, incorporating data-driven prior distributions in its estimates of dispersion and logarithmic fold changes. Independent filtering as implemented in DESeq2, was also used to account for outliers in the data ²⁴. We considered a range of variables that have been associated with gene expression in prior studies: demographic/clinical features (age at death, gender, comorbidities, mode of death, clinical evidence level), genotypic (the first five population components - C1 through C5), and technical covariates (RNA-seq batch and brain bank of origin, post-mortem interval, and RINe) for inclusion in the model (Supplementary Figure 4). Group differences among variables are listed in Supplementary Table 1. To select the variables, we extracted the principal components of the gene expression data and retained the components with eigenvalues above one ²⁵ (R package nFactors, version 2.4.1). The first seven principal components were retained for the ACC, accounting for 68% of variance, and eight principal components were retained for the caudate, accounting for 71% of variance. Spearman correlations tested for associations between these principal components and continuous technical and demographic covariates, while a Kruskal-Wallis test was used for categorical covariates. Covariates associated with any principal component at a Bonferroni corrected p-value < .05 were retained in the final model. For both ACC and Caudate, RINe and brain bank/batch were selected. Finally, we also included variables associated with diagnosis at a Bonferroni corrected p-value < .05, namely, the first population component (C1), comorbidity, and substance abuse. Detailed steps of analytical steps are provided in the supplemental methods.

The full model used was thus:

$$geneExpression~C\mathit{1}+batch\_brain\_bank+RINe+comorbidity+substance\_abuse+Diagnosis$$

where C1 is the first population component, derived from genotype data, and batch_brain_bank is a variable indicating RNA-seq batch/brain bank of origin. The coefficient for Diagnosis determined the significance of each gene in differentiating ADHD from unaffected individuals. We also report DESeq2’s posterior log2 fold changes, considered non-standardized effect sizes that can be comparable across experiments ²⁶

Gene set enrichment analysis

Gene set enrichment analyses were conducted on the gene-level results from the DEG analysis, ranked using the sign of the log-fold change times the negative log₁₀(p value). We tested for enrichment of genes expressed in the ACC and caudate across the lifespan and those confined to different developmental stages (prenatal, infant (0–2 years), child (3–11 years), adolescent (12–19 years), and adult (>19 years). These gene sets (Supplementary Table 2) were derived from the Allen Brain Atlas to define region-specific genes that are differentially expressed in the ACC but not in the caudate, and vice-versa (package ABAEnrichment, version 1.20, quantile cut-off .9, brain structures MFC_anterior/rostral cingulate or medial prefrontal cortex - Allen:10278 and STR_striatum - Allen:10333). In these analyses, the gene set enrichment test for each developmental stage is independent; a statistical comparison across stages is not possible as different genes are expressed at different developmental stages. However, a normalized GSEA enrichment score can be determined for each developmental stage, which allows a contrast of gene enrichment at each stage, corrected for the number of genes expressed. The contrast of normalized GSEA enrichment scores gives an impression of how ADHD related change in the transcriptome aligns with genes expressed at different stages, but should not be interpreted as the results of direct, statistical contrasts.

Additionally, we tested for enrichment of non-redundant genes sets defined in Gene Ontology (http://www.geneontology.org) under the domains of Biological Processes, Cellular Components, and Molecular Functions using WebGestalt (R, version 0.4.4).

Code availability:

Code used for analyses and figures is deposited here: https://zenodo.org/badge/latestdoi/505945767, with the doi: 10.5281/zenodo.6798439. Data is being deposited in NIMH Data Archive under Collection 3151, experiment 2056 (https://nda.nih.gov/edit_collection.html?id=3151), with the doi: 10.15154/1527972.

Transdiagnostic transcriptomic signatures

To contrast the transcriptome in ADHD against other disorder, transdiagnostic Spearman correlations were calculated between the absolute value of the log-fold change in the post-mortem transcriptome of ADHD and the transcriptome for other disorders, using 10,000 bootstraps (sampling with replacement)- (for exact brain regions used for each diagnosis see Supplementary Table 3). The p-value was calculated against the null distribution of studies of the same disorder.

Common variant risk for ADHD and gene expression

We tested the hypothesis that DEGs in ACC and caudate would overlap with genes implicated in ADHD through GWAS. The 2019 Psychiatric Genomics Consortium ADHD GWAS data release was used along with the concatenation of European and African-American samples from 1000Genomes as the reference data to estimate linkage disequilibrium between SNPs ²⁷. These analyses used Multi-marker Analysis of GenoMic Annotation (MAGMA) ²⁸, which relates GWAS SNPs ¹⁸ to expressed genes by their genomic positions. MAGMA then employs a SNP-wise sum model to test mean SNP association to a gene-level continuous variable (the same gene ranks used for GSEA). Reference data from the 1000Genomes project, matching the specific GWAS populations being studied, were used for linkage disequilibrium estimation used by MAGMA.

We also investigated whether DEGs for ADHD in ACC and caudate would overlap with genes implicated through GWAS for other disorders, specifically autistic spectrum disorders ²⁹, major depression ^{30, 31}, bipolar affective disorder ³², schizophrenia ³³, Tourette Syndrome ³⁴, obsessive compulsive disorder ³⁵, post-traumatic stress disorder ³⁶. Finally, we used GSEA to ask whether copy-number variants in genes previously implicated in ADHD would overlap with our differential expression post-mortem results (list in Supplementary Table 4).

Genotyping

Genotyping was conducted on DNA extracted from brain tissue using the Illumina HumanOmniExpressExome-8v1–4 array with genome build GRCh38. Quality control measures are given in the Supplementary Methods and 646,902 SNPs were retained for further analysis. Multi-dimensional scaling in PLINK was used to characterize the population structure (Supplementary Figure 5) and the first five dimensions were retained in analyses.

Robustness analyses

We repeated analyses using different subgroups: the largest genotypic subpopulation (white, non-Hispanic, defined genotypically), and subjects without major depressive disorders.

We also re-ran the analysis without entering comorbidity and substance abuse as covariates, as these share genetic risk with ADHD, and controlling for comorbidity could thus impact the detection of such shared genetic/transcriptomic contributions. Finally, we also repeated the TWAS using a leave-one-out cross validation approach, and report on the stability of the association between ADHD and gene expression levels across the different iterations. Data and code sharing information can be found in the supplementary methods.

Results

Cortico-striatal transcriptome profiling in ADHD

RNA sequencing was used to quantify gene expression in the caudate and ACC, both key hubs in the cognitive control and attentional brain systems that are disrupted in ADHD. We obtained an average of 109 million reads per sample (min 69, max 178, standard deviation 40) in the ACC, detecting 24,945 genes and 17.1% unmapped reads. In the caudate, there was an average 114 million reads per sample (min 76 million, max 196 million, SD 33) with 24,767 genes detected and 16.1% unmapped reads.

Fourteen genes showed significant (FDR q < .05) diagnostic differential expression of genes (DEG) in the ACC (Figure 1, Supplementary File 1 [which include posterior log₂ fold changes, which are considered non-standardized effect sizes], and Supplementary Table 5). Twelve were protein coding and two were long non-coding RNA; all showed upregulation in ADHD (Supplementary Figure 6). Two genes were previously implicated in ADHD (JPH2, Junctophilin Type 2) ³⁷ and alcohol dependence (KIAA0040, uncharacterized protein) ^38–40. The differentially expressed gene ADAM-TS9 (A Disintegrin-Like And Metalloprotease, Reprolysin Type, With Thrombospondin Type 1 Motif, 9), has been associated with a range of neurocognitive features including white matter integrity in bipolar disorder ⁴¹, neuroticism ⁴², brain surface area and subcortical volume ⁴³, and general cognitive ability ⁴⁴. Several genes are involved in neuroplasticity: ANGPTL4 promotes angiogenesis and neurogenesis in a mouse model of acute ischemic stroke, PDPN is involved in neural progenitor cells proliferation, and ADAMTS9 mediates synaptic plasticity, neurorepair ^45–47.

Figure 1.

A) Volcano plots for ACC and Caudate. Horizontal lines indicate FDR q < .05. Vertical lines show log₂ fold change greater than +−1, and can be read as non-standardized effect sizes. Significant genes highlighted in red. B) Manhattan plots for ACC and Caudate. Horizontal red lines indicate FDR q < .05. Protein coding genes show their HGNC symbols, and lncRNA hits are marked by their Ensemble Gene ID.

The gene showing significant differential expression in the caudate (HILPDA, Hypoxia Inducible Lipid Droplet Associated) also showed differential expression in the ACC. HILPDA has also been found to be differentially expressed in the hippocampus of those with major depressive disorder ⁴⁸ and to be upregulated in white matter lesions in multiple sclerosis ⁴⁹.

Gene pathway analyses were performed to interrogate gene expression differences beyond those attaining stringent levels of significance, and to explore possible biological effects of differentially expressed sets of genes. These analyses consistently pointed to gene sets involved in neurotransmission. Using Molecular Function Gene Ontology, DEGs in the caudate enriched glutamate receptor genes, DEGs in the ACC implicated serotonin and GABA receptor activity, and the broad gene set of neurotransmitter receptor activity was enriched by DEGs in both regions. (Figure 2A, Supplemental File 1).

Figure 2.

A) Gene Set Enrichment Analyses (GSEA) for Molecular Function Gene Ontology. All significant sets (FDR q < .05) for each region are displayed. B) GSEA results for top 10 Cellular Components Gene Ontology. Gene sets pertaining to neural features and neurotransmitters are indicated in red. Dotted line indicates FDR q < .05, dashed line FDR q < .1.

When examining sets in a Cellular Component Gene Ontology, DEGs in the caudate also implicated processes pertaining to the glutamatergic and GABAergic synapses, as well as synaptic processes in general, and the neuronal spine (Figure 2B, Supplementary Figure 7, Supplemental File 1 for full results). Semantic space analysis (REVIGO)⁵⁰ highlighted the salient biological processes, including postsynaptic specialization, the neuron spine, and glutamatergic transmission (Supplementary Figure 8).

Finally, GSEA results for Biological Processes largely followed a similar pattern. DEGs in the caudate enriched gene sets related to cognition, neuronal projection, dendritic development, synaptic processes, as well as glutamate and dopamine receptor signaling pathways (Supplementary File 1). ACC significant results (FDR q < .05) included serotonin receptor signaling pathway. The results of a leave-one-out cross validation approach in the main association analyses are given in Supplemental Table 6, and point to stability in the results across the different iterations.

DEGs and genes expressed at different developmental stages

A prediction stemming from a leading developmental model of ADHD is that caudate DEGs would preferentially enrich genes preferentially expressed in the prenatal/infantile caudate, whereas ACC DEGs would enrich genes preferentially expressed in the ACC in adolescence and adulthood. Consistent with this prediction, the caudate DEGs were preferentially expressed in the prenatal and infant caudate (Figure 3, Supplementary File 1). In contrast, the ACC DEGs are expressed in primarily during adulthood, although enrichment in infant life also hovered at significance levels. Normalized GSEA enrichment scores are given (Supplemental File 1; GSEA developmental tabs) to provide an impression of how ADHD related change in the transcriptome aligns with genes expressed at different stages. This dissociation is broadly consistent with concept that altered gene expression in the caudate is more readily detected among genes preferentially expressed in early life, whereas altered gene expression in the ACC may also encompass genes expressed at later developmental stages.

Figure 3.

A) GSEA results for region-specific developmental sets in ACC and Caudate. Color scale indicates the absolute log₁₀ of the nominal p-value, using the sign of the normalized enrichment score (blue meaning upregulation in ADHD during that developmental period, and red downregulation in ADHD). Results marked with thicker borders are significant after correction for multiple comparisons (FDR, q < .05).

Common and rare genetic risk for ADHD and differential gene expression

In line with expectations, by using MAGMA analyses we found DEGs implicated in ADHD through GWAS overlapped significantly with DEGs in both the post-mortem ACC (p = 0.003), and caudate (p = 0.02). Considering rare variant risk for ADHD, caudate DEGs show a trend towards enrichment among genes that harbor CNVs in ADHD (p=0.06) but this relationship was not found for ACC DEGs (p=0 .91).

Transdiagnostic transcriptomic signatures

We looked at correlations between DEGs in ADHD and other psychiatric disorders (Figure 4A). For the caudate, significant correlations were found between the ADHD transcriptome and autistic spectrum disorder (median rho=.19; 95^th centiles 0.121 to 0.252, p < .0001), bipolar affective disorder (.164 [0.079 to 0.377], p=.008), major depression (.07 [0.053 to 0.084], p<.0001), obsessive compulsive disorder (.305 [0.288 to 0.321], p<.0001), and schizophrenia (.189 [0.111 to 0.265], p<.0001). For the cingulate cortex, significant correlations emerged between the ADHD transcriptome and the transcriptome in autistic spectrum disorder (median rho=.166; 95^th centiles 0.102 to 0.257, p < .0001), and major depression (.186 [0.066 to 0.397], p=.02), but not for schizophrenia (.216 [0.053 to 0.461], p = .056) or bipolar affective disorder (.197 [0.031 to 0.447], p=.22). This finding adds ADHD to the list of disorders that show similarities in their neural transcriptomic profile.

Figure 4.

A) Correlation between differentially expressed genes in the ACC and Caudate in our dataset and other disorders. Number indicates the data source: [1] Gandal et al. 2018 (microarray), [2] Gandal et al. 2018 (RNAseq), [3] Akula et al. 2020, [4] Benjamin et al. 2020, [5] Pacifico and Davis, 2017, [6] Piantadosi et al. 2021, [7] Wright et al. 2017, [8] Parikshak et al. 2016.

B) Enrichment of DEGs in the ACC and caudate among genes implicated by GWAS of other disorders (analyses using MAGMA). The ‘X’ indicates non-significant enrichment (p < .05). The circle size indicates the absolute log₁₀ of the nominal p-value.

ASD: Autism Spectrum Disorder; OCD: Obsessive Compulsive Disorder; BD: Bipolar Disorder; SCZ: Schizophrenia; TS: Tourette Syndrome; MDD: Major Depressive Disorder; PTSD: Post-traumatic Stress Disorder; AUD: Alcohol Use Disorders. ASD, TS, and AUD results use European population only for estimating linkage disequilibrium.

The DEGs in the post-mortem ADHD brain were also significantly related to SNP risk for schizophrenia (ACC: p = .002; caudate: p = 5.2*10⁻⁹) and bipolar disorder (ACC: p = 3.1*10⁻⁵; caudate: p = 2.8*10⁻⁵) – Figure 4B. Notably, none of the ADHD participants in this study had schizophrenia and only one had bipolar affective disorder as a comorbidity. Genes differentially expressed in the post-mortem ACC, but not the caudate were nominally significant when tested for association with SNP risk for major depressive disorder (p = .04).

Analyses confined to the largest subpopulation.

We repeated analyses using the largest subpopulation (white, non-Hispanic). A similar pattern of results emerged (Supplementary Table 5, Supplementary Figures 9–12, Supplementary File 2). Specifically, two genes showed significant differential expression in the ACC at FDR q<0.05 for both the entire cohort and white non-Hispanic subpopulation (CCDC13 and HILPDA, and six genes showed differential expression at FDR q<0.1 for both cohorts (MYO1G, ANGPTL4, CCL2, KIAA0040, JPH2, and ITPK). The single significant caudate DEG in the initial analysis did not hold for the white, non-Hispanic cohort only. GSEA analyses also very similar results when performed on the white, non-Hispanic cohort compared to the entire cohort. Thus, for the white, non-Hispanic cohort alone, DEG in the ACC significantly enriched genes expressed in the ACC in adult life, and DEG in caudate enriched genes expressed in the infant caudate (Supplementary Figure 9, Supplementary File 2). Similarly, DEG in the white, non-Hispanic cohort showed enrichment of gene sets pertaining to neurotransmitter receptor activity, including serotonin and GABA in the ACC, and glutamate in the caudate (Supplementary Figure 10, Supplementary File 2). Association between DEG in the caudate and SNP risk for bipolar affective disorder (p=0.0007) and schizophrenia (p=0.03) held, as did associations between DEG in the ACC and SNP risk for bipolar affective disorder (p=0.001) and major depressive disorder (p=0.04) (Supplementary Figure 11). Finally, the correlations between transcriptome across different neuropsychiatric disorders held for the white, non-Hispanic group (Supplementary Figure 12).

Results also held in analyses that did not covary for comorbid disorders (Supplementary File 3). At the single gene level, nine genes remained significantly associated with ADHD (Supplementary Table 5). Additionally, DEG in the ACC significantly enriched genes expressed in the ACC in adult life, whereas DEG in caudate enriched genes expressed in the infant caudate (Supplementary Figure 13). The pattern of enrichment of neurotransmitter gene sets (Supplementary Figure 14) also closely resembled that seen for the main analysis. Correlations between transcriptome across different neuropsychiatric disorders held, as did the pattern of associated SNP risk (Supplementary Figures 15, 16). Findings were also robust when those with comorbid MDD were removed (Supplementary File 4, Supplementary Tables 5, 7 and Supplementary Figures 17 to 20). A summary of all robustness analyses is given as Supplementary Table 7.

Discussion.

The study has four central findings. First, we report significant differential expression of genes in the post-mortem ACC and to a lesser extent the caudate nucleus among those with ADHD, which includes genes previously associated ADHD, other neuropsychiatric disorders, and neuroplasticity. Secondly, DEGs in both the caudate and ACC enriched gene sets associated with neurotransmitter activity, most prominently glutamate, but also serotonin, GABA, and dopamine. Third, in keeping with an influential developmental model of ADHD ^15–17, DEG in the caudate was most pronounced among genes showing preferential early life expression in the caudate, whereas DEG in the ACC was most prominently found in genes expressed in adulthood. Finally, we report transdiagnostic similarities in the post-mortem transcriptome between ADHD and other neurodevelopmental and mood disorders, mirroring their genetic overlap. The overall findings held when analyses were confined to the study’s largest, white, non-Hispanic subpopulation, as well as in other robustness analyses.

Fourteen genes showed transcriptome wide significant differential gene expression in the ACC, with upregulation in those with ADHD. Several of these genes have been implicated in other mental disorders and in cognitive processes related to ADHD. For example, KIAA0040 has been implicated in alcohol dependence ³⁸, MYO1G in bipolar affective disorder ³², and JPH2 has been associated with symptoms of inattention ³⁷. Associations have also been reported with cognitive domains pertinent to ADHD such as reaction time (KIAA0040, MYO1G), cognitive flexibility (CRISPLD1) and general cognitive ability (ADAMTS9) ^{44, 51, 52}. Furthermore, both the caudate and ACC showed significant overexpression of HILPDA, a gene previously associated with major depressive disorder ⁴⁸, and this association held for the ACC even after removing those with comorbid MDD.

Genes showing altered expression consistently enriched neurotransmitter pathways, most strikingly, glutamatergic gene pathways in the caudate. Glutamate is the major excitatory neurotransmitter in pathways that link the caudate with both the cortex and thalamus. These loops mediate many of the cognitive and affective processes disrupted in ADHD, including the control of attention and reward processing ⁵³. Altered expression of glutamatergic genes in the caudate thus fits well with current models of ADHD. The finding also complements genetic studies that point to glutamatergic genes in ADHD. For example, there is an enrichment in ADHD of rare, copy number variants in metabotropic glutamate receptor gene networks ¹⁴. These glutamatergic CNVs have been associated with cognitive and clinical impairments seen in ADHD⁵⁴ and a glutamate receptor activator drug, fasoracetam, reduces ADHD symptoms among those carrying CNVs in glutamatergic gene pathways⁵⁵. Additionally, GWAS of ADHD have also found that common variant risk for ADHD enriches both metabotropic and ionotropic glutamatergic genes and pathways ^{2, 56}. In short, anomalies of glutamatergic neurotransmission have been implicated in ADHD through pharmacological and genomic studies: here we add the transcriptome.

Other major transmitters of the cortico-striato-thalamic circuits including GABA, also showed downregulation in both the ACC and caudate. A role for GABA has been suggested through the finding of a GABA mediated reduction of short interval intra-cortical inhibition in the motor cortex in those with ADHD ⁵⁷. In addition, some though not all studies find reductions of GABA in those with ADHD as measured by in vivo magnetic resonance spectroscopy ⁵⁸. Combined with our finding of downregulation of GABA gene pathways, it seems possible that reduced GABAergic inhibition that may contribute to the poor inhibitory control that is central to ADHD.

We find downregulation of dopaminergic pathways in the caudate that are pivotal in the pathogenesis of ADHD. Dopamine has long been implicated in ADHD through the action of psychostimulant treatment for ADHD in altering striatal dopamine tone, in vivo imaging of dopamine receptors, and theoretical models ^{12, 13}. There are also rich interactions between dopaminergic and glutamatergic pathways. Midbrain dopaminergic neurons act as neuromodulators that can alter glutamatergic synaptic transmission and this interaction is pivotal in cortico-striatal function ⁵⁹. Finally, ephrin receptor activity was also implicated, a notable finding given that these receptors are pivotal during embryonic development in axonal guidance ⁶⁰. Ephrin signaling also plays a role in adult neuroplasticity, as interactions between ephrin and post-synaptic glutamate receptor trafficking support processes such as memory formation and learning. Our findings of altered expression of ephrin and dopaminergic processes are consistent with implication of these gene pathways by GWAS at nominally significant levels ².

Our transcriptome findings speak to models of the distinct roles of subcortical and cortical regions in ADHD across the lifespan. Specifically, events in deep brain structures such as the caudate are held to trigger the onset of ADHD. In support of this view, functional imaging studies have found that subcortical anomalies are associated with a childhood history of ADHD, reflecting early developmental events in the caudate tied to the onset of ADHD ^{16, 17, 61, 62}. Consistent with this model, we find DEGs in the post-mortem caudate enrich genes that are known to be preferentially expressed in the infantile caudate. By contrast, neuroimaging studies suggest that processes reflecting the adult outcomes of ADHD and not just its onset, occur at mainly prefrontal cortical levels ¹⁶. This model predicts, and indeed we find, that DEGs in the post-mortem ACC would most prominently enrich genes that are known to be expressed in the adult ACC. As noted earlier, our developmental analyses give an impression of how ADHD related change in the transcriptome aligns with genes known to be preferentially expressed at different stages of development. The ideal analysis would directly test for differences in gene expression in post-mortem tissue that is acquired in early childhood, late childhood, adolescence and early adulthood. Such analyses are not possible with the current sample as a restricted post-mortem age range was covered, concentrated around adolescence and early adulthood with under-sampling of early childhood. However, a developmental focus on the post-mortem transcriptome would be a valuable direction for future work.

We can speculate about the role of specific genes and gene pathways within this developmental framework. For example, glutamatergic synapses play a pivotal role in neuroplasticity ⁶³, mediating long-term potentiation relating to memory among other processes. Additionally, in vivo neuroimaging suggests that atypical concentrations of cortical glutamate may affect microstructural and functional organization of the cerebral cortex ⁶⁴ further suggesting a possible mechanism for cortical plasticity in processes relating to remission in ADHD. In this regard, it is notable that several of the most significantly differentially expressed genes play a role in neuroplastic processes, promoting neurogenesis (ANGPTL4), neural progenitor cell proliferation (PDPN) and synaptic plasticity (ADAMTS9) ^{45, 46}

We find that polygenic risk for ADHD was significantly associated with transcriptomic profiles both in the caudate and ACC. This alignment of genetic and transcriptomic findings in ADHD could reflect a common underlying pathophysiology and hints at an etiological role for the gene expression changes. This finding also makes alternatives, such as the altered transcriptome merely reflecting a downstream consequence of early symptoms of ADHD, less plausible. There was also an association between DEGs and polygenic risk for other mental disorders, including autistic spectrum disorder, schizophrenia, bipolar affective and major depressive disorder. Similarly, while it is well-established that many psychiatric disorders share genomic risk ¹⁹, we demonstrate an overlapping transcriptomic profile between ADHD and other psychiatric conditions.

The study has limitations. The sample size is modest, mainly due to the challenges in obtaining post-mortem tissue from those who have a clear history of ADHD. This inevitably increases the risk of false positives and negatives and expansion of this post-mortem line of research is needed to replicate and extend these initial findings. Additionally, the cohort included individuals with psychiatric comorbidities, most commonly major depressive or substance use disorders. However, we controlled for the presence of these comorbidities in our analytic model, thus attenuating confounding effects. The cohort was not racially and ethnically uniform, and while diversity is a strength, it presents analytic challenges and we thus repeated analyses confined to the cohort’s largest subpopulation, defined genotypically. Findings held from the level of individual DEG findings through gene sets that were enriched to the transdiagnostic similarities in the transcriptome and underlying genetic risks. Finally, RNA sequencing was performed on bulk tissue homogenates, an important first step in transcriptomics that can inform more focused sequencing of single nuclei or enriched neuronal populations (e.g. glutamatergic). While integrating post-mortem specimens from different biobanks is challenging, tissues were handled in a uniform manner across sites and a uniform processing pipeline at one site was used for the extraction and sequencing of RNA. Furthermore, our analyses consider potential technical confounders including originating brain bank, batch, and RIN quality.

In conclusion, this study of the ADHD post-mortem transcriptome demonstrates alterations in cortico-striatal gene expression that inform both developmental and neurotransmitter-based models of the disorder.