Introduction

Attention-deficit/hyperactivity disorder (ADHD) affects ~3–5% of children worldwide1 and places a significant burden on individuals, their families, and the community2. ADHD is highly heritable (~70–80%)3, so identifying risk genes will increase our understanding of underlying biological processes. Recent case–control genome-wide studies have identified ADHD risk loci by assessing common single-nucleotide polymorphisms (SNPs) through genome-wide association studies (GWAS)4,5. However, to date, SNP-heritability has only accounted for a small portion (~15–30%) of overall heritability estimates, suggesting that other genetic factors, including rare genetic variants, may play an important role in ADHD risk6. Indeed, previous studies have demonstrated that rare copy number variants7 and very rare protein-truncating variants in evolutionarily constrained genes8 are enriched in ADHD. Despite previous research considering these different categories of genetic variation, no specific high-confidence risk genes have yet been identified for ADHD.

Detecting rare de novo genetic variants using parent–child trios has proven to be a powerful approach for risk gene discovery in other neurodevelopmental disorders such as autism spectrum disorder (ASD)9, developmental delay/intellectual disability10, and Tourette’s disorder11, leading to the identification of a plethora of risk genes. Since the background rate of de novo variants in the population is low, finding an elevated rate of damaging de novo variants suggests that we can leverage these variants to identify risk genes and underlying biological pathways. Studies examining de novo copy number variants (CNVs) indicate a greater rate of these variants in ADHD cases compared to controls12,13, but given the large number of genes disrupted by CNVs, it is challenging to identify specific risk genes from these variants. Whole-exome DNA sequencing studies enable the identification of de novo sequence variants affecting single genes. A few small DNA sequencing studies of parent–child trios with ADHD14,15,16,17 have identified rare de novo sequence variants, supporting the discovery potential of applying this approach in larger ADHD cohorts.

Here, we conduct whole-exome DNA sequencing in 152 parent–child trios (456 individuals in total), comprising a child with ADHD and both biological parents, and demonstrate that rare and ultra-rare de novo protein-truncating and damaging missense variants are enriched in ADHD cases compared to unaffected controls. Combining our results with a large independent case–control DNA sequencing study (3206 ADHD cases and 5002 typically developing controls)8, we identify lysine demethylase 5B (KDM5B) as a high-confidence risk gene for ADHD and identify three other potential risk genes. Finally, we show overlap among genes harboring de novo damaging variants in ADHD with previously reported risk genes for other psychiatric conditions, and we conduct exploratory analyses to identify biological pathway enrichment. These findings provide a critical step forward toward improving our etiologic understanding of ADHD, which may, in the future, inform the treatment of this common and impairing condition.

Results

Rare and ultra-rare de novo damaging variants are enriched in ADHD probands

We performed whole-exome DNA sequencing in 152 parent–child trios with ADHD collected from four sites (Supplementary Data 1). We pooled this sequencing data and performed joint variant calling with whole-exome sequencing from 788 parent–child trios without ADHD, already sequenced as part of the Simons Simplex Collection. After applying our quality control methods, we compared rates of de novo variants in 147 ADHD parent–child trios and 780 control parent–child trios. Based on studies of other childhood-onset neuropsychiatric conditions11,18,19,20, we expected to find a greater rate of rare de novo damaging variants in ADHD probands versus controls. Damaging variants include protein-truncating variants (PTVs, including premature stop codons, frameshift, and splice site variants) and missense variants predicted to be damaging (Mis-D) by a “missense badness, PolyPhen-2, constraint” (MPC) score > 221.

Results from this burden analysis demonstrate a greater rate of both rare and ultra-rare de novo damaging variants (PTVs + Mis-D) in ADHD cases versus unaffected controls (Fig. 1, Supplementary Table 1, Supplementary Data 2). For rare de novo damaging variants (non-neuro gnomAD allele frequency < 0.001), the rate ratio was 1.67 (95% CI 1.08–2.53, p = 0.03). We found a greater difference between cases and controls when narrowing our analysis to ultra-rare de novo damaging variants (non-neuro gnomAD allele frequency < 0.00005), with a rate ratio of 1.93 (95% CI 1.24–2.97, p = 0.007) (Fig. 1, Supplementary Table 1). Within the subset of ultra-rare de novo damaging variants, we found a greater rate of PTVs (rate ratio 1.94, 95% CI 1.13–3.24, p = 0.02) and a trend towards an increased rate of Mis-D variants in cases versus controls (rate ratio 1.91, 95% CI 0.78–4.36, p = 0.12). As anticipated, we did not find differences in rates of de novo variants between cases and controls when including all (damaging and non-damaging) rare or all ultra-rare variants (Supplementary Table 1).

Fig. 1: De novo mutation rates in ADHD versus control subjects.
figure 1

Rates of a rare and b ultra-rare de novo damaging mutations are enriched in ADHD probands (n = 147) compared to controls (n = 780). Rare variants have an allele frequency <0.001 (0.1%) in the non-neuro subset of the Genome Aggregation Database (gnomAD), and ultra-rare de novo variants have an allele frequency of <0.00005 (0.005%) in the non-neuro subset of gnomAD. For each variant class, the mutation rate was calculated by dividing the total number of mutations in that class by the “callable” loci across all families meeting sequencing depth and quality score thresholds. Mutation rates were calculated separately for ADHD cases (red) and controls (blue). Error bars show 95% confidence intervals for mutation rates calculated based on a Poisson distribution. Mutation rates between cases and controls were compared with a one-tailed rate ratio test with a significance threshold of p < 0.05. Bold p-values are significant. Values are provided in Supplementary Table 1 and Supplementary Data 2. PTVs, protein-truncating variants, including frameshift, splice site, and stop-gain variants. Mis-D, missense variants predicted to be damaging with “missense badness, PolypPhen-2 constraint” (MPC) score > 2. Source data are provided as a Source Data file.

Full size image

Recurrent ultra-rare damaging variants identify ADHD risk genes

Among 147 ADHD parent–child trios passing quality control, we identified 24 ultra-rare de novo damaging variants in 23 individuals (Table 1, Supplementary Data 2). Among 780 control parent–child trios passing quality control, we identified 51 ultra-rare de novo damaging variants in 50 individuals (Supplementary Data 2). The list of genes harboring rare or ultra-rare de novo damaging variants in ADHD cases did not overlap with genes harboring rare or ultra-rare de novo damaging variants in control parent–child trios passing quality control (Supplementary Data 2). One gene, KDM5B, had two de novo PTVs in unrelated individuals in our ADHD trio cohort. To identify ADHD risk genes (genes harboring damaging variants more often than expected by chance), we combined our de novo parent–child trio findings with counts of ultra-rare PTVs and Mis-D (MPC > 2) variants identified in a large independent ADHD case–control dataset (3206 ADHD cases and 5002 typically developing controls)8. Using this combined dataset, we applied the Transmission And De novo Association test (extTADA)22. We identified KDM5B as a high-confidence risk gene (FDR = 0.04) and three potential risk genes for ADHD: YLPM1 (FDR = 0.20), CTNND2 (FDR = 0.26), and GNB2L1 (FDR = 0.30) (Fig. 2, Supplementary Data 3). This extTADA analysis estimates that 1057 genes (95% CI 219–2791) contribute to ADHD risk.

Table 1 Ultra-rare de novo damaging variants identified in ADHD probands
Full size table
Fig. 2: Gene-based test results for ADHD, combining ultra-rare de novo damaging variants and independent case–control data.
figure 2

Results from the extension of the Transmission And De novo Association test (extTADA) examining ultra-rare de novo protein-truncating variants and missense variants predicted to be damaging (MPC score > 2) from 147 ADHD parent–child trios and an independent group of 3206 ADHD cases and 5002 typically developing controls. Genes are organized by chromosome with the green dots indicating genes on odd-numbered chromosomes and the blue dots indicating genes on the even-numbered chromosomes. The top 5 gene symbols with the lowest q values are listed. Only one gene, KDM5B, is classified as a high-confidence risk gene (FDR, false discovery rate < 0.1). Three genes, YLPM1 (FDR = 0.20), CTNND2 (FDR = 0.26), and GNB2L1 (FDR = 0.30) are classified as potential risk genes (FDR < 0.3). Values are provided in Supplementary Data. Source data are provided as a Source Data file.

Full size image

Genes harboring de novo damaging variants in ADHD overlap with risk genes for other psychiatric conditions

Using the list of 23 genes with ultra-rare de novo damaging variants (PTV and Mis-D) in 147 ADHD probands (Table 1, Supplementary Data 2), we identified overlap with risk genes for other conditions. Six of these 23 genes were recently reported as likely risk genes (FDR < 0.05) for neurodevelopmental disorders (NDD) in the largest meta-analysis of ASD and developmental delay9, including FBXO11 (FDR = 0), KDM5B (FDR = 0), STAG1 (FDR = 1.98 × 10−7), CTNNA2 (FDR = 9.49 × 10−5), EML6 (FDR = 0.002), and PAK1 (FDR = 0.006) (Table 1, Supplementary Data 4). Using the Gene4Denovo database23, KDM5B is also a risk gene for ASD (FDR = 0), undiagnosed developmental disorders (FDR = 0), congenital heart disease (FDR = 0.005), and across all disorders in the Gene4Denovo database (FDR = 0). FBXO11 and STAG1 were also both associated with undiagnosed developmental disorders (FDR = 0 and 0.00002, respectively) and across all disorders (FDR = 0 for both), and PAK1 was a risk gene across all disorders (FDR = 0.009) (Supplementary Data 4). Additionally, we identified an overlap between our list of 23 genes with ultra-rare de novo damaging variants in ADHD probands and gene-mapped loci from common variant GWAS studies in neuropsychiatric disorders in the GWAS Catalog (Supplementary Data 5).

Exploratory gene ontology and pathway enrichment

Using this same list of 23 genes harboring ultra-rare de novo damaging variants in ADHD trios, we also conducted exploratory analyses to identify enriched gene ontology and biological pathways. Several gene ontology and pathway-based sets were enriched for these 23 genes identified in ADHD (Supplementary Data 6). The top pathway-based sets were CXCR4-mediated signaling events (q = 0.004), Sema3A PAK-dependent Axon repulsion (q = 0.004), and ectoderm differentiation (q = 0.009).

Discussion

In this largest parent–child trio whole-exome DNA sequencing study of ADHD to date, we found a significantly greater rate of rare and ultra-rare de novo damaging variants in children with ADHD compared to unaffected controls (Fig. 1). Combining our trio sequencing data with results from a large independent case–control DNA sequencing dataset, we identified KDM5B as a high-confidence risk gene for ADHD and three other potential risk genes, YLPM1, CTNND2, and GNB2L1 (Fig. 2).

Our sequencing data identified a 1.67-fold enrichment of rare de novo damaging variants in ADHD cases compared to unaffected controls, and this enrichment was greater (1.93-fold) when narrowing to ultra-rare de novo damaging variants (Fig. 1, Supplementary Table 1). It is important to note that these estimated enrichments have wide confidence intervals, so caution is warranted in interpreting these results, and replication in larger ADHD parent–child trio cohorts is needed. Nevertheless, our observed mutation rates and our enrichment of rare de novo PTV and Mis-D variants are of a similar magnitude to those reported in other parent-offspring trio studies examining de novo variation in other neurodevelopmental disorders, including ASD and Tourette’s disorder11,20. Our enrichment of rare and ultra-rare de novo damaging variants in ADHD cases compared to controls is also consistent with findings from the largest case–control DNA sequencing study examining rare variations in ADHD. This study also reported enrichment of ultra-rare PTVs in constrained genes in ADHD cases, and the rate of these variants in ADHD was comparable to ASD cases8. Similar to ASD20, we found that PTVs comprise a greater proportion of rare de novo variants than transmitted variants in ADHD (Supplementary Data 2). Our finding of enriched rare de novo damaging variants in ADHD adds information about the genomic architecture of the disorder and supports the value of DNA sequencing studies in larger ADHD parent–child trio cohorts to identify risk genes in a manner that has led to the identification of over 100 high-confidence risk genes in ASD. Similarly, our study suggests at least hundreds of genes contributing to ADHD risk, highlighting an efficient path toward systematic risk gene discovery in ADHD.

Our study identified ultra-rare de novo PTV variants in KDM5B in two unrelated individuals with ADHD (Table 1, Supplementary Data 2). KDM5B is a histone-modifying enzyme that specifically removes methyl groups from lysine 4 on histone 3 (H3K4 demethylase), leading to epigenetic regulation of gene expression. The gene is often studied in association with cancer, but rare damaging variants in KDM5B have been more recently associated with various other conditions and functions, including congenital heart disease, embryonic development, muscle strength, DNA repair, primary complex motor stereotypies in children, cognitive functioning in adults, ASD, and developmental disorders more broadly10,20,24,25,26,27,28,29,30,31. In individuals with an intellectual disability or developmental delay, KDM5B mutations often follow a recessive inheritance pattern with homozygous or compound heterozygous mutations32,33. Heterozygous damaging mutations have been reported in the Deciphering Developmental Disorders Study34 and in ASD probands20,30. However, individuals with ADHD in our study harboring ultra-rare de novo PTV variants in KDM5B did not have diagnoses of ASD or intellectual disability. Consistent with our findings, evidence suggests considerable pleiotropy and gene dosage effects associated with this gene, in contrast to most neurodevelopmental disorder risk genes28,29. Our findings add to this evidence and suggest that ADHD is included in the spectrum of phenotypic changes that may occur in the context of rare damaging variants in KDM5B.

In addition, we identified individuals with ADHD with de novo damaging variants in the genes FBXO11, STAG1, and CTNNA2. These genes have high constraint (pLI) scores and have been previously identified as high-confidence risk genes for neurodevelopmental disorders9 (Table 1, Supplementary Data 2). FBXO11 encodes an F-box protein, and de novo variants have been associated with syndromic intellectual disability and behavioral difficulties, including ADHD35,36. STAG1 encodes a component of cohesion involved with the separation of sister chromatids and has been associated with syndromic intellectual disability37. CTNNA2 encodes a brain-expressed alpha-catenin protein involved with neuronal migration and synaptic plasticity38, and SNPs within this gene and its regulatory region have been associated with impulsivity39, excitement seeking40, and perseverative negative thinking41. Our study identified ultra-rare damaging de novo variants in these genes in children with ADHD who did not have intellectual disability or other known genetic syndromes. We did not see damaging variants in FBXO11 or STAG1 in any controls (Table 1, Supplementary Data 2, Supplementary Data 3), while one out of 5002 control subjects from the large ADHD case–control dataset8 was found to have a PTV in CTNNA2 (Supplementary Data 3). This highlights the potential range of clinical manifestations that may occur due to de novo damaging variants in these genes and suggests potential clinical implications for identifying de novo damaging variants. Several additional genes with rare de novo damaging variants in ADHD probands are discussed in the Supplementary Discussion.

Genes harboring rare de novo gene-damaging variants in the ADHD cases not only overlapped with high-confidence risk genes identified in previous DNA sequencing studies of other neuropsychiatric conditions (Supplementary Data 4) but also overlapped with genes mapped from genome-wide significant common variants identified in previous GWA studies (Supplementary Data 5). Although there was no overlap with the 76 prioritized risk genes by positional and functional annotation or the 45 exome-wide significant genes identified in the recent large ADHD GWAS4, there was overlap between genes mapped from externalizing-related disorders more broadly42. These findings add to the growing evidence supporting the convergence of common and rare variants in ADHD4 and psychiatric disorders in general20,43.

Finally, we conducted exploratory ontology and pathway analyses of genes harboring de novo damaging variants in our ADHD cases. In interpreting these results, it is important to note that many of these genes may not be true ADHD risk genes, and replication of these exploratory findings is needed as more high-confidence risk genes are identified. Nevertheless, we observed a significant enrichment of several biological processes. Of note, one of the top pathways is ectoderm differentiation (Supplementary Data 6), suggesting early neurodevelopmental underpinnings of ADHD. In the largest recent GWAS study of ADHD, gene-linked loci were enriched for expression in early brain development4, also suggesting the possible role of early embryonic changes in the development of ADHD.

Aside from those already mentioned, this study has additional limitations that should be considered. For comparing mutation rates, the ideal controls would have been sequenced simultaneously as the cases and assessed for ADHD. This study prioritized sequencing ADHD parent–child trios and used controls previously sequenced as part of the Simons simplex collection (SSC) using similar methods and scored in the normal range of the ADHD subscale of the child behavioral checklist (CBCL) or the adult behavioral checklist (ABCL). By focusing on the intersection of the capture platforms, we tried to minimize batch effects as done in other DNA sequencing studies11,19,44. It is important to note that these SSC control siblings may have an elevated genetic variant load compared to a population cohort; mutation rates and gene enrichments reported in our ADHD cases would have to overcome this potential elevation in controls to achieve statistical significance. However, these SSC control siblings are often used in parent–child trio studies, offering an advantage for future cross-disorder comparisons. Finally, our study focused on the coding region of the genome, and it is possible that relevant rare variants also occur in the noncoding region. Currently, understanding the biological and clinical relevance of non-coding variation remains challenging, but future studies of ADHD may utilize whole-genome sequencing technologies.

Despite these limitations, our study is important because it demonstrates enrichment of rare and ultra-rare de novo damaging variants in ADHD cases compared to unaffected controls and identifies KDM5B as a high-confidence risk gene for ADHD as well as other candidate risk genes for future study. These findings reinforce the value of DNA sequencing of parent–child trios in larger cohorts to identify additional risk genes for ADHD. Identifying risk genes that can be studied in model systems may offer further insight into the underlying biology of ADHD and can potentially inform clinical care for individuals and families.

Methods

Participants

This research complies with all relevant ethical regulations. This study protocol examining de-identified genetic data was reviewed by the Yale Institutional Review Board, Human Investigation Committee, and Human Subjects Committee, and determined not to be human subjects research (IRB Protocol ID 2000023609). This protocol did not include consent or compensation. Informed consent/assent was obtained at the time that the samples were collected from the participating sites. A total of 152 parent–child trios (456 individuals in total), comprising a child meeting DSM-IV or DSM-5 criteria for the diagnosis of ADHD and both biological parents, were included in this study. Trios were identified from four sites: the University of São Paulo School of Medicine (n = 30), the Center for Addiction and Mental Health in Toronto (n = 37), Florida International University (n = 13), and the Genizon biobank from Génome Québec (n = 72). All subjects were assessed by structured clinical interviews. The average age at evaluation was 4.9 years (s.d. 0.7 years) for the University of São Paulo School of Medicine site, 8.2 years (s.d. 1.9 years) for the Florida International University site, and 8.3 years (s.d. 1.6 years) for the Genizon biobank from Génome Québec site. Age data was unavailable for samples from the Center for Addiction and Mental Health in Toronto. Exclusion criteria included a diagnosis of ASD, intellectual disability, psychosis, mood disorders (including bipolar disorder), and clinically significant medical or neurological disease. No exclusions were made based on self-reported gender or biological sex, which was verified by DNA sequencing data (individual-level data in Supplementary Data 1 and 2). We prioritized the study of simplex (no known family history of ADHD) parent–child trios to increase the likelihood of detecting deleterious de novo variants. Control subjects were 788 unaffected parent–child trios, selected from the Simons Simplex Collection from the National Institutes of Health Data Archive (https://nda.nih.gov/edit_collection.html?id=2042)45. Control subjects did not have ASD and were selected to be in the normal range for the attentional problems subscale from the CBCL or the ABCL (t score < 64.5), which predicts ADHD diagnosis46.

Whole-exome DNA sequencing

Exome capture and whole-exome DNA sequencing of DNA from 80 children with ADHD and their parents were conducted at the Yale Center for Genomic Analysis (YCGA) using the IDT xGen V1 capture and the Illumina NovaSeq6000 sequencing instrument. An additional 72 ADHD parent–child trios were sequenced by Genome Quebec using the Agilent SureSelect All Exon V7 capture and the Illumina NovaSeq6000 sequencing instrument. 788 control parent–child trios were previously sequenced as part of the Simons Simplex Collection, using the NimbleGen SeqCap EzExomeV2 capture and the Illumina HiSeq 2000 sequencing instrument. We performed joint variant calling with sequencing data from all cases and controls (940 trios, 2820 individuals in total).

Sequencing alignment and variant identification

Alignment and variant calling of the DNA sequencing read followed the Genome Analysis Toolkit (GATK) best practice guidelines (https://software.broadinstitute.org/gatk/best-practices/)47. Default parameters were used for BWA-mem (http://bio-bwa.sourceforge.net/) and Picard Tools MarkDuplicates (https://broadinstitute.github.io/picard/Variant) to align reads and to mark PCR duplicates, respectively. GATK was used to realign indels, recalibrate quality scores, and generate GVCF files for each sample using the HaplotypeCaller tool. To minimize the potential downstream effects of differential coverage between the different capture platforms, a target bed file was created using the intersection of target regions of the three capture platforms (IDT xGen V1, Agilent SureSelect All Exon V7, and SeqCap EzExome V2). Case and control samples were called jointly using GATK GenotypeGVCF tools, and variant score recalibration was applied to all called variants. Passing variants were then annotated using the RefSeq hg19 gene definitions and databases using ANNOVAR48.

Quality control of de novo variants

Parent–child trios were excluded if unexpected family relationships were identified using relatedness statistics49. Trios were also omitted if the children were observed to have an outlier number of de novo variants (>20). PLINK/SEQ istats (https://zzz.bwh.harvard.edu/plinkseq/) was used to generate quality control statistics for both cases and controls, and principal component analyses were used to remove outliers (see Supplementary Fig. 1 and Supplementary Data 1 for details). After these quality control steps, we analyzed 147 parent–child trios with ADHD and 780 unaffected parent–child trios for de novo variants. The probands with ADHD included 25 females and 122 males, and the controls included 431 females and 349 males.

We then used stringent thresholds to assess de novo variants44. Specifically, we identified de novo variants as those that were heterozygous in the child (with an alternate allele frequency between 0.3 and 0.7) and not present in both parents (with an alternate allele frequency < 0.05). For variants on the X chromosome, de novo variants in male children were absent in the mother; de novo variants in female children were absent in both parents. For all de novo variants, we also required a sequencing depth of ≥20 in all family members at the variant position, alternate allele depth ≥5, and mapping quality ≥30. Calls were limited to one variant per person per gene, retaining variants with the most severe consequence20. We filtered to include rare de novo variants with an allele frequency <0.001 (0.1%) in the “non-neuro” subset of the Genome Aggregation Database (gnomAD v2.2.1). Within this set of rare de novo variants, we defined an ultra-rare subset as having an allele frequency of <0.00005 in the non-neuro subset of gnomAD50. The gnomAD v2.2.1 non-neuro dataset contains exome sequencing data from 104,068 individuals who were not ascertained for having a neurologic or psychiatric condition in case–control studies. All rare de novo damaging variants entering into our analyses were confirmed as present in the proband and absent in the parents by visualizing aligned sequencing reads from binary alignment map (.bam) files using the Integrative Genomics Viewer (IGV, https://igv.org/)51,52 (see Supplementary Methods, Supplementary Figs. 2 and 3).

Mutation rate analysis

To minimize potential bias in variant calling that may occur between different exome capture platforms and sequencing batches, especially between cases and controls, our primary comparisons are between mutation rates (per bp) within the “callable” exome per family. To calculate the callable exome denominator for our rates, we first used the GATK DepthofCoverage tool to count the number of bp in each trio that met the following criteria: sequenced at ≥20×, base quality ≥20, and mapping quality ≥30; these thresholds are the minimum required for de novo variant calling, described above. Additionally, a “callable” bp must be located within the intersection of the capture platforms (target intervals bed files) used for whole-exome DNA sequencing of ADHD cases and controls. The number of callable bases per family is listed in Supplementary Data 1 (1.1, “countable_coverage”). For each mutation class (e.g., synonymous, missense, PTV), the number of mutations was divided by the sum of callable bp for all trios; this rate was divided by 2 to calculate the haploid mutation rate for each mutation class. This was calculated separately for cases and controls. Confidence intervals were calculated (pois.conf.int, pois.exact functions from epitools v0.5.10.1 in R), and we used a one-tailed rate ratio test to compare de novo mutation rates between cases and controls (rateratio.test v1.1 in R).

Based on studies of other childhood-onset neuropsychiatric conditions11,18,19,20, we hypothesized that mutation rates for rare and ultra-rare de novo protein-truncating variants (PTV) and damaging missense variants (Mis-D) would be greater for cases compared to controls. Mis-D variants were identified using the integrated “missense badness, PolyPhen-2, constraint” (MPC)21 score > 2 as done in other recent studies8,19,53. The combined group of de novo PTV and Mis-D variants were considered “damaging” variants.

Transmission and de novo association test analysis

We used a Bayesian extension of the original transmission and de novo association test (extTADA)22 to integrate de novo and case–control variants in a hierarchical model to increase the power of identification of risk genes for ADHD. We obtained mutation counts for PTVs and Mis-D variants (MPC > 2) from an independent case–control study including 3206 individuals with ADHD and 5002 typically developing controls who did not have diagnoses of autism or intellectual disability8. We ran extTADA following the code outlined at https://github.com/hoangtn/extTADA/blob/master/examples/extTADA_OneStep.ipynb22. ExtTADA uses a Markov chain Monte Carlo approach to calculate all parameters used as input in the traditional TADA54 through sampling from the posterior in one step with resulting credible intervals. Parameter estimation led to the following estimates of (1) proportion of risk genes (\(\pi\)) (lower-upper credible intervals): 5.50% (1.24–14.27%); (2) average relative risk (\(\gamma\)) (lower-upper credible intervals): Mis-D de novo = 20.34 (1.05–66.16), PTV de novo = 21.35 (3.08–66.56), Mis-D case–control = 1.61 (1.00–4.64), PTV case–control = 1.78 (1.08–4.91); and (3) variability in relative risk estimates per gene (\(\beta\)) (lower-upper credible intervals): Mis-D de novo = 0.83, PTV de novo = 0.82, Mis-D case–control = 6.17, PTV case–control = 3.98. These parameters were used by the extTADA function to calculate the Bayes factor and q-values (false discovery rate, FDR) for each gene (Supplementary Data 3). We applied commonly used statistical thresholds to define”potential” (FDR < 0.3) and”high confidence” (FDR < 0.1) risk genes18.

To calculate the absolute number of ADHD risk genes, we multiplied the total number of genes included in the extTADA analysis (19,560) by the proportion of risk genes estimated by the extTADA pipeline. All genes from the list generated by denovolyzeR55 except for American College of Medical Genetics genes (ACTA2, ACTC1, APC, APOB, ATP7B, BMPR1A, BRCA1, BRCA2, CACNA1S, COL3A1, DSC2, DSG2, DSP, FBN1, GLA, KCNH2, KCNQ1, LDLR, LMNA, MEN1, MLH1, MSH2, MSH6, MUTYH, MYBPC3, MYH11, MYH7, MYL2, MYL3, NF2, OTC, PCSK9, PKP2, PMS2, PRKAG2, PTEN, RB1, RET, RYR1, RYR2, SCN5A, SDHAF2, SDHB, SDHC, SDHD, SMAD3, SMAD4, STK11, TGFBR1, TGFBR2, TMEM43, TNNI3, TNNT2, TP53, TPM1, TSC1, TSC2, VHL) were included in the exTADA analysis.

Gene set overlap

We examined if our list of genes with ultra-rare de novo damaging variants (PTV or Mis-D) in the ADHD probands overlapped with genes implicated in other DNA sequencing studies and genome-wide association studies. The Gene4Denovo database23 (http://www.genemed.tech/gene4denovo/home) integrates de novo mutations from 68,404 individuals across 37 different phenotypes, including several neuropsychiatric conditions, but not including ADHD. We assessed the overlap between the Gene4Denovo gene list (release version updated 07/08/2022) and our list of genes with ultra-rare damaging de novo variants. The GWAS Catalog56,57 was used to examine if this same list of genes harboring de novo damaging variants overlapped with loci mapped to genes in previous genome-wide association studies of neuropsychiatric phenotypes. The GWAS Catalog identifies past studies through weekly PubMed searches and extracts data for SNPs with p < 1 × 10−5 in the overall (initial GWAS + replication) population. All curated trait descriptions in the GWAS Catalog are mapped to terms from the experimental factor ontology (EFO), which provides a systematic description of traits to support the annotation, analysis, and visualization of data. We limited our overlap analysis to traits in the GWAS Catalog categorized under the umbrella terms ‘nervous system disease’ or ‘psychiatric disorder’ (additional details found at https://www.ebi.ac.uk/gwas//docs).

Exploratory pathway analysis

We used ConsensusPathDB58 (http://cpdb.molgen.mpg.de/, Latest Release 35, 05.06.2021) to assess whether our list of 23 genes with ultra-rare de novo damaging variants in ADHD probands (Table 1) was over-represented in gene-ontology and biological pathway sets. This network tool integrates information from 31 public databases. The following default settings were used for the exploratory gene set over-representation analysis: pathways as defined by pathway databases; select all databases; minimum overlap with input list = 2; p-value cutoff = 0.01; Gene ontology categories levels 2–5; select all biological process, molecular function, and cellular component categories; p-value cutoff = 0.01. P-values within each database are calculated using Fisher’s exact test, corrected for multiple comparisons. In addition, ConsensusPathDB calculates q-values that are corrected for the number of tests performed across all databases. Q-values are computed as Benjamini–Hochberg (BH)-corrected values from the p-values of Fisher’s exact tests, using the formula q = (p*N)/r, where N is the total number of tests performed (i.e., number of gene-ontology categories or pathways tested across all databases), and r is the rank of the p-value in the sorted list of all p-values. Supplementary Data 6 provides p-values < 0.01 and q-values for all enriched gene-ontology and pathway sets).

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.