Gene burden meta‐analysis of 748 879 individuals identifies LGI1‐ADAM23 protein complex association with epilepsy

Abstract

Epilepsy affects more than 50 million individuals globally and has a substantial genetic component that remains to be completely understood. Traditional studies have focused on severe, early onset cases enrolled through clinical or research settings. Recent biobank‐based approaches, leveraging large‐scale population datasets, offer opportunities to explore genetic associations in broader epilepsy phenotypes, including milder, later onset forms. We analyzed data from more than 750 000 individuals across the UK Biobank, All of Us, and Massachusetts General Brigham Biobank, including 20 026 individuals with epilepsy. Rare coding variant burden testing revealed a significant association with LGI1, a known epilepsy gene. Among the other top 10 associated genes, seven had prior evidence linking them to epilepsy (GABRG2, ATP1A3), neurological disorders with comorbid seizures (HTRA2, KRIT1, STAG1), possible involvement in seizure phenotypes (ADAM23), or roles in neuronal function (PDCD4). Thus, we provide the first statistical evidence for ADAM23 as a candidate gene for epilepsy, based on the suggestive association signal combined with prior biological evidence from both animal (canine and murine) and one recent human epilepsy case study, potentially contributing to human epilepsy through its direct interaction with LGI1. Phenome‐wide analyses highlighted the pleiotropic effects of epilepsy genes, with LGI1 and ADAM23 predominantly associated with epilepsy, whereas other genes such as KRIT1, TSC1, and TSC2 exhibited broader systemic involvement. Our study shows the potential of population‐scale genomic data and suggests that integrating these datasets with deep phenotyping will uncover more novel insights into epilepsy genetics in the future.

1. INTRODUCTION

Epilepsy is a common neurological disorder that affects more than 50 million individuals worldwide and is characterized by a significant but incompletely understood genetic component. ¹ , ² , ³ Conventional clinical research strategies, for example, gene burden association studies on clinically ascertained cohorts, have predominantly focused on severe, early onset forms, for example, developmental and epileptic encephalopathies (DEEs), elucidating key pathways such as ion channel dysfunction, synaptic communication, and mTOR signaling. ² , ⁴ , ⁵ In parallel, large‐scale research efforts, such as the Epi25 Collaborative, have demonstrated the value of deeply phenotyped datasets in uncovering subtype‐specific genetic associations, including ultrarare variant contributions across DEEs, genetic generalized epilepsies, and nonacquired focal epilepsies. ² These recent large‐scale research cohort studies underscore the critical role of integrating genetic data with detailed phenotypic characterization to reveal both shared and distinct molecular mechanisms across epilepsy subtypes. ² A potential major barrier to discovering new genes in epilepsy is that many potential genetic patients are excluded because they do not meet the commonly used clinical criteria, such as those defined by the International League Against Epilepsy (ILAE) syndrome definitions (https://epi‐25.org/epi25‐eligibility‐criteria). ⁶ , ⁷

Biobank‐based approaches, which draw on population‐scale datasets, offer new opportunities to identify genes linked to more broadly defined epilepsies or seizure‐associated disorders, as most phenotype classifications are deliberately broadly based on billing codes, which may include cases of provoked or isolated seizures that do not meet strict ILAE criteria for epilepsy. The large sample size increases statistical power to detect associations with mild or low‐penetrance genetic variants that might not emerge in smaller studies, and are considered to have less of a Mendelian genetic contribution. ⁸ , ⁹ , ¹⁰ Meanwhile, population‐based studies, like those leveraging the UK Biobank (UKB) and other global resources, have revealed the power of panancestry analyses to uncover rare genetic associations across diverse populations. ⁹ The massive cohort sizes of recent cross‐biobank meta‐analyses have shown great success in identifying novel disease‐associated genes and may also enable the identification of novel genes associated with epilepsy. ¹¹

Here, we explore publicly available genomic resources for exome‐wide gene–epilepsy associations in a cohort of more than 750 000 individuals, including 20 026 people with epilepsy, drawn from three biobank resources: All of Us (AoU), the UKB, and the Massachusetts General Brigham (MGB) Biobank. By leveraging the scale and diversity of these datasets, we can identify both well‐established and previously unknown epilepsy‐associated genes.

2. MATERIALS AND METHODS

2.1. Dataset, origin, and composition

The dataset utilized in this analysis originates from a comprehensive panancestry rare coding variant study across three major biobanks—the UKB, the AoU Research Program, and the MGB Biobank ¹¹ —that utilized either exome or genome sequencing to identify rare variants. We reanalyzed publicly available gene‐level association results through a website provided by the authors in December 2024 (https://hugeampkpn.org/research.html?pageid=600_diseases_home). A detailed description of the cohort and methods can be found in the referenced study. ⁹ Our work focuses on reanalyzing and interpreting these results in the context of epilepsy in combination with previous biological evidence and additional analyses, leveraging our domain‐specific expertise. In brief, the combined dataset includes sequencing data for 748 879 individuals: 454162 from UKB (whole‐exome sequencing), 242 902 from AoU (whole‐genome sequencing), and 51 815 from MGB (whole‐exome sequencing). Participants were grouped into principal ancestry categories using genetic similarity to 1000 Genomes reference samples, with 20.7% of individuals classified as non‐European ancestry (notably African, admixed‐American, East Asian, and South Asian).

The cohort is expected to primarily include adults with a spectrum of seizure‐associated disorders, ranging from mild to moderate cases, with a likely underrepresentation of severe epilepsy syndromes and individuals with significant cognitive disabilities. This reflects the recruitment biases of population‐based biobanks, which suffer from a healthy volunteer bias and tend to include fewer individuals with profound neurological impairments. ¹² , ¹³ Phenotypes are classified according to billing codes, which limits the ability of epilepsy diagnosis according to clinical standards or seizure‐disorder subclassifications.

2.2. Phenotypic endpoints and association testing

Phenotypic endpoints were derived from International Classification of Disease (ICD) codes mapped into 1866 phecodes and subsequently pruned to 601 disease phenotypes using hierarchical clustering to minimize redundancy. Disease prevalence was standardized across cohorts, with adjustments for potential differences in coding systems (ICD in UKB vs. ICD‐CM in AoU and MGB). The only epilepsy‐related phecode that was mappable across all three studies was #345 "Epilepsy, recurrent seizures, convulsions," annotated for 20 026 people. The final dataset underwent gene‐based burden testing for rare variants (minor allele frequencies ranging from <.001% to <.1%) across six predefined masks (loss‐of‐function [LOF] variants and missense variants with functional scores) as previously described. ¹¹ For a given phenotype, rare variants in each mask were analyzed with a two‐sided logistic mixed‐effect score model. Briefly, the following rare variant masks were applied in the discovery cohort:

1. "Rare LOF" mask restricting to LOF variants with mean allele frequency (MAF) population maximum (popmax) < .1% (i.e., MAF < .1% in the dataset and gnomAD popmax < .1%).

2. "Rare LOF + missense0.8" mask including both LOF variants and predicted‐deleterious missense variants with missense score > .8 and MAF popmax < .1%.

3. "Rare LOF + missense0.5" mask including both LOF variants and predicted‐deleterious missense variants with missense score > .5 and MAF popmax < .1%.

4. "Ultrarare LOF + missense0.5" mask including both LOF variants and predicted‐deleterious missense variants with missense score > .5 and MAF popmax < .001% (for within‐dataset filtering, we used minor allele count (MAC) < 5 if more inclusive).

5. "Ultrarare missense0.5" mask restricting to missense variants with missense score > .5 and MAF popmax < .001% (for within‐dataset filtering, we used MAC < 5 if more inclusive).

6. "Ultrarare missense0.2" mask restricting to missense variants with missense score > .2 and MAF popmax < .001% (for within‐dataset filtering, we used MAC < 5 if more inclusive).

To assess pleiotropic effects, we utilized Phenome‐wide association study (PheWAS) results available through the online resource (see Table S2). These analyses examined phenotypic associations for selected epilepsy‐related genes (top 50) across 601 disease endpoints. All statistical analyses were conducted by the original study authors. Our contribution was limited to summarizing and interpreting the associations for epilepsy and related traits.

2.3. Evaluation of gene burden association tests

The gene burden association tests, as performed by Jurgens et al., were used in a binomial statistical analysis to assess the probability of observing established epilepsy genes within the top‐ranked genes associated with the phecode: "Epilepsy, recurrent seizures, convulsions" (Table S1). ¹¹ Of the approximately 20 000 human protein‐coding genes, 1016 are known to be associated with epilepsy to date based on the Genes4Epilepsy database (https://github.com/bahlolab/genes4epilepsy, accessed December 2024). ¹⁴ Using a binomial test, we compared the observed number of established epilepsy genes (four of the top five associations) to the expected probability under random chance, calculated as the proportion of epilepsy‐associated genes (1017/20 000). A one‐sided t‐test was performed to determine whether the observed enrichment was statistically significant.

2.4. Gene expression analysis

Gene expression data for ADAM23 was analyzed to examine tissue‐specific and developmental expression patterns across humans and other species. In humans, tissue‐specific expression was quantified using transcripts per million values obtained from the GTEx database (https://www.gtexportal.org/home/gene/ADAM23). Developmental expression profiles in humans (ENSG00000114948), mice (ENSMUSG00000025964), rats (ENSRNOG00000012424), and chickens (ENSGALG00000008582) were retrieved from the Evo‐Devo App (https://apps.kaessmannlab.org/evodevoapp/). Reads per kilobase of transcript per million mapped reads values were used to measure expression levels across pre‐ and postnatal developmental stages.

2.5. Other bioinformatic resources utilized

Rare variant constraint information for genes, including ADAM23, was retrieved from the gnomAD V4 database (https://gnomad.broadinstitute.org/gene/ENSG00000114948?dataset=gnomad_r4). Enrichment analyses for pathway overrepresentation were conducted using the KEGG pathway annotation via the Enrichr platform (https://maayanlab.cloud/Enrichr/).

3. RESULTS

In this meta‐analysis of 748 879 individuals, the phecode #345 “Epilepsy, recurrent seizures, convulsions” was annotated for 20 026 individuals (Figure 1). The strongest rare variant burden association, and the only one surpassing the multiple testing correction threshold, was observed for LGI1 (p = 1.85 × 10⁻⁷, combined MAF [cMAF] = .0000875; see Table S1) under the LOF plus rare missense variant model (Figure 1). LGI1 is a well‐established gene associated with autosomal dominant lateral temporal lobe epilepsy. ¹⁵ , ¹⁶ , ¹⁷ , ¹⁸ Given that only 1016 of ~20 000 human genes have been associated with epilepsy to date (Genes4Epilepsy; see Materials and Methods), the probability of observing this level of enrichment by random chance was (p = 3.2 × 10⁻⁵). Although no other genes reached statistical significance after multiple testing correction, the cross‐biobank results demonstrated strong validity and appeared well calibrated, as four additional established epilepsy‐associated genes ranked among the top five associations (Figure 1). These four additional established epilepsy‐associated genes among the top five included GABRG2 (p = 3.99 × 0⁻⁶, cMAF = .000124) and ATP1A3 (p = 1.44 × 10⁻⁵, cMAF = .000316), both under the LOF plus rare missense variant model. Interestingly, the fourth strongest association was found for truncating variants in ADAM23 (p = 5.54 × 10⁻⁶, LOF model, cMAF = .0000438), a gene constrained for truncating variants in humans (gnomAD; see Materials and Methods) that has previously been associated with epilepsy in dogs and is a direct interaction partner of LGI1. ¹⁹ , ²⁰ Expression analysis further demonstrated that ADAM23 is highly enriched in the brain across multiple species, with a sharp increase in expression postnatally (Figure 2). This developmental regulation aligns with its expected role in epilepsy pathophysiology. ¹⁹ , ²¹ , ²² , ²³ Additionally, ADAM23 expression in humans was observed to be strongly brain‐specific, with negligible expression in other tissues (Figure 2A). Additional well‐established genes in the top 50 include KRIT1, TSC1, TSC2, DEPDC5, and GRIN2A (Figure 1). This group of genes indicates a strong overrepresentation of the mTOR signaling pathway (p _adjusted = 9.26 × 10⁻⁴, KEGG pathway hsa04150; see Materials and Methods).

FIGURE 1.

Manhattan plot depicting the results of the gene burden association test. The y‐axis represents the ‐log10(p‐value) of the associations, indicating the statistical significance of each gene's burden test. The x‐axis is organized by chromosomal positions, with each data point corresponding to a specific gene. Genes identified as well‐established human epilepsy genes and ranking among the top associations are labeled. Notably, ADAM23—an established canine epilepsy gene and a direct protein–protein interaction partner of LGI1, the gene with the strongest association in this study—is highlighted with an asterisk.

FIGURE 2.

ADAM23 is mainly expressed in the postnatal brain in humans. (A) ADAM23 human gene expression across organs of adult donors. (B) ADAM23 gene developmental expression in human, mouse, rat, and chicken. Data visualization includes boxplots to depict human tissue‐specific expression and line plots to illustrate developmental trends across species. RPKM, reads per kilobase of transcript per million mapped reads; TPM, transcripts per million.

To assess the specificity of the observed associations, we utilized the PheWAS results provided by Jurgens et al. ¹¹ for gene burden signals across >600 phecodes (see Table S2). LGI1, ADAM23, GABRG2, and ATP1A3 showed epilepsy/recurrent seizures as the strongest association, consistent with their established roles in epilepsy. For KRIT1, the strongest associations were observed for hemangioma and lymphangioma, surpassing epilepsy/recurrent seizures in significance. For TSC1 and TSC2, the strongest associations were observed for other and unspecified congenital anomalies, with atrial fibrillation/flutter also reaching significance for TSC2. In both cases, epilepsy/recurrent seizures remained significant but secondary. DEPDC5 showed stronger associations with nystagmus and irregular eye movements, herpes zoster, and malignant uterine neoplasms, with epilepsy/recurrent seizures ranking below these phenotypes. For GRIN2A, the strongest association was observed with developmental delays and disorders, followed by epilepsy. These findings highlight the strong specificity of epilepsy for LGI1, ADAM23, GABRG2, and ATP1A3, whereas genes such as KRIT1, TSC1, TSC2, DEPDC5, and GRIN2A exhibited additional associations reflective of their broader phenotypic impact.

4. DISCUSSION

Large population biobanks predominantly recruit middle‐aged to older, well‐educated individuals, leading to underrepresentation of individuals with developmental disabilities or severe early onset conditions. ¹³ Consequently, epilepsy cases captured in these cohorts differ markedly from those in specialized genetic studies, which typically emphasize early onset, severe, and syndromic forms of epilepsy. By reanalyzing a public dataset of rare coding variant associations in more than 750 000 individuals derived from three independent biobanks, including 20 026 with epilepsy, we identified both established and novel possible epilepsy‐associated genes in patient samples. These findings suggest that canonical epilepsy genes also contribute to susceptibility in milder or later onset forms of the disease.

Importantly, we report a novel possible association of ADAM23 with epilepsy in human patients, supported by prior biological evidence implicating ADAM23 in multiple canine breeds as a common risk gene for epilepsy with low penetrance. ¹⁹ , ²³ Furthermore, homozygous knockout of ADAM23 in mice leads to neonatal epilepsy and death by 2 weeks of age, whereas heterozygous mice exhibit a reduced seizure threshold. ²² More recently, prompted by our initial report as a preprint, a homozygous pathogenic ADAM23 frameshift variant was identified in one family, leading to a lethal neonatal onset epilepsy form. ²⁴ The ADAM23 variant followed a recessive inheritance pattern, with unaffected heterozygous carrier parents, and displayed similar clinical features, including early death and an inheritance pattern similar to that of pathogenic homozygous variants in LGI1, as reported for the first time in the same study. ADAM23 (a disintegrin and metalloproteinase 23) encodes a synaptic adhesion molecule that forms a complex with LGI1, is exclusively expressed in the central nervous system, and has been shown to influence neuronal development through an ADAM23‐dependent mechanism. ²² , ²⁵ LGI1, the gene showing the strongest association in our study, is a well‐established epilepsy gene. ²⁰ LGI1 is implicated in a broad phenotypic spectrum, ranging from autosomal dominant lateral temporal lobe epilepsy to other forms of focal epilepsies. ¹⁵ , ¹⁶ , ¹⁷ , ¹⁸ , ²⁶ , ²⁷ Its direct interaction partner, ADAM23, may similarly exhibit a wide, albeit subtler, range of epilepsy presentations. ²⁰ Although the pathogenic role of ADAM23 in human epilepsy remains to be confirmed, the statistical support provided by population‐scale data combined with emerging functional evidence from animal models and human studies raises the possibility of involvement in milder or later onset forms of epilepsy, potentially modulated by additional genetic or environmental factors.

The results from our meta‐analysis compared to the largest exome study in epilepsy to date, the Epi25 exome sequencing study, reveal both overlapping and distinct genetic associations with epilepsy, highlighting methodological, phenotypic, and cohort differences. Both studies reinforced the significance of LGI1, DEPDC5, and GABRG2. Among the top 10 association results in our study, we observed association signals for three genes known to cause neurological disorders with comorbid seizures (HTRA2, KRIT1, STAG1) that were not captured by the Epi25 study. Our findings emphasized the mTOR pathway (e.g., DEPDC5, TSC1, TSC2) and ion channels (KCNQ3 and GRIN2A), whereas the Epi25 study highlighted additional pathways, such as γ‐aminobutyric acid type A receptor complexes. ² Differences likely stem from the broader phenotype used in our study, which grouped diverse epilepsy presentations under a single phenotype code, compared to the subtype‐specific approach of the Epi25 study, which focused on clinically confirmed epilepsy forms, such as severe forms, including DEEs. Furthermore, general population‐based studies may capture neurological disorders with seizure comorbidities that are not clinically defined as epilepsy, yet may present shared genetic risk, likely explaining the lack of signal for ADAM23 in the Epi25 study (odds ratio = 1.03, p = .927). ²⁸ , ²⁹

Other genes that have been previously reported as associated with epilepsy were also represented with association signals in our study, including KCNQ3, GRIN2A, and ZNF667. In addition to our primary findings, our analysis may also help elucidate the broader landscape of epilepsy genetics. For example, EFHC1 and CLCN2 have been largely refuted, yet they are still reported as epilepsy genes. ³⁰ Our study failed to replicate these genes, further supporting the initial refutation. Future analyses could explore the stratification of the Epi25 cohort into mild versus severe phenotypes. Although our current data do not permit this subclassification, we acknowledge this as a valuable avenue for subsequent research. Furthermore, we posit that specialty studies, like Epi25, could consider secondary analyses to capture a broader spectrum of seizure disorders (instead of strictly defined epilepsies) to capture genetic risk factors across epilepsy and seizure‐related disorders. Such genetic risk factors may have higher pleiotropy and possibly be only identifiable in research designs with a broad phenotype definition. In addition to comparing biobank and specialist cohorts, future research could also focus on differentiating individuals with benign, later onset epilepsy from those with severe, treatment‐resistant forms. Genome‐wide association studies may have higher power than rare‐variant association studies, such as Epi25 and our study, to identify genetic factors contributing to mild versus severe forms of epilepsy. However, such case/case association studies have not yet been attempted, and there is no consensus on how to define epilepsy severity. Such stratification could lead to more tailored therapeutic strategies and a better understanding of the underlying genetic architecture.

Phenome‐wide association analyses revealed the pleiotropic effects of identified epilepsy genes. Whereas LGI1, ADAM23, GABRG2, and ATP1A3 were predominantly associated with epilepsy, genes such as KRIT1, TSC1, and TSC2 demonstrated multisystemic involvement. These findings are consistent with prior evidence linking KRIT1 to vascular phenotypes, including cerebral cavernous malformations, and associating TSC1 and TSC2 with congenital anomalies and arrhythmias characteristic of tuberous sclerosis complex. ³¹ , ³² Similarly, DEPDC5 was linked not only to neurological traits but also to unexpected phenotypes such as herpes zoster and uterine neoplasms, reflecting its broader biological roles. The association of GRIN2A with developmental delays further reinforces its contribution to neurodevelopmental disorders.

Our study has notable limitations. Phenotypes derived from billing codes lack the precision of specialist‐defined cohorts, potentially leading to the inclusion of misclassified individuals. The underrepresentation of individuals with developmental disabilities and refractory epilepsy likely biases findings toward milder cases. Furthermore, validating novel associations in patients such as ADAM23 will require functional studies and targeted recruitment. The focus on rare coding variants excludes noncoding and structural variants that may also influence epilepsy susceptibility. Despite these limitations, our reanalysis of population‐scale data identified ADAM23 as a candidate epilepsy gene not previously associated with the disorder in humans. Integrating genetic data from large‐scale public resources with deep phenotyping, functional validation, and longitudinal follow‐up will be essential for refining genotype–phenotype correlations and advancing personalized treatment strategies. Together, findings from population‐based and research‐ascertained cohorts are broadening our understanding of epilepsy genetics, moving beyond traditional syndromic boundaries.

AUTHOR CONTRIBUTIONS

Jessica Castrillon Lal and Dennis Lal conceived the project. Jessica Castrillon Lal, Costin Leu, and Dennis Lal performed the data analysis and interpretation. Jessica Castrillon Lal, Costin Leu, Christian M. Boßelmann, Alina Ivaniuk, Eduardo Pérez‐Palma, and Dennis Lal provided expertise and advised on data interpretation. Dennis Lal conceived the study, planned and supervised the work, performed the analysis, and wrote the manuscript collaborating with all the other authors. All authors contributed to the interpretation of the results and commented on and approved the final manuscript.

FUNDING INFORMATION

J.C.L. is a TRIUMPH Fellow in the CPRIT Training Program (RP210028). C.M.B. is supported by the MINT‐Clinician Scientist program of the Medical Faculty Tübingen, funded by the Deutsche Forschungsgemeinschaft (German Research Foundation; 493665037). E.P.‐P. is supported by the Chilean National Agency for Investigation and Development, FONDECYT grant 1221464.

CONFLICT OF INTEREST STATEMENT

None of the authors has any conflict of interest to disclose.

ETHICS STATEMENT

This study used publicly available, deidentified data. No new human participants were enrolled, and no identifiable private information was accessed; therefore, additional institutional ethics approval was not required. We confirm that we have read the Journal's position on issues involved in ethical publication and affirm that this report is consistent with those guidelines.

Supporting information

TABLE S1. Gene burden results of the top 50 genes correlated with the phecode “Epilepsy, recurrent seizures, and convulsions” and corresponding epilepsy subtype reported in the Epi25 study. For each gene, the corresponding gene ensemble ID, best performing variant mask, odds ratio, p‐value, and combined minor allele frequency threshold used are given. Corresponding effect size and p‐values from the independent Epi25 study are provided for overall epilepsy and for major epilepsy subtypes. Abbreviations: DEE, developmental and epileptic encephalopathies; GGE, genetic generalized epilepsy; MAF, mean allele frequency; NAFE, nonacquired focal epilepsies; OR, odds ratio.

TABLE S2. Complete phecode association results for the top 50 epilepsy‐associated genes.

DATA AVAILABILITY STATEMENT

The dataset utilized in this analysis originates from a comprehensive panancestry rare coding variant study across three major biobanks—the UKB, the AoU Research Program, and the MGB Biobank ¹¹ —that utilized either exome or genome sequencing to identify rare variants. We reanalyzed publicly available gene‐level association results through a website provided by the authors in December 2024 (https://hugeampkpn.org/research.html?pageid=600_diseases_home).