From All to One: Can a Large Cohort of Individuals with SCN1A-Related Epilepsy Offer Predictions for Other Individuals With SCN1A Variants?

Abstract

Genotype–Phenotype Associations in 1018 Individuals With SCN1A-Related Epilepsies

Gallagher D, Pérez-Palma E, Bruenger T, Ghanty I, Brilstra E, Ceulemans B, Chemaly N, de Lange I, Depienne C, Guerrini R, Mei D, Møller RS, Nabbout R, Regan BM, Schneider AL, Scheffer IE, Schoonjans A-S, Symonds JD, Weckhuysen S, Zuberi SM, Lal D, Brunklaus A. Epilepsia, 2024;65(4), 1046–1059. doi: 10.1111/epi.17882

Objective: SCN1A variants are associated with epilepsy syndromes ranging from mild genetic epilepsy with febrile seizures plus (GEFS+) to severe Dravet syndrome (DS). Many variants are de novo, making early phenotype prediction difficult, and genotype–phenotype associations remain poorly understood. Methods: We assessed data from a retrospective cohort of 1018 individuals with SCN1A-related epilepsies. We explored relationships between variant characteristics (position, in silico prediction scores: combined annotation dependent depletion (CADD), rare exome variant ensemble learner (REVEL), SCN1A-genetic score), seizure characteristics, and epilepsy phenotype. Results: DS had earlier seizure onset than other GEFS+ phenotypes (5.3 vs 12.0 months, p < .001). In silico variant scores were higher in DS versus GEFS+ (p < .001). Patients with missense variants in functionally important regions (conserved N-terminus, S4–S6) exhibited earlier seizure onset (6.0 vs 7.0 months, p = .003) and were more likely to have DS (280/340); those with missense variants in nonconserved regions had later onset (10.0 vs 7.0 months, p = .036) and were more likely to have GEFS+ (15/29, χ2 = 19.16, p < .001). A minority of protein-truncating variants were associated with GEFS+ (10/393) and more likely to be located in the proximal first and last exon coding regions than elsewhere in the gene (9.7% vs 1.0%, p < .001). Carriers of the same missense variant exhibited less variability in age at seizure onset compared with carriers of different missense variants for both DS (1.9 vs 2.9 months, p = .001) and GEFS+ (8.0 vs 11.0 months, p = .043). Status epilepticus as presenting seizure type is a highly specific (95.2%) but nonsensitive (32.7%) feature of DS. Significance: Understanding genotype–phenotype associations in SCN1A-related epilepsies is critical for early diagnosis and management. We demonstrate an earlier disease onset in patients with missense variants in important functional regions, the occurrence of GEFS+ truncating variants, and the value of in silico prediction scores. Status epilepticus as initial seizure type is a highly specific, but not sensitive, early feature of DS.

Commentary

Dravet syndrome is one of the best-studied and most prevalent of the rare genetic epilepsies. The majority of Dravet cases are attributed to loss of function variants in SCN1A. However, disease-causing variants in SCN1A can be associated with a range of different phenotypes. Loss of function variants, including truncating variants and deletions, but more commonly missense variants, are associated with Dravet as well as milder phenotypes including familial febrile seizures and genetic epilepsy with febrile seizures plus (GEFS+). In practice, we also encounter many patients with SCN1A-related epilepsy that lie somewhere between the well-defined extremes of the spectrum. This phenotypic heterogeneity can make prognostication a challenge, particularly as genetic diagnosis occurs earlier and earlier in a patient's course. The Dravet community is also on the cutting edge of precision medicine with two gene therapies currently in clinical trials and more in development. These opportunities make early identification of Dravet and other severe SCN1A-related epilepsies imperative. ¹

The manuscript by Gallagher et al ² presented here uses an international cohort of 1018 individuals with validated SCN1A pathogenic variants to improve our understanding of genotype–phenotype correlations with an eye toward phenotype prediction. The cohort included 823 individuals classified as Dravet and 195 classified as GEFS+. This is the authors’ second publication from this cohort. In their first manuscript, Brunklaus et al ³ , they framed their research from the perspective of a family whose child has had one or more febrile seizures and has been diagnosed with a pathogenic SCN1A variant, asking: Can we predict this child's syndrome based on genotype and clinical variables? They developed a quantitative model for the prediction of Dravet syndrome (vs GEFS+) in this setting. The model included two key elements: (1) the age of first seizures, and (2) a “SCN1A-genetic score.” The latter was developed by the authors and is intended to predict the extent to which a given SCN1A variant disrupts protein function. It integrates the biochemical properties of the variant (Grantham score) and the degree of evolutionary conservation of the amino acid position across 10 sodium channel genes (paralog score). Using a supervised machine-learning process, the authors trained and tested a phenotype prediction model on subcohorts of the population. Their final model, which combined the “SCN1A-genetic score” with the age of seizure onset, predicted ultimate phenotype better than the age of onset alone and better than the age of onset combined with other in silico predictors. However, in the final model, the age of onset carried significantly greater weight (2-fold) than the genetic score in the model's performance.

This model created by Brunklaus et al ³ is publicly available and valuable in the individual who has already started to have seizures, but the model would have less predictive value if the age of first seizure was not available. One can imagine a situation in the near future where neonatal screening for SCN1A variants might be considered before the onset of seizures, particularly if there is a disease-modifying treatment available. Early identification of SCN1A may also occur due to increased utilization of genomic sequencing in prenatal diagnosis. In fact, there has already been one case of in-utero incidental diagnosis of a de novo SCN1A variant that contributed to reproductive decision-making. ⁴ This begs the question: Do we have enough information to make phenotype predictions on genotype alone?

In their recent paper presented here, Gallagher et al ² used the same large SCN1A cohort to ask if the features of a genetic variant alone can have important prognostic utility. They first demonstrated that in silico prediction models correlate with phenotypic severity within the group of known pathogenic SCN1A variants. The authors’ own SCN1A-genetic score, as well as three other scores (combined annotation dependent depletion [CADD], rare exome variant ensemble learner [REVEL], and Grantham), were all significantly higher among individuals with SCN1A variants and Dravet, compared with those with GEFS+. This suggests that there are variant characteristics that are strongly associated with disease severity but stop short of providing a cut-off above which a severe phenotype is certain.

Most of the variants in the dataset were missense variants. The authors used the positions of missense variants in the cohort to define regions of the SCN1A protein as “dense” and “sparse” with respect to the prevalence of pathogenic variants. This strategy was used to identify domains of the SCN1A gene that are more strongly associated with disease. These “dense” regions included the highly conserved domains of the N-terminus and voltage sensor as well as pore-lining domains. There was a significant association between the location of a given SCN1A missense variant and phenotype: 82.3% of those associated with Dravet came from the “dense” regions, whereas 51.7% of those associated with GEFS+ came from the “sparse” regions. This study also included 393 protein-truncating variants (PTVs). Typically, these variants, which can result in loss of protein function through nonsense-mediated decay (NMD), are associated with more severe phenotypes. In this cohort, the majority of the PTVs were associated with Dravet, but there were 10 patients with PTVs and GEFS+, accounting for 5% of the GEFS+ cases. The authors explain that 6 of these 10 truncating variants were located in the first 100 nucleotides or in the last exon of the gene. They suggest that in the PTV GEFS+ cases, the milder phenotype is due to the reduced efficiency of NMD at these locations. Thus, the authors again demonstrate that a detailed understanding of the type and location of a variant may be critical for phenotype prediction.

In my practice which focuses on adult epilepsy, one of the most important aspects of genotype–phenotype predictions is the likelihood that the phenotype will be consistent across individuals. Imagine counseling a prospective parent with GEFS+ and a pathogenic SCN1A variant. This patient has a 50% chance of passing on the variant to each child, but is a child who inherits the variant likely to have GEFS+ or Dravet? Variable penetrance and expressivity across families with SCN1A variants have been reported.^5–7 In the current cohort there were 14 variants that were identified in more than one individual. Of these, 6 variants, all missense, had to be excluded from several analyses because of discrepant phenotypes. Four variants were seen in “mixed” families where they were associated with both Dravet and GEFS+ in different individuals.

Overall, this study by Gallagher et al ² demonstrates that there is clear value in decoding the characteristics that may contribute to the phenotypic expression of SCN1A variants. However, there is still much to consider before putting this information to use as the sole arbiter of an individual's prognosis. There is mounting evidence that SCN1A-related epilepsy is not a purely monogenic condition. Genomic background, as measured by polygenic risk scores of common variants, may play a role in the Dravet phenotype ⁸ and in determining disease severity in GEFS+ families with or without SCN1A pathogenic variants. ⁹ Additionally, some of the variable penetrance and expressivity in individuals with pathogenic SCN1A variants has been shown to be due to the cooccurrence of additional rare variants in SCN1A or other genes that can modify the phenotype.^7,8,10 Thus, while the type and location of a single SCN1A variant may carry a lot of prognostic weight, there are other important variables to explore. This likely explains why, in the original model, a clinical variable (age of onset) outperforms the genotype alone in phenotype prediction. Future research on this and other large datasets will benefit from further study of individuals with discordant genotype–phenotype correlations. The genomic differences between these individuals may serve to fine-tune multimodal predictive algorithms.