The First Step Towards Personalized Risk Prediction for Common Epilepsies

Thomas F. Hansen; Rikke S. Møller


Brain. 2019;142(11):3316-3318. 

Epilepsy is one of the most common neurological disorders, with a lifetime prevalence of ~3% (Hesdorffer et al., 2011). Epilepsy is defined by recurrent, unprovoked seizures resulting from abnormal, synchronized neuronal firing in the brain, and encompasses a group of disorders with diverse aetiologies and outcomes. Genetic factors are known to play a role in up to 70–80% of cases, either as single-gene mutations in rare monogenic epilepsies, or as multiple genetic risk factors in common epilepsies, including genetic generalized and focal epilepsies (Thomas and Berkovic, 2014). The latter two groups account for >90% of all cases. While early-onset epilepsies, including developmental and epileptic encephalopathies, comprise a genetically heterogeneous group of rare monogenic disorders mainly caused by de novo variants (Moller et al., 2016; Heyne et al., 2018), this is not the case for common epilepsies. These are thought to be complex traits with a polygenic origin, meaning that the genetic component is likely to arise from the combined effect of many common risk variants, each with small effect size. More than 100 genes have been associated with monogenic epilepsies (McTague et al., 2016), and improved understanding of the pathophysiology of these disorders has led to an increasing awareness of the importance of personalized medicine. Indeed, targeted treatment approaches have already been developed for a few of these rare conditions (Wolff et al., 2017; Moller et al., 2019). Although personalized medicine for common epilepsies is still lacking, genetic studies have also led to a better understanding of the pathophysiology of these disorders. A recent meta-analysis of genome-wide association studies (GWAS) reported 16 genome-wide significant loci (International League Against Epilepsy Consortium on Complex, 2018). Beyond these 16 loci, estimates of heritability that can be attributed to single nucleotide polymorphisms (SNPs) showed that, although each variant only explained a small proportion of the genetic variance, collectively all variants could explain 32% and 9% of the liability for generalized and focal epilepsy phenotypes, respectively (International League Against Epilepsy Consortium on Complex, 2018). This is often referred to as the SNP or chip heritability. This finding prompted the use of a polygenic risk score, which can leverage the combined effect of thousands of variants into a single instrument used to predict individual genetic disease risk. Polygenic risk scores can also be used to test the genetic overlap among traits, and are therefore useful in the investigation of shared genetic risk with neurological and psychiatric co-morbidities, and in studying the relationship between epilepsy subtypes. Polygenic risk scores may also pave the way for more personalized treatment approaches.

In this issue of Brain, Leu and co-workers report the first study on genetic burden within and among common epilepsies (Leu et al., 2019). The authors estimated the genetic risk for generalized and focal epilepsy in three independent cohorts, totalling 6774 patients (Leu et al., 2019). The study was conducted by calculating polygenic risk scores using effect measures of common variants derived from the largest GWAS in epilepsy to date (International League Against Epilepsy Consortium on Complex, 2018). The authors identified a significantly higher genetic burden for epilepsy compared to population controls and assessed how much phenotypic variance of the common epilepsies could be explained by the polygenic risk score. By comparing logistic regression models with and without polygenic risk scores, the authors estimated the Nagelkerke's pseudo R2-values—an analogue to the R2 variance used for continuous measures—for the polygenic risk score, after correcting for gender and population structure. The polygenic risk scores for generalized epilepsy (GE-PRS) were able to explain a higher proportion of phenotypic variation in generalized compared to focal epilepsy (~2.8% versus 0.5%) (Figure 1A), showing that patients with generalized epilepsy have a significantly higher burden of common risk variants associated with generalized epilepsy than patients with focal epilepsy (Leu et al., 2019). The authors also showed that polygenic risk scores for focal epilepsy (PRS-FE) were significantly higher in patients with focal epilepsy compared to controls. Interestingly, the GE-PRS also explained more phenotypic variation in focal epilepsy (>1.7%) than vice versa (<0.2%) (Leu et al., 2019). This could reflect the fact that the original GWAS for focal epilepsies had less power than that for generalized epilepsies to identify common risk variants. Alternatively, given the low SNP heritability for focal epilepsy and its reduced ability to predict generalized epilepsy, it could also indicate that focal epilepsies have a lower genetic component than generalized epilepsies. Furthermore, Leu et al. assessed three biobanks, in which ICD-10 G40.3 codes were used to identify individuals with generalized epilepsy, and G40.0 to G40.2 codes to identify individuals with focal epilepsy. Although they found a significant effect in biobanks of European ancestry, the phenotypic variance explained was much reduced (<0.2%), indicating that epilepsy polygenic risk scores have limited value in biobank-derived cohorts in comparison with clinically assessed cohorts.

Figure 1.

Use of polygenetic risk scores for predicting generalized and focal epilepsy. (A) Nagelkerke's R2 values were estimated by comparing regression models with and without GE/FE-PRS against controls, and by comparing generalized and focal epilepsies. The values shown are derived from the results of the predictions from the different target cohorts. Values were corrected for gender and age. (B) Odds ratios derived from comparison of the top 5th percentile of the GE/FE-PRS versus the remaining 95%, from the different target cohorts. (C) Receiver operating characteristic (ROC) curve, illustrating the clinical value of polygenic risk scores (PRS), clinical assessment and the putative impact of combining polygenic risk scores with clinical observations. Numbers shown are based on the results from Leu et al. (2019).

Lastly, in an effort to guide future studies, the authors estimated the sensitivity and specificity of the polygenic risk score as a predictive test, given different thresholds for GE/FE-PRS, to identify individuals at high risk (Figure 1B). While the specificity was relatively high at all thresholds, the sensitivity (true positive) rate, was low. For example, 13.5% of all cases would be correctly identified as individuals with epilepsy, when setting the 95th percentile as the threshold to identify significant generalized epilepsy polygenic risk score burden. However, there was still a large portion of patients with lower polygenic risk scores, and 41% of patients had scores outside the top 20% of polygenic risk score values.

Overall, Leu et al. have shown that it is possible to measure a common polygenic variant burden for epilepsy, and that the burden is differentially distributed among patients with epilepsy and controls, as well as between epilepsy subtypes. The study provides the first step towards a possible clinical use of polygenic risk scores in estimating an individual's overall genetic risk for epilepsy, which may in the future aid in early diagnosis, differential diagnosis and personalized treatment approaches. However, it is important to stress that this genetic predictor should be combined with clinical observations, as it seems unlikely that the polygenic risk score alone can provide enough sensitivity to predict a 'true' epilepsy case. There is a possibility that inclusion of rare variants may improve predictability; however, so far the evidence is sparse. Another future use of polygenic risk scores in epilepsy diagnostics could be in the 'first seizure clinic', where a genetic predictor could be useful in identifying individuals with epilepsy versus those that have had a single non-recurrent seizure (Figure 1C). However, before polygenic risk scores can be used in a diagnostic setting there is a need for larger discovery cohorts, which will be able to improve the SNP-effect estimates and improve the sensitivity of the score. Furthermore, it will also be important to produce genome-wide association studies on epilepsy subtypes, endophenotypes, and treatment outcomes, and to extend the exploration of the polygenic risk score in non-European populations, in order to increase the specificity of SNP-effect estimates used for the polygenic risk score. For common complex traits, substitutes for precise clinical classifications by experts are often used to allow implementation of large cohorts e.g. biobanks. This strategy usually results in a large gain of statistical power, despite a potentially higher clinical heterogeneity of the analysed cohort. The current results suggest that this strategy fails to increase the power of analyses in epilepsy subtypes, and use of biobank data to distinguish epilepsy subtypes should therefore be considered carefully until improved strategies to curate biobank phenotypes have been developed. Further research is clearly still warranted. However, development of models that can combine common variants, i.e. polygenic risk scores with clinical/diagnostic observations, may in the future enable personalized risk prediction and precision medicine for common epilepsies.