No major role of common SV2A variation for predisposition or levetiracetam response in epilepsy

Introduction 

Levetiracetam (LEV) is currently licensed in Europe and the US for adults and children of specified age groups as adjunctive therapy for partial epilepsies with or without secondary generalization, for myoclonic seizures in juvenile myoclonic epilepsy, and for primary generalized tonic–clonic seizures in patients with idiopathic generalized epilepsy. Licensing in Europe as first-line monotherapy in partial epilepsy occurred subsequent to the completion of this study.

The efficacy of LEV has been demonstrated in a number of regulatory clinical trials (Ben-Menachem and Falter, 2000, Cereghino et al., 2000, Shorvon et al., 2000) and it is effective in a wide range of seizure types and syndromes (Depondt et al., 2006, Kumar and Smith, 2004, Weber and Beran, 2004). Furthermore, some previously refractory patients become seizure-free in response to LEV treatment (Kinirons et al., 2006). It binds to the synaptic vesicle protein SV2A (Lynch et al., 2004), the predominant of three SV2 isoforms (SV2A, SV2B, and SV2C) in the brain. There are no accepted clinical predictors of response, and it is possible that some of the variation in response may be due to genetic differences between patients in SV2A, SV2B or SV2C.

SV2A is an essential protein as homozygous SV2A knockout mice appear normal at birth but fail to grow, experience severe seizures and die by 3 weeks (Crowder et al., 1999). Heterozygous knockout mice, although viable, are 10 times more likely to have seizures than wild-type animals (Crowder et al., 1999). These observations suggest the possibility that SV2A may influence predisposition to epilepsy (although not necessarily response to LEV) and therefore an assessment of human genetic variation in SV2A may also contribute to understanding aspects of the biological basis of epilepsy. In this report we consider and present the SV2 data for single SNP analysis only. As part of a much larger study of epilepsy predisposition (Cavalleri et al., 2007) which included additional cohorts (n=4 in total) and genes (n=279 in total), we considered the SV2A, SV2B and SV2C genes in a separate and different analysis for the possibility of enrichment for P-values in the context of a large set of genes, but not for single SNP analysis as presented here.

We have therefore related genetic variation in SV2A, SV2B and SV2C to LEV response in two separate cohorts of 247 and 290 patients with epilepsy, and to predisposition to epilepsy (including a range of epilepsy syndromes) and to mesial temporal lobe epilepsy associated with hippocampal sclerosis (mTLE+HS) in two cohorts of 803 (133 mTLE+HS) and 801 (139 mTLE+HS) patients with epilepsy.

Results 

Results are summarised in Table 7 (Supplementary Material). All P-values reported are uncorrected for multiple testing. Cell counts are available on request. Seven polymorphisms were associated with response to LEV in the UK cohort (P<0.05). None was significant after Bonferroni correction for 86 multiple comparisons (correcting within UK cohort only). One of these seven polymorphisms, rs2937720, an intronic polymorphism in SV2C, was associated with LEV response in the Irish cohort (P=0.027). The direction of association was different between cohorts. In the UK cohort the association appears driven by a deficit of major allele homozygotes in the “honeymoon” category and an excess of minor allele homozygotes in the “Partial 20–50%” category whereas in the Irish cohort the opposite occurs. A further four polymorphisms were associated with response in the Irish cohort alone (0.01<P<0.05); none remained significant after Bonferroni correction. There were no further significant associations with response to LEV in the joint analyses.

As there is no consensus for the optimal representation of AED response, we explored our data further by devising five additional phenotyping schemes, collapsing and/or excluding categories from the original clinically driven scheme (Table 7). 16 SNPs were associated with response (P<0.05) in the UK cohort in at least one of the five modified classification schemes. Two of these polymorphisms were also associated with the same response category in the Irish cohort: rs17651293, an intronic SNP in SV2C and rs17594138, an intronic SNP in SV2B. However, the directions of the associations were again different between cohorts and P-values did not remain significant after correction for multiple testing.

As recently described by Cavalleri et al. (2007), in the context of a larger candidate gene study, no polymorphisms were associated with epilepsy predisposition in the UK cohort. Seven polymorphisms were associated with epilepsy predisposition in the Irish cohort (P<0.05). However, none survived correction for multiple testing. Six polymorphisms were associated with mTLE+HS in the UK cohort. One of these, rs2913261, an intronic SNP in SV2C, was also associated with mTLE+HS in the Irish cohort (P=0.027). The direction of the association was different between the two cohorts and the result did not remain significant after correction for six independent comparisons in the Irish cohort. In the joint analysis one further polymorphism, rs7498036, an intronic SV2B polymorphism, was associated with overall epilepsy predisposition and two further polymorphisms, rs7172040 (intronic) in SV2B and rs2431872, 40kb upstream of SV2C, were associated with predisposition to mTLE+HS. None of the associations remain significant after Bonferroni correction.

We assessed the power of our study to detect significant causal variants under various scenarios as follows. First, we considered our ability to detect a variant perfectly tagged (r2=1) by one of the SNPs in our study. We modified the power calculation method of Purcell et al. (2003) to consider association testing in a 3×5 (genotype×phenotype) contingency table for LEV response, and a 3×2 table (for epilepsy predisposition), and we assessed joint power over both cohorts in our study. We considered two extreme situations (in terms of power): a very common variant (allele frequency=0.5) affecting the risk in the larger “no response” category only, and a less common variant (allele frequency=0.1) affecting the risk in the smaller “seizure-free” category only. In both cases we assumed a multiplicative allelic risk model. Using an uncorrected Type I error level of 0.05, the allelic relative risks for LEV response required for 80% power were 1.6 in the very common variant scenario and 2.4 in the less common variant scenario, and using a Bonferroni-corrected Type I error level of 0.05/86 these risks rose to 2.0 in the very common variant scenario and 3.1 the less common variant scenario. For epilepsy predisposition (all cases), the allelic relative risks required for 80% Bonferroni-corrected power were 1.3 in the very common variant scenario and 1.6 in the less common variant calculation; for mTLE+HS, these were 1.5 in the very common variant scenario, and 1.8 in the less common variant calculation.

Genotyping failure of some tags and subsequent additions of SNPs to the HapMap database required us to reassess tagging coverage for each gene using Release 21a. SV2A is still well covered, both by a reassessment of HapMap tags on Release 21a data and by a reassessment of tags from SNPs discovered by sequencing on separate CEPH data. In the former set, all 7 HapMap SNPs with minor allele frequency >0.05 are covered with pairwise r2>0.9. In the latter set, 15 out of 18 SNPs from HapMap release 21a are covered with r2>0.7, two SNPs have r2≈0.5, and one has r2<0.1. For SV2B, genotyping failures resulted in worse coverage: 136 out of 228 SNPs (60%) from HapMap release 21a are covered with r2>0.7, 39 (17%) are covered with 0.3<r2<0.7, and 53 (23%) have r2<0.3. For SV2C, 198 out of 268 SNPs (74%) from HapMap release 21a are covered with r2>0.7, 52 (19%) are covered with 0.3<r2<0.7, and 18 (7%) have r2<0.3. When we modified the power calculation to consider a causal variant with r2=0.7 with any one tag, the allelic relative risks for LEV response required for 80% power were 1.8 in the common variant scenario, and 2.7 in the rare variant scenario (no Bonferroni correction), and rose to 2.4 for the common variant scenario and 3.7 in the rare variant scenario with Bonferroni correction. For epilepsy predisposition, the allelic relative risks required for 80% Bonferroni-corrected power were 1.4 for the common variant scenario and 1.7 for the rare variant calculation; for mTLE+HS, these were 1.6 for the common variant and 1.9 for the rare variant scenario.

Discussion 

We focused on the gene encoding the known binding site of LEV as most known pharmacogenetic variants reside in obvious candidate genes encoding drug targets, drug metabolising enzymes and drug transporters (Goldstein et al., 2003). A study design which considers candidate genes based on the biological action of a drug is a logical and economic way of initiating exploration of genetic predictors of drug response.

We tested 86 common polymorphisms in SV2A, SV2B and SV2C for association with LEV response but none of the associations remained significant after correction for multiple testing. All patients were well-phenotyped with extensive data collected from tertiary referral centres and thus possibly representing a relatively treatment-refractory cohort. Response was classified in the same way for both cohorts and the clinical characteristics of the two cohorts were broadly similar with no significant differences in the proportions of individuals within each response category between the two cohorts. In the UK cohort the mean seizure frequency prior to initiation of LEV treatment is higher in the groups with “no response” or “partial (20–50%) response” but the means are skewed.

Since our study provided very good coverage of variation within SV2A, these findings coupled with our power calculations allow us to state that common variation in SV2A does not have a major effect on LEV response, though we cannot exclude effects of lesser magnitude (allelic relative risk<2 for allele frequency=0.5; allelic relative risk<3.1 for allele frequency=0.1), nor an effect of rare variation.

We also investigated predisposition to epilepsy. As recently described by Cavalleri et al. (2007) who performed an analysis for the possibility of enrichment for P-values in the context of a large set of genes rather than a single SNP analysis as presented here, we did not find any common polymorphisms to be associated with overall epilepsy predisposition in both cohorts. The two cohorts used for investigating predisposition were large and our study was well-powered to detect minimum allelic relative risks of between 1.4 and 1.9 for epilepsy predisposition, thus suggesting our results here are likely to be true negatives.

There are several reasons why we may have failed to find an association between SV2 variation and LEV response ranging from the biological to the methodological. It is possible that common genetic variation in SV2A, SV2B and SV2C does not in fact contribute to clinically important differences in response to LEV: LEV binds SV2A, but it is not proven that this binding directly mediates its clinical effects. Other genetic and non-genetic factors may be more important.

This study design addresses other important phenotyping issues in pharmacogenetics. Firstly, consideration must be given to what is the most relevant clinical question and what question may be addressed with the cohort available. LEV was initially licensed as adjuvant therapy during this study, and therefore, given the cohorts are from two tertiary referral centres it is very relevant to include capture of partial improvement rather than restricting focus to extreme responders. It remains possible that our representation of LEV response does not allow detection of a genetic influence. However, we did investigate several schemes, although all highly correlated, including a scheme focusing on extreme phenotype comparing responders vs. non-responders. Secondly, the allelic architecture underlying LEV, or indeed any AED, response is unknown and therefore we consider it appropriate to use a variety of phenotyping schemes to increase the likelihood of detecting a genetic contribution. It is possible for instance that a SNP contributing to seizure-freedom may have little or no effect on partial improvement and vice versa. Resolving the phenotype, e.g. comparing extreme responders, or segregating focal from generalized epilepsies, has to be balanced against the resultant decline in power with fewer individuals in each category. Thirdly, whether a pharmacogenetic study restricts design to the extreme phenotypes or captures the full complexity of AED response, it will remain a prerequisite for any initial positive finding to be replicated in a similarly phenotyped but different cohort. Larger, homogeneously phenotyped, multicentre studies will be necessary to address these issues. We suggest the phenotypic categorisation and replication cohort design described here may be used in future pharmacogenetic studies examining AED efficacy. Given the large number of polymorphisms considered, any association would have to be strong in order to survive appropriate statistical correction. Our power calculations suggest that even a perfectly tagged causal variant would require a relatively high allelic risk ratio effect size (of at least 2.0) to have a good chance of being detected after correction for multiple testing. This would rise for an imperfectly tagged causal variant (to at least 2.4 if tagged with r2=0.7). Due to genotyping failure and additions to the HapMap database, reassessment of actual tagging coverage suggests that some gaps do exist, especially for SV2B and SV2C. A causal SNP falling into one of these gaps would have a very low chance of being detected. Finally, it is also possible that rare genetic variation may influence LEV response in which case the effect would not be detectable with the current tagging SNP set. As part of a much larger study of epilepsy predisposition (Cavalleri et al., 2007) we considered the SV2A, SV2B and SV2C genes in a separate and different analysis for the possibility of enrichment for P-values in a large set of genes, but not for single SNP analysis as presented here.

Whilst the binding site of a drug is a good starting point for investigating genetic determinants of response, pharmacogenetic variants have been found in genes encoding DMEs (drug metabolising enzymes), drug transporters, and other elements of drug target pathways (Goldstein et al., 2003). As LEV is not extensively metabolised and is not known to be a substrate for any drug transporter, the next obvious step would be to consider genetic variation in the wider pool of candidate genes encoding other elements of the synaptic vesicle pathway. Choosing between sequence vs. map based approaches is discussed in literature (Botstein and Risch, 2003) and is beyond the scope of this study. However, given recent advances in genotyping technology (both in terms of reduced cost and throughput volume) we suggest further study should be genome-wide.