Cohort phenotyping

The CENet cohort is composed of patients with Genetic Generalized Epilepsy (GGE) or Non-Acquired Focal Epilepsy (NAFE) collected in CHUM Research Center in Montreal and controls (unaffected Developmental Epileptic Encephalopathy (DEE) trio parents) collected in CHU Ste-Justine in Montreal and the Hospital for Sick Children in Toronto [23–26]. The patients were recruited between 2002 and 2014. Patients were diagnosed by epileptologists. The clinical epilepsy phenotype was classified according to the current classification by the International League against Epilepsy (ILAE) [27]. More specifically for NAFE, patients were at least five years of age and have experienced at least two unprovoked seizures in the six months prior to starting treatment, an MRI scan of the brain that did not demonstrate any potentially epileptogenic lesions, other than mesial temporal sclerosis. Patients with clinical and EEG characteristics meeting the 1989 ILAE syndrome definitions for GGE were included. An MRI of the brain was not required for participation. All patients were at least four years of age at the time of diagnosis. In GGE, we also included patients with Jeavons syndrome, which is an idiopathic generalized form of reflex epilepsy characterized by childhood onset, unique seizure manifestations, striking light sensitivity and possible occurrence of generalized tonic-clonic seizures. Certain cases were found with an epilepsy phenotype different from the other affected family members, hence they were marked as ‘mixed’. Only one affected GGE or NAFE patient was used per family, therefore all the individuals in this study are unrelated. We used WGS from 377 controls and 247 patients, 123 GGE, 112 NAFE and 12 mixed patients (Table 1). This study was approved by the CHUM research Center (CRCHUM) ethics committee and written informed consent was obtained for all patients (2003–1394, ND 02.058 -BSP (CA)). We did not have access to information that could allow us to identify the patients.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Number of individuals for each phenotype.

https://doi.org/10.1371/journal.pone.0291935.t001

Sequencing

DNA was extracted at the time of recruitment. All samples were sequenced at the same time in 2015. Samples were sequenced for the whole genome at 30X coverage at Genome Quebec Innovation Center in Montreal. gDNA was cleaned using ZR-96 DNA Clean & ConcentratorTM-5 Kit (Zymo) prior to being quantified using the Quant-iTTM PicoGreen dsDNA Assay Kit (Life Technologies) and its integrity assessed on agarose gels. Libraries were generated using the TruSeq DNA PCR-Free Library Preparation Kit (Illumina) according to the manufacturer’s recommendations. Libraries were quantified using the Quant-iTTM PicoGreen dsDNA Assay Kit (Life Technologies) and the Kapa Illumina GA with Revised Primers-SYBR Fast Universal kit (Kapa Biosystems). The average size fragment was determined using a LabChip GX (PerkinElmer) instrument. The libraries were denatured in 0.05N NaOH and diluted to 8pM using HT1 buffer. The clustering was done on an Illumina cBot and the flowcell was run on a HiSeq 2500 for 2×125 cycles (paired-end mode) using v4 chemistry and following the manufacturer’s instructions. A phiX library was used as a control and mixed with libraries at 0.01 level. The Illumina control software used was HCS 2.2.58 and the real-time analysis program used was RTA v. 1.18.64. bcl2fastq v1.8.4 was used to demultiplex samples and generate fastq reads. The filtered reads were aligned to reference Homo_sapiens assembly b37. Each readset was aligned using BWA-MEM version 0.7.10 to create a Binary Alignment Map file (.bam). Bam files were processed to gvcf files and we performed joint calling of gvcf files that were merged into a single vcf file using GATK version 3.7–0 [28]. The vcf file was recalibrated and filtered following the GATK best practice guidelines.

Data cleaning

Cleaning was made using plink v2.0 [29]. First, Single Nucleotide Variants (SNV) with a call rate below 98% were removed using ‘—geno 0.02’. Afterward, individuals with a genotype rate below 98% were excluded using ‘—mind 0.02’. Next, SNV that did not follow the Hardy-Weinberg equilibrium were removed using ‘—hwe 0.001’. Finally, individuals with unknown biological sex were excluded. To make the principal component analysis, further cleaning was required. Only common variants were used in the PCA (maf > 0.05) and SNV that were not in linkage disequilibrium by using ‘—indep-pairwise 50 5 0.2’. Single nucleotide variants (SNVs) were filtered based on their minor allele frequency (maf), only rare variants (maf <0.01) as assessed in our cohort as well as in gnomAD v3.1.2 (non-Finnish European) were kept [30]. Variants were liftover to GRCh38 to get the most recent frequency data with an increased non-coding DNA coverage.

Statistical analyses

We applied ExPecto on rare variants (MAF<0.01) for three tissues related to epilepsy: hippocampus, amygdala and brain cortex (GTEx V6) for which we calculated the gene expression change median [16, 31]. ExPecto computes gene expression changes by using a neural network to predict the effect of variants on features such as transcription factors, histone marks and DNA accessibility. It then transforms those feature predictions in tissue specific gene expression changes with L2-regularized linear regression models. Gene expression changes calculated by ExPecto were used to compute a Constraint Violation Score (CVS) in accordance with the methods described in Zhou et al. [21]. The CVS quantifies how deleterious the gene expression change is: the higher the score, the more deleterious the variant.

The accuracy of ExPecto predictions was validated using known eQTLs from the GTEx V6p release (S1 Fig) [16]. To do so, we used two parameters, the prediction’s directionality and magnitude. Directionality is defined as whether the Single Nucleotide Variant (SNV) increases or decreases gene expression. Magnitude is defined as the absolute size of the effect in gene expression fold change (natural log). We determined that the magnitude above which the accuracy of the prediction’s directionality was perfect was 0.2 so we kept only variants with a median (for the cortex, hippocampus, and amygdala) above this threshold (S1 Fig). Analyses were replicated with the median of three non-neurological tissues: artery aorta, colon transverse and skin of body to validate the tissue specificity of the model (S2 Fig).

A binomial logistic regression was performed with the python package statsmodels v0.12.2. Sex was used as a covariate as well as a two-dimension UMAP (umap-learn v0.5.1) based on the first five principal components (plink v2.0) (S3 Fig). We had access to self-declared ethnicity from the affected individuals, which allowed us to confirm that the UMAP was accurate. Analyses were replicated by using only individuals of French-Canadian and European descent to validate that the findings were not related to population structure (S4 Fig). Those individuals were selected based on the cluster that corresponded with self-declared French-Canadians and Europeans.

Limitations

The study has two main limitations. First, the use of deep learning, as useful as it is, has the limitation of being predictions, not observations. Nonetheless, we validated that those predictions were accurate by using experimental data from GTEx (S1 Fig). Additionally, ExPecto is not able to compute long range (>40kb) sequence effects on gene expression, which limits this work to the study of short-range interactions. The second limitation is our small sample size. Despite this, we were successful in identifying a genome-wide effect in individuals of European descent, but we lacked power to investigate those effects at a gene or variant level resolution.

Conclusion

This study reveals the importance of non-coding regions in the etiology of epilepsy. The effect was specific for GGE, whereas NAFE showed no significant difference with controls. Therefore, our results indicate that the differences between those subphenotypes extends to non-coding genetic mechanisms. Larger WGS cohorts will be needed to deepen our understanding of the role of non-coding regions in epilepsy.