Development of Pharmacogenomic Classifier Models for Epilepsy Treatment Outcomes

Terry and Melanie's presentation 4MB

Approximately 30% of patients treated for epilepsy, the most common chronic neurological disease, will not achieve seizure control. The identification of genetic markers that provide accurate prediction in an individual patient of their biological chance of seizure control with AED treatment would have significant clinical and scientific value. However, it is unlikely in complex diseases such as epilepsy, affecting heterogeneous populations, that a single SNP will adequately explain treatment outcomes. Our group has developed and published an approach to developing multigenic models to classify treatment outcomes - utilizing dimension reduction, machine learning (kNN) and cross-validation approaches ^[1]. This was developed on a dataset of 4,041 SNPs in 115 prospectively followed newly treated epilepsy patients, and the derived classifier (using 5 SNPs) provided positive predictive values >80% for diagnosing successful seizure control for both the development cohort and subsequently in two different validation cohorts. In all three cohorts the multi-SNP model performed better than any of the five individual SNPs alone. This represents "proof of concept" for such approaches.

As a VLSCI Stage 0 project we are applying this approach in an attempt to develop a multigenic classifier for patients in a group of epilepsies, the idiopathic generalised epilepsies (IGE), which have a strong genetic contribution to their pathogenesis. The study cohort of consists of 412 patients genotyped for the same 4,041 SNPs on the same Illumina™ platform as for the original newly treated cohort. The primary aim is to identify a multigenic classifier for treatment response (i.e. seizure control) in patients with IGE. A secondary aim of the study will be to identify multiSNP combinations which distinguish patients with IGE from those with non-familial cryptogenic focal epilepsy (n=156). It will also prepare the computational and bioinformatic framework for our proposed VLSCI Stage 1 project which will utilize a unique international collaborative dataset of almost 1900 newly treated epilepsy patients with treatment outcomes who will each be genotyped on the Illumina 660™ platform for 550,000 SNPs genome wide. This large phenotype-genotype dataset will provide a powerful opportunity to identify multigenic classifiers that predict: (i) the relative chance of seizure control with different commonly prescribed AEDs, (ii) the chance of seizure control in patients with different epilepsy syndromes, (iii) interactions between both. A successful outcome would represent a very important step forward in the quest to realise the dream of personalised medicine for epilepsy, as well as inform bioinformatic approaches more broadly to identifying multigenomic classifiers for complex diseases.

High throughput genotyping platforms have given us the ability to identify risk loci for many common complex diseases. These platforms can also be used to identify loci that determine drug response. The identification of risk loci is of great importance in itself but the ability to predict, with a reasonable sensitivity and specificity, the likely drug response for a patient based on their genetic profile is of greater clinical relevance and may be of benefit in the clinic. Epilepsy treatment is a good example where such information would be of great use to avoid toxicity or lack of response.

There are some important differences between the two approaches. The drug response prediction problem is a different beast to the identification of the underlying loci since we merely wish to achieve the highest sensitivity and specificity. We do not need to know what drives the prediction signature. All GWAS require two stages designs for publication: a discovery phase followed by a validation phase. The same is true for prediction models. We can test performance of prediction models in the discovery phase using strategies such as cross validation.

Dr. Bahlo will outline our approaches that her group have used on a pharmacogenomics Hepatitis C study. These are mainly based on standard statistical approaches such as data reduction and logistic model building. Professor O'Brien will discuss current problems with this approach and specific issues that plague genetics data.

1. Petrovski S, Szoeke CEI, Sheffield LJ, D'Souza W, Huggins R, O'Brien TJ. A multi-SNP pharmacogenomic classifier is superior to single SNP models for predicting drug outcome in complex diseases. Pharmacogenetics and Genomics 2009;19:147-152.