Objectives: Most pharmacogenomic studies have attempted to identify single nucleotide polymorphism (SNP) markers that are predictive for treatment outcomes. It is, however, unlikely in complex diseases such as epilepsy, affecting heterogeneous populations, that a single SNP will adequately explain treatment outcomes. This study reports an approach to develop a multi-SNP model to classify treatment outcomes for such a disease and compares this with single-SNP models.
Methods: A prospectively collected dataset of outcomes in 115 patients newly treated for epilepsy, with genotyping for 4041 SNPs in 279 candidate genes, was used for the model development. A cross-validation-based methodology identified SNPs most influential in predicting seizure control after 1 year of drug treatment and then incorporated these into a multi-SNP classification model; using the k-Nearest Neighbour (kNN) supervised learning approach. The classifier was cross-validated to determine its effectiveness in predicting treatment outcome in the developmental cohort and then in two independent validation cohorts. In each, the classification by the multi-SNP model was compared with that of models using the individual SNPs alone.
Results: Five SNPs were selected for the multi-SNP model. Cross-validation showed that the multi-SNP model had a predictive accuracy of 83.5% in the developmental cohort and sensitivity and positive predictive values above 80% in both the independent validation cohorts. In all cases, the multi-SNP model classified the treatment outcomes better than those using any individual SNPs alone.
Conclusion: The results show that a classifier using multiple SNPs can predict treatment outcome more reliably than single-SNP models. This multi-SNP classifier should be tested on data from newly diagnosed epilepsy populations to determine its broad clinical validity. Our method to developing a multi-SNP classifier could be applied to pharmacogenomic studies of other complex diseases.