Abstract:
|
Lack of replication on findings and missing heritability are two of the many challenges in Pharmacogenetics (PGx) studies in building predictive models for common disease prognosis and drug response. To build up a predictive model from genomic SNP (single nucleotide polymorphism) data usually needs a two-step process: first scan and rank top SNPs to a manageable size (feature selection) followed by a second step model building (two step/stage modeling). In our first simulation study, we compared one step and two step model building with five approaches: Elastic Net (EN), genome wide association study (GWAS) + EN, Principal Component Regression (PCR), Random Forest (RF) and Support Vector Machine (SVM). We used genotype data of 9,968 SNPs on Chromosome One by Illumina Infimium Omni5Exome array on 535 real samples. We randomly select 5 SNPs to generate a quantitative phenotype using a linear model and used a fivefold cross validation (CV) and 250 replications. The results have shown that EN has the smallest test MSE, highest sensitivity and causal %. In the second simulation, we compared three two-step approaches, GWAS+EN, GWAS+RF and GWAS+SVM. The GWAS+RF has the smallest test MSE (mean squared error) and best accuracy in picking up the seeded causal variants. In the third simulation study, we compared two cross validation procedures: GWAS +EN and modified CV GWAS +EN. The results show that the CV GWAS + EN has better prediction accuracy but at a huge computational cost.
|