Predicting Type 1 Diabetes Status From Genetics

Genome wide association studies (GWAS) have identified many single nucleotide polymorphisms (SNPs) that are associated with disease. GWAS typically compare allele frequencies of ∼ 500K individual SNPs in case patients to those of control patients using single-feature statistics and hundreds or thousands of patients to gain enough power to overcome the serious multiple testing hypothesis correction burden. The original Wellcome Trust Case Control Consortium (WTCCC) paper published 24 high-confidence SNPs that are associated with 7 major diseases in a study with 14,000 cases and 3,000 shared controls ([1]) Currently, identifying SNPs associated with disease is motivated by two goals: first, to find a parsimonious panel of testable biomarkers for disease, and second, to learn about the biology of disease by examining which DNA markers are correlated with incidence. Given that the cost of genome sequencing is dropping exceptionally fast, and the cost of genotyping tag SNPs on a SNP-chip is even cheaper, the motivation for this first goal is waning—in the future, physicians will not have to prioritize which SNPs to genotype, but will have access to whole-chip or sequencing data for each patient. Additionally, it is becoming increasingly clear that multiple genetic loci, and the interactions between them, are associated with disease states, and the individual-SNP approach does not capture the biological intricacies of many disorders. In this report, I suggest that we can build whole-chip classifiers for disease, with selected features and interaction terms that will help elucidate the biology of disease. I also hope to show that using all of the data results in a better clinical predictor than the state-of-the-art predictors that include only the top individually associated SNPs.