The Promise and Computational Challenges of Personalized Medicine

High hopes for using genetic profiling for personalized medicine have been, in part, driven by the rapid progress of genome-wide association studies such as the recent Collaborative Oncological Gene-environment Study (COGS). We consolidated these genetic variants into a list of 77 SNPs which reflect the state of the art of breast cancer GWAS. We incorporated these SNPs with the descriptors that radiologists observe on mammograms and built naive Bayes models to predict which mammograms indicate a malignancy. We observed that the inclusion of the genetic variants significantly improved the breast cancer diagnostic model, resulting in reduced false positives and alleviated risk of overdiagnosis. Despite great promise, there are still many computational challenges. One computational challenge is that we usually have to perform large-scale multiple testing under dependence due to the correlation structure inherent in genetic data. To solve this problem, we propose a multiple testing procedure which is based on hidden Markov random field model. This procedure can control false discovery rate at the nominal level, and reduce false non-discovery rate by leveraging the dependence structure.