I will discuss the application of graphical models to two problems in biomedicine. The first problem is HIV vaccine design. I will describe how graphical models can be used to discern the evolution of HIV in an individual in response to their immune system, a key ingredient to successful vaccine design. I will discuss how working on this problem exposed a weakness in the process of learning graphical models from data — namely, the inability to quantify how many arcs in a learned graphical model are spurious, and describe one solution based on the False Discovery Rate.

The second problem is genome-wide association studies — the search for correlations between an individual’s single nucleotide polymorphisms (SNPs) and phenotypes of interest, such as whether he or she gets a particular disease or how he or she will react to a drug treatment. One of the key weaknesses of standard methods for genome-wide association studies is lack of statistical power. I will describe how methods based on graphical models yield more power than these approaches.