Reconciling Cancer Genotype and Phenotypes by Learning Structured Priors

Опубликовано 20 апреля 2018, 23:44

Cancer is often viewed as a genetic disease because accumulation of mutations in the DNA are said to “drive” the progression of the disease. Depending on the mutations, an individual may exhibit some or many “hallmarks” of cancer, such as genome instability, angiogenesis, etc. However, it is not necessary to accumulate multiple mutations leading to the same hallmark for the disease to progress. This leads to the notion of mutual exclusivity in a mutation profile: different genes may lead to the same disruption in phenotype, but it is unlikely that an individual will have more than one of these mutations. One research goal is to learn a latent variable model of phenotypes such that the latent variables represented by phenotype matches those represented by genotype. In this work, we attempt to reconcile this goal by constructing a structured prior that conditions on mutations, providing a soft version of the notion of mutual exclusivity constraint. We use a variational auto-encoder (VAE) of RNA expression levels using discrete latent factors which naturally fits our research goal of reconstructing phenotypes from latent factors, while enforcing the factors to match from both the phenotypic or genotypic perspectives. Early results from this ongoing work will be presented.

See more at microsoft.com/en-us/research/v...