Patterns of somatic mutations in the tumor may affect prognosis of cancer patients, and many studies routinely collect tumor tissues and sequence known commonly mutated genes. We consider a problem of building a prognostic model that will define a patient’s survival risk based on the somatic mutations that are observed in the patient’s tumor. These mutations might co-occur or be mutually exclusive, and they might affect the survival through both main effects and interactions. Various decision trees methods are available for such problem. We develop an alternative approach where we consider each possible combination of mutations, and then we cluster their survival curves into few risk groups. Various distance metrics between survival curves and clustering methods are compared using simulated data.
This is joint work with Mithat Gonen and Sean Devlin.