Optimal Bayesian cancer prognosis with model-constrained robust intervention

This work addresses optimal cancer prognosis under robust control relative to an uncertainty class of different subtypes or stages of cancer. Specifically, we model the inherent unpredictability of somatic gene mutations and aberrant pathway functioning in cancer by assuming that the precise regulatory relationships between genes, which relate to prognosis, belong to an uncertainty class of plausible mutations of some known healthy network. We implement model-constrained robust intervention relative to this uncertainty class, and train an optimal classifier to predict prognosis under this robust treatment given a snapshot of the patient gene activity profile. While accurate prognosis is possible, we show that performance depends on many factors.