On the limits of data-driven cancer prognosis prediction

We show that a widely used training-data classification rule fails to provide reliable cancer prognosis prediction even in a very simple case with binary data from a relatively small number of genes. We also show that this limitation can be overcome by optimally integrating prior biological knowledge into a model-based framework with a sufficient amount of patient-specific data. This highlights the importance of utilizing knowledge of the dynamical interactions of genes along with personalized data, in contrast with current data-driven efforts to infer cancer prognosis using population data as training samples, in fulfilling the promise of precision medicine.