Statistical Methods in Precision Medicine: Employing Systems Biology forCancer Survival Prediction

[摘要] Many cancer drugs showed limited therapeutic effects in fighting thetumor in a certain proportional of patients, due to the heterogeneity oftumors. Nevertheless, the current clinical pathological factors have notreached the expectation of accuracy in discriminating cancer patients.It is believed that this heterogeneity is, for a large part, geneticallydetermined and rooted in molecular profile of the patient. Precisionmedicine has been initiated by White House to expand cancer genomicsas a short-term goal to develop better prevention and treatmentmethods for more cancers. Recent high-throughput technologiescan easily and robustly generate large-scale molecular profiling data,offering extraordinary opportunities to develop molecular signatureor biomarkers through predictive modeling on the patients’ survivalor metastatic status. Notably, in analyzing these large-scale data, apotential statistical challenge arises in which the number of predictorvariables greatly exceeds the sample size. The classical Cox proportionalhazard model cannot simultaneously analyze a large number of and/or correlated predictors, due to the problems of non-identifiabilityand possibly overfitting. To date, various statistical approaches havebeen applied in analyzing large-scale molecular profiling data to buildpredictive models for cancer survival prediction and prognosis, whichwill be discussed in the following section.