Constructing Prognostic Gene Expression Profiles for Breast Cancer Survival

Derick R. Peterson, PhD
Associate Professor of Biostatistics & Computational Biology and Oncology
Director, Biostatistics Shared Resource for the James P. Wilmot Cancer Center

Modern micro-array technologies allow us to simultaneously measure the expressions of a huge 
number of genes, some of which are likely to be associated with cancer survival. While such 
gene expressions are unlikely to ever completely replace important clinical covariates, 
evidence is already beginning to mount that they can provide significant additional predictive 
information. The difficult task is to search among an enormous number of potential predictors 
and to correctly identify most of the important ones, without mistakenly identifying too many 
spurious associations. Many commonly used screening procedures unfortunately over-fit the 
training data, leading to subsets of selected genes that are unrelated to survival in the 
target population, despite appearing associated with the outcome in the particular sample of 
data used for subset selection. And some genes might only be useful when used in concert with 
certain other genes and/or with clinical covariates, yet most available screening methods are 
inherently univariate in nature, based only on the marginal associations between each 
predictor and the outcome. While it is impossible to simultaneously adjust for a huge number 
of predictors in an unconstrained way, we propose a method that offers a middle ground where 
some partial adjustments can be made in an adaptive way, regardless of the number of 
candidate predictors.