Simpler evaluation of predictions and signature stability for gene expression data.

Scientific advances are raising expectations that patient-tailored treatment will soon be available. The development of resulting clinical approaches needs to be based on well-designed experimental and observational procedures that provide data to which proper biostatistical analyses are applied. Gene expression microarray and related technology are rapidly evolving. It is providing extremely large gene expression profiles containing many thousands of measurements. Choosing a subset from these gene expression measurements to include in a gene expression signature is one of the many challenges needing to be met. Choice of this signature depends on many factors, including the selection of patients in the training set. So the reliability and reproducibility of the resultant prognostic gene signature needs to be evaluated, in such a way as to be relevant to the clinical setting. A relatively straightforward approach is based on cross validation, with separate selection of genes at each iteration to avoid selection bias. Within this approach we developed two different methods, one based on forward selection, the other on genes that were statistically significant in all training blocks of data. We demonstrate our approach to gene signature evaluation with a well-known breast cancer data set.