Prediction of stage, grade, and survival in bladder cancer using genome-wide expression data: a validation study

To evaluate performances of published gene signatures for the assessment of urothelial carcinoma.

We evaluated 28 published gene signatures designed for diagnostic and prognostic purposes of urothelial cancer. The investigated signatures include eight signatures for stage, five for grade, four for progression, and six for survival. We used two algorithms for classification, nearest centroid classification and support vector machine, and Cox regression to evaluate signature performance in four independent data sets.

The overlap of genes among the signatures was low, ranging from 11% among stage signatures to 0.6% among survival signatures. The published signatures predicted muscle-invasive and high-grade tumors with accuracies in the range of 70% to 90%. The performance for a given signature varied considerably with the validation data set used, and interestingly, some of the best performing signatures were not designed for the tested classification problem. In addition, several nonbladder-derived gene signatures performed equally well. Large randomly selected gene signatures performed better than the published signatures, and by systematically increasing signature size, we show that signatures with >150 genes are needed to obtain robust performance in independent validation data sets. None of the published survival signatures performed better than random assignments when applied to independent validation data.

We conclude that gene expression signatures with >150 genes predict muscle-invasive growth and high-grade tumors with robust accuracies. Special considerations have to be taken when designing gene signatures for outcome in bladder cancer.