Recently, several classifiers that combine primary tumor data, like gene expression data, and secondary data sources, such as protein-protein interaction networks, have been proposed for predicting outcome in breast cancer. In these approaches, new composite features are typically constructed by aggregating the expression levels of several genes. The secondary data sources are employed to guide this aggregation. Although many studies claim that these approaches improve classification performance over single genes classifiers, the gain in performance is difficult to assess. This stems mainly from the fact that different breast cancer data sets and validation procedures are employed to assess the performance. Here we address these issues by employing a large cohort of six breast cancer data sets as benchmark set and by performing an unbiased evaluation of the classification accuracies of the different approaches. Contrary to previous claims, we find that composite feature classifiers do not outperform simple single genes classifiers. We investigate the effect of (1) the number of selected features; (2) the specific gene set from which features are selected; (3) the size of the training set and (4) the heterogeneity of the data set on the performance of composite feature and single genes classifiers. Strikingly, we find that randomization of secondary data sources, which destroys all biological information in these sources, does not result in a deterioration in performance of composite feature classifiers. Finally, we show that when a proper correction for gene set size is performed, the stability of single genes sets is similar to the stability of composite feature sets. Based on these results there is currently no reason to prefer prognostic classifiers based on composite features over single genes classifiers for predicting outcome in breast cancer.

Integrating gene expression data with secondary data such as pathway or protein-protein interaction data has been proposed as a promising approach for improved outcome prediction of cancer patients. Methods employing this approach usually aggregate the expression of genes into new composite features, while the secondary data guide this aggregation. Previous studies were limited to few data sets with a small number of patients. Moreover, each study used different data and evaluation procedures. This makes it difficult to objectively assess the gain in classification performance. Here we introduce the Amsterdam Classification Evaluation Suite (ACES). ACES is a Python package to objectively evaluate classification and feature-selection methods and contains methods for pooling and normalizing Affymetrix microarrays from different studies. It is simple to use and therefore facilitates the comparison of new approaches to best-in-class approaches. In addition to the methods described in our earlier study (Staiger et al., 2012), we have included two prominent prognostic gene signatures specific for breast cancer outcome, one more composite feature selection method and two network-based gene ranking methods. Employing the evaluation pipeline we show that current composite-feature classification methods do not outperform simple single-genes classifiers in predicting outcome in breast cancer. Furthermore, we find that also the stability of features across different data sets is not higher for composite features. Most stunningly, we observe that prediction performances are not affected when extracting features from randomized PPI networks.

Supplementary data are available at Bioinformatics online.