Subnetwork-based prognostic biomarkers exhibit performance and robustness superior to gene-based biomarkers in breast cancer

Background: Effective classification of cancer patients into groups with differential survival remains an important and unsolved challenge. Biomarkers have been developed based on mRNA abundance data, but their replicability and clinical utility is modest. Integrating functional information, such as pathway data, has been suggested to improve biomarker performance. To date, however, the advantages of subnetwork-based biomarkers have not been quantified. Results: We deeply sampled the population of prognostic gene-based and subnetwork-based biomarkers in a breast cancer meta-dataset of 4,960 patients. Analysing the performance and robustness of 22,000,000 gene biomarkers and 6,250,000 subnetwork biomarkers across twenty different training:testing cohort partitions of the meta-dataset revealed that subnetwork biomarkers exhibit superior overall performance and higher concordance across partitions. We find evidence of an upper bound for optimal biomarker size of ~200 genes or ~100 subnetworks. Additionally, with both biomarker feature types, larger biomarkers tend to show less consistency in performance across partitions, suggestive of over-fitting. Finally, an evaluation of varying training cohort sizes quantifies the effects of training cohort size. Conclusions: Many groups are developing techniques for exploiting network-based representations of biological pathways to characterize cancer and other diseases. By considering the distribution of gene- and subnetwork-based biomarkers, we show that pathway data improves performance and replicability, and that smaller biomarkers are more robust across patient cohorts. These insights may facilitate development of clinically useful biomarkers.