bioRxiv2018
Robustly predicting outcome for cancer patients from gene expression is an important challenge on the road to better personalized treatment. Network-based outcome predictors (NOPs), which considers the cellular wiring diagram in the classification, hold much promise to improve performance, stability and interpretability of identified marker genes. Problematically, reports on the efficacy of NOPs are conflicting and for instance suggest that utilizing random networks performs on par to networks that describe biologically relevant interactions. In this paper we turn the prediction problem around: instead of using a given biological network in the NOP, we aim to identify the network of genes that truly improves outcome prediction. To this end, we propose SyNet, a gene network constructed ab initio from synergistic gene pairs derived from survival-labelled gene expression data. To obtain SyNet, we evaluate synergy for all 69 million pairwise combinations of genes resulting in a network that is specific to the dataset and phenotype under study and can be used to in a NOP model. We evaluated SyNet and 11 other networks on a compendium dataset of >4000 survival-labelled breast cancer samples. For this purpose, we used cross-study validation which more closely emulates real world application of these outcome predictors. We find that SyNet is the only network that truly improves performance, stability and interpretability in several existing NOPs. We show that SyNet overlaps significantly with existing gene networks, and can be confidently predicted (~85% AUC) from graph-topological descriptions of these networks, in particular the breast tissue-specific network. Due to its data-driven nature, SyNet is not biased to well-studied genes and thus facilitates post-hoc interpretation. We find that SyNet is highly enriched for known breast cancer genes and genes related to e.g. histological grade and tamoxifen resistance, suggestive of a role in determining breast cancer outcome.\n\nAuthor SummaryCancer is caused by disrupted activity of several pathways. Therefore, outcome predictors analyze patients expression profiles from perspective of gene groups collected from interactomes (e.g. protein interaction networks). These Network based Outcome Predictors (NOPs) hold potential to facilitate identification of dysregulated pathways and delivering improved prognosis. Nonetheless, recent studies revealed that compared to classical models, neither performance nor consistency can be improved using NOPs.\n\nWe argue that NOPs can only perform well under guidance of suitable networks. The commonly used networks may miss associations specially for under-studied genes. Additionally, these networks are often generic with low resemblance to perturbations that arise in cancer.\n\nTo address this issue, we exploit ~4100 samples and infer a disease specific network called SyNet linking synergistic gene pairs that collectively show predictivity beyond individual performance of genes.\n\nUsing identical datasets, we show that a NOP yields superior performance merely by considering groups of genes in SyNet. Further, NOP performance severely reduces if SyNet nodes are shuffled, confirming relevance of SyNet links.\n\nDue to simplicity of our approach, this framework can be used for any phenotype of interest. Our findings represent the value of network-based models and crucial role of interactome in their performance.