PRER: A patient representation with pairwise relative expression of proteins on biological networks

Official URL: http://dx.doi.org/10.1101/2020.06.16.153999

Alterations in protein and gene expression levels are often used as features to predictivemodels such as clinical outcome prediction. A common strategy to combine signals on individualproteins is to integrate alterations with biological knowledge. In this work, we propose a novelpatient representation where we integrate the expression levels of proteins with the biological net-works. Patient representation with PRER (Pairwise Relative Expressions with Random walks)operates in the neighborhood of a protein and aims to capture the dysregulation patterns in pro-tein abundance for proteins that are known to interact. This neighborhood of the source proteinis derived using a biased random-walk strategy on the network. Specifically, PRER computes afeature vector for a patient by comparing the protein expression level of the source protein withother proteins’ levels in its neighborhood. We test PRER’s performance through a survival predic-tion task in 10 different cancers using random forest survival models. PRER representation yieldsa statistically significant predictive performance in 8 out of 10 cancer types when compared toa representation based on individual protein expression. We also identify the set of proteins thatare important not because of alteration of its expression values but due to the alteration in theirpairwise relative expression values. The set of identified relations provides a valuable collection ofbiomarkers with high prognostic value. PRER representation can be used for other complex diseasesand prediction tasks that use molecular expression profiles as input