Logic and Linear Programs to Understand Cancer Response

Abstract

Understanding which are the key components of a system that distinguish a normal from a cancerous cell has been approached widely in the recent years using machine learning and statistical approaches to detect gene signatures and predict cell growth. Recently, the idea of using gene regulatory and signaling networks, in the form of logic programs has been introduced in order to detect the mechanisms that control cells state change. Complementary to this, a large literature deals with constraint-based methods for analyzing genome-scale metabolic networks. One of the major outcome of these methods concern the quantitative prediction of growth rates under both given environmental conditions and the presence or absence of a given set of enzymes which catalyze biochemical reactions. It is of high importance to plug logic regulatory models into metabolic networks by using a gene-enzyme logical interaction rule. In this work, our aim is first to review previously proposed logic programs to discover key components in the graph-based causal models that distinguish different variants of cell types. These variants represent either cancerous versus healthy cell types, multiple cancer cell lines, or patients with different treatment response. With these approaches, we can handle experimental data coming from transcriptomic profiles, gene expression micro-arrays or RNAseq, and (multi-perturbation) phosphoproteomics measurements. In a second part, we deal with the problem of combining both, regulatory and signaling, logic models within metabolic networks. Such a combination allows us to obtain quantitative prediction of tumor cell growth. Our results point to logic program models built for three cancer types: Multiple Myeloma, Acute Myeloid Leukemia, and Breast Cancer. Experimental data for these studies was collected through DREAM challenges and in collaboration with biologists that produced them. The networks were built using several publicly available pathway databases, such as Pathways Interaction Database [39], KEGG [18], Reactome [10], and Trrust [13]. We show how these models allow us to identify the key mechanisms distinguishing a cancerous cell. In complement to this, we sketch a methodology, based on currently available frameworks and datasets, that relates both the linear component of the metabolic networks and the logical part of logic programing-based methods.