DrugLogics: logical models for drug screen prioritization
Åsmund Flobak, Tonje S. Steigedal, Barbara Niederdorfer, Liv Thommesen, Martin T. R. Kuiper, Astrid Lægreid · 2016
Multi-drug precision oncology is in need of approaches that enable drug combination prioritization, since the combinatorial explosion renders traditional trial-and-error screening approaches ineffective. Our computation-assisted approach contributes by highly efficient prediction of drug responses while relying only on characterizing the experimental system (cell line, tumor) at baseline conditions. Logical models are derived from cancer signaling topologies, calibrated to particular cell types or tumors by steady state biomarkers from unperturbed cells. Based on a proof-of-concept model (Flobak et al. PLoS Comp Biol, 2015) we now explore a pipeline for automated causal network topology assembly, logical model parameterization and model ensemble evaluation. Prior knowledge is integrated from databases on cell signaling (Reactome, Signor, SignaLink etc), and multi-omic data is integrated to describe patterns of signaling entity activities characterizing a given experimental system. While parameterization constraints are traditionally obtained from drug response experiments, we explore self-contained topologies as a means of constraining possible parameterizations. Genetic algorithms are employed to optimize logical equations to obtain models where the attractor recapitulates steady state biomarkers from biological assays. Model simulations suggest prioritization of drugs that targets proteins to disrupt disease phenotypes, leading to restoration and activation of regulatory anti-survival phenotypes present in healthy cells. The pipeline correctly classified 20 of 21 combinations as synergistic or non-synergistic (Flobak 2015), with one novel synergy validated in vivo . When applied to a manually curated topology, models automatically parameterized predicted five synergies (four true positives, no false negatives) when normalized to topology-intrinsic synergies. A Reactome-topology-derived model predicted five synergies (three true positives). In ongoing work, model predictions are challenged with a dataset of 171 drug combinations (19 individual drugs) across 8 cell lines. Our prototype computational-experimental pipeline demonstrates the potential to economize pre-clinical drug combination synergy discovery and to provide clinical decision support for personalized therapy.