DrugLogics: An ensemble model analysis related to drug prediction performance
John Zobolas, Miguél Vázquez, Martin T. R. Kuiper, Åsmund Flobak · 2019
We have developed a computational pipeline which processes two main inputs: a cancer cell fate decision network and the activity state profile of a particular cell line, derived from gene expression and copy number variations data from the Cancer Cell Line Encyclopedia. Using a logical modeling framework coupled with a genetic algorithm approach, our pipeline trains an ensemble of models to fit to the corresponding steady state activity profile of a specific cancer cell line. Then, based on a given drug panel, a drug response analysis is performed on these trained logical model ensembles and the pairwise drug combination in-silico predictions are compared to the experimentally observed ones. In this study, we show that even though the ensembles of models are trained to a specific consensus steady state profile and exhibit a match (fitness) of more than 50% to it, they still show a diverse individual stable state pattern landscape which leaves room for uncertainties in our steady state inference and thus allows for a more robust drug response analysis. We also devise strategies to split and compare the cell-specific trained models based on individual performance characteristics (number of true positive predictions) or the prediction of specific drug combination sets. This enables a mechanistic approach to find nodes whose state is decisive for the global behavior of the model, and thus represent potential biomarkers.