Development of constrained fuzzy logic for modeling biological regulatory networks and predicting contextual therapeutic effects

Abstract

Upon exposure to environmental cues, protein modifications form a complex signaling network that dictates cellular response. In this thesis, we develop methods for using continuous logic-based models to aide our understanding of these signaling networks and facilitate data interpretation. We present a novel modeling framework called constrained fuzzy logic (cFL) that maintains a simple logic-based description of interactions with AND, OR, and NOT gates, but allows for intermediate species activities with mathematical functions relating input and output values (transfer functions). We first train a prior knowledge network (PKN) to data with cFL, which reveals what aspects of the dataset agree or disagree with prior knowledge. The cFL models are trained to a dataset describing signaling proteins in a hepatocellular carcinoma cell line after exposure to ligand cues in the presence or absence of small molecule inhibitors. We find that multiple models with differing topology and parameters explain the data equally well, and it is crucial to consider this non-identifiability during model training and subsequence analysis. Our trained models generate new biological understanding of network crosstalk as well as quantitative predictions of signaling protein activation. In our next applications of cFL, we explore the ability of models either constructed based solely on prior knowledge or trained to dedicated biochemical data to make predictions that answer the following questions: 1) What perturbations to species in the system are effective at accomplishing a clinical goal? and 2) In what environmental conditions are these perturbations effective? We find that we are able to make accurate predictions in both cases. Thus, we offer cFL as a flexible modeling methodology to assist data interpretation and hypothesis generation for choice of therapeutic targets.