Integrating Mechanistic Signaling Models with Multi-Omics Data to Predict Cell Fate Outcomes in Cancer Biology

Cancer is a complex and diverse disease---so much that two people with the same cancer type (lung cancer, for example) may not respond the same way to the same therapy. This happens in large part because each tumor possesses a unique set of mutations, and each set may enable sensitivity to different drugs. Consequently, one of the major obstacles in cancer therapy today is selecting an effective drug cocktail for a given patient. The complexity of cancer transcends many levels. First, there is much inherent complexity in normal biochemical signaling, including feed-back and feed-forward loops as well as significant pathway crosstalk, that make it difficult to predict how a given perturbation will affect the signaling network as a whole, not to mention the ramifications for cell fate. Although there is still much new biochemistry to learn, there is a lot known about the signaling that drives cell proliferation and death, especially for the core pan-cancer related pathways.;Here, we use this understanding to formalize biochemical mechanisms using ordinary differential equation-based mechanistic mathematical models that are rooted in chemical kinetics theory and principles of mass action. We first explore dimerization affinity-based control over cooperativity behavior in simple biochemical systems, proving for the first time that dimerization affinity can completely reverse the cooperativity behavior of a system, in contrast to widely held beliefs. Next, we build a large pan-cancer driver network model that incorporates many of the major signaling pathways implicated across human cancer---receptor tyrosine kinases, Ras/RAF/ERK, PI3K/AKT, mTOR, cell cycle, and apoptosis---as indicated by pan-caner analyses by The Cancer Genome Atlas (TCGA). We develop methods for how to tailor the model to multi-omics data from a specific biological context (here, MCF10A cells) and devise a novel stochastic algorithm to induce non-genetic cell-to-cell fluctuations in mRNA and protein quantities over time. We train the model against a wealth of biochemical (western blot) and cell fate (flow cytometry) data, and gain mechanistic insight into the stochastic control of proliferation and death.;One day, we hope models of this kind could be tailored to patient-derived tumor mRNA sequencing data and used to select patient-specific drug regimens. Towards this goal, we find that protein quantities can be reasonably estimated from mRNA quantities based on a pan-tissue gene-specific protein-to-mRNA ratio; the results of which allows us to use mRNA, instead of protein, data as a model input. We also find the drug response profiles between two large pharmacogenomics databases---the Cancer Cell Line Encyclopedia (CCLE) and the Cancer Genome Project (CGP)---to be reasonably consistent, in contrast to one report, reviving confidence in their use for model building and/or other applications. Finally, as an initial proof-of-concept application, we tailor our pan-cancer driver network model to patient-derived tumor data from TCGA and make predictions regarding drug responsiveness for a group of patients. This works lays the foundation for building context-tailored systems pharmacology models, integrating multi-omics data with mechanistic modeling approaches, with the potential to enable a more rational and personalized drug selection process for patients.