Optimization of logical networks for the modelling of cancer signalling pathways

Cancer is one of the main causes of death throughout the world. The survival of patients diagnosed with various cancer types remains low despite the numerous progresses of the last decades. Some of the reasons for this unmet clinical need are the high heterogeneity between patients, the differentiation of cancer cells within a single tumor, the persistence of cancer stem cells, and the high number of possible clinical phenotypes arising from the combination of the genetic and epigenetic insults that confer to cells the functional characteristics enabling them to proliferate, evade the immune system and programmed cell death, and give rise to neoplasms. To identify new therapeutic options, a better understanding of the mechanisms that generate and maintain these functional characteristics is needed. As many of the alterations that characterize cancerous lesions relate to the signaling pathways that ensure the adequacy of cellular behavior in a specific micro-environment and in response to molecular cues, it is likely that increased knowledge about these signaling pathways will result in the identification of new pharmacological targets towards which new drugs can be designed.

As such, the modeling of the cellular regulatory networks can play a prominent role in this understanding, as computational modeling allows the integration of large quantities of data and the simulation of large systems. Logical modeling is well adapted to the large-scale modeling of regulatory networks. Different types of logical network modeling have been used successfully to study cancer signaling pathways and investigate specific hypotheses. In this work we propose a Dynamic Bayesian Network framework to contextualize network models of signaling pathways. We implemented FALCON, a Matlab toolbox to formulate the parametrization of a prior-knowledge interaction network given a set of biological measurements under different experimental conditions. The FALCON toolbox allows a systems-level analysis of the model with the aim of identifying the most sensitive nodes and interactions of the inferred regulatory network and point to possible ways to modify its functional properties. The resulting hypotheses can be tested in the form of virtual knock-out experiments. We also propose a series of regularization schemes, materializing biological assumptions, to incorporate relevant research questions in the optimization procedure. These questions include the detection of the active signaling pathways in a specific context, the identification of the most important differences within a group of cell lines, or the time-frame of network rewiring.

We used the toolbox and its extensions on a series of toy models and biological examples. We showed that our pipeline is able to identify cell type-specific parameters that are predictive of drug sensitivity, using a regularization scheme based on local parameter densities in the parameter space. We applied FALCON to the analysis of the resistance mechanism in A375 melanoma cells adapted to low doses of a TNFR agonist, and we accurately predict the re-sensitization and successful induction of apoptosis in the adapted cells via the silencing of XIAP and the down-regulation of NFkB. We further point to specific drug combinations that could be applied in the clinics. Overall, we demonstrate that our approach is able to identify the most relevant changes between sensitive and resistant cancer clones.