Computational Learning Strategies for Assessing Modular Influences on Biological Phenotypes

Biological organisms are highly complex systems composed of multiple levels of organization. Modeling biological systems is a crucial approach for describing processes from these organisms, and understanding characteristics of critical importance. Modeling approaches have traditionally targeted well confined biological processes located at low levels of organization; these include models that describe interactions among genes, metabolites, or biomolecules. New biological and engineering challenges, however, demand the comprehension of larger and more complex processes, and overall, the study of the interaction among processes at several levels of abstraction. Combining multiple models of these different levels of abstraction to generate multihierarchical models has been identified as a path to describe meaningful phenomena. This goal, however, raises several mathematical and computational challenges. The high complexity associated to biological models impedes the combination of large numbers of models. Mathematical and computational strategies are needed to provide more intuitive information of the models, simplify models, and relate models at several levels of abstraction and to fulfill the demands of the new challenges. Furthermore, I are sometimes not able to relate the low level models that have been developed over the years to meaningful higher level biological phenomena. This results in a need for better understanding how lower level biomolecular components impact higher level phenotypic characteristics.;I present in this document two computer based algorithms that aim to 1) provide functional information needed to understand the underlying structure of models, which can be a critical tool in comprehending the impact of biomolecular components on higher level phenomena, and 2) find a relationship between models of low level processes and more abstract meaningful phenomena, by identifying key groups of components to control or manipulate these phenomena. Particularly, the first algorithm decomposes dynamic models into functional units, using fuzzy clustering to identify primary relationships, secondary relationships, and changes of these relationships over time. The second algorithm leverages machine learning tools and experimental data to find a relationship between a model and a phenomena of interest, and it uses sensitivity analysis to find groups of components that can be used to manipulate the phenomena. This approach is used in combination with the decomposition algorithm to identify the functional modules from complex models that can impact those abstract phenomena.;I show the application of the presented algorithms to two different dynamic biological models, the EGF induced MAPK, and the C3 photosynthesis pathway. I validate the results of my applications by comparing them with previous literature. By showing that the obtained results are similar to others that were experimentally acquired, I demonstrate the potential that these new computational methods can have on the biological perspective.