REPOSITORY > RESULTS

Doctoral dissertation

Deterministic and stochastic process-based modeling and design of dynamic systems in biology

Author(s): Jovan Tanevski (Author), Sašo Džeroski (Supervisor), Ljupčo Todorovski (Co-Supervisor)

Thesis defense date: 24.08.2016

Organization: MPŠ - Mednarodna podiplomska šola Jožefa Stefana

PID: 20.500.12556/ReVIS-15034

Views: 0 | Downloads: 0

Abstract

The work presented in this thesis is situated at the intersection of three research areas: systems
and synthetic biology, mathematical modeling and machine learning. Systems biology
is concerned with understanding the principles of life and the emergence of its complexity
from a systems-level perspective. Synthetic biology is the application of knowledge learned
from systems biology for the purpose of design of novel living systems that behave in an
expected and beneficial manner. Mathematical modeling is an important aspect of systems
and synthetic biology. It is an established approach to acquiring knowledge about
the structure, function and behavior of dynamical systems. Process-based modeling is a
machine learning approach to computational scientific discovery that constructs complete
mathematical models of dynamical systems from knowledge and data.
The major limitations of process-based modeling are the coarse and deterministic formal
description of models and domain knowledge, the limited use of knowledge for model
selection and the confinement to completely data-driven scenarios. In this work, we overcome
these limitations of process-based modeling approaches and make them more suitable
for application to tasks of modeling and design of dynamical biological systems. In particular,
we extend process-based modeling towards representing and learning of stochastic
models, the use of complex and domain-specific model selection criteria and, finally, towards
process-based design.
We present an improved formalism for process-based modeling, which uses a finer
grained representation of models and domain specific modeling knowledge. This formalism
describes interactions, i.e., processes as sets of reaction equations. The new formalism
allows for both deterministic and stochastic interpretation of models and modeling knowledge.
We address the problem of model selection by adapting regularization to the specifics of
process-based modeling. We further propose to strengthen the evaluation bias of the learning
process by introducing domain-specific criteria and demonstrate that this outperforms
standard approaches to model selection in scenarios with limited observability.
We extend the scope of process-based modeling towards the task of designing dynamical
biological systems that can achieve a desired behavior. We relax the dependence of the
approach to data, make a shift towards completely knowledge-driven learning and take
advantage of methods for simultaneous optimization of multiple problem-specic objective
functions. This allows for process-based design of novel biological systems with desired
properties and behaviors.
The presented improvements are evaluated on multiple tasks of modeling and design
from the domains of systems and synthetic biology. The evaluation shows that the processbased
approach can successfully reconstruct, in an automated manner, results from the literature
obtained by manual or related computational approaches. Our approach performs
well on both synthetic and real world problems of modeling and design of deterministic
and stochastic dynamical systems. Using the new formalism, our process-based modeling approach was successfully applied
to the reconstruction of the structure and dynamics of gene regulatory networks with
global and local kinetic rates and the learning of compartmental epidemiological models for
the Eyam plague outbreak and the Tristan da Cunha influenza outbreak. The strengthening
of the evaluation bias by domain-specific criteria was applied to solve the model
selection problem in the task of modeling the dynamics of the Rab5-Rab7 switch in endocytosis.
Finally, our process-based approach was also successfully evaluated by applying it
to two design tasks of constructing a stochastic toggle switch without cooperativity and a
deterministic oscillator.

Attachments

Cite this work