Abstract:

Advances in genomics, proteomics, metabolomics, and systems biology have recently provided the nascent possibility to conduct systems biology studies that aim to understand the quantitative dynamical behavior of complex biological networks and to optimally control/manipulate their properties. To fulfill these promises, however, the technological capabilities must be integrated with suitable mathematical and algorithmic tools to handle the complexities in studying bionetworks, such as network nonlinearity, large degree of biological noise, and limited manipulation and measurement capabilities.

This dissertation introduces a number of special concepts and computational protocols to deal with the above issues. The goals are to enable reliable and effective identification, analysis, and control of complex bionetworks. Unlike most systems biology studies, this dissertation extensively utilizes the concepts and techniques in optimal control engineering so that these goals can be achieved with minimal laboratory and computational cost.

A closed-loop identification protocol (CLIP) is first developed to extract bionetwork model parameters from noisy laboratory measurements. This protocol utilizes global inversion algorithms to extract the full distribution of model parameters consistent with the laboratory data. More importantly, it iteratively seeks out the optimal laboratory perturbations and measurements that most effectively filter out the data noise and enhance sensitivity to the desired parameters. In this fashion, the highest quality model parameters can be efficiently attained from inverting the tailored laboratory data.

This dissertation then introduces a special random-sampling high dimensional model representation (RS-HDMR) algorithm to unravel the nonlinear, hierarchical input-output relationships of complex bionetworks. RS-HDMR is first utilized to assist the optimization of artificial gene circuits. Several algorithmic enhancements are then made to allow for reliable paxametric analysis of a dynamically heterogeneous neuron model.

Lastly, a closed-loop learning algorithm is utilized to optimally control the quantitative properties of complex bionetworks. This algorithm is first applied to optimize the dynamical behavior of a Hodgkin-Huxley neuron. It is then utilized, in a feasibility study, to achieve model-free optimization of deep brain stimulation of the subthalamic nucleus for treating Parkinson's disease.