Files
Abstract
The analysis of gene regulatory networks has emerged as a leading paradigm for understanding how molecules inside cells communicate, process inputs, and coordinate responses to perform processes that sustain cell’s life. Such networks are subject to internal noise, which occurs due to small number of molecules taking part in some reactions within a cell. Noise implies that cellular information processing must go through biochemical networks whose components can fluctuate significantly and unpredictably. This inherent stochasticity in regulatory networks can have major effects on a cell’s fate. It can produce different phenotypes for genetically identical organisms and can lead to different cellular behaviors. Understanding the effects of noise and how cells adapt to it became an important issue in systems biology. There are two broad challenges posed by stochastic biological networks:
(1) Optimally finding the network characteristics/parameters, such as the molecular
numbers of species in the network and reaction rates;
(2) Quantifying the impact of noise on the observed dynamics.
These are fundamental questions that straddle computer science, stochastic analysis and systems biology. However, they raise several new challenges due to the noisy nature of the data, computational intensity required for simulating stochastic behaviors and the inherently interdisciplinary nature of the problems. The aim of this thesis is to address these challenges by proposing new efficient optimization algorithms and developing new methods to tackle the stochastic nature of these problems.
To address the inference of a stochastic network’s parameters, two ensemble Markov Chain Monte Carlo (MCMC) methods were developed. They were used to find parameter sets that best describe the observed behavior of a stochastic oscillatory gene network, the clock network of model organism Neurosporra crassa. Both methods used the average periodogram as a fitting criterion for simulating the behavior observed in the single cell data. The high computational intensity of the methods required the use of Graphical Processing Units (GPUs) to simulate the stochastic behavior of the proposed models. In the first method, Parallel Tempering was successfully used in modeling the clock behavior of single cells observed in the dark (D/D). In the second method a combination of Particle Swarm Optimization algorithms was used to simulate the light entrainment/synchronization of single cells by light observed in a series of light/dark (L/D) experiments.
To quantify the impact of noise on the observed behavior a model that accounts for detection/experimental error in a stochastic oscillatory network was developed. The way this error propagates into the periodogram of oscillatory trajectories was also determined. This allowed us to separate the effects of detection noise and intrinsic noise.
We were able to test the Stochastic Resonance Hypothesis (SRH) to see whether the intrinsic noise was the main driver of the oscillations. SRH was tested for models simulating the network behavior observed in a dark experiment and for models derived from a series of 4 dark/dark and light/dark experiments. In the latter case the test showed that stochastic resonance is produced for a single optimal level of noise across all four experiments, which is pretty remarkable.
(1) Optimally finding the network characteristics/parameters, such as the molecular
numbers of species in the network and reaction rates;
(2) Quantifying the impact of noise on the observed dynamics.
These are fundamental questions that straddle computer science, stochastic analysis and systems biology. However, they raise several new challenges due to the noisy nature of the data, computational intensity required for simulating stochastic behaviors and the inherently interdisciplinary nature of the problems. The aim of this thesis is to address these challenges by proposing new efficient optimization algorithms and developing new methods to tackle the stochastic nature of these problems.
To address the inference of a stochastic network’s parameters, two ensemble Markov Chain Monte Carlo (MCMC) methods were developed. They were used to find parameter sets that best describe the observed behavior of a stochastic oscillatory gene network, the clock network of model organism Neurosporra crassa. Both methods used the average periodogram as a fitting criterion for simulating the behavior observed in the single cell data. The high computational intensity of the methods required the use of Graphical Processing Units (GPUs) to simulate the stochastic behavior of the proposed models. In the first method, Parallel Tempering was successfully used in modeling the clock behavior of single cells observed in the dark (D/D). In the second method a combination of Particle Swarm Optimization algorithms was used to simulate the light entrainment/synchronization of single cells by light observed in a series of light/dark (L/D) experiments.
To quantify the impact of noise on the observed behavior a model that accounts for detection/experimental error in a stochastic oscillatory network was developed. The way this error propagates into the periodogram of oscillatory trajectories was also determined. This allowed us to separate the effects of detection noise and intrinsic noise.
We were able to test the Stochastic Resonance Hypothesis (SRH) to see whether the intrinsic noise was the main driver of the oscillations. SRH was tested for models simulating the network behavior observed in a dark experiment and for models derived from a series of 4 dark/dark and light/dark experiments. In the latter case the test showed that stochastic resonance is produced for a single optimal level of noise across all four experiments, which is pretty remarkable.