EMBRACING NOISE IN BIOINFORMATICS

Abstract

In 1953, James Watson and Francis Crick discovered the structure of DNA. This eventually led to the Human Genome Project, which was completed in 2003. The post- genomic era opens up exciting possibilities, along with grand challenges to overcome. One of which is to build a mathematical model of the whole cell.

The first part of this thesis focuses on building efficient and practical tools for model calibration and validation that are scalable to handle models of massive sizes. We built two powerful and easy-to-use software (DA and MIRACH) for estimating parameters¿ distribution of a given biological system and testing whether certain given properties are satisfied by a given biological system. We then combined the technology of these two software to design a framework that allows us to perform parameter estimation, even when time series data are not available, by using known biological properties and model checking.

In building these tools, we utilized state-of-the-art hypothesis testing algorithms, which are necessary for interpreting the stochastic output of biological systems, and discovered that they came with practical limitations. This leads us to the second part of the thesis, where we developed algorithms to overcome these limitations. Specifically, we developed two novel algorithms for sequential hypothesis testing that are compu- tationally faster and more memory efficient. In addition, by integrating sequential hypothesis testing algorithms with bagging, we developed a new powerful algorithm which we named dynamic bagging. This algorithm supersedes standard bagging by having all the benefits of standard bagging but is more efficient and removes the need to arbitrarily fix a priori the number of bootstrap replicates. We first used dynamic bagging in gene expression profile analysis to overcome batch effects that have plagued many gene expression analysis projects. We then went on to show that its usefulness is not limited to any problem domain. We also show that predictions from dynamic bag- ging is consistent to standard bagging with much larger number of bootstrap replicates. Finally, we offered an alternative and more direct explanation of bagging¿s effectiveness than the classical explanation based on bias-variance decomposition.