| المستخلص: |
To address one of the most challenging biological systems issues at the cellular level, this thesis surveys the soft computing methods used in gene regulatory networks (GRNs) inference while stressing the major applications and issues related to usability, complexity and tractability of these methods. The work is motivated by the fact that soft computing methods, such as neural networks (NNs), genetic algorithms (GAs) and fuzzy systems (FSs) hold a noticeable position among the computer science approaches used in bioinformatics. GRNs represent causal relationships between genes that have a direct influence on the life and the development of living organisms, and provide a useful contribution to the understanding of the cellular functions as well as the mechanisms of diseases. It must be stressed that the most common pathologies are not triggered by the alteration of a single gene, rather these are multifaceted infections that arise owing to the dynamic interaction of many genes along with other environmental factors. To infer the dynamic behavior of GRNs, we therefore need to understand the interplay between many genes. One important issue in computational biology, is thus, that of finding the best possible set of regulatory interactions between genes - inferring the GRN - from partial knowledge, as given for example by means of gene expression profile (GEP) data.
GRN inference from GEP is now rapidly evolving as a discipline in its own right. As far as GRN inference using GEPs is concerned, it is important to note the metabolites and proteins are considered as hidden variables. Communication between genes is mediated by these variables and their effects appear as edges between the observed variables. Thus, in an implicit way, the GRN provides communication between genes along with all regulatory process of living cells, and offers a thorough description of cellular regulation concerning most of the gene’s activities. Furthermore, GRN as a formalism can be considered as phenomenological since the underlying mechanisms are mostly unknown up to now as they correspond to complex tracks through metabolites and proteins. In this thesis, and within the framework of GRN inference, we will review and compare five standard soft computing methods, namely fuzzy systems (FSs), artificial neural networks (ANNs), genetic algorithms (GAs), support vector machine (SVM) and rough sets (RSs). In addition, we will also consider the main hybrid methods such as neuro-fuzzy and neuro-genetic methods. We show that these soft computing methods provide viable computational tools for inferring GRNs from GEP data which might contribute to the discovery of gene faulty interactions responsible for specific diseases and/or ad-hoc correcting therapies. Technologically speaking, GRN inference is greatly enhanced by increasing computational power and high throughput devices that have provided powerful means to manage these challenging biological systems at different levels from cell to society – globally. The thesis objectives can be summarized as follows: 1. Make a comprehensive review of the GRN field with special emphasis on computational issues by reporting, presenting and discussing the main contributions of this multidisciplinary field in a coherent and structured novel scientific framework. 2. Provide the main advantages and disadvantages of each of soft computing method addressed in solving GRN inference problem and make quantitative and qualitative comparisons. 3. Make provisions for implementing one of the computational methods studied and identify some open research questions. For instance, use genetic algorithm (GA) as a heuristic search through the space of qualitative causal networks for inferring the causal relationships between genes using partial correlation. Search for GRN candidate (s) whose scores are the highest possible for network reconstruction. 4. Finally, draw conclusions and future recommendations as to the usability of the studied methods in GRN inference. On the practical level, four simulation experiments on artificial data are undertaken using the causal inference of GRN using GAs, and their results are discussed. In order to reduce complexity and with no loss of generality, for each experiment, the GRN is presented as a directed acyclic graph (DAG) i.e. with no auto-regulation. Indeed, acyclicity is essential for ensuring consistency. For the sake of simplicity, only a part of the obtained GRN is shown. The causal modeling of GRN combined with GA is implemented in MATLAB™ environment. The thesis represents a useful contribution to the field of GRN inference using soft computing methods; it offers a coherent structure of the field and pave the way toward the integration of more powerful methods for GRN inference.
|
