Discovering gene interactions by image analysis and by reverse engineering genetic networks.

摘要

The study of gene expression patterns in developing embryos forms a major part of understanding biological systems. Such gene expression patterns offer insights into the regulation and differentiation processes that occur during development. Spatially similar expression patterns from both wild-type and mutant embryos are used to predict and infer gene regulation. From individual gene interactions, gene regulatory networks are built and the structure and dynamics of biological systems are studied.; Currently, tens of thousands of fruit fly embryo gene expression pattern images are available. To find spatially similar patterns within this enormous dataset in order to understand gene regulations, computational methods that are based only on image features are needed.; Existing techniques for such automated gene expression pattern matching include the Binary Feature Vector (BFV) representation, which is able to retrieve biologically significant images with respect to a query gene expression pattern. In order to find ways to enhance the performance of the BFV representation, this dissertation offers a new metric for similarity measurement and also proposes the separation of multi-domain images into many individual images with single domains. Moreover, the use of Invariant Moment Vectors (IMV), which have been successfully applied in natural image processing, for analyzing fruit fly embryo gene expression pattern images is studied.; Another major contribution of this dissertation is to systems biology for understanding the structure of gene regulatory networks. Conflicting Interactions (CIs) are a pair of contradictory interactions, one positive and one negative, from the same regulating gene to the same target gene. None of the existing directed graph approaches for reconstructing regulatory networks models Conflicting Interactions (CIs). This work proposes a modified approach that includes CIs and also determines the effect of their incorporation on the inferred network. In general, the effect of a given gene is not observed more than two levels downstream from where it is transcribed. This was modeled by restricting path traversals to those of length two or less in the network retrieval process. A study of such a restriction, which could potentially decrease the run time, is also given.