| The gene expression pattern images of Drosophila capture the spatial and temporal dynamics of gene expression, and they provide valuable information for studying gene interactions during development. The current practice of manually annotating these biological images with spatial keywords and temporal stages does not scale with the increasing number of images, hindering the pace of biological discoveries. In this dissertation, I propose computational approaches for the automated annotation and analysis of Drosophila gene expression pattern images. In the spatial keyword annotation, I develop a novel feature representation scheme called the sparse coding to overcome the limitations of existing methods. I also develop a set of multi-label and multi-task learning models to integrate multiple types of image features and exploit the correlations among different spatial keywords. In temporal stage annotation, I propose to extract morphological texture features by applying a filter bank followed by a local neighborhood pooling. I then propose to apply a group sparsity regularized model for the annotation. Based on the spatial and temporal expression information captured by the gene expression pattern images, I propose a regression-based model for learning the global gene interaction network. The developed computational methods are evaluated using the gene expression data from the FlyExpress database in comparison with baseline methods. Experimental results demonstrate that the developed representations, formulations, and algorithms are effective and efficient in learning from the genome-wide biological data. |
