A Computational Framework for Learning from Complex Data: Formulations, Algorithms, and Applications

Abstract

Many real-world processes are dynamically changing over time. As a consequence, the observed complex data generated by these processes also evolve smoothly. For example, in computational biology, the expression data matrices are evolving, since gene expression controls are deployed sequentially during development in many biological processes. Investigations into the spatial and temporal gene expression dynamics are essential for understanding the regulatory biology governing development. In this dissertation, I mainly focus on two types of complex data: genome-wide spatial gene expression patterns in the model organism fruit fly and Allen Brain Atlas mouse brain data. I provide a framework to explore spatiotemporal regulation of gene expression during development. I develop evolutionary co-clustering formulation to identify co-expressed domains and the associated genes simultaneously over different temporal stages using a mesh-generation pipeline. I also propose to employ the deep convolutional neural networks as a multi-layer feature extractor to generate generic representations for gene expression pattern in situ hybridization (ISH) images. Furthermore, I employ the multi-task learning method to fine-tune the pre-trained models with labeled ISH images. My proposed computational methods are evaluated using synthetic data sets and real biological data sets including the gene expression data from the fruit fly BDGP data sets and Allen Developing Mouse Brain Atlas in comparison with baseline existing methods. Experimental results indicate that the proposed representations, formulations, and methods are efficient and effective in annotating and analyzing the large-scale biological data sets.