| Transcriptional regulation is the control of gene expression, involving interactions between protein transcription factors (TFs), transcription factor binding sites, DNA packing material, and associated genes. Many of the key insights in understanding transcriptional regulation have resulted from wet lab experimentation, but these efforts are often laborious and time consuming. The biological complexity of higher organisms, in terms of larger upstream regions and a larger number of TFs, TF binding sites, and interactions between them, complicates research in this field. Computational modeling, therefore, serves as an important complement to experimental efforts.; This thesis furthers the state of the art in transcription factor binding site modeling, particularly within higher organisms, making use of large-scale computing, machine learning/optimization methods, and high throughput experimental data. The work contributes three important biological results: (1) upstream regions of coexpressed human genes are quantitatively related to the repetitive element structures embedded within these upstream regions; (2) a fast, deterministic motif finder applied to human not only finds annotated binding motifs but also finds biologically relevant co-occurring motifs; and (3) a quantitative model of the TF binding site using a neural network recognizes binding sites better than its position weight matrix counterpart by its ability to model positional interdependencies in the binding site. The work was supported by the development of a platform of tools that are well-suited to index-based whole-genome characterizations, which form the basis for very fast algorithms that support direct computation of essentially exact expectation frequencies for large n-mers.; It is hoped that these efforts using large-scale datasets and efficient algorithms will allow further advances in understanding mammalian transcriptional regulation at the sequence level. |
