- Title
- Transcription factor binding specificity and occupancy : elucidation, modelling and evaluation
- Creator
- Kibet, Caleb Kipkurui
- ThesisAdvisor
- Machanick, Philip
- Subject
- Transcription factors
- Subject
- Transcription factors -- Data processing
- Subject
- Motif Assessment and Ranking Suite
- Date
- 2017
- Type
- Thesis
- Type
- Doctoral
- Type
- PhD
- Identifier
- vital:21185
- Identifier
- http://hdl.handle.net/10962/6838
- Description
- The major contributions of this thesis are addressing the need for an objective quality evaluation of a transcription factor binding model, demonstrating the value of the tools developed to this end and elucidating how in vitro and in vivo information can be utilized to improve TF binding specificity models. Accurate elucidation of TF binding specificity remains an ongoing challenge in gene regulatory research. Several in vitro and in vivo experimental techniques have been developed followed by a proliferation of algorithms, and ultimately, the binding models. This increase led to a choice problem for the end users: which tools to use, and which is the most accurate model for a given TF? Therefore, the first section of this thesis investigates the motif assessment problem: how scoring functions, choice and processing of benchmark data, and statistics used in evaluation affect motif ranking. This analysis revealed that TF motif quality assessment requires a systematic comparative analysis, and that scoring functions used have a TF-specific effect on motif ranking. These results advised the design of a Motif Assessment and Ranking Suite MARS, supported by PBM and ChIP-seq benchmark data and an extensive collection of PWM motifs. MARS implements consistency, enrichment, and scoring and classification-based motif evaluation algorithms. Transcription factor binding is also influenced and determined by contextual factors: chromatin accessibility, competition or cooperation with other TFs, cell line or condition specificity, binding locality (e.g. proximity to transcription start sites) and the shape of the binding site (DNA-shape). In vitro techniques do not capture such context; therefore, this thesis also combines PBM and DNase-seq data using a comparative k-mer enrichment approach that compares open chromatin with genome-wide prevalence, achieving a modest performance improvement when benchmarked on ChIP-seq data. Finally, since statistical and probabilistic methods cannot capture all the information that determine binding, a machine learning approach (XGBooost) was implemented to investigate how the features contribute to TF specificity and occupancy. This combinatorial approach improves the predictive ability of TF specificity models with the most predictive feature being chromatin accessibility, while the DNA-shape and conservation information all significantly improve on the baseline model of k-mer and DNase data. The results and the tools introduced in this thesis are useful for systematic comparative analysis (via MARS) and a combinatorial approach to modelling TF binding specificity, including appropriate feature engineering practices for machine learning modelling.
- Format
- 194 pages, pdf
- Publisher
- Rhodes University, Faculty of Science, Computer Science
- Language
- English
- Rights
- Kibet, Caleb Kipkurui
- Hits: 1163
- Visitors: 1220
- Downloads: 225
Thumbnail | File | Description | Size | Format | |||
---|---|---|---|---|---|---|---|
View Details | SOURCE1 | 8 MB | Adobe Acrobat PDF | View Details |