The explosion of high throughput genomic data in recent years has already altered our view of the extent and complexity of biology. Technologically specific features, heterogeneous data structures and massive sample sizes present great challenges and opportunities to develop novel statistical methodologies in computational biology. This dissertation presents three Bayesian modeling methods in high throughput genomic data analysis.

In chapter 2, we develop a model-based gene expression query algorithm built under the Bayesian model selection framework. This algorithm is capable of detecting co-expression profiles under a subset of samples/experimental conditions. In addition, it allows linearly transformed expression patterns to be recognized and is robust in the presence of sporadic outliers in the data. Our simulation studies suggest that this method outperforms existing query tools. When we apply this new method to the Escherichia coli microarray compendium data, it identifies a majority of known regulons, as well as novel potential target genes of numerous key transcription factors.

In chapter 3, we introduce a novel computational algorithm named Hybrid Motif Sampler (HMS), specifically designed for transcription factor binding sites (TFBS) motif discovery in ChIP-Seq data. HMS incorporates sequencing depth information to aid motif identification, allows intra-motif dependency to describe more accurately the underlying motif pattern and combines stochastic sampling and deterministic search to accelerate the computation process. Simulation studies demonstrate favorable performance of HMS compared to other existing methods. When applying HMS to real ChIP-Seq datasets, we find that the accuracy of existing TFBS motif patterns can be significantly improved.

In chapter 4, we propose a spatial Poisson regression model to provide a portrait of base-level sequencing depth in RNA-Seq data. The model utilizes two random effects to explain the spatial correlation and the non-spatial variation and incorporates GC content effects into the mean structure for better fitting. Both simulation study and real data analysis demonstrate that this method can capture local genomic features that affect coverage depth, and therefore, offers improved quantification of the true underlying expression levels.

The research in this dissertation demonstrates that Bayesian modeling methods have achieved great success and have the potential to accelerate biomedical research.