Advances in microarray and sequencing technology enable the era of personalized medicine. With increasing availability of genomic assays, clinicians have started to utilize genetics and gene expression of patients to guide clinical care. Signatures of gene expression and genetic variation in genes have been associated with disease risks and response to clinical treatment. It is therefore not difficult to envision a future where each patient will have clinical care that is optimized based on his or her genetic background and genomic profiles. However, many challenges exist towards the full realization of the potential personalized medicine. The human genome is complex and we have yet to gain a better understanding of how to associate genomic data with phenotype. First, the human genome is very complex: more than 50 million sequence variants and more than 20,000 genes have been reported. Many efforts have been devoted to genome-wide association studies (GWAS) in the last decade, associating common genetic variants with common complex traits and diseases. While many associations have been identified by genome-wide association studies, most of our phenotypic variation remains unexplained, both at the level of the variants involved and the underlying mechanism. Finally, interaction between genetics and environment presents additional layer of complexity governing phenotypic variation. Currently, there is much research developing computational methods to help associate genomic features with phenotypic variation. Modeling techniques such as machine learning have been very useful in uncovering the intricate relationships between genomics and phenotype. Despite some early successes, the performance of most models is disappointing. Many models lack robustness and predictions do not replicate. In addition, many successful models work as a black box, giving good predictions of phenotypic variation but unable to reveal the underlying mechanism. In this thesis I propose two methods addressing this challenge. First, I describe an algorithm that focuses on identifying causal genomic features of phenotype. My approach assumes genomic features predictive of phenotype are more likely to be causal. The algorithm builds models that not only accurately predict the traits, but also uncover molecular mechanisms that are responsible for these traits. . The algorithm gains its power by combining regularized linear regression, causality testing and Bayesian statistics. I demonstrate the application of the algorithm on a yeast dataset, where genotype and gene expression are used to predict drug sensitivity and elucidate the underlying mechanisms. The accuracy and robustness of the algorithm are both evaluated statistically and experimentally validated. The second part of the thesis takes on a much more complicated system: cancer. The availability of genomic and drug sensitivity data of cancer cell lines has recently been made available. The challenge here is not only the increasing complexity of the system (e.g. size of genome), but also the fundamental differences between cancers and tissues. Different cancers or tissues provide different contexts influencing regulatory networks and signaling pathways. In order to account for this, I propose a method to associate contextual genomic features with drug sensitivity. The algorithm is based on information theory, Bayesian statistics, and transfer learning. The algorithm demonstrates the importance of context specificity in predictive modeling of cancer pharmacogenomics. The two complementary algorithms highlight the challenges faced in personalized medicine and the potential solutions. This thesis detailed the results and analysis that demonstrate the importance of causality and context specificity in predictive modeling of drug response, which will be crucial for us towards bringing personalized medicine in practice.
| Identifer | oai:union.ndltd.org:columbia.edu/oai:academiccommons.columbia.edu:10.7916/D8J109C0 |
| Date | January 2013 |
| Creators | Chen, Bo-Juen |
| Source Sets | Columbia University |
| Language | English |
| Detected Language | English |
| Type | Theses |
Page generated in 0.0023 seconds