|dc.description.abstract||Access to large genome-wide biological datasets has now enabled computational researchers to tackle long-standing questions in Biomedicine through the lens of Machine Learning (ML) and Artificial Intelligence (AI). The potential benefits of such computational approaches to biological research are immense. For example, efficient, and yet interpretable, machine learning models of disease/drug response/phenotype can impact our life at both personal and social levels. However, heterogeneity is found at multiple scales in biology, manifested as the context-specificity of biological processes. This context-specific heterogeneity poses a major challenge to ML models. Even though context-specific models are often trained, this is mostly done without the benefit of mechanistic insights about the biological processes being modeled, and as such do not help improve our biological understanding.
This dissertation addresses these challenges and their limitations by: a) designing appropriate features and ML models motivated by the current biological hypothesis at hand, b) building pipelines to analyze multiple context-specific models together, and c) developing data integration and imputation methods to address the problems of insufficient and missing data.
The first project studies loss of methylation or hypo-methylation in large blocks causing aberrant gene activity, a well-known phenomenon in cancer. To find the associated markers, I designed a classification model of hypo-methylated block boundaries and non-boundaries in colon cancer.
The second project models binding of transcription factor (TF) to specific DNA element to the genome, one of the principal components of gene regulation. Since condition specificity of TF binding is not yet well understood, this dissertation examines a design of cell type-specific models for transcription factor (TF) binding using ChIPSeq data. A meta-analysis pipeline, called TRISECT, is applied for multiple TF binding models to understand heterogeneity of cell specificity across those models.
Next, models for breast cancer metastasis using gene expression data are discussed. In breast cancer metastasis, the affinity towards distant tissues called secondary tissues has not been comprehended. Therefore, going beyond mere discriminatory models, I propose another meta-analysis pipeline, MONTAGE intending to understand the organotropism of breast cancer metastasis across secondary tissues.
Building ML models can be hindered by the data size, specially, for rare diseases. Therefore, by necessity, molecular data have been merged across multiple studies, and across multiple technical platforms which has vulnerability of so called batch effects diluting the actual biological signal. Existing methods are not capable of removing multi-variate confounding artifacts leading to inaccurate models. To circumvent this issue, this dissertation examines a deep learning based technique (deepSavior) which ‘translates’ the gene expression profile from samples of one technical platform to another platform.
To summarize, this dissertation makes three distinct contributions, a) designing effective ML model to explore the determinants of cancer-associated hypomethlation, b) designing meta-analysis pipelines to compare multiple related but context-specific ML models to understand heterogeneous relations among biological processes, and b) developing new method to overcome the data integration and imputation challenges.||en_US