Causal Graph Methods for Heterogeneous Data and Application to MUC1 Cancer Vaccine Response

The rapid advancement of biomedical technology has ushered in a new era of research that integrates powerful analytical approaches with traditional experimental lab work. To tackle the increasingly complex multi-modal datasets, numerous innovative new computational algorithms have been developed. These algorithms can accurately perform a range of tasks, from providing comprehensive insights and building complex models of biological systems to designing detailed personalized treatment plans and streamlining clinician workflow. However, despite their success, many of these algorithms are limited in their ability to infer causal relationships, a core component of scientific research. To bridge this gap, researchers have started to build new or adapt existing algorithms to incorporate causal strategies. In this dissertation, I will present several strategies to improve causal search approaches and demonstrate how we can use these and other ML approaches on the MUC1 cancer vaccine data. First, I will present how to use dimensionality reduction and latent factor models to improve the performance of causal search algorithms. Building upon this idea, we propose several ways to incorporate longitudinal data into existing causal models using time aggregation metrics. Next, I will introduce two new approaches to edge selection that improve efficiency of constraint-based causal search algorithms. After the computational work, the dissertation presents experimental work with the MUC1 cancer vaccine. I showcase how we can use causal search and machine learning algorithms to facilitate the experimental process and final data analysis.