Bayesian causal inference of cell signal transduction from proteomics experiments
Cell signal transduction describes how a cell senses and processes signals from the environment using networks of interacting proteins. In computational systems biology, investigators apply machine learning methods for causal inference to develop causal Bayesian network models of signal transduction from experimental data. Directed edges in the network represent causal regulatory relationships, and the model can be used to predict the effects of interventions to signal transduction. Causal inference approaches applied to proteomics experiments use statistical associations between observed signaling protein concentrations to infer a causal Bayesian network model, but there is no experimental and analysis framework for applying these methods to this experimental context. ^ The goal of this dissertation is to provide a Bayesian experimental design and modeling framework for causal inference of signal transduction. We evaluate how different high-throughput experimental settings affect the performance of algorithms that detect conditional dependence relationships between proteins. We present a Bayesian active learning approach for designing intervention experiments that reveal the direction of causal influence between proteins. Finally, we present a Bayesian model for inferring the parameters of the conditional probability density functions in a causal Bayesian network. The parameters are directly interpretable as a function of the rate constants in the biochemical reactions between interacting proteins. ^ The work pays special attention to analysis of single-cell "snapshot" data such as mass cytometry, where each cell is a multivariate cell-level replicate of signal transduction at a single time point. We also address the role of large-scale bulk experiments such as mass-spectrometry-based proteomics, and small-scale time-course experiments in causal inference.^
Hyonho Chun, Purdue University, Olga Vitek, Purdue University.