Statistical Methods for Inferring Latent Factors in Biological Networks

Abstract

The development of high-throughput sequencing technologies has significantly improved our understanding of the genetic basis of disease, allowing for rapid sequencing of entire genomes and identification of disease-associated genetic variants. However, despite the abundance of sequencing data, explaining the incidence of disease phenotypes remains challenging due to the involvement of complex interactions among multiple genes. Network biology has emerged as a powerful approach for investigating biological systems as networks, enabling researchers to model interactions between components in a comprehensive and systematic manner. A fundamental question in network biology is the estimation of latent factors, such as protein or pathway activity, which play a critical role in driving disease phenotypes but are not easily measurable in the laboratory. My research focuses on utilizing large-scale sequencing data and statistical modeling to estimate these essential latent variables, with the ultimate goal of enhancing our understanding of complex diseases. In the first project, I present a novel algorithm - TIGER to improve the estimation of transcription factor (TF) activities. A transcription factor is a type of protein that can bind to DNA and alter the expression of certain genes. By integrating context-specific gene expression profiles with static transcriptional interaction databases into a Bayesian matrix factorization framework, TIGER obtains more accurate estimates of TF activity. The TIGER package is available on GitHub (https://github.com/cchen22/TIGER). My second project investigates various approaches for evaluating the statistical significance of groups of genes (network “communities”) that work together to drive disease. The main challenge is finding an appropriate null distribution for large and complex transcription networks. To address this issue, we proposed two new methods. Firstly, we developed a rapid solution for generating samples of the bipartite weighted configuration model. Secondly, we created CRANE, a novel algorithm that is able to mimic the distribution of networks derived from subsampling data. Our results demonstrate that building statistical methodology into network analysis can improve the identification of biologically relevant pathways. In my third project, I propose a novel change point detection method for transcription networks that can detect the dramatic module structure change in the time series transcription network. Our approach extends the stochastic block model to handle multilayer bipartite weighted networks, which I term the global stochastic block model (GSBM). Notably, when applied to a simulation study, my method can identify changes in community structure even if the mean node degree or centrality remains constant. Moreover, by applying our method to cellular quiescence data, we demonstrate its efficacy in detecting significant time points at which phenotypic transitions occur.