Statistical Methods for Inferring Gene-Phenotype Associations Using Omic Data from Gene Knockout and Human Phenotype Studies

点击翻译按钮获取中文摘要

Project Summary The phase 1 of the Molecular Phenotypes of Null Alleles in Cells (MorPhiC) consortium will produce a catalog of molecular and cellular phenotypes for null alleles of ~1000 human genes using in vitro cellular systems. These rich resources will allow us to study the gene functions in several multicellular systems that often model early human development. The impact of a gene knockout on complex human phenotypes can be highly dependent on the corresponding cell type, cell stage, and tissue microenvironment. Therefore, to generalize the insights from MorPhiC studies to in vivo settings, we need to harmonize MorPhiC resources and the molecular/cellular phenotypes of appropriate cell types or tissues, by a flexible and robust computational framework. We aim to achieve this goal by two complementary approaches. First, we will develop a dynamic gene regulatory network named moDAG: multi-omic Directed Acyclic Graph. MoDAG combines multi-omic data from MorPhiC and other studies and the state-of-the-art statistical methods to estimate a gene-regulatory network. MoDAG models cell types characterized by genome-wide epigenetic or gene expression data. It also accounts for signals from tissue microenvironment by modeling a set of signaling proteins. MoDAG can be used to predict the effect of gene knock out in the in vitro cellular systems, and thus help prioritize the genes to be targeted in future MorPhiC studies. Second, we propose a biologically informed deep learning method named as SDAN: Supervised Deep learning with gene Annotation. SDAN combines molecular phenotype of gene knockout with gene annotation to identify gene sets associated with gene knock out. Gene sets provide more robust characterization of gene knockout than individual genes and thus are more generalizable to different cell types or tissues. The gene annotation used by SDAN is gene-gene interaction network that can be modified according to relevant cell types or tissues. Finally, we apply these two methods to predict the phenotypic outcomes of gene knockouts and assess the association between gene knockouts and human phenotypes. Our computational framework bridges MorPhiC’s resource with accumulating omic data in various human cell types and tissues and provide effective solutions to generate new insights or hypothesis for future studies.