ABI Innovation: A Probabilistic Approach to Meta-Analysis of Biological Network Interface

点击翻译按钮获取中文摘要

Genes do not work alone, but rather in an intricate network of interactions to regulate fundamental cellular processes. Biological network inference has become a key analysis tool in modern biology allowing scientists to gain a better understanding of basic biology such as evolution, molecular biology and genetics. Network inference algorithms have played significant roles in understanding gene networks from molecular data but are limited by the large search space of the network structure and data insufficient for statistical power. The goal of this project is to develop a computational framework, based on statistical and machine learning techniques, for effectively integrating multiple heterogeneous data sets to infer gene networks accurately. The developed methods will be applied to answer important biological questions, such as: How is a gene network reshaped in the evolutionary process of yeast species?, How do genes and proteins interact with one another underlying a certain biological trait?, and Which genes contain causative sequence variations that influence important biological traits?. This project is expected to dramatically increase the applicability of network learning algorithms in a wide variety of applications, especially those with smaller sample sizes. The implementation of the developed methods will be made publicly available, which can help many other biologists to study gene networks in their research problems. This project is interdisciplinary in nature and has significant emphasis on interdisciplinary education, through project courses and outreach activities. It will have a long-term effect of advancing the field of biology, by increasing the number of students in computer science inspired to solve biology problems.Network inference from high-throughput biological data has significantly contributed to advancing our knowledge of molecular biology, evolution and genetics. Its major drawback is that, due to the exponentially large search space of the network structure, the sample size provided by a single dataset is often not large enough to obtain valid inference results. However, simply appending datasets from different studies is unlikely to be successful, because in many cases they contain different variables and overly heterogeneous samples. The goal of this project is to develop an innovative probabilistic approach to integrate multiple heterogeneous datasets, by jointly modeling one or more networks represented by these datasets. This project addresses the problem of how to integrate datasets containing heterogeneous samples (Aim 1), different sets of variables (Aim 2) and fundamentally different types of measurements (Aim 3). The developed algorithms in each Aim is applied to the following problems: 1) learning a network model that can represents evolutionary rewiring of gene regulatory networks across Saccharomyces species; 2) learning a joint latent network model that can infer a very high-dimensional network underlying biofilm phenotypes in S. cerevisiae by combining datasets with different variables; and 3) identifying hubs in the gene network from many expression datasets, to guide quantitative trait loci studies in S. cerevisiae.