Learning Latent Dynamic Bayesian Networks from High Dimensional InterventionEffects and Applications in Systems Biology

点击翻译按钮获取中文摘要

Structure learning of causal Dynamic Bayesian Networks (DBNs) is relevant in many application domains, including medicine and systems biology. Accordingly, different methods have been proposed to learn DBNs from interventional data. So far little attention has been paid to situations in which the network of interest cannot be observed directly, but only through its effects on a layer of observable variables. The need for these models arises in systems biology, where protein signaling networks can often not be measured directly, but the response of thousands of molecular species to perturbations, like gene knock-downs, can be recorded cheaply.The aim of this project is to establish a general framework for structure learning of hidden DBNs from static and time series perturbation data, which is practically applicable to this situation. Our goal will require a realistic probabilistic modeling of signal propagation in a network of hidden variables. Moreover, we will investigate intervention schemes which account for uncertainty and simultaneous targeting of multiple hidden variables. A major task is the development and implementation of an efficient structure learning algorithm for our model, which allows for application to data sets of realistic size that appear, for instance, in systems biology.We will validate our approach in extensive simulations and use it for structure learning of biological networks from experimental data. We expect that our method will significantly improve the automated reconstruction of such networks and thus improve their causal understanding. Our work might impact other scientific domains as well, in which causal inference of unobservable variables is of relevance, e.g. psychology or sociology.