BiComp: Nonparametric Bayesian Models for Genetic Variations and Their Associations to Diseases and Population Demography

更好地了解细菌生理过程可能对感染性疾病的治疗方法和生物技术中的代谢工程应用产生巨大影响。在这个项目中，dr。分别来自加州克莱蒙特凯克研究所（Keck Graduate Institute）和纽约州布法罗纽约州立大学（State University of New York）的戴维·l·怀尔德（David L. Wild）和马修·j·比尔（Matthew J. Beal）提议建立时间序列数据的统计模型，以期利用复杂的贝叶斯方法，从基因表达和代谢物浓度的原始测量中“逆向工程”生物体复杂的遗传调控网络。Wild和Beal将把他们的技术应用于一个理想的实验系统：大肠杆菌对酸胁迫的反应，这使得致病性大肠杆菌能够通过胃和胃肠道的酸性环境存活下来。他们将与实验家博士合作。英国伯明翰大学的Francesco Falciani和Mark Viant说，他们将提供这种细菌致病性和非致病性菌株的数据。然后，Wild和Beal的模型所做出的预测可以在实验室中进行测试和探索。功能基因组学技术的最新进展使生物学家能够前所未有地测量复杂生物有机体的内部运作。使用微阵列表达谱，现在可以在一个单一的生物实验中测量成千上万个基因的表达水平，以时间序列的形式进行几天。与此形成鲜明对比的是，仅仅在十年前，生物学家要测量不止一两个精心挑选的基因的表达是相当不寻常的。与高通量基因表达方法一样，“代谢组学”的新技术为以数百种代谢物的浓度形式测量更多信息打开了大门，这些代谢物在研究中的复杂细胞过程中也起着至关重要的作用。如此庞大的数据量挑战了传统的分析方法，尤其是考虑到时间因素时，因为现在必须考虑某些基因如何从实验的一个时间点调节到下一个时间点的其他基因的表达。博士的一个关键成分。怀尔德和比尔的模型包含了有助于解释观测到的测量的相关结构的“隐藏因素”。这些因素可能与实验中未捕获的未测量数量相对应，并且经常减少直接基因对基因依赖的数量，使最终的网络对生物学家来说更容易解释。一个自然的问题出现了：应该使用多少隐藏因素来解释观察到的数据中的依赖性？这个问题可以通过贝叶斯模型选择来回答，贝叶斯模型选择是机器学习和统计学中使用的一个有充分根据的原则，用于在不同复杂性的模型之间进行选择。他们的模型还使用了一种称为自动相关性测定的技术来进一步简化模型，以便在最终模型中只保留那些参与过程的基因和代谢物。贝叶斯框架的另一个优点是，从文献或商业数据库中获得的关于已知网络连接和交互的现有信息可以包含在模型中。建模过程的输出是对哪些基因调控网络是可信的或不可信的概率推算。这些概率可以用来设计未来针对特定基因的生物实验，以证实模型的计算机预测，或者只是探索一个相对未知的网络。

Carnegie Mellon University Better understanding of the processes involved in the physiology of bacteria can potentially have tremendous impact on both therapeutic approaches to infectious diseases and metabolic engineering applications in biotechnology.In this project, Drs. David L. Wild and Matthew J. Beal, of the Keck Graduate Institute in Claremont, California and the State University of New York in Buffalo, New York, respectively, are proposing to build statistical models of time series data, with a view to leveraging sophisticated Bayesian methods to "reverse-engineer" an organism's complex genetic regulatory networks from the raw measurements of gene expression and metabolite concentration.Drs. Wild and Beal will apply their techniques to an ideal experimental system: the response of the bacterium E coli to acid stress, which enables pathogenic E. coli to survive passage through the acidic environment of the stomach and gastro-intestinal tract. They will collaborate with experimentalists Drs. Francesco Falciani and Mark Viant at the University of Birmingham, UK, who will provide data from both pathogenic and non-pathogenic strains of this bacterium. Predictions made by Wild and Beal's models can then be tested and explored back in the laboratory.Recent advances in functional genomics technologies have given biologists unprecedented access to measurements of the inner workings of complex biological organisms. Using microarray expression profiling, it is now possible to measure the expression levels of tens of thousands of genes in just a single biological experiment, conducted over several days in the form of a time series. Contrast this to the situation only ten years ago when it was rather unusual for a biologist to measure the expression of more than just one or two carefully chosen genes. As well as high-throughput gene expression methods, the new technology of "metabolomics" has opened the door to measuring even more information in the form of the concentration of hundreds of metabolites that are also crucial players in the complex cellular processes under study.This overwhelming amount of data challenges traditional methods of analysis, especially when one considers the element of time, because now one must consider how certain genes regulate the expression of other genes from one time point in the experiment to the next. A key ingredient in Drs. Wild and Beal's models is the inclusion of "hidden factors" that help to explain the correlation structure of the observed measurements. These factors may correspond to unmeasured quantities that were not captured during the experiment and often reduce the number of direct gene-to-gene dependencies, leaving the resulting networks much more interpretable for the biologist. A natural question arises: how many hidden factors should be used to account for the dependencies in the observed data? This is answered by employing Bayesian model selection, a well-founded principle used in machine learning and statistics to choose between models of differing complexities. Their models also use a technique called Automatic Relevance Determination to further simplify the models so that only those genes and metabolites that are participating players in the process are retained in the final model.Another advantage of the Bayesian framework is that existing information about known network connections and interactions, derived from the literature or commercial databases, can be included in the model. The output of the modeling procedure is a probabilistic reckoning of which genetic regulatory networks are plausible or not. These probabilities can be used to design future biological experiments targeted at specific genes, with a view to corroborating the model's in silico predictions or to simply probe a relatively uncharted network.