Novel and Robust Methods for Differential Protein Network Analysis of Proteomics Data in Schizophrenia Research

摘要 生物网络，如蛋白质网络，提供了关于蛋白质如何协同工作的综合视角 正在成为研究精神分裂症等神经精神障碍的重要工具。质量 基于光谱(MS)的蛋白质组学正在迅速发展，现在能够用 更高的灵敏度和吞吐量，这为蛋白质网络提供了关键的数据源，并已 在精神疾病研究中出现了重要的应用。例如，在我们最近的研究中， 突触蛋白共表达网络在精神分裂症患者的听皮质中发生了改变。 鉴于现在已经为微阵列数据开发了各种网络分析方法、方法学 根据蛋白质组数据定制的数据远远落后。此外，这些方法主要集中在成对问题上 在构建网络时忽略其他基因的联合影响的边际相关性，未能 通过中间基因区分因果交互作用和相关性。此外，大多数现有的方法用于 网络测试是基于排列的，如果基于排列的p值为空，则p值可能无效 分配不准确。基于概率图模型的差分网络推理 可取的，因为它通过调整来自所有其他蛋白质和 当分布假设得到满足时，保证是有效和强大的。 我们建议的研究目标是开发、验证和应用新的和稳健的统计方法 从两个流行的蛋白质组平台构建、分析和推断蛋白质网络，即靶向- 和不偏不倚的差别--MS。新的方法将立即应用于正在进行的 匹兹堡大学的精神分裂症项目，以促进新的分析来识别蛋白质变化 对疾病的病理有贡献。首先，我们将基于以下原则开发新的网络建设方法 基于偏相关的方法，它在高斯图形模型(GGM)框架下 在排除其他蛋白质的影响后，量化两个蛋白质之间的相关性，对于蛋白质网络 建筑。然后，我们将在最近的基础上，开发一种新的差分网络推理过程 发展GGM理论和关联推理，形式化地测试网络差异。最后，我们会 使用统计模拟数据和实际数据对所提出的方法进行了全面的验证 具有良好特征的网络相互作用的生物模型。将评估网络的稳健性 用严格设计的重复性实验对正常人死后脑组织标本进行研究 研究对象。总而言之，这项研究的新方法和发现将为 利用蛋白质组学方法设计、分析和验证正在进行的和未来的网络研究 精神障碍，这将大大提高随后科学发现的敏感性和有效性。

Abstract Biological networks such as protein networks provide an integrated perspective on how proteins work together and are becoming important tools to study neuropsychiatric disorders such as schizophrenia. Mass spectrometry (MS) based proteomics are rapidly advancing and are now capable of quantifying proteins with increased sensitivity and throughput, which provide critical data sources for protein networks and have been emerging as important application in the study of psychiatric diseases. For example, in our recent study, the synaptic protein co-expression network was found to be altered in the auditory cortex of schizophrenia patients. Whereas a variety of network analysis methods have now been developed for microarray data, methodologies customized to proteomic data are lagging far behind. In addition, these methods mainly focus on pairwise marginal correlations while ignoring the joint effects from other genes when constructing the network, failing to distinguish causal interactions from correlations via intermediate genes. Moreover, most existing methods for network testing are permutation based, from which the p-values could be invalid if the permutation-based null distribution is inaccurate. The probabilistic graphical model based differential network inference is more desirable as it infers conditional dependency by adjusting for the joint effects from all other proteins and guarantees to be valid and powerful when the distributional assumptions are satisfied. The objective of our proposed research is to develop, validate and apply novel and robust statistical methods to construct, analyze and infer protein networks from two popular proteomic platforms, namely, the targeted- MS and the unbiased differential-MS. The novel methodology will be immediately applied to the ongoing schizophrenia projects at the University of Pittsburgh, to facilitate novel analyses to identify protein alterations contributing to the disease pathology. First, we will develop novel network construction methodology based on a partial-correlation-based approach, which is under the Gaussian Graphical Model (GGM) framework and quantifies the correlation between two proteins after excluding the effects of other proteins, for protein network construction. Then, we will develop a novel differential network inference procedure, based on the recent development of GGM theory and associated inference, to formally test network differences. Finally, we will thoroughly validate the proposed methods using both statistically simulated data and the real data from a biological model with well characterized network interactions. Robustness of the networks will be assessed using rigorously designed replicate experiments with samples from post-mortem brain tissues of normal subjects. In summary, the novel methods and findings from this research will provide critical guidance for the design, analysis and validation of ongoing and future network studies that utilize proteomics approaches in psychiatric disorders, which will greatly improve the sensitivity and validity of the consequent scientific findings.