Machine learning analysis of tandem mass spectra

描述（由申请人提供）：蛋白质是活细胞中的主要功能分子，串联质谱法提供了高通量研究蛋白质的最有效手段。该提案旨在使用机器学习，统计学和自然语言处理领域的最先进方法来提高我们理解大型串联质谱数据集的能力。该提议的核心是一种被称为动态贝叶斯网络的概率模型，它允许我们对复杂的序列数据集进行有效和准确的推理。该建模框架利用了自然语言处理和语音识别领域的大量相关工作。许多先前的工作尚未被计算生物学家利用，因此该提议代表了跨学科的有价值的交叉受精。更具体地说，该项目采用一系列合作的动态贝叶斯网络来共同模拟整个质谱实验。现有的大多数质谱数据分析方法倾向于将实验分析划分为一系列小的独立子任务，而本文提出的统一模型综合考虑了所有可用数据。因此，这种方法可以利用光谱之间和数据的各个维度之间有价值的依赖关系。动态贝叶斯网络还为从观察数据和定性专家知识的组合中执行推理提供了严格的框架。该项目分为五个目标，每个目标都涉及一种特定类型的质谱实验。这些实验包括(1)使用标准质谱法鉴定给定复杂生物样品中的所有蛋白质；(2)使用改进的方案鉴定蛋白质，其中质谱仪以系统而非数据依赖的方式对数据进行采样，目的是鉴定低丰度蛋白质；(3)定量测定生物样品内或样品间蛋白质的相对丰度；(4)鉴定翻译后修饰蛋白或含有序列变异的蛋白；(5)对一组特定的蛋白质进行靶向定量，例如感兴趣途径中的蛋白质或蛋白质生物标志物。本提案中描述的方法有可能极大地提高我们从高通量霰弹枪蛋白质组学实验中得出结论和制定假设的能力。例如，上面描述的实验可以识别参与基本疾病过程的蛋白质，识别以前未知的蛋白质同种异构体，或量化蛋白质对环境压力源或疾病状态的反应。

DESCRIPTION (provided by applicant): Proteins are the primary functional molecules in living cells, and tandem mass spectrometry provides the most efficient means of studying proteins in a high-throughput fashion. The proposal aims to use state-of-the-art methods from the fields of machine learning, statistics and natural language processing to improve our ability to make sense of large tandem mass spectrometry data sets. The core of the proposal is a type of probabilistic model, known as a dynamic Bayesian network that allows us to reason efficiently and accurately about complex sequential data sets. This modeling framework leverages a large body of related work from the fields of natural language processing and speech recognition. Much of this prior work has not yet been exploited by computational biologists, so the proposal represents a valuable cross-fertilization across disciplines. More specifically, this project employs a collection of cooperating dynamic Bayesian networks to model jointly an entire mass spectrometry experiment. Relative to most existing methods for analyzing mass spectrometry data, which tend to divide the analysis of an experiment into a series of small independent subtasks, the proposed unified model jointly, considers all of the available data. This approach can thus exploit valuable dependencies among spectra and along various dimensions of the data. Dynamic Bayesian networks also provide a rigorous framework for performing inference from a combination of observed data and qualitative expert knowledge. The project is divided into five aims, each of which concerns a particular type of mass spectrometry experiment. These experiments involve (1) identifying all of the proteins in a given complex biological sample using a standard mass spectrometry protocol; (2) identifying proteins using a modified protocol in which the mass spectrometer samples the data in a systematic, rather than data-dependent, fashion, with the goal of identifying lower abundance proteins; (3) quantifying the relative abundance of proteins within or between biological samples; (4) identifying post-translational modified proteins or proteins that contain sequence variation; and (5) performing targeted quantification of a specified set of proteins, such as proteins in a pathway of interest or protein biomarkers. The methods described in this proposal have the potential to dramatically improve our ability to draw conclusions from and formulate hypotheses on the basis of high-throughput shotgun proteomics experiments. Experiments like the ones described above can, for example, identify proteins involved in fundamental disease processes, identify previously unknown protein isoforms, or quantify the re- sponses of proteins to environmental stressors or disease states.