Novel Chemometric Strategies for Multivariate Measurements in Chemistry and the Omics Sciences

化学测量引领科学发现，现代测量由信号的矢量或矩阵主导，称为多变量数据。这些数据集可以包括光学光谱(可见光、红外、拉曼等)、质谱图、核磁共振谱、色谱数据、传感器阵列信号及其组合，以及反映系统状态(例如生态系统、细胞、化学过程)或其来源(例如食品、药品)的组成数据。使用这些数据的应用非常广泛，包括医疗诊断、药物发现、代谢组学、微观/宏观层面的成像、法医学、威胁检测、食品质量和安全、药品认证、环境监测、工业过程控制等领域。许多分析仪器公司以及知名的IT领导者都在从事研究，以开发智能手持式传感器和工具，用于个性化医学，利用已经到位的微型化多元传感技术。 由于所研究系统的复杂性，从多变量数据中提取相关化学信息通常需要先进的工具。扩展这些工具是这项提案的主要目标，该提案按照两条大线组织。第一是开发更好的数据可视化方法，这是化学中几乎所有多变量分析策略的关键，特别是在“组学”领域。分析多维数据的一种常见方法是将其投影到较低的维度，同时保留重要信息，以便更容易地可视化样本之间的关系(例如，健康与疾病患者、食品的地理来源)。这项工作将开发新的投影方法，使用更有效的标准来保存信号中的信息。第二条研究路线寻求更好地了解与化学测量本身有关的误差，目的是利用这种知识来提高提取的信息的质量。所有的数据分析方法都试图将有意义的化学信号从“噪声”中分离出来，当对后者有更好的理解时，这会更有效。拟议的研究将开发工具来描述特定技术的误差结构(方差和协方差)，并改进利用这一知识的方法，以获得更准确和相关的结果。 这项工作依赖于化学体系和测量的背景知识，以及多变量统计应用方面的专业知识。学生将学习分析化学和先进数据科学的跨学科技能，为进入知识经济做好准备，更加重视大数据和机器学习，特别是与化学系统有关的知识经济。

Chemical measurements lead scientific discovery and modern measurements are dominated by vectors or matrices of signals, referred to as multivariate data. These data sets can consist of optical spectra (visible, infrared, Raman, etc.), mass spectra, NMR spectra, chromatographic data, sensor array signals, and combinations of these, as well as compositional data reflecting the state of a system (e.g. ecosystem, cell, chemical process) or its origins (e.g. foods, drugs). Applications employing such data are pervasive, encompassing areas such as medical diagnostics, drug discovery, metabolomics, imaging at microscopic/macroscopic levels, forensics, threat detection, food quality and security, pharmaceutical authentication, environmental monitoring, industrial process control and many more. Many analytical instrument companies, as well as well-known IT leaders, are engaged in research to develop smart hand-held sensors and tools for personalized medicine that exploit miniaturized multivariate sensing technologies already in place. The extraction of relevant chemical information from multivariate data typically requires advanced tools due to the complexity of the systems under study. The expansion of these tools is the primary goal of this proposal, which is organized along two general lines. The first is the development of better data visualization methods, which are key to virtually all multivariate analysis strategies in chemistry, especially in the “omics” fields. A common approach in the analysis of multidimensional data is to project it into lower dimensions while preserving the important information so that the relationships among samples (e.g. healthy vs diseased patients, geographic origin of foods) can be more easily visualized. This work will develop new projection methods which use more effective criteria to preserve the information in the signals. The second line of research seeks a better understanding of the errors associated with the chemical measurements themselves, with the goal of using that knowledge to improve the quality of information extracted. All data analysis methods attempt to separate meaningful chemical signals from the “noise”, and this is more effective when the latter is better understood. The proposed research will develop tools to characterize the structure of the errors (variance and covariance) for particular techniques and refine methods that exploit this knowledge to obtain more accurate and relevant results. This work relies on contextual knowledge of the chemical systems and measurements, as well as expertise in the application of multivariate statistics. Students will receive interdisciplinary skills in analytical chemistry and advanced data science that will prepare them for entry into a knowledge-based economy with an increased emphasis on big data and machine learning, specifically as it relates to chemical systems.