Advanced Computational Approaches for NMR Data-mining

摘要 基于核磁共振波谱的代谢组学是一种有效的鉴定方法 代谢扰动，报告不同的生物状态和样本类型。与质量相比 核磁共振在几分钟内提供了健壮和高度重复性的定量数据，这是 使其非常适合一线临床诊断。虽然已知代谢组提供了一种 细胞、组织和生物体生物状态的瞬时快照，核磁共振在临床上的应用 繁琐的数据分析阻碍了实践。主要挑战包括数据的高维性， 信号重叠、共振频率的可变性(化学位移)、信号形状不理想和低 低浓度代谢产物的信噪比(SNR)。现有的方法无法解决这些问题 挑战和样本分析非常耗时，需要手动完成，并且需要相当多的知识 核磁共振波谱。稀疏方法在机器学习领域的最新发展和加速 用于高维问题的凸优化以及基于核的空间聚类在以下方面显示出前景 使我们能够克服这些挑战，实现完全自动化、独立于操作员的分析。我们 正在开发两种新的、强大的和自动化的算法，它们利用这些最新的发展 机器学习。在目标1中，我们描述了用于自动识别和量化 带注释的代谢物，与化学位移、低信噪比和信号形状变异性无关。在目标2中，我们 描述‘SPA-STOCSY’自动从头识别未知、非... 带注释的代谢物。基于大量的初步数据，我们建议对这些算法进行评估 对体模、生物样本和临床样本的敏感性、特异性、稳定性和抗噪性，比较 将它们添加到当前的方法。我们将通过实验2D核磁共振、添加和 质谱学。拟议的努力将产生新的核磁共振分析软件，用于发现这两个 注释和非注释代谢物，大大提高了核磁共振的准确性和重复性 分析。这种分析能力将改变现有的基于核磁共振的代谢组学范式，并提供 这是对目前基于质谱学的方法的更强大的补充。这种方法，一旦彻底 经过验证，将使核磁共振在生物医学、制药和营养领域获得广泛的应用 研究和临床医学。

ABSTRACT Nuclear magnetic resonance spectroscopy (NMR)-based metabolomics is a powerful method for identifying metabolic perturbations that report on different biological states and sample types. Compared to mass spectrometry, NMR provides robust and highly reproducible quantitative data in a matter of minutes, which makes it very suitable for first-line clinical diagnostics. Although the metabolome is known to provide an instantaneous snap-shot of the biological status of a cell, tissue, and organism, the utilization of NMR in clinical practice is hindered by cumbersome data analysis. Major challenges include high-dimensionality of the data, overlapping signals, variability of resonance frequencies (chemical shift), non-ideal shapes of signals, and low signal-to-noise ratio (SNR) for low concentration metabolites. Existing approaches fail to address these challenges and sample analysis is time-consuming, manually done, and requires considerable knowledge of NMR spectroscopy. Recent developments in the field of sparse methods for machine learning and accelerated convex optimization for high dimensional problems, as well as kernel-based spatial clustering show promise at enabling us to overcome these challenges and achieve fully automated, operator-independent analysis. We are developing two novel, powerful, and automated algorithms that capitalize on these recent developments in machine learning. In Aim 1, we describe ‘NMRQuant’ for automated identification and quantification of annotated metabolites irrespective of the chemical shift, low SNR, and signal shape variability. In Aim 2, we describe ‘SPA-STOCSY’ for automated de-novo identification of molecular fragments of unknown, non- annotated metabolites. Based on substantial preliminary data, we propose to evaluate these algorithms' sensitivity, specificity, stability, and resistance to noise on phantom, biological, and clinical samples, comparing them to current methods. We will validate the accuracy of analyses by experimental 2D NMR, spike-in, and mass spectrometry. The proposed efforts will produce new NMR analytical software for discovery of both annotated and non-annotated metabolites, substantially improving accuracy and reproducibility of NMR analysis. Such analytical ability would change the existing paradigm of NMR-based metabolomics and provide an even stronger complement to current mass spectrometry-based methods. This approach, once thoroughly validated, will enable NMR to reach wide network of applications in biomedical, pharmaceutical, and nutritional research and clinical medicine.