Division of Biomedical Engineering, the Hong Kong University of Science and Technology, Clear Water Bay, Hong Kong, China.
J Proteome Res. 2013 Jul 5;12(7):3223-32. doi: 10.1021/pr400080b. Epub 2013 Jun 6.
With the rapid accumulation of data from shotgun proteomics experiments, it has become feasible to build comprehensive and high-quality spectral libraries of tandem mass spectra of peptides. A spectral library condenses experimental data into a retrievable format and can be used to aid peptide identification by spectral library searching. A key step in spectral library building is spectrum denoising, which is best accomplished by merging multiple replicates of the same peptide ion into a consensus spectrum. However, this approach cannot be applied to "singleton spectra," for which only one observed spectrum is available for the peptide ion. We developed a method, based on a Bayesian classifier, for denoising peptide tandem mass spectra. The classifier accounts for relationships between peaks, and can be trained on the fly from consensus spectra and immediately applied to denoise singleton spectra, without hard-coded knowledge about peptide fragmentation. A linear regression model was also trained to predict the number of useful "signal" peaks in a spectrum, thereby obviating the need for arbitrary thresholds for peak filtering. This Bayesian approach accumulates weak evidence systematically to boost the discrimination power between signal and noise peaks, and produces readily interpretable conditional probabilities that offer valuable insights into peptide fragmentation behaviors. By cross validation, spectra denoised by this method were shown to retain more signal peaks, and have higher spectral similarities to replicates, than those filtered by intensity only.
随着 shotgun 蛋白质组学实验所产生的数据的快速积累，构建全面、高质量肽段串联质谱谱库已成为可能。谱库将实验数据压缩成可检索的格式，并可通过谱库检索来辅助肽鉴定。谱库构建的关键步骤是谱图去噪，这最好通过将同一肽离子的多个重复合并到共识谱中来实现。然而，这种方法不能应用于“单峰谱”，因为对于肽离子只有一个可观测的谱图。我们开发了一种基于贝叶斯分类器的肽串联质谱去噪方法。该分类器考虑了峰之间的关系，可以从共识谱中实时训练，并立即应用于去噪单峰谱，而无需对肽片段有硬编码的知识。我们还训练了一个线性回归模型来预测谱图中有用的“信号”峰的数量，从而避免了对峰过滤进行任意阈值的需要。这种贝叶斯方法系统地积累微弱的证据，以提高信号峰和噪声峰之间的区分能力，并生成易于解释的条件概率，为肽片段化行为提供有价值的见解。通过交叉验证，与仅通过强度过滤的谱图相比，这种方法去噪的谱图保留了更多的信号峰，并且与重复谱图的相似度更高。
