Transferable retention time prediction for Liquid Chromatography-Mass Spectrometry-based metabolomics

代谢物鉴定仍然是代谢组学研究的主要瓶颈。液相色谱-质谱（LC-MS）是目前非靶向代谢组学中应用最多的分析技术。目前，在一个典型的非靶向实验中，只有不到10%的光谱可以被注释。因此，迫切需要改进代谢物鉴定工具。而质量本身不能识别分子，串联质谱产生碎片光谱，可用于结构阐明。最近，代谢组学社区已经开发并越来越多地使用计算机方法，允许在PubChem和ChemSpider等分子结构数据库中进行搜索。这样的结构数据库比任何光谱库都要大很多数量级，因此，分子结构的覆盖范围要广得多。但即使采用串联质谱进行鉴定，也会产生大量的伪鉴定。为了提高鉴定质量，必须报告两个独立的参数，如化学标准品的质量和保留时间。如今，保留时间主要用于鉴别管道的后期，并且主要基于与化学参考标准的比较。但是，如果在早期阶段使用保留时间，特别是对于计算机方法，显然是有益的；在这里，我们可以过滤候选人，或者更好的是，根据比较预测和观察到的留存时间来修改候选人的分数。本项目旨在利用可转移的保留时间预测，更好地利用保留时间在基于LC-MS的非靶向代谢组学中鉴定小生物分子。预测将基于两步方法。首先，机器学习将用于预测给定分子结构的保留顺序数；培训将基于从公共可用数据集中广泛收集的保留时间数据，以及参考代谢物标准的系统内部测量。与其质量相反，保留时间不是代谢物的特征，而是代谢物、固定相和流动相结合的特征。因此，我们将利用所采用的色谱系统的特性以及代谢物的分子指纹进行机器学习。在第二步中，将保留顺序编号映射到保留时间，使用已知和已识别的物质作为映射的锚点。保留顺序和保留时间预测将用于过滤假阳性反应对，并应用于秀丽隐杆线虫次级代谢的独立生物数据集。所有整理和获取的数据，用于预测保留顺序和保留时间的开源软件将免费提供给代谢组学社区。最后，将保留时间预测整合到CSI:FingerID评分中，以提高其代谢物识别率。

Metabolite identification still represent the major bottleneck in metabolomics. Liquid Chromatography-Mass Spectrometry (LC-MS) is the currently most employed analytical technique in untargeted metabolomics. Currently, less than 10% of spectra in a typical untargeted experiment can be annotated. Therefore, there is a strong need for improved tools for metabolite identification. While mass alone cannot identify molecules, tandem MS yields fragmentation spectra which can be used for structural elucidation. Recently, in silico approaches have been developed and are increasingly used by the metabolomics community, that allow to search in molecular structure databases such as PubChem and ChemSpider. Such structure databases are many orders of magnitude larger than any spectral library and, hence, have a much wider coverage of molecular structures. But even identification by tandem MS will result in numerous spurious identifications. To improve identification quality, two independent parameters, e.g. mass and retention time of a chemical reference standard have to be reported. Today, retention time is mainly used at a later stage of the identification pipeline, and mainly based on comparison with chemical reference standards. However, it would clearly be beneficial if retention times were used at an early stage, in particular for in silico methods; here, we could filter candidates or, even better, modify the scores of candidates based on comparing predicted and observed retention times.This project aims to make better use of retention times for the identification of small biomolecules in LC-MS based untargeted metabolomics, using transferable retention time prediction. Prediction will be based on a two-step approach. First, Machine Learning will be used to predict retention order numbers for give molecular structures; training will be based on an extensively curated collection of retention time data from public available datasets, as well as systematic in-house measurements for reference metabolite standards. In contrast to its mass, retention time is not a feature of a metabolite, but of the combination of metabolite, stationary and mobile phase. Therefore, we will use properties of the employed chromatographic system in addition to molecular fingerprints of metabolites for machine learning. In the second step, retention order numbers will be mapped to retention times, using known and identified substances as anchors of the mapping. Retention order and retention time prediction will be used to filter false positive reaction pairs, and applied to an independent biological dataset from C. elegans secondary metabolism.All curated and acquired data, open-source software for prediction of retention order and retention times will be made freely available to the metabolomics community. Finally, retention time prediction will be integrated into the CSI:FingerID scoring in order to improve its metabolite identification rates.