Development of Kernel-based and ensemble machine learning methods for binary/multi-class processing of environmental LC-HRMS/MS data and multiset modeling of fused data

该项目的主要目标如下：（a）开发一个有效的数据挖掘协议，用于在高效液相色谱-轨道阱-质谱/质谱中使用DIA（AIF）模式评价复杂的环境数据集。将多集/张量分解的化学计量学算法纳入非靶向分析（NTA）的潜力将评估使用AIF增加复杂数据集的MS 2光谱覆盖范围的工作流程。这些张量分解可用于映射低维空间中的数据，并在所有测量模式中分离变量。这些过程将在LC-MS 1和LC-MS 2数据的不同复杂程度的水样上进行，以分离模式进行，然后以融合模式进行（工作包1）。(b)采用支持向量机和RF，在他们的原始形式，以及结合递归特征消除（RFE），空间/时间的地表水样品的LC-HRMS数据的二进制/多分类。目标将是污染物的优先级/排名，并为高精度的预测模型找到最佳的特征子集。设计这些分类器是为了应对以前报告的问题，如对地表水流中的纵向污染模式进行分类49，有限的水样复制和样本大小对时空环境研究中优先污染物的再现性和稳定性的影响。上述方法的输出（初始特征排名、选定污染物子集和分类准确度）将与PLS-DA的可变投影重要性（VIP）和选择性比（SR）提供的污染物优先级列表进行比较。因此，将对空间/时间变化的地表水样本进行更全面的评估。因此，可以确定与特定时间框架有关的污染物的优先次序并加以识别。最后，通过根据高等级污染物的真实标准制定定量方法，将提供关于地表水和环境解释中潜在危险污染物（作为“混合物接触”）相对浓度的可靠信息，然后可用于进一步研究（环境监测和风险评估研究）。(Work包装2）。

The main objectives of this project are as follows: (a) Developing an efficient data mining protocol for the evaluation of complex environmental data sets using DIA (AIF) mode in HPLC-Orbitrap-MS/MS. Due to the high value of accurate-mass MS/MS spectra for elucidation and confirmation purposes, the potential of chemometrics algorithms of multiset/tensor decomposition incorporated into the non-targeted analysis (NTA) workflow to increase MS2 spectral coverage for complex datasets using AIF will be evaluated. These tensor decompositions can be used to map data in low dimensional spaces and to separate variables in all modes of measurement. These processes will be performed on water samples with different degrees of complexity on the LC-MS1 and LC-MS2 data, in the separate mode and then the fused mode (Work Package 1).(b) Employing SVM and RF, in their original form as well as in combination with recursive feature elimination (RFE), for binary/multi-classification of LC-HRMS data of spatial/temporal surface water samples. The objective will be pollutant prioritization/ranking and finding the best subset of features for a predictive model with high accuracy. These classifiers are designed to be exploited in response to previously reported issues such as classifying longitudinal pollution patterns in surface water streams49, the impact of limited replication of water samples and sample size on reproducibility and stability of prioritized pollutants in spatiotemporal environmental studies. The output of the mentioned methods (initial feature ranking, subset of selected pollutants, and classification accuracy) will be compared with the prioritized list of pollutants provided by variable importance on projection (VIP) and selectivity ratio (SR) for PLS-DA. So, a more comprehensive evaluation of spatially/temporally varying surface water samples will be carried out. As a result, pollutants that link to a specific temporal frame can be prioritized and identified. Ultimately, by developing a quantitative method based on authentic standards for highly ranked pollutants, reliable information about the relative concentrations of potentially hazardous pollutants (as "mixture exposure”) in the set of surface water and environmental interpretation will be provided, which can then be utilized for further research (environmental monitoring and risk assessment studies). (Work Package 2).