Data Processing Thresholds for Abundance and Sparsity and Missed Biological Insights in an Untargeted Chemical Analysis of Blood Specimens for Exposomics

Abstract

Background: An untargeted chemical analysis of bio-fluids provides semi-quantitative data for thousands of chemicals for expanding our understanding about relationships among metabolic pathways, diseases, phenotypes and exposures. During the processing of mass spectral and chromatography data, various signal thresholds are used to control the number of peaks in the final data matrix that is used for statistical analyses. However, commonly used stringent thresholds generate constrained data matrices which may under-represent the detected chemical space, leading to missed biological insights in the exposome research.Methods: We have re-analyzed a liquid chromatography high resolution mass spectrometry data set for a publicly available epidemiology study (n = 499) of human cord blood samples using the MS-DIAL software with minimally possible thresholds during the data processing steps. Peak list for individual files and the data matrix after alignment and gap-filling steps were summarized for different peak height and detection frequency thresholds. Correlations between birth weight and LC/MS peaks in the newly generated data matrix were computed using the spearman correlation coefficient.Results: MS-DIAL software detected on average 23,156 peaks for individual LC/MS file and 63,393 peaks in the aligned peak table. A combination of peak height and detection frequency thresholds that was used in the original publication at the individual file and the peak alignment levels can reject 90% peaks from the untargeted chemical analysis dataset that was generated by MS-DIAL. Correlation analysis for birth weight data suggested that up to 80% of the significantly associated peaks were rejected by the data processing thresholds that were used in the original publication. The re-analysis with minimum possible thresholds recovered metabolic insights about C19 steroids and hydroxy-acyl-carnitines and their relationships with birth weight.Conclusions: Data processing thresholds for peak height and detection frequencies at individual data file and at the alignment level should be used at minimal possible level or completely avoided for mining untargeted chemical analysis data in the exposome research for discovering new biomarkers and mechanisms.

IDSL.CCDB: a database for exploring inter-chemical correlations in metabolomics and exposomics datasets

“…We will also add text mining 23 , chemoinformatics and other bioinformatics resources 81 which can help in interpreting the inter-chemical correlations in an automatic manner. For the untargeted metabolomics dataset, we need to re-generate those datasets using MS-DIAL 87 or other software which can generate comprehensive data matrices so that the chances of missing any biological insights will be minimal 21 .…”

“…For the MTBLS1684 study, a comprehensive untargeted data matrix was generated using the MS-DIAL software. Parameters were described in the previous report 21 . We applied a log2 transformation before computing the correlation statistics for this study.…”

“…A systematic understanding and cataloging of targeted and untargeted analyses of small molecules measured in biospecimens is needed, as such datasets are critical to translate the information gathered from exposomics and metabolomics projects 20 . These key datasets include: 1) population-scale biomonitoring surveys; 2) targeted analysis of multiple analytes in hypothesis-driven studies (typically 10-100); and 3) untargeted analysis of thousands of chemicals using a high-resolution mass spectrometry instrument 21, 22 . They cover key high priority exposome chemicals 23 including carcinogens 24, 25 , endocrine disrupters 26 and industry chemicals 27 .…”

Commentary: Data Processing Thresholds for Abundance and Sparsity and Missed Biological Insights in an Untargeted Chemical Analysis of Blood Specimens for Exposomics

“…We read with interest the paper by Barupal et al on the effect of untargeted metabolomics data filtering thresholds that was recently published in ( 1 ). The authors used publicly available liquid chromatography-mass spectrometry data of 499 newborn cord blood samples.…”

Benchmarking feature quality assurance strategies for non-targeted metabolomics

“…Reproducing results on different tools and optimizing parameters has been recognized as one of the main challenges in non-targeted data processing, impacting data interpretation and ultimately study outcome. [9][10][11][12][13][14][15] The emerging unease regarding the under-utilization of data 12,13 , calls for evaluation strategies that enable benchmarking of different tools, algorithms, and parameter choices on a regular basis. Undeniably, validation of LC-HRMS NPP poses several challenges.…”