Filtern
Erscheinungsjahr
- 2019 (2)
Dokumenttyp
- Posterpräsentation (2) (entfernen)
Sprache
- Englisch (2)
Referierte Publikation
- nein (2)
Schlagworte
Organisationseinheit der BAM
Nontargeted high-resolution mass spectrometry (HRMS) is widely used for small molecule screening in biotic or abiotic samples. However, HRMS approaches like metabolomics or environmental nontarget screening currently still lack confidence in chemical annotation, i.e. computational structure assignment to all measured mass signals. As a crucial step within the annotation pipeline, molecular weight inference (MWI) deduces a compound’s intact mass from diagnostic mass differences between MS1 peaks, allowing precise database queries in subsequent steps. As the common practice of considering all possible ionization products such as adducts, multimers, multiple charges etc. in MWI suffers from high false positive rates, we aimed at selecting candidate ionization products in a chemically sensitive way. Generally, electrospray ionization produces different types of adducts depending on chromatographic system and sample matrix, necessitating application-specific optimization for optimum MWI performance. To avoid, however, the tedious and potentially biased manual data curation connected to optimization, we established an R-based workflow for automating this task. The workflow consists of two parts. Part 1 creates an MS1 spectral library by performing peak detection, spectral deconvolution and target peak assignment based on density estimation. Part 2 analyzes ion relationships within the library and returns a list of detected ionization products ranked by their frequency. We applied the workflow to a commercial 634-compound library that was acquired for two chromatographic methods (reverse phase, RP; hydrophilic liquid interaction chromatographic, HILIC) and the two ESI modes (positive, negative). As expected, different frequency distributions of ionization products were found for the two chromatographies. Interestingly, however, some of the differences were expected in terms of solvent chemistries (e.g. [M+NH4]+ in ammonium formate-buffered HILIC) while others indicated more complex ion competition (e.g. abundant [M+K]+, [M+2K-H]+ in HILIC). This demonstrated the relevance of this empirical approach. We further show that MWI accuracy clearly benefitted from derived optimized adduct lists – by adding filters or weighting terms – and present FDR calculations supporting this observation. We conclude that chemistry-aware compound annotation based on the combination of high-throughput library acquisition and statistical analysis holds significant potential for further improvements in nontargeted small molecule HRMS.
Comparing Nontargeted LC-MS Methods by Co-visualizing Linear Dynamic Range and Chemical Coverage
(2019)
INTRODUCTION
Biological and environmental samples contain thousands of small molecule species that all vary in chemical properties and concentration range. Identifying and quantifying all these chemical entities remains a long-term goal in metabolomics and related systems approaches. Due to its broad selectivity, nontargeted LC-MS is usually the method of choice for broad chemical screening. Optimizing nontargeted LC-MS methods, however, is less straightforward than for targeted methods where sensitivity, specificity, linearity etc. serve as well-established performance criteria. We therefore investigated linear dynamic range (LDR) and chemical classification as alternative performance criteria to guide nontargeted method development.
EXPERIMENTAL METHODS
LDR was defined as the linear portion of a feature’s response curve over multiple concentration levels. Comparing the LDR of features across methods can be expected to be significantly more robust than comparing signal intensities for a single concentration. To determine LDR for all features, a computational workflow was implemented in the R programming language. For estimating the linear portion of a curve, several mathematical approaches including linear, non-linear and piecewise linear regression were evaluated. Chemical classification was based on ClassyFire, which computes chemical classes for a given structure. To avoid false classifications for incorrectly annotated compounds, we took the following statistical approach. For each compound, multiple likely annotation hypotheses were derived using a recently described workflow[2]. All annotation hypotheses were submitted to ClassyFire and obtained classifications were ranked by frequency. The most frequently suggested class was kept for further analysis. Finally, LDR and chemical classes were visualized together on a molecular network, which was constructed using the well-established MS/MS similarity approach.
RESULTS AND DISCUSSION
For technical validation of the workflow, several hundred curve fits obtained from the different regression models were reviewed visually. Piecewise linear regression performed the most reliably with respect to the heterogeneous curve shapes of ‘real-life’ features. Validation of chemical classification was performed against a compound library, which showed that 90% of ~450 library compounds were correctly classified using the described approach. Two liquid chromatography methods (HILIC, RPC) as well as two electrospray ionization variants (low/high-temperature ESI) applied to urinary metabolomics were exemplarily studied to test the workflow. Molecular network visualization indicated that of all analytical setups, HILIC/high temperature ESI performed best in terms of high LDR achieved over a wide range of compound classes. Despite one order of magnitude lower sensitivity, HILIC/low temperature ESI showed similar chemical coverage, except for organic nitrogen compounds that were underrepresented compared to high-temperature ESI. Both RPC setups were inferior to the HILIC setups in terms of high-LDR features, supporting previous findings for the given matrix. The higher relative representation of benzenoids and lipids in RPC demonstrated that the workflow successfully captured expected selectivity differences between chromatographies.
CONCLUSION
When comparing nontargeted LC-MS methods for optimization purposes, ideally all available quantitative and qualitative information should be integrated. The present workflow follows this idea. Visualizing LDR and chemical classes of all features on a molecular network quickly indicated differences in method selectivity that were otherwise difficult to spot. As an automated approach, it is easily applied to repeated optimization steps, enabling effective optimization strategies.