Chemie und Prozesstechnik
Filtern
Erscheinungsjahr
- 2023 (2)
Dokumenttyp
Sprache
- Englisch (2)
Referierte Publikation
- nein (2)
Organisationseinheit der BAM
Non-target (NT) mass spectrometry techniques play a crucial role in metabolomics applications, particularly in areas such as environmental safety. Soft ionization methods, such as electrospray ionization (ESI), are commonly employed due to their ability to generate spectra containing molecular ions, aiding in the identification of unknown peaks. However, ESI may fail to ionize certain compounds effectively, resulting in their absence in NT approaches. In contrast, hard ionization methods like electron impact (EI) can ionize a wide range of compounds but often lack information about the molecular ion. Although extensive databases of EI spectra exist, their nominal mass resolution (NR) limits their utility for modern high-resolution (HR) EI mass spectrometers. Here, the often-applied conversion of HR spectra to nominal mass leads to the loss of unique characteristics, i.e. by grouping distinct masses into more ambiguous nominal masses.
Our study aims to (i) quantify the negative impact of such a binning approach and (ii) develop a machine learning (ML) tool capable of enhancing existing nominal mass spectral libraries. In the initial phase, we employed the RECETOX Exposome HR-[EI+]-MS library to assess the influence of HR spectra on identification. We compared the dot product of each spectrum against all others, utilizing bin sizes of 0.001 and 1 Da. The difference in dot product between the second-best candidate and the query spectrum (the best candidate, with a score of 1) was calculated for both HR and nominal mass spectra. Subsequently, we explored the application of ML techniques to predict HR spectra from nominal mass spectra using the before mentioned dataset.
Preliminary findings demonstrate the potential of high-resolution spectral libraries. As anticipated, HR spectra consistently exhibited lower similarity scores for the second candidate. This observation likely stems from the high redundancy and resulting ambiguity associated with nominal masses. Even after eliminating spectra containing multiple HR masses mapped to the same nominal mass, the aforementioned trend persisted. Furthermore, initial investigations into ML have revealed its ability to predict up to 40% of HR masses within a 10 mDa precision window.
Gas Chromatography coupled with Electron Ionisation Mass Spectrometry (GC-EI-MS) is a well-established technique which, in combination with spectral libraries, has the potential to identify compounds in a sample. Nevertheless, most libraries are dominated by spectra with nominal mass resolution which does not allow to make full use of the data generated by modern high-resolution instruments. The production of high-resolution spectral library is time consuming and expensive, while in silico fragmentation tools that are capable of generating HR mass spectra are still too computationally intensive. We explored the alternative of using ML models to upscale existing spectral libraries. The models were trained with spectra from the RECETOX metabolome (DOI: 10.5281/zenodo.5483564), RECETOX Exposome (DOI: 10.5281/zenodo.4471216), and MassBank (10.5281/zenodo.7436394) HR-GC-MS libraries. The model was used to generate a synthetic library which was compared with a synthetic library simulated with CFM-ID (DOI: 10.1021/acs.analchem.6b01622).