LC-HRMS non-target screening is considered one of the most comprehensive approaches for the detection and identification of chemicals of emerging concern (CECs) in complex environmental samples. These techniques, depending on the data acquisition strategy, generate a set of MS2 or pseudo MS2 spectra that are used for structural elucidation via spectral library matching. There are several library search algorithms used for the structural elucidation based on the measured MS2 spectra. These algorithms use one or more parameters (e.g., direct and reverse match scores; mass error; and/or number of matched fragments) to assess the similarity between the measured spectra and the library entries. One major limitation of such algorithms is that they are not able provide the confidence levels associated with each identified measured spectrum. As a consequence, analysts must determine confidence levels manually and based on their prior knowledge of the samples and the initial hypothesis. To overcome this limitation, we developed a random forest filter (RFF) that extracts nine different parameters from the measured and the library spectra in order to determine the confidence levels associated with the structural elucidation based on the spectral matching. This algorithm was trained using 48,000 semi-synthetic spectra generated using 12,000 MassBank-EU spectra (i.e., experimental spectra). Four different portions (i.e., 10, 40, 60, 90%) of the experimental spectra were sampled to mimic the false negative ion cases. In the next step, random noise (mass error) ranging between -30 mDa and 30 mDa was added to each fragment in the modified spectra to imitate the cases of false positive ions. These semi-synthetic spectra were then used for training, validation, and testing of the RFF. The validated RFF was able to correctly identify 82% of the test spectra. Finally, we used the validated RFF for the chemical characterization of 48 wastewater influent samples. In total 36 CECs were correctly identified in those samples and 11 were confirmed via target analysis.

Confidence assessment of LC-HRMS chemical identification employing machine learning

ACS Spring Meeting 2021

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.