| Timetable |
| Download Conference Program |
| Download All Abstracts |
| Zoom Access |
| Corporate Program |
Poster Presentations
Day 3, June 24（Tue.）
Room P (Maesato East, Foyer, Ocean Wing)
- 3P-PM-28（4B-O1-1155）
Prioritizing Candidate Structures in Non-Targeted LC/ESI/HRMS Analysis by Combining Machine Learning Predictions
(1Stockholm Univ. Kemikum, 2Stockholm Univ. ACES)
oWei-Chieh Wang1, Lucas Ferrando Plo1, Chimnaz Emrah1, Amina Souihi1, Pilleriin Peets1, Anneli Kruve1,2
Liquid-chromatography/high-resolution mass spectrometry (LC/HRMS) enables non-targeted screening (NTS) of novel chemical substances; however, the lack of analytical standards and multidimensional data hamper the unequivocal structural annotation. Machine learning (ML) is commonly used in structure annotation from tandem mass spectra (MS2). Still, additional prioritization is required, which is commonly executed by matching empirical analytical properties to increase the confidence of the structure identifications. To date, utilizing ML predictions for different analytical properties of the candidate structures remained unexplored. Moreover, the uncertainty of the predictions is rarely considered.
In the study, different ML models for predicting retention time, ionizability, collision cross-section (CCS), in silico MS spectra similarity, and sodium adduct formation are trained and evaluated for the prioritized process. The preliminary results showed that a strict combination of predictions resulted in the undesired removal of up to 68% of the correct candidates. Therefore, to thoroughly examine the potential of the models, all models will be retrained and complemented with uncertainty estimated with a Mondrian conformal predictive system. The degree of the matching will be combined to estimate the probability of the candidate structures.