The 10th Asia-Oceania Mass Spectrometry Conference (AOMSC2025) - organized by the Mass Spectrometry Society of Japan

Abstract

Timetable
Download Conference Program
Download All Abstracts
Zoom Access
Corporate Program

Oral Sessions

Day 3, June 24(Tue.) 14:40-14:55

Room B (Maesato Center)

  • 3B-O2-1440
  • PDF

Improving structure elucidation using machine learning for non-target analysis using Gas Chromatograph-Mass Spectrometer

(JEOL)
oAyumi Kubo, Azusa Kubota, Kenji Nagatomo, Masaaki Ubukata

Large-scale chemical databases such as PubChem contains over 100 million compounds, whereas commercial EI mass spectral libraries support several hundred thousand compounds, resulting in a significant numerical gap. Therefore, the EI mass spectra of most compounds remain unknown. As a result, database searches often result in non-identification or misidentification of compounds. To solve this problem, we have developed a qualitative analysis method that combines a machine learning model, soft ionization method, and accurate mass analysis for the purpose of estimating the structure of such compounds. We have improved the accuracy of structural formula estimation using this method by making some improvements. The first improvement is narrowing down the structural formula candidates using the retention index (RI). Before comparing EI mass spectra, we added a process to compare the RI predicted from the structural formula and the measured RI and exclude candidates with large RI differences. The second improvement is to improve the EI mass spectrum prediction accuracy by optimizing hyperparameters and feature vectors. We investigated the effect of each parameter using an orthogonal array and determined the optimal level for each.