Program

Poster Presentations

Day 3, June 12（Fri.）　　Room P (5F 501+502)

3P-12

PDF

MassBank in silico: Construction of a Structural Annotation Analysis Pipeline for Low-Molecular-Weight Compounds

(¹UOsaka, ²Kyushu Univ., ³Kyushu Univ., ⁴Tokyo Univ. Agr. Tech., ⁵Niigata Univ., ⁶AIST, ⁷UOsaka)
^oTaihei Torigoe¹, Masatomo Takahashi¹, Heravizadeh Omidreza³, Taiki Hirose⁴, Kazuki Ikeda², Kohta Nakatani⁵, Yuki Soma⁶, Takeshi Bamba^2,3, Fumio Matsuda⁷, Hiroshi Tsugawa⁴, Yoshihiro Izumi¹

Although recent LC/MS-based metabolomics can detect thousands of metabolic features, 80–90% of them remain unidentified. To address this issue, we developed an ensemble-based structural annotation pipeline that integrates multiple structural annotation tools to improve the transparency and accuracy of MassBank spectra. Additionally, we built a unified-HILIC/AEX retention time prediction model in Python using scikit-learn and evaluated it with the SRM 1950 plasma sample. The model was developed with three elements in mind: a high-quality training dataset, calculation of 12,420 molecular descriptors, and systematic descriptor selection and model construction. The pipeline was implemented in Python 3.11.5 on Ubuntu (WSL2). Retention time-based filtering eliminated approximately 50% of candidate structures as false positives in plasma data evaluation, demonstrating its effectiveness in reducing annotation errors. We further integrated CFM-ID, MetFrag, MS-FINDER, and SIRIUS into the ensemble framework. Future work will validate the practical utility of the pipeline for retention time prediction and expand it into a new framework called MassBank in silico, which combines the complementary strengths of existing annotation tools.

---