MassFormer: uma abordagem de aprendizado profundo para a predição de espectros de massa

Abstract

This work presents a detailed investigation into the use of the MassFormer, a transformer based model, for predicting mass spectra from molecular structures. Utilizing the public dataset MassBank of North America (MoNA), this study explores the efficacy of the MassFormer model, which is refined from a pre-trained Graphormer model. Graphormer, known for its ability to efficiently model structured data like graphs, provides a solid foundation for the MassFormer by capturing the complex interactions between nodes representing atoms in a molecule. Throughout this work, two approaches to data splitting were implemented and compared: a simple random split and a more complex Scaffold-based split. Both strategies produced similar results, demonstrating the robustness of the model under varied conditions. The performance of the MassFormer, although inferior to commercial models due to the limitations in publicly available data, proved promising, especially considering the potential for continuous improvement through community contributions to the MoNA database. This study not only demonstrates the applicability of the MassFormer in academic and research contexts but also highlights the potential for future improvements and expansions in mass spectrum modeling, promoting a deeper understanding and practical application in mass spectrometry.