Abstract
Over the past decades, mass spectrometry has served as a fundamental technique for molecular identification in the field of metabolomics, widely applied to the analysis and characterization of biomolecules. However, the complexity of experimental conditions and the structural similarities among compounds pose significant challenges for accurate identification. Recent advancements in deep learning have offered new solutions to address these challenges, particularly demonstrating great potential in generating high-quality spectral embeddings. In this study, we propose a novel method, SpecEmbedding, which leverages a transformer encoder architecture and employs replicated spectra of compounds as positive samples, trained under a supervised contrastive learning framework. By mapping complex mass spectra into low-dimensional vector representations, SpecEmbedding significantly enhances the comparability between spectra, thereby improving identification accuracy. We trained SpecEmbedding on the GNPS training subset and evaluated its performance on the GNPS test subset, the MoNA data set, and the MTBLS1572 data set. Experimental results show that SpecEmbedding achieves a Top-1 hit ratio of 81.73% on the GNPS test subset, substantially outperforming MSBERT (77.81%) and DreaMS (71.90%). The source code for this study is publicly available at https://github.com/sword-nan/SpecEmbedding. A web service is provided at https://huggingface.co/spaces/xp113280/SpecEmbedding.