CSU-EP: Contrastive Learning for Unifying Experimental and Predicted EI-MS Spectra

Spectral library matching is the most common method for compound identification in mass spectrometry (MS), but it is limited by the coverage of experimental libraries. In-silico libraries can expand the searchable chemical space, yet their utility is hindered by the experiment-to-prediction gap arising from the discrepancies between experimental and predicted spectra. Here, we propose CSU-EP, a Contrastive spectral unification framework that bridges this gap with a two-stage training strategy. First, it is self-supervisedly pretrained on 1,883,697 predicted electron ionization (EI) mass spectra using masked peak prediction. Second, it is fine-tuned via contrastive learning to unify representations of paired experimental and predicted spectra. Using this fine-tuned encoder, we assemble 2.24 million NEIMS-predicted spectra into a spectral embedding database (CSU-EP-DB). CSU-EP achieves a Recall@1 of 47.10% on the NIST replib benchmark when searching against CSU-EP-DB, outperforming LLM4MS by 7.28% and FastEI by 12.47%. It also demonstrates excellent ability to identify compounds absent from experimental libraries. In a plasma metabolomics application, incorporating a molecular weight filter boosts Recall@5 to 92.86%. To ensure broad accessibility, a user-friendly web server is deployed. By effectively bridging the experiment-to-prediction gap, CSU-EP establishes a new paradigm for reliable, scalable compound identification using in-silico libraries. The source code, models, and CSU-EP web server are accessible at https://github.com/tingxiecsu/CSU-EP.